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/Optional.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/PriorityQueue.h"
#include "llvm/ADT/STLExtras.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/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/DebugLoc.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/NoFolder.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"
#include "llvm/IR/Verifier.h"
#include "llvm/InitializePasses.h"
#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/LoopUtils.h"
#include "llvm/Transforms/Vectorize.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <iterator>
#include <memory>
#include <set>
#include <string>
#include <tuple>
#include <utility>
#include <vector>
 
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");
 
cl::opt<bool> RunSLPVectorization("vectorize-slp", cl::init(true), cl::Hidden,
                                  cl::desc("Run the SLP vectorization passes"));
 
static cl::opt<int>
    SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
                     cl::desc("Only vectorize if you gain more than this "
                              "number "));
 
static cl::opt<bool>
ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden,
                   cl::desc("Attempt to vectorize horizontal reductions"));
 
static cl::opt<bool> ShouldStartVectorizeHorAtStore(
    "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
    cl::desc(
        "Attempt to vectorize horizontal reductions feeding into a store"));
 
static cl::opt<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)"));
 
static cl::opt<int>
MaxStoreLookup("slp-max-store-lookup", cl::init(32), cl::Hidden,
    cl::desc("Maximum depth of the lookup for consecutive stores."));
 
/// 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 Look-ahead heuristic goes through the users of the bundle to calculate
// the users cost in getExternalUsesCost(). To avoid compilation time increase
// we limit the number of users visited to this value.
static cl::opt<unsigned> LookAheadUsersBudget(
    "slp-look-ahead-users-budget", cl::init(2), cl::Hidden,
    cl::desc("The maximum number of users to visit while visiting the "
             "predecessors. This prevents compilation time increase."));
 
static cl::opt<bool>
    ViewSLPTree("view-slp-tree", cl::Hidden,
                cl::desc("Display the SLP trees with Graphviz"));
 
// Limit the number of alias checks. The limit is chosen so that
// it has no negative effect on the llvm benchmarks.
static const unsigned AliasedCheckLimit = 10;
 
// Another limit for the alias checks: The maximum distance between load/store
// instructions where alias checks are done.
// This limit is useful for very large basic blocks.
static const unsigned MaxMemDepDistance = 160;
 
/// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling
/// regions to be handled.
static const int MinScheduleRegionSize = 16;
 
/// 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) {
  return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() &&
         !Ty->isPPC_FP128Ty();
}
 
/// \returns True if the value is a constant (but not globals/constant
/// expressions).
static bool isConstant(Value *V) {
  return isa<Constant>(V) && !isa<ConstantExpr>(V) && !isa<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 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();
  // TODO: This should check for generic Instruction::isCommutative(), but
  //       we need to confirm that the caller code correctly handles Intrinsics
  //       for example (does not have 2 operands).
  return false;
}
 
/// Checks if the given value is actually an undefined constant vector.
static bool isUndefVector(const Value *V) {
  if (isa<UndefValue>(V))
    return true;
  auto *C = dyn_cast<Constant>(V);
  if (!C)
    return false;
  if (!C->containsUndefOrPoisonElement())
    return false;
  auto *VecTy = dyn_cast<FixedVectorType>(C->getType());
  if (!VecTy)
    return false;
  for (unsigned I = 0, E = VecTy->getNumElements(); I != E; ++I) {
    if (Constant *Elem = C->getAggregateElement(I))
      if (!isa<UndefValue>(Elem))
        return false;
  }
  return true;
}
 
/// 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
/// We convert this initially to something 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
/// %1 = insertelement <4 x i8> poison, i8 %x0, i32 0
/// %2 = insertelement <4 x i8> %1, i8 %x3, i32 1
/// %3 = insertelement <4 x i8> %2, i8 %y1, i32 2
/// %4 = insertelement <4 x i8> %3, i8 %y2, i32 3
/// %5 = mul <4 x i8> %4, %4
/// %6 = extractelement <4 x i8> %5, i32 0
/// %ins1 = insertelement <4 x i8> poison, i8 %6, i32 0
/// %7 = extractelement <4 x i8> %5, i32 1
/// %ins2 = insertelement <4 x i8> %ins1, i8 %7, i32 1
/// %8 = extractelement <4 x i8> %5, i32 2
/// %ins3 = insertelement <4 x i8> %ins2, i8 %8, i32 2
/// %9 = extractelement <4 x i8> %5, i32 3
/// %ins4 = insertelement <4 x i8> %ins3, i8 %9, i32 3
/// ret <4 x i8> %ins4
/// InstCombiner transforms this into a shuffle and vector mul
/// Mask will return the Shuffle Mask equivalent to the extracted elements.
/// TODO: Can we split off and reuse the shuffle mask detection from
/// TargetTransformInfo::getInstructionThroughput?
static Optional<TargetTransformInfo::ShuffleKind>
isFixedVectorShuffle(ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask) {
  const auto *It =
      find_if(VL, [](Value *V) { return isa<ExtractElementInst>(V); });
  if (It == VL.end())
    return None;
  auto *EI0 = cast<ExtractElementInst>(*It);
  if (isa<ScalableVectorType>(EI0->getVectorOperandType()))
    return None;
  unsigned Size =
      cast<FixedVectorType>(EI0->getVectorOperandType())->getNumElements();
  Value *Vec1 = nullptr;
  Value *Vec2 = nullptr;
  enum ShuffleMode { Unknown, Select, Permute };
  ShuffleMode CommonShuffleMode = Unknown;
  Mask.assign(VL.size(), UndefMaskElem);
  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 None;
    auto *Vec = EI->getVectorOperand();
    // We can extractelement from undef or poison vector.
    if (isUndefVector(Vec))
      continue;
    // All vector operands must have the same number of vector elements.
    if (cast<FixedVectorType>(Vec->getType())->getNumElements() != Size)
      return None;
    if (isa<UndefValue>(EI->getIndexOperand()))
      continue;
    auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand());
    if (!Idx)
      return None;
    // Undefined behavior if Idx is negative or >= Size.
    if (Idx->getValue().uge(Size))
      continue;
    unsigned IntIdx = Idx->getValue().getZExtValue();
    Mask[I] = IntIdx;
    // 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 None;
    }
    if (CommonShuffleMode == Permute)
      continue;
    // If the extract index is not the same as the operation number, it is a
    // permutation.
    if (IntIdx != 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;
}
 
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 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;
}
 
/// \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,
                                       unsigned BaseIndex = 0) {
  // 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]);
  unsigned Opcode = cast<Instruction>(VL[BaseIndex])->getOpcode();
  unsigned AltOpcode = Opcode;
  unsigned AltIndex = BaseIndex;
 
  // Check for one alternate opcode from another BinaryOperator.
  // TODO - generalize to support all operators (types, calls etc.).
  for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) {
    unsigned InstOpcode = cast<Instruction>(VL[Cnt])->getOpcode();
    if (IsBinOp && isa<BinaryOperator>(VL[Cnt])) {
      if (InstOpcode == Opcode || InstOpcode == AltOpcode)
        continue;
      if (Opcode == AltOpcode && isValidForAlternation(InstOpcode) &&
          isValidForAlternation(Opcode)) {
        AltOpcode = InstOpcode;
        AltIndex = Cnt;
        continue;
      }
    } else if (IsCastOp && isa<CastInst>(VL[Cnt])) {
      Type *Ty0 = cast<Instruction>(VL[BaseIndex])->getOperand(0)->getType();
      Type *Ty1 = cast<Instruction>(VL[Cnt])->getOperand(0)->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 (InstOpcode == Opcode || InstOpcode == AltOpcode)
      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[0]->getType();
  for (int i = 1, e = VL.size(); i < e; i++)
    if (VL[i]->getType() != Ty)
      return false;
 
  return true;
}
 
/// \returns True if Extract{Value,Element} instruction extracts element Idx.
static 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 None;
    return CI->getZExtValue();
  }
  ExtractValueInst *EI = cast<ExtractValueInst>(E);
  if (EI->getNumIndices() != 1)
    return None;
  return *EI->idx_begin();
}
 
/// \returns True if in-tree use also needs extract. This refers to
/// possible scalar operand in vectorized instruction.
static bool InTreeUserNeedToExtract(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);
    for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) {
      if (hasVectorInstrinsicScalarOpd(ID, i))
        return (CI->getArgOperand(i) == Scalar);
    }
    LLVM_FALLTHROUGH;
  }
  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.
static void addMask(SmallVectorImpl<int> &Mask, ArrayRef<int> SubMask) {
  if (SubMask.empty())
    return;
  if (Mask.empty()) {
    Mask.append(SubMask.begin(), SubMask.end());
    return;
  }
  SmallVector<int> NewMask(SubMask.size(), UndefMaskElem);
  int TermValue = std::min(Mask.size(), SubMask.size());
  for (int I = 0, E = SubMask.size(); I < E; ++I) {
    if (SubMask[I] >= TermValue || SubMask[I] == UndefMaskElem ||
        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(SmallVectorImpl<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);
  }
}
 
namespace llvm {
 
static void inversePermutation(ArrayRef<unsigned> Indices,
                               SmallVectorImpl<int> &Mask) {
  Mask.clear();
  const unsigned E = Indices.size();
  Mask.resize(E, UndefMaskElem);
  for (unsigned I = 0; I < E; ++I)
    Mask[Indices[I]] = I;
}
 
/// \returns inserting index of InsertElement or InsertValue instruction,
/// using Offset as base offset for index.
static Optional<int> getInsertIndex(Value *InsertInst, unsigned Offset) {
  int Index = Offset;
  if (auto *IE = dyn_cast<InsertElementInst>(InsertInst)) {
    if (auto *CI = dyn_cast<ConstantInt>(IE->getOperand(2))) {
      auto *VT = cast<FixedVectorType>(IE->getType());
      if (CI->getValue().uge(VT->getNumElements()))
        return UndefMaskElem;
      Index *= VT->getNumElements();
      Index += CI->getZExtValue();
      return Index;
    }
    if (isa<UndefValue>(IE->getOperand(2)))
      return UndefMaskElem;
    return None;
  }
 
  auto *IV = cast<InsertValueInst>(InsertInst);
  Type *CurrentType = IV->getType();
  for (unsigned I : IV->indices()) {
    if (auto *ST = dyn_cast<StructType>(CurrentType)) {
      Index *= ST->getNumElements();
      CurrentType = ST->getElementType(I);
    } else if (auto *AT = dyn_cast<ArrayType>(CurrentType)) {
      Index *= AT->getNumElements();
      CurrentType = AT->getElementType();
    } else {
      return None;
    }
    Index += I;
  }
  return Index;
}
 
/// Reorders the list of scalars in accordance with the given \p Order and then
/// the \p Mask. \p Order - is the original order of the scalars, need to
/// reorder scalars into an unordered state at first according to the given
/// order. Then the ordered scalars are shuffled once again in accordance with
/// the provided mask.
static void reorderScalars(SmallVectorImpl<Value *> &Scalars,
                           ArrayRef<int> Mask) {
  assert(!Mask.empty() && "Expected non-empty mask.");
  SmallVector<Value *> Prev(Scalars.size(),
                            UndefValue::get(Scalars.front()->getType()));
  Prev.swap(Scalars);
  for (unsigned I = 0, E = Prev.size(); I < E; ++I)
    if (Mask[I] != UndefMaskElem)
      Scalars[Mask[I]] = Prev[I];
}
 
namespace slpvectorizer {
 
/// Bottom Up SLP Vectorizer.
class BoUpSLP {
  struct TreeEntry;
  struct ScheduleData;
 
public:
  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)
      : F(Func), SE(Se), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), AC(AC),
        DB(DB), DL(DL), ORE(ORE), Builder(Se->getContext()) {
    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)
              .getFixedSize();
 
    if (MinVectorRegSizeOption.getNumOccurrences())
      MinVecRegSize = MinVectorRegSizeOption;
    else
      MinVecRegSize = TTI->getMinVectorRegisterBitWidth();
  }
 
  /// Vectorize the tree that starts with the elements in \p VL.
  /// Returns the vectorized root.
  Value *vectorizeTree();
 
  /// Vectorize the tree but with the list of externally used values \p
  /// ExternallyUsedValues. Values in this MapVector can be replaced but the
  /// generated extractvalue instructions.
  Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues);
 
  /// \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 = None);
 
  /// 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,
                 ArrayRef<Value *> UserIgnoreLst = None);
 
  /// 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 = {});
 
  /// Clear the internal data structures that are created by 'buildTree'.
  void deleteTree() {
    VectorizableTree.clear();
    ScalarToTreeEntry.clear();
    MustGather.clear();
    ExternalUses.clear();
    for (auto &Iter : BlocksSchedules) {
      BlockScheduling *BS = Iter.second.get();
      BS->clear();
    }
    MinBWs.clear();
    InstrElementSize.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.
  Optional<OrdersType> findReusedOrderedScalars(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.
  /// \param TopToBottom If true, include the order of vectorized stores and
  /// insertelement nodes, otherwise skip them.
  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 DataLayout &DL) 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() 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
  };
 
  /// 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 DataLayout &DL;
    ScalarEvolution &SE;
    const BoUpSLP &R;
 
    /// \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]);
    }
 
    // 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;
    /// Loads from reversed memory addresses, e.g. load(A[i+1]), load(A[i]).
    static const int ScoreReversedLoads = 3;
    /// 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;
    /// User exteranl to the vectorized code.
    static const int ExternalUseCost = 1;
    /// The user is internal but in a different lane.
    static const int UserInDiffLaneCost = ExternalUseCost;
 
    /// \returns the score of placing \p V1 and \p V2 in consecutive lanes.
    static int getShallowScore(Value *V1, Value *V2, const DataLayout &DL,
                               ScalarEvolution &SE, int NumLanes) {
      if (V1 == V2)
        return VLOperands::ScoreSplat;
 
      auto *LI1 = dyn_cast<LoadInst>(V1);
      auto *LI2 = dyn_cast<LoadInst>(V2);
      if (LI1 && LI2) {
        if (LI1->getParent() != LI2->getParent())
          return VLOperands::ScoreFail;
 
        Optional<int> Dist = getPointersDiff(
            LI1->getType(), LI1->getPointerOperand(), LI2->getType(),
            LI2->getPointerOperand(), DL, SE, /*StrictCheck=*/true);
        if (!Dist)
          return VLOperands::ScoreFail;
        // The distance is too large - still may be profitable to use masked
        // loads/gathers.
        if (std::abs(*Dist) > NumLanes / 2)
          return VLOperands::ScoreAltOpcodes;
        // 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) ? VLOperands::ScoreConsecutiveLoads
                           : VLOperands::ScoreReversedLoads;
      }
 
      auto *C1 = dyn_cast<Constant>(V1);
      auto *C2 = dyn_cast<Constant>(V2);
      if (C1 && C2)
        return VLOperands::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.
        if (isa<UndefValue>(V2))
          return VLOperands::ScoreConsecutiveExtracts;
        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 VLOperands::ScoreConsecutiveExtracts;
          if (isUndefVector(EV2) && EV2->getType() == EV1->getType())
            return VLOperands::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) > NumLanes / 2)
              return VLOperands::ScoreAltOpcodes;
            return (Dist > 0) ? VLOperands::ScoreConsecutiveExtracts
                              : VLOperands::ScoreReversedExtracts;
          }
        }
      }
 
      auto *I1 = dyn_cast<Instruction>(V1);
      auto *I2 = dyn_cast<Instruction>(V2);
      if (I1 && I2) {
        if (I1->getParent() != I2->getParent())
          return VLOperands::ScoreFail;
        InstructionsState S = getSameOpcode({I1, I2});
        // Note: Only consider instructions with <= 2 operands to avoid
        // complexity explosion.
        if (S.getOpcode() && S.MainOp->getNumOperands() <= 2)
          return S.isAltShuffle() ? VLOperands::ScoreAltOpcodes
                                  : VLOperands::ScoreSameOpcode;
      }
 
      if (isa<UndefValue>(V2))
        return VLOperands::ScoreUndef;
 
      return VLOperands::ScoreFail;
    }
 
    /// Holds the values and their lanes that are taking part in the look-ahead
    /// score calculation. This is used in the external uses cost calculation.
    /// Need to hold all the lanes in case of splat/broadcast at least to
    /// correctly check for the use in the different lane.
    SmallDenseMap<Value *, SmallSet<int, 4>> InLookAheadValues;
 
