//===-- VPlanTransforms.cpp - Utility VPlan to VPlan transforms -----------===//
//
// 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 file implements a set of utility VPlan to VPlan transformations.
///
//===----------------------------------------------------------------------===//

#include "VPlanTransforms.h"
#include "VPRecipeBuilder.h"
#include "VPlan.h"
#include "VPlanAnalysis.h"
#include "VPlanCFG.h"
#include "VPlanDominatorTree.h"
#include "VPlanHelpers.h"
#include "VPlanPatternMatch.h"
#include "VPlanUtils.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/TypeSwitch.h"
#include "llvm/Analysis/IVDescriptors.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/MemoryLocation.h"
#include "llvm/Analysis/ScalarEvolutionPatternMatch.h"
#include "llvm/Analysis/ScopedNoAliasAA.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/Metadata.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/TypeSize.h"
#include "llvm/Transforms/Utils/LoopUtils.h"

using namespace llvm;
using namespace LoopVectorizationUtils;
using namespace VPlanPatternMatch;
using namespace SCEVPatternMatch;

/// Returns the metadata attached to \p R, or an empty set for a recipe that
/// does not carry any.
static VPIRMetadata getMetadataOf(VPRecipeBase *R) {
  if (auto *MD = dyn_cast<VPIRMetadata>(R))
    return *MD;
  return {};
}

// TODO: Remove this once the partial reduction intrinsics are no worse than
//       normal vector operations.
static cl::opt<bool> UsePartialReductionsByDefault(
    "use-partial-reductions-by-default", cl::init(false), cl::Hidden,
    cl::desc("Use partial reduction intrinsics for "
             "all supported unordered reductions."));

bool VPlanTransforms::tryToConvertVPInstructionsToVPRecipes(
    VPlan &Plan, const TargetLibraryInfo &TLI, PredicatedScalarEvolution &PSE,
    Loop *OuterLoop) {

  // Returns true if the access of \p AccessTy at \p Addr can be widened to a
  // consecutive vector access.
  auto IsConsecutiveAccess = [&](VPValue *Addr, Type *AccessTy) {
    return !hasIrregularType(AccessTy, Plan.getDataLayout()) &&
           vputils::getConstantStride(Addr, AccessTy, PSE, OuterLoop) == 1;
  };

  ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
      Plan.getVectorLoopRegion());
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
    // Skip blocks outside region
    if (!VPBB->getParent())
      break;
    VPRecipeBase *Term = VPBB->getTerminator();
    auto EndIter = Term ? Term->getIterator() : VPBB->end();
    // Introduce each ingredient into VPlan.
    for (VPRecipeBase &Ingredient :
         make_early_inc_range(make_range(VPBB->begin(), EndIter))) {

      VPValue *VPV = Ingredient.getVPSingleValue();
      if (!VPV->getUnderlyingValue())
        continue;

      Instruction *Inst = cast<Instruction>(VPV->getUnderlyingValue());

      VPRecipeBase *NewRecipe = nullptr;
      if (auto *PhiR = dyn_cast<VPPhi>(&Ingredient)) {
        auto *Phi = cast<PHINode>(PhiR->getUnderlyingValue());
        NewRecipe = new VPWidenPHIRecipe(PhiR->operands(), PhiR->getDebugLoc(),
                                         Phi->getName());
      } else if (auto *VPI = dyn_cast<VPInstruction>(&Ingredient)) {
        assert(!isa<PHINode>(Inst) && "phis should be handled above");
        // Create VPWidenMemoryRecipe for loads and stores.
        if (LoadInst *Load = dyn_cast<LoadInst>(Inst)) {
          bool IsConsecutive =
              IsConsecutiveAccess(VPI->getOperand(0), VPI->getScalarType());
          NewRecipe = new VPWidenLoadRecipe(*Load, Ingredient.getOperand(0),
                                            nullptr /*Mask*/, IsConsecutive,
                                            *VPI, Ingredient.getDebugLoc());
        } else if (StoreInst *Store = dyn_cast<StoreInst>(Inst)) {
          bool IsConsecutive = IsConsecutiveAccess(
              VPI->getOperand(1), VPI->getOperand(0)->getScalarType());
          NewRecipe = new VPWidenStoreRecipe(
              *Store, Ingredient.getOperand(1), Ingredient.getOperand(0),
              nullptr /*Mask*/, IsConsecutive, *VPI, Ingredient.getDebugLoc());
        } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
          NewRecipe = new VPWidenGEPRecipe(GEP->getSourceElementType(),
                                           Ingredient.operands(), *VPI,
                                           Ingredient.getDebugLoc(), GEP);
        } else if (CallInst *CI = dyn_cast<CallInst>(Inst)) {
          Intrinsic::ID VectorID = getVectorIntrinsicIDForCall(CI, &TLI);
          if (VectorID == Intrinsic::not_intrinsic)
            return false;

          // The noalias.scope.decl intrinsic declares a noalias scope that
          // is valid for a single iteration. Emitting it as a single-scalar
          // replicate would incorrectly extend the scope across multiple
          // original iterations packed into one vector iteration.
          // FIXME: If we want to vectorize this loop, then we have to drop
          // all the associated !alias.scope and !noalias.
          if (VectorID == Intrinsic::experimental_noalias_scope_decl)
            return false;

          // These intrinsics are recognized by getVectorIntrinsicIDForCall
          // but are not widenable. Emit them as replicate instead of widening.
          if (VectorID == Intrinsic::assume ||
              VectorID == Intrinsic::lifetime_end ||
              VectorID == Intrinsic::lifetime_start ||
              VectorID == Intrinsic::sideeffect ||
              VectorID == Intrinsic::pseudoprobe) {
            // If the operand of llvm.assume holds before vectorization, it will
            // also hold per lane.
            // llvm.pseudoprobe requires to be duplicated per lane for accurate
            // sample count.
            const bool IsSingleScalar = VectorID != Intrinsic::assume &&
                                        VectorID != Intrinsic::pseudoprobe;
            NewRecipe = new VPReplicateRecipe(CI, Ingredient.operands(),
                                              /*IsSingleScalar=*/IsSingleScalar,
                                              /*Mask=*/nullptr, *VPI, *VPI,
                                              Ingredient.getDebugLoc());
          } else {
            NewRecipe = new VPWidenIntrinsicRecipe(
                *CI, VectorID, drop_end(Ingredient.operands()), CI->getType(),
                VPIRFlags(*CI), *VPI, CI->getDebugLoc());
          }
        } else if (auto *CI = dyn_cast<CastInst>(Inst)) {
          NewRecipe = new VPWidenCastRecipe(
              CI->getOpcode(), Ingredient.getOperand(0), CI->getType(), CI,
              VPIRFlags(*CI), VPIRMetadata(*CI));
        } else {
          NewRecipe = new VPWidenRecipe(*Inst, Ingredient.operands(), *VPI,
                                        *VPI, Ingredient.getDebugLoc());
        }
      } else {
        assert(isa<VPWidenIntOrFpInductionRecipe>(&Ingredient) &&
               "inductions must be created earlier");
        continue;
      }

      NewRecipe->insertBefore(&Ingredient);
      if (NewRecipe->getNumDefinedValues() == 1)
        VPV->replaceAllUsesWith(NewRecipe->getVPSingleValue());
      else
        assert(NewRecipe->getNumDefinedValues() == 0 &&
               "Only recpies with zero or one defined values expected");
      Ingredient.eraseFromParent();
    }
  }
  return true;
}

/// Helper for extra no-alias checks via known-safe recipe and SCEV.
class SinkStoreInfo {
  SmallPtrSet<VPReplicateRecipe *, 4> ExcludeRecipes;
  VPReplicateRecipe &GroupLeader;
  PredicatedScalarEvolution *PSE = nullptr;
  const Loop *L = nullptr;

  // Return true if \p A and \p B are known to not alias for all VFs in the
  // plan, checked via the distance between the accesses
  bool isNoAliasViaDistance(VPReplicateRecipe *A, VPReplicateRecipe *B) const {
    if (A->getOpcode() != Instruction::Store ||
        B->getOpcode() != Instruction::Store)
      return false;

    if (!PSE || !L)
      return A == B;

    VPValue *AddrA = A->getOperand(1);
    const SCEV *SCEVA = vputils::getSCEVExprForVPValue(AddrA, *PSE, L);
    VPValue *AddrB = B->getOperand(1);
    const SCEV *SCEVB = vputils::getSCEVExprForVPValue(AddrB, *PSE, L);
    if (isa<SCEVCouldNotCompute>(SCEVA) || isa<SCEVCouldNotCompute>(SCEVB))
      return false;

    const APInt *Distance;
    ScalarEvolution &SE = *PSE->getSE();
    if (!match(SE.getMinusSCEV(SCEVA, SCEVB), m_scev_APInt(Distance)))
      return false;

    const DataLayout &DL = SE.getDataLayout();
    Type *TyA = A->getOperand(0)->getScalarType();
    uint64_t SizeA = DL.getTypeStoreSize(TyA);
    Type *TyB = B->getOperand(0)->getScalarType();
    uint64_t SizeB = DL.getTypeStoreSize(TyB);

    // Use the maximum store size to ensure no overlap from either direction.
    // Currently only handles fixed sizes, as it is only used for
    // replicating VPReplicateRecipes.
    uint64_t MaxStoreSize = std::max(SizeA, SizeB);

    auto VFs = B->getParent()->getPlan()->vectorFactors();
    ElementCount MaxVF = *max_element(VFs, ElementCount::isKnownLT);
    if (MaxVF.isScalable())
      return false;
    return Distance->abs().uge(MaxVF.getFixedValue() * MaxStoreSize);
  }

public:
  SinkStoreInfo(ArrayRef<VPReplicateRecipe *> ExcludeRecipes,
                VPReplicateRecipe &GroupLeader, PredicatedScalarEvolution &PSE,
                const Loop &L)
      : ExcludeRecipes(ExcludeRecipes.begin(), ExcludeRecipes.end()),
        GroupLeader(GroupLeader), PSE(&PSE), L(&L) {}

  SinkStoreInfo(VPReplicateRecipe &GroupLeader) : GroupLeader(GroupLeader) {}

  /// Return true if \p R should be skipped during alias checking, either
  /// because it's in the exclude set or because no-alias can be proven via
  /// SCEV.
  bool shouldSkip(VPRecipeBase &R) const {
    auto *Store = dyn_cast<VPReplicateRecipe>(&R);
    return ExcludeRecipes.contains(Store) ||
           (Store && isNoAliasViaDistance(Store, &GroupLeader));
  }
};

/// Check if a memory operation doesn't alias with memory operations using
/// scoped noalias metadata, in blocks in the single-successor chain between \p
/// FirstBB and \p LastBB. If \p SinkInfo is std::nullopt, only recipes that may
/// write to memory are checked (for load hoisting). Otherwise recipes that both
/// read and write memory are checked, and SCEV is used to prove no-alias
/// between the group leader and other replicate recipes (for store sinking).
static bool
canHoistOrSinkWithNoAliasCheck(const MemoryLocation &MemLoc,
                               VPBasicBlock *FirstBB, VPBasicBlock *LastBB,
                               std::optional<SinkStoreInfo> SinkInfo = {}) {
  bool CheckReads = SinkInfo.has_value();
  for (VPBasicBlock *VPBB :
       VPBlockUtils::blocksInSingleSuccessorChainBetween(FirstBB, LastBB)) {
    for (VPRecipeBase &R : *VPBB) {
      if (SinkInfo && SinkInfo->shouldSkip(R))
        continue;

      // Skip recipes that don't need checking.
      if (!R.mayWriteToMemory() && !(CheckReads && R.mayReadFromMemory()))
        continue;

      auto Loc = vputils::getMemoryLocation(R);
      if (!Loc)
        // Conservatively assume aliasing for memory operations without
        // location.
        return false;

      if (ScopedNoAliasAAResult::alias(*Loc, MemLoc) != AliasResult::NoAlias)
        return false;
    }
  }
  return true;
}

/// Get the value type of the replicate load or store. \p IsLoad indicates
/// whether it is a load.
static Type *getLoadStoreValueType(VPReplicateRecipe *R, bool IsLoad) {
  return (IsLoad ? R : R->getOperand(0))->getScalarType();
}

/// Collect either replicated Loads or Stores grouped by their address SCEV and
/// their load-store type, in a deep-traversal of the vector loop region in \p
/// Plan.
template <unsigned Opcode>
static SmallVector<SmallVector<VPReplicateRecipe *, 4>>
collectGroupedReplicateMemOps(
    VPlan &Plan, PredicatedScalarEvolution &PSE, const Loop *L,
    function_ref<bool(VPReplicateRecipe *)> FilterFn) {
  static_assert(Opcode == Instruction::Load || Opcode == Instruction::Store,
                "Only Load and Store opcodes supported");
  constexpr bool IsLoad = (Opcode == Instruction::Load);
  SmallDenseMap<std::pair<const SCEV *, const Type *>,
                SmallVector<VPReplicateRecipe *, 4>>
      RecipesByAddressAndType;
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_deep(Plan.getVectorLoopRegion()->getEntry()))) {
    for (VPReplicateRecipe &RepR : make_isa_range<VPReplicateRecipe>(*VPBB)) {
      if (RepR.getOpcode() != Opcode || !FilterFn(&RepR))
        continue;

      // For loads, operand 0 is address; for stores, operand 1 is address.
      VPValue *Addr = RepR.getOperand(IsLoad ? 0 : 1);
      const Type *LoadStoreTy = getLoadStoreValueType(&RepR, IsLoad);
      const SCEV *AddrSCEV = vputils::getSCEVExprForVPValue(Addr, PSE, L);
      if (!isa<SCEVCouldNotCompute>(AddrSCEV))
        RecipesByAddressAndType[{AddrSCEV, LoadStoreTy}].push_back(&RepR);
    }
  }
  auto Groups = to_vector(RecipesByAddressAndType.values());
  VPDominatorTree VPDT(Plan);
  for (auto &Group : Groups) {
    // Sort mem ops by dominance order, with earliest (most dominating) first.
    stable_sort(Group, [&VPDT](VPReplicateRecipe *A, VPReplicateRecipe *B) {
      return VPDT.properlyDominates(A, B);
    });
  }
  return Groups;
}

static bool sinkScalarOperands(VPlan &Plan) {
  auto Iter = vp_depth_first_deep(Plan.getEntry());
  bool ScalarVFOnly = Plan.hasScalarVFOnly();
  bool Changed = false;

  SetVector<std::pair<VPBasicBlock *, VPSingleDefRecipe *>> WorkList;
  auto InsertIfValidSinkCandidate = [ScalarVFOnly, &WorkList](
                                        VPBasicBlock *SinkTo, VPValue *Op) {
    auto *Candidate = dyn_cast<VPSingleDefRecipe>(Op);
    if (!isa_and_nonnull<VPReplicateRecipe, VPScalarIVStepsRecipe,
                         VPInstruction>(Candidate))
      return;

    if (Candidate->getParent() == SinkTo ||
        all_of(Candidate->operands(),
               [](VPValue *Op) { return Op->isDefinedOutsideLoopRegions(); }) ||
        vputils::cannotHoistOrSinkRecipe(*Candidate, /*Sinking=*/true))
      return;

    if (!ScalarVFOnly && !vputils::doesGeneratePerAllLanes(Candidate))
      return;

    // Only single-scalar VPInstructions can be sunk.
    if (auto *VPI = dyn_cast<VPInstruction>(Candidate))
      if (!vputils::isSingleScalar(VPI))
        return;

    WorkList.insert({SinkTo, Candidate});
  };

  // First, collect the operands of all recipes in replicate blocks as seeds for
  // sinking.
  for (VPRegionBlock *VPR : VPBlockUtils::blocksOnly<VPRegionBlock>(Iter)) {
    VPBasicBlock *EntryVPBB = VPR->getEntryBasicBlock();
    if (!VPR->isReplicator() || EntryVPBB->getSuccessors().size() != 2)
      continue;
    VPBasicBlock *VPBB = cast<VPBasicBlock>(EntryVPBB->getSuccessors().front());
    if (VPBB->getSingleSuccessor() != VPR->getExitingBasicBlock())
      continue;
    for (auto &Recipe : *VPBB)
      for (VPValue *Op : Recipe.operands())
        InsertIfValidSinkCandidate(VPBB, Op);
  }

  // Try to sink each replicate or scalar IV steps recipe in the worklist.
  for (unsigned I = 0; I != WorkList.size(); ++I) {
    VPBasicBlock *SinkTo;
    VPSingleDefRecipe *SinkCandidate;
    std::tie(SinkTo, SinkCandidate) = WorkList[I];

    // All recipe users of SinkCandidate must be in the same block SinkTo or all
    // users outside of SinkTo must only use the first lane of SinkCandidate. In
    // the latter case, we need to duplicate SinkCandidate.
    auto UsersOutsideSinkTo =
        make_filter_range(SinkCandidate->users(), [SinkTo](VPUser *U) {
          return cast<VPRecipeBase>(U)->getParent() != SinkTo;
        });
    if (any_of(UsersOutsideSinkTo, [SinkCandidate](VPUser *U) {
          return !U->usesFirstLaneOnly(SinkCandidate);
        }))
      continue;
    bool NeedsDuplicating = !UsersOutsideSinkTo.empty();

    if (NeedsDuplicating) {
      if (ScalarVFOnly)
        continue;
      VPSingleDefRecipe *Clone;
      if (auto *SinkCandidateRepR =
              dyn_cast<VPReplicateRecipe>(SinkCandidate)) {
        // TODO: Handle converting to uniform recipes as separate transform,
        // then cloning should be sufficient here.
        Clone = VPBuilder::createSingleScalarOp(
            SinkCandidateRepR->getOpcode(), SinkCandidate->operands(),
            /*Mask=*/nullptr, *SinkCandidateRepR, *SinkCandidateRepR,
            SinkCandidate->getDebugLoc(), SinkCandidate->getScalarType(),
            SinkCandidate->getUnderlyingInstr());
        // TODO: add ".cloned" suffix to name of Clone's VPValue.
      } else {
        Clone = SinkCandidate->clone();
      }

      Clone->insertBefore(SinkCandidate);
      SinkCandidate->replaceUsesWithIf(Clone, [SinkTo](VPUser &U, unsigned) {
        return cast<VPRecipeBase>(&U)->getParent() != SinkTo;
      });
    }
    SinkCandidate->moveBefore(*SinkTo, SinkTo->getFirstNonPhi());
    for (VPValue *Op : SinkCandidate->operands())
      InsertIfValidSinkCandidate(SinkTo, Op);
    Changed = true;
  }
  return Changed;
}

/// If \p R is a triangle region, return the 'then' block of the triangle.
static VPBasicBlock *getPredicatedThenBlock(VPRegionBlock *R) {
  auto *EntryBB = cast<VPBasicBlock>(R->getEntry());
  if (EntryBB->getNumSuccessors() != 2)
    return nullptr;

  auto *Succ0 = dyn_cast<VPBasicBlock>(EntryBB->getSuccessors()[0]);
  auto *Succ1 = dyn_cast<VPBasicBlock>(EntryBB->getSuccessors()[1]);
  if (!Succ0 || !Succ1)
    return nullptr;

  if (Succ0->getNumSuccessors() + Succ1->getNumSuccessors() != 1)
    return nullptr;
  if (Succ0->getSingleSuccessor() == Succ1)
    return Succ0;
  if (Succ1->getSingleSuccessor() == Succ0)
    return Succ1;
  return nullptr;
}

// Merge replicate regions in their successor region, if a replicate region
// is connected to a successor replicate region with the same predicate by a
// single, empty VPBasicBlock.
static bool mergeReplicateRegionsIntoSuccessors(VPlan &Plan) {
  SmallPtrSet<VPRegionBlock *, 4> TransformedRegions;

  // Collect replicate regions followed by an empty block, followed by another
  // replicate region with matching masks to process front. This is to avoid
  // iterator invalidation issues while merging regions.
  SmallVector<VPRegionBlock *, 8> WorkList;
  for (VPRegionBlock *Region1 : VPBlockUtils::blocksOnly<VPRegionBlock>(
           vp_depth_first_deep(Plan.getEntry()))) {
    if (!Region1->isReplicator())
      continue;
    auto *MiddleBasicBlock =
        dyn_cast_or_null<VPBasicBlock>(Region1->getSingleSuccessor());
    if (!MiddleBasicBlock || !MiddleBasicBlock->empty())
      continue;

    auto *Region2 =
        dyn_cast_or_null<VPRegionBlock>(MiddleBasicBlock->getSingleSuccessor());
    if (!Region2 || !Region2->isReplicator())
      continue;

    VPValue *Mask1 = Region1->getEntryBranchOnMask()->getOperand(0);
    VPValue *Mask2 = Region2->getEntryBranchOnMask()->getOperand(0);
    if (!Mask1 || Mask1 != Mask2)
      continue;

    assert(Mask1 && Mask2 && "both region must have conditions");
    WorkList.push_back(Region1);
  }

  // Move recipes from Region1 to its successor region, if both are triangles.
  for (VPRegionBlock *Region1 : WorkList) {
    if (TransformedRegions.contains(Region1))
      continue;
    auto *MiddleBasicBlock = cast<VPBasicBlock>(Region1->getSingleSuccessor());
    auto *Region2 = cast<VPRegionBlock>(MiddleBasicBlock->getSingleSuccessor());

    VPBasicBlock *Then1 = getPredicatedThenBlock(Region1);
    VPBasicBlock *Then2 = getPredicatedThenBlock(Region2);
    if (!Then1 || !Then2)
      continue;

    // The merged region is entered whenever either of the original regions was,
    // so use the higher, i.e. more conservative, of their entry frequencies.
    // If only one of the two is known, the higher one is unknown, so the
    // result must be unknown too.
    VPBranchOnMaskRecipe *Guard2 = Region2->getEntryBranchOnMask();
    std::optional<VPExecutionFrequency> Freq1 =
        Region1->getEntryBranchOnMask()->getExecutionFrequency();
    std::optional<VPExecutionFrequency> Freq2 = Guard2->getExecutionFrequency();
    if (Freq1 && Freq2) {
      if (Freq2->Freq < Freq1->Freq) {
        // Freq1's frequency is taken, but it is only as trustworthy as the
        // less trustworthy of the two.
        Freq1.emplace(Freq1->Freq, Freq1->IsEstimated || Freq2->IsEstimated);
        Guard2->setExecutionFrequency(Freq1, Plan.getContext());
      }
    } else if (Freq2) {
      Guard2->clearExecutionFrequency();
    }

    // Note: No fusion-preventing memory dependencies are expected in either
    // region. Such dependencies should be rejected during earlier dependence
    // checks, which guarantee accesses can be re-ordered for vectorization.
    //
    // Move recipes to the successor region.
    for (VPRecipeBase &ToMove : make_early_inc_range(reverse(*Then1)))
      ToMove.moveBefore(*Then2, Then2->getFirstNonPhi());

    auto *Merge1 = cast<VPBasicBlock>(Then1->getSingleSuccessor());
    auto *Merge2 = cast<VPBasicBlock>(Then2->getSingleSuccessor());

    // Move VPPredInstPHIRecipes from the merge block to the successor region's
    // merge block. Update all users inside the successor region to use the
    // original values.
    for (VPRecipeBase &Phi1ToMove : make_early_inc_range(reverse(*Merge1))) {
      VPValue *PredInst1 =
          cast<VPPredInstPHIRecipe>(&Phi1ToMove)->getOperand(0);
      VPValue *Phi1ToMoveV = Phi1ToMove.getVPSingleValue();
      Phi1ToMoveV->replaceUsesWithIf(PredInst1, [Then2](VPUser &U, unsigned) {
        return cast<VPRecipeBase>(&U)->getParent() == Then2;
      });

      // Remove phi recipes that are unused after merging the regions.
      if (Phi1ToMove.getVPSingleValue()->user_empty()) {
        Phi1ToMove.eraseFromParent();
        continue;
      }
      Phi1ToMove.moveBefore(*Merge2, Merge2->begin());
    }

    // Remove the dead recipes in Region1's entry block.
    for (VPRecipeBase &R :
         make_early_inc_range(reverse(*Region1->getEntryBasicBlock())))
      R.eraseFromParent();

    // Finally, remove the first region.
    for (VPBlockBase *Pred : make_early_inc_range(Region1->getPredecessors())) {
      VPBlockUtils::disconnectBlocks(Pred, Region1);
      VPBlockUtils::connectBlocks(Pred, MiddleBasicBlock);
    }
    VPBlockUtils::disconnectBlocks(Region1, MiddleBasicBlock);
    TransformedRegions.insert(Region1);
  }

  return !TransformedRegions.empty();
}

static VPRegionBlock *createReplicateRegion(VPReplicateRecipe *PredRecipe,
                                            VPRegionBlock *ParentRegion,
                                            VPlan &Plan) {
  Instruction *Instr = PredRecipe->getUnderlyingInstr();
  // Build the triangular if-then region.
  std::string RegionName = (Twine("pred.") + Instr->getOpcodeName()).str();
  assert(Instr->getParent() && "Predicated instruction not in any basic block");
  auto *BlockInMask = PredRecipe->getMask();
  auto *MaskDef = BlockInMask->getDefiningRecipe();
  auto *BOMRecipe = new VPBranchOnMaskRecipe(
      BlockInMask, MaskDef ? MaskDef->getDebugLoc() : DebugLoc::getUnknown());
  auto *Entry =
      Plan.createVPBasicBlock(Twine(RegionName) + ".entry", BOMRecipe);

  // Replace predicated replicate recipe with a replicate recipe without a
  // mask but in the replicate region.
  auto *RecipeWithoutMask = new VPReplicateRecipe(
      PredRecipe->getUnderlyingInstr(), PredRecipe->operandsWithoutMask(),
      PredRecipe->isSingleScalar(), nullptr /*Mask*/, *PredRecipe, *PredRecipe,
      PredRecipe->getDebugLoc());
  // The predicated recipe executes exactly when the guarding branch-on-mask is
  // taken, so move its execution frequency there.
  BOMRecipe->setExecutionFrequency(RecipeWithoutMask->getExecutionFrequency(),
                                   Plan.getContext());
  RecipeWithoutMask->clearExecutionFrequency();
  auto *Pred =
      Plan.createVPBasicBlock(Twine(RegionName) + ".if", RecipeWithoutMask);
  auto *Exiting = Plan.createVPBasicBlock(Twine(RegionName) + ".continue");
  VPRegionBlock *Region =
      Plan.createReplicateRegion(Entry, Exiting, RegionName);

  // Note: first set Entry as region entry and then connect successors starting
  // from it in order, to propagate the "parent" of each VPBasicBlock.
  Region->setParent(ParentRegion);
  VPBlockUtils::insertTwoBlocksAfter(Pred, Exiting, Entry);
  VPBlockUtils::connectBlocks(Pred, Exiting);

  if (!PredRecipe->user_empty()) {
    auto *PHIRecipe = new VPPredInstPHIRecipe(RecipeWithoutMask,
                                              RecipeWithoutMask->getDebugLoc());
    Exiting->appendRecipe(PHIRecipe);
    PredRecipe->replaceAllUsesWith(PHIRecipe);
  }
  PredRecipe->eraseFromParent();
  return Region;
}

static void addReplicateRegions(VPlan &Plan) {
  SmallVector<VPReplicateRecipe *> WorkList;
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_deep(Plan.getEntry()))) {
    for (VPReplicateRecipe &RepR : make_isa_range<VPReplicateRecipe>(*VPBB))
      if (RepR.isPredicated())
        WorkList.push_back(&RepR);
  }

  unsigned BBNum = 0;
  for (VPReplicateRecipe *RepR : WorkList) {
    VPBasicBlock *CurrentBlock = RepR->getParent();
    VPBasicBlock *SplitBlock = CurrentBlock->splitAt(RepR->getIterator());

    BasicBlock *OrigBB = RepR->getUnderlyingInstr()->getParent();
    SplitBlock->setName(
        OrigBB->hasName() ? OrigBB->getName() + "." + Twine(BBNum++) : "");
    // Record predicated instructions for above packing optimizations.
    VPRegionBlock *Region =
        createReplicateRegion(RepR, CurrentBlock->getParent(), Plan);
    VPBlockUtils::insertOnEdge(CurrentBlock, SplitBlock, Region);

    VPRegionBlock *ParentRegion = Region->getParent();
    if (ParentRegion && ParentRegion->getExiting() == CurrentBlock)
      ParentRegion->setExiting(SplitBlock);
  }
}

bool VPlanTransforms::mergeBlocksIntoPredecessors(VPlan &Plan) {
  SmallVector<VPBasicBlock *> WorkList;
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_deep(Plan.getEntry()))) {
    // Don't fold the blocks in the skeleton of the Plan into their single
    // predecessors for now.
    // TODO: Remove restriction once more of the skeleton is modeled in VPlan.
    if (!VPBB->getParent())
      continue;
    auto *PredVPBB =
        dyn_cast_or_null<VPBasicBlock>(VPBB->getSinglePredecessor());
    if (!PredVPBB || PredVPBB->getNumSuccessors() != 1 ||
        isa<VPIRBasicBlock>(PredVPBB))
      continue;
    WorkList.push_back(VPBB);
  }

  for (VPBasicBlock *VPBB : WorkList) {
    VPBasicBlock *PredVPBB = cast<VPBasicBlock>(VPBB->getSinglePredecessor());
    for (VPRecipeBase &R : make_early_inc_range(*VPBB))
      R.moveBefore(*PredVPBB, PredVPBB->end());
    VPBlockUtils::disconnectBlocks(PredVPBB, VPBB);
    auto *ParentRegion = VPBB->getParent();
    if (ParentRegion && ParentRegion->getExiting() == VPBB)
      ParentRegion->setExiting(PredVPBB);
    VPBlockUtils::transferSuccessors(VPBB, PredVPBB);
    // VPBB is now dead and will be cleaned up when the plan gets destroyed.
  }
  return !WorkList.empty();
}

void VPlanTransforms::createAndOptimizeReplicateRegions(VPlan &Plan) {
  // Convert masked VPReplicateRecipes to if-then region blocks.
  addReplicateRegions(Plan);

  bool ShouldSimplify = true;
  while (ShouldSimplify) {
    ShouldSimplify = sinkScalarOperands(Plan);
    ShouldSimplify |= mergeReplicateRegionsIntoSuccessors(Plan);
    ShouldSimplify |= mergeBlocksIntoPredecessors(Plan);
  }
}

/// Remove redundant casts of inductions.
///
/// Such redundant casts are casts of induction variables that can be ignored,
/// because we already proved that the casted phi is equal to the uncasted phi
/// in the vectorized loop. There is no need to vectorize the cast - the same
/// value can be used for both the phi and casts in the vector loop.
static void removeRedundantInductionCasts(VPlan &Plan) {
  for (VPWidenIntOrFpInductionRecipe &IV :
       make_isa_range<VPWidenIntOrFpInductionRecipe>(
           Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis())) {
    if (IV.getTruncInst())
      continue;

    // A sequence of IR Casts has potentially been recorded for IV, which
    // *must be bypassed* when the IV is vectorized, because the vectorized IV
    // will produce the desired casted value. This sequence forms a def-use
    // chain and is provided in reverse order, ending with the cast that uses
    // the IV phi. Search for the recipe of the last cast in the chain and
    // replace it with the original IV. Note that only the final cast is
    // expected to have users outside the cast-chain and the dead casts left
    // over will be cleaned up later.
    ArrayRef<Instruction *> Casts = IV.getInductionDescriptor().getCastInsts();
    VPValue *FindMyCast = &IV;
    for (Instruction *IRCast : reverse(Casts)) {
      VPSingleDefRecipe *FoundUserCast = nullptr;
      for (auto *U : FindMyCast->users()) {
        auto *UserCast = dyn_cast<VPSingleDefRecipe>(U);
        if (UserCast && UserCast->getUnderlyingValue() == IRCast) {
          FoundUserCast = UserCast;
          break;
        }
      }
      // A cast recipe in the chain may have been removed by earlier DCE.
      if (!FoundUserCast)
        break;
      FindMyCast = FoundUserCast;
    }
    if (FindMyCast != &IV)
      FindMyCast->replaceAllUsesWith(&IV);
  }
}

/// If R is a phi-like recipe starting a dead cycle of recipes, erase all
/// reachable recipes of the dead cycle and return true. Otherwise leave the
/// plan unchanged and return false.
static bool tryToRemoveDeadCycle(VPRecipeBase *R) {
  auto *PhiR = dyn_cast<VPSingleDefRecipe>(R);
  if (!PhiR || !isa<VPPhi, VPReductionPHIRecipe>(R))
    return false;

  // The transitive users of PhiR are closed under users, so the cycle is dead
  // if every one of them can be erased.
  for (VPUser *U : vputils::collectUsersRecursively(PhiR)) {
    auto *R = cast<VPRecipeBase>(U);
    // Bail out if a user must be retained, or if it is a phi-like recipe other
    // than PhiR;
    if (R->mayHaveSideEffects() || (R != PhiR && isa<VPPhiAccessors>(R)))
      return false;
  }

  // Break the cycle by replacing PhiR with its first incoming value, which is
  // defined outside the cycle. That leaves the rest of the cycle dead.
  PhiR->replaceAllUsesWith(PhiR->getOperand(0));
  SmallVector<VPValue *> Incoming(PhiR->operands());
  PhiR->eraseFromParent();
  for (VPValue *Op : Incoming)
    vputils::recursivelyDeleteDeadRecipes(Op);
  return true;
}

void VPlanTransforms::removeDeadRecipes(VPlan &Plan) {
  PostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> POT(
      Plan.getEntry());
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(POT)) {
    // The recipes in the block are processed in reverse order, to catch chains
    // of dead recipes.
    for (VPRecipeBase &R : make_early_inc_range(reverse(*VPBB)))
      if (vputils::isDeadRecipe(R))
        R.eraseFromParent();

    // Erase dead cycles starting at one of VPBB's phi-like recipes. Erasing a
    // cycle may also erase other phi-like recipes of VPBB, so restart the scan
    // of the phi section after each removal. This terminates, as each removal
    // erases the cycle's phi.
    bool Changed = true;
    while (Changed) {
      Changed = false;
      for (VPRecipeBase &R : VPBB->phis()) {
        if (tryToRemoveDeadCycle(&R)) {
          Changed = true;
          break;
        }
      }
    }
  }
}

/// Legalize VPWidenPointerInductionRecipe, by replacing it with a PtrAdd
/// (IndStart, ScalarIVSteps (0, Step)) if only its scalar values are used, as
/// VPWidenPointerInductionRecipe will generate vectors only. If some users
/// require vectors while other require scalars, the scalar uses need to extract
/// the scalars from the generated vectors (Note that this is different to how
/// int/fp inductions are handled). Legalize extract-from-ends using uniform
/// VPReplicateRecipe of wide inductions to use regular VPReplicateRecipe, so
/// the correct end value is available. Also optimize
/// VPWidenIntOrFpInductionRecipe, if any of its users needs scalar values, by
/// providing them scalar steps built on the canonical scalar IV and update the
/// original IV's users. This is an optional optimization to reduce the needs of
/// vector extracts.
static void legalizeAndOptimizeInductions(VPlan &Plan) {
  VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
  bool HasOnlyVectorVFs = !Plan.hasScalarVFOnly();

  SmallVector<VPWidenInductionRecipe *> WideIVs;
  for (VPWidenInductionRecipe &PhiR :
       make_isa_range<VPWidenInductionRecipe>(HeaderVPBB->phis()))
    WideIVs.push_back(&PhiR);

  // Try to narrow wide and replicating recipes to uniform recipes, based on
  // VPlan analysis.
  // TODO: Apply to all recipes in the future, to replace legacy uniformity
  // analysis.
  for (VPWidenInductionRecipe *PhiR : WideIVs) {
    auto Users = vputils::collectUsersRecursively(PhiR);
    for (VPUser *U : reverse(Users)) {
      auto *Def = dyn_cast<VPRecipeWithIRFlags>(U);
      auto *RepR = dyn_cast<VPReplicateRecipe>(U);
      // Skip recipes that shouldn't be narrowed.
      if (!Def ||
          !isa<VPReplicateRecipe, VPWidenRecipe, VPWidenGEPRecipe>(Def) ||
          Def->user_empty() || !Def->getUnderlyingValue() ||
          (RepR && (RepR->isSingleScalar() || RepR->isPredicated())))
        continue;

