LICM.cpp

Go to the documentation of this file.

//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This pass performs loop invariant code motion, attempting to remove as much
// code from the body of a loop as possible.  It does this by either hoisting
// code into the preheader block, or by sinking code to the exit blocks if it is
// safe.  This pass also promotes must-aliased memory locations in the loop to
// live in registers, thus hoisting and sinking "invariant" loads and stores.
//
// Hoisting operations out of loops is a canonicalization transform.  It
// enables and simplifies subsequent optimizations in the middle-end.
// Rematerialization of hoisted instructions to reduce register pressure is the
// responsibility of the back-end, which has more accurate information about
// register pressure and also handles other optimizations than LICM that
// increase live-ranges.
//
// This pass uses alias analysis for two purposes:
//
//  1. Moving loop invariant loads and calls out of loops.  If we can determine
//     that a load or call inside of a loop never aliases anything stored to,
//     we can hoist it or sink it like any other instruction.
//  2. Scalar Promotion of Memory - If there is a store instruction inside of
//     the loop, we try to move the store to happen AFTER the loop instead of
//     inside of the loop.  This can only happen if a few conditions are true:
//       A. The pointer stored through is loop invariant
//       B. There are no stores or loads in the loop which _may_ alias the
//          pointer.  There are no calls in the loop which mod/ref the pointer.
//     If these conditions are true, we can promote the loads and stores in the
//     loop of the pointer to use a temporary alloca'd variable.  We then use
//     the SSAUpdater to construct the appropriate SSA form for the value.
//
//===----------------------------------------------------------------------===//
 
#include "llvm/Transforms/Scalar/LICM.h"
#include "ScalarOptions.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/PriorityWorklist.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AliasSetTracker.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/CaptureTracking.h"
#include "llvm/Analysis/DomTreeUpdater.h"
#include "llvm/Analysis/GuardUtils.h"
#include "llvm/Analysis/LazyBlockFrequencyInfo.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopIterator.h"
#include "llvm/Analysis/LoopNestAnalysis.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/MemoryLocation.h"
#include "llvm/Analysis/MemorySSA.h"
#include "llvm/Analysis/MemorySSAUpdater.h"
#include "llvm/Analysis/MustExecute.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/CFG.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/IR/PredIteratorCache.h"
#include "llvm/InitializePasses.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/LoopUtils.h"
#include "llvm/Transforms/Utils/SSAUpdater.h"
#include <algorithm>
#include <utility>
using namespace llvm;
 
namespace llvm {
class LPMUpdater;
} // namespace llvm
 
#define DEBUG_TYPE "licm"
 
STATISTIC(NumSunk, "Number of instructions sunk out of loop");
STATISTIC(NumHoisted, "Number of instructions hoisted out of loop");
STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
STATISTIC(NumPromotionCandidates, "Number of promotion candidates");
STATISTIC(NumLoadPromoted, "Number of load-only promotions");
STATISTIC(NumLoadStorePromoted, "Number of load and store promotions");
STATISTIC(NumMinMaxHoisted,
          "Number of min/max expressions hoisted out of the loop");
STATISTIC(NumGEPsHoisted,
          "Number of geps reassociated and hoisted out of the loop");
STATISTIC(NumAddSubHoisted, "Number of add/subtract expressions reassociated "
                            "and hoisted out of the loop");
STATISTIC(NumFPAssociationsHoisted, "Number of invariant FP expressions "
                                    "reassociated and hoisted out of the loop");
STATISTIC(NumIntAssociationsHoisted,
          "Number of invariant int expressions "
          "reassociated and hoisted out of the loop");
STATISTIC(NumBOAssociationsHoisted, "Number of invariant BinaryOp expressions "
                                    "reassociated and hoisted out of the loop");
 
unsigned
llvm::getLicmMssaOptCap(const cl::LibraryOptionsContext &LibraryOptions) {
  return LibraryOptions.get<ScalarOptions>().licm_mssa_optimization_cap;
}
 
unsigned llvm::getLicmMssaNoAccForPromotionCap(
    const cl::LibraryOptionsContext &LibraryOptions) {
  return LibraryOptions.get<ScalarOptions>().licm_mssa_max_acc_promotion;
}
 
static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
static bool isNotUsedOrFoldableInLoop(const Instruction &I, const Loop *CurLoop,
                                      const LoopSafetyInfo *SafetyInfo,
                                      TargetTransformInfo *TTI,
                                      bool &FoldableInLoop, bool LoopNestMode);
static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
                  BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
                  MemorySSAUpdater &MSSAU, ScalarEvolution *SE,
                  OptimizationRemarkEmitter *ORE);
static bool sink(Instruction &I, const LICMDependencies &Deps,
                 const Loop *CurLoop);
static bool isSafeToExecuteUnconditionally(
    Instruction &Inst, const DominatorTree *DT, const TargetLibraryInfo *TLI,
    const Loop *CurLoop, const LoopSafetyInfo *SafetyInfo,
    OptimizationRemarkEmitter *ORE, const Instruction *CtxI,
    AssumptionCache *AC, bool AllowSpeculation);
static bool noConflictingReadWrites(Instruction *I, MemorySSA *MSSA,
                                    AAResults *AA, Loop *CurLoop,
                                    SinkAndHoistLICMFlags &Flags);
static bool pointerInvalidatedByLoop(const SinkAndHoistDependencies &Deps,
                                     MemoryUse *MU, Loop *CurLoop,
                                     Instruction &I,
                                     SinkAndHoistLICMFlags &Flags,
                                     bool InvariantGroup);
static bool pointerInvalidatedByBlock(BasicBlock &BB, MemorySSA &MSSA,
                                      MemoryUse &MU);
/// Aggregates various functions for hoisting computations out of loop.
static bool hoistArithmetics(const ScalarOptions &Opts, Instruction &I, Loop &L,
                             const LICMDependencies &Deps);
static bool hoistInsertPastInsert(InsertElementInst *Ins, Loop *CurLoop,
                                  DominatorTree *DT, BasicBlock *HoistDest,
                                  ICFLoopSafetyInfo *SafetyInfo,
                                  MemorySSAUpdater &MSSAU, ScalarEvolution *SE,
                                  OptimizationRemarkEmitter *ORE);
static Instruction *cloneInstructionInExitBlock(
    Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
    const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater &MSSAU);
 
static void eraseInstruction(Instruction &I, const LICMDependencies &Deps);
 
static void moveInstructionBefore(Instruction &I, BasicBlock::iterator Dest,
                                  ICFLoopSafetyInfo &SafetyInfo,
                                  MemorySSAUpdater &MSSAU, ScalarEvolution *SE);
 
static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L,
                                function_ref<void(Instruction *)> Fn);
using PointersAndHasReadsOutsideSet =
    std::pair<SmallSetVector<Value *, 8>, bool>;
static SmallVector<PointersAndHasReadsOutsideSet, 0> collectPromotionCandidates(
    MemorySSA *MSSA, AliasAnalysis *AA, DominatorTree *DT,
    ICFLoopSafetyInfo *SafetyInfo,
    const SmallPtrSetImpl<const MDNode *> &LoopLocalAliasScopes, Loop *L);
 
namespace {
struct LoopInvariantCodeMotion {
  bool runOnLoop(Loop *L, const LICMDependencies &Deps,
                 bool LoopNestMode = false);
 
  LoopInvariantCodeMotion(const ScalarOptions &Opts, const LICMOptions &LO)
      : Opts(Opts),
        LicmMssaOptCap(LO.MssaOptCap.value_or(Opts.licm_mssa_optimization_cap)),
        LicmMssaNoAccForPromotionCap(LO.MssaNoAccForPromotionCap.value_or(
            Opts.licm_mssa_max_acc_promotion)),
        LicmAllowSpeculation(LO.AllowSpeculation) {}
 
private:
  const ScalarOptions &Opts;
  unsigned LicmMssaOptCap;
  unsigned LicmMssaNoAccForPromotionCap;
  bool LicmAllowSpeculation;
};
 
struct LegacyLICMPass : public LoopPass {
  static char ID; // Pass identification, replacement for typeid
  LegacyLICMPass() : LoopPass(ID) {
    initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry());
  }
 
  bool runOnLoop(Loop *L, LPPassManager &LPM) override {
    if (skipLoop(L))
      return false;
 
    LLVM_DEBUG(dbgs() << "Perform LICM on Loop with header at block "
                      << L->getHeader()->getNameOrAsOperand() << "\n");
 
    Function *F = L->getHeader()->getParent();
 
    auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
    MemorySSA *MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA();
    MemorySSAUpdater MSSAU(MSSA);
    // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
    // pass. Function analyses need to be preserved across loop transformations
    // but ORE cannot be preserved (see comment before the pass definition).
    OptimizationRemarkEmitter ORE(L->getHeader()->getParent());
    const ScalarOptions &Opts = getScalarOptions(*F);
    LoopInvariantCodeMotion LICM(Opts, LICMOptions());
    ICFLoopSafetyInfo SafetyInfo(L);
    LICMDependencies Deps{
        SinkAndHoistDependencies{
            getAnalysis<AAResultsWrapperPass>().getAAResults(),
            getAnalysis<DominatorTreeWrapperPass>().getDomTree(), MSSAU, &ORE},
        getAnalysis<AssumptionCacheTracker>().getAssumptionCache(*F),
        getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
        getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(*F),
        getAnalysis<TargetTransformInfoWrapperPass>().getTTI(*F),
        SE ? &SE->getSE() : nullptr,
        SafetyInfo};
    return LICM.runOnLoop(L, Deps);
  }
 
  /// This transformation requires natural loop information & requires that
  /// loop preheaders be inserted into the CFG...
  ///
  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addPreserved<DominatorTreeWrapperPass>();
    AU.addPreserved<LoopInfoWrapperPass>();
    AU.addRequired<TargetLibraryInfoWrapperPass>();
    AU.addRequired<MemorySSAWrapperPass>();
    AU.addPreserved<MemorySSAWrapperPass>();
    AU.addRequired<TargetTransformInfoWrapperPass>();
    AU.addRequired<AssumptionCacheTracker>();
    getLoopAnalysisUsage(AU);
    LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
    AU.addPreserved<LazyBlockFrequencyInfoPass>();
    AU.addPreserved<LazyBranchProbabilityInfoPass>();
  }
};
} // namespace
 
PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM,
                                LoopStandardAnalysisResults &AR, LPMUpdater &) {
  if (!AR.MSSA)
    reportFatalUsageError("LICM requires MemorySSA (loop-mssa)");
 
  // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
  // pass.  Function analyses need to be preserved across loop transformations
  // but ORE cannot be preserved (see comment before the pass definition).
  OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
 
  MemorySSAUpdater MSSAU(AR.MSSA);
  ICFLoopSafetyInfo SafetyInfo(&L);
  SinkAndHoistDependencies SHDeps{AR.AA, AR.DT, MSSAU, &ORE};
  LICMDependencies Deps{SHDeps, AR.AC,  AR.LI,     AR.TLI,
                        AR.TTI, &AR.SE, SafetyInfo};
  LoopInvariantCodeMotion LICM(getScalarOptions(L), Opts);
  if (!LICM.runOnLoop(&L, Deps))
    return PreservedAnalyses::all();
 
  auto PA = getLoopPassPreservedAnalyses();
  PA.preserve<MemorySSAAnalysis>();
 
  return PA;
}
 
void LICMPass::printPipeline(
    raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
  static_cast<PassInfoMixin<LICMPass> *>(this)->printPipeline(
      OS, MapClassName2PassName);
 
  OS << '<';
  OS << (Opts.AllowSpeculation ? "" : "no-") << "allowspeculation";
  OS << '>';
}
 
PreservedAnalyses LNICMPass::run(LoopNest &LN, LoopAnalysisManager &AM,
                                 LoopStandardAnalysisResults &AR,
                                 LPMUpdater &) {
  if (!AR.MSSA)
    reportFatalUsageError("LNICM requires MemorySSA (loop-mssa)");
 
  // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
  // pass.  Function analyses need to be preserved across loop transformations
  // but ORE cannot be preserved (see comment before the pass definition).
  OptimizationRemarkEmitter ORE(LN.getParent());
 
  MemorySSAUpdater MSSAU(AR.MSSA);
  Loop &OutermostLoop = LN.getOutermostLoop();
  ICFLoopSafetyInfo SafetyInfo(&OutermostLoop);
  SinkAndHoistDependencies SHDeps{AR.AA, AR.DT, MSSAU, &ORE};
  LICMDependencies Deps{SHDeps, AR.AC,  AR.LI,     AR.TLI,
                        AR.TTI, &AR.SE, SafetyInfo};
  LoopInvariantCodeMotion LICM(getScalarOptions(OutermostLoop), Opts);
 
  bool Changed = LICM.runOnLoop(&OutermostLoop, Deps, true);
 
  if (!Changed)
    return PreservedAnalyses::all();
 
  auto PA = getLoopPassPreservedAnalyses();
 
  PA.preserve<DominatorTreeAnalysis>();
  PA.preserve<LoopAnalysis>();
  PA.preserve<MemorySSAAnalysis>();
 
  return PA;
}
 
void LNICMPass::printPipeline(
    raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
  static_cast<PassInfoMixin<LNICMPass> *>(this)->printPipeline(
      OS, MapClassName2PassName);
 
  OS << '<';
  OS << (Opts.AllowSpeculation ? "" : "no-") << "allowspeculation";
  OS << '>';
}
 
char LegacyLICMPass::ID = 0;
INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",
                      false, false)
INITIALIZE_PASS_DEPENDENCY(LoopPass)
INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
INITIALIZE_PASS_DEPENDENCY(LazyBFIPass)
INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,
                    false)
 
Pass *llvm::createLICMPass() { return new LegacyLICMPass(); }
 
llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(bool IsSink, Loop &L,
                                                   MemorySSA &MSSA)
    : SinkAndHoistLICMFlags(getScalarOptions(L).licm_mssa_optimization_cap,
                            getScalarOptions(L).licm_mssa_max_acc_promotion,
                            IsSink, L, MSSA) {}
 
llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(
    unsigned LicmMssaOptCap, unsigned LicmMssaNoAccForPromotionCap, bool IsSink,
    Loop &L, MemorySSA &MSSA)
    : LicmMssaOptCap(LicmMssaOptCap),
      LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap),
      IsSink(IsSink) {
  unsigned AccessCapCount = 0;
  for (auto *BB : L.getBlocks())
    if (const auto *Accesses = MSSA.getBlockAccesses(BB))
      for (const auto &MA : *Accesses) {
        (void)MA;
        ++AccessCapCount;
        if (AccessCapCount > LicmMssaNoAccForPromotionCap) {
          NoOfMemAccTooLarge = true;
          return;
        }
      }
}
 
/// Hoist expressions out of the specified loop. Note, alias info for inner
/// loop is not preserved so it is not a good idea to run LICM multiple
/// times on one loop.
bool LoopInvariantCodeMotion::runOnLoop(Loop *L, const LICMDependencies &Deps,
                                        bool LoopNestMode) {
  bool Changed = false;
 
  assert(L->isLCSSAForm(Deps.DT) && "Loop is not in LCSSA form.");
 
  // If this loop has metadata indicating that LICM is not to be performed then
  // just exit.
  if (hasDisableLICMTransformsHint(L)) {
    return false;
  }
 
  // Don't sink stores from loops with coroutine suspend instructions.
  // LICM would sink instructions into the default destination of
  // the coroutine switch. The default destination of the switch is to
  // handle the case where the coroutine is suspended, by which point the
  // coroutine frame may have been destroyed. No instruction can be sunk there.
  // FIXME: This would unfortunately hurt the performance of coroutines, however
  // there is currently no general solution for this. Similar issues could also
  // potentially happen in other passes where instructions are being moved
  // across that edge.
  bool HasCoroSuspendInst = false;
 
  // AA metadata declared to be local to each iteration cannot be used to infer
  // alias information when promoting stores.
  SmallPtrSet<const MDNode *, 4> LoopLocalAliasScopes;
 
  for (BasicBlock *BB : L->getBlocks()) {
    for (Instruction &I : *BB) {
      using namespace PatternMatch;
      HasCoroSuspendInst |= match(&I, m_Intrinsic<Intrinsic::coro_suspend>());
 
      if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
        for (const MDOperand &Op : Decl->getScopeList()->operands())
          LoopLocalAliasScopes.insert(cast<MDNode>(Op.get()));
    }
  }
 
  MemorySSA *MSSA = Deps.MSSAU.getMemorySSA();
  SinkAndHoistLICMFlags Flags(LicmMssaOptCap, LicmMssaNoAccForPromotionCap,
                              /*IsSink=*/true, *L, *MSSA);
 
  // Get the preheader block to move instructions into...
  BasicBlock *Preheader = L->getLoopPreheader();
 
  // We want to visit all of the instructions in this loop... that are not parts
  // of our subloops (they have already had their invariants hoisted out of
  // their loop, into this loop, so there is no need to process the BODIES of
  // the subloops).
  //
  // Traverse the body of the loop in depth first order on the dominator tree so
  // that we are guaranteed to see definitions before we see uses.  This allows
  // us to sink instructions in one pass, without iteration.  After sinking
  // instructions, we perform another pass to hoist them out of the loop.
  if (L->hasDedicatedExits())
    Changed |=
        LoopNestMode
            ? sinkRegionForLoopNest(Deps.DT.getNode(L->getHeader()), Deps, L,
                                    Flags)
            : sinkRegion(Deps.DT.getNode(L->getHeader()), Deps, L, Flags);
  Flags.setIsSink(false);
  if (Preheader)
    Changed |= hoistRegion(Deps.DT.getNode(L->getHeader()), Deps, L, Flags,
                           LoopNestMode, LicmAllowSpeculation);
 
  // Now that all loop invariants have been removed from the loop, promote any
  // memory references to scalars that we can.
  // Don't sink stores from loops without dedicated block exits. Exits
  // containing indirect branches are not transformed by loop simplify,
  // make sure we catch that. An additional load may be generated in the
  // preheader for SSA updater, so also avoid sinking when no preheader
  // is available.
  if (!Opts.disable_licm_promotion && Preheader && L->hasDedicatedExits() &&
      !Flags.tooManyMemoryAccesses() && !HasCoroSuspendInst) {
    // Figure out the loop exits and their insertion points
    SmallVector<BasicBlock *, 8> ExitBlocks;
    L->getUniqueExitBlocks(ExitBlocks);
 
    // We can't insert into a catchswitch.
    bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) {
      return isa<CatchSwitchInst>(Exit->getTerminator());
    });
 
    if (!HasCatchSwitch) {
      SmallVector<BasicBlock::iterator, 8> InsertPts;
      SmallVector<MemoryAccess *, 8> MSSAInsertPts;
      InsertPts.reserve(ExitBlocks.size());
      MSSAInsertPts.reserve(ExitBlocks.size());
      PredIteratorCache PIC;
 
      // Promoting one set of accesses may make the pointers for another set
      // loop invariant, so run this in a loop.
      bool Promoted = false;
      bool LocalPromoted;
      do {
        LocalPromoted = false;
 
        // Recompute the insertion points each time we compute the promotion
        // candidates, so we don't sink past a store which was promoted in a
        // previous iteration.
        InsertPts.clear();
        MSSAInsertPts.clear();
        for (BasicBlock *ExitBlock : ExitBlocks) {
          InsertPts.push_back(ExitBlock->getFirstInsertionPt());
          MSSAInsertPts.push_back(nullptr);
        }
 
        for (auto [PointerMustAliases, HasReadsOutsideSet] :
             collectPromotionCandidates(MSSA, &Deps.AA, &Deps.DT,
                                        &Deps.SafetyInfo, LoopLocalAliasScopes,
                                        L)) {
          LocalPromoted |= promoteLoopAccessesToScalars(
              PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC,
              Deps, L, LicmAllowSpeculation, HasReadsOutsideSet);
        }
        Promoted |= LocalPromoted;
      } while (LocalPromoted);
 
      // Once we have promoted values across the loop body we have to
      // recursively reform LCSSA as any nested loop may now have values defined
      // within the loop used in the outer loop.
      // FIXME: This is really heavy handed. It would be a bit better to use an
      // SSAUpdater strategy during promotion that was LCSSA aware and reformed
      // it as it went.
      if (Promoted)
        formLCSSARecursively(*L, Deps.DT, &Deps.LI, Deps.SE);
 
      Changed |= Promoted;
    }
  }
 
  // Check that neither this loop nor its parent have had LCSSA broken. LICM is
  // specifically moving instructions across the loop boundary and so it is
  // especially in need of basic functional correctness checking here.
  assert(L->isLCSSAForm(Deps.DT) && "Loop not left in LCSSA form after LICM!");
  assert((L->isOutermost() || L->getParentLoop()->isLCSSAForm(Deps.DT)) &&
         "Parent loop not left in LCSSA form after LICM!");
 
  if (VerifyMemorySSA)
    MSSA->verifyMemorySSA();
 
  if (Changed && Deps.SE)
    Deps.SE->forgetLoopDispositions();
  return Changed;
}
 