    /// \returns the additional cost due to uses of \p LHS and \p RHS that are
    /// either external to the vectorized code, or require shuffling.
    int getExternalUsesCost(const std::pair<Value *, int> &LHS,
                            const std::pair<Value *, int> &RHS) {
      int Cost = 0;
      std::array<std::pair<Value *, int>, 2> Values = {{LHS, RHS}};
      for (int Idx = 0, IdxE = Values.size(); Idx != IdxE; ++Idx) {
        Value *V = Values[Idx].first;
        if (isa<Constant>(V)) {
          // Since this is a function pass, it doesn't make semantic sense to
          // walk the users of a subclass of Constant. The users could be in
          // another function, or even another module that happens to be in
          // the same LLVMContext.
          continue;
        }
 
        // Calculate the absolute lane, using the minimum relative lane of LHS
        // and RHS as base and Idx as the offset.
        int Ln = std::min(LHS.second, RHS.second) + Idx;
        assert(Ln >= 0 && "Bad lane calculation");
        unsigned UsersBudget = LookAheadUsersBudget;
        for (User *U : V->users()) {
          if (const TreeEntry *UserTE = R.getTreeEntry(U)) {
            // The user is in the VectorizableTree. Check if we need to insert.
            int UserLn = UserTE->findLaneForValue(U);
            assert(UserLn >= 0 && "Bad lane");
            // If the values are different, check just the line of the current
            // value. If the values are the same, need to add UserInDiffLaneCost
            // only if UserLn does not match both line numbers.
            if ((LHS.first != RHS.first && UserLn != Ln) ||
                (LHS.first == RHS.first && UserLn != LHS.second &&
                 UserLn != RHS.second)) {
              Cost += UserInDiffLaneCost;
              break;
            }
          } else {
            // Check if the user is in the look-ahead code.
            auto It2 = InLookAheadValues.find(U);
            if (It2 != InLookAheadValues.end()) {
              // The user is in the look-ahead code. Check the lane.
              if (!It2->getSecond().contains(Ln)) {
                Cost += UserInDiffLaneCost;
                break;
              }
            } else {
              // The user is neither in SLP tree nor in the look-ahead code.
              Cost += ExternalUseCost;
              break;
            }
          }
          // Limit the number of visited uses to cap compilation time.
          if (--UsersBudget == 0)
            break;
        }
      }
      return Cost;
    }
 
    /// Go through the operands of \p LHS and \p RHS recursively until \p
    /// MaxLevel, and return the cummulative score. 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 VLOperands::ScoreConsecutiveLoads, while
    /// {A[0],C[0]} has a score of VLOperands::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(const std::pair<Value *, int> &LHS,
                           const std::pair<Value *, int> &RHS, int CurrLevel,
                           int MaxLevel) {
 
      Value *V1 = LHS.first;
      Value *V2 = RHS.first;
      // Get the shallow score of V1 and V2.
      int ShallowScoreAtThisLevel = std::max(
          (int)ScoreFail, getShallowScore(V1, V2, DL, SE, getNumLanes()) -
                              getExternalUsesCost(LHS, RHS));
      int Lane1 = LHS.second;
      int Lane2 = RHS.second;
 
      // 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>(V1);
      auto *I2 = dyn_cast<Instruction>(V2);
      if (CurrLevel == MaxLevel || !(I1 && I2) || I1 == I2 ||
          ShallowScoreAtThisLevel == VLOperands::ScoreFail ||
          (((isa<LoadInst>(I1) && isa<LoadInst>(I2)) ||
            (isa<ExtractElementInst>(I1) && isa<ExtractElementInst>(I2))) &&
           ShallowScoreAtThisLevel))
        return ShallowScoreAtThisLevel;
      assert(I1 && I2 && "Should have early exited.");
 
      // Keep track of in-tree values for determining the external-use cost.
      InLookAheadValues[V1].insert(Lane1);
      InLookAheadValues[V2].insert(Lane2);
 
      // 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), Lane1},
                                            {I2->getOperand(OpIdx2), Lane2},
                                            CurrLevel + 1, MaxLevel);
          // Look for the best score.
          if (TmpScore > VLOperands::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;
    }
 
    /// \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(const std::pair<Value *, int> &LHS,
                          const std::pair<Value *, int> &RHS) {
      InLookAheadValues.clear();
      return getScoreAtLevelRec(LHS, RHS, 1, LookAheadMaxDepth);
    }
 
    // 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 None.
    Optional<unsigned>
    getBestOperand(unsigned OpIdx, int Lane, int LastLane,
                   ArrayRef<ReorderingMode> ReorderingModes) {
      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];
 
      // 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 {
        Optional<unsigned> Idx = None;
        unsigned Score = 0;
      } BestOp;
 
      // 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::Constant:
        case ReorderingMode::Opcode: {
          bool LeftToRight = Lane > LastLane;
          Value *OpLeft = (LeftToRight) ? OpLastLane : Op;
          Value *OpRight = (LeftToRight) ? Op : OpLastLane;
          unsigned Score =
              getLookAheadScore({OpLeft, LastLane}, {OpRight, Lane});
          if (Score > BestOp.Score) {
            BestOp.Idx = Idx;
            BestOp.Score = Score;
          }
          break;
        }
        case ReorderingMode::Splat:
          if (Op == OpLastLane)
            BestOp.Idx = Idx;
          break;
        case ReorderingMode::Failed:
          return None;
        }
      }
 
      if (BestOp.Idx) {
        getData(BestOp.Idx.getValue(), Lane).IsUsed = true;
        return BestOp.Idx;
      }
      // If we could not find a good match return None.
      return None;
    }
 
    /// 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}).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();
      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.
    /// 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;
      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;
          if (Data.V == Op) {
            FoundCandidate = true;
            Data.IsUsed = true;
            break;
          }
        }
        if (!FoundCandidate)
          return false;
      }
      return true;
    }
 
  public:
    /// Initialize with all the operands of the instruction vector \p RootVL.
    VLOperands(ArrayRef<Value *> RootVL, const DataLayout &DL,
               ScalarEvolution &SE, const BoUpSLP &R)
        : DL(DL), SE(SE), R(R) {
      // 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 (isa<Instruction>(OpLane0)) {
          // Check if OpLane0 should be broadcast.
          if (shouldBroadcast(OpLane0, 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 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.
        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].
              Optional<unsigned> BestIdx =
                  getBestOperand(OpIdx, Lane, LastLane, ReorderingModes);
              // 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.getValue(), Lane);
              } else {
                // We failed to find a best operand, set mode to 'Failed'.
                ReorderingModes[OpIdx] = ReorderingMode::Failed;
                // Enable the second pass.
                StrategyFailed = true;
              }
            }
          }
        }
        // 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
  };
 
  /// Checks if the instruction is marked for deletion.
  bool isDeleted(Instruction *I) const { return DeletedInstructions.count(I); }
 
  /// Marks values operands for later deletion by replacing them with Undefs.
  void eraseInstructions(ArrayRef<Value *> AV);
 
  ~BoUpSLP();
 
private:
  /// Checks if all users of \p I are the part of the vectorization tree.
  bool areAllUsersVectorized(Instruction *I,
                             ArrayRef<Value *> VectorizedVals) const;
 
  /// \returns the cost of the vectorizable entry.
  InstructionCost getEntryCost(const TreeEntry *E,
                               ArrayRef<Value *> VectorizedVals);
 
  /// 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.
  bool canReuseExtract(ArrayRef<Value *> VL, Value *OpValue,
                       SmallVectorImpl<unsigned> &CurrentOrder) const;
 
  /// Vectorize a single entry in the tree.
  Value *vectorizeTree(TreeEntry *E);
 
  /// Vectorize a single entry in the tree, starting in \p VL.
  Value *vectorizeTree(ArrayRef<Value *> VL);
 
  /// \returns the scalarization cost for this type. Scalarization in this
  /// context means the creation of vectors from a group of scalars. If \p
  /// NeedToShuffle is true, need to add a cost of reshuffling some of the
  /// vector elements.
  InstructionCost getGatherCost(FixedVectorType *Ty,
                                const DenseSet<unsigned> &ShuffledIndices,
                                bool NeedToShuffle) const;
 
  /// Checks if the gathered \p VL can be represented as shuffle(s) of previous
  /// tree entries.
  /// \returns ShuffleKind, if gathered values can be represented as shuffles of
  /// previous tree entries. \p Mask is filled with the shuffle mask.
  Optional<TargetTransformInfo::ShuffleKind>
  isGatherShuffledEntry(const TreeEntry *TE, SmallVectorImpl<int> &Mask,
                        SmallVectorImpl<const TreeEntry *> &Entries);
 
  /// \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.
  InstructionCost getGatherCost(ArrayRef<Value *> VL) 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.
  Value *gather(ArrayRef<Value *> VL);
 
  /// \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 DataLayout &DL,
                                             ScalarEvolution &SE,
                                             const BoUpSLP &R);
  struct TreeEntry {
    using VecTreeTy = SmallVector<std::unique_ptr<TreeEntry>, 8>;
    TreeEntry(VecTreeTy &Container) : Container(Container) {}
 
    /// \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 == UndefMaskElem) ||
                                   (Idx != UndefMaskElem && V == Scalars[Idx]);
                          });
      };
      if (!ReorderIndices.empty()) {
        // TODO: implement matching if the nodes are just reordered, still can
        // treat the vector as the same if the list of scalars matches VL
        // directly, without reordering.
        SmallVector<int> Mask;
        inversePermutation(ReorderIndices, Mask);
        if (VL.size() == Scalars.size())
          return IsSame(Scalars, Mask);
        if (VL.size() == ReuseShuffleIndices.size()) {
          ::addMask(Mask, ReuseShuffleIndices);
          return IsSame(Scalars, Mask);
        }
        return false;
      }
      return IsSame(Scalars, ReuseShuffleIndices);
    }
 
    /// \returns true if current entry has same operands as \p TE.
    bool hasEqualOperands(const TreeEntry &TE) const {
      if (TE.getNumOperands() != getNumOperands())
        return false;
      SmallBitVector Used(getNumOperands());
      for (unsigned I = 0, E = getNumOperands(); I < E; ++I) {
        unsigned PrevCount = Used.count();
        for (unsigned K = 0; K < E; ++K) {
          if (Used.test(K))
            continue;
          if (getOperand(K) == TE.getOperand(I)) {
            Used.set(K);
            break;
          }
        }
        // Check if we actually found the matching operand.
        if (PrevCount == Used.count())
          return false;
      }
      return true;
    }
 
    /// \return Final vectorization factor for the node. Defined by the total
    /// number of vectorized scalars, including those, used several times in the
    /// entry and counted in the \a ReuseShuffleIndices, if any.
    unsigned getVectorFactor() const {
      if (!ReuseShuffleIndices.empty())
        return ReuseShuffleIndices.size();
      return Scalars.size();
    };
 
    /// A vector of scalars.
    ValueList Scalars;
 
    /// The Scalars are vectorized into this value. It is initialized to Null.
    Value *VectorizedValue = nullptr;
 
    /// Do we need to gather this sequence or vectorize it
    /// (either with vector instruction or with scatter/gather
    /// intrinsics for store/load)?
    enum EntryState { Vectorize, ScatterVectorize, NeedToGather };
    EntryState State;
 
    /// 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;
    }
 
#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 NeedToGather:
        dbgs() << "NeedToGather\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) const {
    dbgs() << "SLP: Calculated costs for Tree:\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, Optional<ScheduleData *> Bundle,
                          const InstructionsState &S,
                          const EdgeInfo &UserTreeIdx,
                          ArrayRef<int> ReuseShuffleIndices = None,
                          ArrayRef<unsigned> ReorderIndices = None) {
    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,
                          Optional<ScheduleData *> Bundle,
                          const InstructionsState &S,
                          const EdgeInfo &UserTreeIdx,
                          ArrayRef<int> ReuseShuffleIndices = None,
                          ArrayRef<unsigned> ReorderIndices = None) {
    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);
      Last->setOperations(S);
      Last->ReorderIndices.append(ReorderIndices.begin(), ReorderIndices.end());
    }
    if (Last->State != TreeEntry::NeedToGather) {
      for (Value *V : VL) {
        assert(!getTreeEntry(V) && "Scalar already in tree!");
        ScalarToTreeEntry[V] = Last;
      }
      // Update the scheduler bundle to point to this TreeEntry.
      unsigned Lane = 0;
      for (ScheduleData *BundleMember = Bundle.getValue(); BundleMember;
           BundleMember = BundleMember->NextInBundle) {
        BundleMember->TE = Last;
        BundleMember->Lane = Lane;
        ++Lane;
      }
      assert((!Bundle.getValue() || Lane == VL.size()) &&
             "Bundle and VL out of sync");
    } else {
      MustGather.insert(VL.begin(), VL.end());
    }
 
    if (UserTreeIdx.UserTE)
      Last->UserTreeIndices.push_back(UserTreeIdx);
 
    return Last;
  }
 
  /// -- Vectorization State --
  /// Holds all of the tree entries.
  TreeEntry::VecTreeTy VectorizableTree;
 
#ifndef NDEBUG
  /// Debug printer.
  LLVM_DUMP_METHOD void dumpVectorizableTree() const {
    for (unsigned Id = 0, IdE = VectorizableTree.size(); Id != IdE; ++Id) {
      VectorizableTree[Id]->dump();
      dbgs() << "\n";
    }
  }
#endif
 
  TreeEntry *getTreeEntry(Value *V) { return ScalarToTreeEntry.lookup(V); }
 
  const TreeEntry *getTreeEntry(Value *V) const {
    return ScalarToTreeEntry.lookup(V);
  }
 
  /// Maps a specific scalar to its tree entry.
  SmallDenseMap<Value*, TreeEntry *> ScalarToTreeEntry;
 
  /// 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;
 
  /// 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) {
    // First check if the result is already in the cache.
    AliasCacheKey key = std::make_pair(Inst1, Inst2);
    Optional<bool> &result = AliasCache[key];
    if (result.hasValue()) {
      return result.getValue();
    }
    bool aliased = true;
    if (Loc1.Ptr && isSimple(Inst1))
      aliased = isModOrRefSet(AA->getModRefInfo(Inst2, Loc1));
    // Store the result in the cache.
    result = aliased;
    return aliased;
  }
 
  using AliasCacheKey = std::pair<Instruction *, Instruction *>;
 
  /// Cache for alias results.
  /// TODO: consider moving this to the AliasAnalysis itself.
  DenseMap<AliasCacheKey, Optional<bool>> AliasCache;
 
  /// 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.
  /// This 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.
  void eraseInstruction(Instruction *I, bool ReplaceOpsWithUndef = false) {
    auto It = DeletedInstructions.try_emplace(I, ReplaceOpsWithUndef).first;
    It->getSecond() = It->getSecond() && ReplaceOpsWithUndef;
  }
 
  /// Temporary store for deleted instructions. Instructions will be deleted
  /// eventually when the BoUpSLP is destructed.
  DenseMap<Instruction *, bool> DeletedInstructions;
 
  /// 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;
 
  /// Values used only by @llvm.assume calls.
  SmallPtrSet<const Value *, 32> EphValues;
 
  /// Holds all of the instructions that we gathered.
  SetVector<Instruction *> GatherShuffleSeq;
 
  /// A list of blocks that we are going to CSE.
  SetVector<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;
      UnscheduledDepsInBundle = UnscheduledDeps;
      clearDependencies();
      OpValue = OpVal;
      TE = nullptr;
      Lane = -1;
    }
 
    /// Returns true if the dependency information has been calculated.
    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;
    }
 
    /// 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, also updating it for
    /// the whole bundle.
    int incrementUnscheduledDeps(int Incr) {
      UnscheduledDeps += Incr;
      return FirstInBundle->UnscheduledDepsInBundle += Incr;
    }
 
    /// Sets the number of unscheduled dependencies to the number of
    /// dependencies.
    void resetUnscheduledDeps() {
      incrementUnscheduledDeps(Dependencies - UnscheduledDeps);
    }
 