      // Skip recipes that may have other lanes than their first used.
      if (!vputils::isSingleScalar(Def) && !vputils::onlyFirstLaneUsed(Def))
        continue;

      // TODO: Support scalarizing ExtractValue.
      if (match(Def,
                m_Binary<Instruction::ExtractValue>(m_VPValue(), m_VPValue())))
        continue;

      auto *Clone = VPBuilder::createSingleScalarOp(
          Def->getUnderlyingInstr()->getOpcode(), Def->operands(),
          /*Mask=*/nullptr, *Def, getMetadataOf(Def), DebugLoc::getUnknown(),
          Def->getScalarType(), Def->getUnderlyingInstr());
      Clone->insertAfter(Def);
      Def->replaceAllUsesWith(Clone);
      Def->eraseFromParent();
    }
  }

  VPBuilder Builder(HeaderVPBB, HeaderVPBB->getFirstNonPhi());
  for (VPWidenInductionRecipe *PhiR : WideIVs) {
    // Replace wide pointer inductions which have only their scalars used by
    // PtrAdd(IndStart, ScalarIVSteps (0, Step)).
    if (auto *PtrIV = dyn_cast<VPWidenPointerInductionRecipe>(PhiR)) {
      if (!Plan.hasScalarVFOnly() &&
          !PtrIV->onlyScalarsGenerated(Plan.hasScalableVF()))
        continue;

      VPValue *PtrAdd =
          vputils::scalarizeVPWidenPointerInduction(PtrIV, Plan, Builder);
      PtrIV->replaceAllUsesWith(PtrAdd);
      continue;
    }

    // Replace widened induction with scalar steps for users that only use
    // scalars.
    auto *WideIV = cast<VPWidenIntOrFpInductionRecipe>(PhiR);
    if (HasOnlyVectorVFs && none_of(WideIV->users(), [WideIV](VPUser *U) {
          return U->usesScalars(WideIV);
        }))
      continue;

    const InductionDescriptor &ID = WideIV->getInductionDescriptor();
    VPIRFlags::WrapFlagsTy WrapFlags;
    // We can preserve nuw when the step is non-negative.
    const APInt *Step;
    if (match(WideIV->getStepValue(), m_APInt(Step)) && Step->isNonNegative())
      WrapFlags = {static_cast<bool>(WideIV->getNoWrapFlagsOrNone().HasNUW),
                   false};
    VPScalarIVStepsRecipe *Steps = vputils::createScalarIVSteps(
        Plan, ID.getKind(), ID.getInductionOpcode(),
        dyn_cast_or_null<FPMathOperator>(ID.getInductionBinOp()),
        WideIV->getTruncInst(), WideIV->getStartValue(), WideIV->getStepValue(),
        WideIV->getDebugLoc(), Builder, WrapFlags);

    // Update scalar users of IV to use Step instead.
    if (!HasOnlyVectorVFs) {
      assert(!Plan.hasScalableVF() &&
             "plans containing a scalar VF cannot also include scalable VFs");
      WideIV->replaceAllUsesWith(Steps);
    } else {
      bool HasScalableVF = Plan.hasScalableVF();
      WideIV->replaceUsesWithIf(Steps,
                                [WideIV, HasScalableVF](VPUser &U, unsigned) {
                                  if (HasScalableVF)
                                    return U.usesFirstLaneOnly(WideIV);
                                  return U.usesScalars(WideIV);
                                });
    }
  }
}

/// Check if \p VPV is an untruncated wide induction, either before or after the
/// increment. If so return the header IV (before the increment), otherwise
/// return null.
static VPWidenInductionRecipe *
getOptimizableIVOf(VPValue *VPV, PredicatedScalarEvolution &PSE) {
  auto *WideIV = dyn_cast<VPWidenInductionRecipe>(VPV);
  if (WideIV) {
    // VPV itself is a wide induction, separately compute the end value for exit
    // users if it is not a truncated IV.
    auto *IntOrFpIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(WideIV);
    return (IntOrFpIV && IntOrFpIV->getTruncInst()) ? nullptr : WideIV;
  }

  // Check if VPV is an optimizable induction increment.
  VPRecipeBase *Def = VPV->getDefiningRecipe();
  if (!Def || Def->getNumOperands() != 2)
    return nullptr;
  WideIV = dyn_cast<VPWidenInductionRecipe>(Def->getOperand(0));
  if (!WideIV)
    WideIV = dyn_cast<VPWidenInductionRecipe>(Def->getOperand(1));
  if (!WideIV)
    return nullptr;

  auto IsWideIVInc = [&]() {
    auto &ID = WideIV->getInductionDescriptor();

    // Check if VPV increments the induction by the induction step.
    VPValue *IVStep = WideIV->getStepValue();
    switch (ID.getInductionOpcode()) {
    case Instruction::Add:
      return match(VPV, m_c_Add(m_Specific(WideIV), m_Specific(IVStep)));
    case Instruction::FAdd:
      return match(VPV, m_c_FAdd(m_Specific(WideIV), m_Specific(IVStep)));
    case Instruction::FSub:
      return match(VPV, m_Binary<Instruction::FSub>(m_Specific(WideIV),
                                                    m_Specific(IVStep)));
    case Instruction::Sub: {
      // IVStep will be the negated step of the subtraction. Check if Step == -1
      // * IVStep.
      VPValue *Step;
      if (!match(VPV, m_Sub(m_VPValue(), m_VPValue(Step))))
        return false;
      const SCEV *IVStepSCEV = vputils::getSCEVExprForVPValue(IVStep, PSE);
      const SCEV *StepSCEV = vputils::getSCEVExprForVPValue(Step, PSE);
      ScalarEvolution &SE = *PSE.getSE();
      return !isa<SCEVCouldNotCompute>(IVStepSCEV) &&
             !isa<SCEVCouldNotCompute>(StepSCEV) &&
             IVStepSCEV == SE.getNegativeSCEV(StepSCEV);
    }
    default:
      return ID.getKind() == InductionDescriptor::IK_PtrInduction &&
             match(VPV, m_GetElementPtr(m_Specific(WideIV),
                                        m_Specific(WideIV->getStepValue())));
    }
    llvm_unreachable("should have been covered by switch above");
  };
  return IsWideIVInc() ? WideIV : nullptr;
}

/// Attempts to optimize the induction variable exit values for users in the
/// early exit block.
static VPValue *optimizeEarlyExitInductionUser(VPlan &Plan, VPValue *Op,
                                               PredicatedScalarEvolution &PSE) {
  VPValue *Incoming, *Mask;
  if (!match(Op, m_ExtractLane(m_FirstActiveLane(m_VPValue(Mask)),
                               m_VPValue(Incoming))))
    return nullptr;

  auto *WideIV = getOptimizableIVOf(Incoming, PSE);
  if (!WideIV)
    return nullptr;

  // Calculate the final index.
  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
  auto *CanonicalIV = LoopRegion->getCanonicalIV();
  Type *CanonicalIVType = LoopRegion->getCanonicalIVType();
  auto *ExtractR = cast<VPInstruction>(Op);
  VPBuilder B(ExtractR);

  DebugLoc DL = ExtractR->getDebugLoc();
  VPValue *FirstActiveLane = B.createFirstActiveLane(Mask, DL);
  FirstActiveLane =
      B.createScalarZExtOrTrunc(FirstActiveLane, CanonicalIVType, DL);
  VPValue *EndValue = B.createAdd(CanonicalIV, FirstActiveLane, DL);

  // `getOptimizableIVOf()` always returns the pre-incremented IV, so if it
  // changed it means the exit is using the incremented value, so we need to
  // add the step.
  if (Incoming != WideIV) {
    VPValue *One = Plan.getConstantInt(CanonicalIVType, 1);
    EndValue = B.createAdd(EndValue, One, DL);
  }

  if (!match(WideIV, m_CanonicalWidenIV())) {
    const InductionDescriptor &ID = WideIV->getInductionDescriptor();
    VPValue *Start = WideIV->getStartValue();
    VPValue *Step = WideIV->getStepValue();
    EndValue = B.createDerivedIV(
        ID.getKind(), dyn_cast_or_null<FPMathOperator>(ID.getInductionBinOp()),
        Start, EndValue, Step);
  }

  return EndValue;
}

/// Compute the end value for \p WideIV, unless it is truncated. Creates a
/// VPDerivedIVRecipe for non-canonical inductions.
static VPValue *tryToComputeEndValueForInduction(VPWidenInductionRecipe *WideIV,
                                                 VPBuilder &VectorPHBuilder,
                                                 VPValue *VectorTC) {
  auto *WideIntOrFp = dyn_cast<VPWidenIntOrFpInductionRecipe>(WideIV);
  // Truncated wide inductions resume from the last lane of their vector value
  // in the last vector iteration which is handled elsewhere.
  if (WideIntOrFp && WideIntOrFp->getTruncInst())
    return nullptr;

  VPValue *Start = WideIV->getStartValue();
  VPValue *Step = WideIV->getStepValue();
  const InductionDescriptor &ID = WideIV->getInductionDescriptor();
  VPValue *EndValue = VectorTC;
  if (!match(WideIV, m_CanonicalWidenIV())) {
    EndValue = VectorPHBuilder.createDerivedIV(
        ID.getKind(), dyn_cast_or_null<FPMathOperator>(ID.getInductionBinOp()),
        Start, VectorTC, Step);
  }

  // EndValue is derived from the vector trip count (which has the same type as
  // the widest induction) and thus may be wider than the induction here.
  Type *ScalarTypeOfWideIV = WideIV->getScalarType();
  if (ScalarTypeOfWideIV != EndValue->getScalarType()) {
    EndValue = VectorPHBuilder.createScalarCast(Instruction::Trunc, EndValue,
                                                ScalarTypeOfWideIV,
                                                WideIV->getDebugLoc());
  }

  return EndValue;
}

/// Attempts to optimize the induction variable exit values for users in the
/// exit block coming from the latch in the original scalar loop.
static VPValue *
optimizeLatchExitInductionUser(VPlan &Plan, VPValue *Op,
                               DenseMap<VPValue *, VPValue *> &EndValues,
                               PredicatedScalarEvolution &PSE) {
  VPValue *Incoming;
  if (!match(Op, m_CombineOr(m_ExtractLastLaneOfLastPart(m_VPValue(Incoming)),
                             m_ExtractLane(m_LastActiveLane(m_HeaderMask()),
                                           m_VPValue(Incoming)))))
    return nullptr;

  VPWidenInductionRecipe *WideIV = getOptimizableIVOf(Incoming, PSE);
  if (!WideIV)
    return nullptr;

  VPValue *EndValue = EndValues.lookup(WideIV);
  assert(EndValue && "Must have computed the end value up front");

  // `getOptimizableIVOf()` always returns the pre-incremented IV, so if it
  // changed it means the exit is using the incremented value, so we don't
  // need to subtract the step.
  if (Incoming != WideIV)
    return EndValue;

  // Otherwise, subtract the step from the EndValue.
  auto *ExtractR = cast<VPInstruction>(Op);
  VPBuilder B(ExtractR);
  VPValue *Step = WideIV->getStepValue();
  Type *ScalarTy = WideIV->getScalarType();
  if (ScalarTy->isIntegerTy())
    return B.createSub(EndValue, Step, DebugLoc::getUnknown(), "ind.escape");
  if (ScalarTy->isPointerTy()) {
    Type *StepTy = Step->getScalarType();
    auto *Zero = Plan.getZero(StepTy);
    return B.createPtrAdd(EndValue, B.createSub(Zero, Step),
                          DebugLoc::getUnknown(), "ind.escape");
  }
  if (ScalarTy->isFloatingPointTy()) {
    const auto &ID = WideIV->getInductionDescriptor();
    return B.createNaryOp(
        ID.getInductionBinOp()->getOpcode() == Instruction::FAdd
            ? Instruction::FSub
            : Instruction::FAdd,
        {EndValue, Step}, {ID.getInductionBinOp()->getFastMathFlags()});
  }
  llvm_unreachable("all possible induction types must be handled");
  return nullptr;
}

static VPValue *optimizeLatchExitIVUserViaSCEV(VPlan &Plan, VPValue *Op,
                                               PredicatedScalarEvolution &PSE,
                                               VPValue *ResumeTC,
                                               const Loop *L) {
  VPValue *Incoming;
  if (!match(Op, m_CombineOr(m_ExtractLastLaneOfLastPart(m_VPValue(Incoming)),
                             m_ExtractLane(m_LastActiveLane(m_HeaderMask()),
                                           m_VPValue(Incoming)))))
    return nullptr;

  const SCEV *IncomingSCEV = vputils::getSCEVExprForVPValue(Incoming, PSE, L);
  const SCEV *Start, *Step;
  if (!match(IncomingSCEV, m_scev_AffineAddRec(m_SCEV(Start), m_SCEV(Step),
                                               m_SpecificLoop(L))))
    return nullptr;

  auto *ExtractR = cast<VPInstruction>(Op);
  DebugLoc DL = ExtractR->getDebugLoc();
  VPBuilder Builder(ExtractR);
  VPSCEVExpander Expander(Builder, *PSE.getSE(), DL);
  VPValue *StartVPV = Expander.expand(Start);
  VPValue *StepVPV = Expander.expand(Step);

  Type *StartTy = StartVPV->getScalarType();
  assert(StartTy->isIntOrPtrTy() && "The type must be SCEVable");
  InductionDescriptor::InductionKind Kind =
      StartTy->isPointerTy() ? InductionDescriptor::IK_PtrInduction
                             : InductionDescriptor::IK_IntInduction;
  Type *TCTy = ResumeTC->getScalarType();
  VPValue *ExitCount = Builder.createOverflowingOp(
      Instruction::Sub, {ResumeTC, Plan.getConstantInt(TCTy, 1)},
      {/*HasNUW=*/true, /*HasNSW=*/false}, DebugLoc::getUnknown());
  return Builder.createDerivedIV(Kind, /*FPBinOp=*/nullptr, StartVPV, ExitCount,
                                 StepVPV);
}

void VPlanTransforms::optimizeInductionLiveOutUsers(
    VPlan &Plan, PredicatedScalarEvolution &PSE, const Loop *L) {
  // Compute end values for all inductions.
  VPRegionBlock *VectorRegion = Plan.getVectorLoopRegion();
  auto *VectorPH = cast<VPBasicBlock>(VectorRegion->getSinglePredecessor());
  VPBuilder VectorPHBuilder(VectorPH, VectorPH->getFirstNonPhi());
  DenseMap<VPValue *, VPValue *> EndValues;
  VPValue *ResumeTC =
      Plan.hasTailFolded() ? Plan.getTripCount() : &Plan.getVectorTripCount();
  for (VPWidenInductionRecipe &WideIV : make_isa_range<VPWidenInductionRecipe>(
           VectorRegion->getEntryBasicBlock()->phis())) {
    if (VPValue *EndValue = tryToComputeEndValueForInduction(
            &WideIV, VectorPHBuilder, ResumeTC))
      EndValues[&WideIV] = EndValue;
  }

  VPBasicBlock *MiddleVPBB = Plan.getMiddleBlock();
  for (VPRecipeBase &R : make_early_inc_range(*MiddleVPBB)) {
    VPValue *Op;
    if (!match(&R, m_ExitingIVValue(m_VPValue(Op))))
      continue;
    auto *WideIV = cast<VPWidenInductionRecipe>(Op);
    if (VPValue *EndValue = EndValues.lookup(WideIV)) {
      R.getVPSingleValue()->replaceAllUsesWith(EndValue);
      R.eraseFromParent();
    }
  }

  // Then, optimize exit block users.
  for (VPIRBasicBlock *ExitVPBB : Plan.getExitBlocks()) {
    for (VPRecipeBase &R : ExitVPBB->phis()) {
      auto *ExitIRI = cast<VPIRPhi>(&R);

      for (auto [Idx, PredVPBB] : enumerate(ExitVPBB->getPredecessors())) {
        VPValue *Escape = nullptr;
        if (PredVPBB == MiddleVPBB) {
          Escape = optimizeLatchExitInductionUser(
              Plan, ExitIRI->getOperand(Idx), EndValues, PSE);
          if (!Escape)
            Escape = optimizeLatchExitIVUserViaSCEV(
                Plan, ExitIRI->getOperand(Idx), PSE, ResumeTC, L);
        } else {
          Escape = optimizeEarlyExitInductionUser(
              Plan, ExitIRI->getOperand(Idx), PSE);
        }
        if (Escape)
          ExitIRI->setOperand(Idx, Escape);
      }
    }
  }
}

/// Remove redundant ExpandSCEVRecipes in \p Plan's entry block by replacing
/// them with already existing recipes expanding the same SCEV expression.
static void removeRedundantExpandSCEVRecipes(VPlan &Plan) {
  DenseMap<const SCEV *, VPValue *> SCEV2VPV;

  for (VPExpandSCEVRecipe &ExpR :
       make_early_inc_range(make_isa_range<VPExpandSCEVRecipe>(
           *Plan.getEntry()->getEntryBasicBlock()))) {
    const auto &[V, Inserted] = SCEV2VPV.try_emplace(ExpR.getSCEV(), &ExpR);
    if (Inserted)
      continue;

    ExpR.replaceAllUsesWith(V->second);
    if (&ExpR == Plan.getTripCount())
      Plan.resetTripCount(V->second);

    ExpR.eraseFromParent();
  }
}

/// Try to simplify logical and bitwise recipes in \p Def.
static VPValue *simplifyLogicalRecipe(VPlan &Plan, VPSingleDefRecipe *Def) {
  VPValue *X;

  // X | AllOnes -> AllOnes
  if (match(Def, m_c_BinaryOr(m_VPValue(X), m_AllOnes())))
    return Plan.getAllOnesValue(Def->getScalarType());

  // X | 0 -> X
  if (match(Def, m_c_BinaryOr(m_VPValue(X), m_ZeroInt())))
    return X;

  // X | !X -> AllOnes
  if (match(Def, m_c_BinaryOr(m_VPValue(X), m_Not(m_Deferred(X)))))
    return Plan.getAllOnesValue(Def->getScalarType());

  // X & 0 -> 0
  if (match(Def, m_c_BinaryAnd(m_VPValue(X), m_ZeroInt())))
    return Plan.getZero(Def->getScalarType());

  // X & AllOnes -> X
  if (match(Def, m_c_BinaryAnd(m_VPValue(X), m_AllOnes())))
    return X;

  // X && false -> false
  if (match(Def, m_c_LogicalAnd(m_VPValue(X), m_False())))
    return Plan.getFalse();

  // X && true -> X
  if (match(Def, m_c_LogicalAnd(m_VPValue(X), m_True())))
    return X;

  // X && (X && Y) -> X && Y
  if (match(Def, m_LogicalAnd(m_VPValue(X),
                              m_LogicalAnd(m_Deferred(X), m_VPValue()))))
    return Def->getOperand(1);

  // X && !X -> 0
  if (match(Def, m_LogicalAnd(m_VPValue(X), m_Not(m_Deferred(X)))))
    return Plan.getFalse();

  if (match(Def, m_Select(m_VPValue(), m_VPValue(X), m_Deferred(X))))
    return X;

  return nullptr;
}

/// Return an existing value or a live in for VPSingleDefRecipe \p Def if
/// possible. This shouldn't create or modify recipes.
static VPValue *simplifyRecipe(VPlan &Plan, VPSingleDefRecipe *Def) {
  // Simplification of live-in IR values for SingleDef recipes using
  // InstSimplifyFolder.
  const DataLayout &DL = Plan.getDataLayout();
  if (VPValue *V = vputils::tryToFoldLiveIns(*Def, Def->operands(), DL))
    return V;

  // Fold PredPHI LiveIn -> LiveIn.
  if (auto *PredPHI = dyn_cast<VPPredInstPHIRecipe>(Def)) {
    VPValue *Op = PredPHI->getOperand(0);
    if (isa<VPIRValue>(Op))
      return Op;
  }

  if (VPValue *V = simplifyLogicalRecipe(Plan, Def))
    return V;

  VPValue *A, *B;

  if (match(Def, m_c_Add(m_VPValue(A), m_ZeroInt())))
    return A;

  if (match(Def, m_c_Mul(m_VPValue(A), m_One())))
    return A;

  if (match(Def, m_c_Mul(m_VPValue(), m_ZeroInt())))
    return Plan.getZero(Def->getScalarType());

  // A bitcast to the same type is a no-op.
  if (match(Def, m_BitCast(m_VPValue(A))) &&
      Def->getScalarType() == A->getScalarType())
    return A;

  // Shifting by zero is a no-op.
  if (match(Def, m_CombineOr(m_Shl(m_VPValue(A), m_ZeroInt()),
                             m_CombineOr(m_LShr(m_VPValue(A), m_ZeroInt()),
                                         m_AShr(m_VPValue(A), m_ZeroInt())))))
    return A;

  if (match(Def, m_Trunc(m_ZExtOrSExt(m_VPValue(A)))))
    if (Def->getScalarType() == A->getScalarType())
      return A;

  if (match(Def, m_Not(m_Not(m_VPValue(A)))))
    return A;

  // Remove redundant DerviedIVs, that is 0 + A * 1 -> A and 0 + 0 * x -> 0.
  if ((match(Def, m_DerivedIV(m_ZeroInt(), m_VPValue(A), m_One())) ||
       match(Def, m_DerivedIV(m_ZeroInt(), m_VPValue(A, m_ZeroInt()),
                              m_VPValue()))) &&
      A->getScalarType() == Def->getScalarType())
    return A;

  // Simplify MaskedCond with no block mask to its single operand.
  if (match(Def, m_VPInstruction<VPInstruction::MaskedCond>()) &&
      !cast<VPInstruction>(Def)->isMasked())
    return Def->getOperand(0);

  // Look through ExtractLastLane.
  if (match(Def, m_ExtractLastLane(m_VPValue(A)))) {
    if (match(A, m_BuildVector())) {
      auto *BuildVector = cast<VPInstruction>(A);
      return BuildVector->getOperand(BuildVector->getNumOperands() - 1);
    }

    if (match(A, m_Broadcast(m_VPValue(B))))
      return B;

    if (isa<VPInstruction, VPReplicateRecipe>(A) && vputils::isSingleScalar(A))
      return A;

    if (Plan.hasScalarVFOnly())
      return A;
  }

  // Look through ExtractPenultimateElement (BuildVector ....).
  if (match(Def, m_ExtractPenultimateElement(m_BuildVector()))) {
    auto *BuildVector = cast<VPInstruction>(Def->getOperand(0));
    return BuildVector->getOperand(BuildVector->getNumOperands() - 2);
  }

  uint64_t Idx;
  if (match(Def, m_ExtractElement(m_BuildVector(), m_ConstantInt(Idx)))) {
    auto *BuildVector = cast<VPInstruction>(Def->getOperand(0));
    return BuildVector->getOperand(Idx);
  }

  if (isa<VPPhi, VPWidenPHIRecipe, VPHeaderPHIRecipe>(Def)) {
    if (Def->getNumOperands() == 1) {
      return Def->getOperand(0);
    }
    if (auto *Phi = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(Def)) {
      if (all_equal(Phi->incoming_values()))
        return Phi->getOperand(0);
    }
    return nullptr;
  }

  VPIRValue *IRV;
  if (Def->getNumOperands() == 1 &&
      match(Def, m_ComputeReductionResult(m_VPIRValue(IRV))))
    return IRV;

  if (match(Def, m_VPInstruction<VPInstruction::WideIVStep>(m_VPValue(A),
                                                            m_One())) &&
      A->getScalarType() == Def->getScalarType())
    return A;

  // Some simplifications can only be applied after unrolling. Perform them
  // below.
  if (!Plan.isUnrolled())
    return nullptr;

  // Simplify extracts of the same single-scalar.
  if (match(Def, m_VPInstruction<VPInstruction::ExtractLane>()) &&
      all_equal(drop_begin(Def->operands())) &&
      vputils::isSingleScalar(Def->getOperand(1)))
    return Def->getOperand(1);

  // Replace extract-lane(0, canonical-WIDEN-INDUCTION) with the region's
  // scalar canonical IV.
  VPWidenIntOrFpInductionRecipe *WidenIV;
  if (match(Def, m_ExtractLane(m_ZeroInt(), m_CanonicalWidenIV(WidenIV))))
    return WidenIV->getRegion()->getCanonicalIV();

  // Simplify unrolled VectorPointer without offset, or with zero offset, to
  // just the pointer operand.
  if (auto *VPR = dyn_cast<VPVectorPointerRecipe>(Def))
    if (!VPR->getVFxPart() || match(VPR->getVFxPart(), m_ZeroInt()))
      return VPR->getOperand(0);

  // VPScalarIVSteps after unrolling can be replaced by their start value, if
  // the start index is zero and only the first lane 0 is demanded.
  if (auto *Steps = dyn_cast<VPScalarIVStepsRecipe>(Def))
    if (!Steps->getStartIndex() && vputils::onlyFirstLaneUsed(Steps))
      return Steps->getOperand(0);

  if (Plan.getConcreteUF() == 1 && match(Def, m_ExtractLastPart(m_VPValue(A))))
    return A;

  return nullptr;
}

/// Returns true if \p V is available at the end of \p VPBB, i.e. it either is a
/// live-in from the original IR or defined in \p VPBB.
static bool isAvailableAtEndOf(VPValue *V, const VPBasicBlock *VPBB) {
  VPRecipeBase *DefR = V->getDefiningRecipe();
  return DefR ? DefR->getParent() == VPBB : isa<VPIRValue>(V);
}

namespace {
/// Inserter for VPBuilderBase which appends all created VPSingleDefRecipes to a
/// worklist, so they get combined as well.
struct VPCombineInserter {
  SmallVectorImpl<VPSingleDefRecipe *> &Worklist;

  void insertHelper(VPRecipeBase *R, VPBasicBlock *VPBB,
                    VPBasicBlock::iterator It) {
    VPBB->insert(R, It);
    if (auto *Def = dyn_cast<VPSingleDefRecipe>(R))
      Worklist.push_back(Def);
  }
};

using VPCombineBuilder = VPBuilderBase<VPCombineInserter>;
} // namespace

/// Combine \p Def into a simpler recipe. May modify or create new recipes via
/// \p Builder.
static VPSingleDefRecipe *combineRecipe(VPlan &Plan, VPSingleDefRecipe *Def,
                                        VPCombineBuilder &Builder) {
  if (auto *V = simplifyRecipe(Plan, Def)) {
    Def->replaceAllUsesWith(V);
    return Def;
  }

  // Drop the mask of a predicated store masked by the header mask (which is
  // guaranteed to be true at least for the first lane) and both the stored
  // value and the address are uniform across VF and UF. The header mask is
  // still the abstract region value here.
  if (auto *RepR = dyn_cast<VPReplicateRecipe>(Def);
      RepR && RepR->isPredicated() && RepR->getOpcode() == Instruction::Store &&
      all_of(RepR->operandsWithoutMask(), vputils::isUniformAcrossVFsAndUFs) &&
      match(RepR->getMask(), m_HeaderMask())) {
    auto *Unmasked = new VPReplicateRecipe(
        RepR->getUnderlyingInstr(), RepR->operandsWithoutMask(),
        RepR->isSingleScalar(), /*Mask=*/nullptr, *RepR, *RepR,
        RepR->getDebugLoc());
    Builder.insert(Unmasked);
    return Unmasked;
  }

  // Avoid replacing VPInstructions with underlying values with new
  // VPInstructions, as we would fail to create widen/replicate recpes from the
  // new VPInstructions without an underlying value, and miss out on some
  // transformations that only apply to widened/replicated recipes later, by
  // doing so.
  // TODO: We should also not replace non-VPInstructions like VPWidenRecipe with
  // VPInstructions without underlying values, as those will get skipped during
  // cost computation.
  bool CanCreateNewRecipe =
      !isa<VPInstruction>(Def) || !Def->getUnderlyingValue();

  VPValue *X, *Y, *Z;

  // X && (Y && X) -> X && Y
  if (CanCreateNewRecipe &&
      match(Def, m_LogicalAnd(m_VPValue(X),
                              m_LogicalAnd(m_VPValue(Y), m_Deferred(X)))))
    return Builder.createLogicalAnd(X, Y);

  // (X && Y) | (X && Z) -> X && (Y | Z)
  if (CanCreateNewRecipe &&
      match(Def, m_c_BinaryOr(m_LogicalAnd(m_VPValue(X), m_VPValue(Y)),
                              m_LogicalAnd(m_Deferred(X), m_VPValue(Z)))) &&
      // Simplify only if one of the operands has one use to avoid creating an
      // extra recipe.
      (!Def->getOperand(0)->hasMoreThanOneUniqueUser() ||
       !Def->getOperand(1)->hasMoreThanOneUniqueUser()))
    return Builder.createLogicalAnd(X, Builder.createOr(Y, Z));

  // (X && Y) | !X -> !X || Y
  if (CanCreateNewRecipe &&
      match(Def,
            m_c_BinaryOr(m_OneUse(m_LogicalAnd(m_VPValue(X), m_VPValue(Y))),
                         m_VPValue(Z, m_Not(m_Deferred(X))))))
    return Builder.createLogicalOr(Z, Y);

  // select C, false, true -> not C
  VPValue *C;
  if (CanCreateNewRecipe &&
      match(Def, m_Select(m_VPValue(C), m_False(), m_True())))
    return Builder.createNot(C);

  // select !C, X, Y -> select C, Y, X
  if (match(Def, m_Select(m_Not(m_VPValue(C)), m_VPValue(X), m_VPValue(Y)))) {
    Def->setOperand(0, C);
    Def->setOperand(1, Y);
    Def->setOperand(2, X);
    return Def;
  }

  // select X, (i1 Y | Z), Y -> Y | (X && Z)
  if (CanCreateNewRecipe &&
      match(Def, m_Select(m_VPValue(X),
                          m_OneUse(m_c_BinaryOr(m_VPValue(Y), m_VPValue(Z))),
                          m_Deferred(Y))) &&
      Y->getScalarType()->isIntegerTy(1))
    return Builder.createOr(Y, Builder.createLogicalAnd(X, Z));

  // select M0, (select M1, X, Y), Y -> select (M0 && M1), X, Y
  VPValue *Mask0, *Mask1;
  if (CanCreateNewRecipe &&
      match(Def,
            m_SelectLike(m_VPValue(Mask0),
                         m_OneUse(m_SelectLike(m_VPValue(Mask1), m_VPValue(X),
                                               m_VPValue(Y))),
                         m_Deferred(Y))))
    return Builder.createSelect(Builder.createLogicalAnd(Mask0, Mask1), X, Y,
                                Def->getDebugLoc());

  if (match(Def, m_Trunc(m_VPValue(Y, m_ZExtOrSExt(m_VPValue(X)))))) {
    // Don't replace a non-widened cast recipe with a widened cast.
    if (!isa<VPWidenCastRecipe>(Def))
      return nullptr;
    Type *TruncTy = Def->getScalarType();
    Type *XTy = X->getScalarType();
    if (XTy->getScalarSizeInBits() < TruncTy->getScalarSizeInBits()) {

      unsigned ExtOpcode =
          match(Y, m_SExt(m_VPValue())) ? Instruction::SExt : Instruction::ZExt;
      auto *Ext =
          Builder.createWidenCast(Instruction::CastOps(ExtOpcode), X, TruncTy);
      if (auto *UnderlyingExt = Y->getUnderlyingValue()) {
        // UnderlyingExt has distinct return type, used to retain legacy cost.
        Ext->setUnderlyingValue(UnderlyingExt);
      }
      return Ext;
    } else if (XTy->getScalarSizeInBits() > TruncTy->getScalarSizeInBits()) {
      auto *Trunc = Builder.createWidenCast(Instruction::Trunc, X, TruncTy);
      return Trunc;
    }
  }

  if (CanCreateNewRecipe && match(Def, m_c_Mul(m_VPValue(X), m_AllOnes()))) {
    // Preserve nsw from the Mul on the new Sub.
    VPIRFlags::WrapFlagsTy NW = {
        false, cast<VPRecipeWithIRFlags>(Def)->hasNoSignedWrap()};
    return Builder.createSub(Plan.getZero(X->getScalarType()), X,
                             Def->getDebugLoc(), "", NW);
  }

  if (CanCreateNewRecipe &&
      match(Def, m_c_Add(m_VPValue(X),
                         m_VPValue(Z, m_Sub(m_ZeroInt(), m_VPValue(Y)))))) {
    // Preserve nsw from the Add and the Sub, if it's present on both, on the
    // new Sub.
    VPIRFlags::WrapFlagsTy NW = {
        false, cast<VPRecipeWithIRFlags>(Def)->hasNoSignedWrap() &&
                   cast<VPRecipeWithIRFlags>(Z)->hasNoSignedWrap()};
    return Builder.createSub(X, Y, Def->getDebugLoc(), "", NW);
  }

  const APInt *APC;
  if (CanCreateNewRecipe && match(Def, m_URem(m_VPValue(X), m_APInt(APC))) &&
      APC->isPowerOf2())
    return Builder.createAnd(X, Plan.getConstantInt(*APC - 1),
                             Def->getDebugLoc());

  if (CanCreateNewRecipe && match(Def, m_c_Mul(m_VPValue(X), m_APInt(APC))) &&
      APC->isPowerOf2()) {
    auto *MulR = cast<VPRecipeWithIRFlags>(Def);
    unsigned ShiftAmt = APC->exactLogBase2();
    VPIRFlags::WrapFlagsTy NW(MulR->hasNoUnsignedWrap(),
                              MulR->hasNoSignedWrap() &&
                                  ShiftAmt != APC->getBitWidth() - 1);
    return Builder.createNaryOp(
        Instruction::Shl,
        {X, Plan.getConstantInt(APC->getBitWidth(), ShiftAmt)}, NW,
        Def->getDebugLoc());
  }

  if (CanCreateNewRecipe && match(Def, m_UDiv(m_VPValue(X), m_APInt(APC))) &&
      APC->isPowerOf2())
    return Builder.createNaryOp(
        Instruction::LShr,
        {X, Plan.getConstantInt(APC->getBitWidth(), APC->exactLogBase2())},
        *cast<VPRecipeWithIRFlags>(Def), Def->getDebugLoc());

  // (X >> C) << C -> X & (-1 << C).
  if (CanCreateNewRecipe &&
      match(Def, m_Shl(m_LShr(m_VPValue(X), m_VPValue(Y, m_APInt(APC))),
                       m_Deferred(Y))))
    return Builder.createAnd(
        X, Plan.getConstantInt(APInt::getAllOnes(APC->getBitWidth()) << *APC),
        Def->getDebugLoc());

  if (match(Def, m_Not(m_VPValue(X)))) {
    // Try to fold Not into compares by adjusting the predicate in-place.
    CmpPredicate Pred;
    if (match(X, m_Cmp(Pred, m_VPValue(), m_VPValue()))) {
      auto *Cmp = cast<VPRecipeWithIRFlags>(X);
      // Only fold if every user is a Not of the cmp, or a select using the cmp
      // solely as its condition.
      if (all_of(Cmp->users(), [Cmp](VPUser *U) {
            return match(U, m_Not(m_Specific(Cmp))) ||
                   (match(U, m_Select(m_Specific(Cmp), m_VPValue(),
                                      m_VPValue())) &&
                    U->getOperand(1) != Cmp && U->getOperand(2) != Cmp);
          })) {
        Cmp->setPredicate(CmpInst::getInversePredicate(Pred));
        for (VPUser *U : to_vector(Cmp->users())) {
          auto *R = cast<VPSingleDefRecipe>(U);
          if (match(R, m_Select(m_Specific(Cmp), m_VPValue(X), m_VPValue(Y)))) {
            // select (cmp pred), X, Y -> select (cmp inv_pred), Y, X
            R->setOperand(1, Y);
            R->setOperand(2, X);
          } else {
            // not (cmp pred) -> cmp inv_pred
            assert(match(R, m_Not(m_Specific(Cmp))) && "Unexpected user");
            R->replaceAllUsesWith(Cmp);
          }
        }
        // If Cmp doesn't have a debug location, use the one from the negation,
        // to preserve the location.
        if (!Cmp->getDebugLoc() && Def->getDebugLoc())
          Cmp->setDebugLoc(Def->getDebugLoc());
        return Def;
      }
    }
  }

