//===- VPlan.cpp - Vectorizer Plan ----------------------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
///
/// \file
/// This is the LLVM vectorization plan. It represents a candidate for
/// vectorization, allowing to plan and optimize how to vectorize a given loop
/// before generating LLVM-IR.
/// The vectorizer uses vectorization plans to estimate the costs of potential
/// candidates and if profitable to execute the desired plan, generating vector
/// LLVM-IR code.
///
//===----------------------------------------------------------------------===//

#include "VPlan.h"
#include "LoopVectorizationPlanner.h"
#include "VPlanCFG.h"
#include "VPlanDominatorTree.h"
#include "VPlanHelpers.h"
#include "VPlanPatternMatch.h"
#include "VPlanTransforms.h"
#include "VPlanUtils.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/ADT/Twine.h"
#include "llvm/Analysis/DomTreeUpdater.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/CFG.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Value.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/GraphWriter.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/LoopVersioning.h"
#include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h"
#include <cassert>
#include <string>

using namespace llvm;
using namespace llvm::VPlanPatternMatch;

namespace llvm {
extern cl::opt<bool> ProfcheckDisableMetadataFixes;
} // namespace llvm

/// @{
/// Metadata attribute names
const char LLVMLoopVectorizeFollowupAll[] = "llvm.loop.vectorize.followup_all";
const char LLVMLoopVectorizeFollowupVectorized[] =
    "llvm.loop.vectorize.followup_vectorized";
const char LLVMLoopVectorizeFollowupEpilogue[] =
    "llvm.loop.vectorize.followup_epilogue";
/// @}

extern cl::opt<unsigned> ForceTargetInstructionCost;

extern cl::opt<unsigned> NumberOfStoresToPredicate;

static cl::opt<bool> PrintVPlansInDotFormat(
    "vplan-print-in-dot-format", cl::Hidden,
    cl::desc("Use dot format instead of plain text when dumping VPlans"));

#define DEBUG_TYPE "loop-vectorize"

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
raw_ostream &llvm::operator<<(raw_ostream &OS, const VPRecipeBase &R) {
  const VPBasicBlock *Parent = R.getParent();
  VPSlotTracker SlotTracker(Parent ? Parent->getPlan() : nullptr);
  R.print(OS, "", SlotTracker);
  return OS;
}
#endif

Value *VPLane::getAsRuntimeExpr(IRBuilderBase &Builder,
                                const ElementCount &VF) const {
  switch (LaneKind) {
  case VPLane::Kind::ScalableLast:
    // Lane = RuntimeVF - VF.getKnownMinValue() + Lane
    return Builder.CreateSub(getRuntimeVF(Builder, Builder.getInt32Ty(), VF),
                             Builder.getInt32(VF.getKnownMinValue() - Lane));
  case VPLane::Kind::First:
    return Builder.getInt64(Lane);
  }
  llvm_unreachable("Unknown lane kind");
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void VPValue::print(raw_ostream &OS, VPSlotTracker &SlotTracker) const {
  if (const VPRecipeBase *R = getDefiningRecipe())
    R->print(OS, "", SlotTracker);
  else
    printAsOperand(OS, SlotTracker);
}

void VPValue::dump() const {
  const VPRecipeBase *Instr = getDefiningRecipe();
  VPSlotTracker SlotTracker(
      (Instr && Instr->getParent()) ? Instr->getParent()->getPlan() : nullptr);
  print(dbgs(), SlotTracker);
  dbgs() << "\n";
}

void VPRecipeBase::dump() const {
  VPSlotTracker SlotTracker(getParent() ? getParent()->getPlan() : nullptr);
  print(dbgs(), "", SlotTracker);
  dbgs() << "\n";
}
#endif

#if !defined(NDEBUG)
bool VPRecipeValue::isDefinedBy(const VPDef *D) const {
  return getDefiningRecipe() == D;
}
#endif

VPRecipeBase *VPValue::getDefiningRecipe() {
  auto *RecipeValue = dyn_cast<VPRecipeValue>(this);
  if (!RecipeValue)
    return nullptr;
  if (auto *MultiDef = dyn_cast<VPMultiDefValue>(RecipeValue))
    return MultiDef->getDef();
  return static_cast<VPSingleDefRecipe *>(RecipeValue);
}

const VPRecipeBase *VPValue::getDefiningRecipe() const {
  return const_cast<VPValue *>(this)->getDefiningRecipe();
}

Value *VPValue::getLiveInIRValue() const {
  return cast<VPIRValue>(this)->getValue();
}

Type *VPIRValue::getType() const { return getUnderlyingValue()->getType(); }

Type *VPValue::getScalarType() const {
  switch (getVPValueID()) {
  case VPVIRValueSC:
    return cast<VPIRValue>(this)->getType();
  case VPRegionValueSC:
    return cast<VPRegionValue>(this)->getType();
  case VPVSymbolicSC:
    return cast<VPSymbolicValue>(this)->getType();
  case VPVMultiDefValueSC:
  case VPVSingleDefValueSC:
    return cast<VPRecipeValue>(this)->getScalarType();
  }
  llvm_unreachable("Unhandled VPValue subclass");
}

VPRecipeValue::~VPRecipeValue() {
  assert(Users.empty() &&
         "trying to delete a VPRecipeValue with remaining users");
}

VPSingleDefValue::VPSingleDefValue(VPSingleDefRecipe *Def, Value *UV, Type *Ty)
    : VPRecipeValue(VPVSingleDefValueSC, UV, Ty) {
  assert(Def && "VPSingleDefValue requires a defining recipe");
  Def->addDefinedValue(this);
}

VPSingleDefValue::~VPSingleDefValue() {
  getDefiningRecipe()->removeDefinedValue(this);
}

VPMultiDefValue::VPMultiDefValue(VPRecipeBase *Def, Value *UV, Type *Ty)
    : VPRecipeValue(VPVMultiDefValueSC, UV, Ty), Def(Def) {
  assert(Def && "VPMultiDefValue requires a defining recipe");
  Def->addDefinedValue(this);
}

VPMultiDefValue::~VPMultiDefValue() {
  getDefiningRecipe()->removeDefinedValue(this);
}

// Get the top-most entry block of \p Start. This is the entry block of the
// containing VPlan. This function is templated to support both const and non-const blocks
template <typename T> static T *getPlanEntry(T *Start) {
  T *Next = Start;
  T *Current = Start;
  while ((Next = Next->getParent()))
    Current = Next;

  SmallSetVector<T *, 8> WorkList;
  WorkList.insert(Current);

  for (unsigned i = 0; i < WorkList.size(); i++) {
    T *Current = WorkList[i];
    if (!Current->hasPredecessors())
      return Current;
    auto &Predecessors = Current->getPredecessors();
    WorkList.insert_range(Predecessors);
  }

  llvm_unreachable("VPlan without any entry node without predecessors");
}

VPlan *VPBlockBase::getPlan() { return getPlanEntry(this)->Plan; }

const VPlan *VPBlockBase::getPlan() const { return getPlanEntry(this)->Plan; }

/// \return the VPBasicBlock that is the entry of Block, possibly indirectly.
const VPBasicBlock *VPBlockBase::getEntryBasicBlock() const {
  const VPBlockBase *Block = this;
  while (const VPRegionBlock *Region = dyn_cast<VPRegionBlock>(Block))
    Block = Region->getEntry();
  return cast<VPBasicBlock>(Block);
}

VPBasicBlock *VPBlockBase::getEntryBasicBlock() {
  VPBlockBase *Block = this;
  while (VPRegionBlock *Region = dyn_cast<VPRegionBlock>(Block))
    Block = Region->getEntry();
  return cast<VPBasicBlock>(Block);
}

void VPBlockBase::setPlan(VPlan *ParentPlan) {
  assert(ParentPlan->getEntry() == this && "Can only set plan on its entry.");
  Plan = ParentPlan;
}

/// \return the VPBasicBlock that is the exit of Block, possibly indirectly.
const VPBasicBlock *VPBlockBase::getExitingBasicBlock() const {
  const VPBlockBase *Block = this;
  while (const VPRegionBlock *Region = dyn_cast<VPRegionBlock>(Block))
    Block = Region->getExiting();
  return cast<VPBasicBlock>(Block);
}

VPBasicBlock *VPBlockBase::getExitingBasicBlock() {
  VPBlockBase *Block = this;
  while (VPRegionBlock *Region = dyn_cast<VPRegionBlock>(Block))
    Block = Region->getExiting();
  return cast<VPBasicBlock>(Block);
}

VPBlockBase *VPBlockBase::getEnclosingBlockWithSuccessors() {
  if (!Successors.empty() || !Parent)
    return this;
  assert(Parent->getExiting() == this &&
         "Block w/o successors not the exiting block of its parent.");
  return Parent->getEnclosingBlockWithSuccessors();
}

VPBlockBase *VPBlockBase::getEnclosingBlockWithPredecessors() {
  if (!Predecessors.empty() || !Parent)
    return this;
  assert(Parent->getEntry() == this &&
         "Block w/o predecessors not the entry of its parent.");
  return Parent->getEnclosingBlockWithPredecessors();
}

VPBasicBlock::iterator VPBasicBlock::getFirstNonPhi() {
  iterator It = begin();
  while (It != end() && It->isPhi())
    It++;
  return It;
}

VPTransformState::VPTransformState(const TargetTransformInfo *TTI,
                                   ElementCount VF, LoopInfo *LI,
                                   DominatorTree *DT, AssumptionCache *AC,
                                   IRBuilderBase &Builder, VPlan *Plan,
                                   Loop *CurrentParentLoop)
    : TTI(TTI), VF(VF), CFG(DT), LI(LI), AC(AC), Builder(Builder), Plan(Plan),
      CurrentParentLoop(CurrentParentLoop), VPDT(*Plan) {}

Value *VPTransformState::get(const VPValue *Def, const VPLane &Lane) {
  assert(!isa<VPRegionValue>(Def) &&
         "VPRegionValue must be materialized before VPTransformState::get");
  if (isa<VPIRValue, VPSymbolicValue>(Def))
    return Def->getUnderlyingValue();

  if (hasScalarValue(Def, Lane))
    return Data.VPV2Scalars[Def][Lane.mapToCacheIndex(VF)];

  if (!Lane.isFirstLane() && vputils::isSingleScalar(Def) &&
      hasScalarValue(Def, VPLane::getFirstLane())) {
    return Data.VPV2Scalars[Def][0];
  }