    /// Clears all dependency information.
    void clearDependencies() {
      Dependencies = InvalidDeps;
      resetUnscheduledDeps();
      MemoryDependencies.clear();
    }
 
    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;
 
    /// 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;
 
    /// 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;
 
    /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for
    /// single instructions.
    int UnscheduledDepsInBundle = InvalidDeps;
 
    /// True if this instruction is scheduled (or considered as scheduled in the
    /// dry-run).
    bool IsScheduled = false;
 
    /// Opcode of the current instruction in the schedule data.
    Value *OpValue = nullptr;
 
    /// The TreeEntry that this instruction corresponds to.
    TreeEntry *TE = nullptr;
 
    /// The lane of this node in the TreeEntry.
    int Lane = -1;
  };
 
#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.
  struct BlockScheduling {
    BlockScheduling(BasicBlock *BB)
        : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {}
 
    void clear() {
      ReadyInsts.clear();
      ScheduleStart = nullptr;
      ScheduleEnd = nullptr;
      FirstLoadStoreInRegion = nullptr;
      LastLoadStoreInRegion = nullptr;
 
      // 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(Value *V) {
      ScheduleData *SD = ScheduleDataMap[V];
      if (SD && SD->SchedulingRegionID == SchedulingRegionID)
        return SD;
      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[Key];
        if (SD && SD->SchedulingRegionID == SchedulingRegionID)
          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 duiring buildTree(). We therefore need to get its operands
        // through the TreeEntry.
        if (TreeEntry *TE = BundleMember->TE) {
          int Lane = BundleMember->Lane;
          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 = TE->getMainOp();
          assert(In &&
                 (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(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->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");
          }
        }
      }
    }
 
    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 (P.second->SchedulingRegionID == SchedulingRegionID)
            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->isReady()) {
            ReadyList.insert(SD);
            LLVM_DEBUG(dbgs()
                       << "SLP:    initially in ready list: " << *I << "\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 non-None
    /// if \p VL is allowed to be scheduled.
    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.
    std::vector<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<Value *, ScheduleData *> ScheduleDataMap;
 
    /// Attaches ScheduleData to Instruction with the leading key.
    DenseMap<Value *, SmallDenseMap<Value *, ScheduleData *>>
        ExtraScheduleDataMap;
 
    struct ReadyList : SmallVector<ScheduleData *, 8> {
      void insert(ScheduleData *SD) { push_back(SD); }
    };
 
    /// The ready-list for scheduling (only used for the dry-run).
    ReadyList 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;
 
    /// 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.
  ArrayRef<Value *> UserIgnoreList;
 
  /// 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;
  AAResults *AA;
  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<> 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.
  MapVector<Value *, std::pair<uint64_t, bool>> MinBWs;
};
 
} // 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);
    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->State == TreeEntry::NeedToGather)
      return "color=red";
    return "";
  }
};
 
} // end namespace llvm
 
BoUpSLP::~BoUpSLP() {
  for (const auto &Pair : DeletedInstructions) {
    // Replace operands of ignored instructions with Undefs in case if they were
    // marked for deletion.
    if (Pair.getSecond()) {
      Value *Undef = UndefValue::get(Pair.getFirst()->getType());
      Pair.getFirst()->replaceAllUsesWith(Undef);
    }
    Pair.getFirst()->dropAllReferences();
  }
  for (const auto &Pair : DeletedInstructions) {
    assert(Pair.getFirst()->use_empty() &&
           "trying to erase instruction with users.");
    Pair.getFirst()->eraseFromParent();
  }
#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
}
 
void BoUpSLP::eraseInstructions(ArrayRef<Value *> AV) {
  for (auto *V : AV) {
    if (auto *I = dyn_cast<Instruction>(V))
      eraseInstruction(I, /*ReplaceOpsWithUndef=*/true);
  };
}
 
/// 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] != UndefMaskElem)
      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) {
  assert(!Mask.empty() && "Expected non-empty mask.");
  SmallVector<int> MaskOrder;
  if (Order.empty()) {
    MaskOrder.resize(Mask.size());
    std::iota(MaskOrder.begin(), MaskOrder.end(), 0);
  } else {
    inversePermutation(Order, MaskOrder);
  }
  reorderReuses(MaskOrder, Mask);
  if (ShuffleVectorInst::isIdentityMask(MaskOrder)) {
    Order.clear();
    return;
  }
  Order.assign(Mask.size(), Mask.size());
  for (unsigned I = 0, E = Mask.size(); I < E; ++I)
    if (MaskOrder[I] != UndefMaskElem)
      Order[MaskOrder[I]] = I;
  fixupOrderingIndices(Order);
}
 
Optional<BoUpSLP::OrdersType>
BoUpSLP::findReusedOrderedScalars(const BoUpSLP::TreeEntry &TE) {
  assert(TE.State == TreeEntry::NeedToGather && "Expected gather node only.");
  unsigned NumScalars = TE.Scalars.size();
  OrdersType CurrentOrder(NumScalars, NumScalars);
  SmallVector<int> Positions;
  SmallBitVector UsedPositions(NumScalars);
  const TreeEntry *STE = nullptr;
  // Try to find all gathered scalars that are gets vectorized in other
  // vectorize node. Here we can have only one single tree vector node to
  // correctly identify order of the gathered scalars.
  for (unsigned I = 0; I < NumScalars; ++I) {
    Value *V = TE.Scalars[I];
    if (!isa<LoadInst, ExtractElementInst, ExtractValueInst>(V))
      continue;
    if (const auto *LocalSTE = getTreeEntry(V)) {
      if (!STE)
        STE = LocalSTE;
      else if (STE != LocalSTE)
        // Take the order only from the single vector node.
        return None;
      unsigned Lane =
          std::distance(STE->Scalars.begin(), find(STE->Scalars, V));
      if (Lane >= NumScalars)
        return None;
      if (CurrentOrder[Lane] != NumScalars) {
        if (Lane != I)
          continue;
        UsedPositions.reset(CurrentOrder[Lane]);
      }
      // The partial identity (where only some elements of the gather node are
      // in the identity order) is good.
      CurrentOrder[Lane] = I;
      UsedPositions.set(I);
    }
  }
  // Need to keep the order if we have a vector entry and at least 2 scalars or
  // the vectorized entry has just 2 scalars.
  if (STE && (UsedPositions.count() > 1 || STE->Scalars.size() == 2)) {
    auto &&IsIdentityOrder = [NumScalars](ArrayRef<unsigned> CurrentOrder) {
      for (unsigned I = 0; I < NumScalars; ++I)
        if (CurrentOrder[I] != I && CurrentOrder[I] != NumScalars)
          return false;
      return true;
    };
    if (IsIdentityOrder(CurrentOrder)) {
      CurrentOrder.clear();
      return CurrentOrder;
    }
    auto *It = CurrentOrder.begin();
    for (unsigned I = 0; I < NumScalars;) {
      if (UsedPositions.test(I)) {
        ++I;
        continue;
      }
      if (*It == NumScalars) {
        *It = I;
        ++I;
      }
      ++It;
    }
    return CurrentOrder;
  }
  return None;
}
 
Optional<BoUpSLP::OrdersType> BoUpSLP::getReorderingData(const TreeEntry &TE,
                                                         bool TopToBottom) {
  // No need to reorder if need to shuffle reuses, still need to shuffle the
  // node.
  if (!TE.ReuseShuffleIndices.empty())
    return None;
  if (TE.State == TreeEntry::Vectorize &&
      (isa<LoadInst, ExtractElementInst, ExtractValueInst>(TE.getMainOp()) ||
       (TopToBottom && isa<StoreInst, InsertElementInst>(TE.getMainOp()))) &&
      !TE.isAltShuffle())
    return TE.ReorderIndices;
  if (TE.State == TreeEntry::NeedToGather) {
    // TODO: add analysis of other gather nodes with extractelement
    // instructions and other values/instructions, not only undefs.
    if (((TE.getOpcode() == Instruction::ExtractElement &&
          !TE.isAltShuffle()) ||
         (all_of(TE.Scalars,
                 [](Value *V) {
                   return isa<UndefValue, ExtractElementInst>(V);
                 }) &&
          any_of(TE.Scalars,
                 [](Value *V) { return isa<ExtractElementInst>(V); }))) &&
        all_of(TE.Scalars,
               [](Value *V) {
                 auto *EE = dyn_cast<ExtractElementInst>(V);
                 return !EE || isa<FixedVectorType>(EE->getVectorOperandType());
               }) &&
        allSameType(TE.Scalars)) {
      // Check that gather of extractelements can be represented as
      // just a shuffle of a single vector.
      OrdersType CurrentOrder;
      bool Reuse = canReuseExtract(TE.Scalars, TE.getMainOp(), CurrentOrder);
      if (Reuse || !CurrentOrder.empty()) {
        if (!CurrentOrder.empty())
          fixupOrderingIndices(CurrentOrder);
        return CurrentOrder;
      }
    }
    if (Optional<OrdersType> CurrentOrder = findReusedOrderedScalars(TE))
      return CurrentOrder;
  }
  return None;
}
 
void BoUpSLP::reorderTopToBottom() {
  // Maps VF to the graph nodes.
  DenseMap<unsigned, SetVector<TreeEntry *>> VFToOrderedEntries;
  // ExtractElement gather nodes which can be vectorized and need to handle
  // their ordering.
  DenseMap<const TreeEntry *, OrdersType> GathersToOrders;
  // Find all reorderable nodes with the given VF.
  // Currently the are vectorized stores,loads,extracts + some gathering of
  // extracts.
  for_each(VectorizableTree, [this, &VFToOrderedEntries, &GathersToOrders](
                                 const std::unique_ptr<TreeEntry> &TE) {
    if (Optional<OrdersType> CurrentOrder =
            getReorderingData(*TE.get(), /*TopToBottom=*/true)) {
      // Do not include ordering for nodes used in the alt opcode vectorization,
      // better to reorder them during bottom-to-top stage. If follow the order
      // here, it causes reordering of the whole graph though actually it is
      // profitable just to reorder the subgraph that starts from the alternate
      // opcode vectorization node. Such nodes already end-up with the shuffle
      // instruction and it is just enough to change this shuffle rather than
      // rotate the scalars for the whole graph.
      unsigned Cnt = 0;
      const TreeEntry *UserTE = TE.get();
      while (UserTE && Cnt < RecursionMaxDepth) {
        if (UserTE->UserTreeIndices.size() != 1)
          break;
        if (all_of(UserTE->UserTreeIndices, [](const EdgeInfo &EI) {
              return EI.UserTE->State == TreeEntry::Vectorize &&
                     EI.UserTE->isAltShuffle() && EI.UserTE->Idx != 0;
            }))
          return;
        if (UserTE->UserTreeIndices.empty())
          UserTE = nullptr;
        else
          UserTE = UserTE->UserTreeIndices.back().UserTE;
        ++Cnt;
      }
      VFToOrderedEntries[TE->Scalars.size()].insert(TE.get());
      if (TE->State != TreeEntry::Vectorize)
        GathersToOrders.try_emplace(TE.get(), *CurrentOrder);
    }
  });
 
  // Reorder the graph nodes according to their vectorization factor.
  for (unsigned VF = VectorizableTree.front()->Scalars.size(); VF > 1;
       VF /= 2) {
    auto It = VFToOrderedEntries.find(VF);
    if (It == VFToOrderedEntries.end())
      continue;
    // Try to find the most profitable order. We just are looking for the most
    // used order and reorder scalar elements in the nodes according to this
    // mostly used order.
    ArrayRef<TreeEntry *> OrderedEntries = It->second.getArrayRef();
    // All operands are reordered and used only in this node - propagate the
    // most used order to the user node.
    MapVector<OrdersType, unsigned,
              DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>>
        OrdersUses;
    SmallPtrSet<const TreeEntry *, 4> VisitedOps;
    for (const TreeEntry *OpTE : OrderedEntries) {
      // No need to reorder this nodes, still need to extend and to use shuffle,
      // just need to merge reordering shuffle and the reuse shuffle.
      if (!OpTE->ReuseShuffleIndices.empty())
        continue;
      // Count number of orders uses.
      const auto &Order = [OpTE, &GathersToOrders]() -> const OrdersType & {
        if (OpTE->State == TreeEntry::NeedToGather)
          return GathersToOrders.find(OpTE)->second;
        return OpTE->ReorderIndices;
      }();
      // Stores actually store the mask, not the order, need to invert.
      if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() &&
          OpTE->getOpcode() == Instruction::Store && !Order.empty()) {
        SmallVector<int> Mask;
        inversePermutation(Order, Mask);
        unsigned E = Order.size();
        OrdersType CurrentOrder(E, E);
        transform(Mask, CurrentOrder.begin(), [E](int Idx) {
          return Idx == UndefMaskElem ? E : static_cast<unsigned>(Idx);
        });
        fixupOrderingIndices(CurrentOrder);
        ++OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second;
      } else {
        ++OrdersUses.insert(std::make_pair(Order, 0)).first->second;
      }
    }
    // Set order of the user node.
    if (OrdersUses.empty())
      continue;
    // Choose the most used order.
    ArrayRef<unsigned> BestOrder = OrdersUses.front().first;
    unsigned Cnt = OrdersUses.front().second;
    for (const auto &Pair : drop_begin(OrdersUses)) {
      if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) {
        BestOrder = Pair.first;
        Cnt = Pair.second;
      }
    }
    // Set order of the user node.
    if (BestOrder.empty())
      continue;
    SmallVector<int> Mask;
    inversePermutation(BestOrder, Mask);
    SmallVector<int> MaskOrder(BestOrder.size(), UndefMaskElem);
    unsigned E = BestOrder.size();
    transform(BestOrder, MaskOrder.begin(), [E](unsigned I) {
      return I < E ? static_cast<int>(I) : UndefMaskElem;
    });
    // Do an actual reordering, if profitable.
    for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
      // Just do the reordering for the nodes with the given VF.
      if (TE->Scalars.size() != VF) {
        if (TE->ReuseShuffleIndices.size() == VF) {
          // Need to reorder the reuses masks of the operands with smaller VF to
          // be able to find the match between the graph nodes and scalar
          // operands of the given node during vectorization/cost estimation.
          assert(all_of(TE->UserTreeIndices,
                        [VF, &TE](const EdgeInfo &EI) {
                          return EI.UserTE->Scalars.size() == VF ||
                                 EI.UserTE->Scalars.size() ==
                                     TE->Scalars.size();
                        }) &&
                 "All users must be of VF size.");
          // Update ordering of the operands with the smaller VF than the given
          // one.
          reorderReuses(TE->ReuseShuffleIndices, Mask);
        }
        continue;
      }
      if (TE->State == TreeEntry::Vectorize &&
          isa<ExtractElementInst, ExtractValueInst, LoadInst, StoreInst,
              InsertElementInst>(TE->getMainOp()) &&
          !TE->isAltShuffle()) {
        // Build correct orders for extract{element,value}, loads and
        // stores.
        reorderOrder(TE->ReorderIndices, Mask);
        if (isa<InsertElementInst, StoreInst>(TE->getMainOp()))
          TE->reorderOperands(Mask);
      } else {
        // Reorder the node and its operands.
        TE->reorderOperands(Mask);
        assert(TE->ReorderIndices.empty() &&
               "Expected empty reorder sequence.");
        reorderScalars(TE->Scalars, Mask);
      }
      if (!TE->ReuseShuffleIndices.empty()) {
        // Apply reversed order to keep the original ordering of the reused
        // elements to avoid extra reorder indices shuffling.
        OrdersType CurrentOrder;
        reorderOrder(CurrentOrder, MaskOrder);
        SmallVector<int> NewReuses;
        inversePermutation(CurrentOrder, NewReuses);
        addMask(NewReuses, TE->ReuseShuffleIndices);
        TE->ReuseShuffleIndices.swap(NewReuses);
      }
    }
  }
}
 
void BoUpSLP::reorderBottomToTop(bool IgnoreReorder) {
  SetVector<TreeEntry *> OrderedEntries;
  DenseMap<const TreeEntry *, OrdersType> GathersToOrders;
  // Find all reorderable leaf nodes with the given VF.
  // Currently the are vectorized loads,extracts without alternate operands +
  // some gathering of extracts.
  SmallVector<TreeEntry *> NonVectorized;
  for_each(VectorizableTree, [this, &OrderedEntries, &GathersToOrders,
                              &NonVectorized](
                                 const std::unique_ptr<TreeEntry> &TE) {
    if (TE->State != TreeEntry::Vectorize)
      NonVectorized.push_back(TE.get());
    if (Optional<OrdersType> CurrentOrder =
            getReorderingData(*TE.get(), /*TopToBottom=*/false)) {
      OrderedEntries.insert(TE.get());
      if (TE->State != TreeEntry::Vectorize)
        GathersToOrders.try_emplace(TE.get(), *CurrentOrder);
    }
  });
 