  // Fold any-of (fcmp uno A, A), (fcmp uno B, B), ... ->
  //      any-of (fcmp uno A, B), ...
  if (match(Def, m_AnyOf())) {
    SmallVector<VPValue *, 4> NewOps;
    VPRecipeBase *UnpairedCmp = nullptr;
    for (VPValue *Op : Def->operands()) {
      VPValue *X;
      if (Op->getNumUsers() > 1 ||
          !match(Op, m_SpecificCmp(CmpInst::FCMP_UNO, m_VPValue(X),
                                   m_Deferred(X)))) {
        NewOps.push_back(Op);
      } else if (!UnpairedCmp) {
        UnpairedCmp = Op->getDefiningRecipe();
      } else {
        NewOps.push_back(Builder.createFCmp(CmpInst::FCMP_UNO,
                                            UnpairedCmp->getOperand(0), X));
        UnpairedCmp = nullptr;
      }
    }

    if (UnpairedCmp)
      NewOps.push_back(UnpairedCmp->getVPSingleValue());

    if (NewOps.size() < Def->getNumOperands())
      return Builder.createNaryOp(VPInstruction::AnyOf, NewOps);
  }

  // Fold (fcmp uno X, X) | (fcmp uno Y, Y) -> fcmp uno X, Y
  // This is useful for fmax/fmin without fast-math flags, where we need to
  // check if any operand is NaN.
  if (CanCreateNewRecipe &&
      match(Def,
            m_BinaryOr(
                m_SpecificCmp(CmpInst::FCMP_UNO, m_VPValue(X), m_Deferred(X)),
                m_SpecificCmp(CmpInst::FCMP_UNO, m_VPValue(Y), m_Deferred(Y)))))
    return Builder.createFCmp(CmpInst::FCMP_UNO, X, Y);

  if (match(Def, m_VPInstruction<VPInstruction::WideIVStep>(m_VPValue(X),
                                                            m_One())) &&
      X->getScalarType() != Def->getScalarType())
    return Builder.createWidenCast(Instruction::Trunc, X, Def->getScalarType());

  // For i1 vp.merges produced by AnyOf reductions:
  // vp.merge true, (or X, Y), X, evl -> vp.merge Y, true, X, evl
  if (match(Def, m_Intrinsic<Intrinsic::vp_merge>(m_True(), m_VPValue(X),
                                                  m_VPValue(X), m_VPValue())) &&
      match(X, m_c_BinaryOr(m_Specific(X), m_VPValue(Y))) &&
      Def->getScalarType()->isIntegerTy(1)) {
    Def->setOperand(1, Plan.getTrue());
    Def->setOperand(0, Y);
    return Def;
  }

  if (match(Def, m_BuildVector()) && all_equal(Def->operands()))
    return Builder.createNaryOp(VPInstruction::Broadcast, Def->getOperand(0));

  // Replace uses of a BuildVector by users that only use its first lane with
  // its first operand directly.
  if (match(Def, m_BuildVector())) {
    Def->replaceUsesWithIf(Def->getOperand(0), [Def](VPUser &U, unsigned) {
      return U.usesFirstLaneOnly(Def);
    });
    return Def;
  }

  // Look through broadcast of single-scalar when used as select conditions; in
  // that case the scalar condition can be used directly.
  if (match(Def,
            m_Select(m_Broadcast(m_VPValue(Z)), m_VPValue(), m_VPValue()))) {
    assert(vputils::isSingleScalar(Z) &&
           "broadcast operand must be single-scalar");
    Def->setOperand(0, Z);
    return Def;
  }

  if (match(Def, m_Broadcast(m_VPValue(X)))) {
    Def->replaceUsesWithIf(
        X, [Def](const VPUser &U, unsigned) { return U.usesScalars(Def); });
    return Def;
  }

  // Some simplifications can only be applied after unrolling. Perform them
  // below.
  if (!Plan.isUnrolled())
    return nullptr;

  // Simplify extract-lane with single source to extract-element.
  VPValue *LaneToExtract;
  if (match(Def, m_ExtractLane(m_VPValue(LaneToExtract), m_VPValue(X))))
    return Builder.createNaryOp(Instruction::ExtractElement, {X, LaneToExtract},
                                Def->getDebugLoc());

  // Look for cycles where Def is of the form:
  //  X = phi(0, IVInc)  ; used only by IVInc, or by IVInc and Inc = X + Y
  //  IVInc = X + Step   ; used by X and Def
  //  Def = IVInc + Y
  // Fold the increment Y into the phi's start value, replace Def with IVInc,
  // and if Inc exists, replace it with X.
  VPValue *IVInc;
  if (match(Def, m_Add(m_VPValue(IVInc, m_Add(m_VPValue(X), m_VPValue())),
                       m_VPValue(Y))) &&
      match(X, m_VPPhi(m_ZeroInt(), m_Specific(IVInc))) &&
      IVInc->getNumUsers() == 2) {
    auto *Phi = cast<VPPhi>(X);
    // If Phi has a second user (besides IVInc's defining recipe), it must be
    // Inc = Phi + Y for the fold to apply.
    auto *Inc = dyn_cast_if_present<VPSingleDefRecipe>(
        findUserOf(Phi, m_Add(m_Specific(Phi), m_Specific(Y))));
    if ((Phi->getNumUsers() == 1 || (Phi->getNumUsers() == 2 && Inc)) &&
        isAvailableAtEndOf(Y, Phi->getIncomingBlock(0))) {
      Def->replaceAllUsesWith(IVInc);
      if (Inc)
        Inc->replaceAllUsesWith(Phi);
      Phi->setOperand(0, Y);
      return Def;
    }
  }

  // Simplify redundant ReductionStartVector recipes after unrolling.
  VPValue *StartV;
  if (match(Def, m_VPInstruction<VPInstruction::ReductionStartVector>(
                     m_VPValue(StartV), m_VPValue(), m_VPValue()))) {
    Def->replaceUsesWithIf(StartV, [](const VPUser &U, unsigned Idx) {
      auto *PhiR = dyn_cast<VPReductionPHIRecipe>(&U);
      return PhiR && PhiR->isInLoop();
    });
    return Def;
  }

  return nullptr;
}

void VPlanTransforms::combineRecipes(VPlan &Plan) {
  SmallVector<VPSingleDefRecipe *, 256> Worklist;
  PostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> POT(
      Plan.getEntry());
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(POT))
    for (VPSingleDefRecipe &Def :
         make_isa_range<VPSingleDefRecipe>(reverse(*VPBB)))
      Worklist.push_back(&Def);

  [[maybe_unused]] unsigned InitWorklistSize = Worklist.size();

  VPCombineBuilder Builder({Worklist});
  while (!Worklist.empty()) {
    assert(Worklist.size() < InitWorklistSize * 2 &&
           "Worklist is growing large, possible cycle?");
    VPSingleDefRecipe *Def = Worklist.pop_back_val();
    Builder.setInsertPoint(Def);
    VPSingleDefRecipe *New = combineRecipe(Plan, Def, Builder);
    if (!New)
      continue;
    if (New != Def) {
      // Replace the recipe with a new one.
      Def->replaceAllUsesWith(New);
      Def->eraseFromParent();
      // TODO: Append users to the worklist (might need a setvector)
    } else if (vputils::isDeadRecipe(*Def)) {
      // Recipe was modified - it may be dead now.
      Def->eraseFromParent();
    }
  }
}

void VPlanTransforms::simplifyReverses(VPlan &Plan) {
  // Pull out reverses from any elementwise op.
  // binop(reverse(x), reverse(y)) -> reverse(binop(x,y))
  vputils::pullOutPermutations(
      Plan, [](VPValue *&X) { return m_Reverse(m_VPValue(X)); },
      [](auto *X) { return new VPInstruction(VPInstruction::Reverse, X); });

  // reverse(reverse(x)) -> x
  VPValue *X;
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_deep(Plan.getEntry())))
    for (VPRecipeBase &R : make_early_inc_range(*VPBB))
      if (match(&R, m_Reverse(m_Reverse(m_VPValue(X)))))
        R.getVPSingleValue()->replaceAllUsesWith(X);
}

/// Reassociate (headermask && x) && y -> headermask && (x && y) to allow the
/// header mask to be simplified further when tail folding, e.g. in
/// optimizeEVLMasks.
static void reassociateHeaderMask(VPlan &Plan) {
  VPValue *HeaderMask = Plan.getVectorLoopRegion()->getHeaderMask();
  if (!HeaderMask)
    return;

  SmallVector<VPUser *> Worklist;
  for (VPUser *U : HeaderMask->users())
    if (match(U, m_LogicalAnd(m_Specific(HeaderMask), m_VPValue())))
      append_range(Worklist, cast<VPSingleDefRecipe>(U)->users());

  while (!Worklist.empty()) {
    auto *R = dyn_cast<VPSingleDefRecipe>(Worklist.pop_back_val());
    VPValue *X, *Y;
    if (!R || !match(R, m_LogicalAnd(
                            m_LogicalAnd(m_Specific(HeaderMask), m_VPValue(X)),
                            m_VPValue(Y))))
      continue;
    append_range(Worklist, R->users());
    VPBuilder Builder(R);
    R->replaceAllUsesWith(
        Builder.createLogicalAnd(HeaderMask, Builder.createLogicalAnd(X, Y)));
  }
}

static std::optional<Instruction::BinaryOps>
getUnmaskedDivRemOpcode(Intrinsic::ID ID) {
  switch (ID) {
  case Intrinsic::masked_udiv:
    return Instruction::UDiv;
  case Intrinsic::masked_sdiv:
    return Instruction::SDiv;
  case Intrinsic::masked_urem:
    return Instruction::URem;
  case Intrinsic::masked_srem:
    return Instruction::SRem;
  default:
    return {};
  }
}

static void narrowToSingleScalarRecipes(VPlan &Plan) {
  if (Plan.hasScalarVFOnly())
    return;

  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_deep(Plan.getEntry()))) {
    for (VPRecipeBase &R : make_early_inc_range(reverse(*VPBB))) {
      if (!isa<VPWidenRecipe, VPWidenGEPRecipe, VPReplicateRecipe,
               VPWidenIntrinsicRecipe>(&R))
        continue;
      auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
      if (RepR && (RepR->isSingleScalar() || RepR->isPredicated()))
        continue;

      auto *RepOrWidenR = cast<VPRecipeWithIRFlags>(&R);
      if (RepR && RepR->getOpcode() == Instruction::Store &&
          vputils::isSingleScalar(RepR->getOperand(1))) {
        auto *Clone = new VPReplicateRecipe(
            RepOrWidenR->getUnderlyingInstr(), RepOrWidenR->operands(),
            true /*IsSingleScalar*/, nullptr /*Mask*/, *RepR /*Flags*/,
            *RepR /*Metadata*/, RepR->getDebugLoc());
        Clone->insertBefore(RepOrWidenR);
        VPBuilder Builder(Clone);
        VPValue *ExtractOp = Clone->getOperand(0);
        if (vputils::isUniformAcrossVFsAndUFs(RepR->getOperand(1)))
          ExtractOp =
              Builder.createNaryOp(VPInstruction::ExtractLastPart, ExtractOp);
        ExtractOp =
            Builder.createNaryOp(VPInstruction::ExtractLastLane, ExtractOp);
        Clone->setOperand(0, ExtractOp);
        RepR->eraseFromParent();
        continue;
      }

      // Narrow llvm.masked.{u,s}{div,rem} intrinsics with a safe divisor.
      if (auto *IntrR = dyn_cast<VPWidenIntrinsicRecipe>(RepOrWidenR)) {
        if (!vputils::onlyFirstLaneUsed(IntrR))
          continue;
        auto Opc = getUnmaskedDivRemOpcode(IntrR->getVectorIntrinsicID());
        if (!Opc)
          continue;
        VPBuilder Builder(IntrR);
        VPValue *SafeDivisor = Builder.createSelect(
            IntrR->getOperand(2), IntrR->getOperand(1),
            Plan.getConstantInt(IntrR->getScalarType(), 1));
        VPValue *Clone = Builder.createNaryOp(
            *Opc, {IntrR->getOperand(0), SafeDivisor},
            VPIRFlags::getDefaultFlags(*Opc), IntrR->getDebugLoc());
        IntrR->replaceAllUsesWith(Clone);
        IntrR->eraseFromParent();
        continue;
      }

      // Skip recipes that aren't single scalars.
      if (!vputils::isSingleScalar(RepOrWidenR))
        continue;

      // Predicate to check if a user of Op introduces extra broadcasts.
      auto IntroducesBCastOf = [](const VPValue *Op) {
        return [Op](const VPUser *U) {
          if (auto *VPI = dyn_cast<VPInstruction>(U)) {
            if (is_contained({VPInstruction::ExtractLastLane,
                              VPInstruction::ExtractLastPart,
                              VPInstruction::ExtractPenultimateElement},
                             VPI->getOpcode()))
              return false;
          }
          return !U->usesScalars(Op);
        };
      };

      if (any_of(RepOrWidenR->users(), IntroducesBCastOf(RepOrWidenR)) &&
          none_of(RepOrWidenR->operands(), [&](VPValue *Op) {
            if (any_of(
                    make_filter_range(Op->users(), not_equal_to(RepOrWidenR)),
                    IntroducesBCastOf(Op)))
              return false;
            // Non-constant live-ins require broadcasts, while constants do not
            // need explicit broadcasts.
            bool LiveInNeedsBroadcast =
                isa<VPIRValue>(Op) && !isa<VPConstant>(Op);
            auto *OpR = dyn_cast<VPReplicateRecipe>(Op);
            return LiveInNeedsBroadcast || (OpR && OpR->isSingleScalar());
          }))
        continue;

      auto *Clone = VPBuilder::createSingleScalarOp(
          vputils::getOpcode(RepOrWidenR), RepOrWidenR->operands(),
          /*Mask=*/nullptr, *RepOrWidenR, getMetadataOf(RepOrWidenR),
          DebugLoc::getUnknown(), RepOrWidenR->getScalarType(),
          RepOrWidenR->getUnderlyingInstr());
      Clone->insertBefore(RepOrWidenR);
      RepOrWidenR->replaceAllUsesWith(Clone);
      if (vputils::isDeadRecipe(*RepOrWidenR))
        RepOrWidenR->eraseFromParent();
    }
  }
}

/// Try to see if all of \p Blend's masks share a common value logically and'ed
/// and remove it from the masks.
static void removeCommonBlendMask(VPBlendRecipe *Blend) {
  if (Blend->isNormalized())
    return;
  VPValue *CommonEdgeMask;
  if (!match(Blend->getMask(0),
             m_LogicalAnd(m_VPValue(CommonEdgeMask), m_VPValue())))
    return;
  for (unsigned I = 0; I < Blend->getNumIncomingValues(); I++)
    if (!match(Blend->getMask(I),
               m_LogicalAnd(m_Specific(CommonEdgeMask), m_VPValue())))
      return;
  for (unsigned I = 0; I < Blend->getNumIncomingValues(); I++)
    Blend->setMask(I, Blend->getMask(I)->getDefiningRecipe()->getOperand(1));
}

/// Normalize and simplify VPBlendRecipes. Should be run after combineRecipes
/// to make sure the masks are simplified.
static void simplifyBlends(VPlan &Plan) {
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) {
    for (VPBlendRecipe &Blend :
         make_early_inc_range(make_isa_range<VPBlendRecipe>(*VPBB))) {
      removeCommonBlendMask(&Blend);

      // Try to remove redundant blend recipes.
      SmallPtrSet<VPValue *, 4> UniqueValues;
      if (Blend.isNormalized() || !match(Blend.getMask(0), m_False()))
        UniqueValues.insert(Blend.getIncomingValue(0));
      for (unsigned I = 1; I != Blend.getNumIncomingValues(); ++I)
        if (!match(Blend.getMask(I), m_False()))
          UniqueValues.insert(Blend.getIncomingValue(I));

      if (UniqueValues.size() == 1) {
        Blend.replaceAllUsesWith(*UniqueValues.begin());
        Blend.eraseFromParent();
        continue;
      }

      if (Blend.isNormalized())
        continue;

      // Normalize the blend so its first incoming value is used as the initial
      // value with the others blended into it.

      unsigned StartIndex = 0;
      for (unsigned I = 0; I != Blend.getNumIncomingValues(); ++I) {
        // If a value's mask is used only by the blend then is can be deadcoded.
        // TODO: Find the most expensive mask that can be deadcoded, or a mask
        // that's used by multiple blends where it can be removed from them all.
        VPValue *Mask = Blend.getMask(I);
        if (Mask->hasOneUse() && !match(Mask, m_False())) {
          StartIndex = I;
          break;
        }
      }

      SmallVector<VPValue *, 4> OperandsWithMask;
      OperandsWithMask.push_back(Blend.getIncomingValue(StartIndex));

      for (unsigned I = 0; I != Blend.getNumIncomingValues(); ++I) {
        if (I == StartIndex)
          continue;
        OperandsWithMask.push_back(Blend.getIncomingValue(I));
        OperandsWithMask.push_back(Blend.getMask(I));
      }

      auto *NewBlend =
          new VPBlendRecipe(cast_or_null<PHINode>(Blend.getUnderlyingValue()),
                            OperandsWithMask, Blend, Blend.getDebugLoc());
      NewBlend->insertBefore(&Blend);

      VPValue *DeadMask = Blend.getMask(StartIndex);
      Blend.replaceAllUsesWith(NewBlend);
      Blend.eraseFromParent();
      vputils::recursivelyDeleteDeadRecipes(DeadMask);

      /// Simplify BLEND %a, %b, Not(%mask) -> BLEND %b, %a, %mask.
      VPValue *NewMask;
      if (NewBlend->getNumOperands() == 3 &&
          match(NewBlend->getMask(1), m_Not(m_VPValue(NewMask)))) {
        VPValue *Inc0 = NewBlend->getOperand(0);
        VPValue *Inc1 = NewBlend->getOperand(1);
        VPValue *OldMask = NewBlend->getOperand(2);
        NewBlend->setOperand(0, Inc1);
        NewBlend->setOperand(1, Inc0);
        NewBlend->setOperand(2, NewMask);
        if (OldMask->user_empty())
          cast<VPInstruction>(OldMask)->eraseFromParent();
      }
    }
  }
}

/// Optimize the width of vector induction variables in \p Plan based on a known
/// constant Trip Count, \p BestVF and \p BestUF.
static bool optimizeVectorInductionWidthForTCAndVFUF(VPlan &Plan,
                                                     ElementCount BestVF,
                                                     unsigned BestUF) {
  // Only proceed if we have not completely removed the vector region.
  if (!Plan.getVectorLoopRegion())
    return false;

  const APInt *TC;
  if (!BestVF.isFixed() || !match(Plan.getTripCount(), m_APInt(TC)))
    return false;

  // Calculate the minimum power-of-2 bit width that can fit the known TC, VF
  // and UF. Returns at least 8.
  auto ComputeBitWidth = [](APInt TC, uint64_t Align) {
    APInt AlignedTC =
        Align * APIntOps::RoundingUDiv(TC, APInt(TC.getBitWidth(), Align),
                                       APInt::Rounding::UP);
    APInt MaxVal = AlignedTC - 1;
    return std::max<unsigned>(PowerOf2Ceil(MaxVal.getActiveBits()), 8);
  };
  unsigned NewBitWidth =
      ComputeBitWidth(*TC, BestVF.getKnownMinValue() * BestUF);

  LLVMContext &Ctx = Plan.getContext();
  auto *NewIVTy = IntegerType::get(Ctx, NewBitWidth);

  bool MadeChange = false;

  VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
  for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
    // Currently only handle canonical IVs as it is trivial to replace the start
    // and stop values, and we currently only perform the optimization when the
    // IV has a single use.
    VPWidenIntOrFpInductionRecipe *WideIV;
    if (!match(&Phi, m_CanonicalWidenIV(WideIV)))
      continue;
    if (WideIV->hasMoreThanOneUniqueUser() ||
        NewIVTy == WideIV->getScalarType())
      continue;

    // Currently only handle cases where the single user is a header-mask
    // comparison with the backedge-taken-count.
    VPUser *SingleUser = WideIV->getSingleUser();
    if (!SingleUser ||
        !match(SingleUser,
               m_ICmp(m_Specific(WideIV),
                      m_Broadcast(m_Specific(Plan.getBackedgeTakenCount())))))
      continue;

    // Update IV operands and comparison bound to use new narrower type.
    assert(!WideIV->getTruncInst() &&
           "canonical IV is not expected to have a truncation");
    auto *NewWideIV = new VPWidenIntOrFpInductionRecipe(
        WideIV->getPHINode(), Plan.getZero(NewIVTy),
        Plan.getConstantInt(NewIVTy, 1), WideIV->getVFValue(),
        WideIV->getInductionDescriptor(), *WideIV, WideIV->getDebugLoc());
    NewWideIV->insertBefore(WideIV);

    auto *NewBTC = new VPWidenCastRecipe(
        Instruction::Trunc, Plan.getOrCreateBackedgeTakenCount(), NewIVTy,
        nullptr, VPIRFlags::getDefaultFlags(Instruction::Trunc));
    Plan.getVectorPreheader()->appendRecipe(NewBTC);
    auto *Cmp = cast<VPInstruction>(WideIV->getSingleUser());
    Cmp->replaceAllUsesWith(
        VPBuilder(Cmp).createICmp(Cmp->getPredicate(), NewWideIV, NewBTC));

    MadeChange = true;
  }

  return MadeChange;
}

/// Return true if \p Cond is known to be true for given \p BestVF and \p
/// BestUF.
static bool isConditionTrueViaVFAndUF(VPValue *Cond, VPlan &Plan,
                                      ElementCount BestVF, unsigned BestUF,
                                      PredicatedScalarEvolution &PSE) {
  if (match(Cond, m_BinaryOr(m_VPValue(), m_VPValue())))
    return any_of(Cond->getDefiningRecipe()->operands(), [&Plan, BestVF, BestUF,
                                                          &PSE](VPValue *C) {
      return isConditionTrueViaVFAndUF(C, Plan, BestVF, BestUF, PSE);
    });

  auto *CanIV = Plan.getVectorLoopRegion()->getCanonicalIV();
  if (!match(Cond, m_SpecificICmp(
                       CmpInst::ICMP_EQ,
                       m_c_Add(m_Specific(CanIV), m_Specific(&Plan.getVFxUF())),
                       m_Specific(&Plan.getVectorTripCount()))))
    return false;

  // The compare checks CanIV + VFxUF == vector trip count. The vector trip
  // count is not conveniently available as SCEV so far, so we compare directly
  // against the original trip count. This is stricter than necessary, as we
  // will only return true if the trip count == vector trip count.
  const SCEV *VectorTripCount =
      vputils::getSCEVExprForVPValue(&Plan.getVectorTripCount(), PSE);
  if (isa<SCEVCouldNotCompute>(VectorTripCount))
    VectorTripCount = vputils::getSCEVExprForVPValue(Plan.getTripCount(), PSE);
  assert(!isa<SCEVCouldNotCompute>(VectorTripCount) &&
         "Trip count SCEV must be computable");
  ScalarEvolution &SE = *PSE.getSE();
  ElementCount NumElements = BestVF * BestUF;
  const SCEV *C = SE.getElementCount(VectorTripCount->getType(), NumElements);
  return SE.isKnownPredicate(CmpInst::ICMP_EQ, VectorTripCount, C);
}

// Replaces ExtractVectorForPart instructions with ICMP when the VF is scalar
// and the source is a WideActiveLaneMask. The unused mask is removed later
// when removing dead recipes.
static bool replaceMaskWithCompareForScalarPlan(VPlan &Plan,
                                                ElementCount BestVF) {
  if (!BestVF.isScalar())
    return false;

  bool MadeChange = false;
  VPBuilder Builder;
  VPRegionBlock *VectorRegion = Plan.getVectorLoopRegion();
  VPBasicBlock *PreheaderVPBB = Plan.getVectorPreheader();
  VPBasicBlock *ExitingVPBB = VectorRegion->getExitingBasicBlock();

  VPValue *Start, *TC;
  uint64_t Idx;
  for (VPBasicBlock *VPBB : {PreheaderVPBB, ExitingVPBB}) {
    for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
      if (!match(&R, m_ExtractVectorForPart(
                         m_WideActiveLaneMask(m_VPValue(Start), m_VPValue(TC),
                                              m_VPValue()),
                         m_ConstantInt(Idx))))
        continue;

      auto *Extract = cast<VPInstruction>(&R);
      Builder.setInsertPoint(Extract);

      if (Idx > 0)
        Start = Builder.createAdd(
            Start, Plan.getConstantInt(Start->getScalarType(), Idx));

      VPValue *ICmp = Builder.createICmp(CmpInst::ICMP_ULT, Start, TC);
      Extract->replaceAllUsesWith(ICmp);
      Extract->eraseFromParent();
      MadeChange = true;
    }
  }

  return MadeChange;
}

/// Try to simplify the branch condition of \p Plan. This may restrict the
/// resulting plan to \p BestVF and \p BestUF.
static bool simplifyBranchConditionForVFAndUF(VPlan &Plan, ElementCount BestVF,
                                              unsigned BestUF,
                                              PredicatedScalarEvolution &PSE) {
  VPRegionBlock *VectorRegion = Plan.getVectorLoopRegion();
  VPBasicBlock *ExitingVPBB = VectorRegion->getExitingBasicBlock();
  auto *Term = &ExitingVPBB->back();
  VPValue *Cond;
  VPValue *Offset = nullptr;
  auto m_CanIVInc = m_Add(m_VPValue(), m_Specific(&Plan.getVFxUF()));
  // Check if the branch condition compares the canonical IV increment (for main
  // loop), or the canonical IV increment plus an offset (for epilog loop).
  bool MatchedCanIVInc =
      match(Term,
            m_BranchOnCount(
                m_CombineOr(m_CanIVInc, m_c_Add(m_CanIVInc, m_VPValue(Offset))),
                m_VPValue())) &&
      (!Offset || Offset->isDefinedOutsideLoopRegions());
  if (MatchedCanIVInc ||
      match(Term,
            m_BranchOnCond(m_Not(m_ExtractVectorForPart(
                m_WideActiveLaneMask(m_VPValue(), m_VPValue(), m_VPValue()),
                m_ZeroInt()))))) {
    // Try to simplify the branch condition if VectorTC <= VF * UF when the
    // latch terminator is BranchOnCount or
    // BranchOnCond(Not(ExtractVectorForPart(WideActiveLaneMask), 0))
    const SCEV *VectorTripCount =
        vputils::getSCEVExprForVPValue(&Plan.getVectorTripCount(), PSE);
    if (isa<SCEVCouldNotCompute>(VectorTripCount))
      VectorTripCount =
          vputils::getSCEVExprForVPValue(Plan.getTripCount(), PSE);
    assert(!isa<SCEVCouldNotCompute>(VectorTripCount) &&
           "Trip count SCEV must be computable");
    ScalarEvolution &SE = *PSE.getSE();
    ElementCount NumElements = BestVF * BestUF;
    const SCEV *C = SE.getElementCount(VectorTripCount->getType(), NumElements);
    if (!SE.isKnownPredicate(CmpInst::ICMP_ULE, VectorTripCount, C))
      return false;
  } else if (match(Term, m_BranchOnCond(m_VPValue(Cond))) ||
             match(Term, m_BranchOnTwoConds(m_VPValue(), m_VPValue(Cond)))) {
    // For BranchOnCond, check if we can prove the condition to be true using VF
    // and UF.
    if (!isConditionTrueViaVFAndUF(Cond, Plan, BestVF, BestUF, PSE))
      return false;
  } else {
    return false;
  }

  // The vector loop region only executes once. Convert terminator of the
  // exiting block to exit in the first iteration.
  if (match(Term, m_BranchOnTwoConds())) {
    Term->setOperand(1, Plan.getTrue());
    return true;
  }

  auto *BOC = new VPInstruction(VPInstruction::BranchOnCond, Plan.getTrue(), {},
                                {}, Term->getDebugLoc());
  ExitingVPBB->appendRecipe(BOC);
  Term->eraseFromParent();

  return true;
}

void VPlanTransforms::optimizeForVFAndUF(VPlan &Plan, ElementCount BestVF,
                                         unsigned BestUF,
                                         PredicatedScalarEvolution &PSE) {
  assert(Plan.hasVF(BestVF) && "BestVF is not available in Plan");
  assert(Plan.hasUF(BestUF) && "BestUF is not available in Plan");

  bool MadeChange =
      simplifyBranchConditionForVFAndUF(Plan, BestVF, BestUF, PSE);
  MadeChange |= replaceMaskWithCompareForScalarPlan(Plan, BestVF);
  MadeChange |= optimizeVectorInductionWidthForTCAndVFUF(Plan, BestVF, BestUF);

  if (MadeChange) {
    Plan.setVF(BestVF);
    assert(Plan.getConcreteUF() == BestUF && "BestUF must match the Plan's UF");
  }
}

void VPlanTransforms::clearReductionWrapFlags(VPlan &Plan) {
  for (VPReductionPHIRecipe &PhiR : make_isa_range<VPReductionPHIRecipe>(
           Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis())) {
    RecurKind RK = PhiR.getRecurrenceKind();
    if (RK != RecurKind::Add && RK != RecurKind::Mul && RK != RecurKind::Sub &&
        RK != RecurKind::AddChainWithSubs)
      continue;

    for (VPUser *U : vputils::collectUsersRecursively(&PhiR))
      if (auto *RecWithFlags = dyn_cast<VPRecipeWithIRFlags>(U)) {
        RecWithFlags->dropPoisonGeneratingFlags();
      }
  }
}

namespace {
struct VPCSEDenseMapInfo : public DenseMapInfo<VPSingleDefRecipe *> {
  /// If recipe \p R will lower to a GEP with a non-i8 source element type,
  /// return that source element type.
  static Type *getGEPSourceElementType(const VPSingleDefRecipe *R) {
    // All VPInstructions that lower to GEPs must have the i8 source element
    // type (as they are PtrAdds), so we omit it.
    return TypeSwitch<const VPSingleDefRecipe *, Type *>(R)
        .Case([](const VPReplicateRecipe *I) -> Type * {
          if (auto *GEP = dyn_cast<GetElementPtrInst>(I->getUnderlyingValue()))
            return GEP->getSourceElementType();
          return nullptr;
        })
        .Case<VPVectorPointerRecipe, VPWidenGEPRecipe>(
            [](auto *I) { return I->getSourceElementType(); })
        .Default([](auto *) { return nullptr; });
  }

  /// Returns true if recipe \p Def can be safely handed for CSE.
  static bool canHandle(const VPSingleDefRecipe *Def) {
    // We can extend the list of handled recipes in the future,
    // provided we account for the data embedded in them while checking for
    // equality or hashing.
    auto C = vputils::getOpcodeOrIntrinsicID(Def);

    // The issue with (Insert|Extract)Value is that the index of the
    // insert/extract is not a proper operand in LLVM IR, and hence also not in
    // VPlan.
    if (!C || (!C->first && (C->second == Instruction::InsertValue ||
                             C->second == Instruction::ExtractValue)))
      return false;

    // Widened loads (including the EVL variant) are handled, as cse() only
    // reuses them within a block with no intervening memory write. Any other
    // memory access is rejected.
    if (Def->mayWriteToMemory())
      return false;
    return !Def->mayReadFromMemory() ||
           isa<VPWidenLoadRecipe, VPWidenLoadEVLRecipe>(Def);
  }

  /// Hash the underlying data of \p Def.
  static unsigned getHashValue(const VPSingleDefRecipe *Def) {
    hash_code Result = hash_combine(
        Def->getVPRecipeID(), vputils::getOpcodeOrIntrinsicID(Def),
        getGEPSourceElementType(Def), Def->getScalarType(),
        vputils::isSingleScalar(Def), hash_combine_range(Def->operands()));
    if (auto *RFlags = dyn_cast<VPRecipeWithIRFlags>(Def))
      if (RFlags->hasPredicate())
        return hash_combine(Result, RFlags->getPredicate());
    if (auto *SIVSteps = dyn_cast<VPScalarIVStepsRecipe>(Def))
      return hash_combine(Result, SIVSteps->getInductionOpcode());
    // Fold in the separately stored consecutive flag. Alignment is left out and
    // handled by cse.
    if (auto *Load = dyn_cast<VPWidenMemoryRecipe>(Def))
      return hash_combine(Result, Load->isConsecutive());
    return Result;
  }

  /// Check equality of underlying data of \p L and \p R.
  static bool isEqual(const VPSingleDefRecipe *L, const VPSingleDefRecipe *R) {
    if (L->getVPRecipeID() != R->getVPRecipeID() ||
        vputils::getOpcodeOrIntrinsicID(L) !=
            vputils::getOpcodeOrIntrinsicID(R) ||
        getGEPSourceElementType(L) != getGEPSourceElementType(R) ||
        vputils::isSingleScalar(L) != vputils::isSingleScalar(R) ||
        !equal(L->operands(), R->operands()))
      return false;
    assert(vputils::getOpcodeOrIntrinsicID(L) &&
           vputils::getOpcodeOrIntrinsicID(R) &&
           "must have valid opcode info for both recipes");
    if (auto *LFlags = dyn_cast<VPRecipeWithIRFlags>(L))
      if (LFlags->hasPredicate() &&
          LFlags->getPredicate() !=
              cast<VPRecipeWithIRFlags>(R)->getPredicate())
        return false;
    if (auto *LSIV = dyn_cast<VPScalarIVStepsRecipe>(L))
      if (LSIV->getInductionOpcode() !=
          cast<VPScalarIVStepsRecipe>(R)->getInductionOpcode())
        return false;
    // Compare the separately stored consecutive flag. Alignment is left out and
    // handled by cse.
    if (auto *LL = dyn_cast<VPWidenMemoryRecipe>(L))
      if (LL->isConsecutive() != cast<VPWidenMemoryRecipe>(R)->isConsecutive())
        return false;
    // Phi recipes can only be equal if they are in the same VPBB, as they
    // implicitly depend on their predecessors.
    if (isa<VPWidenPHIRecipe>(L) && L->getParent() != R->getParent())
      return false;
    // Recipes in replicate regions implicitly depend on predicate. If either
    // recipe is in a replicate region, only consider them equal if both have
    // the same parent.
    const VPRegionBlock *RegionL = L->getRegion();
    const VPRegionBlock *RegionR = R->getRegion();
    if (((RegionL && RegionL->isReplicator()) ||
         (RegionR && RegionR->isReplicator())) &&
        L->getParent() != R->getParent())
      return false;
    return L->getScalarType() == R->getScalarType();
  }
};
} // end anonymous namespace

/// Perform a common-subexpression-elimination of VPSingleDefRecipes on the \p
/// Plan.
void VPlanTransforms::cse(VPlan &Plan) {
  VPDominatorTree VPDT(Plan);
  DenseMap<VPSingleDefRecipe *, VPSingleDefRecipe *, VPCSEDenseMapInfo> CSEMap;
  // CSE map for widened loads. Must be cleared on recipes that may write to
  // memory, and at the end of each VPBB.
  DenseMap<VPSingleDefRecipe *, VPSingleDefRecipe *, VPCSEDenseMapInfo>
      LoadCSEMap;

  ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
      Plan.getEntry());
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
    for (VPRecipeBase &R : *VPBB) {
      if (R.mayWriteToMemory())
        LoadCSEMap.clear();
      auto *Def = dyn_cast<VPSingleDefRecipe>(&R);
      if (!Def || !VPCSEDenseMapInfo::canHandle(Def))
        continue;
      bool IsLoad = isa<VPWidenLoadRecipe, VPWidenLoadEVLRecipe>(Def);
      auto [It, Inserted] =
          (IsLoad ? LoadCSEMap : CSEMap).try_emplace(Def, Def);
      if (Inserted)
        continue;
      VPSingleDefRecipe *V = It->second;
      // V must dominate Def for a valid replacement.
      if (!VPDT.dominates(V->getParent(), VPBB))
        continue;
      if (IsLoad) {
        auto *EarlierLoad = cast<VPWidenMemoryRecipe>(V);
        auto *Load = cast<VPWidenMemoryRecipe>(Def);
        if (EarlierLoad->getAlign() < Load->getAlign()) {
          // Record Load as the candidate for subsequent loads, as it may be
          // reusable where EarlierLoad is not.
          It->second = Def;
          continue;
        }
        // Keep only metadata common to both loads on the survivor.
        EarlierLoad->intersect(*Load);
      }
      // Only keep flags present on both V and Def.
      if (auto *RFlags = dyn_cast<VPRecipeWithIRFlags>(V))
        RFlags->intersectFlags(*cast<VPRecipeWithIRFlags>(Def));
      Def->replaceAllUsesWith(V);
    }
    LoadCSEMap.clear();
  }
}

/// Return true if we do not know how to (mechanically) hoist or sink a
/// non-memory or memory recipe \p R out of a loop region. When sinking, passing
/// \p Sinking = true ensures that assumes aren't sunk.
static bool cannotHoistOrSinkRecipe(VPRecipeBase &R, VPBasicBlock *FirstBB,
                                    VPBasicBlock *LastBB,
                                    bool Sinking = false) {
  if (!isa<VPReplicateRecipe>(R) || !R.mayReadOrWriteMemory() ||
      match(&R, m_Intrinsic<Intrinsic::assume>()))
    return vputils::cannotHoistOrSinkRecipe(R, Sinking);

  // Check that the memory operation doesn't alias between FirstBB and LastBB.
  auto MemLoc = vputils::getMemoryLocation(R);

  // TODO: Could make use of SinkStoreInfo::isNoAliasViaDistance by collecting
  // stores upfront, and constructing a full SinkStoreInfo.
  auto SinkInfo =
      Sinking ? std::make_optional(SinkStoreInfo(cast<VPReplicateRecipe>(R)))
              : std::nullopt;

  return !MemLoc ||
         !canHoistOrSinkWithNoAliasCheck(*MemLoc, FirstBB, LastBB, SinkInfo);
}

/// Move loop-invariant recipes out of the vector loop region in \p Plan.
static void licm(VPlan &Plan) {
  VPBasicBlock *Preheader = Plan.getVectorPreheader();

  // Hoist any loop invariant recipes from the vector loop region to the
  // preheader. Preform a shallow traversal of the vector loop region, to
  // exclude recipes in replicate regions. Since the top-level blocks in the
  // vector loop region are guaranteed to execute if the vector pre-header is,
  // we don't need to check speculation safety.
  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
  assert(Preheader->getSingleSuccessor() == LoopRegion &&
         "Expected vector prehader's successor to be the vector loop region");
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_shallow(LoopRegion->getEntry()))) {
    for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
      if (any_of(R.operands(), [](VPValue *Op) {
            return !Op->isDefinedOutsideLoopRegions();
          }))
        continue;
      if (cannotHoistOrSinkRecipe(R, LoopRegion->getEntryBasicBlock(),
                                  LoopRegion->getExitingBasicBlock()))
        continue;
      R.moveBefore(*Preheader, Preheader->end());
    }
  }

#ifndef NDEBUG
  VPDominatorTree VPDT(Plan);
#endif
  // Sink recipes with no users inside the vector loop region if all users are
  // in the same exit block of the region.
  // TODO: Extend to sink recipes from inner loops.
  PostOrderTraversal<VPBlockShallowTraversalWrapper<VPBlockBase *>> POT(
      LoopRegion->getEntry());
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(POT)) {
    for (VPRecipeBase &R : make_early_inc_range(reverse(*VPBB))) {
      // TODO: Use R.definedValues() instead of casting to VPSingleDefRecipe to
      // support recipes with multiple defined values (e.g., interleaved loads).
      auto *Def = dyn_cast<VPSingleDefRecipe>(&R);
      if (!Def)
        continue;

      if (auto *RepR = dyn_cast<VPReplicateRecipe>(&R)) {
        assert(!RepR->isPredicated() &&
               "Expected prior transformation of predicated replicates to "
               "replicate regions");
        // narrowToSingleScalarRecipes should have already maximally narrowed
        // replicates to single-scalar replicates.
        // TODO: When unrolling, replicateByVF doesn't handle sunk
        // non-single-scalar replicates correctly.
        if (!RepR->isSingleScalar())
          continue;

        // The pointer operand of stores must be loop-invariant.
        if (RepR->getOpcode() == Instruction::Store &&
            !RepR->getOperand(1)->isDefinedOutsideLoopRegions())
          continue;
      }

      // Cannot sink the recipe if the user is defined in a loop region or a
      // non-successor of the vector loop region. Cannot sink if user is a phi
      // either.
      VPBasicBlock *SinkBB = nullptr;
      if (any_of(Def->users(), [&SinkBB, &LoopRegion](VPUser *U) {
            auto *UserR = cast<VPRecipeBase>(U);
            VPBasicBlock *Parent = UserR->getParent();
            // TODO: Support sinking when users are in multiple blocks.
            if (SinkBB && SinkBB != Parent)
              return true;
            SinkBB = Parent;
            // TODO: If the user is a PHI node, we should check the block of
            // incoming value. Support PHI node users if needed.
            return UserR->isPhi() || Parent->getEnclosingLoopRegion() ||
                   Parent->getSinglePredecessor() != LoopRegion;
          }))
        continue;

      if (cannotHoistOrSinkRecipe(R, LoopRegion->getEntryBasicBlock(),
                                  LoopRegion->getExitingBasicBlock(),
                                  /*Sinking=*/true))
        continue;

      [[maybe_unused]] auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
      assert((!R.mayWriteToMemory() ||
              (RepR && RepR->getOpcode() == Instruction::Store &&
               RepR->getOperand(1)->isDefinedOutsideLoopRegions())) &&
             "The only recipes that may write to memory are expected to be "
             "stores with invariant pointer-operand");

      if (!SinkBB)
        SinkBB = cast<VPBasicBlock>(LoopRegion->getSingleSuccessor());

      // TODO: This will need to be a check instead of a assert after
      // conditional branches in vectorized loops are supported.
      assert(VPDT.properlyDominates(VPBB, SinkBB) &&
             "Defining block must dominate sink block");
      // TODO: Clone the recipe if users are on multiple exit paths, instead of
      // just moving.
      Def->moveBefore(*SinkBB, SinkBB->getFirstNonPhi());
    }
  }
}

void VPlanTransforms::truncateToMinimalBitwidths(
    VPlan &Plan, const MapVector<Instruction *, uint64_t> &MinBWs) {
  if (Plan.hasScalarVFOnly())
    return;
  // Keep track of created truncates, so they can be re-used. Note that we
  // cannot use RAUW after creating a new truncate, as this would could make
  // other uses have different types for their operands, making them invalidly
  // typed.
  DenseMap<VPValue *, VPWidenCastRecipe *> ProcessedTruncs;
  VPBasicBlock *PH = Plan.getVectorPreheader();
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_deep(Plan.getVectorLoopRegion()))) {
    for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
      if (!isa<VPWidenRecipe, VPWidenCastRecipe, VPReplicateRecipe,
               VPWidenLoadRecipe, VPWidenIntrinsicRecipe>(&R))
        continue;

      VPValue *ResultVPV = R.getVPSingleValue();
      auto *UI = cast_or_null<Instruction>(ResultVPV->getUnderlyingValue());
      unsigned NewResSizeInBits = MinBWs.lookup(UI);
      if (!NewResSizeInBits)
        continue;

      // If the value wasn't vectorized, we must maintain the original scalar
      // type. Skip those here, after incrementing NumProcessedRecipes. Also
      // skip casts which do not need to be handled explicitly here, as
      // redundant casts will be removed during recipe simplification.
      if (isa<VPReplicateRecipe, VPWidenCastRecipe>(&R))
        continue;

      Type *OldResTy = ResultVPV->getScalarType();
      unsigned OldResSizeInBits = OldResTy->getScalarSizeInBits();
      assert(OldResTy->isIntegerTy() && "only integer types supported");
      (void)OldResSizeInBits;

      auto *NewResTy = IntegerType::get(Plan.getContext(), NewResSizeInBits);

      // Any wrapping introduced by shrinking this operation shouldn't be
      // considered undefined behavior. So, we can't unconditionally copy
      // arithmetic wrapping flags to VPW.
      if (auto *VPW = dyn_cast<VPRecipeWithIRFlags>(&R))
        VPW->dropPoisonGeneratingFlags();

      assert((OldResSizeInBits != NewResSizeInBits ||
              match(&R, m_ICmp(m_VPValue(), m_VPValue()))) &&
             "Only ICmps should not need extending the result.");
      assert(!isa<VPWidenStoreRecipe>(&R) && "stores cannot be narrowed");

      // Loads/intrinsics are not recreated; they keep producing their original
      // wide result and narrowed users will truncate it as needed below.
      if (isa<VPWidenLoadRecipe, VPWidenIntrinsicRecipe>(&R))
        continue;

      // Shrink operands by introducing truncates as needed.
      unsigned StartIdx =
          match(&R, m_Select(m_VPValue(), m_VPValue(), m_VPValue())) ? 1 : 0;
      SmallVector<VPValue *> NewOperands(R.operands());
      for (VPValue *&Op : drop_begin(NewOperands, StartIdx)) {
        unsigned OpSizeInBits = Op->getScalarType()->getScalarSizeInBits();
        if (OpSizeInBits == NewResSizeInBits)
          continue;
        assert(OpSizeInBits > NewResSizeInBits && "nothing to truncate");
        auto [ProcessedIter, Inserted] = ProcessedTruncs.try_emplace(Op);
        if (Inserted) {
          VPBuilder Builder;
          if (isa<VPIRValue>(Op))
            Builder.setInsertPoint(PH);
          else
            Builder.setInsertPoint(&R);
          ProcessedIter->second =
              Builder.createWidenCast(Instruction::Trunc, Op, NewResTy);
        }
        Op = ProcessedIter->second;
      }

      auto *NWR = cast<VPWidenRecipe>(&R)->cloneWithOperands(NewOperands);
      NWR->insertBefore(&R);

      // Wrap NWR in a ZExt to preserve the original wide type for downstream
      // users. Not needed for ICmps, whose result type is i1 irrespective of
      // the narrowing of their operands.
      VPValue *Replacement = NWR->getVPSingleValue();
      if (Replacement->getScalarType() != OldResTy)
        Replacement =
            VPBuilder::getToInsertAfter(NWR)
                .createWidenCast(Instruction::ZExt, Replacement, OldResTy)
                ->getVPSingleValue();
      ResultVPV->replaceAllUsesWith(Replacement);
      R.eraseFromParent();
    }
  }
}

bool VPlanTransforms::removeBranchOnConst(VPlan &Plan, bool OnlyLatches) {
  std::optional<VPDominatorTree> VPDT;
  if (OnlyLatches)
    VPDT.emplace(Plan);

  // Collect all blocks before modifying the CFG so we can identify unreachable
  // ones after constant branch removal.
  SmallVector<VPBlockBase *> AllBlocks(vp_depth_first_shallow(Plan.getEntry()));

  bool SimplifiedPhi = false;
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(AllBlocks)) {
    VPValue *Cond;
    // Skip blocks that are not terminated by BranchOnCond.
    if (VPBB->empty() || !match(&VPBB->back(), m_BranchOnCond(m_VPValue(Cond))))
      continue;

    if (OnlyLatches && !VPBlockUtils::isLatch(VPBB, *VPDT))
      continue;

    assert(VPBB->getNumSuccessors() == 2 &&
           "Two successors expected for BranchOnCond");
    unsigned RemovedIdx;
    if (match(Cond, m_True()))
      RemovedIdx = 1;
    else if (match(Cond, m_False()))
      RemovedIdx = 0;
    else
      continue;

    VPBasicBlock *RemovedSucc =
        cast<VPBasicBlock>(VPBB->getSuccessors()[RemovedIdx]);
    assert(count(RemovedSucc->getPredecessors(), VPBB) == 1 &&
           "There must be a single edge between VPBB and its successor");
    // Values coming from VPBB into phi recipes of RemovedSucc are removed from
    // these recipes and single-entry header phis are removed.
    for (VPRecipeBase &R : make_early_inc_range(RemovedSucc->phis())) {
      cast<VPPhiAccessors>(&R)->removeIncomingValueFor(VPBB);
      SimplifiedPhi = true;
      // Remove now invalid header phis that are left single-entry after
      // removing their backedges.
      auto *PhiR = dyn_cast<VPHeaderPHIRecipe>(&R);
      if (!PhiR || PhiR->getNumIncoming() != 1)
        continue;
      PhiR->replaceAllUsesWith(PhiR->getOperand(0));
      PhiR->eraseFromParent();
    }

    // Disconnect blocks and remove the terminator.
    VPBlockUtils::disconnectBlocks(VPBB, RemovedSucc);
    VPBB->back().eraseFromParent();
  }

  // Compute which blocks are still reachable from the entry after constant
  // branch removal.
  SmallPtrSet<VPBlockBase *, 16> Reachable(
      llvm::from_range, vp_depth_first_shallow(Plan.getEntry()));

  // Detach all unreachable blocks from their successors, removing their recipes
  // and incoming values from phi recipes.
  VPSymbolicValue Tmp(nullptr);
  for (VPBlockBase *B : AllBlocks) {
    if (Reachable.contains(B))
      continue;
    for (VPBlockBase *Succ : to_vector(B->successors())) {
      if (auto *SuccBB = dyn_cast<VPBasicBlock>(Succ))
        for (VPRecipeBase &R : SuccBB->phis())
          cast<VPPhiAccessors>(&R)->removeIncomingValueFor(B);
      VPBlockUtils::disconnectBlocks(B, Succ);
    }
    for (VPBasicBlock *DeadBB :
         VPBlockUtils::blocksOnly<VPBasicBlock>(vp_depth_first_deep(B))) {
      for (VPRecipeBase &R : make_early_inc_range(*DeadBB)) {
        for (VPValue *Def : R.definedValues())
          Def->replaceAllUsesWith(&Tmp);
        R.eraseFromParent();
      }
    }
  }
  return SimplifiedPhi;
}

void VPlanTransforms::optimize(VPlan &Plan) {
  RUN_VPLAN_PASS(removeRedundantInductionCasts, Plan);

  RUN_VPLAN_PASS(reassociateHeaderMask, Plan);
  RUN_VPLAN_PASS(combineRecipes, Plan);
  RUN_VPLAN_PASS(removeDeadRecipes, Plan);
  RUN_VPLAN_PASS(simplifyBlends, Plan);
  RUN_VPLAN_PASS(legalizeAndOptimizeInductions, Plan);
  RUN_VPLAN_PASS(narrowToSingleScalarRecipes, Plan);
  RUN_VPLAN_PASS(removeRedundantExpandSCEVRecipes, Plan);
  RUN_VPLAN_PASS(reassociateHeaderMask, Plan);
  RUN_VPLAN_PASS(combineRecipes, Plan);
  RUN_VPLAN_PASS(removeBranchOnConst, Plan, /*OnlyLatches=*/false);
  RUN_VPLAN_PASS(simplifyReverses, Plan);
  RUN_VPLAN_PASS(removeDeadRecipes, Plan);

  RUN_VPLAN_PASS(createAndOptimizeReplicateRegions, Plan);
  RUN_VPLAN_PASS(mergeBlocksIntoPredecessors, Plan);
  RUN_VPLAN_PASS(licm, Plan);
}

void VPlanTransforms::simplifyLiveInsWithSCEV(VPlan &Plan,
                                              PredicatedScalarEvolution &PSE) {
  auto GetSimplifiedLiveInViaSCEV = [&](VPValue *VPV) -> VPValue * {
    const SCEV *Expr = vputils::getSCEVExprForVPValue(VPV, PSE);
    const APInt *C;
    if (match(Expr, m_scev_APInt(C)))
      return Plan.getConstantInt(*C);
    return nullptr;
  };

  for (VPValue *LiveIn : to_vector(Plan.getLiveIns())) {
    if (VPValue *SimplifiedLiveIn = GetSimplifiedLiveInViaSCEV(LiveIn))
      LiveIn->replaceAllUsesWith(SimplifiedLiveIn);
  }
}

void VPlanTransforms::replaceSymbolicStrides(
    VPlan &Plan, PredicatedScalarEvolution &PSE,
    const SymbolicStrideMap &StridesMap, const VPDominatorTree &VPDT) {
  // Replace VPValues for known constant strides guaranteed by predicated scalar
  // evolution that are guaranteed to be guarded by the runtime checks; that is,
  // blocks dominated by the vector header.
  assert(!Plan.getVectorLoopRegion() &&
         "expected to run before loop regions are created");
  const auto &[Header, _] = VPBlockUtils::getPlainCFGHeaderAndLatch(Plan);
  auto CanUseVersionedStride = [&VPDT, Header = Header, &Plan](VPUser &U,
                                                               unsigned Idx) {
    auto *R = cast<VPRecipeBase>(&U);
    // Skip phis if the loop if loop is not yet guarded.
    if (isa<VPPhiAccessors>(R) &&
        Header == Plan.getEntry()->getSingleSuccessor())
      return false;
    return VPDT.dominates(Header, R->getParent());
  };
  ValueToSCEVMapTy RewriteMap;
  for (const SCEVUnknown *Stride : StridesMap.values()) {
    Value *StrideV = Stride->getValue();
    const APInt *StrideConst;
    const SCEV *StrideExpr = PSE.getSCEV(StrideV);
    if (!match(StrideExpr, m_scev_APInt(StrideConst)))
      // Only handle constant strides for now.
      continue;
    if (VPValue *StrideVPV = Plan.getLiveIn(StrideV))
      StrideVPV->replaceUsesWithIf(Plan.getConstantInt(*StrideConst),
                                   CanUseVersionedStride);

    // The versioned value may not be used in the loop directly but through an
    // integral cast (sext/zext/trunc). Add new live-ins in those cases.
    for (Value *U : StrideV->users()) {
      if (!isa<SExtInst, ZExtInst, TruncInst>(U))
        continue;
      VPValue *StrideVPV = Plan.getLiveIn(U);
      if (!StrideVPV)
        continue;
      unsigned BW = U->getType()->getScalarSizeInBits();
      APInt C = isa<SExtInst>(U) ? StrideConst->sext(BW)
                                 : StrideConst->zextOrTrunc(BW);
      StrideVPV->replaceUsesWithIf(Plan.getConstantInt(C),
                                   CanUseVersionedStride);
    }
    RewriteMap[StrideV] = StrideExpr;
  }

  for (VPExpandSCEVRecipe &ExpSCEV :
       make_isa_range<VPExpandSCEVRecipe>(*Plan.getEntry())) {
    const SCEV *ScevExpr = ExpSCEV.getSCEV();
    auto *NewSCEV =
        SCEVParameterRewriter::rewrite(ScevExpr, *PSE.getSE(), RewriteMap);
    if (NewSCEV != ScevExpr) {
      VPValue *NewExp = vputils::getOrCreateVPValueForSCEVExpr(Plan, NewSCEV);
      ExpSCEV.replaceAllUsesWith(NewExp);
      if (Plan.getTripCount() == &ExpSCEV)
        Plan.resetTripCount(NewExp);
    }
  }
}

void VPlanTransforms::dropPoisonGeneratingRecipes(VPlan &Plan) {
  // Collect recipes in the backward slice of `Root` that may generate a poison
  // value that is used after vectorization.
  SmallPtrSet<VPRecipeBase *, 16> Visited;
  auto CollectPoisonGeneratingInstrsInBackwardSlice([&](VPRecipeBase *Root) {
    SmallVector<VPRecipeBase *, 16> Worklist;
    Worklist.push_back(Root);

    // Traverse the backward slice of Root through its use-def chain.
    while (!Worklist.empty()) {
      VPRecipeBase *CurRec = Worklist.pop_back_val();

      if (!Visited.insert(CurRec).second)
        continue;

      // Prune search if we find another recipe generating a widen memory
      // instruction. Widen memory instructions involved in address computation
      // will lead to gather/scatter instructions, which don't need to be
      // handled.
      if (isa<VPWidenMemoryRecipe, VPInterleaveRecipe, VPScalarIVStepsRecipe,
              VPHeaderPHIRecipe>(CurRec))
        continue;

      // This recipe contributes to the address computation of a widen
      // load/store. If the underlying instruction has poison-generating flags,
      // drop them directly.
      if (auto *RecWithFlags = dyn_cast<VPRecipeWithIRFlags>(CurRec)) {
        VPValue *A, *B;
        // Dropping disjoint from an OR may yield incorrect results, as some
        // analysis may have converted it to an Add implicitly (e.g. SCEV used
        // for dependence analysis). Instead, replace it with an equivalent Add.
        // This is possible as all users of the disjoint OR only access lanes
        // where the operands are disjoint or poison otherwise.
        if (match(RecWithFlags, m_BinaryOr(m_VPValue(A), m_VPValue(B))) &&
            RecWithFlags->isDisjoint()) {
          VPBuilder Builder(RecWithFlags);
          VPInstruction *New =
              Builder.createAdd(A, B, RecWithFlags->getDebugLoc());
          New->setUnderlyingValue(RecWithFlags->getUnderlyingValue());
          RecWithFlags->replaceAllUsesWith(New);
          RecWithFlags->eraseFromParent();
          CurRec = New;
        } else
          RecWithFlags->dropPoisonGeneratingFlags();
      } else {
        Instruction *Instr = dyn_cast_or_null<Instruction>(
            CurRec->getVPSingleValue()->getUnderlyingValue());
        (void)Instr;
        assert((!Instr || !Instr->hasPoisonGeneratingFlags()) &&
               "found instruction with poison generating flags not covered by "
               "VPRecipeWithIRFlags");
      }

      // Add new definitions to the worklist.
      for (VPValue *Operand : CurRec->operands())
        if (VPRecipeBase *OpDef = Operand->getDefiningRecipe())
          Worklist.push_back(OpDef);
    }
  });

  // We want to exclude the tail folding case, as we don't need to drop flags
  // for operations computing the first lane in this case: the first lane of the
  // header mask must always be true. For reverse memory accesses, the mask is
  // wrapped in a Reverse, which is just a permutation of the header mask, so
  // peel it off before checking. The header mask is still the abstract region
  // value at this point (materialization happens later).
  auto m_UnlessHdrMask = m_Unless( // NOLINT
      m_CombineOr(m_HeaderMask(), m_Reverse(m_HeaderMask())));

  // Traverse all the recipes in the VPlan and collect the poison-generating
  // recipes in the backward slice starting at the address of a VPWidenRecipe or
  // VPInterleaveRecipe.
  auto Iter =
      vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntryBasicBlock());
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) {
    for (VPRecipeBase &Recipe : *VPBB) {
      if (auto *WidenRec = dyn_cast<VPWidenMemoryRecipe>(&Recipe)) {
        VPRecipeBase *AddrDef = WidenRec->getAddr()->getDefiningRecipe();
        if (AddrDef && WidenRec->isConsecutive() && WidenRec->getMask() &&
            match(WidenRec->getMask(), m_UnlessHdrMask))
          CollectPoisonGeneratingInstrsInBackwardSlice(AddrDef);
      } else if (auto *InterleaveRec = dyn_cast<VPInterleaveRecipe>(&Recipe)) {
        VPRecipeBase *AddrDef = InterleaveRec->getAddr()->getDefiningRecipe();
        if (AddrDef && InterleaveRec->getMask() &&
            match(InterleaveRec->getMask(), m_UnlessHdrMask))
          CollectPoisonGeneratingInstrsInBackwardSlice(AddrDef);
      }
    }
  }
}

void VPlanTransforms::createInterleaveGroups(
    VPlan &Plan,
    const SmallPtrSetImpl<const InterleaveGroup<Instruction> *>
        &InterleaveGroups,
    const bool &EpilogueAllowed) {
  if (InterleaveGroups.empty())
    return;

  DenseMap<Instruction *, VPWidenMemoryRecipe *> IRMemberToRecipe;
  for (VPBasicBlock *VPBB :
       VPBlockUtils::blocksOnly<VPBasicBlock>(vp_depth_first_shallow(
           Plan.getVectorLoopRegion()->getEntryBasicBlock())))
    for (VPRecipeBase &R : make_filter_range(*VPBB, [](VPRecipeBase &R) {
           return isa<VPWidenMemoryRecipe>(&R);
         })) {
      auto *MemR = cast<VPWidenMemoryRecipe>(&R);
      IRMemberToRecipe[&MemR->getIngredient()] = MemR;
    }

  // Interleave memory: for each Interleave Group we marked earlier as relevant
  // for this VPlan, replace the Recipes widening its memory instructions with a
  // single VPInterleaveRecipe at its insertion point.
  VPDominatorTree VPDT(Plan);
  for (const auto *IG : InterleaveGroups) {
    VPWidenMemoryRecipe *Start = nullptr;
    Instruction *StartMember = nullptr;
    for (auto *Member : IG->members())
      if (VPWidenMemoryRecipe *R = IRMemberToRecipe.lookup(Member)) {
        StartMember = Member;
        Start = R;
        break;
      }
    if (!StartMember) // All member recipes are dead, so the group is dead.
      continue;
    VPIRMetadata InterleaveMD(*Start);
    SmallVector<VPValue *, 4> StoredValues;
    for (unsigned I = 0; I < IG->getFactor(); ++I) {
      Instruction *MemberI = IG->getMember(I);
      if (!MemberI)
        continue;
      if (VPWidenMemoryRecipe *MemoryR = IRMemberToRecipe.lookup(MemberI)) {
        if (auto *StoreR = dyn_cast<VPWidenStoreRecipe>(MemoryR->getAsRecipe()))
          StoredValues.push_back(StoreR->getStoredValue());
        InterleaveMD.intersect(*MemoryR);
      } else {
        InterleaveMD.intersect(VPIRMetadata(*MemberI));
      }
    }

    bool NeedsMaskForGaps =
        (IG->requiresScalarEpilogue() && !EpilogueAllowed) ||
        (!StoredValues.empty() && !IG->isFull());

    Instruction *IRInsertPos = IG->getInsertPos();
    auto *InsertPos = IRMemberToRecipe.lookup(IRInsertPos);
    if (!InsertPos) {
      // InsertPos member is dead: find a new member that is alive.
      assert(isa<VPWidenLoadRecipe>(Start->getAsRecipe()) &&
             "Dead member in non-load group?");
      InsertPos = Start;
      for (Instruction *Member : IG->members())
        if (VPWidenMemoryRecipe *MemberR = IRMemberToRecipe.lookup(Member))
          if (VPDT.properlyDominates(MemberR->getAsRecipe(),
                                     InsertPos->getAsRecipe()))
            InsertPos = MemberR;
      IRInsertPos = &InsertPos->getIngredient();
    }
    VPRecipeBase *InsertPosR = InsertPos->getAsRecipe();

    GEPNoWrapFlags NW = GEPNoWrapFlags::none();
    if (auto *Gep = dyn_cast<GetElementPtrInst>(
            getLoadStorePointerOperand(IRInsertPos)->stripPointerCasts()))
      NW = Gep->getNoWrapFlags().withoutNoUnsignedWrap();

    // Get or create the start address for the interleave group.
    VPValue *Addr = Start->getAddr();
    VPRecipeBase *AddrDef = Addr->getDefiningRecipe();
    if (IG->getIndex(StartMember) != 0 ||
        (AddrDef && !VPDT.properlyDominates(AddrDef, InsertPosR))) {
      // Either member zero's recipe is dead, or we cannot re-use the address of
      // member zero because it does not dominate the insert position. Instead,
      // use the address of the insert position and create a PtrAdd adjusting it
      // to the address of member zero.
      // TODO: Hoist Addr's defining recipe (and any operands as needed) to
      // InsertPos or sink loads above zero members to join it.
      assert(IG->getIndex(IRInsertPos) != 0 &&
             "index of insert position shouldn't be zero");
      auto &DL = IRInsertPos->getDataLayout();
      APInt Offset(32,
                   DL.getTypeAllocSize(getLoadStoreType(IRInsertPos)) *
                       IG->getIndex(IRInsertPos),
                   /*IsSigned=*/true);
      VPValue *OffsetVPV = Plan.getConstantInt(-Offset);
      VPBuilder B(InsertPosR);
      Addr = B.createNoWrapPtrAdd(InsertPos->getAddr(), OffsetVPV, NW);
    }
    // If the group is reverse, adjust the index to refer to the last vector
    // lane instead of the first. We adjust the index from the first vector
    // lane, rather than directly getting the pointer for lane VF - 1, because
    // the pointer operand of the interleaved access is supposed to be uniform.
    if (IG->isReverse()) {
      auto *ReversePtr = new VPVectorEndPointerRecipe(
          Addr, &Plan.getVF(), getLoadStoreType(IRInsertPos),
          -(int64_t)IG->getFactor(), NW, InsertPosR->getDebugLoc());
      ReversePtr->insertBefore(InsertPosR);
      Addr = ReversePtr;
    }
    auto *VPIG = new VPInterleaveRecipe(
        IG, Addr, StoredValues, InsertPos->getMask(), NeedsMaskForGaps,
        InterleaveMD, InsertPosR->getDebugLoc());
    VPIG->insertBefore(InsertPosR);

    unsigned J = 0;
    for (unsigned i = 0; i < IG->getFactor(); ++i)
      if (Instruction *Member = IG->getMember(i)) {
        VPWidenMemoryRecipe *MemberR = IRMemberToRecipe.lookup(Member);
        if (!Member->getType()->isVoidTy()) {
          if (MemberR) {
            VPValue *OriginalV = MemberR->getAsRecipe()->getVPSingleValue();
            OriginalV->replaceAllUsesWith(VPIG->getVPValue(J));
          }
          J++;
        }
        if (MemberR)
          MemberR->getAsRecipe()->eraseFromParent();
      }
  }
}

/// Matches an exit condition formed by comparing a value loaded from memory
/// with a loop-invariant term. Binds the comparison for the condition.
static auto m_Uncountable(VPValue *&Cond) {
  return m_VPValue(
      Cond,
      m_c_Cmp(m_VPInstruction<Instruction::Load>(m_VPValue()), m_LiveIn()));
}

namespace {
struct CountableConditionMatch {
  VPValue *&Cmp;
  PredicatedScalarEvolution &PSE;
  Loop *L;

  CountableConditionMatch(VPValue *&Cmp, PredicatedScalarEvolution &PSE,
                          Loop *L)
      : Cmp(Cmp), PSE(PSE), L(L) {}

  template <typename ITy> bool match(ITy *V) const {
    VPValue *Update;
    if (!VPlanPatternMatch::match(
            V, m_VPValue(Cmp, m_c_ICmp(m_VPValue(Update, m_Add(m_VPValue(),
                                                               m_VPValue())),
                                       m_LiveIn()))))
      return false;

    const SCEV *S = vputils::getSCEVExprForVPValue(Update, PSE, L);
    return SCEVPatternMatch::match(
        S, m_scev_AffineAddRec(m_SCEV(), m_scev_One(), m_SpecificLoop(L)));
  }
};
} // end anonymous namespace

/// Matches an exit condition formed by comparing the current value of a
/// affine add recurrence in the given loop with a stride of 1 against a
/// loop-invariant term. Binds the comparison for the condition.
static auto m_Countable(VPValue *&Cmp, PredicatedScalarEvolution &PSE,
                        Loop *L) {
  return CountableConditionMatch(Cmp, PSE, L);
}

bool VPlanTransforms::splitCombinedExits(VPlan &Plan,
                                         PredicatedScalarEvolution &PSE,
                                         Loop *L) {
  // Check for a single combined exit in the latch block.
  // TODO: Generalize to other blocks besides the latch.
  // If we don't find a combined condition in the latch, just return true
  // to proceed with vectorization.
  auto [_, LatchVPBB] = VPBlockUtils::getPlainCFGHeaderAndLatch(Plan);

  // We're looking for a conditional branch...
  auto *Term = dyn_cast<VPInstruction>(LatchVPBB->getTerminator());
  if (!Term || Term->getOpcode() != VPInstruction::BranchOnCond)
    return true;

  // ...where the condition is a combination of both a countable and an
  // uncountable comparison.
  VPValue *Uncountable = nullptr;
  VPValue *Countable = nullptr;
  VPValue *Cond = Term->getOperand(0);
  if (!match(Cond, m_OneUse(m_CombineOr(
                       m_c_LogicalOr(m_Uncountable(Uncountable),
                                     m_Countable(Countable, PSE, L)),
                       m_c_BinaryOr(m_Uncountable(Uncountable),
                                    m_Countable(Countable, PSE, L))))))
    return true;

  // If the conditions are combined with a logical or (select), then we'll
  // need to freeze the individual terms when splitting.
  bool NeedsFreeze = match(Cond, m_LogicalOr(m_VPValue(), m_VPValue()));

  // If we do have a combined exit condition, bail out if there's more than
  // one exit block.
  // TODO: Support additional exits.
  ArrayRef<VPIRBasicBlock *> ExitBlocks = Plan.getExitBlocks();
  if (ExitBlocks.size() != 1)
    return false;

  // If there are any live-outs, bail out. The exit block is an existing IR
  // block, and if we split the exiting block then the incoming blocks and
  // values won't be correct.
  // TODO: Support live-outs with combined exits.
  if (!ExitBlocks.front()->phis().empty())
    return false;

  // Split the latch block just before the terminator.
  VPBasicBlock *NewLatch = LatchVPBB->splitAt(Term->getIterator());

  // Create new terminator for uncountable condition.
  VPBuilder EEBuilder(LatchVPBB);
  if (NeedsFreeze)
    Uncountable = EEBuilder.createFreeze(Uncountable);
  EEBuilder.createNaryOp(VPInstruction::BranchOnCond, {Uncountable});

  // We need to connect the uncountable exit to the sole exit block. The
  // latch is expected to connect to the middle block instead.
  // In canonical form, the backedge is the last successor for the latch. So
  // the first successor (true path) should be the exit for both conditions.
  LatchVPBB->clearSuccessors();
  NewLatch->clearPredecessors();
  VPBlockUtils::connectBlocks(LatchVPBB, ExitBlocks.front());
  VPBlockUtils::connectBlocks(LatchVPBB, NewLatch);

  // Set condition for latch block to countable condition.
  if (NeedsFreeze) {
    VPBuilder NewLatchBuilder(Term);
    Countable = NewLatchBuilder.createFreeze(Countable);
  }
  Term->setOperand(0, Countable);