  // Look through BuildVector to avoid redundant extracts.
  // TODO: Remove once replicate regions are unrolled explicitly.
  if (Lane.getKind() == VPLane::Kind::First && match(Def, m_BuildVector())) {
    auto *BuildVector = cast<VPInstruction>(Def);
    return get(BuildVector->getOperand(Lane.getKnownLane()), true);
  }

  assert(hasVectorValue(Def));
  auto *VecPart = Data.VPV2Vector[Def];
  if (!VecPart->getType()->isVectorTy()) {
    assert(Lane.isFirstLane() && "cannot get lane > 0 for scalar");
    return VecPart;
  }
  // TODO: Cache created scalar values.
  Value *LaneV = Lane.getAsRuntimeExpr(Builder, VF);
  auto *Extract = Builder.CreateExtractElement(VecPart, LaneV);
  // set(Def, Extract, Instance);
  return Extract;
}

Value *VPTransformState::get(const VPValue *Def, bool NeedsScalar) {
  assert(!isa<VPRegionValue>(Def) &&
         "VPRegionValue must be materialized before VPTransformState::get");
  if (NeedsScalar) {
    assert((VF.isScalar() || isa<VPIRValue, VPSymbolicValue>(Def) ||
            hasVectorValue(Def) || !vputils::onlyFirstLaneUsed(Def) ||
            (hasScalarValue(Def, VPLane(0)) &&
             Data.VPV2Scalars[Def].size() == 1)) &&
           "Trying to access a single scalar per part but has multiple scalars "
           "per part.");
    return get(Def, VPLane(0));
  }

  // If Values have been set for this Def return the one relevant for \p Part.
  if (hasVectorValue(Def))
    return Data.VPV2Vector[Def];

  auto GetBroadcastInstrs = [this](Value *V) {
    if (VF.isScalar())
      return V;
    // Broadcast the scalar into all locations in the vector.
    Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
    return Shuf;
  };

  Value *ScalarValue = get(Def, VPLane(0));
  VPLane LastLane = VPLane::getLastLaneForVF(VF);
  IRBuilderBase::InsertPointGuard Guard(Builder);
  if (auto *LastInst = dyn_cast<Instruction>(get(Def, LastLane)))
    // Set the insert point after the last scalarized instruction or after the
    // last PHI, if LastInst is a PHI. This ensures the insertelement sequence
    // will directly follow the scalar definitions.
    Builder.SetInsertPoint(isa<PHINode>(LastInst)
                               ? LastInst->getParent()->getFirstNonPHIIt()
                               : std::next(BasicBlock::iterator(LastInst)));
  Value *VectorValue = GetBroadcastInstrs(ScalarValue);
  set(Def, VectorValue);
  return VectorValue;
}

void VPTransformState::setDebugLocFrom(DebugLoc DL) {
  const DILocation *DIL = DL;
  // When a FSDiscriminator is enabled, we don't need to add the multiply
  // factors to the discriminators.
  if (DIL &&
      Builder.GetInsertBlock()
          ->getParent()
          ->shouldEmitDebugInfoForProfiling() &&
      !EnableFSDiscriminator) {
    // FIXME: For scalable vectors, assume vscale=1.
    unsigned UF = Plan->getConcreteUF();
    auto NewDIL =
        DIL->cloneByMultiplyingDuplicationFactor(UF * VF.getKnownMinValue());
    if (NewDIL)
      Builder.SetCurrentDebugLocation(*NewDIL);
    else
      LLVM_DEBUG(dbgs() << "Failed to create new discriminator: "
                        << DIL->getFilename() << " Line: " << DIL->getLine());
  } else
    Builder.SetCurrentDebugLocation(DL);
}

Value *VPTransformState::packScalarIntoVectorizedValue(const VPValue *Def,
                                                       Value *WideValue,
                                                       const VPLane &Lane) {
  Value *ScalarInst = get(Def, Lane);
  Value *LaneExpr = Lane.getAsRuntimeExpr(Builder, VF);
  if (auto *StructTy = dyn_cast<StructType>(WideValue->getType())) {
    // We must handle each element of a vectorized struct type.
    for (unsigned I = 0, E = StructTy->getNumElements(); I != E; I++) {
      Value *ScalarValue = Builder.CreateExtractValue(ScalarInst, I);
      Value *VectorValue = Builder.CreateExtractValue(WideValue, I);
      VectorValue =
          Builder.CreateInsertElement(VectorValue, ScalarValue, LaneExpr);
      WideValue = Builder.CreateInsertValue(WideValue, VectorValue, I);
    }
  } else {
    WideValue = Builder.CreateInsertElement(WideValue, ScalarInst, LaneExpr);
  }
  return WideValue;
}

BasicBlock *VPBasicBlock::createEmptyBasicBlock(VPTransformState &State) {
  auto &CFG = State.CFG;
  // BB stands for IR BasicBlocks. VPBB stands for VPlan VPBasicBlocks.
  // Pred stands for Predessor. Prev stands for Previous - last visited/created.
  BasicBlock *PrevBB = CFG.PrevBB;
  BasicBlock *NewBB = BasicBlock::Create(PrevBB->getContext(), getName(),
                                         PrevBB->getParent(), CFG.ExitBB);
  LLVM_DEBUG(dbgs() << "LV: created " << NewBB->getName() << '\n');

  return NewBB;
}

void VPBasicBlock::connectToPredecessors(VPTransformState &State) {
  auto &CFG = State.CFG;
  BasicBlock *NewBB = CFG.VPBB2IRBB[this];

  // Register NewBB in its loop. In innermost loops its the same for all
  // BB's.
  Loop *ParentLoop = State.CurrentParentLoop;
  // If this block has a sole successor that is an exit block or is an exit
  // block itself then it needs adding to the same parent loop as the exit
  // block.
  VPBlockBase *SuccOrExitVPB = getSingleSuccessor();
  SuccOrExitVPB = SuccOrExitVPB ? SuccOrExitVPB : this;
  if (State.Plan->isExitBlock(SuccOrExitVPB)) {
    ParentLoop = State.LI->getLoopFor(
        cast<VPIRBasicBlock>(SuccOrExitVPB)->getIRBasicBlock());
  }

  if (ParentLoop && !State.LI->getLoopFor(NewBB))
    ParentLoop->addBasicBlockToLoop(NewBB, *State.LI);

  SmallVector<VPBlockBase *> Preds;
  if (VPBlockUtils::isHeader(this, State.VPDT)) {
    // There's no block for the latch yet, connect to the preheader only.
    Preds = {getPredecessors()[0]};
  } else {
    Preds = to_vector(getPredecessors());
  }

  // Hook up the new basic block to its predecessors.
  for (VPBlockBase *PredVPBlock : Preds) {
    VPBasicBlock *PredVPBB = PredVPBlock->getExitingBasicBlock();
    auto &PredVPSuccessors = PredVPBB->getHierarchicalSuccessors();
    assert(CFG.VPBB2IRBB.contains(PredVPBB) &&
           "Predecessor basic-block not found building successor.");
    BasicBlock *PredBB = CFG.VPBB2IRBB[PredVPBB];
    auto *PredBBTerminator = PredBB->getTerminator();
    LLVM_DEBUG(dbgs() << "LV: draw edge from " << PredBB->getName() << '\n');

    if (isa<UnreachableInst>(PredBBTerminator)) {
      assert(PredVPSuccessors.size() == 1 &&
             "Predecessor ending w/o branch must have single successor.");
      DebugLoc DL = PredBBTerminator->getDebugLoc();
      PredBBTerminator->eraseFromParent();
      auto *Br = UncondBrInst::Create(NewBB, PredBB);
      Br->setDebugLoc(DL);
    } else if (auto *UBI = dyn_cast<UncondBrInst>(PredBBTerminator)) {
      UBI->setSuccessor(NewBB);
    } else {
      // Set each forward successor here when it is created, excluding
      // backedges. A backward successor is set when the branch is created.
      // Branches to VPIRBasicBlocks must have the same successors in VPlan as
      // in the original IR, except when the predecessor is the entry block.
      // This enables including SCEV and memory runtime check blocks in VPlan.
      // TODO: Remove exception by modeling the terminator of entry block using
      // BranchOnCond.
      unsigned idx = PredVPSuccessors.front() == this ? 0 : 1;
      auto *TermBr = cast<CondBrInst>(PredBBTerminator);
      assert((!TermBr->getSuccessor(idx) ||
              (isa<VPIRBasicBlock>(this) &&
               (TermBr->getSuccessor(idx) == NewBB ||
                PredVPBlock == getPlan()->getEntry()))) &&
             "Trying to reset an existing successor block.");
      TermBr->setSuccessor(idx, NewBB);
    }
    CFG.DTU.applyUpdates({{DominatorTree::Insert, PredBB, NewBB}});
  }
}

void VPIRBasicBlock::execute(VPTransformState *State) {
  assert(getHierarchicalSuccessors().size() <= 2 &&
         "VPIRBasicBlock can have at most two successors at the moment!");
  // Move completely disconnected blocks to their final position.
  if (IRBB->hasNPredecessors(0) && succ_begin(IRBB) == succ_end(IRBB))
    IRBB->moveAfter(State->CFG.PrevBB);
  State->Builder.SetInsertPoint(IRBB->getTerminator());
  State->CFG.PrevBB = IRBB;
  State->CFG.VPBB2IRBB[this] = IRBB;
  executeRecipes(State, IRBB);
  // Create a branch instruction to terminate IRBB if one was not created yet
  // and is needed.
  if (getSingleSuccessor() && isa<UnreachableInst>(IRBB->getTerminator())) {
    auto *Br = State->Builder.CreateBr(IRBB);
    Br->setOperand(0, nullptr);
    IRBB->getTerminator()->eraseFromParent();
  } else {
    assert((getNumSuccessors() == 0 ||
            isa<UncondBrInst, CondBrInst>(IRBB->getTerminator())) &&
           "other blocks must be terminated by a branch");
  }

  connectToPredecessors(*State);
}

VPIRBasicBlock *VPIRBasicBlock::clone() {
  auto *NewBlock = getPlan()->createEmptyVPIRBasicBlock(IRBB);
  for (VPRecipeBase &R : Recipes)
    NewBlock->appendRecipe(R.clone());
  return NewBlock;
}

void VPBasicBlock::execute(VPTransformState *State) {
  if (VPBlockUtils::isHeader(this, State->VPDT)) {
    // Create and register the new vector loop.
    Loop *PrevParentLoop = State->CurrentParentLoop;
    State->CurrentParentLoop = State->LI->AllocateLoop();