  // Checks if the operands of the users are reordarable and have only single
  // use.
  auto &&CheckOperands =
      [this, &NonVectorized](const auto &Data,
                             SmallVectorImpl<TreeEntry *> &GatherOps) {
        for (unsigned I = 0, E = Data.first->getNumOperands(); I < E; ++I) {
          if (any_of(Data.second,
                     [I](const std::pair<unsigned, TreeEntry *> &OpData) {
                       return OpData.first == I &&
                              OpData.second->State == TreeEntry::Vectorize;
                     }))
            continue;
          ArrayRef<Value *> VL = Data.first->getOperand(I);
          const TreeEntry *TE = nullptr;
          const auto *It = find_if(VL, [this, &TE](Value *V) {
            TE = getTreeEntry(V);
            return TE;
          });
          if (It != VL.end() && TE->isSame(VL))
            return false;
          TreeEntry *Gather = nullptr;
          if (count_if(NonVectorized, [VL, &Gather](TreeEntry *TE) {
                assert(TE->State != TreeEntry::Vectorize &&
                       "Only non-vectorized nodes are expected.");
                if (TE->isSame(VL)) {
                  Gather = TE;
                  return true;
                }
                return false;
              }) > 1)
            return false;
          if (Gather)
            GatherOps.push_back(Gather);
        }
        return true;
      };
  // 1. Propagate order to the graph nodes, which use only reordered nodes.
  // I.e., if the node has operands, that are reordered, try to make at least
  // one operand order in the natural order and reorder others + reorder the
  // user node itself.
  SmallPtrSet<const TreeEntry *, 4> Visited;
  while (!OrderedEntries.empty()) {
    // 1. Filter out only reordered nodes.
    // 2. If the entry has multiple uses - skip it and jump to the next node.
    MapVector<TreeEntry *, SmallVector<std::pair<unsigned, TreeEntry *>>> Users;
    SmallVector<TreeEntry *> Filtered;
    for (TreeEntry *TE : OrderedEntries) {
      if (!(TE->State == TreeEntry::Vectorize ||
            (TE->State == TreeEntry::NeedToGather &&
             GathersToOrders.count(TE))) ||
          TE->UserTreeIndices.empty() || !TE->ReuseShuffleIndices.empty() ||
          !all_of(drop_begin(TE->UserTreeIndices),
                  [TE](const EdgeInfo &EI) {
                    return EI.UserTE == TE->UserTreeIndices.front().UserTE;
                  }) ||
          !Visited.insert(TE).second) {
        Filtered.push_back(TE);
        continue;
      }
      // Build a map between user nodes and their operands order to speedup
      // search. The graph currently does not provide this dependency directly.
      for (EdgeInfo &EI : TE->UserTreeIndices) {
        TreeEntry *UserTE = EI.UserTE;
        auto It = Users.find(UserTE);
        if (It == Users.end())
          It = Users.insert({UserTE, {}}).first;
        It->second.emplace_back(EI.EdgeIdx, TE);
      }
    }
    // Erase filtered entries.
    for_each(Filtered,
             [&OrderedEntries](TreeEntry *TE) { OrderedEntries.remove(TE); });
    for (const auto &Data : Users) {
      // Check that operands are used only in the User node.
      SmallVector<TreeEntry *> GatherOps;
      if (!CheckOperands(Data, GatherOps)) {
        for_each(Data.second,
                 [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) {
                   OrderedEntries.remove(Op.second);
                 });
        continue;
      }
      // All operands are reordered and used only in this node - propagate the
      // most used order to the user node.
      MapVector<OrdersType, unsigned,
                DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>>
          OrdersUses;
      SmallPtrSet<const TreeEntry *, 4> VisitedOps;
      for (const auto &Op : Data.second) {
        TreeEntry *OpTE = Op.second;
        if (!OpTE->ReuseShuffleIndices.empty() ||
            (IgnoreReorder && OpTE == VectorizableTree.front().get()))
          continue;
        const auto &Order = [OpTE, &GathersToOrders]() -> const OrdersType & {
          if (OpTE->State == TreeEntry::NeedToGather)
            return GathersToOrders.find(OpTE)->second;
          return OpTE->ReorderIndices;
        }();
        // Stores actually store the mask, not the order, need to invert.
        if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() &&
            OpTE->getOpcode() == Instruction::Store && !Order.empty()) {
          SmallVector<int> Mask;
          inversePermutation(Order, Mask);
          unsigned E = Order.size();
          OrdersType CurrentOrder(E, E);
          transform(Mask, CurrentOrder.begin(), [E](int Idx) {
            return Idx == UndefMaskElem ? E : static_cast<unsigned>(Idx);
          });
          fixupOrderingIndices(CurrentOrder);
          ++OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second;
        } else {
          ++OrdersUses.insert(std::make_pair(Order, 0)).first->second;
        }
        if (VisitedOps.insert(OpTE).second)
          OrdersUses.insert(std::make_pair(OrdersType(), 0)).first->second +=
              OpTE->UserTreeIndices.size();
        assert(OrdersUses[{}] > 0 && "Counter cannot be less than 0.");
        --OrdersUses[{}];
      }
      // If no orders - skip current nodes and jump to the next one, if any.
      if (OrdersUses.empty()) {
        for_each(Data.second,
                 [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) {
                   OrderedEntries.remove(Op.second);
                 });
        continue;
      }
      // Choose the best order.
      ArrayRef<unsigned> BestOrder = OrdersUses.front().first;
      unsigned Cnt = OrdersUses.front().second;
      for (const auto &Pair : drop_begin(OrdersUses)) {
        if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) {
          BestOrder = Pair.first;
          Cnt = Pair.second;
        }
      }
      // Set order of the user node (reordering of operands and user nodes).
      if (BestOrder.empty()) {
        for_each(Data.second,
                 [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) {
                   OrderedEntries.remove(Op.second);
                 });
        continue;
      }
      // Erase operands from OrderedEntries list and adjust their orders.
      VisitedOps.clear();
      SmallVector<int> Mask;
      inversePermutation(BestOrder, Mask);
      SmallVector<int> MaskOrder(BestOrder.size(), UndefMaskElem);
      unsigned E = BestOrder.size();
      transform(BestOrder, MaskOrder.begin(), [E](unsigned I) {
        return I < E ? static_cast<int>(I) : UndefMaskElem;
      });
      for (const std::pair<unsigned, TreeEntry *> &Op : Data.second) {
        TreeEntry *TE = Op.second;
        OrderedEntries.remove(TE);
        if (!VisitedOps.insert(TE).second)
          continue;
        if (!TE->ReuseShuffleIndices.empty() && TE->ReorderIndices.empty()) {
          // Just reorder reuses indices.
          reorderReuses(TE->ReuseShuffleIndices, Mask);
          continue;
        }
        // Gathers are processed separately.
        if (TE->State != TreeEntry::Vectorize)
          continue;
        assert((BestOrder.size() == TE->ReorderIndices.size() ||
                TE->ReorderIndices.empty()) &&
               "Non-matching sizes of user/operand entries.");
        reorderOrder(TE->ReorderIndices, Mask);
      }
      // For gathers just need to reorder its scalars.
      for (TreeEntry *Gather : GatherOps) {
        assert(Gather->ReorderIndices.empty() &&
               "Unexpected reordering of gathers.");
        if (!Gather->ReuseShuffleIndices.empty()) {
          // Just reorder reuses indices.
          reorderReuses(Gather->ReuseShuffleIndices, Mask);
          continue;
        }
        reorderScalars(Gather->Scalars, Mask);
        OrderedEntries.remove(Gather);
      }
      // Reorder operands of the user node and set the ordering for the user
      // node itself.
      if (Data.first->State != TreeEntry::Vectorize ||
          !isa<ExtractElementInst, ExtractValueInst, LoadInst>(
              Data.first->getMainOp()) ||
          Data.first->isAltShuffle())
        Data.first->reorderOperands(Mask);
      if (!isa<InsertElementInst, StoreInst>(Data.first->getMainOp()) ||
          Data.first->isAltShuffle()) {
        reorderScalars(Data.first->Scalars, Mask);
        reorderOrder(Data.first->ReorderIndices, MaskOrder);
        if (Data.first->ReuseShuffleIndices.empty() &&
            !Data.first->ReorderIndices.empty() &&
            !Data.first->isAltShuffle()) {
          // Insert user node to the list to try to sink reordering deeper in
          // the graph.
          OrderedEntries.insert(Data.first);
        }
      } else {
        reorderOrder(Data.first->ReorderIndices, Mask);
      }
    }
  }
  // If the reordering is unnecessary, just remove the reorder.
  if (IgnoreReorder && !VectorizableTree.front()->ReorderIndices.empty() &&
      VectorizableTree.front()->ReuseShuffleIndices.empty())
    VectorizableTree.front()->ReorderIndices.clear();
}
 
void BoUpSLP::buildExternalUses(
    const ExtraValueToDebugLocsMap &ExternallyUsedValues) {
  // Collect the values that we need to extract from the tree.
  for (auto &TEPtr : VectorizableTree) {
    TreeEntry *Entry = TEPtr.get();
 
    // No need to handle users of gathered values.
    if (Entry->State == TreeEntry::NeedToGather)
      continue;
 
    // For each lane:
    for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
      Value *Scalar = Entry->Scalars[Lane];
      int FoundLane = Entry->findLaneForValue(Scalar);
 
      // Check if the scalar is externally used as an extra arg.
      auto ExtI = ExternallyUsedValues.find(Scalar);
      if (ExtI != ExternallyUsedValues.end()) {
        LLVM_DEBUG(dbgs() << "SLP: Need to extract: Extra arg from lane "
                          << Lane << " from " << *Scalar << ".\n");
        ExternalUses.emplace_back(Scalar, nullptr, FoundLane);
      }
      for (User *U : Scalar->users()) {
        LLVM_DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n");
 
        Instruction *UserInst = dyn_cast<Instruction>(U);
        if (!UserInst)
          continue;
 
        if (isDeleted(UserInst))
          continue;
 
        // Skip in-tree scalars that become vectors
        if (TreeEntry *UseEntry = getTreeEntry(U)) {
          Value *UseScalar = UseEntry->Scalars[0];
          // Some in-tree scalars will remain as scalar in vectorized
          // instructions. If that is the case, the one in Lane 0 will
          // be used.
          if (UseScalar != U ||
              UseEntry->State == TreeEntry::ScatterVectorize ||
              !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) {
            LLVM_DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << *U
                              << ".\n");
            assert(UseEntry->State != TreeEntry::NeedToGather && "Bad state");
            continue;
          }
        }
 
        // Ignore users in the user ignore list.
        if (is_contained(UserIgnoreList, UserInst))
          continue;
 
        LLVM_DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane "
                          << Lane << " from " << *Scalar << ".\n");
        ExternalUses.push_back(ExternalUser(Scalar, U, FoundLane));
      }
    }
  }
}
 
void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
                        ArrayRef<Value *> UserIgnoreLst) {
  deleteTree();
  UserIgnoreList = UserIgnoreLst;
  if (!allSameType(Roots))
    return;
  buildTree_rec(Roots, 0, EdgeInfo());
}
 
namespace {
/// Tracks the state we can represent the loads in the given sequence.
enum class LoadsState { Gather, Vectorize, ScatterVectorize };
} // anonymous namespace
 
/// Checks if the given array of loads can be represented as a vectorized,
/// scatter or just simple gather.
static LoadsState canVectorizeLoads(ArrayRef<Value *> VL, const Value *VL0,
                                    const TargetTransformInfo &TTI,
                                    const DataLayout &DL, ScalarEvolution &SE,
                                    SmallVectorImpl<unsigned> &Order,
                                    SmallVectorImpl<Value *> &PointerOps) {
  // 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();
  PointerOps.resize(VL.size());
  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();
  // Check the order of pointer operands.
  if (llvm::sortPtrAccesses(PointerOps, ScalarTy, DL, SE, Order)) {
    Value *Ptr0;
    Value *PtrN;
    if (Order.empty()) {
      Ptr0 = PointerOps.front();
      PtrN = PointerOps.back();
    } else {
      Ptr0 = PointerOps[Order.front()];
      PtrN = PointerOps[Order.back()];
    }
    Optional<int> Diff =
        getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, DL, SE);
    // Check that the sorted loads are consecutive.
    if (static_cast<unsigned>(*Diff) == VL.size() - 1)
      return LoadsState::Vectorize;
    Align CommonAlignment = cast<LoadInst>(VL0)->getAlign();
    for (Value *V : VL)
      CommonAlignment =
          commonAlignment(CommonAlignment, cast<LoadInst>(V)->getAlign());
    if (TTI.isLegalMaskedGather(FixedVectorType::get(ScalarTy, VL.size()),
                                CommonAlignment))
      return LoadsState::ScatterVectorize;
  }
 
  return LoadsState::Gather;
}
 
void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
                            const EdgeInfo &UserTreeIdx) {
  assert((allConstant(VL) || allSameType(VL)) && "Invalid types!");
 
  SmallVector<int> ReuseShuffleIndicies;
  SmallVector<Value *> UniqueValues;
  auto &&TryToFindDuplicates = [&VL, &ReuseShuffleIndicies, &UniqueValues,
                                &UserTreeIdx,
                                this](const InstructionsState &S) {
    // Check that every instruction appears once in this bundle.
    DenseMap<Value *, unsigned> UniquePositions;
    for (Value *V : VL) {
      if (isConstant(V)) {
        ReuseShuffleIndicies.emplace_back(
            isa<UndefValue>(V) ? UndefMaskElem : UniqueValues.size());
        UniqueValues.emplace_back(V);
        continue;
      }
      auto Res = UniquePositions.try_emplace(V, UniqueValues.size());
      ReuseShuffleIndicies.emplace_back(Res.first->second);
      if (Res.second)
        UniqueValues.emplace_back(V);
    }
    size_t NumUniqueScalarValues = UniqueValues.size();
    if (NumUniqueScalarValues == VL.size()) {
      ReuseShuffleIndicies.clear();
    } else {
      LLVM_DEBUG(dbgs() << "SLP: Shuffle for reused scalars.\n");
      if (NumUniqueScalarValues <= 1 ||
          (UniquePositions.size() == 1 && all_of(UniqueValues,
                                                 [](Value *V) {
                                                   return isa<UndefValue>(V) ||
                                                          !isConstant(V);
                                                 })) ||
          !llvm::isPowerOf2_32(NumUniqueScalarValues)) {
        LLVM_DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
        newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx);
        return false;
      }
      VL = UniqueValues;
    }
    return true;
  };
 
  InstructionsState S = getSameOpcode(VL);
  if (Depth == RecursionMaxDepth) {
    LLVM_DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
    if (TryToFindDuplicates(S))
      newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                   ReuseShuffleIndicies);
    return;
  }
 
  // Don't handle scalable vectors
  if (S.getOpcode() == Instruction::ExtractElement &&
      isa<ScalableVectorType>(
          cast<ExtractElementInst>(S.OpValue)->getVectorOperandType())) {
    LLVM_DEBUG(dbgs() << "SLP: Gathering due to scalable vector type.\n");
    if (TryToFindDuplicates(S))
      newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                   ReuseShuffleIndicies);
    return;
  }
 
  // Don't handle vectors.
  if (S.OpValue->getType()->isVectorTy() &&
      !isa<InsertElementInst>(S.OpValue)) {
    LLVM_DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
    newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx);
    return;
  }
 
  if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue))
    if (SI->getValueOperand()->getType()->isVectorTy()) {
      LLVM_DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n");
      newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx);
      return;
    }
 
  // If all of the operands are identical or constant we have a simple solution.
  // If we deal with insert/extract instructions, they all must have constant
  // indices, otherwise we should gather them, not try to vectorize.
  if (allConstant(VL) || isSplat(VL) || !allSameBlock(VL) || !S.getOpcode() ||
      (isa<InsertElementInst, ExtractValueInst, ExtractElementInst>(S.MainOp) &&
       !all_of(VL, isVectorLikeInstWithConstOps))) {
    LLVM_DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n");
    if (TryToFindDuplicates(S))
      newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                   ReuseShuffleIndicies);
    return;
  }
 
  // We now know that this is a vector of instructions of the same type from
  // the same block.
 