/// Walk the specified region of the CFG (defined by all blocks dominated by
/// the specified block, and that are in the current loop) in reverse depth
/// first order w.r.t the DominatorTree.  This allows us to visit uses before
/// definitions, allowing us to sink a loop body in one pass without iteration.
///
bool llvm::sinkRegion(DomTreeNode *N, const LICMDependencies &Deps,
                      Loop *CurLoop, SinkAndHoistLICMFlags &Flags,
                      Loop *OutermostLoop) {
 
  // Verify inputs.
  assert(N != nullptr && CurLoop != nullptr &&
         "Unexpected input to sinkRegion.");
 
  // We want to visit children before parents. We will enqueue all the parents
  // before their children in the worklist and process the worklist in reverse
  // order.
  SmallVector<BasicBlock *, 16> Worklist =
      collectChildrenInLoop(&Deps.DT, N, CurLoop);
 
  bool Changed = false;
  for (BasicBlock *BB : reverse(Worklist)) {
    // subloop (which would already have been processed).
    if (inSubLoop(BB, CurLoop, &Deps.LI))
      continue;
 
    for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
      Instruction &I = *--II;
 
      // The instruction is not used in the loop if it is dead.  In this case,
      // we just delete it instead of sinking it.
      if (isInstructionTriviallyDead(&I, &Deps.TLI)) {
        LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
        salvageKnowledge(&I);
        salvageDebugInfo(I);
        ++II;
        eraseInstruction(I, Deps);
        Changed = true;
        continue;
      }
 
      // Check to see if we can sink this instruction to the exit blocks
      // of the loop.  We can do this if the all users of the instruction are
      // outside of the loop.  In this case, it doesn't even matter if the
      // operands of the instruction are loop invariant.
      //
      bool FoldableInLoop = false;
      bool LoopNestMode = OutermostLoop != nullptr;
      if (!I.mayHaveSideEffects() &&
          isNotUsedOrFoldableInLoop(I, LoopNestMode ? OutermostLoop : CurLoop,
                                    &Deps.SafetyInfo, &Deps.TTI, FoldableInLoop,
                                    LoopNestMode) &&
          canSinkOrHoistInst(I, Deps, CurLoop, true, Flags)) {
        if (sink(I, Deps, CurLoop)) {
          if (!FoldableInLoop) {
            ++II;
            salvageDebugInfo(I);
            eraseInstruction(I, Deps);
          }
          Changed = true;
        }
      }
    }
  }
  if (VerifyMemorySSA)
    Deps.MSSAU.getMemorySSA()->verifyMemorySSA();
  return Changed;
}
 
bool llvm::sinkRegionForLoopNest(DomTreeNode *N, const LICMDependencies &Deps,
                                 Loop *CurLoop, SinkAndHoistLICMFlags &Flags) {
 
  bool Changed = false;
  SmallPriorityWorklist<Loop *, 4> Worklist;
  Worklist.insert(CurLoop);
  appendLoopsToWorklist(*CurLoop, Worklist);
  while (!Worklist.empty()) {
    Loop *L = Worklist.pop_back_val();
    Changed |=
        sinkRegion(Deps.DT.getNode(L->getHeader()), Deps, L, Flags, CurLoop);
  }
  return Changed;
}
 
/// Walk the specified region of the CFG (defined by all blocks dominated by
/// the specified block, and that are in the current loop) in depth first
/// order w.r.t the DominatorTree.  This allows us to visit definitions before
/// uses, allowing us to hoist a loop body in one pass without iteration.
///
bool llvm::hoistRegion(DomTreeNode *N, const LICMDependencies &Deps,
                       Loop *CurLoop, SinkAndHoistLICMFlags &Flags,
                       bool LoopNestMode, bool AllowSpeculation) {
  // Verify inputs.
  assert(N != nullptr && CurLoop != nullptr &&
         "Unexpected input to hoistRegion.");
 
  const ScalarOptions &Opts = getScalarOptions(*CurLoop);
  LoopBlocksRPO Worklist(CurLoop);
  Worklist.perform(&Deps.LI);
  bool Changed = false;
  BasicBlock *Preheader = CurLoop->getLoopPreheader();
  for (BasicBlock *BB : Worklist) {
    // Only need to process the contents of this block if it is not part of a
    // subloop (which would already have been processed).
    if (!LoopNestMode && inSubLoop(BB, CurLoop, &Deps.LI))
      continue;
 
    for (Instruction &I : llvm::make_early_inc_range(*BB)) {
      // Try hoisting the instruction out to the preheader.  We can only do
      // this if all of the operands of the instruction are loop invariant and
      // if it is safe to hoist the instruction.
      if (CurLoop->hasLoopInvariantOperands(&I) &&
          canSinkOrHoistInst(I, Deps, CurLoop, true, Flags) &&
          isSafeToExecuteUnconditionally(
              I, &Deps.DT, &Deps.TLI, CurLoop, &Deps.SafetyInfo, Deps.ORE,
              Preheader->getTerminator(), &Deps.AC, AllowSpeculation)) {
        hoist(I, &Deps.DT, CurLoop, Preheader, &Deps.SafetyInfo, Deps.MSSAU,
              Deps.SE, Deps.ORE);
        Changed = true;
        continue;
      }
 
      if (auto *Ins = dyn_cast<InsertElementInst>(&I))
        if (hoistInsertPastInsert(Ins, CurLoop, &Deps.DT, Preheader,
                                  &Deps.SafetyInfo, Deps.MSSAU, Deps.SE,
                                  Deps.ORE)) {
          Changed = true;
          continue;
        }
 
      // Attempt to remove floating point division out of the loop by
      // converting it to a reciprocal multiplication.
      if (I.getOpcode() == Instruction::FDiv && I.hasAllowReciprocal() &&
          CurLoop->isLoopInvariant(I.getOperand(1))) {
        auto Divisor = I.getOperand(1);
        auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0);
        auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor);
        ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags());
        Deps.SafetyInfo.insertInstructionTo(ReciprocalDivisor, I.getParent());
        ReciprocalDivisor->insertBefore(I.getIterator());
        ReciprocalDivisor->setDebugLoc(I.getDebugLoc());
 
        auto Product =
            BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor);
        Product->setFastMathFlags(I.getFastMathFlags());
        Deps.SafetyInfo.insertInstructionTo(Product, I.getParent());
        Product->insertAfter(I.getIterator());
        Product->setDebugLoc(I.getDebugLoc());
        I.replaceAllUsesWith(Product);
        eraseInstruction(I, Deps);
 
        hoist(*ReciprocalDivisor, &Deps.DT, CurLoop, Preheader,
              &Deps.SafetyInfo, Deps.MSSAU, Deps.SE, Deps.ORE);
        Changed = true;
        continue;
      }
 
      auto IsInvariantStart = [&](Instruction &I) {
        using namespace PatternMatch;
        return I.use_empty() &&
               match(&I, m_Intrinsic<Intrinsic::invariant_start>());
      };
      auto MustExecuteWithoutWritesBefore = [&](Instruction &I) {
        return Deps.SafetyInfo.isGuaranteedToExecute(I, &Deps.DT) &&
               Deps.SafetyInfo.doesNotWriteMemoryBefore(I);
      };
      if ((IsInvariantStart(I) || isGuard(&I)) &&
          CurLoop->hasLoopInvariantOperands(&I) &&
          MustExecuteWithoutWritesBefore(I)) {
        hoist(I, &Deps.DT, CurLoop, Preheader, &Deps.SafetyInfo, Deps.MSSAU,
              Deps.SE, Deps.ORE);
        Changed = true;
        continue;
      }
 
      // Try to reassociate instructions so that part of computations can be
      // done out of loop.
      if (hoistArithmetics(Opts, I, *CurLoop, Deps)) {
        Changed = true;
        continue;
      }
    }
  }
 
  if (VerifyMemorySSA)
    Deps.MSSAU.getMemorySSA()->verifyMemorySSA();
 
    // Now that we've finished hoisting make sure that LI and DT are still
    // valid.
#ifdef EXPENSIVE_CHECKS
  if (Changed) {
    assert(Deps.DT.verify(DominatorTree::VerificationLevel::Fast) &&
           "Dominator tree verification failed");
    Deps.LI.verify();
  }
#endif
 
  return Changed;
}
 
static std::optional<uint64_t>
getConstantInsertionIndex(InsertElementInst *Ins) {
  // Must have constant insertion lane.
  auto *InsertedIdxCI = dyn_cast<ConstantInt>(Ins->getOperand(2));
  if (!InsertedIdxCI)
    return std::nullopt;
  auto *VecTy = cast<VectorType>(Ins->getType());
 
  // Avoid hoisting past out of bounds inserts.
  if (InsertedIdxCI->isNegative() ||
      InsertedIdxCI->getValue().uge(
          VecTy->getElementCount().getKnownMinValue()))
    return std::nullopt;
  return InsertedIdxCI->getValue().getLimitedValue();
}
 
static bool hoistInsertPastInsert(InsertElementInst *Ins, Loop *CurLoop,
                                  DominatorTree *DT, BasicBlock *HoistDest,
                                  ICFLoopSafetyInfo *SafetyInfo,
                                  MemorySSAUpdater &MSSAU, ScalarEvolution *SE,
                                  OptimizationRemarkEmitter *ORE) {
  // Canonicalize:
  //   %inner = insertelement %base, %variant, C1
  //   %outer = insertelement %inner, %invariant, C2
  // into:
  //   %outer = insertelement %base, %invariant, C2
  //   %inner = insertelement %outer, %variant, C1
  // so we can hoist %outer
 
  // The instruction we are hoisting must have invariant insertion data
  Value *InsertedElt = Ins->getOperand(1);
  if (!CurLoop->isLoopInvariant(InsertedElt))
    return false;
 
  std::optional<uint64_t> HoistIdx = getConstantInsertionIndex(Ins);
  if (!HoistIdx)
    return false;
 
  InsertElementInst *Inner = Ins;
  while (!CurLoop->isLoopInvariant(Inner->getOperand(0))) {
    // If the inner value isn't invariant, check to see if it is another insert
    // All instructions in the chain must be in the same basic block
    auto *InnerIns = dyn_cast<InsertElementInst>(Inner->getOperand(0));
    if (!InnerIns || InnerIns->getParent() != Ins->getParent())
      return false;
 