    // Insert the new loop into the loop nest and register the new basic blocks
    // before calling any utilities such as SCEV that require valid LoopInfo.
    if (PrevParentLoop)
      PrevParentLoop->addChildLoop(State->CurrentParentLoop);
    else
      State->LI->addTopLevelLoop(State->CurrentParentLoop);
  }

  // 1. Create an IR basic block.
  BasicBlock *NewBB = createEmptyBasicBlock(*State);

  State->Builder.SetInsertPoint(NewBB);
  // Temporarily terminate with unreachable until CFG is rewired.
  UnreachableInst *Terminator = State->Builder.CreateUnreachable();
  State->Builder.SetInsertPoint(Terminator);

  State->CFG.PrevBB = NewBB;
  State->CFG.VPBB2IRBB[this] = NewBB;
  connectToPredecessors(*State);

  // 2. Fill the IR basic block with IR instructions.
  executeRecipes(State, NewBB);

  // If this block is a latch, update CurrentParentLoop.
  if (VPBlockUtils::isLatch(this, State->VPDT))
    State->CurrentParentLoop = State->CurrentParentLoop->getParentLoop();
}

VPBasicBlock *VPBasicBlock::clone() {
  auto *NewBlock = getPlan()->createVPBasicBlock(getName());
  for (VPRecipeBase &R : *this)
    NewBlock->appendRecipe(R.clone());
  return NewBlock;
}

void VPBasicBlock::executeRecipes(VPTransformState *State, BasicBlock *BB) {
  LLVM_DEBUG(dbgs() << "LV: vectorizing VPBB: " << getName()
                    << " in BB: " << BB->getName() << '\n');

  State->CFG.PrevVPBB = this;

  for (VPRecipeBase &Recipe : Recipes) {
    State->setDebugLocFrom(Recipe.getDebugLoc());
    Recipe.execute(*State);
  }

  LLVM_DEBUG(dbgs() << "LV: filled BB: " << *BB);
}

VPBasicBlock *VPBasicBlock::splitAt(iterator SplitAt) {
  assert((SplitAt == end() || SplitAt->getParent() == this) &&
         "can only split at a position in the same block");

  // Create new empty block after the block to split.
  auto *SplitBlock = getPlan()->createVPBasicBlock(getName() + ".split");
  VPBlockUtils::insertBlockAfter(SplitBlock, this);

  // If this is the exiting block, make the split the new exiting block.
  auto *ParentRegion = getParent();
  if (ParentRegion && ParentRegion->getExiting() == this)
    ParentRegion->setExiting(SplitBlock);

  // Finally, move the recipes starting at SplitAt to new block.
  for (VPRecipeBase &ToMove :
       make_early_inc_range(make_range(SplitAt, this->end())))
    ToMove.moveBefore(*SplitBlock, SplitBlock->end());

  return SplitBlock;
}

/// Return the enclosing loop region for region \p P. The templated version is
/// used to support both const and non-const block arguments.
template <typename T> static T *getEnclosingLoopRegionForRegion(T *P) {
  if (P && P->isReplicator()) {
    P = P->getParent();
    // Multiple loop regions can be nested, but replicate regions can only be
    // nested inside a loop region or must be outside any other region.
    assert((!P || !P->isReplicator()) && "unexpected nested replicate regions");
  }
  return P;
}

VPRegionBlock *VPBasicBlock::getEnclosingLoopRegion() {
  return getEnclosingLoopRegionForRegion(getParent());
}

const VPRegionBlock *VPBasicBlock::getEnclosingLoopRegion() const {
  return getEnclosingLoopRegionForRegion(getParent());
}

static bool hasConditionalTerminator(const VPBasicBlock *VPBB) {
  if (VPBB->empty()) {
    assert(
        VPBB->getNumSuccessors() < 2 &&
        "block with multiple successors doesn't have a recipe as terminator");
    return false;
  }

  const VPRecipeBase *R = &VPBB->back();
  [[maybe_unused]] bool IsSwitch =
      isa<VPInstruction>(R) &&
      cast<VPInstruction>(R)->getOpcode() == Instruction::Switch;
  [[maybe_unused]] bool IsBranchOnTwoConds = match(R, m_BranchOnTwoConds());
  [[maybe_unused]] bool IsCondBranch =
      isa<VPBranchOnMaskRecipe>(R) ||
      match(R, m_CombineOr(m_BranchOnCond(), m_BranchOnCount()));
  if (VPBB->getNumSuccessors() == 2 ||
      (VPBB->isExiting() && !VPBB->getParent()->isReplicator())) {
    assert((IsCondBranch || IsSwitch || IsBranchOnTwoConds) &&
           "block with multiple successors not terminated by "
           "conditional branch nor switch recipe");

    return true;
  }

  if (VPBB->getNumSuccessors() > 2) {
    assert((IsSwitch || IsBranchOnTwoConds) &&
           "block with more than 2 successors not terminated by a switch or "
           "branch-on-two-conds recipe");
    return true;
  }

  assert(
      !IsCondBranch && !IsBranchOnTwoConds &&
      "block with 0 or 1 successors terminated by conditional branch recipe");
  return false;
}

VPRecipeBase *VPBasicBlock::getTerminator() {
  if (hasConditionalTerminator(this))
    return &back();
  return nullptr;
}

const VPRecipeBase *VPBasicBlock::getTerminator() const {
  if (hasConditionalTerminator(this))
    return &back();
  return nullptr;
}

bool VPBasicBlock::isExiting() const {
  return getParent() && getParent()->getExitingBasicBlock() == this;
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void VPBlockBase::print(raw_ostream &O) const {
  VPSlotTracker SlotTracker(getPlan());
  print(O, "", SlotTracker);
}

void VPBlockBase::printSuccessors(raw_ostream &O, const Twine &Indent) const {
  if (!hasSuccessors()) {
    O << Indent << "No successors\n";
  } else {
    O << Indent << "Successor(s): ";
    ListSeparator LS;
    for (auto *Succ : getSuccessors())
      O << LS << Succ->getName();
    O << '\n';
  }
}

void VPBasicBlock::print(raw_ostream &O, const Twine &Indent,
                         VPSlotTracker &SlotTracker) const {
  O << Indent << getName() << ":\n";

  auto RecipeIndent = Indent + "  ";
  for (const VPRecipeBase &Recipe : *this) {
    Recipe.print(O, RecipeIndent, SlotTracker);
    O << '\n';
  }

  printSuccessors(O, Indent);
}
#endif

std::pair<VPBlockBase *, VPBlockBase *>
VPBlockUtils::cloneFrom(VPBlockBase *Entry) {
  DenseMap<VPBlockBase *, VPBlockBase *> Old2NewVPBlocks;
  VPBlockBase *Exiting = nullptr;
  bool InRegion = Entry->getParent();
  // First, clone blocks reachable from Entry.
  for (VPBlockBase *BB : vp_depth_first_shallow(Entry)) {
    VPBlockBase *NewBB = BB->clone();
    Old2NewVPBlocks[BB] = NewBB;
    if (InRegion && BB->getNumSuccessors() == 0) {
      assert(!Exiting && "Multiple exiting blocks?");
      Exiting = BB;
    }
  }
  assert((!InRegion || Exiting) && "regions must have a single exiting block");

  // Second, update the predecessors & successors of the cloned blocks.
  for (VPBlockBase *BB : vp_depth_first_shallow(Entry)) {
    VPBlockBase *NewBB = Old2NewVPBlocks[BB];
    SmallVector<VPBlockBase *> NewPreds;
    for (VPBlockBase *Pred : BB->getPredecessors()) {
      NewPreds.push_back(Old2NewVPBlocks[Pred]);
    }
    NewBB->setPredecessors(NewPreds);
    SmallVector<VPBlockBase *> NewSuccs;
    for (VPBlockBase *Succ : BB->successors()) {
      NewSuccs.push_back(Old2NewVPBlocks[Succ]);
    }
    NewBB->setSuccessors(NewSuccs);
  }

#if !defined(NDEBUG)
  // Verify that the order of predecessors and successors matches in the cloned
  // version.
  for (const auto &[OldBB, NewBB] :
       zip(vp_depth_first_shallow(Entry),
           vp_depth_first_shallow(Old2NewVPBlocks[Entry]))) {
    for (const auto &[OldPred, NewPred] :
         zip(OldBB->getPredecessors(), NewBB->getPredecessors()))
      assert(NewPred == Old2NewVPBlocks[OldPred] && "Different predecessors");

    for (const auto &[OldSucc, NewSucc] :
         zip(OldBB->successors(), NewBB->successors()))
      assert(NewSucc == Old2NewVPBlocks[OldSucc] && "Different successors");
  }
#endif

  return std::make_pair(Old2NewVPBlocks[Entry],
                        Exiting ? Old2NewVPBlocks[Exiting] : nullptr);
}

VPRegionBlock *VPRegionBlock::clone() {
  const auto &[NewEntry, NewExiting] = VPBlockUtils::cloneFrom(getEntry());
  VPlan &Plan = *getPlan();
  VPRegionValue *CanIV = getCanonicalIV();
  VPRegionBlock *NewRegion =
      CanIV ? Plan.createLoopRegion(CanIV->getType(), CanIV->getDebugLoc(),
                                    getName(), NewEntry, NewExiting)
            : Plan.createReplicateRegion(NewEntry, NewExiting, getName());

  if (getHeaderMask())
    NewRegion->createHeaderMask();

  for (VPBlockBase *Block : vp_depth_first_shallow(NewEntry))
    Block->setParent(NewRegion);
  return NewRegion;
}

void VPRegionBlock::execute(VPTransformState *State) {
  llvm_unreachable("regions must get dissolved before ::execute");
}