  // Don't vectorize ephemeral values.
  for (Value *V : VL) {
    if (EphValues.count(V)) {
      LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *V
                        << ") is ephemeral.\n");
      newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx);
      return;
    }
  }
 
  // Check if this is a duplicate of another entry.
  if (TreeEntry *E = getTreeEntry(S.OpValue)) {
    LLVM_DEBUG(dbgs() << "SLP: \tChecking bundle: " << *S.OpValue << ".\n");
    if (!E->isSame(VL)) {
      LLVM_DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n");
      if (TryToFindDuplicates(S))
        newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                     ReuseShuffleIndicies);
      return;
    }
    // Record the reuse of the tree node.  FIXME, currently this is only used to
    // properly draw the graph rather than for the actual vectorization.
    E->UserTreeIndices.push_back(UserTreeIdx);
    LLVM_DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue
                      << ".\n");
    return;
  }
 
  // Check that none of the instructions in the bundle are already in the tree.
  for (Value *V : VL) {
    auto *I = dyn_cast<Instruction>(V);
    if (!I)
      continue;
    if (getTreeEntry(I)) {
      LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *V
                        << ") is already in tree.\n");
      if (TryToFindDuplicates(S))
        newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                     ReuseShuffleIndicies);
      return;
    }
  }
 
  // The reduction nodes (stored in UserIgnoreList) also should stay scalar.
  for (Value *V : VL) {
    if (is_contained(UserIgnoreList, V)) {
      LLVM_DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n");
      if (TryToFindDuplicates(S))
        newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                     ReuseShuffleIndicies);
      return;
    }
  }
 
  // Check that all of the users of the scalars that we want to vectorize are
  // schedulable.
  auto *VL0 = cast<Instruction>(S.OpValue);
  BasicBlock *BB = VL0->getParent();
 
  if (!DT->isReachableFromEntry(BB)) {
    // Don't go into unreachable blocks. They may contain instructions with
    // dependency cycles which confuse the final scheduling.
    LLVM_DEBUG(dbgs() << "SLP: bundle in unreachable block.\n");
    newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx);
    return;
  }
 
  // Check that every instruction appears once in this bundle.
  if (!TryToFindDuplicates(S))
    return;
 
  auto &BSRef = BlocksSchedules[BB];
  if (!BSRef)
    BSRef = std::make_unique<BlockScheduling>(BB);
 
  BlockScheduling &BS = *BSRef.get();
 
  Optional<ScheduleData *> Bundle = BS.tryScheduleBundle(VL, this, S);
  if (!Bundle) {
    LLVM_DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n");
    assert((!BS.getScheduleData(VL0) ||
            !BS.getScheduleData(VL0)->isPartOfBundle()) &&
           "tryScheduleBundle should cancelScheduling on failure");
    newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                 ReuseShuffleIndicies);
    return;
  }
  LLVM_DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
 
  unsigned ShuffleOrOp = S.isAltShuffle() ?
                (unsigned) Instruction::ShuffleVector : S.getOpcode();
  switch (ShuffleOrOp) {
    case Instruction::PHI: {
      auto *PH = cast<PHINode>(VL0);
 
      // Check for terminator values (e.g. invoke).
      for (Value *V : VL)
        for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) {
          Instruction *Term = dyn_cast<Instruction>(
              cast<PHINode>(V)->getIncomingValueForBlock(
                  PH->getIncomingBlock(I)));
          if (Term && Term->isTerminator()) {
            LLVM_DEBUG(dbgs()
                       << "SLP: Need to swizzle PHINodes (terminator use).\n");
            BS.cancelScheduling(VL, VL0);
            newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                         ReuseShuffleIndicies);
            return;
          }
        }
 
      TreeEntry *TE =
          newTreeEntry(VL, Bundle, S, UserTreeIdx, ReuseShuffleIndicies);
      LLVM_DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
 
      // Keeps the reordered operands to avoid code duplication.
      SmallVector<ValueList, 2> OperandsVec;
      for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) {
        if (!DT->isReachableFromEntry(PH->getIncomingBlock(I))) {
          ValueList Operands(VL.size(), PoisonValue::get(PH->getType()));
          TE->setOperand(I, Operands);
          OperandsVec.push_back(Operands);
          continue;
        }
        ValueList Operands;
        // Prepare the operand vector.
        for (Value *V : VL)
          Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock(
              PH->getIncomingBlock(I)));
        TE->setOperand(I, Operands);
        OperandsVec.push_back(Operands);
      }
      for (unsigned OpIdx = 0, OpE = OperandsVec.size(); OpIdx != OpE; ++OpIdx)
        buildTree_rec(OperandsVec[OpIdx], Depth + 1, {TE, OpIdx});
      return;
    }
    case Instruction::ExtractValue:
    case Instruction::ExtractElement: {
      OrdersType CurrentOrder;
      bool Reuse = canReuseExtract(VL, VL0, CurrentOrder);
      if (Reuse) {
        LLVM_DEBUG(dbgs() << "SLP: Reusing or shuffling extract sequence.\n");
        newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                     ReuseShuffleIndicies);
        // This is a special case, as it does not gather, but at the same time
        // we are not extending buildTree_rec() towards the operands.
        ValueList Op0;
        Op0.assign(VL.size(), VL0->getOperand(0));
        VectorizableTree.back()->setOperand(0, Op0);
        return;
      }
      if (!CurrentOrder.empty()) {
        LLVM_DEBUG({
          dbgs() << "SLP: Reusing or shuffling of reordered extract sequence "
                    "with order";
          for (unsigned Idx : CurrentOrder)
            dbgs() << " " << Idx;
          dbgs() << "\n";
        });
        fixupOrderingIndices(CurrentOrder);
        // Insert new order with initial value 0, if it does not exist,
        // otherwise return the iterator to the existing one.
        newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                     ReuseShuffleIndicies, CurrentOrder);
        // This is a special case, as it does not gather, but at the same time
        // we are not extending buildTree_rec() towards the operands.
        ValueList Op0;
        Op0.assign(VL.size(), VL0->getOperand(0));
        VectorizableTree.back()->setOperand(0, Op0);
        return;
      }
      LLVM_DEBUG(dbgs() << "SLP: Gather extract sequence.\n");
      newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                   ReuseShuffleIndicies);
      BS.cancelScheduling(VL, VL0);
      return;
    }
    case Instruction::InsertElement: {
      assert(ReuseShuffleIndicies.empty() && "All inserts should be unique");
 
      // Check that we have a buildvector and not a shuffle of 2 or more
      // different vectors.
      ValueSet SourceVectors;
      int MinIdx = std::numeric_limits<int>::max();
      for (Value *V : VL) {
        SourceVectors.insert(cast<Instruction>(V)->getOperand(0));
        Optional<int> Idx = *getInsertIndex(V, 0);
        if (!Idx || *Idx == UndefMaskElem)
          continue;
        MinIdx = std::min(MinIdx, *Idx);
      }
 
      if (count_if(VL, [&SourceVectors](Value *V) {
            return !SourceVectors.contains(V);
          }) >= 2) {
        // Found 2nd source vector - cancel.
        LLVM_DEBUG(dbgs() << "SLP: Gather of insertelement vectors with "
                             "different source vectors.\n");
        newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx);
        BS.cancelScheduling(VL, VL0);
        return;
      }
 
      auto OrdCompare = [](const std::pair<int, int> &P1,
                           const std::pair<int, int> &P2) {
        return P1.first > P2.first;
      };
      PriorityQueue<std::pair<int, int>, SmallVector<std::pair<int, int>>,
                    decltype(OrdCompare)>
          Indices(OrdCompare);
      for (int I = 0, E = VL.size(); I < E; ++I) {
        Optional<int> Idx = *getInsertIndex(VL[I], 0);
        if (!Idx || *Idx == UndefMaskElem)
          continue;
        Indices.emplace(*Idx, I);
      }
      OrdersType CurrentOrder(VL.size(), VL.size());
      bool IsIdentity = true;
      for (int I = 0, E = VL.size(); I < E; ++I) {
        CurrentOrder[Indices.top().second] = I;
        IsIdentity &= Indices.top().second == I;
        Indices.pop();
      }
      if (IsIdentity)
        CurrentOrder.clear();
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   None, CurrentOrder);
      LLVM_DEBUG(dbgs() << "SLP: added inserts bundle.\n");
 
      constexpr int NumOps = 2;
      ValueList VectorOperands[NumOps];
      for (int I = 0; I < NumOps; ++I) {
        for (Value *V : VL)
          VectorOperands[I].push_back(cast<Instruction>(V)->getOperand(I));
 
        TE->setOperand(I, VectorOperands[I]);
      }
      buildTree_rec(VectorOperands[NumOps - 1], Depth + 1, {TE, NumOps - 1});
      return;
    }
    case Instruction::Load: {
      // Check that a vectorized load would load the same memory as a scalar
      // load. For example, we don't want to vectorize loads that are smaller
      // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM
      // treats loading/storing it as an i8 struct. If we vectorize loads/stores
      // from such a struct, we read/write packed bits disagreeing with the
      // unvectorized version.
      SmallVector<Value *> PointerOps;
      OrdersType CurrentOrder;
      TreeEntry *TE = nullptr;
      switch (canVectorizeLoads(VL, VL0, *TTI, *DL, *SE, CurrentOrder,
                                PointerOps)) {
      case LoadsState::Vectorize:
        if (CurrentOrder.empty()) {
          // Original loads are consecutive and does not require reordering.
          TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                            ReuseShuffleIndicies);
          LLVM_DEBUG(dbgs() << "SLP: added a vector of loads.\n");
        } else {
          fixupOrderingIndices(CurrentOrder);
          // Need to reorder.
          TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                            ReuseShuffleIndicies, CurrentOrder);
          LLVM_DEBUG(dbgs() << "SLP: added a vector of jumbled loads.\n");
        }
        TE->setOperandsInOrder();
        break;
      case LoadsState::ScatterVectorize:
        // Vectorizing non-consecutive loads with `llvm.masked.gather`.
        TE = newTreeEntry(VL, TreeEntry::ScatterVectorize, Bundle, S,
                          UserTreeIdx, ReuseShuffleIndicies);
        TE->setOperandsInOrder();
        buildTree_rec(PointerOps, Depth + 1, {TE, 0});
        LLVM_DEBUG(dbgs() << "SLP: added a vector of non-consecutive loads.\n");
        break;
      case LoadsState::Gather:
        BS.cancelScheduling(VL, VL0);
        newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                     ReuseShuffleIndicies);
#ifndef NDEBUG
        Type *ScalarTy = VL0->getType();
        if (DL->getTypeSizeInBits(ScalarTy) !=
            DL->getTypeAllocSizeInBits(ScalarTy))
          LLVM_DEBUG(dbgs() << "SLP: Gathering loads of non-packed type.\n");
        else if (any_of(VL, [](Value *V) {
                   return !cast<LoadInst>(V)->isSimple();
                 }))
          LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n");
        else
          LLVM_DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n");
#endif // NDEBUG
        break;
      }
      return;
    }
    case Instruction::ZExt:
    case Instruction::SExt:
    case Instruction::FPToUI:
    case Instruction::FPToSI:
    case Instruction::FPExt:
    case Instruction::PtrToInt:
    case Instruction::IntToPtr:
    case Instruction::SIToFP:
    case Instruction::UIToFP:
    case Instruction::Trunc:
    case Instruction::FPTrunc:
    case Instruction::BitCast: {
      Type *SrcTy = VL0->getOperand(0)->getType();
      for (Value *V : VL) {
        Type *Ty = cast<Instruction>(V)->getOperand(0)->getType();
        if (Ty != SrcTy || !isValidElementType(Ty)) {
          BS.cancelScheduling(VL, VL0);
          newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                       ReuseShuffleIndicies);
          LLVM_DEBUG(dbgs()
                     << "SLP: Gathering casts with different src types.\n");
          return;
        }
      }
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndicies);
      LLVM_DEBUG(dbgs() << "SLP: added a vector of casts.\n");
 
      TE->setOperandsInOrder();
      for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
        ValueList Operands;
        // Prepare the operand vector.
        for (Value *V : VL)
          Operands.push_back(cast<Instruction>(V)->getOperand(i));
 
        buildTree_rec(Operands, Depth + 1, {TE, i});
      }
      return;
    }
    case Instruction::ICmp:
    case Instruction::FCmp: {
      // Check that all of the compares have the same predicate.
      CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate();
      CmpInst::Predicate SwapP0 = CmpInst::getSwappedPredicate(P0);
      Type *ComparedTy = VL0->getOperand(0)->getType();
      for (Value *V : VL) {
        CmpInst *Cmp = cast<CmpInst>(V);
        if ((Cmp->getPredicate() != P0 && Cmp->getPredicate() != SwapP0) ||
            Cmp->getOperand(0)->getType() != ComparedTy) {
          BS.cancelScheduling(VL, VL0);
          newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                       ReuseShuffleIndicies);
          LLVM_DEBUG(dbgs()
                     << "SLP: Gathering cmp with different predicate.\n");
          return;
        }
      }
 
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndicies);
      LLVM_DEBUG(dbgs() << "SLP: added a vector of compares.\n");
 
      ValueList Left, Right;
      if (cast<CmpInst>(VL0)->isCommutative()) {
        // Commutative predicate - collect + sort operands of the instructions
        // so that each side is more likely to have the same opcode.
        assert(P0 == SwapP0 && "Commutative Predicate mismatch");
        reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this);
      } else {
        // Collect operands - commute if it uses the swapped predicate.
        for (Value *V : VL) {
          auto *Cmp = cast<CmpInst>(V);
          Value *LHS = Cmp->getOperand(0);
          Value *RHS = Cmp->getOperand(1);
          if (Cmp->getPredicate() != P0)
            std::swap(LHS, RHS);
          Left.push_back(LHS);
          Right.push_back(RHS);
        }
      }
      TE->setOperand(0, Left);
      TE->setOperand(1, Right);
      buildTree_rec(Left, Depth + 1, {TE, 0});
      buildTree_rec(Right, Depth + 1, {TE, 1});
      return;
    }
    case Instruction::Select:
    case Instruction::FNeg:
    case Instruction::Add:
    case Instruction::FAdd:
    case Instruction::Sub:
    case Instruction::FSub:
    case Instruction::Mul:
    case Instruction::FMul:
    case Instruction::UDiv:
    case Instruction::SDiv:
    case Instruction::FDiv:
    case Instruction::URem:
    case Instruction::SRem:
    case Instruction::FRem:
    case Instruction::Shl:
    case Instruction::LShr:
    case Instruction::AShr:
    case Instruction::And:
    case Instruction::Or:
    case Instruction::Xor: {
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndicies);
      LLVM_DEBUG(dbgs() << "SLP: added a vector of un/bin op.\n");
 
      // Sort operands of the instructions so that each side is more likely to
      // have the same opcode.
      if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
        ValueList Left, Right;
        reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this);
        TE->setOperand(0, Left);
        TE->setOperand(1, Right);
        buildTree_rec(Left, Depth + 1, {TE, 0});
        buildTree_rec(Right, Depth + 1, {TE, 1});
        return;
      }
 