  // Remove the combining or.
  cast<VPInstruction>(Cond)->eraseFromParent();

  return true;
}

/// Returns the VPValue representing the uncountable exit comparison used by
/// AnyOf if the recipes it depends on can be traced back to live-ins and
/// the addresses (in GEP/PtrAdd form) of any (non-masked) load used in
/// generating the values for the comparison. The recipes are stored in
/// \p Recipes.
static VPValue *
getRecipesForUncountableExit(SmallVectorImpl<VPInstruction *> &Recipes,
                             VPBasicBlock *LatchVPBB) {
  // Given a plain CFG VPlan loop with countable latch exiting block
  // \p LatchVPBB, we're looking to match the recipes contributing to the
  // uncountable exit condition comparison (here, vp<%4>) back to either
  // live-ins or the address nodes for the load used as part of the uncountable
  // exit comparison so that we can either move them within the loop, or copy
  // them to the preheader depending on the chosen method for dealing with
  // stores in uncountable exit loops.
  //
  // Currently, the address of the load is restricted to a GEP with 2 operands
  // and a live-in base address. This constraint may be relaxed later.
  //
  // VPlan ' for UF>=1' {
  // Live-in vp<%0> = VF * UF
  // Live-in vp<%1> = vector-trip-count
  // Live-in ir<20> = original trip-count
  //
  // ir-bb<entry>:
  // Successor(s): scalar.ph, vector.ph
  //
  // vector.ph:
  // Successor(s): for.body
  //
  // for.body:
  //   EMIT vp<%2> = phi ir<0>, vp<%index.next>
  //   EMIT-SCALAR ir<%iv> = phi [ ir<0>, vector.ph ], [ ir<%iv.next>, for.inc ]
  //   EMIT ir<%uncountable.addr> = getelementptr inbounds nuw ir<%pred>,ir<%iv>
  //   EMIT ir<%uncountable.val> = load ir<%uncountable.addr>
  //   EMIT ir<%uncountable.cond> = icmp sgt ir<%uncountable.val>, ir<500>
  //   EMIT vp<%3> = masked-cond ir<%uncountable.cond>
  // Successor(s): for.inc
  //
  // for.inc:
  //   EMIT ir<%iv.next> = add nuw nsw ir<%iv>, ir<1>
  //   EMIT ir<%countable.cond> = icmp eq ir<%iv.next>, ir<20>
  //   EMIT vp<%index.next> = add nuw vp<%2>, vp<%0>
  //   EMIT vp<%freeze> = freeze ir<%3>
  //   EMIT vp<%4> = any-of ir<%freeze>
  //   EMIT vp<%5> = icmp eq vp<%index.next>, vp<%1>
  //   EMIT branch-on-two-conds vp<%4>, vp<%5>
  // Successor(s): middle.block, middle.block, for.body
  //
  // middle.block:
  // Successor(s): ir-bb<exit>, scalar.ph
  //
  // ir-bb<exit>:
  // No successors
  //
  // scalar.ph:
  // }

  // Find the uncountable loop exit condition.
  VPValue *UncountableCondition = nullptr;
  if (!match(LatchVPBB->getTerminator(),
             m_BranchOnTwoConds(m_AnyOf(m_VPValue(UncountableCondition)),
                                m_VPValue())))
    return nullptr;

  SmallVector<VPValue *, 4> Worklist;
  Worklist.push_back(UncountableCondition);
  while (!Worklist.empty()) {
    VPValue *V = Worklist.pop_back_val();

    // Any value defined outside the loop does not need to be copied.
    if (V->isDefinedOutsideLoopRegions())
      continue;

    // FIXME: Remove the single user restriction; it's here because we're
    //        starting with the simplest set of loops we can, and multiple
    //        users means needing to add PHI nodes in the transform.
    if (V->getNumUsers() > 1)
      return nullptr;

    VPValue *Op1, *Op2;
    // Walk back through recipes until we find at least one load from memory.
    if (match(V, m_ICmp(m_VPValue(Op1), m_VPValue(Op2)))) {
      Worklist.push_back(Op1);
      Worklist.push_back(Op2);
      Recipes.push_back(cast<VPInstruction>(V->getDefiningRecipe()));
    } else if (match(V, m_VPInstruction<Instruction::Load>(m_VPValue(Op1)))) {
      VPRecipeBase *GepR = Op1->getDefiningRecipe();
      // Only matching base + single offset term for now.
      if (GepR->getNumOperands() != 2)
        return nullptr;
      // Matching a GEP with a loop-invariant base ptr.
      if (!match(GepR, m_VPInstruction<Instruction::GetElementPtr>(
                           m_LiveIn(), m_VPValue())))
        return nullptr;
      Recipes.push_back(cast<VPInstruction>(V->getDefiningRecipe()));
      Recipes.push_back(cast<VPInstruction>(GepR));
    } else if (match(V, m_Freeze(m_VPValue(Op1)))) {
      Worklist.push_back(Op1);
      Recipes.push_back(cast<VPInstruction>(V->getDefiningRecipe()));
    } else if (match(V, m_VPInstruction<VPInstruction::MaskedCond>(
                            m_VPValue(Op1)))) {
      Worklist.push_back(Op1);
      Recipes.push_back(cast<VPInstruction>(V->getDefiningRecipe()));
    } else
      return nullptr;
  }

  // If we couldn't match anything, don't return the condition. It may be
  // defined outside the loop.
  if (Recipes.empty() ||
      none_of(Recipes, match_fn(m_VPInstruction<Instruction::GetElementPtr>())))
    return nullptr;

  return UncountableCondition;
}

struct EarlyExitInfo {
  VPBasicBlock *EarlyExitingVPBB;
  VPIRBasicBlock *EarlyExitVPBB;
  VPValue *CondToExit;
};

/// Update \p Plan to mask memory operations in the loop based on whether the
/// early exit is taken or not.
///
/// We're currently expecting to find a loop with properties similar to the
/// following:
///
/// for.body:
///   ir<%indvars.iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<%0>
///   EMIT ir<%arrayidx> = getelementptr inbounds nuw ir<@c>, ir<%indvars.iv>
///   EMIT-SCALAR ir<%0> = load ir<%arrayidx>
///   EMIT ir<%cmp1> = icmp sgt ir<%0>, ir<5>
///   EMIT vp<%1> = masked-cond ir<%cmp1>
/// Successor(s): if.end
///
/// if.end:
///   EMIT ir<%arrayidx3> = getelementptr inbounds nuw ir<@src>, ir<%indvars.iv>
///   EMIT-SCALAR ir<%2> = load ir<%arrayidx3>
///   EMIT ir<%add> = add nsw ir<%2>, ir<42>
///   EMIT ir<%arrayidx5> = getelementptr inbounds nuw ir<@dst>, ir<%indvars.iv>
///   EMIT store ir<%add>, ir<%arrayidx5>
///   EMIT ir<%indvars.iv.next> = add nuw nsw ir<%indvars.iv>, ir<1>
///   EMIT vp<%freeze> = freeze ir<%1>
///   EMIT vp<%3> = any-of ir<%freeze>
///   EMIT ir<%exitcond.not> = icmp eq ir<%indvars.iv.next>, ir<10000>
///   EMIT branch-on-two-conds vp<%3>, ir<%exitcond.not>
/// Successor(s): middle.block, middle.block, for.body
///
/// We currently expect LoopVectorizationLegality to ensure that:
/// * There must also be a counted exit. We will need to support speculative
///   or first-faulting loads before we can remove this restriction.
/// * Any stores within the loop must not alias with the load used for the
///   uncountable exit. We can relax this a bit with runtime aliasing checks.
/// * Other memory operations in the loop can take place before or after the
///   uncountable exit, but must also be unconditional. We need to support
///   combining the conditions in VPlanPredicator.
/// * The loop must have a single unconditional load contributing to the
///   uncountable exit comparison, and the other term must be loop-invariant.
///   Improving upon this requires work in getRecipesForUncountableExit to
///   handle more complex recipe graphs.
static bool handleUncountableExitsWithSideEffects(
    VPlan &Plan, SmallVectorImpl<EarlyExitInfo> &Exits,
    VPBasicBlock *HeaderVPBB, VPBasicBlock *LatchVPBB, VPBasicBlock *MiddleVPBB,
    OptimizationRemarkEmitter *ORE, Loop *TheLoop,
    PredicatedScalarEvolution &PSE, DominatorTree &DT, AssumptionCache *AC) {

  // Disconnect early exiting blocks from successors, remove branches. We
  // currently don't support multiple uses for recipes involved in creating
  // the uncountable exit condition.
  for (auto &Exit : Exits) {
    if (Exit.EarlyExitingVPBB == LatchVPBB)
      continue;

    for (VPRecipeBase &R : Exit.EarlyExitVPBB->phis())
      cast<VPIRPhi>(&R)->removeIncomingValueFor(Exit.EarlyExitingVPBB);
    Exit.EarlyExitingVPBB->getTerminator()->eraseFromParent();
    VPBlockUtils::disconnectBlocks(Exit.EarlyExitingVPBB, Exit.EarlyExitVPBB);
  }

  VPDominatorTree VPDT(Plan);

  // We can abandon a VPlan entirely if we return false here, so we shouldn't
  // crash if some earlier assumptions on scalar IR don't hold for the vplan
  // version of the loop.
  SmallVector<VPInstruction *, 8> ConditionRecipes;

  VPValue *Cond = getRecipesForUncountableExit(ConditionRecipes, LatchVPBB);
  if (!Cond) {
    reportVectorizationFailure("Unable to determine early exit condition for "
                               "loop with side effects",
                               "EarlyExitSideEffectsCond", ORE, TheLoop);
    return false;
  }

  // Find load contributing to condition.
  // At the moment LoopVectorizationLegality only supports a single
  // early-exit expression with a compare and a single load that must
  // be unconditional.
  // TODO: Support more than one load.
  auto *Load =
      find_singleton<VPInstruction>(ConditionRecipes, [](auto *I, bool _) {
        return match(I, m_VPInstruction<Instruction::Load>(m_VPValue()))
                   ? I
                   : nullptr;
      });
  assert(Load && "Couldn't find exactly one load");
  // TODO: Support conditional loads for uncountable exits.
  assert(VPDT.dominates(Load->getParent(), LatchVPBB) &&
         "Uncountable exit condition load is conditional.");
  VPInstruction *Ptr = cast<VPInstruction>(Load->getOperand(0));

  // Ensure that we are guaranteed to be able to dereference the memory used
  // for determining the uncountable exit for the maximum possible number of
  // scalar iterations of the loop.
  //
  // TODO: Support first-faulting loads in cases where we don't know whether
  //       all possible addresses are dereferenceable.
  {
    SmallVector<const SCEVPredicate *, 4> Predicates;
    const SCEV *PtrSCEV = vputils::getSCEVExprForVPValue(Ptr, PSE, TheLoop);
    const DataLayout &DL = Plan.getDataLayout();
    APInt EltSize(DL.getIndexTypeSizeInBits(Ptr->getScalarType()),
                  DL.getTypeStoreSize(Load->getScalarType()).getFixedValue());
    if (!isDereferenceableAndAlignedInLoop(
            PtrSCEV, cast<LoadInst>(Load->getUnderlyingInstr())->getAlign(),
            PSE.getSE()->getConstant(EltSize), TheLoop, *PSE.getSE(), DT, AC,
            &Predicates)) {
      reportVectorizationFailure("Early exit loop with side effects contains "
                                 "load used by the exit condition that may "
                                 "fault",
                                 "EarlyExitSideEffectsFaultingLoad", ORE,
                                 TheLoop);
      return false;
    }
  }

  // Check for a single GEP for the condition load to see if we can link it to
  // a widen IV recipe with a step of 1; we're only interested in contiguous
  // accesses for the condition load right now.
  auto *IV = cast<VPWidenInductionRecipe>(&HeaderVPBB->front());
  if (!match(IV->getStartValue(), m_SpecificInt(0)) ||
      !match(IV->getStepValue(), m_SpecificInt(1))) {
    reportVectorizationFailure("Early exit loop with side effects contains "
                               "load used by the exit condition with an "
                               "unsupported memory access pattern",
                               "EarlyExitSideEffectsBadLoadAccessPattern", ORE,
                               TheLoop);
    return false;
  }

  if (!match(Ptr, m_VPInstruction<Instruction::GetElementPtr>(
                      m_LiveIn(), m_Specific(IV)))) {
    reportVectorizationFailure("Early exit loop with side effects contains "
                               "load used by the exit condition with an "
                               "unsupported memory access pattern",
                               "EarlyExitSideEffectsBadLoadAccessPattern", ORE,
                               TheLoop);
    return false;
  }

  // We want to guarantee that the uncountable exit condition (and the mask
  // we will generate from it) are available for all operations in the loop
  // that need to be masked. If the condition recipes are not already the first
  // recipes in the header after the last phi, move them there.
  auto InsertIt = HeaderVPBB->getFirstNonPhi();
  while (InsertIt != HeaderVPBB->end() &&
         is_contained(ConditionRecipes, &*InsertIt)) {
    erase(ConditionRecipes, &*InsertIt);
    InsertIt++;
  }
  for (auto *Recipe : reverse(ConditionRecipes))
    Recipe->moveBefore(*HeaderVPBB, InsertIt);

  // Create a mask to represent all lanes that fully execute in the vector loop,
  // stopping short of any early exit.
  VPBuilder MaskBuilder(HeaderVPBB, InsertIt);
  VPValue *FirstActive = MaskBuilder.createFirstActiveLane(Cond);
  Type *IVScalarTy = IV->getScalarType();
  VPValue *Zero = Plan.getZero(IVScalarTy);
  FirstActive =
      MaskBuilder.createScalarZExtOrTrunc(FirstActive, IVScalarTy, DebugLoc());
  VPValue *Mask = MaskBuilder.createNaryOp(VPInstruction::ActiveLaneMask,
                                           {Zero, FirstActive}, DebugLoc(),
                                           "uncountable.exit.mask");

  // Convert all other memory operations to use the mask.
  for (VPBasicBlock *VPBB : vp_rpo_plain_cfg_loop_body(HeaderVPBB))
    for (VPRecipeBase &R : *VPBB)
      if (R.mayReadOrWriteMemory() && &R != Load) {
        // TODO: Handle conditional memory operations in the loop.
        if (!VPDT.dominates(R.getParent(), LatchVPBB)) {
          reportVectorizationFailure(
              "Early exit loop with side effects contains unsupported "
              "conditional memory operations",
              "EarlyExitSideEffectsUnsupportedConditionalMemOps", ORE, TheLoop);
          return false;
        }
        cast<VPInstruction>(&R)->addMask(Mask);
      }

  // Update middle block branch to compare (IV + however many lanes were active)
  // against the full trip count, since we may be exiting the vector loop early.
  // If we didn't take an early exit, we should get the equivalent of VF from
  // the FirstActiveLane.
  assert(match(MiddleVPBB->getTerminator(), m_BranchOnCond()) &&
         "Expected BranchOnCond terminator for MiddleVPBB");
  VPBuilder MiddleBuilder(MiddleVPBB->getTerminator());
  VPValue *ScalarIV = MiddleBuilder.createNaryOp(VPInstruction::ExtractLane,
                                                 {Zero, IV}, DebugLoc());
  VPValue *ExitIV = MiddleBuilder.createAdd(ScalarIV, FirstActive);
  VPValue *FullTC =
      MiddleBuilder.createICmp(CmpInst::ICMP_EQ, ExitIV, Plan.getTripCount());
  MiddleVPBB->getTerminator()->setOperand(0, FullTC);

  // Update resume phi in scalar.ph.
  VPBasicBlock *ScalarPH = Plan.getScalarPreheader();
  auto Phis = ScalarPH->phis();
  // TODO: Handle more than one Phi; re-derive from IV.
  // TODO: Handle reductions.
  if (range_size(Phis) != 1) {
    reportVectorizationFailure(
        "Early exit loop with side effects contains "
        "unsupported reductions, inductions or recurrences",
        "EarlyExitSideEffectsReductions", ORE, TheLoop);
    return false;
  }
  VPPhi *ContinueIV = cast<VPPhi>(Phis.begin());
  // Make sure we're referring to the same IV.
  assert(
      match(ContinueIV->getOperand(0),
            m_VPInstruction<VPInstruction::ExitingIVValue>(m_Specific(IV))) &&
      "Continuing from different IV");
  ContinueIV->setOperand(0, ExitIV);
  return true;
}

bool VPlanTransforms::handleUncountableEarlyExits(
    VPlan &Plan, OptimizationRemarkEmitter *ORE, Loop *TheLoop,
    PredicatedScalarEvolution &PSE, DominatorTree &DT, AssumptionCache *AC,
    UncountableExitStyle Style) {
#ifndef NDEBUG
  VPDominatorTree VPDT(Plan);
#endif

  auto *MiddleVPBB = VPBlockUtils::getPlainCFGMiddleBlock(Plan);
  auto [HeaderVPBB, LatchVPBB] = VPBlockUtils::getPlainCFGHeaderAndLatch(Plan);

  // Dereferenceability is checked separately for uncountable exit loops with
  // stores, as only the loads contributing to the exit condition need to
  // be checked.
  if (Style == UncountableExitStyle::ReadOnly &&
      !areAllLoadsDereferenceable(HeaderVPBB, TheLoop, PSE, DT, AC)) {
    reportVectorizationFailure(
        "Auto-vectorization of early exit loops with potentially "
        "faulting loads is not supported",
        "EarlyExitFaultingLoads", ORE, TheLoop);
    return false;
  }

  VPBuilder LatchBuilder(LatchVPBB->getTerminator());
  SmallVector<EarlyExitInfo> Exits;
  for (auto [EarlyExitingVPBB, ExitBlock] :
       vputils::getEarlyExits(Plan, MiddleVPBB)) {
    // Collect condition for this early exit.
    VPBlockBase *TrueSucc = EarlyExitingVPBB->getSuccessors()[0];
    VPValue *CondOfEarlyExitingVPBB;
    [[maybe_unused]] bool Matched =
        match(EarlyExitingVPBB->getTerminator(),
              m_BranchOnCond(m_VPValue(CondOfEarlyExitingVPBB)));
    assert(Matched && "Terminator must be BranchOnCond");

    // Insert the MaskedCond in the EarlyExitingVPBB so the predicator adds
    // the correct block mask.
    VPBuilder EarlyExitingBuilder(EarlyExitingVPBB->getTerminator());
    auto *CondToEarlyExit = EarlyExitingBuilder.createNaryOp(
        VPInstruction::MaskedCond,
        TrueSucc == ExitBlock
            ? CondOfEarlyExitingVPBB
            : EarlyExitingBuilder.createNot(CondOfEarlyExitingVPBB));
    assert((isa<VPIRValue>(CondOfEarlyExitingVPBB) ||
            !VPDT.properlyDominates(EarlyExitingVPBB, LatchVPBB) ||
            VPDT.properlyDominates(
                CondOfEarlyExitingVPBB->getDefiningRecipe()->getParent(),
                LatchVPBB)) &&
           "exit condition must dominate the latch");
    Exits.push_back({
        EarlyExitingVPBB,
        ExitBlock,
        CondToEarlyExit,
    });
  }

  assert(!Exits.empty() && "must have at least one early exit");
  // Sort exits by RPO order to get correct program order. RPO gives a
  // topological ordering of the CFG, ensuring upstream exits are checked
  // before downstream exits in the dispatch chain.
  ReversePostOrderTraversal<VPBlockShallowTraversalWrapper<VPBlockBase *>> RPOT(
      HeaderVPBB);
  DenseMap<VPBlockBase *, unsigned> RPOIdx;
  for (const auto &[Num, VPB] : enumerate(RPOT))
    RPOIdx[VPB] = Num;
  llvm::sort(Exits, [&RPOIdx](const EarlyExitInfo &A, const EarlyExitInfo &B) {
    return RPOIdx[A.EarlyExitingVPBB] < RPOIdx[B.EarlyExitingVPBB];
  });
#ifndef NDEBUG
  // After RPO sorting, verify that for any pair where one exit dominates
  // another, the dominating exit comes first. This is guaranteed by RPO
  // (topological order) and is required for the dispatch chain correctness.
  for (unsigned I = 0; I + 1 < Exits.size(); ++I)
    for (unsigned J = I + 1; J < Exits.size(); ++J)
      assert(!VPDT.properlyDominates(Exits[J].EarlyExitingVPBB,
                                     Exits[I].EarlyExitingVPBB) &&
             "RPO sort must place dominating exits before dominated ones");
#endif

  // Build the AnyOf condition for the latch terminator using logical OR
  // to avoid poison propagation from later exit conditions when an earlier
  // exit is taken.
  VPValue *Combined = Exits[0].CondToExit;
  for (const EarlyExitInfo &Info : drop_begin(Exits))
    Combined = LatchBuilder.createLogicalOr(Combined, Info.CondToExit);
  Combined = LatchBuilder.createFreeze(Combined);

  // Even though the logical or prevents posion propagation, we need to freeze
  // Combined to prevent poisoning the entire AnyOf result:
  //
  // Exits[0].CondToExit = [0,1,0,0]
  // Exits[1].CondToExit = [0,0,p,p]
  //            Combined = [0,1,p,p]
  //    freeze(Combined) = [0,1,?,?]
  //               AnyOf = 1
  VPValue *IsAnyExitTaken =
      LatchBuilder.createNaryOp(VPInstruction::AnyOf, Combined);

  // Create a comparison for the latch exit condition and replace the
  // BranchOnCond with a BranchOnTwoConds. The original BranchOnCond's condition
  // is used as the latch-exit condition; canonical IV recipes have not been
  // introduced yet, so there is no BranchOnCount to derive the condition from.
  auto *LatchExitingBranch = cast<VPInstruction>(LatchVPBB->getTerminator());
  assert(LatchExitingBranch->getOpcode() == VPInstruction::BranchOnCond &&
         "Unexpected terminator");
  VPValue *IsLatchExitTaken = LatchExitingBranch->getOperand(0);
  DebugLoc LatchDL = LatchExitingBranch->getDebugLoc();
  LatchExitingBranch->eraseFromParent();
  LatchBuilder.setInsertPoint(LatchVPBB);
  LatchBuilder.createNaryOp(VPInstruction::BranchOnTwoConds,
                            {IsAnyExitTaken, IsLatchExitTaken}, LatchDL);
  LatchVPBB->clearSuccessors();

  if (Style == UncountableExitStyle::MaskedHandleExitInScalarLoop) {
    // If handling the exiting lane in the scalar loop, combine the exit
    // conditions into a single BranchOnCond.
    LatchVPBB->setSuccessors({MiddleVPBB, MiddleVPBB, HeaderVPBB});
    MiddleVPBB->clearPredecessors();
    MiddleVPBB->setPredecessors({LatchVPBB, LatchVPBB});
    return handleUncountableExitsWithSideEffects(Plan, Exits, HeaderVPBB,
                                                 LatchVPBB, MiddleVPBB, ORE,
                                                 TheLoop, PSE, DT, AC);
  }

  // Create the vector.early.exit blocks.
  SmallVector<VPBasicBlock *> VectorEarlyExitVPBBs(Exits.size());
  for (unsigned Idx = 0; Idx != Exits.size(); ++Idx) {
    Twine BlockSuffix = Exits.size() == 1 ? "" : Twine(".") + Twine(Idx);
    VPBasicBlock *VectorEarlyExitVPBB =
        Plan.createVPBasicBlock("vector.early.exit" + BlockSuffix);
    VectorEarlyExitVPBBs[Idx] = VectorEarlyExitVPBB;
  }

  // Create the dispatch block (or reuse the single exit block if only one
  // exit). The dispatch block computes the first active lane of the combined
  // condition and, for multiple exits, chains through conditions to determine
  // which exit to take.
  VPBasicBlock *DispatchVPBB =
      Exits.size() == 1 ? VectorEarlyExitVPBBs[0]
                        : Plan.createVPBasicBlock("vector.early.exit.check");
  DispatchVPBB->setPredecessors({LatchVPBB});
  LatchVPBB->setSuccessors({DispatchVPBB, MiddleVPBB, HeaderVPBB});
  VPBuilder DispatchBuilder(DispatchVPBB, DispatchVPBB->begin());
  VPValue *FirstActiveLane = DispatchBuilder.createFirstActiveLane(
      {Combined}, DebugLoc::getUnknown(), "first.active.lane");

  // For each early exit, disconnect the original exiting block
  // (early.exiting.I) from the exit block (ir-bb<exit.I>) and route through a
  // new vector.early.exit block. Update ir-bb<exit.I>'s phis to extract their
  // values at the first active lane:
  //
  // Input:
  //  early.exiting.I:
  //     ...
  //    EMIT branch-on-cond vp<%cond.I>
  //  Successor(s): in.loop.succ, ir-bb<exit.I>
  //
  //  ir-bb<exit.I>:
  //    IR %phi = phi [ vp<%incoming.I>, early.exiting.I ], ...
  //
  // Output:
  //  early.exiting.I:
  //    ...
  //  Successor(s): in.loop.succ
  //
  //  vector.early.exit.I:
  //    EMIT vp<%exit.val> = extract-lane vp<%first.lane>, vp<%incoming.I>
  //  Successor(s): ir-bb<exit.I>
  //
  //  ir-bb<exit.I>:
  //    IR %phi = phi ... (extra operand: vp<%exit.val> from
  //                                      vector.early.exit.I)
  //
  for (auto [Exit, VectorEarlyExitVPBB] :
       zip_equal(Exits, VectorEarlyExitVPBBs)) {
    auto &[EarlyExitingVPBB, EarlyExitVPBB, _] = Exit;
    // Adjust the phi nodes in EarlyExitVPBB.
    //   1. remove incoming values from EarlyExitingVPBB,
    //   2. extract the incoming value at FirstActiveLane
    //   3. add back the extracts as last operands for the phis
    // Then adjust the CFG, removing the edge between EarlyExitingVPBB and
    // EarlyExitVPBB and adding a new edge between VectorEarlyExitVPBB and
    // EarlyExitVPBB. The extracts at FirstActiveLane are now the incoming
    // values from VectorEarlyExitVPBB.
    for (VPRecipeBase &R : EarlyExitVPBB->phis()) {
      auto *ExitIRI = cast<VPIRPhi>(&R);
      VPValue *IncomingVal =
          ExitIRI->getIncomingValueForBlock(EarlyExitingVPBB);
      VPValue *NewIncoming = IncomingVal;
      if (!isa<VPIRValue>(IncomingVal)) {
        VPBuilder EarlyExitBuilder(VectorEarlyExitVPBB);
        NewIncoming = EarlyExitBuilder.createNaryOp(
            VPInstruction::ExtractLane, {FirstActiveLane, IncomingVal},
            DebugLoc::getUnknown(), "early.exit.value");
      }
      ExitIRI->removeIncomingValueFor(EarlyExitingVPBB);
      ExitIRI->addIncoming(NewIncoming);
    }

    EarlyExitingVPBB->getTerminator()->eraseFromParent();
    VPBlockUtils::disconnectBlocks(EarlyExitingVPBB, EarlyExitVPBB);
    VPBlockUtils::connectBlocks(VectorEarlyExitVPBB, EarlyExitVPBB);
  }

  // Chain through exits: for each exit, check if its condition is true at
  // the first active lane. If so, take that exit; otherwise, try the next.
  // The last exit needs no check since it must be taken if all others fail.
  //
  // For 3 exits (cond.0, cond.1, cond.2), this creates:
  //
  // latch:
  //   ...
  //   EMIT vp<%combined> = logical-or vp<%cond.0>, vp<%cond.1>, vp<%cond.2>
  //   EMIT vp<%combined.freeze> = freeze vp<%combined>
  //   ...
  //
  // vector.early.exit.check:
  //   EMIT vp<%first.lane> = first-active-lane vp<%combined.freeze>
  //   EMIT vp<%at.cond.0> = extract-lane vp<%first.lane>, vp<%cond.0>
  //   EMIT branch-on-cond vp<%at.cond.0>
  // Successor(s): vector.early.exit.0, vector.early.exit.check.0
  //
  // vector.early.exit.check.0:
  //   EMIT vp<%at.cond.1> = extract-lane vp<%first.lane>, vp<%cond.1>
  //   EMIT branch-on-cond vp<%at.cond.1>
  // Successor(s): vector.early.exit.1, vector.early.exit.2
  VPBasicBlock *CurrentBB = DispatchVPBB;
  for (auto [I, Exit] : enumerate(ArrayRef(Exits).drop_back())) {
    VPValue *LaneVal = DispatchBuilder.createNaryOp(
        VPInstruction::ExtractLane, {FirstActiveLane, Exit.CondToExit},
        DebugLoc::getUnknown(), "exit.cond.at.lane");

    // For the last dispatch, branch directly to the last exit on false;
    // otherwise, create a new check block.
    bool IsLastDispatch = (I + 2 == Exits.size());
    VPBasicBlock *FalseBB =
        IsLastDispatch ? VectorEarlyExitVPBBs.back()
                       : Plan.createVPBasicBlock(
                             Twine("vector.early.exit.check.") + Twine(I));

    DispatchBuilder.createNaryOp(VPInstruction::BranchOnCond, {LaneVal});
    CurrentBB->setSuccessors({VectorEarlyExitVPBBs[I], FalseBB});
    VectorEarlyExitVPBBs[I]->setPredecessors({CurrentBB});
    FalseBB->setPredecessors({CurrentBB});

    CurrentBB = FalseBB;
    DispatchBuilder.setInsertPoint(CurrentBB);
  }

  return true;
}

/// This function tries convert extended in-loop reductions to
/// VPExpressionRecipe and clamp the \p Range if it is beneficial and
/// valid. The created recipe must be decomposed to its constituent
/// recipes before execution.
static VPExpressionRecipe *
tryToMatchAndCreateExtendedReduction(VPReductionRecipe *Red, VPCostContext &Ctx,
                                     VFRange &Range) {
  Type *RedTy = Red->getScalarType();
  VPValue *VecOp = Red->getVecOp();

  // We don't handle partial reductions here.
  if (Red->isPartialReduction())
    return nullptr;

  // Clamp the range if using extended-reduction is profitable.
  auto IsExtendedRedValidAndClampRange =
      [&](unsigned Opcode, Instruction::CastOps ExtOpc, Type *SrcTy) -> bool {
    return LoopVectorizationPlanner::getDecisionAndClampRange(
        [&](ElementCount VF) {
          auto *SrcVecTy = cast<VectorType>(toVectorTy(SrcTy, VF));
          TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;

          InstructionCost ExtRedCost = InstructionCost::getInvalid();
          InstructionCost ExtCost =
              cast<VPWidenCastRecipe>(VecOp)->computeCost(VF, Ctx);
          InstructionCost RedCost = Red->computeCost(VF, Ctx);

          assert(!RedTy->isFloatingPointTy() &&
                 "getExtendedReductionCost only supports integer types");
          ExtRedCost = Ctx.TTI.getExtendedReductionCost(
              Opcode, ExtOpc == Instruction::CastOps::ZExt, RedTy, SrcVecTy,
              Red->getFastMathFlagsOrNone(), CostKind);
          return ExtRedCost.isValid() && ExtRedCost < ExtCost + RedCost;
        },
        Range);
  };

  VPValue *A;
  // Match reduce(ext)).
  if (match(VecOp, m_Isa<VPWidenCastRecipe>(m_ZExtOrSExt(m_VPValue(A)))) &&
      IsExtendedRedValidAndClampRange(
          RecurrenceDescriptor::getOpcode(Red->getRecurrenceKind()),
          cast<VPWidenCastRecipe>(VecOp)->getOpcode(), A->getScalarType()))
    return new VPExpressionRecipe(cast<VPWidenCastRecipe>(VecOp), Red);

  return nullptr;
}

/// This function tries convert extended in-loop reductions to
/// VPExpressionRecipe and clamp the \p Range if it is beneficial
/// and valid. The created VPExpressionRecipe must be decomposed to its
/// constituent recipes before execution. Patterns of the
/// VPExpressionRecipe:
///   reduce.add(mul(...)),
///   reduce.add(mul(ext(A), ext(B))),
///   reduce.add(ext(mul(ext(A), ext(B)))).
///   reduce.fadd(fmul(ext(A), ext(B)))
static VPExpressionRecipe *
tryToMatchAndCreateMulAccumulateReduction(VPReductionRecipe *Red,
                                          VPCostContext &Ctx, VFRange &Range) {
  unsigned Opcode = RecurrenceDescriptor::getOpcode(Red->getRecurrenceKind());
  if (Opcode != Instruction::Add && Opcode != Instruction::Sub &&
      Opcode != Instruction::FAdd)
    return nullptr;

  // We don't handle partial reductions here.
  if (Red->isPartialReduction())
    return nullptr;

  Type *RedTy = Red->getScalarType();

  // Clamp the range if using multiply-accumulate-reduction is profitable.
  auto IsMulAccValidAndClampRange =
      [&](VPWidenRecipe *Mul, VPWidenCastRecipe *Ext0, VPWidenCastRecipe *Ext1,
          VPWidenCastRecipe *OuterExt) -> bool {
    return LoopVectorizationPlanner::getDecisionAndClampRange(
        [&](ElementCount VF) {
          TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
          Type *SrcTy = Ext0 ? Ext0->getOperand(0)->getScalarType() : RedTy;
          InstructionCost MulAccCost;

          // getMulAccReductionCost for in-loop reductions does not support
          // mixed or floating-point extends.
          if (Ext0 && Ext1 &&
              (Ext0->getOpcode() != Ext1->getOpcode() ||
               Ext0->getOpcode() == Instruction::CastOps::FPExt))
            return false;

          bool IsZExt =
              !Ext0 || Ext0->getOpcode() == Instruction::CastOps::ZExt;
          auto *SrcVecTy = cast<VectorType>(toVectorTy(SrcTy, VF));
          MulAccCost = Ctx.TTI.getMulAccReductionCost(IsZExt, Opcode, RedTy,
                                                      SrcVecTy, CostKind);

          InstructionCost MulCost = Mul->computeCost(VF, Ctx);
          InstructionCost RedCost = Red->computeCost(VF, Ctx);
          InstructionCost ExtCost = 0;
          if (Ext0)
            ExtCost += Ext0->computeCost(VF, Ctx);
          if (Ext1)
            ExtCost += Ext1->computeCost(VF, Ctx);
          if (OuterExt)
            ExtCost += OuterExt->computeCost(VF, Ctx);

          return MulAccCost.isValid() &&
                 MulAccCost < ExtCost + MulCost + RedCost;
        },
        Range);
  };

  VPValue *VecOp = Red->getVecOp();
  VPRecipeBase *Sub = nullptr;
  VPValue *A, *B;
  VPValue *Tmp = nullptr;

  if (RedTy->isFloatingPointTy())
    return nullptr;

  // Sub reductions could have a sub between the add reduction and vec op.
  if (match(VecOp, m_Sub(m_ZeroInt(), m_VPValue(Tmp)))) {
    Sub = VecOp->getDefiningRecipe();
    VecOp = Tmp;
  }

  // If ValB is a constant and can be safely extended, truncate it to the same
  // type as ExtA's operand, then extend it to the same type as ExtA. This
  // creates two uniform extends that can more easily be matched by the rest of
  // the bundling code. The ExtB reference, ValB and operand 1 of Mul are all
  // replaced with the new extend of the constant.
  auto ExtendAndReplaceConstantOp = [](VPWidenCastRecipe *ExtA,
                                       VPWidenCastRecipe *&ExtB, VPValue *&ValB,
                                       VPWidenRecipe *Mul) {
    if (!ExtA || ExtB || !isa<VPIRValue>(ValB))
      return;
    Type *NarrowTy = ExtA->getOperand(0)->getScalarType();
    Instruction::CastOps ExtOpc = ExtA->getOpcode();
    const APInt *Const;
    if (!match(ValB, m_APInt(Const)) ||
        !llvm::canConstantBeExtended(
            Const, NarrowTy, TTI::getPartialReductionExtendKind(ExtOpc)))
      return;
    // The truncate ensures that the type of each extended operand is the
    // same, and it's been proven that the constant can be extended from
    // NarrowTy safely. Necessary since ExtA's extended operand would be
    // e.g. an i8, while the const will likely be an i32. This will be
    // elided by later optimisations.
    VPBuilder Builder(Mul);
    auto *Trunc =
        Builder.createWidenCast(Instruction::CastOps::Trunc, ValB, NarrowTy);
    Type *WideTy = ExtA->getScalarType();
    ValB = ExtB = Builder.createWidenCast(ExtOpc, Trunc, WideTy);
    Mul->setOperand(1, ExtB);
  };

  // Try to match reduce.add(mul(...)).
  if (match(VecOp, m_Mul(m_VPValue(A), m_VPValue(B)))) {
    auto *RecipeA = dyn_cast<VPWidenCastRecipe>(A);
    auto *RecipeB = dyn_cast<VPWidenCastRecipe>(B);
    auto *Mul = cast<VPWidenRecipe>(VecOp);

    // Convert reduce.add(mul(ext, const)) to reduce.add(mul(ext, ext(const)))
    ExtendAndReplaceConstantOp(RecipeA, RecipeB, B, Mul);