InstructionCost VPBasicBlock::cost(ElementCount VF, VPCostContext &Ctx) {
  InstructionCost Cost = 0;
  for (VPRecipeBase &R : Recipes)
    Cost += R.cost(VF, Ctx);
  return Cost;
}

const VPBasicBlock *VPBasicBlock::getCFGPredecessor(unsigned Idx) const {
  const VPBlockBase *Pred = nullptr;
  if (hasPredecessors()) {
    Pred = getPredecessors()[Idx];
  } else {
    auto *Region = getParent();
    assert(Region && !Region->isReplicator() && Region->getEntry() == this &&
           "must be in the entry block of a non-replicate region");
    assert(Idx < 2 && Region->getNumPredecessors() == 1 &&
           "loop region has a single predecessor (preheader), its entry block "
           "has 2 incoming blocks");

    // Idx ==  0 selects the predecessor of the region, Idx == 1 selects the
    // region itself whose exiting block feeds the phi across the backedge.
    Pred = Idx == 0 ? Region->getSinglePredecessor() : Region;
  }
  return Pred->getExitingBasicBlock();
}

InstructionCost VPRegionBlock::cost(ElementCount VF, VPCostContext &Ctx) {
  if (!isReplicator()) {
    // Neglect the cost of canonical IV, matching the legacy cost model.
    InstructionCost Cost = 0;
    for (VPBlockBase *Block : vp_depth_first_shallow(getEntry()))
      Cost += Block->cost(VF, Ctx);
    InstructionCost BackedgeCost =
        ForceTargetInstructionCost.getNumOccurrences()
            ? InstructionCost(ForceTargetInstructionCost)
            : Ctx.TTI.getCFInstrCost(Instruction::UncondBr, Ctx.CostKind);
    LLVM_DEBUG(dbgs() << "Cost of " << BackedgeCost << " for VF " << VF
                      << ": vector loop backedge\n");
    Cost += BackedgeCost;
    return Cost;
  }

  // Compute the cost of a replicate region. Replicating isn't supported for
  // scalable vectors, return an invalid cost for them.
  // TODO: Discard scalable VPlans with replicate recipes earlier after
  // construction.
  if (VF.isScalable())
    return InstructionCost::getInvalid();

  // Compute and return the cost of the conditionally executed recipes.
  assert(VF.isVector() && "Can only compute vector cost at the moment.");
  VPBasicBlock *Then = cast<VPBasicBlock>(getEntry()->getSuccessors()[0]);
  return Then->cost(VF, Ctx);
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void VPRegionBlock::print(raw_ostream &O, const Twine &Indent,
                          VPSlotTracker &SlotTracker) const {
  O << Indent << (isReplicator() ? "<xVFxUF> " : "<x1> ") << getName() << ": {";
  auto NewIndent = Indent + "  ";
  if (auto *CanIV = getCanonicalIV()) {
    O << '\n';
    CanIV->print(O, SlotTracker);
    O << " = CANONICAL-IV\n";
  }
  if (auto *HdrMask = getUsedHeaderMask()) {
    HdrMask->print(O, SlotTracker);
    O << " = HEADER-MASK\n";
  }
  for (auto *BlockBase : vp_depth_first_shallow(Entry)) {
    O << '\n';
    BlockBase->print(O, NewIndent, SlotTracker);
  }
  O << Indent << "}\n";

  printSuccessors(O, Indent);
}
#endif

void VPRegionBlock::dissolveToCFGLoop() {
  auto *Header = cast<VPBasicBlock>(getEntry());
  auto *ExitingLatch = cast<VPBasicBlock>(getExiting());
  auto *CanIV = getCanonicalIV();
  if (!CanIV->user_empty()) {
    VPlan &Plan = *getPlan();
    auto *Zero = Plan.getZero(CanIV->getType());
    DebugLoc DL = CanIV->getDebugLoc();
    VPInstruction *CanIVInc = getOrCreateCanonicalIVIncrement();
    VPBuilder HeaderBuilder(Header, Header->begin());
    auto *ScalarR =
        HeaderBuilder.createScalarPhi({Zero, CanIVInc}, DL, "index");
    CanIV->replaceAllUsesWith(ScalarR);
  }

  VPBlockBase *Preheader = getSinglePredecessor();
  VPBlockUtils::disconnectBlocks(Preheader, this);

  for (VPBlockBase *VPB : vp_depth_first_shallow(Entry))
    VPB->setParent(getParent());

  VPBlockUtils::connectBlocks(Preheader, Header);
  VPBlockUtils::transferSuccessors(this, ExitingLatch);
  VPBlockUtils::connectBlocks(ExitingLatch, Header);
}

VPInstruction *VPRegionBlock::getOrCreateCanonicalIVIncrement() {
  // TODO: Represent the increment as VPRegionValue as well.
  VPRegionValue *CanIV = getCanonicalIV();
  assert(CanIV && "Expected a canonical IV");

  if (auto *Inc = vputils::findCanonicalIVIncrement(*getPlan()))
    return Inc;

  assert(!getPlan()->getVFxUF().isMaterialized() &&
         "VFxUF can be used only before it is materialized.");
  auto *ExitingLatch = cast<VPBasicBlock>(getExiting());
  return VPBuilder(ExitingLatch->getTerminator())
      .createOverflowingOp(Instruction::Add, {CanIV, &getPlan()->getVFxUF()},
                           {hasCanonicalIVNUW(), /* HasNSW */ false},
                           CanIV->getDebugLoc(), "index.next");
}

VPlan::VPlan(Loop *L, Type *IdxTy)
    : VectorTripCount(IdxTy), VF(IdxTy), UF(IdxTy), VFxUF(IdxTy) {
  setEntry(createVPIRBasicBlock(L->getLoopPreheader()));
  ScalarHeader = createVPIRBasicBlock(L->getHeader());

  SmallVector<BasicBlock *> IRExitBlocks;
  L->getUniqueExitBlocks(IRExitBlocks);
  for (BasicBlock *EB : IRExitBlocks)
    ExitBlocks.push_back(createVPIRBasicBlock(EB));
}

VPlan::~VPlan() {
  VPSymbolicValue DummyValue(nullptr);

  // Redirect all recipe operands to DummyValue before deleting blocks.
  for (VPBasicBlock *VPBB :
       VPBlockUtils::blocksOnly<VPBasicBlock>(CreatedBlocks))
    for (VPRecipeBase &R : *VPBB)
      for (unsigned I = 0, E = R.getNumOperands(); I != E; I++)
        R.setOperand(I, &DummyValue);

  for (auto *VPB : CreatedBlocks)
    delete VPB;
  for (VPValue *VPV : getLiveIns())
    delete VPV;
  delete BackedgeTakenCount;
}

bool VPlan::isExitBlock(VPBlockBase *VPBB) {
  return is_contained(ExitBlocks, VPBB);
}

/// To make RUN_VPLAN_PASS print final VPlan.
static void printFinalVPlan(VPlan &) {}

/// Generate the code inside the preheader and body of the vectorized loop.
/// Assumes a single pre-header basic-block was created for this. Introduce
/// additional basic-blocks as needed, and fill them all.
void VPlan::execute(VPTransformState *State) {
  assert(none_of(vp_depth_first_shallow(getEntry()), IsaPred<VPRegionBlock>) &&
         "all region blocks must be dissolved before ::execute");

  // Initialize CFG state.
  State->CFG.PrevVPBB = nullptr;
  State->CFG.ExitBB = State->CFG.PrevBB->getSingleSuccessor();

  // Update VPDominatorTree since VPBasicBlock may be removed after State was
  // constructed.
  State->VPDT.recalculate(*this);

  // Disconnect VectorPreHeader from ExitBB in both the CFG and DT.
  BasicBlock *VectorPreHeader = State->CFG.PrevBB;
  cast<UncondBrInst>(VectorPreHeader->getTerminator())->setSuccessor(nullptr);
  State->CFG.DTU.applyUpdates(
      {{DominatorTree::Delete, VectorPreHeader, State->CFG.ExitBB}});

  LLVM_DEBUG(dbgs() << "Executing best plan with VF=" << State->VF
                    << ", UF=" << getConcreteUF() << '\n');
  setName("Final VPlan");
  // TODO: RUN_VPLAN_PASS/VPlanTransforms::runPass should automatically dump
  // VPlans after some specific stages when "-debug" is specified, but that
  // hasn't been implemented yet. For now, just do both:
  LLVM_DEBUG(dump());
  RUN_VPLAN_PASS(printFinalVPlan, *this);

  BasicBlock *ScalarPh = State->CFG.ExitBB;
  VPBasicBlock *ScalarPhVPBB = getScalarPreheader();
  if (ScalarPhVPBB) {
    // Disconnect scalar preheader and scalar header, as the dominator tree edge
    // will be updated as part of VPlan execution. This allows keeping the DTU
    // logic generic during VPlan execution.
    State->CFG.DTU.applyUpdates(
        {{DominatorTree::Delete, ScalarPh, ScalarPh->getSingleSuccessor()}});
  }
  ReversePostOrderTraversal<VPBlockShallowTraversalWrapper<VPBlockBase *>> RPOT(
      Entry);
  // Generate code for the VPlan, in parts of the vector skeleton, loop body and
  // successor blocks including the middle, exit and scalar preheader blocks.
  for (VPBlockBase *Block : RPOT)
    Block->execute(State);

  if (hasEarlyExit()) {
    // Fix up LoopInfo for extra dispatch blocks when vectorizing loops with
    // early exits. For dispatch blocks, we need to find the smallest common
    // loop of all successors that are in a loop. Note: we only need to update
    // loop info for blocks after the middle block, but there is no easy way to
    // get those at this point.
    for (VPBlockBase *VPB : reverse(RPOT)) {
      auto *VPBB = dyn_cast<VPBasicBlock>(VPB);
      if (!VPBB || isa<VPIRBasicBlock>(VPBB))
        continue;
      BasicBlock *BB = State->CFG.VPBB2IRBB[VPBB];
      Loop *L = State->LI->getLoopFor(BB);
      if (!L || any_of(successors(BB),
                       [L](BasicBlock *Succ) { return L->contains(Succ); }))
        continue;
      // Find the innermost loop containing all successors that are in a loop.
      // Successors not in any loop don't constrain the target loop.
      Loop *Target = nullptr;
      for (BasicBlock *Succ : successors(BB)) {
        Loop *SuccLoop = State->LI->getLoopFor(Succ);
        if (!SuccLoop)
          continue;
        if (!Target)
          Target = SuccLoop;
        else
          Target = State->LI->getSmallestCommonLoop(Target, SuccLoop);
      }
      State->LI->removeBlock(BB);
      if (Target)
        Target->addBasicBlockToLoop(BB, *State->LI);
    }
  }