      TE->setOperandsInOrder();
      for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
        ValueList Operands;
        // Prepare the operand vector.
        for (Value *V : VL)
          Operands.push_back(cast<Instruction>(V)->getOperand(i));
 
        buildTree_rec(Operands, Depth + 1, {TE, i});
      }
      return;
    }
    case Instruction::GetElementPtr: {
      // We don't combine GEPs with complicated (nested) indexing.
      for (Value *V : VL) {
        if (cast<Instruction>(V)->getNumOperands() != 2) {
          LLVM_DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n");
          BS.cancelScheduling(VL, VL0);
          newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                       ReuseShuffleIndicies);
          return;
        }
      }
 
      // We can't combine several GEPs into one vector if they operate on
      // different types.
      Type *Ty0 = VL0->getOperand(0)->getType();
      for (Value *V : VL) {
        Type *CurTy = cast<Instruction>(V)->getOperand(0)->getType();
        if (Ty0 != CurTy) {
          LLVM_DEBUG(dbgs()
                     << "SLP: not-vectorizable GEP (different types).\n");
          BS.cancelScheduling(VL, VL0);
          newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                       ReuseShuffleIndicies);
          return;
        }
      }
 
      // We don't combine GEPs with non-constant indexes.
      Type *Ty1 = VL0->getOperand(1)->getType();
      for (Value *V : VL) {
        auto Op = cast<Instruction>(V)->getOperand(1);
        if (!isa<ConstantInt>(Op) ||
            (Op->getType() != Ty1 &&
             Op->getType()->getScalarSizeInBits() >
                 DL->getIndexSizeInBits(
                     V->getType()->getPointerAddressSpace()))) {
          LLVM_DEBUG(dbgs()
                     << "SLP: not-vectorizable GEP (non-constant indexes).\n");
          BS.cancelScheduling(VL, VL0);
          newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                       ReuseShuffleIndicies);
          return;
        }
      }
 
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndicies);
      LLVM_DEBUG(dbgs() << "SLP: added a vector of GEPs.\n");
      SmallVector<ValueList, 2> Operands(2);
      // Prepare the operand vector for pointer operands.
      for (Value *V : VL)
        Operands.front().push_back(
            cast<GetElementPtrInst>(V)->getPointerOperand());
      TE->setOperand(0, Operands.front());
      // Need to cast all indices to the same type before vectorization to
      // avoid crash.
      // Required to be able to find correct matches between different gather
      // nodes and reuse the vectorized values rather than trying to gather them
      // again.
      int IndexIdx = 1;
      Type *VL0Ty = VL0->getOperand(IndexIdx)->getType();
      Type *Ty = all_of(VL,
                        [VL0Ty, IndexIdx](Value *V) {
                          return VL0Ty == cast<GetElementPtrInst>(V)
                                              ->getOperand(IndexIdx)
                                              ->getType();
                        })
                     ? VL0Ty
                     : DL->getIndexType(cast<GetElementPtrInst>(VL0)
                                            ->getPointerOperandType()
                                            ->getScalarType());
      // Prepare the operand vector.
      for (Value *V : VL) {
        auto *Op = cast<Instruction>(V)->getOperand(IndexIdx);
        auto *CI = cast<ConstantInt>(Op);
        Operands.back().push_back(ConstantExpr::getIntegerCast(
            CI, Ty, CI->getValue().isSignBitSet()));
      }
      TE->setOperand(IndexIdx, Operands.back());
 
      for (unsigned I = 0, Ops = Operands.size(); I < Ops; ++I)
        buildTree_rec(Operands[I], Depth + 1, {TE, I});
      return;
    }
    case Instruction::Store: {
      // Check if the stores are consecutive or if we need to swizzle them.
      llvm::Type *ScalarTy = cast<StoreInst>(VL0)->getValueOperand()->getType();
      // Avoid types that are padded when being allocated as scalars, while
      // being packed together in a vector (such as i1).
      if (DL->getTypeSizeInBits(ScalarTy) !=
          DL->getTypeAllocSizeInBits(ScalarTy)) {
        BS.cancelScheduling(VL, VL0);
        newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                     ReuseShuffleIndicies);
        LLVM_DEBUG(dbgs() << "SLP: Gathering stores of non-packed type.\n");
        return;
      }
      // Make sure all stores in the bundle are simple - we can't vectorize
      // atomic or volatile stores.
      SmallVector<Value *, 4> PointerOps(VL.size());
      ValueList Operands(VL.size());
      auto POIter = PointerOps.begin();
      auto OIter = Operands.begin();
      for (Value *V : VL) {
        auto *SI = cast<StoreInst>(V);
        if (!SI->isSimple()) {
          BS.cancelScheduling(VL, VL0);
          newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                       ReuseShuffleIndicies);
          LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple stores.\n");
          return;
        }
        *POIter = SI->getPointerOperand();
        *OIter = SI->getValueOperand();
        ++POIter;
        ++OIter;
      }
 
      OrdersType CurrentOrder;
      // Check the order of pointer operands.
      if (llvm::sortPtrAccesses(PointerOps, ScalarTy, *DL, *SE, CurrentOrder)) {
        Value *Ptr0;
        Value *PtrN;
        if (CurrentOrder.empty()) {
          Ptr0 = PointerOps.front();
          PtrN = PointerOps.back();
        } else {
          Ptr0 = PointerOps[CurrentOrder.front()];
          PtrN = PointerOps[CurrentOrder.back()];
        }
        Optional<int> Dist =
            getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, *DL, *SE);
        // Check that the sorted pointer operands are consecutive.
        if (static_cast<unsigned>(*Dist) == VL.size() - 1) {
          if (CurrentOrder.empty()) {
            // Original stores are consecutive and does not require reordering.
            TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S,
                                         UserTreeIdx, ReuseShuffleIndicies);
            TE->setOperandsInOrder();
            buildTree_rec(Operands, Depth + 1, {TE, 0});
            LLVM_DEBUG(dbgs() << "SLP: added a vector of stores.\n");
          } else {
            fixupOrderingIndices(CurrentOrder);
            TreeEntry *TE =
                newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                             ReuseShuffleIndicies, CurrentOrder);
            TE->setOperandsInOrder();
            buildTree_rec(Operands, Depth + 1, {TE, 0});
            LLVM_DEBUG(dbgs() << "SLP: added a vector of jumbled stores.\n");
          }
          return;
        }
      }
 
      BS.cancelScheduling(VL, VL0);
      newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                   ReuseShuffleIndicies);
      LLVM_DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
      return;
    }
    case Instruction::Call: {
      // Check if the calls are all to the same vectorizable intrinsic or
      // library function.
      CallInst *CI = cast<CallInst>(VL0);
      Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
 
      VFShape Shape = VFShape::get(
          *CI, ElementCount::getFixed(static_cast<unsigned int>(VL.size())),
          false /*HasGlobalPred*/);
      Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape);
 
      if (!VecFunc && !isTriviallyVectorizable(ID)) {
        BS.cancelScheduling(VL, VL0);
        newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                     ReuseShuffleIndicies);
        LLVM_DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
        return;
      }
      Function *F = CI->getCalledFunction();
      unsigned NumArgs = CI->arg_size();
      SmallVector<Value*, 4> ScalarArgs(NumArgs, nullptr);
      for (unsigned j = 0; j != NumArgs; ++j)
        if (hasVectorInstrinsicScalarOpd(ID, j))
          ScalarArgs[j] = CI->getArgOperand(j);
      for (Value *V : VL) {
        CallInst *CI2 = dyn_cast<CallInst>(V);
        if (!CI2 || CI2->getCalledFunction() != F ||
            getVectorIntrinsicIDForCall(CI2, TLI) != ID ||
            (VecFunc &&
             VecFunc != VFDatabase(*CI2).getVectorizedFunction(Shape)) ||
            !CI->hasIdenticalOperandBundleSchema(*CI2)) {
          BS.cancelScheduling(VL, VL0);
          newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                       ReuseShuffleIndicies);
          LLVM_DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *V
                            << "\n");
          return;
        }
        // Some intrinsics have scalar arguments and should be same in order for
        // them to be vectorized.
        for (unsigned j = 0; j != NumArgs; ++j) {
          if (hasVectorInstrinsicScalarOpd(ID, j)) {
            Value *A1J = CI2->getArgOperand(j);
            if (ScalarArgs[j] != A1J) {
              BS.cancelScheduling(VL, VL0);
              newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                           ReuseShuffleIndicies);
              LLVM_DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CI
                                << " argument " << ScalarArgs[j] << "!=" << A1J
                                << "\n");
              return;
            }
          }
        }
        // Verify that the bundle operands are identical between the two calls.
        if (CI->hasOperandBundles() &&
            !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(),
                        CI->op_begin() + CI->getBundleOperandsEndIndex(),
                        CI2->op_begin() + CI2->getBundleOperandsStartIndex())) {
          BS.cancelScheduling(VL, VL0);
          newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                       ReuseShuffleIndicies);
          LLVM_DEBUG(dbgs() << "SLP: mismatched bundle operands in calls:"
                            << *CI << "!=" << *V << '\n');
          return;
        }
      }
 
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndicies);
      TE->setOperandsInOrder();
      for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) {
        // For scalar operands no need to to create an entry since no need to
        // vectorize it.
        if (hasVectorInstrinsicScalarOpd(ID, i))
          continue;
        ValueList Operands;
        // Prepare the operand vector.
        for (Value *V : VL) {
          auto *CI2 = cast<CallInst>(V);
          Operands.push_back(CI2->getArgOperand(i));
        }
        buildTree_rec(Operands, Depth + 1, {TE, i});
      }
      return;
    }
    case Instruction::ShuffleVector: {
      // If this is not an alternate sequence of opcode like add-sub
      // then do not vectorize this instruction.
      if (!S.isAltShuffle()) {
        BS.cancelScheduling(VL, VL0);
        newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                     ReuseShuffleIndicies);
        LLVM_DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n");
        return;
      }
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndicies);
      LLVM_DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n");
 
      // Reorder operands if reordering would enable vectorization.
      if (isa<BinaryOperator>(VL0)) {
        ValueList Left, Right;
        reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this);
        TE->setOperand(0, Left);
        TE->setOperand(1, Right);
        buildTree_rec(Left, Depth + 1, {TE, 0});
        buildTree_rec(Right, Depth + 1, {TE, 1});
        return;
      }
 
      TE->setOperandsInOrder();
      for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
        ValueList Operands;
        // Prepare the operand vector.
        for (Value *V : VL)
          Operands.push_back(cast<Instruction>(V)->getOperand(i));
 
        buildTree_rec(Operands, Depth + 1, {TE, i});
      }
      return;
    }
    default:
      BS.cancelScheduling(VL, VL0);
      newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx,
                   ReuseShuffleIndicies);
      LLVM_DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
      return;
  }
}
 
unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const {
  unsigned N = 1;
  Type *EltTy = T;
 
  while (isa<StructType>(EltTy) || isa<ArrayType>(EltTy) ||
         isa<VectorType>(EltTy)) {
    if (auto *ST = dyn_cast<StructType>(EltTy)) {
      // Check that struct is homogeneous.
      for (const auto *Ty : ST->elements())
        if (Ty != *ST->element_begin())
          return 0;
      N *= ST->getNumElements();
      EltTy = *ST->element_begin();
    } else if (auto *AT = dyn_cast<ArrayType>(EltTy)) {
      N *= AT->getNumElements();
      EltTy = AT->getElementType();
    } else {
      auto *VT = cast<FixedVectorType>(EltTy);
      N *= VT->getNumElements();
      EltTy = VT->getElementType();
    }
  }
 
  if (!isValidElementType(EltTy))
    return 0;
  uint64_t VTSize = DL.getTypeStoreSizeInBits(FixedVectorType::get(EltTy, N));
  if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T))
    return 0;
  return N;
}
 
bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, Value *OpValue,
                              SmallVectorImpl<unsigned> &CurrentOrder) const {
  const auto *It = find_if(VL, [](Value *V) {
    return isa<ExtractElementInst, ExtractValueInst>(V);
  });
  assert(It != VL.end() && "Expected at least one extract instruction.");
  auto *E0 = cast<Instruction>(*It);
  assert(all_of(VL,
                [](Value *V) {
                  return isa<UndefValue, ExtractElementInst, ExtractValueInst>(
                      V);
                }) &&
         "Invalid opcode");
  // Check if all of the extracts come from the same vector and from the
  // correct offset.
  Value *Vec = E0->getOperand(0);
 
  CurrentOrder.clear();
 
  // We have to extract from a vector/aggregate with the same number of elements.
  unsigned NElts;
  if (E0->getOpcode() == Instruction::ExtractValue) {
    const DataLayout &DL = E0->getModule()->getDataLayout();
    NElts = canMapToVector(Vec->getType(), DL);
    if (!NElts)
      return false;
    // Check if load can be rewritten as load of vector.
    LoadInst *LI = dyn_cast<LoadInst>(Vec);
    if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size()))
      return false;
  } else {
    NElts = cast<FixedVectorType>(Vec->getType())->getNumElements();
  }
 
  if (NElts != VL.size())
    return false;
 
  // Check that all of the indices extract from the correct offset.
  bool ShouldKeepOrder = true;
  unsigned E = VL.size();
  // Assign to all items the initial value E + 1 so we can check if the extract
  // instruction index was used already.
  // Also, later we can check that all the indices are used and we have a
  // consecutive access in the extract instructions, by checking that no
  // element of CurrentOrder still has value E + 1.
  CurrentOrder.assign(E, E);
  unsigned I = 0;
  for (; I < E; ++I) {
    auto *Inst = dyn_cast<Instruction>(VL[I]);
    if (!Inst)
      continue;
    if (Inst->getOperand(0) != Vec)
      break;
    if (auto *EE = dyn_cast<ExtractElementInst>(Inst))
      if (isa<UndefValue>(EE->getIndexOperand()))
        continue;
    Optional<unsigned> Idx = getExtractIndex(Inst);
    if (!Idx)
      break;
    const unsigned ExtIdx = *Idx;
    if (ExtIdx != I) {
      if (ExtIdx >= E || CurrentOrder[ExtIdx] != E)
        break;
      ShouldKeepOrder = false;
      CurrentOrder[ExtIdx] = I;
    } else {
      if (CurrentOrder[I] != E)
        break;
      CurrentOrder[I] = I;
    }
  }
  if (I < E) {
    CurrentOrder.clear();
    return false;
  }
  if (ShouldKeepOrder)
    CurrentOrder.clear();
 
  return ShouldKeepOrder;
}
 
bool BoUpSLP::areAllUsersVectorized(Instruction *I,
                                    ArrayRef<Value *> VectorizedVals) const {
  return (I->hasOneUse() && is_contained(VectorizedVals, I)) ||
         all_of(I->users(), [this](User *U) {
           return ScalarToTreeEntry.count(U) > 0 || MustGather.contains(U);
         });
}
 
static std::pair<InstructionCost, InstructionCost>
getVectorCallCosts(CallInst *CI, FixedVectorType *VecTy,
                   TargetTransformInfo *TTI, TargetLibraryInfo *TLI) {
  Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
 
  // Calculate the cost of the scalar and vector calls.
  SmallVector<Type *, 4> VecTys;
  for (Use &Arg : CI->args())
    VecTys.push_back(
        FixedVectorType::get(Arg->getType(), VecTy->getNumElements()));
  FastMathFlags FMF;
  if (auto *FPCI = dyn_cast<FPMathOperator>(CI))
    FMF = FPCI->getFastMathFlags();
  SmallVector<const Value *> Arguments(CI->args());
  IntrinsicCostAttributes CostAttrs(ID, VecTy, Arguments, VecTys, FMF,
                                    dyn_cast<IntrinsicInst>(CI));
  auto IntrinsicCost =
    TTI->getIntrinsicInstrCost(CostAttrs, TTI::TCK_RecipThroughput);
 
  auto Shape = VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>(
                                     VecTy->getNumElements())),
                            false /*HasGlobalPred*/);
  Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape);
  auto LibCost = IntrinsicCost;
  if (!CI->isNoBuiltin() && VecFunc) {
    // Calculate the cost of the vector library call.
    // If the corresponding vector call is cheaper, return its cost.
    LibCost = TTI->getCallInstrCost(nullptr, VecTy, VecTys,
                                    TTI::TCK_RecipThroughput);
  }
  return {IntrinsicCost, LibCost};
}
 