    // Make sure not hoisting past insertions into the same lane
    std::optional<uint64_t> InsertIdx = getConstantInsertionIndex(InnerIns);
    if (!InsertIdx || *InsertIdx == *HoistIdx)
      return false;
 
    // Instruction being hoisted past must only have one use
    if (!InnerIns->hasOneUse())
      return false;
 
    Inner = InnerIns;
  }
 
  // Base case of `insertelement <4 x i8> %invar0, i8 %invar1, i32 2` handled in
  // base LICM logic
  if (Inner == Ins)
    return false;
 
  Ins->replaceAllUsesWith(Ins->getOperand(0));
  Ins->moveBefore(Inner->getIterator());
  Ins->setOperand(0, Inner->getOperand(0));
  Inner->setOperand(0, Ins);
  hoist(*Ins, DT, CurLoop, HoistDest, SafetyInfo, MSSAU, SE, ORE);
  return true;
}
 
// Return true if LI is invariant within scope of the loop. LI is invariant if
// CurLoop is dominated by an invariant.start representing the same memory
// location and size as the memory location LI loads from, and also the
// invariant.start has no uses.
static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT,
                                  Loop *CurLoop) {
  Value *Addr = LI->getPointerOperand();
  const DataLayout &DL = LI->getDataLayout();
  const TypeSize LocSizeInBits = DL.getTypeSizeInBits(LI->getType());
 
  // It is not currently possible for clang to generate an invariant.start
  // intrinsic with scalable vector types because we don't support thread local
  // sizeless types and we don't permit sizeless types in structs or classes.
  // Furthermore, even if support is added for this in future the intrinsic
  // itself is defined to have a size of -1 for variable sized objects. This
  // makes it impossible to verify if the intrinsic envelops our region of
  // interest. For example, both <vscale x 32 x i8> and <vscale x 16 x i8>
  // types would have a -1 parameter, but the former is clearly double the size
  // of the latter.
  if (LocSizeInBits.isScalable())
    return false;
 
  // If we've ended up at a global/constant, bail. We shouldn't be looking at
  // uselists for non-local Values in a loop pass.
  if (isa<Constant>(Addr))
    return false;
 
  const ScalarOptions &Opts = getScalarOptions(*LI->getFunction());
  unsigned UsesVisited = 0;
  // Traverse all uses of the load operand value, to see if invariant.start is
  // one of the uses, and whether it dominates the load instruction.
  for (auto *U : Addr->users()) {
    // Avoid traversing for Load operand with high number of users.
    if (++UsesVisited > Opts.licm_max_num_uses_traversed)
      return false;
    IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
    // If there are escaping uses of invariant.start instruction, the load maybe
    // non-invariant.
    if (!II || II->getIntrinsicID() != Intrinsic::invariant_start ||
        !II->use_empty())
      continue;
    ConstantInt *InvariantSize = cast<ConstantInt>(II->getArgOperand(0));
    // The intrinsic supports having a -1 argument for variable sized objects
    // so we should check for that here.
    if (InvariantSize->isNegative())
      continue;
    uint64_t InvariantSizeInBits = InvariantSize->getSExtValue() * 8;
    // Confirm the invariant.start location size contains the load operand size
    // in bits. Also, the invariant.start should dominate the load, and we
    // should not hoist the load out of a loop that contains this dominating
    // invariant.start.
    if (LocSizeInBits.getFixedValue() <= InvariantSizeInBits &&
        DT->properlyDominates(II->getParent(), CurLoop->getHeader()))
      return true;
  }
 
  return false;
}
 
/// Return true if-and-only-if we know how to (mechanically) both hoist and
/// sink a given instruction out of a loop.  Does not address legality
/// concerns such as aliasing or speculation safety.
static bool isHoistableAndSinkableInst(Instruction &I) {
  // Only these instructions are hoistable/sinkable.
  return (isa<LoadInst>(I) || isa<StoreInst>(I) || isa<CallInst>(I) ||
          isa<FenceInst>(I) || isa<CastInst>(I) || isa<UnaryOperator>(I) ||
          isa<BinaryOperator>(I) || isa<SelectInst>(I) ||
          isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
          isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
          isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) ||
          isa<InsertValueInst>(I) || isa<FreezeInst>(I));
}
 
/// Return true if I is the only Instruction with a MemoryAccess in L.
static bool isOnlyMemoryAccess(const Instruction *I, const Loop *L,
                               const MemorySSAUpdater &MSSAU) {
  for (auto *BB : L->getBlocks())
    if (auto *Accs = MSSAU.getMemorySSA()->getBlockAccesses(BB)) {
      int NotAPhi = 0;
      for (const auto &Acc : *Accs) {
        if (isa<MemoryPhi>(&Acc))
          continue;
        const auto *MUD = cast<MemoryUseOrDef>(&Acc);
        if (MUD->getMemoryInst() != I || NotAPhi++ == 1)
          return false;
      }
    }
  return true;
}
 
static MemoryAccess *getClobberingMemoryAccess(MemorySSA &MSSA,
                                               BatchAAResults &BAA,
                                               SinkAndHoistLICMFlags &Flags,
                                               MemoryUseOrDef *MA) {
  // See declaration of SetLicmMssaOptCap for usage details.
  if (Flags.tooManyClobberingCalls())
    return MA->getDefiningAccess();
 
  MemoryAccess *Source =
      MSSA.getSkipSelfWalker()->getClobberingMemoryAccess(MA, BAA);
  Flags.incrementClobberingCalls();
  return Source;
}
 
bool llvm::canHoistLoad(LoadInst &LI, const SinkAndHoistDependencies &Deps,
                        Loop *CurLoop, bool TargetExecutesOncePerLoop,
                        SinkAndHoistLICMFlags &Flags) {
  if (!LI.isUnordered())
    return false; // Don't sink/hoist volatile or ordered atomic loads!
 
  // Loads from constant memory are always safe to move, even if they end up
  // in the same alias set as something that ends up being modified.
  if (!isModSet(Deps.AA.getModRefInfoMask(LI.getOperand(0))))
    return true;
  if (LI.hasMetadata(LLVMContext::MD_invariant_load))
    return true;
 
  if (LI.isAtomic() && !TargetExecutesOncePerLoop)
    return false; // Don't risk duplicating unordered loads
 
  // This checks for an invariant.start dominating the load.
  if (isLoadInvariantInLoop(&LI, &Deps.DT, CurLoop))
    return true;
 
  auto *MU = cast<MemoryUse>(Deps.MSSAU.getMemorySSA()->getMemoryAccess(&LI));
 
  bool InvariantGroup = LI.hasMetadata(LLVMContext::MD_invariant_group);
 
  bool Invalidated =
      pointerInvalidatedByLoop(Deps, MU, CurLoop, LI, Flags, InvariantGroup);
  // Check loop-invariant address because this may also be a sinkable load
  // whose address is not necessarily loop-invariant.
  if (Deps.ORE && Invalidated &&
      CurLoop->isLoopInvariant(LI.getPointerOperand()))
    Deps.ORE->emit([&]() {
      return OptimizationRemarkMissed(
                 DEBUG_TYPE, "LoadWithLoopInvariantAddressInvalidated", &LI)
             << "failed to move load with loop-invariant address "
                "because the loop may invalidate its value";
    });
 
  return !Invalidated;
}
 
bool llvm::canSinkOrHoistInst(Instruction &I,
                              const SinkAndHoistDependencies &Deps,
                              Loop *CurLoop, bool TargetExecutesOncePerLoop,
                              SinkAndHoistLICMFlags &Flags) {
  // If we don't understand the instruction, bail early.
  if (!isHoistableAndSinkableInst(I))
    return false;
 
  MemorySSA *MSSA = Deps.MSSAU.getMemorySSA();
  // Loads have extra constraints we have to verify before we can hoist them.
  if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
    return canHoistLoad(*LI, Deps, CurLoop, TargetExecutesOncePerLoop, Flags);
  } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
    // Don't sink calls which can throw.
    if (CI->mayThrow())
      return false;
 
    // Convergent attribute has been used on operations that involve
    // inter-thread communication which results are implicitly affected by the
    // enclosing control flows. It is not safe to hoist or sink such operations
    // across control flow.
    if (CI->isConvergent())
      return false;
 
    // FIXME: Current LLVM IR semantics don't work well with coroutines and
    // thread local globals. We currently treat getting the address of a thread
    // local global as not accessing memory, even though it may not be a
    // constant throughout a function with coroutines. Remove this check after
    // we better model semantics of thread local globals.
    if (CI->getFunction()->isPresplitCoroutine())
      return false;
 
    using namespace PatternMatch;
    if (match(CI, m_Intrinsic<Intrinsic::assume>()))
      // Assumes don't actually alias anything or throw
      return true;
 
    // Handle simple cases by querying alias analysis.
    MemoryEffects Behavior = Deps.AA.getMemoryEffects(CI);
 
    if (Behavior.doesNotAccessMemory())
      return true;
    if (Behavior.onlyReadsMemory()) {
      // Might have stale MemoryDef for call that was later inferred to be
      // read-only.
      auto *MU = dyn_cast<MemoryUse>(MSSA->getMemoryAccess(CI));
      if (!MU)
        return false;
 
      // If we can prove there are no writes to the memory read by the call, we
      // can hoist or sink.
      return !pointerInvalidatedByLoop(Deps, MU, CurLoop, I, Flags,
                                       /*InvariantGroup=*/false);
    }
 
    if (Behavior.onlyWritesMemory()) {
      // can hoist or sink if there are no conflicting read/writes to the
      // memory location written to by the call.
      return noConflictingReadWrites(CI, MSSA, &Deps.AA, CurLoop, Flags);
    }
 
    return false;
  } else if (auto *FI = dyn_cast<FenceInst>(&I)) {
    // Fences alias (most) everything to provide ordering.  For the moment,
    // just give up if there are any other memory operations in the loop.
    return isOnlyMemoryAccess(FI, CurLoop, Deps.MSSAU);
  } else if (auto *SI = dyn_cast<StoreInst>(&I)) {
    if (!SI->isUnordered())
      return false; // Don't sink/hoist volatile or ordered atomic store!
 