  // If the original loop is unreachable, delete it and all its blocks.
  if (!ScalarPhVPBB) {
    // DeleteDeadBlocks will remove single-entry phis. Remove them from the exit
    // VPIRBBs in VPlan as well, otherwise we would retain references to deleted
    // IR instructions.
    for (VPIRBasicBlock *EB : getExitBlocks()) {
      for (VPRecipeBase &R : make_early_inc_range(EB->phis())) {
        if (R.getNumOperands() == 1)
          R.eraseFromParent();
      }
    }

    Loop *OrigLoop =
        State->LI->getLoopFor(getScalarHeader()->getIRBasicBlock());
    SmallVector<BasicBlock *> Blocks(OrigLoop->block_begin(),
                                     OrigLoop->block_end());
    Blocks.push_back(ScalarPh);
    while (!OrigLoop->isInnermost())
      State->LI->erase(*OrigLoop->begin());
    State->LI->erase(OrigLoop);
    for (auto *BB : Blocks)
      State->LI->removeBlock(BB);
    DeleteDeadBlocks(Blocks, &State->CFG.DTU);
  }

  State->CFG.DTU.flush();

  // Fix the latch (backedge) value of all header phis in all loop headers.
  for (VPBlockBase *VPB : vp_depth_first_shallow(getEntry())) {
    if (!VPBlockUtils::isHeader(VPB, State->VPDT))
      continue;
    auto *Header = cast<VPBasicBlock>(VPB);
    auto *LatchVPBB = cast<VPBasicBlock>(Header->getPredecessors()[1]);
    BasicBlock *VectorLatchBB = State->CFG.VPBB2IRBB[LatchVPBB];

    for (VPRecipeBase &R : Header->phis()) {
      auto *PhiR = cast<VPSingleDefRecipe>(&R);
      bool NeedsScalar =
          isa<VPPhi>(PhiR) || (isa<VPReductionPHIRecipe>(PhiR) &&
                               cast<VPReductionPHIRecipe>(PhiR)->isInLoop());

      Value *Phi = State->get(PhiR, NeedsScalar);
      Value *Val = State->get(PhiR->getOperand(1), NeedsScalar);
      cast<PHINode>(Phi)->addIncoming(Val, VectorLatchBB);
    }
  }
}

InstructionCost VPlan::cost(ElementCount VF, VPCostContext &Ctx) {
  // For now only return the cost of the vector loop region, ignoring any other
  // blocks, like the preheader or middle blocks, expect for checking them for
  // recipes with invalid costs.
  InstructionCost Cost = getVectorLoopRegion()->cost(VF, Ctx);

  // If the cost of the loop region is invalid or any recipe in the skeleton
  // outside loop regions are invalid return an invalid cost.
  if (!Cost.isValid() || any_of(VPBlockUtils::blocksOnly<VPBasicBlock>(
                                    vp_depth_first_shallow(getEntry())),
                                [&VF, &Ctx](VPBasicBlock *VPBB) {
                                  return !VPBB->cost(VF, Ctx).isValid();
                                }))
    return InstructionCost::getInvalid();

  return Cost;
}

VPRegionBlock *VPlan::getVectorLoopRegion() {
  // TODO: Cache if possible.
  for (VPBlockBase *B : vp_depth_first_shallow(getEntry()))
    if (auto *R = dyn_cast<VPRegionBlock>(B))
      return R->isReplicator() ? nullptr : R;
  return nullptr;
}

const VPRegionBlock *VPlan::getVectorLoopRegion() const {
  for (const VPBlockBase *B : vp_depth_first_shallow(getEntry()))
    if (auto *R = dyn_cast<VPRegionBlock>(B))
      return R->isReplicator() ? nullptr : R;
  return nullptr;
}

bool VPlan::isOuterLoop() const {
  const VPRegionBlock *LoopRegion = getVectorLoopRegion();
  assert(LoopRegion && "expected a vector loop region");
  return any_of(VPBlockUtils::blocksOnly<const VPRegionBlock>(
                    vp_depth_first_shallow(LoopRegion->getEntry())),
                [](const VPRegionBlock *R) { return !R->isReplicator(); });
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void VPlan::printLiveIns(raw_ostream &O) const {
  VPSlotTracker SlotTracker(this);

  if (!VF.user_empty()) {
    O << "\nLive-in ";
    VF.printAsOperand(O, SlotTracker);
    O << " = VF";
  }

  if (!UF.user_empty()) {
    O << "\nLive-in ";
    UF.printAsOperand(O, SlotTracker);
    O << " = UF";
  }

  if (!VFxUF.user_empty()) {
    O << "\nLive-in ";
    VFxUF.printAsOperand(O, SlotTracker);
    O << " = VF * UF";
  }

  if (!VectorTripCount.user_empty()) {
    O << "\nLive-in ";
    VectorTripCount.printAsOperand(O, SlotTracker);
    O << " = vector-trip-count";
  }

  if (BackedgeTakenCount && !BackedgeTakenCount->user_empty()) {
    O << "\nLive-in ";
    BackedgeTakenCount->printAsOperand(O, SlotTracker);
    O << " = backedge-taken count";
  }

  O << "\n";
  if (TripCount && !TripCount->user_empty()) {
    if (isa<VPIRValue>(TripCount))
      O << "Live-in ";
    TripCount->printAsOperand(O, SlotTracker);
    O << " = original trip-count";
    O << "\n";
  }
}

LLVM_DUMP_METHOD
void VPlan::print(raw_ostream &O) const {
  VPSlotTracker SlotTracker(this);

  O << "VPlan '" << getName() << "' {";

  printLiveIns(O);

  ReversePostOrderTraversal<VPBlockShallowTraversalWrapper<const VPBlockBase *>>
      RPOT(getEntry());
  for (const VPBlockBase *Block : RPOT) {
    O << '\n';
    Block->print(O, "", SlotTracker);
  }

  O << "}\n";
}

std::string VPlan::getName() const {
  std::string Out;
  raw_string_ostream RSO(Out);
  RSO << Name << " for ";
  if (!VFs.empty()) {
    RSO << "VF={" << VFs[0];
    for (ElementCount VF : drop_begin(VFs))
      RSO << "," << VF;
    RSO << "},";
  }

  if (UFs.empty()) {
    RSO << "UF>=1";
  } else {
    RSO << "UF={" << UFs[0];
    for (unsigned UF : drop_begin(UFs))
      RSO << "," << UF;
    RSO << "}";
  }

  return Out;
}

LLVM_DUMP_METHOD
void VPlan::printDOT(raw_ostream &O) const {
  VPlanPrinter Printer(O, *this);
  Printer.dump();
}

LLVM_DUMP_METHOD
void VPlan::dump() const { print(dbgs()); }
#endif

static void remapOperands(VPBlockBase *Entry, VPBlockBase *NewEntry,
                          DenseMap<VPValue *, VPValue *> &Old2NewVPValues) {
  // Update the operands of all cloned recipes starting at NewEntry. This
  // traverses all reachable blocks. This is done in two steps, to handle cycles
  // in PHI recipes.
  ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>>
      OldDeepRPOT(Entry);
  ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>>
      NewDeepRPOT(NewEntry);
  // First, collect all mappings from old to new VPValues defined by cloned
  // recipes.
  for (const auto &[OldBB, NewBB] :
       zip(VPBlockUtils::blocksOnly<VPBasicBlock>(OldDeepRPOT),
           VPBlockUtils::blocksOnly<VPBasicBlock>(NewDeepRPOT))) {
    assert(OldBB->getRecipeList().size() == NewBB->getRecipeList().size() &&
           "blocks must have the same number of recipes");
    for (const auto &[OldR, NewR] : zip(*OldBB, *NewBB)) {
      assert(OldR.getNumOperands() == NewR.getNumOperands() &&
             "recipes must have the same number of operands");
      assert(OldR.getNumDefinedValues() == NewR.getNumDefinedValues() &&
             "recipes must define the same number of operands");
      for (const auto &[OldV, NewV] :
           zip(OldR.definedValues(), NewR.definedValues()))
        Old2NewVPValues[OldV] = NewV;
    }
  }

  // Update all operands to use cloned VPValues.
  for (VPBasicBlock *NewBB :
       VPBlockUtils::blocksOnly<VPBasicBlock>(NewDeepRPOT)) {
    for (VPRecipeBase &NewR : *NewBB)
      for (unsigned I = 0, E = NewR.getNumOperands(); I != E; ++I) {
        VPValue *NewOp = Old2NewVPValues.lookup(NewR.getOperand(I));
        NewR.setOperand(I, NewOp);
      }
  }
}

VPlan *VPlan::duplicate() {
  unsigned NumBlocksBeforeCloning = CreatedBlocks.size();
  // Clone blocks.
  const auto &[NewEntry, __] = VPBlockUtils::cloneFrom(Entry);

  BasicBlock *ScalarHeaderIRBB = getScalarHeader()->getIRBasicBlock();
  VPIRBasicBlock *NewScalarHeader = nullptr;
  if (getScalarHeader()->hasPredecessors()) {
    NewScalarHeader = cast<VPIRBasicBlock>(*find_if(
        vp_depth_first_shallow(NewEntry), [ScalarHeaderIRBB](VPBlockBase *VPB) {
          auto *VPIRBB = dyn_cast<VPIRBasicBlock>(VPB);
          return VPIRBB && VPIRBB->getIRBasicBlock() == ScalarHeaderIRBB;
        }));
  } else {
    NewScalarHeader = createVPIRBasicBlock(ScalarHeaderIRBB);
  }
  // Create VPlan, clone live-ins and remap operands in the cloned blocks.
  auto *NewPlan =
      new VPlan(cast<VPBasicBlock>(NewEntry), NewScalarHeader, getIndexType());
  DenseMap<VPValue *, VPValue *> Old2NewVPValues;
  for (VPIRValue *OldLiveIn : getLiveIns())
    Old2NewVPValues[OldLiveIn] = NewPlan->getOrAddLiveIn(OldLiveIn);

  if (auto *TripCountIRV = dyn_cast_or_null<VPIRValue>(TripCount))
    Old2NewVPValues[TripCountIRV] = NewPlan->getOrAddLiveIn(TripCountIRV);
  // else NewTripCount will be created and inserted into Old2NewVPValues when
  // TripCount is cloned. In any case NewPlan->TripCount is updated below.