/// Compute the cost of creating a vector of type \p VecTy containing the
/// extracted values from \p VL.
static InstructionCost
computeExtractCost(ArrayRef<Value *> VL, FixedVectorType *VecTy,
                   TargetTransformInfo::ShuffleKind ShuffleKind,
                   ArrayRef<int> Mask, TargetTransformInfo &TTI) {
  unsigned NumOfParts = TTI.getNumberOfParts(VecTy);
 
  if (ShuffleKind != TargetTransformInfo::SK_PermuteSingleSrc || !NumOfParts ||
      VecTy->getNumElements() < NumOfParts)
    return TTI.getShuffleCost(ShuffleKind, VecTy, Mask);
 
  bool AllConsecutive = true;
  unsigned EltsPerVector = VecTy->getNumElements() / NumOfParts;
  unsigned Idx = -1;
  InstructionCost Cost = 0;
 
  // Process extracts in blocks of EltsPerVector to check if the source vector
  // operand can be re-used directly. If not, add the cost of creating a shuffle
  // to extract the values into a vector register.
  for (auto *V : VL) {
    ++Idx;
 
    // Need to exclude undefs from analysis.
    if (isa<UndefValue>(V) || Mask[Idx] == UndefMaskElem)
      continue;
 
    // Reached the start of a new vector registers.
    if (Idx % EltsPerVector == 0) {
      AllConsecutive = true;
      continue;
    }
 
    // Check all extracts for a vector register on the target directly
    // extract values in order.
    unsigned CurrentIdx = *getExtractIndex(cast<Instruction>(V));
    if (!isa<UndefValue>(VL[Idx - 1]) && Mask[Idx - 1] != UndefMaskElem) {
      unsigned PrevIdx = *getExtractIndex(cast<Instruction>(VL[Idx - 1]));
      AllConsecutive &= PrevIdx + 1 == CurrentIdx &&
                        CurrentIdx % EltsPerVector == Idx % EltsPerVector;
    }
 
    if (AllConsecutive)
      continue;
 
    // Skip all indices, except for the last index per vector block.
    if ((Idx + 1) % EltsPerVector != 0 && Idx + 1 != VL.size())
      continue;
 
    // If we have a series of extracts which are not consecutive and hence
    // cannot re-use the source vector register directly, compute the shuffle
    // cost to extract the a vector with EltsPerVector elements.
    Cost += TTI.getShuffleCost(
        TargetTransformInfo::SK_PermuteSingleSrc,
        FixedVectorType::get(VecTy->getElementType(), EltsPerVector));
  }
  return Cost;
}
 
/// Build shuffle mask for shuffle graph entries and lists of main and alternate
/// operations operands.
static void
buildSuffleEntryMask(ArrayRef<Value *> VL, ArrayRef<unsigned> ReorderIndices,
                     ArrayRef<int> ReusesIndices,
                     const function_ref<bool(Instruction *)> IsAltOp,
                     SmallVectorImpl<int> &Mask,
                     SmallVectorImpl<Value *> *OpScalars = nullptr,
                     SmallVectorImpl<Value *> *AltScalars = nullptr) {
  unsigned Sz = VL.size();
  Mask.assign(Sz, UndefMaskElem);
  SmallVector<int> OrderMask;
  if (!ReorderIndices.empty())
    inversePermutation(ReorderIndices, OrderMask);
  for (unsigned I = 0; I < Sz; ++I) {
    unsigned Idx = I;
    if (!ReorderIndices.empty())
      Idx = OrderMask[I];
    auto *OpInst = cast<Instruction>(VL[Idx]);
    if (IsAltOp(OpInst)) {
      Mask[I] = Sz + Idx;
      if (AltScalars)
        AltScalars->push_back(OpInst);
    } else {
      Mask[I] = Idx;
      if (OpScalars)
        OpScalars->push_back(OpInst);
    }
  }
  if (!ReusesIndices.empty()) {
    SmallVector<int> NewMask(ReusesIndices.size(), UndefMaskElem);
    transform(ReusesIndices, NewMask.begin(), [&Mask](int Idx) {
      return Idx != UndefMaskElem ? Mask[Idx] : UndefMaskElem;
    });
    Mask.swap(NewMask);
  }
}
 
InstructionCost BoUpSLP::getEntryCost(const TreeEntry *E,
                                      ArrayRef<Value *> VectorizedVals) {
  ArrayRef<Value*> VL = E->Scalars;
 
  Type *ScalarTy = VL[0]->getType();
  if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
    ScalarTy = SI->getValueOperand()->getType();
  else if (CmpInst *CI = dyn_cast<CmpInst>(VL[0]))
    ScalarTy = CI->getOperand(0)->getType();
  else if (auto *IE = dyn_cast<InsertElementInst>(VL[0]))
    ScalarTy = IE->getOperand(1)->getType();
  auto *VecTy = FixedVectorType::get(ScalarTy, VL.size());
  TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
 
  // If we have computed a smaller type for the expression, update VecTy so
  // that the costs will be accurate.
  if (MinBWs.count(VL[0]))
    VecTy = FixedVectorType::get(
        IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size());
  unsigned EntryVF = E->getVectorFactor();
  auto *FinalVecTy = FixedVectorType::get(VecTy->getElementType(), EntryVF);
 
  bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty();
  // FIXME: it tries to fix a problem with MSVC buildbots.
  TargetTransformInfo &TTIRef = *TTI;
  auto &&AdjustExtractsCost = [this, &TTIRef, CostKind, VL, VecTy,
                               VectorizedVals, E](InstructionCost &Cost) {
    DenseMap<Value *, int> ExtractVectorsTys;
    SmallPtrSet<Value *, 4> CheckedExtracts;
    for (auto *V : VL) {
      if (isa<UndefValue>(V))
        continue;
      // If all users of instruction are going to be vectorized and this
      // instruction itself is not going to be vectorized, consider this
      // instruction as dead and remove its cost from the final cost of the
      // vectorized tree.
      // Also, avoid adjusting the cost for extractelements with multiple uses
      // in different graph entries.
      const TreeEntry *VE = getTreeEntry(V);
      if (!CheckedExtracts.insert(V).second ||
          !areAllUsersVectorized(cast<Instruction>(V), VectorizedVals) ||
          (VE && VE != E))
        continue;
      auto *EE = cast<ExtractElementInst>(V);
      Optional<unsigned> EEIdx = getExtractIndex(EE);
      if (!EEIdx)
        continue;
      unsigned Idx = *EEIdx;
      if (TTIRef.getNumberOfParts(VecTy) !=
          TTIRef.getNumberOfParts(EE->getVectorOperandType())) {
        auto It =
            ExtractVectorsTys.try_emplace(EE->getVectorOperand(), Idx).first;
        It->getSecond() = std::min<int>(It->second, Idx);
      }
      // Take credit for instruction that will become dead.
      if (EE->hasOneUse()) {
        Instruction *Ext = EE->user_back();
        if ((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
            all_of(Ext->users(),
                   [](User *U) { return isa<GetElementPtrInst>(U); })) {
          // Use getExtractWithExtendCost() to calculate the cost of
          // extractelement/ext pair.
          Cost -=
              TTIRef.getExtractWithExtendCost(Ext->getOpcode(), Ext->getType(),
                                              EE->getVectorOperandType(), Idx);
          // Add back the cost of s|zext which is subtracted separately.
          Cost += TTIRef.getCastInstrCost(
              Ext->getOpcode(), Ext->getType(), EE->getType(),
              TTI::getCastContextHint(Ext), CostKind, Ext);
          continue;
        }
      }
      Cost -= TTIRef.getVectorInstrCost(Instruction::ExtractElement,
                                        EE->getVectorOperandType(), Idx);
    }
    // Add a cost for subvector extracts/inserts if required.
    for (const auto &Data : ExtractVectorsTys) {
      auto *EEVTy = cast<FixedVectorType>(Data.first->getType());
      unsigned NumElts = VecTy->getNumElements();
      if (Data.second % NumElts == 0)
        continue;
      if (TTIRef.getNumberOfParts(EEVTy) > TTIRef.getNumberOfParts(VecTy)) {
        unsigned Idx = (Data.second / NumElts) * NumElts;
        unsigned EENumElts = EEVTy->getNumElements();
        if (Idx + NumElts <= EENumElts) {
          Cost +=
              TTIRef.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector,
                                    EEVTy, None, Idx, VecTy);
        } else {
          // Need to round up the subvector type vectorization factor to avoid a
          // crash in cost model functions. Make SubVT so that Idx + VF of SubVT
          // <= EENumElts.
          auto *SubVT =
              FixedVectorType::get(VecTy->getElementType(), EENumElts - Idx);
          Cost +=
              TTIRef.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector,
                                    EEVTy, None, Idx, SubVT);
        }
      } else {
        Cost += TTIRef.getShuffleCost(TargetTransformInfo::SK_InsertSubvector,
                                      VecTy, None, 0, EEVTy);
      }
    }
  };
  if (E->State == TreeEntry::NeedToGather) {
    if (allConstant(VL))
      return 0;
    if (isa<InsertElementInst>(VL[0]))
      return InstructionCost::getInvalid();
    SmallVector<int> Mask;
    SmallVector<const TreeEntry *> Entries;
    Optional<TargetTransformInfo::ShuffleKind> Shuffle =
        isGatherShuffledEntry(E, Mask, Entries);
    if (Shuffle.hasValue()) {
      InstructionCost GatherCost = 0;
      if (ShuffleVectorInst::isIdentityMask(Mask)) {
        // Perfect match in the graph, will reuse the previously vectorized
        // node. Cost is 0.
        LLVM_DEBUG(
            dbgs()
            << "SLP: perfect diamond match for gather bundle that starts with "
            << *VL.front() << ".\n");
        if (NeedToShuffleReuses)
          GatherCost =
              TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc,
                                  FinalVecTy, E->ReuseShuffleIndices);
      } else {
        LLVM_DEBUG(dbgs() << "SLP: shuffled " << Entries.size()
                          << " entries for bundle that starts with "
                          << *VL.front() << ".\n");
        // Detected that instead of gather we can emit a shuffle of single/two
        // previously vectorized nodes. Add the cost of the permutation rather
        // than gather.
        ::addMask(Mask, E->ReuseShuffleIndices);
        GatherCost = TTI->getShuffleCost(*Shuffle, FinalVecTy, Mask);
      }
      return GatherCost;
    }
    if ((E->getOpcode() == Instruction::ExtractElement ||
         all_of(E->Scalars,
                [](Value *V) {
                  return isa<ExtractElementInst, UndefValue>(V);
                })) &&
        allSameType(VL)) {
      // Check that gather of extractelements can be represented as just a
      // shuffle of a single/two vectors the scalars are extracted from.
      SmallVector<int> Mask;
      Optional<TargetTransformInfo::ShuffleKind> ShuffleKind =
          isFixedVectorShuffle(VL, Mask);
      if (ShuffleKind.hasValue()) {
        // Found the bunch of extractelement instructions that must be gathered
        // into a vector and can be represented as a permutation elements in a
        // single input vector or of 2 input vectors.
        InstructionCost Cost =
            computeExtractCost(VL, VecTy, *ShuffleKind, Mask, *TTI);
        AdjustExtractsCost(Cost);
        if (NeedToShuffleReuses)
          Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc,
                                      FinalVecTy, E->ReuseShuffleIndices);
        return Cost;
      }
    }
    if (isSplat(VL)) {
      // Found the broadcasting of the single scalar, calculate the cost as the
      // broadcast.
      assert(VecTy == FinalVecTy &&
             "No reused scalars expected for broadcast.");
      return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy);
    }
    InstructionCost ReuseShuffleCost = 0;
    if (NeedToShuffleReuses)
      ReuseShuffleCost = TTI->getShuffleCost(
          TTI::SK_PermuteSingleSrc, FinalVecTy, E->ReuseShuffleIndices);
    // Improve gather cost for gather of loads, if we can group some of the
    // loads into vector loads.
    if (VL.size() > 2 && E->getOpcode() == Instruction::Load &&
        !E->isAltShuffle()) {
      BoUpSLP::ValueSet VectorizedLoads;
      unsigned StartIdx = 0;
      unsigned VF = VL.size() / 2;
      unsigned VectorizedCnt = 0;
      unsigned ScatterVectorizeCnt = 0;
      const unsigned Sz = DL->getTypeSizeInBits(E->getMainOp()->getType());
      for (unsigned MinVF = getMinVF(2 * Sz); VF >= MinVF; VF /= 2) {
        for (unsigned Cnt = StartIdx, End = VL.size(); Cnt + VF <= End;
             Cnt += VF) {
          ArrayRef<Value *> Slice = VL.slice(Cnt, VF);
          if (!VectorizedLoads.count(Slice.front()) &&
              !VectorizedLoads.count(Slice.back()) && allSameBlock(Slice)) {
            SmallVector<Value *> PointerOps;
            OrdersType CurrentOrder;
            LoadsState LS = canVectorizeLoads(Slice, Slice.front(), *TTI, *DL,
                                              *SE, CurrentOrder, PointerOps);
            switch (LS) {
            case LoadsState::Vectorize:
            case LoadsState::ScatterVectorize:
              // Mark the vectorized loads so that we don't vectorize them
              // again.
              if (LS == LoadsState::Vectorize)
                ++VectorizedCnt;
              else
                ++ScatterVectorizeCnt;
              VectorizedLoads.insert(Slice.begin(), Slice.end());
              // If we vectorized initial block, no need to try to vectorize it
              // again.
              if (Cnt == StartIdx)
                StartIdx += VF;
              break;
            case LoadsState::Gather:
              break;
            }
          }
        }
        // Check if the whole array was vectorized already - exit.
        if (StartIdx >= VL.size())
          break;
        // Found vectorizable parts - exit.
        if (!VectorizedLoads.empty())
          break;
      }
      if (!VectorizedLoads.empty()) {
        InstructionCost GatherCost = 0;
        unsigned NumParts = TTI->getNumberOfParts(VecTy);
        bool NeedInsertSubvectorAnalysis =
            !NumParts || (VL.size() / VF) > NumParts;
        // Get the cost for gathered loads.
        for (unsigned I = 0, End = VL.size(); I < End; I += VF) {
          if (VectorizedLoads.contains(VL[I]))
            continue;
          GatherCost += getGatherCost(VL.slice(I, VF));
        }
        // The cost for vectorized loads.
        InstructionCost ScalarsCost = 0;
        for (Value *V : VectorizedLoads) {
          auto *LI = cast<LoadInst>(V);
          ScalarsCost += TTI->getMemoryOpCost(
              Instruction::Load, LI->getType(), LI->getAlign(),
              LI->getPointerAddressSpace(), CostKind, LI);
        }
        auto *LI = cast<LoadInst>(E->getMainOp());
        auto *LoadTy = FixedVectorType::get(LI->getType(), VF);
        Align Alignment = LI->getAlign();
        GatherCost +=
            VectorizedCnt *
            TTI->getMemoryOpCost(Instruction::Load, LoadTy, Alignment,
                                 LI->getPointerAddressSpace(), CostKind, LI);
        GatherCost += ScatterVectorizeCnt *
                      TTI->getGatherScatterOpCost(
                          Instruction::Load, LoadTy, LI->getPointerOperand(),
                          /*VariableMask=*/false, Alignment, CostKind, LI);
        if (NeedInsertSubvectorAnalysis) {
          // Add the cost for the subvectors insert.
          for (int I = VF, E = VL.size(); I < E; I += VF)
            GatherCost += TTI->getShuffleCost(TTI::SK_InsertSubvector, VecTy,
                                              None, I, LoadTy);
        }
        return ReuseShuffleCost + GatherCost - ScalarsCost;
      }
    }
    return ReuseShuffleCost + getGatherCost(VL);
  }
  InstructionCost CommonCost = 0;
  SmallVector<int> Mask;
  if (!E->ReorderIndices.empty()) {
    SmallVector<int> NewMask;
    if (E->getOpcode() == Instruction::Store) {
      // For stores the order is actually a mask.
      NewMask.resize(E->ReorderIndices.size());
      copy(E->ReorderIndices, NewMask.begin());
    } else {
      inversePermutation(E->ReorderIndices, NewMask);
    }
    ::addMask(Mask, NewMask);
  }
  if (NeedToShuffleReuses)
    ::addMask(Mask, E->ReuseShuffleIndices);
  if (!Mask.empty() && !ShuffleVectorInst::isIdentityMask(Mask))
    CommonCost =
        TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, FinalVecTy, Mask);
  assert((E->State == TreeEntry::Vectorize ||
          E->State == TreeEntry::ScatterVectorize) &&
         "Unhandled state");
  assert(E->getOpcode() && allSameType(VL) && allSameBlock(VL) && "Invalid VL");
  Instruction *VL0 = E->getMainOp();
  unsigned ShuffleOrOp =
      E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode();
  switch (ShuffleOrOp) {
    case Instruction::PHI:
      return 0;
 