    // Match reduce.add/sub(mul(ext, ext)).
    if (RecipeA && RecipeB && match(RecipeA, m_ZExtOrSExt(m_VPValue())) &&
        match(RecipeB, m_ZExtOrSExt(m_VPValue())) &&
        IsMulAccValidAndClampRange(Mul, RecipeA, RecipeB, nullptr)) {
      if (Sub)
        return new VPExpressionRecipe(RecipeA, RecipeB, Mul,
                                      cast<VPWidenRecipe>(Sub), Red);
      return new VPExpressionRecipe(RecipeA, RecipeB, Mul, Red);
    }
    // TODO: Add an expression type for this variant with a negated mul
    if (!Sub && IsMulAccValidAndClampRange(Mul, nullptr, nullptr, nullptr))
      return new VPExpressionRecipe(Mul, Red);
  }
  // TODO: Add an expression type for negated versions of other expression
  // variants.
  if (Sub)
    return nullptr;

  // Match reduce.add(ext(mul(A, B))).
  if (match(VecOp, m_ZExtOrSExt(m_Mul(m_VPValue(A), m_VPValue(B))))) {
    auto *Ext = cast<VPWidenCastRecipe>(VecOp);
    auto *Mul = cast<VPWidenRecipe>(Ext->getOperand(0));
    auto *Ext0 = dyn_cast<VPWidenCastRecipe>(A);
    auto *Ext1 = dyn_cast<VPWidenCastRecipe>(B);

    // reduce.add(ext(mul(ext, const)))
    // -> reduce.add(ext(mul(ext, ext(const))))
    ExtendAndReplaceConstantOp(Ext0, Ext1, B, Mul);

    // reduce.add(ext(mul(ext(A), ext(B))))
    // -> reduce.add(mul(wider_ext(A), wider_ext(B)))
    // The inner extends must either have the same opcode as the outer extend or
    // be the same, in which case the multiply can never result in a negative
    // value and the outer extend can be folded away by doing wider
    // extends for the operands of the mul.
    if (Ext0 && Ext1 &&
        (Ext->getOpcode() == Ext0->getOpcode() || Ext0 == Ext1) &&
        Ext0->getOpcode() == Ext1->getOpcode() &&
        IsMulAccValidAndClampRange(Mul, Ext0, Ext1, Ext) && Mul->hasOneUse()) {
      auto *NewExt0 = new VPWidenCastRecipe(
          Ext0->getOpcode(), Ext0->getOperand(0), Ext->getScalarType(), nullptr,
          *Ext0, *Ext0, Ext0->getDebugLoc());
      NewExt0->insertBefore(Ext0);

      VPWidenCastRecipe *NewExt1 = NewExt0;
      if (Ext0 != Ext1) {
        NewExt1 = new VPWidenCastRecipe(Ext1->getOpcode(), Ext1->getOperand(0),
                                        Ext->getScalarType(), nullptr, *Ext1,
                                        *Ext1, Ext1->getDebugLoc());
        NewExt1->insertBefore(Ext1);
      }
      auto *NewMul = Mul->cloneWithOperands({NewExt0, NewExt1});
      NewMul->insertBefore(Mul);
      Ext->replaceAllUsesWith(NewMul);
      Ext->eraseFromParent();
      Mul->eraseFromParent();
      return new VPExpressionRecipe(NewExt0, NewExt1, NewMul, Red);
    }
  }
  return nullptr;
}

/// This function tries to create abstract recipes from the reduction recipe for
/// following optimizations and cost estimation.
static void tryToCreateAbstractReductionRecipe(VPReductionRecipe *Red,
                                               VPCostContext &Ctx,
                                               VFRange &Range) {
  // Creation of VPExpressions for partial reductions is entirely handled in
  // transformToPartialReduction.
  if (Red->isPartialReduction())
    return;

  VPExpressionRecipe *AbstractR = nullptr;
  auto IP = std::next(Red->getIterator());
  auto *VPBB = Red->getParent();
  if (auto *MulAcc = tryToMatchAndCreateMulAccumulateReduction(Red, Ctx, Range))
    AbstractR = MulAcc;
  else if (auto *ExtRed = tryToMatchAndCreateExtendedReduction(Red, Ctx, Range))
    AbstractR = ExtRed;
  // Cannot create abstract inloop reduction recipes.
  if (!AbstractR)
    return;

  AbstractR->insertBefore(*VPBB, IP);
  Red->replaceAllUsesWith(AbstractR);
}

void VPlanTransforms::convertToAbstractRecipes(VPlan &Plan, VPCostContext &Ctx,
                                               VFRange &Range) {
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_deep(Plan.getVectorLoopRegion()))) {
    for (VPReductionRecipe &Red :
         make_early_inc_range(make_isa_range<VPReductionRecipe>(*VPBB)))
      tryToCreateAbstractReductionRecipe(&Red, Ctx, Range);
  }
}

// Collect common metadata from a group of replicate recipes by intersecting
// metadata from all recipes in the group.
static VPIRMetadata getCommonMetadata(ArrayRef<VPReplicateRecipe *> Recipes) {
  VPIRMetadata CommonMetadata = *Recipes.front();
  for (VPReplicateRecipe *Recipe : drop_begin(Recipes))
    CommonMetadata.intersect(*Recipe);
  // The recipe using the common metadata is not predicated, so it does not
  // share the group's execution frequency.
  CommonMetadata.clearExecutionFrequency();
  return CommonMetadata;
}

template <unsigned Opcode>
static SmallVector<SmallVector<VPReplicateRecipe *, 4>>
collectComplementaryPredicatedMemOps(VPlan &Plan,
                                     PredicatedScalarEvolution &PSE,
                                     const Loop *L) {
  static_assert(Opcode == Instruction::Load || Opcode == Instruction::Store,
                "Only Load and Store opcodes supported");
  [[maybe_unused]] constexpr bool IsLoad = (Opcode == Instruction::Load);

  // For each address, collect operations with the same or complementary masks.
  SmallVector<SmallVector<VPReplicateRecipe *, 4>> AllGroups;
  auto Groups = collectGroupedReplicateMemOps<Opcode>(
      Plan, PSE, L,
      [](VPReplicateRecipe *RepR) { return RepR->isPredicated(); });
  for (auto Recipes : Groups) {
    if (Recipes.size() < 2)
      continue;

    assert(all_equal(
               map_range(Recipes, bind_back<getLoadStoreValueType>(IsLoad))) &&
           "Expected all recipes in group to have the same load-store type");

    // Collect groups with the same or complementary masks.
    for (VPReplicateRecipe *&RecipeI : Recipes) {
      if (!RecipeI)
        continue;

      VPValue *MaskI = RecipeI->getMask();
      SmallVector<VPReplicateRecipe *, 4> Group;
      Group.push_back(RecipeI);
      RecipeI = nullptr;

      // Find all operations with the same or complementary masks.
      bool HasComplementaryMask = false;
      for (VPReplicateRecipe *&RecipeJ : Recipes) {
        if (!RecipeJ)
          continue;

        VPValue *MaskJ = RecipeJ->getMask();
        // Check if any operation in the group has a complementary mask with
        // another, that is M1 == NOT(M2) or M2 == NOT(M1).
        HasComplementaryMask |= match(MaskI, m_Not(m_Specific(MaskJ))) ||
                                match(MaskJ, m_Not(m_Specific(MaskI)));
        Group.push_back(RecipeJ);
        RecipeJ = nullptr;
      }

      if (HasComplementaryMask) {
        assert(Group.size() >= 2 && "must have at least 2 entries");
        AllGroups.push_back(std::move(Group));
      }
    }
  }

  return AllGroups;
}

// Find the recipe with minimum alignment in the group.
template <typename InstType>
static VPReplicateRecipe *
findRecipeWithMinAlign(ArrayRef<VPReplicateRecipe *> Group) {
  return *min_element(Group, [](VPReplicateRecipe *A, VPReplicateRecipe *B) {
    return cast<InstType>(A->getUnderlyingInstr())->getAlign() <
           cast<InstType>(B->getUnderlyingInstr())->getAlign();
  });
}

void VPlanTransforms::hoistPredicatedLoads(VPlan &Plan,
                                           PredicatedScalarEvolution &PSE,
                                           const Loop *L) {
  auto Groups =
      collectComplementaryPredicatedMemOps<Instruction::Load>(Plan, PSE, L);
  if (Groups.empty())
    return;

  // Process each group of loads.
  for (auto &Group : Groups) {
    // Try to use the earliest (most dominating) load to replace all others.
    VPReplicateRecipe *EarliestLoad = Group[0];
    VPBasicBlock *FirstBB = EarliestLoad->getParent();
    VPBasicBlock *LastBB = Group.back()->getParent();

    // Check that the load doesn't alias with stores between first and last.
    auto LoadLoc = vputils::getMemoryLocation(*EarliestLoad);
    if (!LoadLoc || !canHoistOrSinkWithNoAliasCheck(*LoadLoc, FirstBB, LastBB))
      continue;

    // Collect common metadata from all loads in the group.
    VPIRMetadata CommonMetadata = getCommonMetadata(Group);

    // Find the load with minimum alignment to use.
    auto *LoadWithMinAlign = findRecipeWithMinAlign<LoadInst>(Group);

    bool IsSingleScalar = EarliestLoad->isSingleScalar();
    assert(all_of(Group,
                  [IsSingleScalar](VPReplicateRecipe *R) {
                    return R->isSingleScalar() == IsSingleScalar;
                  }) &&
           "all members in group must agree on IsSingleScalar");

    // Create an unpredicated version of the earliest load with common
    // metadata.
    auto *UnpredicatedLoad = new VPReplicateRecipe(
        LoadWithMinAlign->getUnderlyingInstr(), {EarliestLoad->getOperand(0)},
        IsSingleScalar, /*Mask=*/nullptr, *EarliestLoad, CommonMetadata);

    UnpredicatedLoad->insertBefore(EarliestLoad);

    // Replace all loads in the group with the unpredicated load.
    for (VPReplicateRecipe *Load : Group) {
      Load->replaceAllUsesWith(UnpredicatedLoad);
      Load->eraseFromParent();
    }
  }
}

static bool
canSinkStoreWithNoAliasCheck(ArrayRef<VPReplicateRecipe *> StoresToSink,
                             PredicatedScalarEvolution &PSE, const Loop &L) {
  auto StoreLoc = vputils::getMemoryLocation(*StoresToSink.front());
  if (!StoreLoc || !StoreLoc->AATags.Scope)
    return false;

  // When sinking a group of stores, all members of the group alias each other.
  // Skip them during the alias checks.
  VPBasicBlock *FirstBB = StoresToSink.front()->getParent();
  VPBasicBlock *LastBB = StoresToSink.back()->getParent();
  SinkStoreInfo SinkInfo(StoresToSink, *StoresToSink[0], PSE, L);
  return canHoistOrSinkWithNoAliasCheck(*StoreLoc, FirstBB, LastBB, SinkInfo);
}

void VPlanTransforms::sinkPredicatedStores(VPlan &Plan,
                                           PredicatedScalarEvolution &PSE,
                                           const Loop *L) {
  auto Groups =
      collectComplementaryPredicatedMemOps<Instruction::Store>(Plan, PSE, L);
  if (Groups.empty())
    return;

  for (auto &Group : Groups) {
    if (!canSinkStoreWithNoAliasCheck(Group, PSE, *L))
      continue;

    // Use the last (most dominated) store's location for the unconditional
    // store.
    VPReplicateRecipe *LastStore = Group.back();
    VPBasicBlock *InsertBB = LastStore->getParent();

    // Collect common alias metadata from all stores in the group.
    VPIRMetadata CommonMetadata = getCommonMetadata(Group);

    // Build select chain for stored values.
    VPValue *SelectedValue = Group[0]->getOperand(0);
    VPBuilder Builder(InsertBB, LastStore->getIterator());

    bool IsSingleScalar = Group[0]->isSingleScalar();
    for (unsigned I = 1; I < Group.size(); ++I) {
      assert(IsSingleScalar == Group[I]->isSingleScalar() &&
             "all members in group must agree on IsSingleScalar");
      VPValue *Mask = Group[I]->getMask();
      VPValue *Value = Group[I]->getOperand(0);
      SelectedValue = Builder.createSelect(
          Mask, Value, SelectedValue, Group[I]->getDebugLoc(), "",
          VPIRFlags::getDefaultFlags(Instruction::Select,
                                     Value->getScalarType()));
    }

    // Find the store with minimum alignment to use.
    auto *StoreWithMinAlign = findRecipeWithMinAlign<StoreInst>(Group);

    // Create unconditional store with selected value and common metadata.
    auto *UnpredicatedStore = new VPReplicateRecipe(
        StoreWithMinAlign->getUnderlyingInstr(),
        {SelectedValue, LastStore->getOperand(1)}, IsSingleScalar,
        /*Mask=*/nullptr, *LastStore, CommonMetadata);
    UnpredicatedStore->insertBefore(*InsertBB, LastStore->getIterator());

    // Remove all predicated stores from the group.
    for (VPReplicateRecipe *Store : Group)
      Store->eraseFromParent();
  }
}

/// Returns true if \p V is VPWidenLoadRecipe or VPInterleaveRecipe that can be
/// converted to a narrower recipe. \p V is used by a wide recipe that feeds a
/// store interleave group at index \p Idx, \p WideMember0 is the recipe feeding
/// the same interleave group at index 0. A VPWidenLoadRecipe can be narrowed to
/// an index-independent load if it feeds all wide ops at all indices (\p OpV
/// must be the operand at index \p OpIdx for both the recipe at lane 0, \p
/// WideMember0). A VPInterleaveRecipe can be narrowed to a wide load, if \p V
/// is defined at \p Idx of a load interleave group.
/// A live-in or recipe defined outside the loop region can be converted, if it
/// is the same across all lanes, or we can create a BuildVector for it.
static bool canNarrowLoad(VPSingleDefRecipe *WideMember0, unsigned OpIdx,
                          VPValue *OpV, unsigned Idx, bool IsScalable) {
  VPValue *Member0Op = WideMember0->getOperand(OpIdx);
  if (Member0Op->isDefinedOutsideLoopRegions()) {
    // Operand matches Member0, broadcast across all fields for both live-ins
    // and recipes.
    if (Member0Op == OpV)
      return true;
    // Otherwise distinct per-field VPValues are assembled into a BuildVector.
    return !IsScalable && OpV->isDefinedOutsideLoopRegions() &&
           OpV->getScalarType() == Member0Op->getScalarType();
  }
  VPRecipeBase *Member0OpR = Member0Op->getDefiningRecipe();
  if (auto *W = dyn_cast<VPWidenLoadRecipe>(Member0OpR))
    // For scalable VFs, the narrowed plan processes vscale iterations at once,
    // so a shared wide load cannot be narrowed to a uniform scalar; bail out.
    return !IsScalable && !W->getMask() && W->isConsecutive() &&
           Member0Op == OpV;
  if (auto *IR = dyn_cast<VPInterleaveRecipe>(Member0OpR))
    return IR->getInterleaveGroup()->isFull() && IR->getVPValue(Idx) == OpV;
  return false;
}

static bool canNarrowOps(ArrayRef<VPValue *> Ops, bool IsScalable) {
  SmallVector<VPValue *> Ops0;
  auto *WideMember0 = dyn_cast<VPRecipeWithIRFlags>(Ops[0]);
  if (!WideMember0)
    return false;
  for (VPValue *V : Ops) {
    if (!isa<VPWidenRecipe, VPWidenCastRecipe>(V))
      return false;
    auto *R = cast<VPRecipeWithIRFlags>(V);
    if (vputils::getOpcode(R) != vputils::getOpcode(WideMember0))
      return false;
    if (R->getScalarType() != WideMember0->getScalarType())
      return false;
    if (R->hasPredicate() && R->getPredicate() != WideMember0->getPredicate())
      return false;
  }

  for (unsigned Idx = 0; Idx != WideMember0->getNumOperands(); ++Idx) {
    SmallVector<VPValue *> OpsI;
    for (VPValue *Op : Ops)
      OpsI.push_back(Op->getDefiningRecipe()->getOperand(Idx));

    if (canNarrowOps(OpsI, IsScalable))
      continue;

    if (any_of(enumerate(OpsI), [WideMember0, Idx, IsScalable](const auto &P) {
          const auto &[OpIdx, OpV] = P;
          return !canNarrowLoad(WideMember0, Idx, OpV, OpIdx, IsScalable);
        }))
      return false;
  }

  return true;
}

/// Returns VF from \p VFs if \p IR is a full interleave group with factor and
/// number of members both equal to VF. The interleave group must also access
/// the full vector width.
static std::optional<ElementCount>
isConsecutiveInterleaveGroup(VPInterleaveRecipe *InterleaveR,
                             ArrayRef<ElementCount> VFs,
                             const TargetTransformInfo &TTI) {
  if (!InterleaveR || InterleaveR->getMask())
    return std::nullopt;

  Type *GroupElementTy = nullptr;
  if (InterleaveR->getStoredValues().empty()) {
    GroupElementTy = InterleaveR->getVPValue(0)->getScalarType();
    if (!all_of(InterleaveR->definedValues(), [GroupElementTy](VPValue *Op) {
          return Op->getScalarType() == GroupElementTy;
        }))
      return std::nullopt;
  } else {
    GroupElementTy = InterleaveR->getStoredValues()[0]->getScalarType();
    if (!all_of(InterleaveR->getStoredValues(), [GroupElementTy](VPValue *Op) {
          return Op->getScalarType() == GroupElementTy;
        }))
      return std::nullopt;
  }

  auto IG = InterleaveR->getInterleaveGroup();
  if (IG->getFactor() != IG->getNumMembers())
    return std::nullopt;

  auto GetVectorBitWidthForVF = [&TTI](ElementCount VF) {
    TypeSize Size = TTI.getRegisterBitWidth(
        VF.isFixed() ? TargetTransformInfo::RGK_FixedWidthVector
                     : TargetTransformInfo::RGK_ScalableVector);
    assert(Size.isScalable() == VF.isScalable() &&
           "if Size is scalable, VF must be scalable and vice versa");
    return Size.getKnownMinValue();
  };

  for (ElementCount VF : VFs) {
    unsigned MinVal = VF.getKnownMinValue();
    unsigned GroupSize = GroupElementTy->getScalarSizeInBits() * MinVal;
    if (IG->getFactor() == MinVal && GroupSize == GetVectorBitWidthForVF(VF))
      return {VF};
  }
  return std::nullopt;
}

/// Returns true if \p VPValue is a narrow VPValue.
static bool isAlreadyNarrow(VPValue *VPV) {
  if (isa<VPIRValue>(VPV))
    return true;
  auto *RepR = dyn_cast<VPReplicateRecipe>(VPV);
  return RepR && RepR->isSingleScalar();
}

// Convert the wide recipes defining the VPValues in \p Members feeding an
// interleave group to a single narrow variant. The first member is reused as
// the narrowed recipe. BuildVectors for live-in operands are inserted into \p
// Preheader.
static VPValue *narrowInterleaveGroupOp(ArrayRef<VPValue *> Members,
                                        SmallPtrSetImpl<VPValue *> &NarrowedOps,
                                        VPBasicBlock *Preheader) {
  VPValue *V = Members.front();
  if (NarrowedOps.contains(V))
    return V;

  if (V->isDefinedOutsideLoopRegions()) {
    assert(all_of(Members,
                  [V](VPValue *M) {
                    return M->isDefinedOutsideLoopRegions() &&
                           M->getScalarType() == V->getScalarType();
                  }) &&
           "expected distinct loop-invariant values of matching scalar type");
    auto *BV = new VPInstruction(VPInstruction::BuildVector, Members);
    Preheader->appendRecipe(BV);
    NarrowedOps.insert(BV);
    return BV;
  }

  if (isAlreadyNarrow(V))
    return V;

  VPRecipeBase *R = V->getDefiningRecipe();
  if (isa<VPWidenRecipe, VPWidenCastRecipe>(R)) {
    auto *WideMember0 = cast<VPRecipeWithIRFlags>(R);
    for (VPValue *Member : Members.drop_front())
      WideMember0->intersectFlags(*cast<VPRecipeWithIRFlags>(Member));
    for (unsigned Idx = 0, E = WideMember0->getNumOperands(); Idx != E; ++Idx) {
      SmallVector<VPValue *> OpsI;
      for (VPValue *Member : Members)
        OpsI.push_back(Member->getDefiningRecipe()->getOperand(Idx));
      WideMember0->setOperand(
          Idx, narrowInterleaveGroupOp(OpsI, NarrowedOps, Preheader));
    }
    return V;
  }

  if (auto *LoadGroup = dyn_cast<VPInterleaveRecipe>(R)) {
    // Narrow interleave group to wide load, as transformed VPlan will only
    // process one original iteration.
    auto *LI = cast<LoadInst>(LoadGroup->getInterleaveGroup()->getInsertPos());
    auto *L = VPBuilder(LoadGroup).createWidenLoad(
        *LI, LoadGroup->getAddr(), LoadGroup->getMask(), /*Consecutive=*/true,
        *LoadGroup, LoadGroup->getDebugLoc());
    NarrowedOps.insert(L);
    return L;
  }

  if (auto *RepR = dyn_cast<VPReplicateRecipe>(R)) {
    assert(RepR->isSingleScalar() && RepR->getOpcode() == Instruction::Load &&
           "must be a single scalar load");
    NarrowedOps.insert(RepR);
    return RepR;
  }

  auto *WideLoad = cast<VPWidenLoadRecipe>(R);
  VPValue *PtrOp = WideLoad->getAddr();
  if (auto *VecPtr = dyn_cast<VPVectorPointerRecipe>(PtrOp))
    PtrOp = VecPtr->getOperand(0);
  // Narrow wide load to uniform scalar load, as transformed VPlan will only
  // process one original iteration.
  auto *N = new VPReplicateRecipe(&WideLoad->getIngredient(), {PtrOp},
                                  /*IsUniform*/ true,
                                  /*Mask*/ nullptr, {}, *WideLoad);
  N->insertBefore(WideLoad);
  NarrowedOps.insert(N);
  return N;
}

std::unique_ptr<VPlan>
VPlanTransforms::narrowInterleaveGroups(VPlan &Plan,
                                        const TargetTransformInfo &TTI) {
  VPRegionBlock *VectorLoop = Plan.getVectorLoopRegion();

  if (!VectorLoop)
    return nullptr;

  // Only handle single-block loops for now.
  if (VectorLoop->getEntryBasicBlock() != VectorLoop->getExitingBasicBlock())
    return nullptr;

  // Skip plans when we may not be able to properly narrow.
  VPBasicBlock *Exiting = VectorLoop->getExitingBasicBlock();
  if (!match(&Exiting->back(), m_BranchOnCount()))
    return nullptr;

  assert(match(&Exiting->back(),
               m_BranchOnCount(m_Add(m_VPValue(), m_Specific(&Plan.getVFxUF())),
                               m_Specific(&Plan.getVectorTripCount()))) &&
         "unexpected branch-on-count");

  SmallVector<VPInterleaveRecipe *> StoreGroups;
  std::optional<ElementCount> VFToOptimize;
  for (auto &R : *VectorLoop->getEntryBasicBlock()) {
    if (isa<VPDerivedIVRecipe, VPScalarIVStepsRecipe>(&R) &&
        vputils::onlyFirstLaneUsed(cast<VPSingleDefRecipe>(&R)))
      continue;

    // Bail out on recipes not supported at the moment:
    //  * phi recipes other than the canonical induction
    //  * recipes writing to memory except interleave groups
    // Only support plans with a canonical induction phi.
    if (R.isPhi())
      return nullptr;

    auto *InterleaveR = dyn_cast<VPInterleaveRecipe>(&R);
    if (R.mayWriteToMemory() && !InterleaveR)
      return nullptr;

    // Bail out if any recipe defines a vector value used outside the
    // vector loop region.
    if (any_of(R.definedValues(), [&](VPValue *V) {
          return any_of(V->users(), [&](VPUser *U) {
            auto *UR = cast<VPRecipeBase>(U);
            return UR->getParent()->getParent() != VectorLoop;
          });
        }))
      return nullptr;

    // All other ops are allowed, but we reject uses that cannot be converted
    // when checking all allowed consumers (store interleave groups) below.
    if (!InterleaveR)
      continue;

    // Try to find a single VF, where all interleave groups are consecutive and
    // saturate the full vector width. If we already have a candidate VF, check
    // if it is applicable for the current InterleaveR, otherwise look for a
    // suitable VF across the Plan's VFs.
    SmallVector<ElementCount> VFs =
        VFToOptimize ? SmallVector<ElementCount>({*VFToOptimize})
                     : to_vector(Plan.vectorFactors());
    std::optional<ElementCount> NarrowedVF =
        isConsecutiveInterleaveGroup(InterleaveR, VFs, TTI);
    if (!NarrowedVF || (VFToOptimize && NarrowedVF != VFToOptimize))
      return nullptr;
    VFToOptimize = NarrowedVF;

    // Skip read interleave groups.
    if (InterleaveR->getStoredValues().empty())
      continue;

    // Narrow interleave groups, if all operands are already matching narrow
    // ops.
    auto *Member0 = InterleaveR->getStoredValues()[0];
    if (isAlreadyNarrow(Member0) &&
        all_of(InterleaveR->getStoredValues(), equal_to(Member0))) {
      StoreGroups.push_back(InterleaveR);
      continue;
    }

    // For now, we only support full interleave groups storing load interleave
    // groups.
    if (all_of(enumerate(InterleaveR->getStoredValues()), [](auto Op) {
          VPRecipeBase *DefR = Op.value()->getDefiningRecipe();
          if (!DefR)
            return false;
          auto *IR = dyn_cast<VPInterleaveRecipe>(DefR);
          return IR && IR->getInterleaveGroup()->isFull() &&
                 IR->getVPValue(Op.index()) == Op.value();
        })) {
      StoreGroups.push_back(InterleaveR);
      continue;
    }

    // Check if all values feeding InterleaveR are matching wide recipes, which
    // operands that can be narrowed.
    if (!canNarrowOps(InterleaveR->getStoredValues(),
                      VFToOptimize->isScalable()))
      return nullptr;
    StoreGroups.push_back(InterleaveR);
  }

  if (StoreGroups.empty())
    return nullptr;

  VPBasicBlock *MiddleVPBB = Plan.getMiddleBlock();
  bool RequiresScalarEpilogue =
      MiddleVPBB->getNumSuccessors() == 1 &&
      MiddleVPBB->getSingleSuccessor() == Plan.getScalarPreheader();
  // Bail out for tail-folding (middle block with a single successor to exit).
  if (MiddleVPBB->getNumSuccessors() != 2 && !RequiresScalarEpilogue)
    return nullptr;

  // All interleave groups in Plan can be narrowed for VFToOptimize. Split the
  // original Plan into 2: a) a new clone which contains all VFs of Plan, except
  // VFToOptimize, and b) the original Plan with VFToOptimize as single VF.
  // TODO: Handle cases where only some interleave groups can be narrowed.
  std::unique_ptr<VPlan> NewPlan;
  if (size(Plan.vectorFactors()) != 1) {
    NewPlan = std::unique_ptr<VPlan>(Plan.duplicate());
    Plan.setVF(*VFToOptimize);
    NewPlan->removeVF(*VFToOptimize);
  }

  // Convert InterleaveGroup \p R to a single VPWidenLoadRecipe.
  SmallPtrSet<VPValue *, 4> NarrowedOps;
  VPBasicBlock *Preheader = Plan.getVectorPreheader();
  // Narrow operation tree rooted at store groups.
  for (auto *StoreGroup : StoreGroups) {
    VPValue *Res = narrowInterleaveGroupOp(StoreGroup->getStoredValues(),
                                           NarrowedOps, Preheader);
    auto *SI =
        cast<StoreInst>(StoreGroup->getInterleaveGroup()->getInsertPos());
    VPBuilder(StoreGroup)
        .createWidenStore(*SI, StoreGroup->getAddr(), Res, nullptr,
                          /*Consecutive=*/true, *StoreGroup,
                          StoreGroup->getDebugLoc());
    StoreGroup->eraseFromParent();
  }

  // Adjust induction to reflect that the transformed plan only processes one
  // original iteration.
  VPInstruction *CanIVInc = vputils::findCanonicalIVIncrement(Plan);
  Type *CanIVTy = VectorLoop->getCanonicalIVType();
  VPBasicBlock *VectorPH = Plan.getVectorPreheader();
  VPBuilder PHBuilder(VectorPH, VectorPH->getFirstNonPhi());

  VPValue *UF = &Plan.getUF();
  VPValue *Step;
  if (VFToOptimize->isScalable()) {
    VPValue *VScale =
        PHBuilder.createElementCount(CanIVTy, ElementCount::getScalable(1));
    Step = PHBuilder.createOverflowingOp(Instruction::Mul, {VScale, UF},
                                         {true, false});
    Plan.getVF().replaceAllUsesWith(VScale);
  } else {
    Step = UF;
    Plan.getVF().replaceAllUsesWith(Plan.getConstantInt(CanIVTy, 1));
  }
  // Materialize vector trip count with the narrowed step.
  materializeVectorTripCount(Plan, VectorPH, /*TailByMasking=*/false,
                             RequiresScalarEpilogue, Step);

  CanIVInc->setOperand(1, Step);
  Plan.getVFxUF().replaceAllUsesWith(Step);

  removeDeadRecipes(Plan);
  assert(none_of(*VectorLoop->getEntryBasicBlock(),
                 IsaPred<VPVectorPointerRecipe>) &&
         "All VPVectorPointerRecipes should have been removed");
  return NewPlan;
}

void VPlanTransforms::adjustFirstOrderRecurrenceMiddleUsers(VPlan &Plan,
                                                            VFRange &Range) {
  VPRegionBlock *VectorRegion = Plan.getVectorLoopRegion();
  auto *MiddleVPBB = Plan.getMiddleBlock();
  VPBuilder MiddleBuilder(MiddleVPBB, MiddleVPBB->getFirstNonPhi());

  auto IsScalableOne = [](ElementCount VF) -> bool {
    return VF == ElementCount::getScalable(1);
  };

  for (VPFirstOrderRecurrencePHIRecipe &FOR :
       make_isa_range<VPFirstOrderRecurrencePHIRecipe>(
           VectorRegion->getEntryBasicBlock()->phis())) {
    assert(VectorRegion->getSingleSuccessor() == Plan.getMiddleBlock() &&
           "Cannot handle loops with uncountable early exits");

    // Find the existing splice for this FOR, created in
    // createHeaderPhiRecipes. All uses of FOR have already been replaced with
    // RecurSplice there; only RecurSplice itself still references FOR.
    auto *RecurSplice =
        findUserOf<VPInstruction::FirstOrderRecurrenceSplice>(&FOR);
    assert(RecurSplice && "expected FirstOrderRecurrenceSplice");

    // For VF vscale x 1, if vscale = 1, we are unable to extract the
    // penultimate value of the recurrence. Instead we rely on the existing
    // extract of the last element from the result of
    // VPInstruction::FirstOrderRecurrenceSplice.
    // TODO: Consider vscale_range info and UF.
    if (any_of(RecurSplice->users(),
               [](VPUser *U) { return !cast<VPRecipeBase>(U)->getRegion(); }) &&
        LoopVectorizationPlanner::getDecisionAndClampRange(IsScalableOne,
                                                           Range))
      return;

    // This is the second phase of vectorizing first-order recurrences, creating
    // extracts for users outside the loop. An overview of the transformation is
    // described below. Suppose we have the following loop with some use after
    // the loop of the last a[i-1],
    //
    //   for (int i = 0; i < n; ++i) {
    //     t = a[i - 1];
    //     b[i] = a[i] - t;
    //   }
    //   use t;
    //
    // There is a first-order recurrence on "a". For this loop, the shorthand
    // scalar IR looks like:
    //
    //   scalar.ph:
    //     s.init = a[-1]
    //     br scalar.body
    //
    //   scalar.body:
    //     i = phi [0, scalar.ph], [i+1, scalar.body]
    //     s1 = phi [s.init, scalar.ph], [s2, scalar.body]
    //     s2 = a[i]
    //     b[i] = s2 - s1
    //     br cond, scalar.body, exit.block
    //
    //   exit.block:
    //     use = lcssa.phi [s1, scalar.body]
    //
    // In this example, s1 is a recurrence because it's value depends on the
    // previous iteration. In the first phase of vectorization, we created a
    // VPFirstOrderRecurrencePHIRecipe v1 for s1. Now we create the extracts
    // for users in the scalar preheader and exit block.
    //
    //   vector.ph:
    //     v_init = vector(..., ..., ..., a[-1])
    //     br vector.body
    //
    //   vector.body
    //     i = phi [0, vector.ph], [i+4, vector.body]
    //     v1 = phi [v_init, vector.ph], [v2, vector.body]
    //     v2 = a[i, i+1, i+2, i+3]
    //     v1' = splice(v1(3), v2(0, 1, 2))
    //     b[i, i+1, i+2, i+3] = v2 - v1'
    //     br cond, vector.body, middle.block
    //
    //   middle.block:
    //     vector.recur.extract.for.phi = v2(2)
    //     vector.recur.extract = v2(3)
    //     br cond, scalar.ph, exit.block
    //
    //   scalar.ph:
    //     scalar.recur.init = phi [vector.recur.extract, middle.block],
    //                             [s.init, otherwise]
    //     br scalar.body
    //
    //   scalar.body:
    //     i = phi [0, scalar.ph], [i+1, scalar.body]
    //     s1 = phi [scalar.recur.init, scalar.ph], [s2, scalar.body]
    //     s2 = a[i]
    //     b[i] = s2 - s1
    //     br cond, scalar.body, exit.block
    //
    //   exit.block:
    //     lo = lcssa.phi [s1, scalar.body],
    //                    [vector.recur.extract.for.phi, middle.block]
    //
    // Update extracts of the splice in the middle block: they extract the
    // penultimate element of the recurrence.
    for (VPRecipeBase &R : make_early_inc_range(
             make_range(MiddleVPBB->getFirstNonPhi(), MiddleVPBB->end()))) {
      if (!match(&R, m_ExtractLastLaneOfLastPart(m_Specific(RecurSplice))))
        continue;

      auto *ExtractR = cast<VPInstruction>(&R);
      VPValue *PenultimateElement = MiddleBuilder.createNaryOp(
          VPInstruction::ExtractPenultimateElement, RecurSplice->getOperand(1),
          {}, "vector.recur.extract.for.phi");
      for (VPUser *ExitU : to_vector(ExtractR->users())) {
        if (auto *ExitPhi = dyn_cast<VPIRPhi>(ExitU))
          ExitPhi->replaceUsesOfWith(ExtractR, PenultimateElement);
      }
    }
  }
}

/// Check if \p V is a binary expression of a widened IV and a loop-invariant
/// value. Returns the widened IV if found, nullptr otherwise.
static VPWidenIntOrFpInductionRecipe *getExpressionIV(VPValue *V) {
  auto *BinOp = dyn_cast<VPWidenRecipe>(V);
  if (!BinOp || !Instruction::isBinaryOp(BinOp->getOpcode()) ||
      Instruction::isIntDivRem(BinOp->getOpcode()))
    return nullptr;

  VPValue *WidenIVCandidate = BinOp->getOperand(0);
  VPValue *InvariantCandidate = BinOp->getOperand(1);
  if (!isa<VPWidenIntOrFpInductionRecipe>(WidenIVCandidate))
    std::swap(WidenIVCandidate, InvariantCandidate);

  if (!InvariantCandidate->isDefinedOutsideLoopRegions())
    return nullptr;

  return dyn_cast<VPWidenIntOrFpInductionRecipe>(WidenIVCandidate);
}

/// Create a scalar version of \p BinOp, with its \p WidenIV operand replaced
/// by \p ScalarIV, and place it after \p ScalarIV's defining recipe.
static VPValue *cloneBinOpForScalarIV(VPWidenRecipe *BinOp, VPValue *ScalarIV,
                                      VPWidenIntOrFpInductionRecipe *WidenIV) {
  assert(Instruction::isBinaryOp(BinOp->getOpcode()) &&
         BinOp->getNumOperands() == 2 && "BinOp must have 2 operands");
  auto *ClonedOp = BinOp->clone();
  if (ClonedOp->getOperand(0) == WidenIV) {
    ClonedOp->setOperand(0, ScalarIV);
  } else {
    assert(ClonedOp->getOperand(1) == WidenIV && "one operand must be WideIV");
    ClonedOp->setOperand(1, ScalarIV);
  }
  ClonedOp->insertAfter(ScalarIV->getDefiningRecipe());
  return ClonedOp;
}