  assert(none_of(Old2NewVPValues.keys(), IsaPred<VPSymbolicValue>) &&
         "All VPSymbolicValues must be handled below");

  if (auto *LoopRegion = getVectorLoopRegion()) {
    auto *NewLoopRegion = NewPlan->getVectorLoopRegion();
    for (auto [Old, New] : zip_equal(LoopRegion->getRegionValues(),
                                     NewLoopRegion->getRegionValues())) {
      Old2NewVPValues[Old] = New;
      if (Old->isMaterialized())
        New->markMaterialized();
    }
  }

  if (BackedgeTakenCount)
    NewPlan->BackedgeTakenCount =
        new VPSymbolicValue(BackedgeTakenCount->getType());

  // Map and propagate materialized state for symbolic values.
  for (auto [OldSV, NewSV] :
       {std::pair{&VectorTripCount, &NewPlan->VectorTripCount},
        {&VF, &NewPlan->VF},
        {&UF, &NewPlan->UF},
        {&VFxUF, &NewPlan->VFxUF},
        {BackedgeTakenCount, NewPlan->BackedgeTakenCount}}) {
    if (!OldSV)
      continue;
    Old2NewVPValues[OldSV] = NewSV;
    if (OldSV->isMaterialized())
      NewSV->markMaterialized();
  }

  remapOperands(Entry, NewEntry, Old2NewVPValues);

  // Initialize remaining fields of cloned VPlan.
  NewPlan->VFs = VFs;
  NewPlan->UFs = UFs;
  // TODO: Adjust names.
  NewPlan->Name = Name;
  if (TripCount) {
    assert(Old2NewVPValues.contains(TripCount) &&
           "TripCount must have been added to Old2NewVPValues");
    NewPlan->TripCount = Old2NewVPValues[TripCount];
  }

  // Transfer all cloned blocks (the second half of all current blocks) from
  // current to new VPlan.
  unsigned NumBlocksAfterCloning = CreatedBlocks.size();
  for (unsigned I :
       seq<unsigned>(NumBlocksBeforeCloning, NumBlocksAfterCloning))
    NewPlan->CreatedBlocks.push_back(this->CreatedBlocks[I]);
  CreatedBlocks.truncate(NumBlocksBeforeCloning);

  // Update ExitBlocks of the new plan.
  for (VPBlockBase *VPB : NewPlan->CreatedBlocks) {
    if (VPB->getNumSuccessors() == 0 && isa<VPIRBasicBlock>(VPB) &&
        VPB != NewScalarHeader)
      NewPlan->ExitBlocks.push_back(cast<VPIRBasicBlock>(VPB));
  }

  return NewPlan;
}

VPIRBasicBlock *VPlan::createEmptyVPIRBasicBlock(BasicBlock *IRBB) {
  auto *VPIRBB = new VPIRBasicBlock(IRBB);
  CreatedBlocks.push_back(VPIRBB);
  return VPIRBB;
}

VPIRBasicBlock *VPlan::createVPIRBasicBlock(BasicBlock *IRBB) {
  auto *VPIRBB = createEmptyVPIRBasicBlock(IRBB);
  for (Instruction &I :
       make_range(IRBB->begin(), IRBB->getTerminator()->getIterator()))
    VPIRBB->appendRecipe(VPIRInstruction::create(I));
  return VPIRBB;
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)

Twine VPlanPrinter::getUID(const VPBlockBase *Block) {
  return (isa<VPRegionBlock>(Block) ? "cluster_N" : "N") +
         Twine(getOrCreateBID(Block));
}

void VPlanPrinter::dump() {
  Depth = 1;
  bumpIndent(0);
  OS << "digraph VPlan {\n";
  OS << "graph [labelloc=t, fontsize=30; label=\"Vectorization Plan";
  if (!Plan.getName().empty())
    OS << "\\n" << DOT::EscapeString(Plan.getName());

  {
    // Print live-ins.
  std::string Str;
  raw_string_ostream SS(Str);
  Plan.printLiveIns(SS);
  SmallVector<StringRef, 0> Lines;
  StringRef(Str).rtrim('\n').split(Lines, "\n");
  for (auto Line : Lines)
    OS << DOT::EscapeString(Line.str()) << "\\n";
  }

  OS << "\"]\n";
  OS << "node [shape=rect, fontname=Courier, fontsize=30]\n";
  OS << "edge [fontname=Courier, fontsize=30]\n";
  OS << "compound=true\n";

  for (const VPBlockBase *Block : vp_depth_first_shallow(Plan.getEntry()))
    dumpBlock(Block);

  OS << "}\n";
}

void VPlanPrinter::dumpBlock(const VPBlockBase *Block) {
  if (const VPBasicBlock *BasicBlock = dyn_cast<VPBasicBlock>(Block))
    dumpBasicBlock(BasicBlock);
  else if (const VPRegionBlock *Region = dyn_cast<VPRegionBlock>(Block))
    dumpRegion(Region);
  else
    llvm_unreachable("Unsupported kind of VPBlock.");
}

void VPlanPrinter::drawEdge(const VPBlockBase *From, const VPBlockBase *To,
                            bool Hidden, const Twine &Label) {
  // Due to "dot" we print an edge between two regions as an edge between the
  // exiting basic block and the entry basic of the respective regions.
  const VPBlockBase *Tail = From->getExitingBasicBlock();
  const VPBlockBase *Head = To->getEntryBasicBlock();
  OS << Indent << getUID(Tail) << " -> " << getUID(Head);
  OS << " [ label=\"" << Label << '\"';
  if (Tail != From)
    OS << " ltail=" << getUID(From);
  if (Head != To)
    OS << " lhead=" << getUID(To);
  if (Hidden)
    OS << "; splines=none";
  OS << "]\n";
}

void VPlanPrinter::dumpEdges(const VPBlockBase *Block) {
  auto &Successors = Block->getSuccessors();
  if (Successors.size() == 1)
    drawEdge(Block, Successors.front(), false, "");
  else if (Successors.size() == 2) {
    drawEdge(Block, Successors.front(), false, "T");
    drawEdge(Block, Successors.back(), false, "F");
  } else {
    unsigned SuccessorNumber = 0;
    for (auto *Successor : Successors)
      drawEdge(Block, Successor, false, Twine(SuccessorNumber++));
  }
}

void VPlanPrinter::dumpBasicBlock(const VPBasicBlock *BasicBlock) {
  // Implement dot-formatted dump by performing plain-text dump into the
  // temporary storage followed by some post-processing.
  OS << Indent << getUID(BasicBlock) << " [label =\n";
  bumpIndent(1);
  std::string Str;
  raw_string_ostream SS(Str);
  // Use no indentation as we need to wrap the lines into quotes ourselves.
  BasicBlock->print(SS, "", SlotTracker);

  // We need to process each line of the output separately, so split
  // single-string plain-text dump.
  SmallVector<StringRef, 0> Lines;
  StringRef(Str).rtrim('\n').split(Lines, "\n");

  auto EmitLine = [&](StringRef Line, StringRef Suffix) {
    OS << Indent << '"' << DOT::EscapeString(Line.str()) << "\\l\"" << Suffix;
  };

  // Don't need the "+" after the last line.
  for (auto Line : make_range(Lines.begin(), Lines.end() - 1))
    EmitLine(Line, " +\n");
  EmitLine(Lines.back(), "\n");

  bumpIndent(-1);
  OS << Indent << "]\n";

  dumpEdges(BasicBlock);
}

void VPlanPrinter::dumpRegion(const VPRegionBlock *Region) {
  OS << Indent << "subgraph " << getUID(Region) << " {\n";
  bumpIndent(1);
  OS << Indent << "fontname=Courier\n"
     << Indent << "label=\""
     << DOT::EscapeString(Region->isReplicator() ? "<xVFxUF> " : "<x1> ")
     << DOT::EscapeString(Region->getName()) << "\"\n";

  if (auto *CanIV = Region->getCanonicalIV()) {
    OS << Indent << "\"";
    std::string Op;
    raw_string_ostream S(Op);
    CanIV->printAsOperand(S, SlotTracker);
    OS << DOT::EscapeString(Op);
    OS << " = CANONICAL-IV\"\n";
  }

  // Dump the blocks of the region.
  assert(Region->getEntry() && "Region contains no inner blocks.");
  for (const VPBlockBase *Block : vp_depth_first_shallow(Region->getEntry()))
    dumpBlock(Block);
  bumpIndent(-1);
  OS << Indent << "}\n";
  dumpEdges(Region);
}

#endif

/// Returns true if there is a vector loop region and \p VPV is defined in a
/// loop region.
static bool isDefinedInsideLoopRegions(const VPValue *VPV) {
  if (isa<VPRegionValue>(VPV))
    return true;
  const VPRecipeBase *DefR = VPV->getDefiningRecipe();
  return DefR && (!DefR->getParent()->getPlan()->getVectorLoopRegion() ||
                  DefR->getParent()->getEnclosingLoopRegion());
}

bool VPValue::isDefinedOutsideLoopRegions() const {
  return !isDefinedInsideLoopRegions(this);
}
void VPValue::replaceAllUsesWith(VPValue *New) {
  replaceUsesWithIf(New, [](VPUser &, unsigned) { return true; });
  if (auto *SV = dyn_cast<VPSymbolicValue>(this))
    SV->markMaterialized();
}

void VPValue::replaceUsesWithIf(
    VPValue *New,
    llvm::function_ref<bool(VPUser &U, unsigned Idx)> ShouldReplace) {
  assertNotMaterialized();
  // Note that this early exit is required for correctness; the implementation
  // below relies on the number of users for this VPValue to decrease, which
  // isn't the case if this == New.
  if (this == New)
    return;

  for (unsigned J = 0; J < getNumUsers();) {
    VPUser *User = Users[J];
    bool RemovedUser = false;
    for (unsigned I = 0, E = User->getNumOperands(); I < E; ++I) {
      if (User->getOperand(I) != this || !ShouldReplace(*User, I))
        continue;