    // We can only hoist a store that we can prove writes a value which is not
    // read or overwritten within the loop.  For those cases, we fallback to
    // load store promotion instead.  TODO: We can extend this to cases where
    // there is exactly one write to the location and that write dominates an
    // arbitrary number of reads in the loop.
    if (isOnlyMemoryAccess(SI, CurLoop, Deps.MSSAU))
      return true;
    return noConflictingReadWrites(SI, MSSA, &Deps.AA, CurLoop, Flags);
  }
 
  assert(!I.mayReadOrWriteMemory() && "unhandled aliasing");
 
  // We've established mechanical ability and aliasing, it's up to the caller
  // to check fault safety
  return true;
}
 
/// Returns true if a PHINode is a trivially replaceable with an
/// Instruction.
/// This is true when all incoming values are that instruction.
/// This pattern occurs most often with LCSSA PHI nodes.
///
static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) {
  for (const Value *IncValue : PN.incoming_values())
    if (IncValue != &I)
      return false;
 
  return true;
}
 
/// Return true if the instruction is foldable in the loop.
static bool isFoldableInLoop(const Instruction &I, const Loop *CurLoop,
                         const TargetTransformInfo *TTI) {
  if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
    InstructionCost CostI =
        TTI->getInstructionCost(&I, TargetTransformInfo::TCK_SizeAndLatency);
    if (CostI != TargetTransformInfo::TCC_Free)
      return false;
    // For a GEP, we cannot simply use getInstructionCost because currently
    // it optimistically assumes that a GEP will fold into addressing mode
    // regardless of its users.
    const BasicBlock *BB = GEP->getParent();
    for (const User *U : GEP->users()) {
      const Instruction *UI = cast<Instruction>(U);
      if (CurLoop->contains(UI) &&
          (BB != UI->getParent() ||
           (!isa<StoreInst>(UI) && !isa<LoadInst>(UI))))
        return false;
    }
    return true;
  }
 
  return false;
}
 
/// Return true if the only users of this instruction are outside of
/// the loop. If this is true, we can sink the instruction to the exit
/// blocks of the loop.
///
/// We also return true if the instruction could be folded away in lowering.
/// (e.g.,  a GEP can be folded into a load as an addressing mode in the loop).
static bool isNotUsedOrFoldableInLoop(const Instruction &I, const Loop *CurLoop,
                                      const LoopSafetyInfo *SafetyInfo,
                                      TargetTransformInfo *TTI,
                                      bool &FoldableInLoop, bool LoopNestMode) {
  bool IsFoldable = isFoldableInLoop(I, CurLoop, TTI);
  for (const User *U : I.users()) {
    const Instruction *UI = cast<Instruction>(U);
    if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
      const BasicBlock *BB = PN->getParent();
      // We cannot sink uses in catchswitches.
      if (isa<CatchSwitchInst>(BB->getTerminator()))
        return false;
 
      // We need to sink a callsite to a unique funclet.  Avoid sinking if the
      // phi use is too muddled.
      if (isa<CallInst>(I)) {
        const auto &BlockColors = SafetyInfo->getBlockColors();
        if (!BlockColors.empty() &&
            BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
          return false;
      }
 
      if (LoopNestMode) {
        while (isa<PHINode>(UI) && UI->hasOneUser() &&
               UI->getNumOperands() == 1) {
          if (!CurLoop->contains(UI))
            break;
          UI = cast<Instruction>(UI->user_back());
        }
      }
    }
 
    if (CurLoop->contains(UI)) {
      if (IsFoldable) {
        FoldableInLoop = true;
        continue;
      }
      return false;
    }
  }
  return true;
}
 
static Instruction *cloneInstructionInExitBlock(
    Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
    const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater &MSSAU) {
  Instruction *New;
  if (auto *CI = dyn_cast<CallInst>(&I)) {
    const auto &BlockColors = SafetyInfo->getBlockColors();
 
    // Sinking call-sites need to be handled differently from other
    // instructions.  The cloned call-site needs a funclet bundle operand
    // appropriate for its location in the CFG.
    SmallVector<OperandBundleDef, 1> OpBundles;
    for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
         BundleIdx != BundleEnd; ++BundleIdx) {
      OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
      if (Bundle.getTagID() == LLVMContext::OB_funclet)
        continue;
 
      OpBundles.emplace_back(Bundle);
    }
 
    if (!BlockColors.empty()) {
      const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
      assert(CV.size() == 1 && "non-unique color for exit block!");
      BasicBlock *BBColor = CV.front();
      BasicBlock::iterator EHPad = BBColor->getFirstNonPHIIt();
      if (EHPad->isEHPad())
        OpBundles.emplace_back("funclet", &*EHPad);
    }
 
    New = CallInst::Create(CI, OpBundles);
    New->copyMetadata(*CI);
  } else {
    New = I.clone();
  }
 
  New->insertInto(&ExitBlock, ExitBlock.getFirstInsertionPt());
  if (!I.getName().empty())
    New->setName(I.getName() + ".le");
 
  if (MSSAU.getMemorySSA()->getMemoryAccess(&I)) {
    // Create a new MemoryAccess and let MemorySSA set its defining access.
    // After running some passes, MemorySSA might be outdated, and the
    // instruction `I` may have become a non-memory touching instruction.
    MemoryAccess *NewMemAcc = MSSAU.createMemoryAccessInBB(
        New, nullptr, New->getParent(), MemorySSA::Beginning,
        /*CreationMustSucceed=*/false);
    if (NewMemAcc) {
      if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc))
        MSSAU.insertDef(MemDef, /*RenameUses=*/true);
      else {
        auto *MemUse = cast<MemoryUse>(NewMemAcc);
        MSSAU.insertUse(MemUse, /*RenameUses=*/true);
      }
    }
  }
 
  // Build LCSSA PHI nodes for any in-loop operands (if legal).  Note that
  // this is particularly cheap because we can rip off the PHI node that we're
  // replacing for the number and blocks of the predecessors.
  // OPT: If this shows up in a profile, we can instead finish sinking all
  // invariant instructions, and then walk their operands to re-establish
  // LCSSA. That will eliminate creating PHI nodes just to nuke them when
  // sinking bottom-up.
  for (Use &Op : New->operands())
    if (LI->wouldBeOutOfLoopUseRequiringLCSSA(Op.get(), PN.getParent())) {
      auto *OInst = cast<Instruction>(Op.get());
      PHINode *OpPN =
          PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
                          OInst->getName() + ".lcssa");
      OpPN->insertBefore(ExitBlock.begin());
      for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
        OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
      Op = OpPN;
    }
  return New;
}
 
static void eraseInstruction(Instruction &I, const LICMDependencies &Deps) {
  Deps.MSSAU.removeMemoryAccess(&I);
  Deps.SafetyInfo.removeInstruction(&I);
  I.eraseFromParent();
}
 
static void moveInstructionBefore(Instruction &I, BasicBlock::iterator Dest,
                                  ICFLoopSafetyInfo &SafetyInfo,
                                  MemorySSAUpdater &MSSAU,
                                  ScalarEvolution *SE) {
  SafetyInfo.removeInstruction(&I);
  SafetyInfo.insertInstructionTo(&I, Dest->getParent());
  I.moveBefore(*Dest->getParent(), Dest);
  if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>(
          MSSAU.getMemorySSA()->getMemoryAccess(&I)))
    MSSAU.moveToPlace(OldMemAcc, Dest->getParent(),
                      MemorySSA::BeforeTerminator);
  if (SE)
    SE->forgetBlockAndLoopDispositions(&I);
}
 
static Instruction *sinkThroughTriviallyReplaceablePHI(
    PHINode *TPN, Instruction *I, LoopInfo *LI,
    SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies,
    const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop,
    MemorySSAUpdater &MSSAU) {
  assert(isTriviallyReplaceablePHI(*TPN, *I) &&
         "Expect only trivially replaceable PHI");
  BasicBlock *ExitBlock = TPN->getParent();
  auto [It, Inserted] = SunkCopies.try_emplace(ExitBlock);
  if (Inserted)
    It->second = cloneInstructionInExitBlock(*I, *ExitBlock, *TPN, LI,
                                             SafetyInfo, MSSAU);
  return It->second;
}
 
static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) {
  BasicBlock *BB = PN->getParent();
  if (!BB->canSplitPredecessors())
    return false;
  // It's not impossible to split EHPad blocks, but if BlockColors already exist
  // it require updating BlockColors for all offspring blocks accordingly. By
  // skipping such corner case, we can make updating BlockColors after splitting
  // predecessor fairly simple.
  if (!SafetyInfo->getBlockColors().empty() &&
      BB->getFirstNonPHIIt()->isEHPad())
    return false;
  for (BasicBlock *BBPred : predecessors(BB)) {
    if (isa<IndirectBrInst>(BBPred->getTerminator()))
      return false;
  }
  return true;
}
 
static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT,
                                        LoopInfo *LI, const Loop *CurLoop,
                                        LoopSafetyInfo *SafetyInfo,
                                        MemorySSAUpdater *MSSAU) {
#ifndef NDEBUG
  SmallVector<BasicBlock *, 32> ExitBlocks;
  CurLoop->getUniqueExitBlocks(ExitBlocks);
  SmallPtrSet<BasicBlock *, 32> ExitBlockSet(llvm::from_range, ExitBlocks);
#endif
  BasicBlock *ExitBB = PN->getParent();
  assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.");
 
  // Split predecessors of the loop exit to make instructions in the loop are
  // exposed to exit blocks through trivially replaceable PHIs while keeping the
  // loop in the canonical form where each predecessor of each exit block should
  // be contained within the loop. For example, this will convert the loop below
  // from
  //
  // LB1:
  //   %v1 =
  //   br %LE, %LB2
  // LB2:
  //   %v2 =
  //   br %LE, %LB1
  // LE:
  //   %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable
  //
  // to
  //
  // LB1:
  //   %v1 =
  //   br %LE.split, %LB2
  // LB2:
  //   %v2 =
  //   br %LE.split2, %LB1
  // LE.split:
  //   %p1 = phi [%v1, %LB1]  <-- trivially replaceable
  //   br %LE
  // LE.split2:
  //   %p2 = phi [%v2, %LB2]  <-- trivially replaceable
  //   br %LE
  // LE:
  //   %p = phi [%p1, %LE.split], [%p2, %LE.split2]
  //
  const auto &BlockColors = SafetyInfo->getBlockColors();
  SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB));
  DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
  while (!PredBBs.empty()) {
    BasicBlock *PredBB = *PredBBs.begin();
    assert(CurLoop->contains(PredBB) &&
           "Expect all predecessors are in the loop");
    if (PN->getBasicBlockIndex(PredBB) >= 0) {
      BasicBlock *NewPred = SplitBlockPredecessors(
          ExitBB, PredBB, ".split.loop.exit", &DTU, LI, MSSAU, true);
      // Since we do not allow splitting EH-block with BlockColors in
      // canSplitPredecessors(), we can simply assign predecessor's color to
      // the new block.
      if (!BlockColors.empty())
        // Grab a reference to the ColorVector to be inserted before getting the
        // reference to the vector we are copying because inserting the new
        // element in BlockColors might cause the map to be reallocated.
        SafetyInfo->copyColors(NewPred, PredBB);
    }
    PredBBs.remove(PredBB);
  }
}
 