/// If \p S is an affine AddRec, returns true if its step is known to be
/// positive and false if it is known to be negative. Returns std::nullopt if
/// \p S is not an affine AddRec, or if the sign of its step cannot be
/// determined.
static std::optional<bool> getStepDirection(const SCEV *S,
                                            ScalarEvolution &SE) {
  const SCEV *Step;
  if (!match(S, m_scev_AffineAddRec(m_SCEV(), m_SCEV(Step))))
    return std::nullopt;
  if (SE.isKnownPositive(Step))
    return true;
  if (SE.isKnownNegative(Step))
    return false;
  return std::nullopt;
}

void VPlanTransforms::optimizeFindIVReductions(VPlan &Plan,
                                               PredicatedScalarEvolution &PSE,
                                               Loop &L) {
  ScalarEvolution &SE = *PSE.getSE();
  VPRegionBlock *VectorLoopRegion = Plan.getVectorLoopRegion();

  // Helper lambda to check if the IV range excludes the sentinel value. Try
  // signed first, then unsigned. Return an excluded sentinel if found,
  // otherwise return std::nullopt.
  auto CheckSentinel = [&SE](const SCEV *IVSCEV,
                             bool UseMax) -> std::optional<APSInt> {
    unsigned BW = IVSCEV->getType()->getScalarSizeInBits();
    for (bool Signed : {true, false}) {
      APSInt Sentinel = UseMax ? APSInt::getMinValue(BW, /*Unsigned=*/!Signed)
                               : APSInt::getMaxValue(BW, /*Unsigned=*/!Signed);

      ConstantRange IVRange =
          Signed ? SE.getSignedRange(IVSCEV) : SE.getUnsignedRange(IVSCEV);
      if (!IVRange.contains(Sentinel))
        return Sentinel;
    }
    return std::nullopt;
  };

  VPValue *HeaderMask = VectorLoopRegion->getHeaderMask();
  for (VPRecipeBase &Phi :
       make_early_inc_range(VectorLoopRegion->getEntryBasicBlock()->phis())) {
    auto *PhiR = dyn_cast<VPReductionPHIRecipe>(&Phi);
    if (!PhiR || !RecurrenceDescriptor::isFindLastRecurrenceKind(
                     PhiR->getRecurrenceKind()))
      continue;

    Type *PhiTy = PhiR->getScalarType();
    if (PhiTy->isPointerTy() || PhiTy->isFloatingPointTy())
      continue;

    // If there's a header mask, the backedge select will not be the find-last
    // select.
    VPValue *BackedgeVal = PhiR->getBackedgeValue();
    auto *FindLastSelect = cast<VPSingleDefRecipe>(BackedgeVal);
    if (HeaderMask &&
        !match(BackedgeVal,
               m_Select(m_Specific(HeaderMask),
                        m_VPSingleDefRecipe(FindLastSelect), m_Specific(PhiR))))
      continue;

    // Get the find-last expression from the find-last select of the reduction
    // phi. The find-last select should be a select between the phi and the
    // find-last expression.
    VPValue *Cond, *FindLastExpression;
    if (!match(FindLastSelect, m_SelectLike(m_VPValue(Cond), m_Specific(PhiR),
                                            m_VPValue(FindLastExpression))) &&
        !match(FindLastSelect,
               m_SelectLike(m_VPValue(Cond), m_VPValue(FindLastExpression),
                            m_Specific(PhiR))))
      continue;

    // Check if FindLastExpression is a simple expression of a widened IV. If
    // so, we can track the underlying IV instead and sink the expression.
    auto *IVOfExpressionToSink = getExpressionIV(FindLastExpression);
    const SCEV *IVSCEV = vputils::getSCEVExprForVPValue(
        IVOfExpressionToSink ? IVOfExpressionToSink : FindLastExpression, PSE,
        &L);
    if (!match(IVSCEV, m_scev_AffineAddRec(m_SCEV(), m_SCEV()))) {
      assert(!match(vputils::getSCEVExprForVPValue(FindLastExpression, PSE, &L),
                    m_scev_AffineAddRec(m_SCEV(), m_SCEV())) &&
             "IVOfExpressionToSink not being an AddRec must imply "
             "FindLastExpression not being an AddRec.");
      continue;
    }

    // Determine direction from the step of IVSCEV, if possible.
    std::optional<bool> StepDirection = getStepDirection(IVSCEV, SE);
    if (!StepDirection)
      continue;

    bool UseMax = *StepDirection;
    std::optional<APSInt> SentinelVal = CheckSentinel(IVSCEV, UseMax);
    bool UseSigned = SentinelVal && SentinelVal->isSigned();

    // Sinking an expression will disable epilogue vectorization. Only use it,
    // if FindLastExpression cannot be vectorized via a sentinel. Sinking may
    // also prevent vectorizing using a sentinel (e.g., if the expression is a
    // multiply or divide by large constant, respectively), which also makes
    // sinking undesirable.
    if (IVOfExpressionToSink) {
      const SCEV *FindLastExpressionSCEV =
          vputils::getSCEVExprForVPValue(FindLastExpression, PSE, &L);
      if (std::optional<bool> NewUseMax =
              getStepDirection(FindLastExpressionSCEV, SE)) {
        if (auto NewSentinel =
                CheckSentinel(FindLastExpressionSCEV, *NewUseMax)) {
          // The original expression already has a sentinel, so prefer not
          // sinking to keep epilogue vectorization possible.
          SentinelVal = *NewSentinel;
          UseSigned = NewSentinel->isSigned();
          UseMax = *NewUseMax;
          IVSCEV = FindLastExpressionSCEV;
          IVOfExpressionToSink = nullptr;
        }
      }
    }

    // If no sentinel was found, fall back to a boolean AnyOf reduction to track
    // if the condition was ever true. Requires the IV to not wrap, otherwise we
    // cannot use min/max.
    if (!SentinelVal) {
      auto *AR = cast<SCEVAddRecExpr>(IVSCEV);
      if (AR->hasNoSignedWrap())
        UseSigned = true;
      else if (AR->hasNoUnsignedWrap())
        UseSigned = false;
      else
        continue;
    }

    VPInstruction *RdxResult = cast<VPInstruction>(vputils::findRecipe(
        BackedgeVal,
        match_fn(m_VPInstruction<VPInstruction::ComputeReductionResult>())));

    VPValue *NewFindLastSelect = BackedgeVal;
    VPValue *SelectCond = Cond;
    if (!SentinelVal || IVOfExpressionToSink) {
      // When we need to create a new select, normalize the condition so that
      // PhiR is the last operand and include the header mask if needed.
      DebugLoc DL = FindLastSelect->getDefiningRecipe()->getDebugLoc();
      VPBuilder LoopBuilder(FindLastSelect->getDefiningRecipe());
      if (match(FindLastSelect,
                m_SelectLike(m_VPValue(Cond), m_Specific(PhiR), m_VPValue())))
        SelectCond = LoopBuilder.createNot(SelectCond);

      // When tail folding, mask the condition with the header mask to prevent
      // propagating poison from inactive lanes in the last vector iteration.
      if (HeaderMask)
        SelectCond = LoopBuilder.createLogicalAnd(HeaderMask, SelectCond);

      if (SelectCond != Cond || IVOfExpressionToSink) {
        NewFindLastSelect = LoopBuilder.createSelect(
            SelectCond,
            IVOfExpressionToSink ? IVOfExpressionToSink : FindLastExpression,
            PhiR, DL);
      }
    }

    // Create the reduction result in the middle block using sentinel directly.
    RecurKind MinMaxKind =
        UseMax ? (UseSigned ? RecurKind::SMax : RecurKind::UMax)
               : (UseSigned ? RecurKind::SMin : RecurKind::UMin);
    VPIRFlags Flags(MinMaxKind, /*IsOrdered=*/false, /*IsInLoop=*/false,
                    FastMathFlags());
    DebugLoc ExitDL = RdxResult->getDebugLoc();
    VPBuilder MiddleBuilder(RdxResult);
    VPValue *ReducedIV =
        MiddleBuilder.createNaryOp(VPInstruction::ComputeReductionResult,
                                   NewFindLastSelect, Flags, ExitDL);

    // If IVOfExpressionToSink is an expression to sink, sink it now.
    VPValue *VectorRegionExitingVal = ReducedIV;
    if (IVOfExpressionToSink)
      VectorRegionExitingVal =
          cloneBinOpForScalarIV(cast<VPWidenRecipe>(FindLastExpression),
                                ReducedIV, IVOfExpressionToSink);

    VPValue *NewRdxResult;
    VPValue *StartVPV = PhiR->getStartValue();
    if (SentinelVal) {
      // Sentinel-based approach: reduce IVs with min/max, compare against
      // sentinel to detect if condition was ever true, select accordingly.
      VPValue *Sentinel = Plan.getConstantInt(*SentinelVal);
      auto *Cmp = MiddleBuilder.createICmp(CmpInst::ICMP_NE, ReducedIV,
                                           Sentinel, ExitDL);
      NewRdxResult = MiddleBuilder.createSelect(Cmp, VectorRegionExitingVal,
                                                StartVPV, ExitDL);
      StartVPV = Sentinel;
    } else {
      // Introduce a boolean AnyOf reduction to track if the condition was ever
      // true in the loop. Use it to select the initial start value, if it was
      // never true.
      auto *AnyOfPhi = new VPReductionPHIRecipe(
          /*Phi=*/nullptr, RecurKind::Or, *Plan.getFalse(), *Plan.getFalse(),
          RdxUnordered{1}, {}, /*HasUsesOutsideReductionChain=*/false);
      AnyOfPhi->insertAfter(PhiR);

      VPBuilder LoopBuilder(BackedgeVal->getDefiningRecipe());
      VPValue *OrVal = LoopBuilder.createOr(AnyOfPhi, SelectCond);
      AnyOfPhi->setOperand(1, OrVal);

      NewRdxResult = MiddleBuilder.createAnyOfReduction(
          OrVal, VectorRegionExitingVal, StartVPV, ExitDL);

      // Initialize the IV reduction phi with the neutral element, not the
      // original start value, to ensure correct min/max reduction results.
      StartVPV = Plan.getOrAddLiveIn(
          getRecurrenceIdentity(MinMaxKind, IVSCEV->getType(), {}));
    }
    RdxResult->replaceAllUsesWith(NewRdxResult);
    RdxResult->eraseFromParent();

    auto *NewPhiR = new VPReductionPHIRecipe(
        cast<PHINode>(PhiR->getUnderlyingInstr()), RecurKind::FindIV, *StartVPV,
        *NewFindLastSelect, RdxUnordered{1}, {},
        PhiR->hasUsesOutsideReductionChain());
    NewPhiR->insertBefore(PhiR);
    PhiR->replaceAllUsesWith(NewPhiR);
    PhiR->eraseFromParent();
  }
}

namespace {

using ExtendKind = TTI::PartialReductionExtendKind;
struct ReductionExtend {
  Type *SrcType = nullptr;
  ExtendKind Kind = ExtendKind::PR_None;
};

/// Describes the extends used to compute the extended reduction operand.
/// ExtendB is optional. If ExtendB is present, ExtendsUser is a binary
/// operation.
struct ExtendedReductionOperand {
  /// The recipe that consumes the extends.
  VPWidenRecipe *ExtendsUser = nullptr;
  /// Extend descriptions (inputs to getPartialReductionCost).
  ReductionExtend ExtendA, ExtendB;
};

/// A chain of recipes that form a partial reduction. Matches either
///   reduction_bin_op (extended op, accumulator), or
///   reduction_bin_op (accumulator, extended op).
/// The possible forms of the "extended op" are listed in
/// matchExtendedReductionOperand.
struct VPPartialReductionChain {
  /// The top-level binary operation that forms the reduction to a scalar
  /// after the loop body.
  VPWidenRecipe *ReductionBinOp = nullptr;
  /// The user of the extends that is then reduced.
  ExtendedReductionOperand ExtendedOp;
  /// The recurrence kind for the entire partial reduction chain.
  /// This allows distinguishing between Sub and AddWithSub recurrences,
  /// when the ReductionBinOp is a Instruction::Sub.
  RecurKind RK;
  /// The index of the accumulator operand of ReductionBinOp. The extended op
  /// is `1 - AccumulatorOpIdx`.
  unsigned AccumulatorOpIdx;
  unsigned ScaleFactor;
  /// Optional blend to represent predication for the block that updates the
  /// reduction.
  VPBlendRecipe *Blend = nullptr;
};

// Return the incoming index of the single-use value in the blend, which is
// expected to be the predicated reduction update.
static std::optional<unsigned>
getBlendReductionUpdateValueIdx(VPBlendRecipe *Blend) {
  assert(Blend && !Blend->isNormalized() &&
         Blend->getNumIncomingValues() == 2 &&
         "Expected a non-normalized blend with two incoming values");
  bool FirstIncomingHasOneUse = Blend->getIncomingValue(0)->hasOneUse();

  // Only the update value should have one use (the blend). The previous
  // value should always have at least two uses, the blend and the reduction.
  if (FirstIncomingHasOneUse == Blend->getIncomingValue(1)->hasOneUse())
    return std::nullopt;
  return FirstIncomingHasOneUse ? 0 : 1;
}

static VPSingleDefRecipe *
optimizeExtendsForPartialReduction(VPSingleDefRecipe *Op) {
  // reduce.add(mul(ext(A), C))
  // -> reduce.add(mul(ext(A), ext(trunc(C))))
  const APInt *Const;
  if (match(Op, m_Mul(m_ZExtOrSExt(m_VPValue()), m_APInt(Const)))) {
    auto *ExtA = cast<VPWidenCastRecipe>(Op->getOperand(0));
    Instruction::CastOps ExtOpc = ExtA->getOpcode();
    Type *NarrowTy = ExtA->getOperand(0)->getScalarType();
    if (!Op->hasOneUse() ||
        !llvm::canConstantBeExtended(
            Const, NarrowTy, TTI::getPartialReductionExtendKind(ExtOpc)))
      return Op;

    VPBuilder Builder(Op);
    auto *Trunc = Builder.createWidenCast(Instruction::CastOps::Trunc,
                                          Op->getOperand(1), NarrowTy);
    Type *WideTy = ExtA->getScalarType();
    Op->setOperand(1, Builder.createWidenCast(ExtOpc, Trunc, WideTy));
    return Op;
  }

  // reduce.add(abs(sub(ext(A), ext(B))))
  // -> reduce.add(ext(absolute-difference(A, B)))
  VPValue *X, *Y;
  if (match(Op, m_WidenIntrinsic<Intrinsic::abs>(m_Sub(
                    m_ZExtOrSExt(m_VPValue(X)), m_ZExtOrSExt(m_VPValue(Y)))))) {
    auto *Sub = Op->getOperand(0)->getDefiningRecipe();
    auto *Ext = cast<VPWidenCastRecipe>(Sub->getOperand(0));
    assert(Ext->getOpcode() ==
               cast<VPWidenCastRecipe>(Sub->getOperand(1))->getOpcode() &&
           "Expected both the LHS and RHS extends to be the same");
    bool IsSigned = Ext->getOpcode() == Instruction::SExt;
    VPBuilder Builder(Op);
    Type *SrcTy = X->getScalarType();
    auto *FreezeX = Builder.insert(new VPWidenRecipe(Instruction::Freeze, {X}));
    auto *FreezeY = Builder.insert(new VPWidenRecipe(Instruction::Freeze, {Y}));
    auto *Max = Builder.insert(
        new VPWidenIntrinsicRecipe(IsSigned ? Intrinsic::smax : Intrinsic::umax,
                                   {FreezeX, FreezeY}, SrcTy));
    auto *Min = Builder.insert(
        new VPWidenIntrinsicRecipe(IsSigned ? Intrinsic::smin : Intrinsic::umin,
                                   {FreezeX, FreezeY}, SrcTy));
    auto *AbsDiff = Builder.insert(
        new VPWidenRecipe(Instruction::Sub, {Max, Min},
                          VPIRFlags::getDefaultFlags(Instruction::Sub)));
    return Builder.createWidenCast(Instruction::CastOps::ZExt, AbsDiff,
                                   Op->getScalarType());
  }

  // reduce.add(ext(mul(ext(A), ext(B))))
  // -> reduce.add(mul(wider_ext(A), wider_ext(B)))
  // TODO: Support this optimization for float types.
  if (match(Op, m_ZExtOrSExt(m_Mul(m_ZExtOrSExt(m_VPValue()),
                                   m_ZExtOrSExt(m_VPValue()))))) {
    auto *Ext = cast<VPWidenCastRecipe>(Op);
    auto *Mul = cast<VPWidenRecipe>(Ext->getOperand(0));
    auto *MulLHS = cast<VPWidenCastRecipe>(Mul->getOperand(0));
    auto *MulRHS = cast<VPWidenCastRecipe>(Mul->getOperand(1));
    if (!Mul->hasOneUse() ||
        (Ext->getOpcode() != MulLHS->getOpcode() && MulLHS != MulRHS) ||
        MulLHS->getOpcode() != MulRHS->getOpcode())
      return Op;
    VPBuilder Builder(Mul);
    auto *NewLHS = Builder.createWidenCast(
        MulLHS->getOpcode(), MulLHS->getOperand(0), Ext->getScalarType());
    auto *NewRHS = MulLHS == MulRHS
                       ? NewLHS
                       : Builder.createWidenCast(MulRHS->getOpcode(),
                                                 MulRHS->getOperand(0),
                                                 Ext->getScalarType());
    auto *NewMul = Mul->cloneWithOperands({NewLHS, NewRHS});
    Builder.insert(NewMul);
    Op->replaceAllUsesWith(NewMul);
    Op->eraseFromParent();
    Mul->eraseFromParent();
    return NewMul;
  }

  return Op;
}

static VPExpressionRecipe *
createPartialReductionExpression(VPReductionRecipe *Red) {
  VPValue *VecOp = Red->getVecOp();

  // reduce.[f]add(ext(op))
  //  -> VPExpressionRecipe(op, red)
  if (match(VecOp, m_WidenAnyExtend(m_VPValue())))
    return new VPExpressionRecipe(cast<VPWidenCastRecipe>(VecOp), Red);

  // reduce.[f]add(neg(ext(op)))
  // -> VPExpressionRecipe(op, sub/neg, red)
  if (match(VecOp, m_AnyNeg(m_WidenAnyExtend(m_VPValue())))) {
    auto *Neg = cast<VPWidenRecipe>(VecOp);
    auto *Ext =
        cast<VPWidenCastRecipe>(Neg->getOperand(Neg->getNumOperands() - 1));
    return new VPExpressionRecipe(Ext, Neg, Red);
  }

  // reduce.[f]add([f]mul(ext(a), ext(b)))
  //  -> VPExpressionRecipe(a, b, mul, red)
  if (match(VecOp, m_FMul(m_FPExt(m_VPValue()), m_FPExt(m_VPValue()))) ||
      match(VecOp,
            m_Mul(m_ZExtOrSExt(m_VPValue()), m_ZExtOrSExt(m_VPValue())))) {
    auto *Mul = cast<VPWidenRecipe>(VecOp);
    auto *ExtA = cast<VPWidenCastRecipe>(Mul->getOperand(0));
    auto *ExtB = cast<VPWidenCastRecipe>(Mul->getOperand(1));
    return new VPExpressionRecipe(ExtA, ExtB, Mul, Red);
  }

  // reduce.fadd(fneg(fmul(fpext(a), fpext(b))))
  //  -> VPExpressionRecipe(a, b, fmul, fsub, red)
  if (match(VecOp,
            m_FNeg(m_FMul(m_FPExt(m_VPValue()), m_FPExt(m_VPValue()))))) {
    auto *FNeg = cast<VPWidenRecipe>(VecOp);
    auto *FMul = cast<VPWidenRecipe>(FNeg->getOperand(0));
    auto *ExtA = cast<VPWidenCastRecipe>(FMul->getOperand(0));
    auto *ExtB = cast<VPWidenCastRecipe>(FMul->getOperand(1));
    return new VPExpressionRecipe(ExtA, ExtB, FMul, FNeg, Red);
  }

  // reduce.add(neg(mul(ext(a), ext(b))))
  //  -> VPExpressionRecipe(a, b, mul, sub, red)
  if (match(VecOp, m_Sub(m_ZeroInt(), m_Mul(m_ZExtOrSExt(m_VPValue()),
                                            m_ZExtOrSExt(m_VPValue()))))) {
    auto *Sub = cast<VPWidenRecipe>(VecOp);
    auto *Mul = cast<VPWidenRecipe>(Sub->getOperand(1));
    auto *ExtA = cast<VPWidenCastRecipe>(Mul->getOperand(0));
    auto *ExtB = cast<VPWidenCastRecipe>(Mul->getOperand(1));
    return new VPExpressionRecipe(ExtA, ExtB, Mul, Sub, Red);
  }

  llvm_unreachable("Unsupported expression");
}

// Helper to transform a partial reduction chain into a partial reduction
// recipe. Assumes profitability has been checked.
static void transformToPartialReduction(const VPPartialReductionChain &Chain,
                                        VPlan &Plan,
                                        VPReductionPHIRecipe *RdxPhi) {
  VPWidenRecipe *WidenRecipe = Chain.ReductionBinOp;
  assert(WidenRecipe->getNumOperands() == 2 && "Expected binary operation");

  VPValue *Accumulator = WidenRecipe->getOperand(Chain.AccumulatorOpIdx);
  auto *ExtendedOp = cast<VPSingleDefRecipe>(
      WidenRecipe->getOperand(1 - Chain.AccumulatorOpIdx));

  // FIXME: Do these transforms before invoking the cost-model.
  ExtendedOp = optimizeExtendsForPartialReduction(ExtendedOp);

  // Sub-reductions can be implemented in two ways:
  // (1) negate the operand in the vector loop (the default way).
  // (2) subtract the reduced value from the init value in the middle block.
  // Both ways keep the reduction itself as an 'add' reduction.
  //
  // The ISD nodes for partial reductions don't support folding the
  // sub/negation into its operands because the following is not a valid
  // transformation:
  //      sub(0, mul(ext(a), ext(b)))
  //   -> mul(ext(a), ext(sub(0, b)))
  //
  // It's therefore better to choose option (2) such that the partial
  // reduction is always positive (starting at '0') and to do a final
  // subtract in the middle block.
  if ((WidenRecipe->getOpcode() == Instruction::Sub &&
       Chain.RK != RecurKind::Sub) ||
      (WidenRecipe->getOpcode() == Instruction::FSub &&
       Chain.RK != RecurKind::FSub)) {
    VPBuilder Builder(WidenRecipe);
    Type *ElemTy = ExtendedOp->getScalarType();
    VPWidenRecipe *NegRecipe;
    if (WidenRecipe->getOpcode() == Instruction::FSub) {
      NegRecipe =
          new VPWidenRecipe(Instruction::FNeg, {ExtendedOp},
                            VPIRFlags::getDefaultFlags(Instruction::FNeg),
                            VPIRMetadata(), DebugLoc::getUnknown());
    } else {
      auto *Zero = Plan.getZero(ElemTy);
      NegRecipe =
          new VPWidenRecipe(Instruction::Sub, {Zero, ExtendedOp},
                            VPIRFlags::getDefaultFlags(Instruction::Sub),
                            VPIRMetadata(), DebugLoc::getUnknown());
    }
    Builder.insert(NegRecipe);
    ExtendedOp = NegRecipe;
  }

  // Check if WidenRecipe is the final result of the reduction. If so, look
  // through the Select recipe introduced by tail-folding, otherwise look
  // through any Blend recipe introduced by predication for the block.
  VPValue *ExitSearch =
      Chain.Blend ? cast<VPValue>(Chain.Blend) : cast<VPValue>(WidenRecipe);

  VPValue *Cond = nullptr;
  VPValue *ExitValue = cast_or_null<VPInstruction>(
      findUserOf(ExitSearch, m_Select(m_VPValue(Cond), m_Specific(ExitSearch),
                                      m_Specific(RdxPhi))));

  if (Chain.Blend) {
    std::optional<unsigned> BlendReductionIdx =
        getBlendReductionUpdateValueIdx(Chain.Blend);
    assert(BlendReductionIdx &&
           Chain.Blend->getIncomingValue(*BlendReductionIdx) == WidenRecipe &&
           "Expected blend to contain the reduction update");
    VPValue *BlendCond = Chain.Blend->getMask(*BlendReductionIdx);
    Cond = ExitValue ? VPBuilder(WidenRecipe)
                           .createLogicalAnd(Cond, BlendCond,
                                             WidenRecipe->getDebugLoc())
                     : BlendCond;
  }

  // When folding the tail, the inactive lanes of the reduction update are
  // computed from values that do not correspond to any scalar iteration
  // and must not be accumulated.
  if (!Cond)
    Cond = Plan.getVectorLoopRegion()->getHeaderMask();

  bool IsLastInChain = RdxPhi->getBackedgeValue() == WidenRecipe ||
                       RdxPhi->getBackedgeValue() == ExitValue ||
                       RdxPhi->getBackedgeValue() == Chain.Blend;
  assert((!ExitValue || IsLastInChain) &&
         "if we found ExitValue, it must match RdxPhi's backedge value");

  Type *PhiType = RdxPhi->getScalarType();
  RecurKind RdxKind =
      PhiType->isFloatingPointTy() ? RecurKind::FAdd : RecurKind::Add;
  auto *PartialRed = new VPReductionRecipe(
      RdxKind,
      RdxKind == RecurKind::FAdd ? WidenRecipe->getFastMathFlagsOrNone()
                                 : FastMathFlags(),
      WidenRecipe->getUnderlyingInstr(), Accumulator, ExtendedOp, Cond,
      RdxUnordered{/*VFScaleFactor=*/Chain.ScaleFactor});
  PartialRed->insertBefore(WidenRecipe);

  if (ExitValue)
    ExitValue->replaceAllUsesWith(PartialRed);
  if (Chain.Blend)
    Chain.Blend->replaceAllUsesWith(PartialRed);
  WidenRecipe->replaceAllUsesWith(PartialRed);

  // For cost-model purposes, fold this into a VPExpression.
  VPExpressionRecipe *E = createPartialReductionExpression(PartialRed);
  E->insertBefore(WidenRecipe);
  PartialRed->replaceAllUsesWith(E);

  // We only need to update the PHI node once, which is when we find the
  // last reduction in the chain.
  if (!IsLastInChain)
    return;

  // Scale the PHI and ReductionStartVector by the VFScaleFactor
  assert(RdxPhi->getVFScaleFactor() == 1 && "scale factor must not be set");
  RdxPhi->setVFScaleFactor(Chain.ScaleFactor);

  auto *StartInst = cast<VPInstruction>(RdxPhi->getStartValue());
  assert(StartInst->getOpcode() == VPInstruction::ReductionStartVector);
  auto *NewScaleFactor = Plan.getConstantInt(32, Chain.ScaleFactor);
  StartInst->setOperand(2, NewScaleFactor);

  // If this is the last value in a sub-reduction chain, then update the PHI
  // node to start at `0` and update the reduction-result to subtract from
  // the PHI's start value.
  if (Chain.RK != RecurKind::Sub && Chain.RK != RecurKind::FSub)
    return;

  VPValue *OldStartValue = StartInst->getOperand(0);
  StartInst->setOperand(0, StartInst->getOperand(1));

  // Replace reduction_result by 'sub (startval, reductionresult)'.
  VPInstruction *RdxResult = vputils::findComputeReductionResult(RdxPhi);
  assert(RdxResult && "Could not find reduction result");

  VPBuilder Builder = VPBuilder::getToInsertAfter(RdxResult);
  unsigned SubOpc = Chain.RK == RecurKind::FSub ? Instruction::BinaryOps::FSub
                                                : Instruction::BinaryOps::Sub;
  VPInstruction *NewResult = Builder.createNaryOp(
      SubOpc, {OldStartValue, RdxResult}, VPIRFlags::getDefaultFlags(SubOpc),
      RdxPhi->getDebugLoc());
  RdxResult->replaceUsesWithIf(
      NewResult,
      [&NewResult](VPUser &U, unsigned Idx) { return &U != NewResult; });
}

/// Returns the cost of a link in a partial-reduction chain for a given VF.
static InstructionCost
getPartialReductionLinkCost(VPCostContext &CostCtx,
                            const VPPartialReductionChain &Link,
                            ElementCount VF) {
  Type *RdxType = Link.ReductionBinOp->getScalarType();
  const ExtendedReductionOperand &ExtendedOp = Link.ExtendedOp;
  std::optional<unsigned> BinOpc = std::nullopt;
  // If ExtendB is not none, then the "ExtendsUser" is the binary operation.
  if (ExtendedOp.ExtendB.Kind != ExtendKind::PR_None)
    BinOpc = ExtendedOp.ExtendsUser->getOpcode();

  std::optional<llvm::FastMathFlags> Flags;
  if (RdxType->isFloatingPointTy())
    Flags = Link.ReductionBinOp->getFastMathFlagsOrNone();

  auto GetLinkOpcode = [&Link]() -> unsigned {
    switch (Link.RK) {
    case RecurKind::Sub:
      return Instruction::Add;
    case RecurKind::FSub:
      return Instruction::FAdd;
    default:
      return Link.ReductionBinOp->getOpcode();
    }
  };

  return CostCtx.TTI.getPartialReductionCost(
      GetLinkOpcode(), ExtendedOp.ExtendA.SrcType, ExtendedOp.ExtendB.SrcType,
      RdxType, VF, ExtendedOp.ExtendA.Kind, ExtendedOp.ExtendB.Kind, BinOpc,
      CostCtx.CostKind, Flags);
}

static ExtendKind getPartialReductionExtendKind(VPWidenCastRecipe *Cast) {
  return TTI::getPartialReductionExtendKind(Cast->getOpcode());
}

/// Checks if \p Op (which is an operand of \p UpdateR) is an extended reduction
/// operand. This is an operand where the source of the value (e.g. a load) has
/// been extended (sext, zext, or fpext) before it is used in the reduction.
///
/// Possible forms matched by this function:
///  - UpdateR(PrevValue, ext(...))
///  - UpdateR(PrevValue, mul(ext(...), ext(...)))
///  - UpdateR(PrevValue, mul(ext(...), Constant))
///  - UpdateR(PrevValue, ext(mul(ext(...), ext(...))))
///  - UpdateR(PrevValue, ext(mul(ext(...), Constant)))
///  - UpdateR(PrevValue, abs(sub(ext(...), ext(...)))
///
/// Note: The second operand of UpdateR corresponds to \p Op in the examples.
static std::optional<ExtendedReductionOperand>
matchExtendedReductionOperand(VPWidenRecipe *UpdateR, VPValue *Op) {
  assert(is_contained(UpdateR->operands(), Op) &&
         "Op should be operand of UpdateR");

  // Try matching an absolute difference operand of the form
  // `abs(sub(ext(A), ext(B)))`. This will be later transformed into
  // `ext(absolute-difference(A, B))`. This allows us to perform the absolute
  // difference on a wider type and get the extend for "free" from the partial
  // reduction.
  VPValue *X, *Y;
  if (Op->hasOneUse() &&
      match(Op, m_WidenIntrinsic<Intrinsic::abs>(
                    m_OneUse(m_Sub(m_WidenAnyExtend(m_VPValue(X)),
                                   m_WidenAnyExtend(m_VPValue(Y))))))) {
    auto *Abs = cast<VPWidenIntrinsicRecipe>(Op);
    auto *Sub = cast<VPWidenRecipe>(Abs->getOperand(0));
    auto *LHSExt = cast<VPWidenCastRecipe>(Sub->getOperand(0));
    auto *RHSExt = cast<VPWidenCastRecipe>(Sub->getOperand(1));
    Type *LHSInputType = X->getScalarType();
    Type *RHSInputType = Y->getScalarType();
    if (LHSInputType != RHSInputType ||
        LHSExt->getOpcode() != RHSExt->getOpcode())
      return std::nullopt;
    // Note: This is essentially the same as matching ext(...) as we will
    // rewrite this operand to ext(absolute-difference(A, B)).
    return ExtendedReductionOperand{
        Sub,
        /*ExtendA=*/{LHSInputType, getPartialReductionExtendKind(LHSExt)},
        /*ExtendB=*/{}};
  }

  std::optional<TTI::PartialReductionExtendKind> OuterExtKind;
  if (match(Op, m_WidenAnyExtend(m_VPValue()))) {
    auto *CastRecipe = cast<VPWidenCastRecipe>(Op);
    VPValue *CastSource = CastRecipe->getOperand(0);
    OuterExtKind = getPartialReductionExtendKind(CastRecipe);
    if (match(CastSource, m_Mul(m_VPValue(), m_VPValue())) ||
        match(CastSource, m_FMul(m_VPValue(), m_VPValue()))) {
      // Match: ext(mul(...))
      // Record the outer extend kind and set `Op` to the mul. We can then match
      // this as a binary operation. Note: We can optimize out the outer extend
      // by widening the inner extends to match it. See
      // optimizeExtendsForPartialReduction.
      Op = CastSource;
    } else {
      return ExtendedReductionOperand{
          UpdateR,
          /*ExtendA=*/{CastSource->getScalarType(), *OuterExtKind},
          /*ExtendB=*/{}};
    }
  }

  if (!Op->hasOneUse())
    return std::nullopt;

  VPWidenRecipe *MulOp = dyn_cast<VPWidenRecipe>(Op);
  if (!MulOp ||
      !is_contained({Instruction::Mul, Instruction::FMul}, MulOp->getOpcode()))
    return std::nullopt;

  // The rest of the matching assumes `Op` is a (possibly extended) mul
  // operation.