    case Instruction::ExtractValue:
    case Instruction::ExtractElement: {
      // The common cost of removal ExtractElement/ExtractValue instructions +
      // the cost of shuffles, if required to resuffle the original vector.
      if (NeedToShuffleReuses) {
        unsigned Idx = 0;
        for (unsigned I : E->ReuseShuffleIndices) {
          if (ShuffleOrOp == Instruction::ExtractElement) {
            auto *EE = cast<ExtractElementInst>(VL[I]);
            CommonCost -= TTI->getVectorInstrCost(Instruction::ExtractElement,
                                                  EE->getVectorOperandType(),
                                                  *getExtractIndex(EE));
          } else {
            CommonCost -= TTI->getVectorInstrCost(Instruction::ExtractElement,
                                                  VecTy, Idx);
            ++Idx;
          }
        }
        Idx = EntryVF;
        for (Value *V : VL) {
          if (ShuffleOrOp == Instruction::ExtractElement) {
            auto *EE = cast<ExtractElementInst>(V);
            CommonCost += TTI->getVectorInstrCost(Instruction::ExtractElement,
                                                  EE->getVectorOperandType(),
                                                  *getExtractIndex(EE));
          } else {
            --Idx;
            CommonCost += TTI->getVectorInstrCost(Instruction::ExtractElement,
                                                  VecTy, Idx);
          }
        }
      }
      if (ShuffleOrOp == Instruction::ExtractValue) {
        for (unsigned I = 0, E = VL.size(); I < E; ++I) {
          auto *EI = cast<Instruction>(VL[I]);
          // Take credit for instruction that will become dead.
          if (EI->hasOneUse()) {
            Instruction *Ext = EI->user_back();
            if ((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
                all_of(Ext->users(),
                       [](User *U) { return isa<GetElementPtrInst>(U); })) {
              // Use getExtractWithExtendCost() to calculate the cost of
              // extractelement/ext pair.
              CommonCost -= TTI->getExtractWithExtendCost(
                  Ext->getOpcode(), Ext->getType(), VecTy, I);
              // Add back the cost of s|zext which is subtracted separately.
              CommonCost += TTI->getCastInstrCost(
                  Ext->getOpcode(), Ext->getType(), EI->getType(),
                  TTI::getCastContextHint(Ext), CostKind, Ext);
              continue;
            }
          }
          CommonCost -=
              TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, I);
        }
      } else {
        AdjustExtractsCost(CommonCost);
      }
      return CommonCost;
    }
    case Instruction::InsertElement: {
      assert(E->ReuseShuffleIndices.empty() &&
             "Unique insertelements only are expected.");
      auto *SrcVecTy = cast<FixedVectorType>(VL0->getType());
 
      unsigned const NumElts = SrcVecTy->getNumElements();
      unsigned const NumScalars = VL.size();
      APInt DemandedElts = APInt::getZero(NumElts);
      // TODO: Add support for Instruction::InsertValue.
      SmallVector<int> Mask;
      if (!E->ReorderIndices.empty()) {
        inversePermutation(E->ReorderIndices, Mask);
        Mask.append(NumElts - NumScalars, UndefMaskElem);
      } else {
        Mask.assign(NumElts, UndefMaskElem);
        std::iota(Mask.begin(), std::next(Mask.begin(), NumScalars), 0);
      }
      unsigned Offset = *getInsertIndex(VL0, 0);
      bool IsIdentity = true;
      SmallVector<int> PrevMask(NumElts, UndefMaskElem);
      Mask.swap(PrevMask);
      for (unsigned I = 0; I < NumScalars; ++I) {
        Optional<int> InsertIdx = getInsertIndex(VL[PrevMask[I]], 0);
        if (!InsertIdx || *InsertIdx == UndefMaskElem)
          continue;
        DemandedElts.setBit(*InsertIdx);
        IsIdentity &= *InsertIdx - Offset == I;
        Mask[*InsertIdx - Offset] = I;
      }
      assert(Offset < NumElts && "Failed to find vector index offset");
 
      InstructionCost Cost = 0;
      Cost -= TTI->getScalarizationOverhead(SrcVecTy, DemandedElts,
                                            /*Insert*/ true, /*Extract*/ false);
 
      if (IsIdentity && NumElts != NumScalars && Offset % NumScalars != 0) {
        // FIXME: Replace with SK_InsertSubvector once it is properly supported.
        unsigned Sz = PowerOf2Ceil(Offset + NumScalars);
        Cost += TTI->getShuffleCost(
            TargetTransformInfo::SK_PermuteSingleSrc,
            FixedVectorType::get(SrcVecTy->getElementType(), Sz));
      } else if (!IsIdentity) {
        auto *FirstInsert =
            cast<Instruction>(*find_if(E->Scalars, [E](Value *V) {
              return !is_contained(E->Scalars,
                                   cast<Instruction>(V)->getOperand(0));
            }));
        if (isUndefVector(FirstInsert->getOperand(0))) {
          Cost += TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, SrcVecTy, Mask);
        } else {
          SmallVector<int> InsertMask(NumElts);
          std::iota(InsertMask.begin(), InsertMask.end(), 0);
          for (unsigned I = 0; I < NumElts; I++) {
            if (Mask[I] != UndefMaskElem)
              InsertMask[Offset + I] = NumElts + I;
          }
          Cost +=
              TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVecTy, InsertMask);
        }
      }
 
      return Cost;
    }
    case Instruction::ZExt:
    case Instruction::SExt:
    case Instruction::FPToUI:
    case Instruction::FPToSI:
    case Instruction::FPExt:
    case Instruction::PtrToInt:
    case Instruction::IntToPtr:
    case Instruction::SIToFP:
    case Instruction::UIToFP:
    case Instruction::Trunc:
    case Instruction::FPTrunc:
    case Instruction::BitCast: {
      Type *SrcTy = VL0->getOperand(0)->getType();
      InstructionCost ScalarEltCost =
          TTI->getCastInstrCost(E->getOpcode(), ScalarTy, SrcTy,
                                TTI::getCastContextHint(VL0), CostKind, VL0);
      if (NeedToShuffleReuses) {
        CommonCost -= (EntryVF - VL.size()) * ScalarEltCost;
      }
 
      // Calculate the cost of this instruction.
      InstructionCost ScalarCost = VL.size() * ScalarEltCost;
 
      auto *SrcVecTy = FixedVectorType::get(SrcTy, VL.size());
      InstructionCost VecCost = 0;
      // Check if the values are candidates to demote.
      if (!MinBWs.count(VL0) || VecTy != SrcVecTy) {
        VecCost = CommonCost + TTI->getCastInstrCost(
                                   E->getOpcode(), VecTy, SrcVecTy,
                                   TTI::getCastContextHint(VL0), CostKind, VL0);
      }
      LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost));
      return VecCost - ScalarCost;
    }
    case Instruction::FCmp:
    case Instruction::ICmp:
    case Instruction::Select: {
      // Calculate the cost of this instruction.
      InstructionCost ScalarEltCost =
          TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, Builder.getInt1Ty(),
                                  CmpInst::BAD_ICMP_PREDICATE, CostKind, VL0);
      if (NeedToShuffleReuses) {
        CommonCost -= (EntryVF - VL.size()) * ScalarEltCost;
      }
      auto *MaskTy = FixedVectorType::get(Builder.getInt1Ty(), VL.size());
      InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost;
 
      // Check if all entries in VL are either compares or selects with compares
      // as condition that have the same predicates.
      CmpInst::Predicate VecPred = CmpInst::BAD_ICMP_PREDICATE;
      bool First = true;
      for (auto *V : VL) {
        CmpInst::Predicate CurrentPred;
        auto MatchCmp = m_Cmp(CurrentPred, m_Value(), m_Value());
        if ((!match(V, m_Select(MatchCmp, m_Value(), m_Value())) &&
             !match(V, MatchCmp)) ||
            (!First && VecPred != CurrentPred)) {
          VecPred = CmpInst::BAD_ICMP_PREDICATE;
          break;
        }
        First = false;
        VecPred = CurrentPred;
      }
 
      InstructionCost VecCost = TTI->getCmpSelInstrCost(
          E->getOpcode(), VecTy, MaskTy, VecPred, CostKind, VL0);
      // Check if it is possible and profitable to use min/max for selects in
      // VL.
      //
      auto IntrinsicAndUse = canConvertToMinOrMaxIntrinsic(VL);
      if (IntrinsicAndUse.first != Intrinsic::not_intrinsic) {
        IntrinsicCostAttributes CostAttrs(IntrinsicAndUse.first, VecTy,
                                          {VecTy, VecTy});
        InstructionCost IntrinsicCost =
            TTI->getIntrinsicInstrCost(CostAttrs, CostKind);
        // If the selects are the only uses of the compares, they will be dead
        // and we can adjust the cost by removing their cost.
        if (IntrinsicAndUse.second)
          IntrinsicCost -=
              TTI->getCmpSelInstrCost(Instruction::ICmp, VecTy, MaskTy,
                                      CmpInst::BAD_ICMP_PREDICATE, CostKind);
        VecCost = std::min(VecCost, IntrinsicCost);
      }
      LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost));
      return CommonCost + VecCost - ScalarCost;
    }
    case Instruction::FNeg:
    case Instruction::Add:
    case Instruction::FAdd:
    case Instruction::Sub:
    case Instruction::FSub:
    case Instruction::Mul:
    case Instruction::FMul:
    case Instruction::UDiv:
    case Instruction::SDiv:
    case Instruction::FDiv:
    case Instruction::URem:
    case Instruction::SRem:
    case Instruction::FRem:
    case Instruction::Shl:
    case Instruction::LShr:
    case Instruction::AShr:
    case Instruction::And:
    case Instruction::Or:
    case Instruction::Xor: {
      // Certain instructions can be cheaper to vectorize if they have a
      // constant second vector operand.
      TargetTransformInfo::OperandValueKind Op1VK =
          TargetTransformInfo::OK_AnyValue;
      TargetTransformInfo::OperandValueKind Op2VK =
          TargetTransformInfo::OK_UniformConstantValue;
      TargetTransformInfo::OperandValueProperties Op1VP =
          TargetTransformInfo::OP_None;
      TargetTransformInfo::OperandValueProperties Op2VP =
          TargetTransformInfo::OP_PowerOf2;
 
      // If all operands are exactly the same ConstantInt then set the
      // operand kind to OK_UniformConstantValue.
      // If instead not all operands are constants, then set the operand kind
      // to OK_AnyValue. If all operands are constants but not the same,
      // then set the operand kind to OK_NonUniformConstantValue.
      ConstantInt *CInt0 = nullptr;
      for (unsigned i = 0, e = VL.size(); i < e; ++i) {
        const Instruction *I = cast<Instruction>(VL[i]);
        unsigned OpIdx = isa<BinaryOperator>(I) ? 1 : 0;
        ConstantInt *CInt = dyn_cast<ConstantInt>(I->getOperand(OpIdx));
        if (!CInt) {
          Op2VK = TargetTransformInfo::OK_AnyValue;
          Op2VP = TargetTransformInfo::OP_None;
          break;
        }
        if (Op2VP == TargetTransformInfo::OP_PowerOf2 &&
            !CInt->getValue().isPowerOf2())
          Op2VP = TargetTransformInfo::OP_None;
        if (i == 0) {
          CInt0 = CInt;
          continue;
        }
        if (CInt0 != CInt)
          Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
      }
 
      SmallVector<const Value *, 4> Operands(VL0->operand_values());
      InstructionCost ScalarEltCost =
          TTI->getArithmeticInstrCost(E->getOpcode(), ScalarTy, CostKind, Op1VK,
                                      Op2VK, Op1VP, Op2VP, Operands, VL0);
      if (NeedToShuffleReuses) {
        CommonCost -= (EntryVF - VL.size()) * ScalarEltCost;
      }
      InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost;
      InstructionCost VecCost =
          TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind, Op1VK,
                                      Op2VK, Op1VP, Op2VP, Operands, VL0);
      LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost));
      return CommonCost + VecCost - ScalarCost;
    }
    case Instruction::GetElementPtr: {
      TargetTransformInfo::OperandValueKind Op1VK =
          TargetTransformInfo::OK_AnyValue;
      TargetTransformInfo::OperandValueKind Op2VK =
          TargetTransformInfo::OK_UniformConstantValue;
 
      InstructionCost ScalarEltCost = TTI->getArithmeticInstrCost(
          Instruction::Add, ScalarTy, CostKind, Op1VK, Op2VK);
      if (NeedToShuffleReuses) {
        CommonCost -= (EntryVF - VL.size()) * ScalarEltCost;
      }
      InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost;
      InstructionCost VecCost = TTI->getArithmeticInstrCost(
          Instruction::Add, VecTy, CostKind, Op1VK, Op2VK);
      LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost));
      return CommonCost + VecCost - ScalarCost;
    }
    case Instruction::Load: {
      // Cost of wide load - cost of scalar loads.
      Align Alignment = cast<LoadInst>(VL0)->getAlign();
      InstructionCost ScalarEltCost = TTI->getMemoryOpCost(
          Instruction::Load, ScalarTy, Alignment, 0, CostKind, VL0);
      if (NeedToShuffleReuses) {
        CommonCost -= (EntryVF - VL.size()) * ScalarEltCost;
      }
      InstructionCost ScalarLdCost = VecTy->getNumElements() * ScalarEltCost;
      InstructionCost VecLdCost;
      if (E->State == TreeEntry::Vectorize) {
        VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, Alignment, 0,
                                         CostKind, VL0);
      } else {
        assert(E->State == TreeEntry::ScatterVectorize && "Unknown EntryState");
        Align CommonAlignment = Alignment;
        for (Value *V : VL)
          CommonAlignment =
              commonAlignment(CommonAlignment, cast<LoadInst>(V)->getAlign());
        VecLdCost = TTI->getGatherScatterOpCost(
            Instruction::Load, VecTy, cast<LoadInst>(VL0)->getPointerOperand(),
            /*VariableMask=*/false, CommonAlignment, CostKind, VL0);
      }
      LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecLdCost, ScalarLdCost));
      return CommonCost + VecLdCost - ScalarLdCost;
    }
    case Instruction::Store: {
      // We know that we can merge the stores. Calculate the cost.
      bool IsReorder = !E->ReorderIndices.empty();
      auto *SI =
          cast<StoreInst>(IsReorder ? VL[E->ReorderIndices.front()] : VL0);
      Align Alignment = SI->getAlign();
      InstructionCost ScalarEltCost = TTI->getMemoryOpCost(
          Instruction::Store, ScalarTy, Alignment, 0, CostKind, VL0);
      InstructionCost ScalarStCost = VecTy->getNumElements() * ScalarEltCost;
      InstructionCost VecStCost = TTI->getMemoryOpCost(
          Instruction::Store, VecTy, Alignment, 0, CostKind, VL0);
      LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecStCost, ScalarStCost));
      return CommonCost + VecStCost - ScalarStCost;
    }
    case Instruction::Call: {
      CallInst *CI = cast<CallInst>(VL0);
      Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
 
      // Calculate the cost of the scalar and vector calls.
      IntrinsicCostAttributes CostAttrs(ID, *CI, 1);
      InstructionCost ScalarEltCost =
          TTI<