/// When an instruction is found to only be used outside of the loop, this
/// function moves it to the exit blocks and patches up SSA form as needed.
/// This method is guaranteed to remove the original instruction from its
/// position, and may either delete it or move it to outside of the loop.
///
static bool sink(Instruction &I, const LICMDependencies &Deps,
                 const Loop *CurLoop) {
  bool Changed = false;
  LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
 
  // Iterate over users to be ready for actual sinking. Replace users via
  // unreachable blocks with undef and make all user PHIs trivially replaceable.
  SmallPtrSet<Instruction *, 8> VisitedUsers;
  for (Instruction::user_iterator UI = I.user_begin(), UE = I.user_end();
       UI != UE;) {
    auto *User = cast<Instruction>(*UI);
    Use &U = UI.getUse();
    ++UI;
 
    if (VisitedUsers.count(User) || CurLoop->contains(User))
      continue;
 
    if (!Deps.DT.isReachableFromEntry(User->getParent())) {
      U = PoisonValue::get(I.getType());
      Changed = true;
      continue;
    }
 
    // The user must be a PHI node.
    PHINode *PN = cast<PHINode>(User);
 
    // Surprisingly, instructions can be used outside of loops without any
    // exits.  This can only happen in PHI nodes if the incoming block is
    // unreachable.
    BasicBlock *BB = PN->getIncomingBlock(U);
    if (!Deps.DT.isReachableFromEntry(BB)) {
      U = PoisonValue::get(I.getType());
      Changed = true;
      continue;
    }
 
    VisitedUsers.insert(PN);
    if (isTriviallyReplaceablePHI(*PN, I))
      continue;
 
    if (!canSplitPredecessors(PN, &Deps.SafetyInfo))
      return Changed;
 
    // Split predecessors of the PHI so that we can make users trivially
    // replaceable.
    splitPredecessorsOfLoopExit(PN, &Deps.DT, &Deps.LI, CurLoop,
                                &Deps.SafetyInfo, &Deps.MSSAU);
 
    // Should rebuild the iterators, as they may be invalidated by
    // splitPredecessorsOfLoopExit().
    UI = I.user_begin();
    UE = I.user_end();
  }
 
  if (VisitedUsers.empty())
    return Changed;
 
  Deps.ORE->emit([&]() {
    return OptimizationRemark(DEBUG_TYPE, "InstSunk", &I)
           << "sinking " << ore::NV("Inst", &I);
  });
  if (isa<LoadInst>(I))
    ++NumMovedLoads;
  else if (isa<CallInst>(I))
    ++NumMovedCalls;
  ++NumSunk;
 
#ifndef NDEBUG
  SmallVector<BasicBlock *, 32> ExitBlocks;
  CurLoop->getUniqueExitBlocks(ExitBlocks);
  SmallPtrSet<BasicBlock *, 32> ExitBlockSet(llvm::from_range, ExitBlocks);
#endif
 
  // Clones of this instruction. Don't create more than one per exit block!
  SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
 
  // If this instruction is only used outside of the loop, then all users are
  // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
  // the instruction.
  // First check if I is worth sinking for all uses. Sink only when it is worth
  // across all uses.
  SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end());
  for (auto *UI : Users) {
    auto *User = cast<Instruction>(UI);
 
    if (CurLoop->contains(User))
      continue;
 
    PHINode *PN = cast<PHINode>(User);
    assert(ExitBlockSet.count(PN->getParent()) &&
           "The LCSSA PHI is not in an exit block!");
 
    // The PHI must be trivially replaceable.
    Instruction *New = sinkThroughTriviallyReplaceablePHI(
        PN, &I, &Deps.LI, SunkCopies, &Deps.SafetyInfo, CurLoop, Deps.MSSAU);
    // As we sink the instruction out of the BB, drop its debug location.
    New->dropLocation();
    PN->replaceAllUsesWith(New);
    eraseInstruction(*PN, Deps);
    Changed = true;
  }
  return Changed;
}
 
/// When an instruction is found to only use loop invariant operands that
/// is safe to hoist, this instruction is called to do the dirty work.
///
static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
                  BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
                  MemorySSAUpdater &MSSAU, ScalarEvolution *SE,
                  OptimizationRemarkEmitter *ORE) {
  LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getNameOrAsOperand() << ": "
                    << I << "\n");
  ORE->emit([&]() {
    return OptimizationRemark(DEBUG_TYPE, "Hoisted", &I) << "hoisting "
                                                         << ore::NV("Inst", &I);
  });
 
  // Metadata can be dependent on conditions we are hoisting above.
  // Conservatively strip all metadata on the instruction unless we were
  // guaranteed to execute I if we entered the loop, in which case the metadata
  // is valid in the loop preheader.
  // Similarly, If I is a call and it is not guaranteed to execute in the loop,
  // then moving to the preheader means we should strip attributes on the call
  // that can cause UB since we may be hoisting above conditions that allowed
  // inferring those attributes. They may not be valid at the preheader.
  if ((I.hasMetadataOtherThanDebugLoc() || isa<CallInst>(I)) &&
      // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning
      // time in isGuaranteedToExecute if we don't actually have anything to
      // drop.  It is a compile time optimization, not required for correctness.
      !SafetyInfo->isGuaranteedToExecute(I, DT)) {
    I.dropUBImplyingAttrsAndMetadata();
  }
 
  if (isa<PHINode>(I))
    // Move the new node to the end of the phi list in the destination block.
    moveInstructionBefore(I, Dest->getFirstNonPHIIt(), *SafetyInfo, MSSAU, SE);
  else
    // Move the new node to the destination block, before its terminator.
    moveInstructionBefore(I, Dest->getTerminator()->getIterator(), *SafetyInfo,
                          MSSAU, SE);
 
  I.updateLocationAfterHoist();
 
  if (isa<LoadInst>(I))
    ++NumMovedLoads;
  else if (isa<CallInst>(I))
    ++NumMovedCalls;
  ++NumHoisted;
}
 
/// Only sink or hoist an instruction if it is not a trapping instruction,
/// or if the instruction is known not to trap when moved to the preheader.
/// or if it is a trapping instruction and is guaranteed to execute.
static bool isSafeToExecuteUnconditionally(
    Instruction &Inst, const DominatorTree *DT, const TargetLibraryInfo *TLI,
    const Loop *CurLoop, const LoopSafetyInfo *SafetyInfo,
    OptimizationRemarkEmitter *ORE, const Instruction *CtxI,
    AssumptionCache *AC, bool AllowSpeculation) {
  if (AllowSpeculation &&
      isSafeToSpeculativelyExecute(&Inst, CtxI, AC, DT, TLI))
    return true;
 
  bool GuaranteedToExecute = SafetyInfo->isGuaranteedToExecute(Inst, DT);
 
  if (!GuaranteedToExecute) {
    auto *LI = dyn_cast<LoadInst>(&Inst);
    if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand()))
      ORE->emit([&]() {
        return OptimizationRemarkMissed(
                   DEBUG_TYPE, "LoadWithLoopInvariantAddressCondExecuted", LI)
               << "failed to hoist load with loop-invariant address "
                  "because load is conditionally executed";
      });
  }
 
  return GuaranteedToExecute;
}
 
namespace {
class LoopPromoter : public LoadAndStorePromoter {
  Value *SomePtr; // Designated pointer to store to.
  SmallVectorImpl<BasicBlock *> &LoopExitBlocks;
  SmallVectorImpl<BasicBlock::iterator> &LoopInsertPts;
  SmallVectorImpl<MemoryAccess *> &MSSAInsertPts;
  PredIteratorCache &PredCache;
  MemorySSAUpdater &MSSAU;
  LoopInfo &LI;
  DebugLoc DL;
  Align Alignment;
  bool UnorderedAtomic;
  AAMDNodes AATags;
  ICFLoopSafetyInfo &SafetyInfo;
  bool CanInsertStoresInExitBlocks;
  ArrayRef<const Instruction *> Uses;
 
  // We're about to add a use of V in a loop exit block.  Insert an LCSSA phi
  // (if legal) if doing so would add an out-of-loop use to an instruction
  // defined in-loop.
  Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
    if (!LI.wouldBeOutOfLoopUseRequiringLCSSA(V, BB))
      return V;
 
    Instruction *I = cast<Instruction>(V);
    // We need to create an LCSSA PHI node for the incoming value and
    // store that.
    PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB),
                                  I->getName() + ".lcssa");
    PN->insertBefore(BB->begin());
    for (BasicBlock *Pred : PredCache.get(BB))
      PN->addIncoming(I, Pred);
    return PN;
  }
 
public:
  LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S,
               SmallVectorImpl<BasicBlock *> &LEB,
               SmallVectorImpl<BasicBlock::iterator> &LIP,
               SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC,
               MemorySSAUpdater &MSSAU, LoopInfo &li, DebugLoc dl,
               Align Alignment, bool UnorderedAtomic, const AAMDNodes &AATags,
               ICFLoopSafetyInfo &SafetyInfo, bool CanInsertStoresInExitBlocks)
      : LoadAndStorePromoter(Insts, S), SomePtr(SP), LoopExitBlocks(LEB),
        LoopInsertPts(LIP), MSSAInsertPts(MSSAIP), PredCache(PIC), MSSAU(MSSAU),
        LI(li), DL(std::move(dl)), Alignment(Alignment),
        UnorderedAtomic(UnorderedAtomic), AATags(AATags),
        SafetyInfo(SafetyInfo),
        CanInsertStoresInExitBlocks(CanInsertStoresInExitBlocks), Uses(Insts) {}
 
  void insertStoresInLoopExitBlocks() {
    // Insert stores after in the loop exit blocks.  Each exit block gets a
    // store of the live-out values that feed them.  Since we've already told
    // the SSA updater about the defs in the loop and the preheader
    // definition, it is all set and we can start using it.
    DIAssignID *NewID = nullptr;
    for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
      BasicBlock *ExitBlock = LoopExitBlocks[i];
      Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
      LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
      Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
      BasicBlock::iterator InsertPos = LoopInsertPts[i];
      StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
      if (UnorderedAtomic)
        NewSI->setOrdering(AtomicOrdering::Unordered);
      NewSI->setAlignment(Alignment);
      NewSI->setDebugLoc(DL);
      // Attach DIAssignID metadata to the new store, generating it on the
      // first loop iteration.
      if (i == 0) {
        // NewSI will have its DIAssignID set here if there are any stores in
        // Uses with a DIAssignID attachment. This merged ID will then be
        // attached to the other inserted stores (in the branch below).
        NewSI->mergeDIAssignID(Uses);
        NewID = cast_or_null<DIAssignID>(
            NewSI->getMetadata(LLVMContext::MD_DIAssignID));
      } else {
        // Attach the DIAssignID (or nullptr) merged from Uses in the branch
        // above.
        NewSI->setMetadata(LLVMContext::MD_DIAssignID, NewID);
      }
 
      if (AATags)
        NewSI->setAAMetadata(AATags);
 
      MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i];
      MemoryAccess *NewMemAcc;
      if (!MSSAInsertPoint) {
        NewMemAcc = MSSAU.createMemoryAccessInBB(
            NewSI, nullptr, NewSI->getParent(), MemorySSA::Beginning);
      } else {
        NewMemAcc =
            MSSAU.createMemoryAccessAfter(NewSI, nullptr, MSSAInsertPoint);
      }
      MSSAInsertPts[i] = NewMemAcc;
      MSSAU.insertDef(cast<MemoryDef>(NewMemAcc), true);
      // FIXME: true for safety, false may still be correct.
    }
  }
 
  void doExtraRewritesBeforeFinalDeletion() override {
    if (CanInsertStoresInExitBlocks)
      insertStoresInLoopExitBlocks();
  }
 
  void instructionDeleted(Instruction *I) const override {
    SafetyInfo.removeInstruction(I);
    MSSAU.removeMemoryAccess(I);
  }
 
  bool shouldDelete(Instruction *I) const override {
    if (isa<StoreInst>(I))
      return CanInsertStoresInExitBlocks;
    return true;
  }
};
 
bool isNotCapturedBeforeOrInLoop(const Value *V, const Loop *L,
                                 DominatorTree *DT) {
  // We can perform the captured-before check against any instruction in the
  // loop header, as the loop header is reachable from any instruction inside
  // the loop.
  // TODO: ReturnCaptures=true shouldn't be necessary here.
  return capturesNothing(PointerMayBeCapturedBefore(
      V, /*ReturnCaptures=*/true, L->getHeader()->getTerminator(), DT,
      /*IncludeI=*/false, CaptureComponents::Provenance));
}
 
/// Return true if we can prove that a caller cannot inspect the object if an
/// unwind occurs inside the loop.
bool isNotVisibleOnUnwindInLoop(const Value *Object, const Loop *L,
                                DominatorTree *DT) {
  bool RequiresNoCaptureBeforeUnwind;
  if (!isNotVisibleOnUnwind(Object, RequiresNoCaptureBeforeUnwind))
    return false;
 
  return !RequiresNoCaptureBeforeUnwind ||
         isNotCapturedBeforeOrInLoop(Object, L, DT);
}
 
bool isThreadLocalObject(const Value *Object, const Loop *L,
                         DominatorTree *DT) {
  // The object must be function-local to start with, and then not captured
  // before/in the loop.
  if (isIdentifiedFunctionLocal(Object) &&
      isNotCapturedBeforeOrInLoop(Object, L, DT))
    return true;
 
  // In a single-threaded environment, all objects are effectively thread-local.
  const Module *M = L->getHeader()->getModule();
  return M->getThreadModel() == ThreadModel::Single;
}
 
} // namespace
 
/// Try to promote memory values to scalars by sinking stores out of the
/// loop and moving loads to before the loop.  We do this by looping over
/// the stores in the loop, looking for stores to Must pointers which are
/// loop invariant.
///
bool llvm::promoteLoopAccessesToScalars(
    const SmallSetVector<Value *, 8> &PointerMustAliases,
    SmallVectorImpl<BasicBlock *> &ExitBlocks,
    SmallVectorImpl<BasicBlock::iterator> &InsertPts,
    SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC,
    const LICMDependencies &Deps, Loop *CurLoop, bool AllowSpeculation,
    bool HasReadsOutsideSet) {
  // Verify inputs.
  assert(CurLoop != nullptr &&
         "Unexpected Input to promoteLoopAccessesToScalars");
 
  LLVM_DEBUG({
    dbgs() << "Trying to promote set of must-aliased pointers:\n";
    for (Value *Ptr : PointerMustAliases)
      dbgs() << "  " << *Ptr << "\n";
  });
  ++NumPromotionCandidates;
 
  Value *SomePtr = *PointerMustAliases.begin();
  BasicBlock *Preheader = CurLoop->getLoopPreheader();
 
  // It is not safe to promote a load/store from the loop if the load/store is
  // conditional.  For example, turning:
  //
  //    for () { if (c) *P += 1; }
  //
  // into:
  //
  //    tmp = *P;  for () { if (c) tmp +=1; } *P = tmp;
  //
  // is not safe, because *P may only be valid to access if 'c' is true.
  //
  // The safety property divides into two parts:
  // p1) The memory may not be dereferenceable on entry to the loop.  In this
  //    case, we can't insert the required load in the preheader.
  // p2) The memory model does not allow us to insert a store along any dynamic
  //    path which did not originally have one.
  //
  // If at least one store is guaranteed to execute, both properties are
  // satisfied, and promotion is legal.
  //
  // This, however, is not a necessary condition. Even if no store/load is
  // guaranteed to execute, we can still establish these properties.
  // We can establish (p1) by proving that hoisting the load into the preheader
  // is safe (i.e. proving dereferenceability on all paths through the loop). We
  // can use any access within the alias set to prove dereferenceability,
  // since they're all must alias.
  //
  // There are two ways establish (p2):
  // a) Prove the location is thread-local. In this case the memory model
  // requirement does not apply, and stores are safe to insert.
  // b) Prove a store dominates every exit block. In this case, if an exit
  // blocks is reached, the original dynamic path would have taken us through
  // the store, so inserting a store into the exit block is safe. Note that this
  // is different from the store being guaranteed to execute. For instance,
  // if an exception is thrown on the first iteration of the loop, the original
  // store is never executed, but the exit blocks are not executed either.
 
  bool DereferenceableInPH = false;
  bool StoreIsGuaranteedToExecute = false;
  bool LoadIsGuaranteedToExecute = false;
  bool FoundLoadToPromote = false;
 
  // Goes from Unknown to either Safe or Unsafe, but can't switch between them.
  enum {
    StoreSafe,
    StoreUnsafe,
    StoreSafetyUnknown,
  } StoreSafety = StoreSafetyUnknown;
 
  SmallVector<Instruction *, 64> LoopUses;
 
  // We start with an alignment of one and try to find instructions that allow
  // us to prove better alignment.
  Align Alignment;
  // Keep track of which types of access we see
  bool SawUnorderedAtomic = false;
  bool SawNotAtomic = false;
  AAMDNodes AATags;
 
  const DataLayout &MDL = Preheader->getDataLayout();
 
  // If there are reads outside the promoted set, then promoting stores is
  // definitely not safe.
  if (HasReadsOutsideSet)
    StoreSafety = StoreUnsafe;
 
  if (StoreSafety == StoreSafetyUnknown && Deps.SafetyInfo.anyBlockMayThrow()) {
    // If a loop can throw, we have to insert a store along each unwind edge.
    // That said, we can't actually make the unwind edge explicit. Therefore,
    // we have to prove that the store is dead along the unwind edge.  We do
    // this by proving that the caller can't have a reference to the object
    // after return and thus can't possibly load from the object.
    Value *Object = getUnderlyingObject(SomePtr);
    if (!isNotVisibleOnUnwindInLoop(Object, CurLoop, &Deps.DT))
      StoreSafety = StoreUnsafe;
  }
 
  // Check that all accesses to pointers in the alias set use the same type.
  // We cannot (yet) promote a memory location that is loaded and stored in
  // different sizes.  While we are at it, collect alignment and AA info.
  Type *AccessTy = nullptr;
  for (Value *ASIV : PointerMustAliases) {
    for (Use &U : ASIV->uses()) {
      // Ignore instructions that are outside the loop.
      Instruction *UI = dyn_cast<Instruction>(U.getUser());
      if (!UI || !CurLoop->contains(UI))
        continue;
 
      // If there is an non-load/store instruction in the loop, we can't promote
      // it.
      if (LoadInst *Load = dyn_cast<LoadInst>(UI)) {
        if (!Load->isUnordered())
          return false;
 
        SawUnorderedAtomic |= Load->isAtomic();
        SawNotAtomic |= !Load->isAtomic();
        FoundLoadToPromote = true;
 
        Align InstAlignment = Load->getAlign();
 
        if (!LoadIsGuaranteedToExecute)
          LoadIsGuaranteedToExecute =
              Deps.SafetyInfo.isGuaranteedToExecute(*UI, &Deps.DT);
 
        // Note that proving a load safe to speculate requires proving
        // sufficient alignment at the target location.  Proving it guaranteed
        // to execute does as well.  Thus we can increase our guaranteed
        // alignment as well.
        if (!DereferenceableInPH || (InstAlignment > Alignment))
          if (isSafeToExecuteUnconditionally(
                  *Load, &Deps.DT, &Deps.TLI, CurLoop, &Deps.SafetyInfo,
                  Deps.ORE, Preheader->getTerminator(), &Deps.AC,
                  AllowSpeculation)) {
            DereferenceableInPH = true;
            Alignment = std::max(Alignment, InstAlignment);
          }
      } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) {
        // Stores *of* the pointer are not interesting, only stores *to* the
        // pointer.
        if (U.getOperandNo() != StoreInst::getPointerOperandIndex())
          continue;
        if (!Store->isUnordered())
          return false;
 
        SawUnorderedAtomic |= Store->isAtomic();
        SawNotAtomic |= !Store->isAtomic();
 
        // If the store is guaranteed to execute, both properties are satisfied.
        // We may want to check if a store is guaranteed to execute even if we
        // already know that promotion is safe, since it may have higher
        // alignment than any other guaranteed stores, in which case we can
        // raise the alignment on the promoted store.
        Align InstAlignment = Store->getAlign();
        bool GuaranteedToExecute =
            Deps.SafetyInfo.isGuaranteedToExecute(*UI, &Deps.DT);
        StoreIsGuaranteedToExecute |= GuaranteedToExecute;
        if (GuaranteedToExecute) {
          DereferenceableInPH = true;
          if (StoreSafety == StoreSafetyUnknown)
            StoreSafety = StoreSafe;
          Alignment = std::max(Alignment, InstAlignment);
        }
 