      RemovedUser = true;
      User->setOperand(I, New);
    }
    // If a user got removed after updating the current user, the next user to
    // update will be moved to the current position, so we only need to
    // increment the index if the number of users did not change.
    if (!RemovedUser)
      J++;
  }
}

void VPUser::replaceUsesOfWith(VPValue *From, VPValue *To) {
  for (unsigned Idx = 0; Idx != getNumOperands(); ++Idx) {
    if (getOperand(Idx) == From)
      setOperand(Idx, To);
  }
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void VPValue::printAsOperand(raw_ostream &OS, VPSlotTracker &Tracker) const {
  OS << Tracker.getOrCreateName(this);
}

void VPUser::printOperands(raw_ostream &O, VPSlotTracker &SlotTracker) const {
  interleaveComma(operands(), O, [&O, &SlotTracker](VPValue *Op) {
    Op->printAsOperand(O, SlotTracker);
  });
}
#endif

void VPSlotTracker::assignName(const VPValue *V) {
  assert(!VPValue2Name.contains(V) && "VPValue already has a name!");
  auto *UV = V->getUnderlyingValue();
  auto *VPI = dyn_cast_or_null<VPInstruction>(V);
  if (!UV && !(VPI && !VPI->getName().empty())) {
    VPValue2Name[V] = (Twine("vp<%") + Twine(NextSlot) + ">").str();
    NextSlot++;
    return;
  }

  // Use the name of the underlying Value, wrapped in "ir<>", and versioned by
  // appending ".Number" to the name if there are multiple uses.
  std::string Name;
  if (UV)
    Name = getName(UV);
  else
    Name = VPI->getName();

  assert(!Name.empty() && "Name cannot be empty.");
  StringRef Prefix = UV ? "ir<" : "vp<%";
  std::string BaseName = (Twine(Prefix) + Name + Twine(">")).str();

  // First assign the base name for V.
  const auto &[A, _] = VPValue2Name.try_emplace(V, BaseName);
  // Integer or FP constants with different types will result in the same string
  // due to stripping types.
  if (isa<VPIRValue>(V) && isa<ConstantInt, ConstantFP>(UV))
    return;

  // If it is already used by C > 0 other VPValues, increase the version counter
  // C and use it for V.
  const auto &[C, UseInserted] = BaseName2Version.try_emplace(BaseName, 0);
  if (!UseInserted) {
    C->second++;
    A->second = (BaseName + Twine(".") + Twine(C->second)).str();
  }
}

void VPSlotTracker::assignNames(const VPlan &Plan) {
  if (!Plan.VF.user_empty())
    assignName(&Plan.VF);
  if (!Plan.UF.user_empty())
    assignName(&Plan.UF);
  if (!Plan.VFxUF.user_empty())
    assignName(&Plan.VFxUF);
  assignName(&Plan.VectorTripCount);
  if (Plan.BackedgeTakenCount)
    assignName(Plan.BackedgeTakenCount);
  for (VPValue *LI : Plan.getLiveIns())
    assignName(LI);

  ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<const VPBlockBase *>>
      RPOT(VPBlockDeepTraversalWrapper<const VPBlockBase *>(Plan.getEntry()));
  for (const VPBlockBase *VPB : RPOT) {
    if (auto *VPBB = dyn_cast<VPBasicBlock>(VPB))
      assignNames(VPBB);
    else
      for (auto *RV : cast<VPRegionBlock>(VPB)->getRegionValues())
        assignName(RV);
  }
}

void VPSlotTracker::assignNames(const VPBasicBlock *VPBB) {
  for (const VPRecipeBase &Recipe : *VPBB)
    for (VPValue *Def : Recipe.definedValues())
      assignName(Def);
}

std::string VPSlotTracker::getName(const Value *V) {
  std::string Name;
  raw_string_ostream S(Name);
  if (V->hasName() || !isa<Instruction>(V)) {
    V->printAsOperand(S, false);
    return Name;
  }

  if (!MST) {
    // Lazily create the ModuleSlotTracker when we first hit an unnamed
    // instruction.
    auto *I = cast<Instruction>(V);
    // This check is required to support unit tests with incomplete IR.
    if (I->getParent()) {
      MST = std::make_unique<ModuleSlotTracker>(I->getModule());
      MST->incorporateFunction(*I->getFunction());
    } else {
      MST = std::make_unique<ModuleSlotTracker>(nullptr);
    }
  }
  V->printAsOperand(S, false, *MST);
  return Name;
}

std::string VPSlotTracker::getOrCreateName(const VPValue *V) const {
  std::string Name = VPValue2Name.lookup(V);
  if (!Name.empty())
    return Name;

  // If no name was assigned, no VPlan was provided when creating the slot
  // tracker or it is not reachable from the provided VPlan. This can happen,
  // e.g. when trying to print a recipe that has not been inserted into a VPlan
  // in a debugger.
  // TODO: Update VPSlotTracker constructor to assign names to recipes &
  // VPValues not associated with a VPlan, instead of constructing names ad-hoc
  // here.
  const VPRecipeBase *DefR = V->getDefiningRecipe();
  (void)DefR;
  assert((!DefR || !DefR->getParent() || !DefR->getParent()->getPlan()) &&
         "VPValue defined by a recipe in a VPlan?");

  // Use the underlying value's name, if there is one.
  if (auto *UV = V->getUnderlyingValue()) {
    std::string Name;
    raw_string_ostream S(Name);
    UV->printAsOperand(S, false);
    return (Twine("ir<") + Name + ">").str();
  }

  return "<badref>";
}

VPInstruction *VPBuilder::createAnyOfReduction(VPValue *ChainOp,
                                               VPValue *TrueVal,
                                               VPValue *FalseVal, DebugLoc DL) {
  assert(ChainOp->getScalarType()->isIntegerTy(1) &&
         "ChainOp must be i1 for AnyOf reduction");
  VPIRFlags Flags(RecurKind::Or, /*IsOrdered=*/false, /*IsInLoop=*/false,
                  FastMathFlags());
  auto *OrReduce =
      createNaryOp(VPInstruction::ComputeReductionResult, {ChainOp}, Flags, DL);
  auto *Freeze = createNaryOp(Instruction::Freeze, {OrReduce}, DL);
  return createSelect(Freeze, TrueVal, FalseVal, DL, "rdx.select");
}

bool LoopVectorizationPlanner::getDecisionAndClampRange(
    const std::function<bool(ElementCount)> &Predicate, VFRange &Range) {
  assert(!Range.isEmpty() && "Trying to test an empty VF range.");
  bool PredicateAtRangeStart = Predicate(Range.Start);

  for (ElementCount TmpVF : VFRange(Range.Start * 2, Range.End))
    if (Predicate(TmpVF) != PredicateAtRangeStart) {
      Range.End = TmpVF;
      break;
    }

  return PredicateAtRangeStart;
}

VPSingleDefRecipe *
VPBuilder::createConsecutiveVectorPointer(VPValue *Ptr, Type *SourceElementTy,
                                          bool Reverse, DebugLoc DL) {
  VPlan &Plan = getPlan();
  GEPNoWrapFlags Flags = vputils::getGEPFlagsForPtr(Ptr);
  if (Reverse) {
    // When folding the tail, we may compute an address that we don't in the
    // original scalar loop: drop the GEP no-wrap flags in this case. Otherwise
    // preserve existing flags without no-unsigned-wrap, as we will emit
    // negative indices.
    GEPNoWrapFlags ReverseFlags = Plan.hasTailFolded()
                                      ? GEPNoWrapFlags::none()
                                      : Flags.withoutNoUnsignedWrap();
    return tryInsertInstruction(new VPVectorEndPointerRecipe(
        Ptr, &Plan.getVF(), SourceElementTy, /*Stride=*/-1, ReverseFlags, DL));
  }
  Type *StrideTy = Plan.getDataLayout().getIndexType(Ptr->getScalarType());
  VPValue *StrideOne = Plan.getConstantInt(StrideTy, 1);
  return createVectorPointer(Ptr, SourceElementTy, StrideOne, Flags, DL);
}

VPlan &LoopVectorizationPlanner::getPlanFor(ElementCount VF) const {
  assert(count_if(VPlans,
                  [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) ==
             1 &&
         "Multiple VPlans for VF.");

  for (const VPlanPtr &Plan : VPlans) {
    if (Plan->hasVF(VF))
      return *Plan.get();
  }
  llvm_unreachable("No plan found!");
}

static void addRuntimeUnrollDisableMetaData(Loop *L) {
  SmallVector<Metadata *, 4> MDs;
  // Reserve first location for self reference to the LoopID metadata node.
  MDs.push_back(nullptr);
  bool IsUnrollMetadata = false;
  MDNode *LoopID = L->getLoopID();
  if (LoopID) {
    // First find existing loop unrolling disable metadata.
    for (unsigned I = 1, IE = LoopID->getNumOperands(); I < IE; ++I) {
      auto *MD = dyn_cast<MDNode>(LoopID->getOperand(I));
      if (MD) {
        const auto *S = dyn_cast<MDString>(MD->getOperand(0));
        if (!S)
          continue;
        if (S->getString().starts_with("llvm.loop.unroll.runtime.disable"))
          continue;
        IsUnrollMetadata =
            S->getString().starts_with("llvm.loop.unroll.disable");
      }
      MDs.push_back(LoopID->getOperand(I));
    }
  }

  if (!IsUnrollMetadata) {
    // Add runtime unroll disable metadata.
    LLVMContext &Context = L->getHeader()->getContext();
    SmallVector<Metadata *, 1> DisableOperands;
    DisableOperands.push_back(
        MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
    MDNode *DisableNode = MDNode::get(Context, DisableOperands);
    MDs.push_back(DisableNode);
    MDNode *NewLoopID = MDNode::get(Context, MDs);
    // Set operand 0 to refer to the loop id itself.
    NewLoopID->replaceOperandWith(0, NewLoopID);
    L->setLoopID(NewLoopID);
  }
}

void LoopVectorizationPlanner::updateLoopMetadataAndProfileInfo(
    Loop *VectorLoop, VPBasicBlock *HeaderVPBB, const VPlan &Plan,
    bool VectorizingEpilogue, MDNode *OrigLoopID,
    std::optional<unsigned> OrigAverageTripCount,
    unsigned OrigLoopInvocationWeight, unsigned EstimatedVFxUF,
    bool DisableRuntimeUnroll) {
  // Update the metadata of the scalar loop. Skip the update when vectorizing
  // the epilogue loop to ensure it is updated only once. Also skip the update
  // when the scalar loop became unreachable.
  auto *ScalarPH = Plan.getScalarPreheader();
  if (ScalarPH && !VectorizingEpilogue) {
    std::optional<MDNode *> RemainderLoopID =
        makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll,
                                        LLVMLoopVectorizeFollowupEpilogue});
    if (RemainderLoopID) {
      OrigLoop->setLoopID(*RemainderLoopID);
    } else {
      if (DisableRuntimeUnroll)
        addRuntimeUnrollDisableMetaData(OrigLoop);