  VPValue *LHS = MulOp->getOperand(0);
  VPValue *RHS = MulOp->getOperand(1);

  // The LHS of the operation must always be an extend.
  if (!match(LHS, m_WidenAnyExtend(m_VPValue())))
    return std::nullopt;

  auto *LHSCast = cast<VPWidenCastRecipe>(LHS);
  Type *LHSInputType = LHSCast->getOperand(0)->getScalarType();
  ExtendKind LHSExtendKind = getPartialReductionExtendKind(LHSCast);

  // The RHS of the operation can be an extend or a constant integer.
  const APInt *RHSConst = nullptr;
  VPWidenCastRecipe *RHSCast = nullptr;
  if (match(RHS, m_WidenAnyExtend(m_VPValue())))
    RHSCast = cast<VPWidenCastRecipe>(RHS);
  else if (!match(RHS, m_APInt(RHSConst)) ||
           !canConstantBeExtended(RHSConst, LHSInputType, LHSExtendKind))
    return std::nullopt;

  // The outer extend kind must match the inner extends for folding.
  for (VPWidenCastRecipe *Cast : {LHSCast, RHSCast})
    if (Cast && OuterExtKind &&
        getPartialReductionExtendKind(Cast) != OuterExtKind)
      return std::nullopt;

  Type *RHSInputType = LHSInputType;
  ExtendKind RHSExtendKind = LHSExtendKind;
  if (RHSCast) {
    RHSInputType = RHSCast->getOperand(0)->getScalarType();
    RHSExtendKind = getPartialReductionExtendKind(RHSCast);
  }

  return ExtendedReductionOperand{
      MulOp, {LHSInputType, LHSExtendKind}, {RHSInputType, RHSExtendKind}};
}

/// Examines each operation in the reduction chain corresponding to \p RedPhiR,
/// and determines if the target can use a cheaper operation with a wider
/// per-iteration input VF and narrower PHI VF. If successful, returns the chain
/// of operations in the reduction.
static std::optional<SmallVector<VPPartialReductionChain>>
getScaledReductions(VPReductionPHIRecipe *RedPhiR) {
  // Get the backedge value from the reduction PHI and find the
  // ComputeReductionResult that uses it (directly or through a select for
  // predicated reductions).
  auto *RdxResult = vputils::findComputeReductionResult(RedPhiR);
  if (!RdxResult)
    return std::nullopt;
  VPValue *ExitValue = RdxResult->getOperand(0);
  match(ExitValue, m_Select(m_VPValue(), m_VPValue(ExitValue), m_VPValue()));

  SmallVector<VPPartialReductionChain> Chain;
  RecurKind RK = RedPhiR->getRecurrenceKind();
  Type *PhiType = RedPhiR->getScalarType();
  TypeSize PHISize = PhiType->getPrimitiveSizeInBits();

  // Work backwards from the ExitValue examining each reduction operation.
  VPValue *CurrentValue = ExitValue;
  while (CurrentValue != RedPhiR) {
    VPBlendRecipe *Blend = dyn_cast<VPBlendRecipe>(CurrentValue);
    std::optional<unsigned> BlendReductionIdx;
    if (Blend) {
      assert(!Blend->isNormalized() && "Expect Blend not to be normalized.");
      if (Blend->getNumIncomingValues() != 2)
        return std::nullopt;

      BlendReductionIdx = getBlendReductionUpdateValueIdx(Blend);
      if (!BlendReductionIdx)
        return std::nullopt;

      CurrentValue = Blend->getIncomingValue(*BlendReductionIdx);
    }

    auto *UpdateR = dyn_cast<VPWidenRecipe>(CurrentValue);
    if (!UpdateR || !Instruction::isBinaryOp(UpdateR->getOpcode()))
      return std::nullopt;

    VPValue *Op = UpdateR->getOperand(1);
    VPValue *PrevValue = UpdateR->getOperand(0);

    // Find the extended operand. The other operand (PrevValue) is the next link
    // in the reduction chain.
    std::optional<ExtendedReductionOperand> ExtendedOp =
        matchExtendedReductionOperand(UpdateR, Op);
    if (!ExtendedOp) {
      ExtendedOp = matchExtendedReductionOperand(UpdateR, PrevValue);
      if (!ExtendedOp)
        return std::nullopt;
      std::swap(Op, PrevValue);
    }

    // Look for VPBlend(reduce(PrevValue, Op), PrevValue), where
    // reduce is equal to CurrentValue. This can be lowered as
    // a conditional reduction by hoisting the select to the inputs.
    if (Blend && Blend->getIncomingValue(1 - *BlendReductionIdx) != PrevValue)
      return std::nullopt;

    Type *ExtSrcType = ExtendedOp->ExtendA.SrcType;
    TypeSize ExtSrcSize = ExtSrcType->getPrimitiveSizeInBits();
    if (!PHISize.hasKnownScalarFactor(ExtSrcSize))
      return std::nullopt;

    VPPartialReductionChain Link(
        {UpdateR, *ExtendedOp, RK,
         PrevValue == UpdateR->getOperand(0) ? 0U : 1U,
         static_cast<unsigned>(PHISize.getKnownScalarFactor(ExtSrcSize)),
         Blend});
    Chain.push_back(Link);
    CurrentValue = PrevValue;
  }

  // The chain links were collected by traversing backwards from the exit value.
  // Reverse the chains so they are in program order.
  std::reverse(Chain.begin(), Chain.end());
  return Chain;
}
} // namespace

void VPlanTransforms::createPartialReductions(VPlan &Plan,
                                              VPCostContext &CostCtx,
                                              VFRange &Range) {
  // Find all possible valid partial reductions, grouping chains by their PHI.
  // This grouping allows invalidating the whole chain, if any link is not a
  // valid partial reduction.
  MapVector<VPReductionPHIRecipe *, SmallVector<VPPartialReductionChain>>
      ChainsByPhi;
  VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
  SmallVector<VPReductionPHIRecipe *, 4> UnorderedReductions;
  for (VPReductionPHIRecipe &RedPhiR :
       make_isa_range<VPReductionPHIRecipe>(HeaderVPBB->phis())) {
    if (auto Chains = getScaledReductions(&RedPhiR))
      ChainsByPhi.try_emplace(&RedPhiR, std::move(*Chains));
    else if (UsePartialReductionsByDefault &&
             (RedPhiR.getRecurrenceKind() == RecurKind::Add ||
              (RedPhiR.getRecurrenceKind() == RecurKind::FAdd &&
               !RedPhiR.isOrdered() && !RedPhiR.isInLoop())))
      UnorderedReductions.push_back(&RedPhiR);
  }

  // For general unordered reductions which aren't part of a candidate chain for
  // a scaled partial reduction, we can still use the intrinsic to allow for
  // more optimization later on.
  for (auto *Rdx : UnorderedReductions) {
    auto *Backedge = dyn_cast<VPWidenRecipe>(Rdx->getBackedgeValue());
    VPValue *OtherOp;
    if (!Backedge ||
        !match(Backedge,
               m_CombineOr(m_c_FAdd(m_Specific(Rdx), m_VPValue(OtherOp)),
                           m_c_Add(m_Specific(Rdx), m_VPValue(OtherOp)))))
      continue;

    // If the target indicates that the intrinsic is as cheap as (or cheaper
    // than) the add, then prefer the intrinsic.
    if (!LoopVectorizationPlanner::getDecisionAndClampRange(
            [&CostCtx, Rdx, Backedge](ElementCount VF) {
              InstructionCost CurrentCost = Backedge->computeCost(VF, CostCtx);
              Type *ScalarTy = Backedge->getScalarType();
              auto FMF = ScalarTy->isFloatingPointTy()
                             ? std::make_optional(Rdx->getFastMathFlagsOrNone())
                             : std::nullopt;

              InstructionCost PRCost = CostCtx.TTI.getPartialReductionCost(
                  Backedge->getOpcode(), ScalarTy, /*InputTypeB=*/nullptr,
                  ScalarTy, VF, TTI::PR_None, TTI::PR_None,
                  /*BinOp=*/std::nullopt, CostCtx.CostKind, FMF);
              return PRCost <= CurrentCost;
            },
            Range))
      continue;

    auto *Partial = new VPReductionRecipe(
        Rdx->getRecurrenceKind(), Rdx->getFastMathFlagsOrNone(),
        Backedge->getUnderlyingInstr(), Rdx, OtherOp, nullptr,
        getReductionStyle(/*InLoop=*/false, /*Ordered=*/false,
                          /*ScaleFactor=*/1));
    Partial->insertBefore(Backedge);
    Backedge->replaceAllUsesWith(Partial);
    Backedge->eraseFromParent();
  }

  if (ChainsByPhi.empty())
    return;

  // Build set of partial reduction operations and blends for user validation
  // and a map of reduction bin ops to their scale factors for scale validation.
  SmallPtrSet<VPRecipeBase *, 4> PartialReductionOps;
  SmallPtrSet<VPBlendRecipe *, 4> PartialReductionBlends;
  DenseMap<VPSingleDefRecipe *, unsigned> ScaledReductionMap;
  for (const auto &[_, Chains] : ChainsByPhi)
    for (const VPPartialReductionChain &Chain : Chains) {
      PartialReductionOps.insert(Chain.ExtendedOp.ExtendsUser);
      if (Chain.Blend)
        PartialReductionBlends.insert(Chain.Blend);
      ScaledReductionMap[Chain.ReductionBinOp] = Chain.ScaleFactor;
    }

  // A partial reduction is invalid if any of its extends are used by
  // something that isn't another partial reduction. This is because the
  // extends are intended to be lowered along with the reduction itself.
  auto ExtendUsersValid = [&](VPValue *Ext) {
    return !isa<VPWidenCastRecipe>(Ext) || all_of(Ext->users(), [&](VPUser *U) {
      return PartialReductionOps.contains(cast<VPRecipeBase>(U));
    });
  };

  auto IsProfitablePartialReductionChainForVF =
      [&](ArrayRef<VPPartialReductionChain> Chain, ElementCount VF) -> bool {
    InstructionCost PartialCost = 0, RegularCost = 0;

    // The chain is a profitable partial reduction chain if the cost of handling
    // the entire chain is cheaper when using partial reductions than when
    // handling the entire chain using regular reductions.
    for (const VPPartialReductionChain &Link : Chain) {
      const ExtendedReductionOperand &ExtendedOp = Link.ExtendedOp;
      InstructionCost LinkCost = getPartialReductionLinkCost(CostCtx, Link, VF);
      if (!LinkCost.isValid())
        return false;

      PartialCost += LinkCost;
      RegularCost += Link.ReductionBinOp->computeCost(VF, CostCtx);
      // If ExtendB is not none, then the "ExtendsUser" is the binary operation.
      if (ExtendedOp.ExtendB.Kind != ExtendKind::PR_None)
        RegularCost += ExtendedOp.ExtendsUser->computeCost(VF, CostCtx);
      for (VPValue *Op : ExtendedOp.ExtendsUser->operands())
        if (auto *Extend = dyn_cast<VPWidenCastRecipe>(Op))
          RegularCost += Extend->computeCost(VF, CostCtx);
    }
    return PartialCost.isValid() && PartialCost < RegularCost;
  };

  // Validate chains: check that extends are only used by partial reductions,
  // and that reduction bin ops are only used by other partial reductions with
  // matching scale factors, are outside the loop region or the select
  // introduced by tail-folding. Otherwise we would create users of scaled
  // reductions where the types of the other operands don't match.
  for (auto &[RedPhiR, Chains] : ChainsByPhi) {
    for (const VPPartialReductionChain &Chain : Chains) {
      if (!all_of(Chain.ExtendedOp.ExtendsUser->operands(), ExtendUsersValid)) {
        Chains.clear();
        break;
      }
      auto UseIsValid = [&, RedPhiR = RedPhiR](VPUser *U) {
        if (auto *PhiR = dyn_cast<VPReductionPHIRecipe>(U))
          return PhiR == RedPhiR;
        auto *R = cast<VPSingleDefRecipe>(U);

        if (auto *Blend = dyn_cast<VPBlendRecipe>(R))
          return Blend == Chain.Blend || PartialReductionBlends.contains(Blend);

        return Chain.ScaleFactor == ScaledReductionMap.lookup_or(R, 0) ||
               match(R, m_ComputeReductionResult(
                            m_Specific(Chain.ReductionBinOp))) ||
               match(R, m_Select(m_VPValue(), m_Specific(Chain.ReductionBinOp),
                                 m_Specific(RedPhiR)));
      };
      if (!all_of(Chain.ReductionBinOp->users(), UseIsValid)) {
        Chains.clear();
        break;
      }

      // Check if the compute-reduction-result is used by a sunk store.
      // TODO: Also form partial reductions in those cases.
      if (auto *RdxResult = vputils::findComputeReductionResult(RedPhiR)) {
        if (any_of(RdxResult->users(), [](VPUser *U) {
              auto *RepR = dyn_cast<VPReplicateRecipe>(U);
              return RepR && RepR->getOpcode() == Instruction::Store;
            })) {
          Chains.clear();
          break;
        }
      }
    }

    // Clear the chain if it is not profitable.
    if (!LoopVectorizationPlanner::getDecisionAndClampRange(
            [&, &Chains = Chains](ElementCount VF) {
              return IsProfitablePartialReductionChainForVF(Chains, VF);
            },
            Range))
      Chains.clear();
  }

  for (auto &[Phi, Chains] : ChainsByPhi)
    for (const VPPartialReductionChain &Chain : Chains)
      transformToPartialReduction(Chain, Plan, Phi);
}

void VPlanTransforms::makeMemOpWideningDecisions(VPlan &Plan, VFRange &Range,
                                                 VPRecipeBuilder &RecipeBuilder,
                                                 VPCostContext &CostCtx) {
  // Collect all loads/stores first. We will start with ones having simpler
  // decisions followed by more complex ones that are potentially
  // guided/dependent on the simpler ones.
  SmallVector<VPInstruction *> MemOps;
  for (VPBasicBlock *VPBB :
       VPBlockUtils::blocksOnly<VPBasicBlock>(vp_depth_first_shallow(
           Plan.getVectorLoopRegion()->getEntryBasicBlock()))) {
    for (VPInstruction &VPI : make_isa_range<VPInstruction>(*VPBB)) {
      if (VPI.getUnderlyingValue() &&
          is_contained({Instruction::Load, Instruction::Store},
                       VPI.getOpcode()))
        MemOps.push_back(&VPI);
    }
  }

  // Few helpers to process different kinds of memory operations.

  // To be used as argument to `VPlanTransforms::runPass` which explicitly
  // specified pass name, hence `VPlan &` parameter.
  auto ProcessSubset = [&](VPlan &, auto ProcessVPInst) {
    SmallVector<VPInstruction *> RemainingMemOps;
    for (VPInstruction *VPI : MemOps) {
      if (!ProcessVPInst(VPI))
        RemainingMemOps.push_back(VPI);
    }

    MemOps.clear();
    std::swap(MemOps, RemainingMemOps);
  };

  auto ReplaceWith = [&](VPInstruction *VPI, VPRecipeBase *New) {
    assert(New->getParent() && "New recipe must have been inserted");
    if (VPI->getOpcode() == Instruction::Load)
      VPI->replaceAllUsesWith(New->getVPSingleValue());
    VPI->eraseFromParent();

    // VPI has been processed.
    return true;
  };

  auto Scalarize = [&](VPInstruction *VPI) {
    return ReplaceWith(VPI, VPBuilder(VPI).insert(
                                RecipeBuilder.handleReplication(VPI, Range)));
  };

  VPBasicBlock *MiddleVPBB = Plan.getMiddleBlock();
  VPBuilder FinalRedStoresBuilder(MiddleVPBB, MiddleVPBB->getFirstNonPhi());
  VPlanTransforms::runPass(
      "lowerMemoryIdioms", ProcessSubset, Plan, [&](VPInstruction *VPI) {
        if (RecipeBuilder.replaceWithFinalIfReductionStore(
                VPI, FinalRedStoresBuilder))
          return true;

        // Filter out scalar VPlan for the remaining idioms.
        if (LoopVectorizationPlanner::getDecisionAndClampRange(
                [](ElementCount VF) { return VF.isScalar(); }, Range))
          return false;

        if (VPHistogramRecipe *Histogram = RecipeBuilder.widenIfHistogram(VPI))
          return ReplaceWith(VPI, VPBuilder(VPI).insert(Histogram));

        return false;
      });

  // Filter out scalar VPlan for the remaining memory operations.
  if (LoopVectorizationPlanner::getDecisionAndClampRange(
          [](ElementCount VF) { return VF.isScalar(); }, Range))
    return;

  // If the instruction's allocated size doesn't equal it's type size, it
  // requires padding and will be scalarized.
  VPlanTransforms::runPass(
      "scalarizeMemOpsWithIrregularTypes", ProcessSubset, Plan,
      [&](VPInstruction *VPI) {
        Instruction *I = VPI->getUnderlyingInstr();
        if (hasIrregularType(getLoadStoreType(I), I->getDataLayout()))
          return Scalarize(VPI);

        return false;
      });

  if (!RecipeBuilder.prefersVectorizedAddressing()) {
    VPlanTransforms::runPass(
        "makeVPlanMemOpDecision", ProcessSubset, Plan, [&](VPInstruction *VPI) {
          Instruction *I = VPI->getUnderlyingInstr();
          bool IsLoad = VPI->getOpcode() == Instruction::Load;
          if (RecipeBuilder.isPredicatedInst(I) || !IsLoad ||
              !vputils::isUsedByLoadStoreAddress(VPI))
            return false;

          // Scalarize loads used as addresses, matching the legacy CM. The load
          // is single-scalar if the pointer is loop-invariant, otherwise it is
          // replicated per-lane. No mask is needed as the load is not
          // predicated.
          VPValue *Ptr = VPI->getOperand(0);
          const SCEV *PtrSCEV =
              vputils::getSCEVExprForVPValue(Ptr, CostCtx.PSE, CostCtx.L);
          bool IsSingleScalarLoad =
              !isa<SCEVCouldNotCompute>(PtrSCEV) &&
              CostCtx.PSE.getSE()->isLoopInvariant(PtrSCEV, CostCtx.L);

          ReplaceWith(VPI,
                      VPBuilder(VPI).insert(new VPReplicateRecipe(
                          I, Ptr, /*IsSingleScalar=*/IsSingleScalarLoad,
                          /*Mask=*/nullptr, *VPI, *VPI, VPI->getDebugLoc())));
          return true;
        });
  }

  // Widen unit-stride consecutive accesses, matching the legacy CM. Both
  // forward (stride +1) and reverse (stride -1) accesses are handled.
  VPlanTransforms::runPass(
      "widenConsecutiveMemOps", ProcessSubset, Plan, [&](VPInstruction *VPI) {
        Instruction *I = VPI->getUnderlyingInstr();
        bool IsLoad = VPI->getOpcode() == Instruction::Load;
        VPValue *Ptr = VPI->getOperand(!IsLoad);
        Type *ScalarTy =
            IsLoad ? VPI->getScalarType() : VPI->getOperand(0)->getScalarType();
        std::optional<int64_t> Stride =
            vputils::getConstantStride(Ptr, ScalarTy, CostCtx.PSE, CostCtx.L);
        if (Stride != 1 && Stride != -1)
          return false;
        bool Reverse = Stride == -1;

        // A predicated access can only be widened (rather than scalarized) if
        // the target supports a masked load/store for it.
        // TODO: Determine if a load/store needs predication directly in VPlan.
        bool IsPredicated = RecipeBuilder.isPredicatedInst(I);
        if (IsPredicated && !CostCtx.Config.isLegalMaskedLoadOrStore(
                                IsLoad, ScalarTy, getLoadStoreAlignment(I),
                                getLoadStoreAddressSpace(I)))
          return false;

        VPBuilder Builder(VPI);
        VPSingleDefRecipe *VectorPtr = Builder.createConsecutiveVectorPointer(
            Ptr, ScalarTy, Reverse, VPI->getDebugLoc());

        VPValue *Mask = IsPredicated ? VPI->getMask() : nullptr;
        // Reverse the mask so it matches the reversed access order.
        if (Reverse && Mask)
          Mask = Builder.createNaryOp(VPInstruction::Reverse, Mask,
                                      VPI->getDebugLoc());

        if (IsLoad) {
          VPSingleDefRecipe *Load = Builder.createWidenLoad(
              *cast<LoadInst>(I), VectorPtr, Mask,
              /*Consecutive=*/true, *VPI, VPI->getDebugLoc());
          // Reverse the loaded values back into program order.
          if (Reverse)
            Load = Builder.createNaryOp(VPInstruction::Reverse, Load,
                                        VPI->getDebugLoc());
          return ReplaceWith(VPI, Load);
        }

        VPValue *StoredVal = VPI->getOperand(0);
        if (Reverse)
          // Reverse the stored values so they are written in descending order.
          StoredVal = Builder.createNaryOp(VPInstruction::Reverse, StoredVal,
                                           VPI->getDebugLoc());

        auto *StoreR = Builder.createWidenStore(
            *cast<StoreInst>(I), VectorPtr, StoredVal, Mask,
            /*Consecutive=*/true, *VPI, VPI->getDebugLoc());
        return ReplaceWith(VPI, StoreR);
      });

  VPlanTransforms::runPass("delegateMemOpWideningToLegacyCM", ProcessSubset,
                           Plan, [&](VPInstruction *VPI) {
                             if (VPRecipeBase *Recipe =
                                     RecipeBuilder.tryToWidenMemory(VPI, Range))
                               return ReplaceWith(VPI, Recipe);

                             return Scalarize(VPI);
                           });
}

void VPlanTransforms::makeScalarizationDecisions(VPlan &Plan, VFRange &Range) {
  if (LoopVectorizationPlanner::getDecisionAndClampRange(
          [&](ElementCount VF) { return VF.isScalar(); }, Range))
    return;

  PostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> POT(
      Plan.getEntry());
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(POT)) {
    for (VPInstruction &VPI :
         make_early_inc_range(make_isa_range<VPInstruction>(reverse(*VPBB)))) {
      auto *I = cast_or_null<Instruction>(VPI.getUnderlyingValue());
      // Wouldn't be able to create a `VPReplicateRecipe` anyway.
      if (!I)
        continue;

      // If executing other lanes produces side-effects we can't avoid them.
      if (VPI.mayHaveSideEffects())
        continue;

      // We want to drop the mask operand, verify we can safely do that.
      if (VPI.isMasked() && !VPI.isSafeToSpeculativelyExecute())
        continue;

      // Avoid rewriting IV increment as that interferes with
      // `removeRedundantCanonicalIVs`.
      if (VPI.getOpcode() == Instruction::Add &&
          any_of(VPI.operands(), IsaPred<VPWidenIntOrFpInductionRecipe>))
        continue;

      // Other lanes are needed - can't drop them.
      if (!vputils::onlyFirstLaneUsed(&VPI))
        continue;

      auto *Recipe = VPBuilder::createSingleScalarOp(
          VPI.getOpcode(), VPI.operandsWithoutMask(), /*Mask=*/nullptr, VPI,
          VPI, VPI.getDebugLoc(), VPI.getScalarType(), I);
      Recipe->insertBefore(&VPI);
      VPI.replaceAllUsesWith(Recipe);
      VPI.eraseFromParent();
    }
  }
}

/// Returns true if \p Info's parameter kinds are compatible with \p Args.
static bool areVFParamsOk(const VFInfo &Info, ArrayRef<VPValue *> Args,
                          PredicatedScalarEvolution &PSE, const Loop *L) {
  ScalarEvolution *SE = PSE.getSE();
  return all_of(Info.Shape.Parameters, [&](VFParameter Param) {
    switch (Param.ParamKind) {
    case VFParamKind::Vector:
    case VFParamKind::GlobalPredicate:
      return true;
    case VFParamKind::OMP_Uniform:
      return SE->isSCEVable(Args[Param.ParamPos]->getScalarType()) &&
             SE->isLoopInvariant(
                 vputils::getSCEVExprForVPValue(Args[Param.ParamPos], PSE, L),
                 L);
    case VFParamKind::OMP_Linear:
      return match(vputils::getSCEVExprForVPValue(Args[Param.ParamPos], PSE, L),
                   m_scev_AffineAddRec(
                       m_SCEV(), m_scev_SpecificSInt(Param.LinearStepOrPos),
                       m_SpecificLoop(L)));
    default:
      return false;
    }
  });
}

/// Find a vector variant of \p CI for \p VF, respecting \p MaskRequired.
/// Returns the variant function, or nullptr. Masked variants are assumed to
/// take the mask as a trailing parameter.
static Function *findVectorVariant(CallInst *CI, ArrayRef<VPValue *> Args,
                                   ElementCount VF, bool MaskRequired,
                                   PredicatedScalarEvolution &PSE,
                                   const Loop *L) {
  if (CI->isNoBuiltin())
    return nullptr;
  auto Mappings = VFDatabase::getMappings(*CI);
  const auto *It = find_if(Mappings, [&](const VFInfo &Info) {
    return Info.Shape.VF == VF && (!MaskRequired || Info.isMasked()) &&
           areVFParamsOk(Info, Args, PSE, L);
  });
  if (It == Mappings.end())
    return nullptr;
  return CI->getModule()->getFunction(It->VectorName);
}

namespace {
/// The outcome of choosing how to widen a call at a given VF.
struct CallWideningDecision {
  enum class KindTy { Scalarize, Intrinsic, VectorVariant };
  CallWideningDecision(KindTy Kind, Function *Variant = nullptr)
      : Kind(Kind), Variant(Variant) {}
  KindTy Kind;

  /// Set when Kind == VectorVariant.
  Function *Variant;

  bool operator==(const CallWideningDecision &Other) const {
    return Kind == Other.Kind && Variant == Other.Variant;
  }
};
} // namespace

/// Pick the cheapest widening for the call \p VPI at \p VF among scalarization,
/// vector intrinsic, and vector library variant.
static CallWideningDecision decideCallWidening(VPInstruction &VPI,
                                               ArrayRef<VPValue *> Ops,
                                               ElementCount VF,
                                               VPCostContext &CostCtx) {
  auto *CI = cast<CallInst>(VPI.getUnderlyingInstr());

  // Scalar VFs and calls forced or known to scalarize always replicate.
  if (VF.isScalar() || CostCtx.willBeScalarized(CI, VF))
    return CallWideningDecision::KindTy::Scalarize;

  auto *CalledFn = cast<Function>(
      VPI.getOperand(VPI.getNumOperandsWithoutMask() - 1)->getLiveInIRValue());
  Type *ResultTy = VPI.getScalarType();
  Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, &CostCtx.TLI);
  bool MaskRequired = CostCtx.isMaskRequired(CI);

  // Pseudo intrinsics (assume, lifetime, ...) are always scalarized.
  if (ID && VPCostContext::isFreeScalarIntrinsic(ID))
    return CallWideningDecision::KindTy::Scalarize;

  InstructionCost ScalarCost =
      VPReplicateRecipe::computeCallCost(CalledFn, ResultTy, Ops,
                                         /*IsSingleScalar=*/false, VF, CostCtx);

  Function *VecFunc =
      findVectorVariant(CI, Ops, VF, MaskRequired, CostCtx.PSE, CostCtx.L);
  InstructionCost VecCallCost = InstructionCost::getInvalid();
  if (VecFunc)
    VecCallCost = VPWidenCallRecipe::computeCallCost(VecFunc, CostCtx);

  // Prefer the intrinsic if it is at least as cheap as scalarizing and any
  // available vector variant.
  if (ID) {
    InstructionCost IntrinsicCost =
        VPWidenIntrinsicRecipe::computeCallCost(ID, Ops, VPI, VF, CostCtx);
    if (IntrinsicCost.isValid() && ScalarCost >= IntrinsicCost &&
        (!VecFunc || VecCallCost >= IntrinsicCost))
      return CallWideningDecision::KindTy::Intrinsic;
  }

  // Otherwise, use a vector library variant when it beats scalarizing.
  if (VecFunc && ScalarCost >= VecCallCost)
    return {CallWideningDecision::KindTy::VectorVariant, VecFunc};

  return CallWideningDecision::KindTy::Scalarize;
}

void VPlanTransforms::makeCallWideningDecisions(VPlan &Plan, VFRange &Range,
                                                VPRecipeBuilder &RecipeBuilder,
                                                VPCostContext &CostCtx) {
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksAs<VPBasicBlock>(
           vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) {
    for (VPInstruction &VPI :
         make_early_inc_range(make_isa_range<VPInstruction>(*VPBB))) {
      if (!VPI.getUnderlyingValue() || VPI.getOpcode() != Instruction::Call)
        continue;

      auto *CI = cast<CallInst>(VPI.getUnderlyingInstr());
      SmallVector<VPValue *, 4> Ops(VPI.op_begin(),
                                    VPI.op_begin() + CI->arg_size());

      CallWideningDecision Decision =
          decideCallWidening(VPI, Ops, Range.Start, CostCtx);
      LoopVectorizationPlanner::getDecisionAndClampRange(
          [&](ElementCount VF) {
            return Decision == decideCallWidening(VPI, Ops, VF, CostCtx);
          },
          Range);

      VPSingleDefRecipe *Replacement = nullptr;
      switch (Decision.Kind) {
      case CallWideningDecision::KindTy::Intrinsic: {
        Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, &CostCtx.TLI);
        Type *ResultTy = VPI.getScalarType();
        Replacement = new VPWidenIntrinsicRecipe(*CI, ID, Ops, ResultTy, VPI,
                                                 VPI, VPI.getDebugLoc());
        break;
      }
      case CallWideningDecision::KindTy::VectorVariant: {
        // Masked variants take the mask as a trailing parameter, so they have
        // one more parameter than the original call's arguments.
        if (Decision.Variant->arg_size() > Ops.size()) {
          VPValue *Mask = VPI.isMasked() ? VPI.getMask() : Plan.getTrue();
          Ops.push_back(Mask);
        }
        Ops.push_back(VPI.getOperand(VPI.getNumOperandsWithoutMask() - 1));
        Replacement = new VPWidenCallRecipe(CI, Decision.Variant, Ops, VPI, VPI,
                                            VPI.getDebugLoc());
        break;
      }
      case CallWideningDecision::KindTy::Scalarize:
        Replacement = RecipeBuilder.handleReplication(&VPI, Range);
        break;
      }

      Replacement->insertBefore(&VPI);
      VPI.replaceAllUsesWith(Replacement);
      VPI.eraseFromParent();
    }
  }
}

void VPlanTransforms::narrowInductionTruncates(VPlan &Plan, VFRange &Range,
                                               const TargetTransformInfo &TTI,
                                               PredicatedScalarEvolution &PSE) {
  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
  VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock();
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_shallow(LoopRegion->getEntry()))) {
    for (VPInstruction &VPI :
         make_early_inc_range(make_isa_range<VPInstruction>(*VPBB))) {
      // Only truncates are handled, as sext/zext may wrap, FP conversions lose
      // precision and other casts depend on the pointer size.
      if (VPI.getOpcode() != Instruction::Trunc)
        continue;

      // Underlying Trunc is necessary to create VPWidenIntOrFpInductionRecipe.
      auto *Trunc = cast_or_null<TruncInst>(VPI.getUnderlyingValue());
      if (!Trunc)
        continue;

      // A truncate that is not widened is left to the scalarization decisions
      // made earlier.
      if (vputils::onlyFirstLaneUsed(&VPI))
        continue;

      VPValue *Op = VPI.getOperand(0);
      auto *WideIV = getOptimizableIVOf(Op, PSE);
      if (!WideIV)
        continue;

      // getOptimizableIVOf also matches an add of the IV and its step, which
      // is not handled here.
      // TODO: Also narrow truncates of the incremented IV.
      if (Op != WideIV)
        continue;

      // Replacing a free truncate would add an induction update instruction to
      // each iteration of the loop. The canonical induction is exempt, as it
      // needs an update instruction regardless.
      auto IsNarrowingProfitable = [&](ElementCount VF) {
        return match(WideIV, m_CanonicalWidenIV()) ||
               !TTI.isTruncateFree(
                   toVectorTy(VPI.getOperand(0)->getScalarType(), VF),
                   toVectorTy(VPI.getScalarType(), VF));
      };
      if (!LoopVectorizationPlanner::getDecisionAndClampRange(
              IsNarrowingProfitable, Range))
        continue;

      // Wrap flags of the original induction do not hold in the truncated
      // type, so do not propagate them.
      auto *NarrowIV = new VPWidenIntOrFpInductionRecipe(
          WideIV->getPHINode(), WideIV->getStartValue(), WideIV->getStepValue(),
          WideIV->getVFValue(), WideIV->getInductionDescriptor(), Trunc,
          VPIRFlags::WrapFlagsTy(false, false), VPI.getDebugLoc());
      NarrowIV->insertBefore(*HeaderVPBB, HeaderVPBB->getFirstNonPhi());
      VPI.replaceAllUsesWith(NarrowIV);
      VPI.eraseFromParent();
    }
  }
}

void VPlanTransforms::convertToStridedAccesses(VPlan &Plan,
                                               PredicatedScalarEvolution &PSE,
                                               Loop &L, VPCostContext &Ctx,
                                               VFRange &Range) {
  if (Plan.hasScalarVFOnly())
    return;

  VPRegionBlock *VectorLoop = Plan.getVectorLoopRegion();
  VPValue *I32VF = nullptr;
  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
           vp_depth_first_shallow(VectorLoop->getEntry()))) {
    for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
      auto *MemR = dyn_cast<VPWidenMemoryRecipe>(&R);
      // TODO: Transform reverse access into strided access with -1 stride.
      // TODO: Transform gather/scatter with uniform address into strided access
      // with 0 stride.
      // TODO: Transform interleave access into multiple strided accesses.
      if (!MemR || MemR->isConsecutive())
        continue;

      VPValue *Ptr = MemR->getAddr();
      // Check if this is a strided access by analyzing the address SCEV for an
      // affine addRec.
      const SCEV *PtrSCEV = vputils::getSCEVExprForVPValue(Ptr, PSE, &L);
      const SCEV *Start;
      const SCEVConstant *Step;
      // TODO: Support non-constant loop invariant stride.
      if (!match(PtrSCEV,
                 m_scev_AffineAddRec(m_SCEV(Start), m_SCEVConstant(Step),
                                     m_SpecificLoop(&L))))
        continue;

      VPValue *StoredValue = nullptr;
      Type *DataTy;
      Intrinsic::ID IntrinID;
      if (auto *StoreR = dyn_cast<VPWidenStoreRecipe>(&R)) {
        StoredValue = StoreR->getStoredValue();
        DataTy = StoredValue->getScalarType();
        IntrinID = Intrinsic::experimental_vp_strided_store;
      } else {
        auto *LoadR = cast<VPWidenLoadRecipe>(&R);
        DataTy = LoadR->getScalarType();
        IntrinID = Intrinsic::experimental_vp_strided_load;
      }

      Align Alignment = MemR->getAlign();
      auto IsProfitable = [&](ElementCount VF) {
        Type *VectorTy = toVectorTy(DataTy, VF);
        if (!Ctx.TTI.isLegalStridedLoadStore(VectorTy, Alignment))
          return false;
        const InstructionCost CurrentCost = MemR->computeCost(VF, Ctx);
        const InstructionCost StridedLoadStoreCost =
            VPWidenMemIntrinsicRecipe::computeMemIntrinsicCost(
                IntrinID, VectorTy, MemR->isMasked(), Alignment, Ctx);
        return StridedLoadStoreCost < CurrentCost;
      };

      if (!LoopVectorizationPlanner::getDecisionAndClampRange(IsProfitable,
                                                              Range))
        continue;

      // Invalidate the legacy widening decision so the cost of replaced load is
      // not counted during precomputeCosts.
      // TODO: Remove once the legacy exit cost computation is retired.
      for (ElementCount VF : Range)
        Ctx.invalidateWideningDecision(&MemR->getIngredient(), VF);

      // Get VF as i32 for the vector length operand.
      if (!I32VF) {
        VPBuilder Builder(Plan.getVectorPreheader());
        I32VF = Builder.createScalarZExtOrTrunc(
            &Plan.getVF(), Type::getInt32Ty(Plan.getContext()),
            DebugLoc::getUnknown());
      }

      VPBuilder Builder(&R);
      // Create the base pointer of strided access.
      // TODO: reuse VPDerivedIVRecipe for base pointer computation when it
      // supports a general VPValue as the start value.
      VPValue *StartVPV =
          VPSCEVExpander(Builder, *PSE.getSE(), R.getDebugLoc()).expand(Start);
      VPValue *StrideInBytes = Plan.getOrAddLiveIn(Step->getValue());
      Type *IndexTy = Plan.getDataLayout().getIndexType(Ptr->getScalarType());
      assert(IndexTy == StrideInBytes->getScalarType() &&
             "Stride type from SCEV must match the index type");
      VPValue *CanIV = Builder.createScalarZExtOrTrunc(
          VectorLoop->getCanonicalIV(), IndexTy, DebugLoc::getUnknown());
      auto *AddRecPtr = cast<SCEVAddRecExpr>(PtrSCEV);
      auto *Offset = Builder.createOverflowingOp(
          Instruction::Mul, {CanIV, StrideInBytes},
          {AddRecPtr->hasNoUnsignedWrap(), /*HasNSW=*/false});
      GEPNoWrapFlags NWFlags = AddRecPtr->hasNoUnsignedWrap()
                                   ? GEPNoWrapFlags::noUnsignedWrap()
                                   : GEPNoWrapFlags::none();
      VPValue *BasePtr = Builder.createNoWrapPtrAdd(StartVPV, Offset, NWFlags);

      // Create a new vector pointer for strided access.
      VPValue *NewPtr = Builder.createVectorPointer(
          BasePtr, Type::getInt8Ty(Plan.getContext()), StrideInBytes, NWFlags,
          R.getDebugLoc());

      VPValue *Mask = MemR->getMask();
      if (!Mask)
        Mask = Plan.getTrue();
      SmallVector<VPValue *, 5> Ops;
      if (StoredValue)
        Ops.push_back(StoredValue);
      Ops.append({NewPtr, StrideInBytes, Mask, I32VF});

      auto *StridedR = Builder.createWidenMemIntrinsic(
          IntrinID, Ops,
          StoredValue ? Type::getVoidTy(Plan.getContext()) : DataTy, Alignment,
          *MemR, R.getDebugLoc());
      if (!StoredValue)
        cast<VPWidenLoadRecipe>(&R)->replaceAllUsesWith(StridedR);
      R.eraseFromParent();
    }
  }
}