      LoopVectorizeHints Hints(OrigLoop, /*InterleaveOnlyWhenForced*/ false,
                               *ORE);
      Hints.setAlreadyVectorized();
    }
  }
  // Tag the scalar remainder so downstream passes (e.g. the unroller and
  // WarnMissedTransforms) can produce more informative remarks.  Only emit
  // when remarks are enabled.
  if (ORE->enabled() && ScalarPH && ScalarPH->hasPredecessors())
    OrigLoop->addIntLoopAttribute("llvm.loop.vectorize.epilogue", 1);

  if (!VectorLoop)
    return;

  if (std::optional<MDNode *> VectorizedLoopID = makeFollowupLoopID(
          OrigLoopID, {LLVMLoopVectorizeFollowupAll,
                       LLVMLoopVectorizeFollowupVectorized})) {
    VectorLoop->setLoopID(*VectorizedLoopID);
  } else {
    // Keep all loop hints from the original loop on the vector loop (we'll
    // replace the vectorizer-specific hints below).
    if (OrigLoopID)
      VectorLoop->setLoopID(OrigLoopID);

    if (!VectorizingEpilogue) {
      LoopVectorizeHints Hints(VectorLoop, /*InterleaveOnlyWhenForced*/ false,
                               *ORE);
      Hints.setAlreadyVectorized();
    }
  }
  // Tag the vector loop body so downstream passes can identify it.  Only
  // emit when remarks are enabled.
  if (ORE->enabled())
    VectorLoop->addIntLoopAttribute("llvm.loop.vectorize.body", 1);
  TargetTransformInfo::UnrollingPreferences UP;
  TTI.getUnrollingPreferences(VectorLoop, *PSE.getSE(), UP, ORE);
  if (!UP.UnrollVectorizedLoop || VectorizingEpilogue)
    addRuntimeUnrollDisableMetaData(VectorLoop);

  // Set/update profile weights for the vector and remainder loops as original
  // loop iterations are now distributed among them. Note that original loop
  // becomes the scalar remainder loop after vectorization.
  //
  // For cases like foldTailByMasking() and requiresScalarEpiloque() we may
  // end up getting slightly roughened result but that should be OK since
  // profile is not inherently precise anyway. Note also possible bypass of
  // vector code caused by legality checks is ignored, assigning all the weight
  // to the vector loop, optimistically.
  //
  // For scalable vectorization we can't know at compile time how many
  // iterations of the loop are handled in one vector iteration, so instead
  // use the value of vscale used for tuning.
  unsigned AverageVectorTripCount = 0;
  unsigned RemainderAverageTripCount = 0;
  auto EC = VectorLoop->getLoopPreheader()->getParent()->getEntryCount();
  auto IsProfiled = EC && *EC != 0;
  if (!OrigAverageTripCount) {
    if (!IsProfiled)
      return;
    auto &SE = *PSE.getSE();
    AverageVectorTripCount = SE.getSmallConstantTripCount(VectorLoop);
    if (ProfcheckDisableMetadataFixes || !AverageVectorTripCount)
      return;
    if (ScalarPH)
      RemainderAverageTripCount =
          SE.getSmallConstantTripCount(OrigLoop) % EstimatedVFxUF;
    // Setting to 1 should be sufficient to generate the correct branch weights.
    OrigLoopInvocationWeight = 1;
  } else {
    // Calculate number of iterations in unrolled loop.
    AverageVectorTripCount = *OrigAverageTripCount / EstimatedVFxUF;
    // Calculate number of iterations for remainder loop.
    RemainderAverageTripCount = *OrigAverageTripCount % EstimatedVFxUF;
  }
  if (HeaderVPBB) {
    setLoopEstimatedTripCount(VectorLoop, AverageVectorTripCount,
                              OrigLoopInvocationWeight);
  }

  if (ScalarPH) {
    setLoopEstimatedTripCount(OrigLoop, RemainderAverageTripCount,
                              OrigLoopInvocationWeight);
  }
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void LoopVectorizationPlanner::printPlans(raw_ostream &O) {
  if (VPlans.empty()) {
    O << "LV: No VPlans built.\n";
    return;
  }
  for (const auto &Plan : VPlans)
    if (PrintVPlansInDotFormat)
      Plan->printDOT(O);
    else
      Plan->print(O);
}
#endif

bool llvm::canConstantBeExtended(const APInt *C, Type *NarrowType,
                                 TTI::PartialReductionExtendKind ExtKind) {
  APInt TruncatedVal = C->trunc(NarrowType->getScalarSizeInBits());
  unsigned WideSize = C->getBitWidth();
  APInt ExtendedVal = ExtKind == TTI::PR_SignExtend
                          ? TruncatedVal.sext(WideSize)
                          : TruncatedVal.zext(WideSize);
  return ExtendedVal == *C;
}

TargetTransformInfo::OperandValueInfo
VPCostContext::getOperandInfo(VPValue *V) const {
  if (auto *IRV = dyn_cast<VPIRValue>(V))
    return TTI::getOperandInfo(IRV->getValue());

  return {};
}

InstructionCost VPCostContext::getScalarizationOverhead(
    Type *ResultTy, ArrayRef<const VPValue *> Operands, ElementCount VF,
    TTI::VectorInstrContext VIC, bool AlwaysIncludeReplicatingR) {
  if (VF.isScalar())
    return 0;

  assert(!VF.isScalable() &&
         "Scalarization overhead not supported for scalable vectors");

  InstructionCost ScalarizationCost = 0;
  // Compute the cost of scalarizing the result if needed.
  if (!ResultTy->isVoidTy()) {
    for (Type *VectorTy :
         to_vector(getContainedTypes(toVectorizedTy(ResultTy, VF)))) {
      ScalarizationCost += TTI.getScalarizationOverhead(
          cast<VectorType>(VectorTy), APInt::getAllOnes(VF.getFixedValue()),
          /*Insert=*/true, /*Extract=*/false, CostKind,
          /*ForPoisonSrc=*/true, {}, VIC);
    }
  }
  // Compute the cost of scalarizing the operands, skipping ones that do not
  // require extraction/scalarization and do not incur any overhead.
  SmallPtrSet<const VPValue *, 4> UniqueOperands;
  SmallVector<Type *> Tys;
  for (auto *Op : Operands) {
    if (isa<VPIRValue>(Op) ||
        (!AlwaysIncludeReplicatingR &&
         isa<VPReplicateRecipe, VPPredInstPHIRecipe>(Op)) ||
        (isa<VPReplicateRecipe>(Op) &&
         cast<VPReplicateRecipe>(Op)->getOpcode() == Instruction::Load) ||
        !UniqueOperands.insert(Op).second)
      continue;
    Tys.push_back(toVectorizedTy(Op->getScalarType(), VF));
  }
  return ScalarizationCost +
         TTI.getOperandsScalarizationOverhead(Tys, CostKind, VIC);
}

bool VPCostContext::useEmulatedMaskMemRefHack(const VPReplicateRecipe *R,
                                              ElementCount VF) {
  const Instruction *UI = R->getUnderlyingInstr();
  if (isa<LoadInst>(UI))
    return true;
  assert(isa<StoreInst>(UI) && "R must either be a load or store");

  if (!NumPredStores) {
    // Count the number of predicated stores in the VPlan, caching the result.
    // Only stores where scatter is not legal are counted, matching the legacy
    // cost model behavior.
    const VPlan &Plan = *R->getParent()->getPlan();
    NumPredStores = 0;
    for (const VPRegionBlock *VPRB :
         VPBlockUtils::blocksOnly<const VPRegionBlock>(
             vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) {
      assert(VPRB->isReplicator() && "must only contain replicate regions");
      for (const VPBasicBlock *VPBB :
           VPBlockUtils::blocksOnly<const VPBasicBlock>(
               vp_depth_first_shallow(VPRB->getEntry()))) {
        for (const VPRecipeBase &Recipe : *VPBB) {
          auto *RepR = dyn_cast<VPReplicateRecipe>(&Recipe);
          if (!RepR)
            continue;
          if (!isa<StoreInst>(RepR->getUnderlyingInstr()))
            continue;
          // Check if scatter is legal for this store. If so, don't count it.
          Type *Ty = RepR->getOperand(0)->getScalarType();
          auto *VTy = VectorType::get(Ty, VF);
          const Align Alignment =
              getLoadStoreAlignment(RepR->getUnderlyingInstr());
          if (!TTI.isLegalMaskedScatter(VTy, Alignment))
            ++(*NumPredStores);
        }
      }
    }
  }
  return *NumPredStores > NumberOfStoresToPredicate;
}

bool VPCostContext::isFreeScalarIntrinsic(Intrinsic::ID ID) {
  return is_contained({Intrinsic::assume, Intrinsic::lifetime_end,
                       Intrinsic::lifetime_start, Intrinsic::sideeffect,
                       Intrinsic::pseudoprobe,
                       Intrinsic::experimental_noalias_scope_decl},
                      ID);
}