        // If a store dominates all exit blocks, it is safe to sink.
        // As explained above, if an exit block was executed, a dominating
        // store must have been executed at least once, so we are not
        // introducing stores on paths that did not have them.
        // Note that this only looks at explicit exit blocks. If we ever
        // start sinking stores into unwind edges (see above), this will break.
        if (StoreSafety == StoreSafetyUnknown &&
            llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) {
              return Deps.DT.dominates(Store->getParent(), Exit);
            }))
          StoreSafety = StoreSafe;
 
        // If the store is not guaranteed to execute, we may still get
        // deref info through it.
        if (!DereferenceableInPH) {
          DereferenceableInPH = isDereferenceableAndAlignedPointer(
              Store->getPointerOperand(), Store->getValueOperand()->getType(),
              Store->getAlign(),
              SimplifyQuery(MDL, &Deps.TLI, &Deps.DT, &Deps.AC,
                            Preheader->getTerminator()));
        }
      } else
        continue; // Not a load or store.
 
      if (!AccessTy)
        AccessTy = getLoadStoreType(UI);
      else if (AccessTy != getLoadStoreType(UI))
        return false;
 
      // Merge the AA tags.
      if (LoopUses.empty()) {
        // On the first load/store, just take its AA tags.
        AATags = UI->getAAMetadata();
      } else if (AATags) {
        AATags = AATags.merge(UI->getAAMetadata());
      }
 
      LoopUses.push_back(UI);
    }
  }
 
  // If we found both an unordered atomic instruction and a non-atomic memory
  // access, bail.  We can't blindly promote non-atomic to atomic since we
  // might not be able to lower the result.  We can't downgrade since that
  // would violate memory model.  Also, align 0 is an error for atomics.
  if (SawUnorderedAtomic && SawNotAtomic)
    return false;
 
  // If we're inserting an atomic load in the preheader, we must be able to
  // lower it.  We're only guaranteed to be able to lower naturally aligned
  // atomics.
  if (SawUnorderedAtomic && Alignment < MDL.getTypeStoreSize(AccessTy))
    return false;
 
  // If we couldn't prove we can hoist the load, bail.
  if (!DereferenceableInPH) {
    LLVM_DEBUG(dbgs() << "Not promoting: Not dereferenceable in preheader\n");
    return false;
  }
 
  // We know we can hoist the load, but don't have a guaranteed store.
  // Check whether the location is writable and thread-local. If it is, then we
  // can insert stores along paths which originally didn't have them without
  // violating the memory model.
  if (StoreSafety == StoreSafetyUnknown) {
    Value *Object = getUnderlyingObject(SomePtr);
    bool ExplicitlyDereferenceableOnly;
    // The dereferenceability query here is only required to satisfy the
    // writable contract, actual dereferenceability has already been proven
    // above. As such, we can ignore frees.
    if (isWritableObject(Object, ExplicitlyDereferenceableOnly) &&
        (!ExplicitlyDereferenceableOnly ||
         isDereferenceablePointer(SomePtr, AccessTy, MDL,
                                  /*IgnoreFree=*/true)) &&
        isThreadLocalObject(Object, CurLoop, &Deps.DT))
      StoreSafety = StoreSafe;
  }
 
  // If we've still failed to prove we can sink the store, hoist the load
  // only, if possible.
  if (StoreSafety != StoreSafe && !FoundLoadToPromote)
    // If we cannot hoist the load either, give up.
    return false;
 
  // Lets do the promotion!
  if (StoreSafety == StoreSafe) {
    LLVM_DEBUG(dbgs() << "LICM: Promoting load/store of the value: " << *SomePtr
                      << '\n');
    ++NumLoadStorePromoted;
  } else {
    LLVM_DEBUG(dbgs() << "LICM: Promoting load of the value: " << *SomePtr
                      << '\n');
    ++NumLoadPromoted;
  }
 
  Deps.ORE->emit([&]() {
    return OptimizationRemark(DEBUG_TYPE, "PromoteLoopAccessesToScalar",
                              LoopUses[0])
           << "Moving accesses to memory location out of the loop";
  });
 
  // Look at all the loop uses, and try to merge their locations.
  std::vector<DebugLoc> LoopUsesLocs;
  for (auto U : LoopUses)
    LoopUsesLocs.push_back(U->getDebugLoc());
  auto DL = DebugLoc::getMergedLocations(LoopUsesLocs);
 
  // We use the SSAUpdater interface to insert phi nodes as required.
  SmallVector<PHINode *, 16> NewPHIs;
  SSAUpdater SSA(&NewPHIs);
  LoopPromoter Promoter(SomePtr, LoopUses, SSA, ExitBlocks, InsertPts,
                        MSSAInsertPts, PIC, Deps.MSSAU, Deps.LI, DL, Alignment,
                        SawUnorderedAtomic,
                        StoreIsGuaranteedToExecute ? AATags : AAMDNodes(),
                        Deps.SafetyInfo, StoreSafety == StoreSafe);
 
  // Set up the preheader to have a definition of the value.  It is the live-out
  // value from the preheader that uses in the loop will use.
  LoadInst *PreheaderLoad = nullptr;
  if (FoundLoadToPromote || !StoreIsGuaranteedToExecute) {
    PreheaderLoad =
        new LoadInst(AccessTy, SomePtr, SomePtr->getName() + ".promoted",
                     Preheader->getTerminator()->getIterator());
    if (SawUnorderedAtomic)
      PreheaderLoad->setOrdering(AtomicOrdering::Unordered);
    PreheaderLoad->setAlignment(Alignment);
    PreheaderLoad->setDebugLoc(DebugLoc::getDropped());
    if (AATags && LoadIsGuaranteedToExecute)
      PreheaderLoad->setAAMetadata(AATags);
 
    MemoryAccess *PreheaderLoadMemoryAccess = Deps.MSSAU.createMemoryAccessInBB(
        PreheaderLoad, nullptr, PreheaderLoad->getParent(), MemorySSA::End);
    MemoryUse *NewMemUse = cast<MemoryUse>(PreheaderLoadMemoryAccess);
    Deps.MSSAU.insertUse(NewMemUse, /*RenameUses=*/true);
    SSA.AddAvailableValue(Preheader, PreheaderLoad);
  } else {
    SSA.AddAvailableValue(Preheader, PoisonValue::get(AccessTy));
  }
 
  if (VerifyMemorySSA)
    Deps.MSSAU.getMemorySSA()->verifyMemorySSA();
  // Rewrite all the loads in the loop and remember all the definitions from
  // stores in the loop.
  Promoter.run(LoopUses);
 
  if (VerifyMemorySSA)
    Deps.MSSAU.getMemorySSA()->verifyMemorySSA();
  // If the SSAUpdater didn't use the load in the preheader, just zap it now.
  if (PreheaderLoad && PreheaderLoad->use_empty())
    eraseInstruction(*PreheaderLoad, Deps);
 
  return true;
}
 
static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L,
                                function_ref<void(Instruction *)> Fn) {
  for (const BasicBlock *BB : L->blocks())
    if (const auto *Accesses = MSSA->getBlockAccesses(BB))
      for (const auto &Access : *Accesses)
        if (const auto *MUD = dyn_cast<MemoryUseOrDef>(&Access))
          Fn(MUD->getMemoryInst());
}
 
/// Returns whether \p I is a memory access that may be a candidate for
/// promotion out of the loop \p L.
static bool isPotentiallyPromotable(const Instruction *I, const Loop *L) {
  if (const auto *SI = dyn_cast<StoreInst>(I)) {
    const Value *PtrOp = SI->getPointerOperand();
    if (isStrongerThanMonotonic(SI->getOrdering()))
      return false;
    return !isa<ConstantData>(PtrOp) && L->isLoopInvariant(PtrOp);
  }
  if (const auto *LI = dyn_cast<LoadInst>(I)) {
    const Value *PtrOp = LI->getPointerOperand();
    if (isStrongerThanMonotonic(LI->getOrdering()))
      return false;
    return !isa<ConstantData>(PtrOp) && L->isLoopInvariant(PtrOp);
  }
  return false;
}
 
/// Returns whether \p N has any operand from the set \p Operands.
static bool
hasAnyMDOperandsFrom(const MDNode *N,
                     const SmallPtrSetImpl<const MDNode *> &Operands) {
  return N && llvm::any_of(N->operands(), [&](const MDOperand &Op) {
           return Operands.contains(cast<MDNode>(Op.get()));
         });
}
 
/// Returns the potentially promotable stores with AA tags that are valid along
/// all non-unwinding execution paths of the loop \p L, which allows for the AA
/// tags to be used when deciding promotions.
static SmallPtrSet<const StoreInst *, 8> collectStoresWithInvariantAATags(
    MemorySSA *MSSA, DominatorTree *DT,
    const SmallPtrSetImpl<const MDNode *> &LoopLocalAliasScopes, Loop *L) {
  SmallDenseMap<MemoryLocation, SmallVector<const StoreInst *, 1>, 4>
      StoresByLoc;
  foreachMemoryAccess(MSSA, L, [&](Instruction *I) {
    const auto *SI = dyn_cast<StoreInst>(I);
    if (SI && SI->getAAMetadata() && isPotentiallyPromotable(SI, L))
      StoresByLoc[MemoryLocation::get(SI)].push_back(SI);
  });
 
  // This only looks at explicit exiting blocks. If we ever start sinking
  // stores into unwind edges, this will break.
  SmallVector<BasicBlock *, 4> ExitingBlocks;
  L->getExitingBlocks(ExitingBlocks);
 
  SmallPtrSet<const StoreInst *, 8> StoresWithInvariantAATags;
  for (const auto &Pair : StoresByLoc) {
    const MemoryLocation &Loc = Pair.first;
    const SmallVector<const StoreInst *, 1> &Stores = Pair.second;
 
    // A scope declared inside the loop denotes a different scope on each
    // iteration, and thus should not be preserved.
    if (hasAnyMDOperandsFrom(Loc.AATags.Scope, LoopLocalAliasScopes) ||
        hasAnyMDOperandsFrom(Loc.AATags.NoAlias, LoopLocalAliasScopes))
      continue;
 
    // Without exiting blocks the loop is never left, and promotion has no
    // exit block to insert a store into either.
    if (llvm::all_of(ExitingBlocks, [&](BasicBlock *ExitingBB) {
          return llvm::any_of(Stores, [&](const StoreInst *SI) {
            return DT->dominates(SI->getParent(), ExitingBB);
          });
        }))
      StoresWithInvariantAATags.insert_range(Stores);
  }
  return StoresWithInvariantAATags;
}
 
// The bool indicates whether there might be reads outside the set, in which
// case only loads may be promoted.
static SmallVector<PointersAndHasReadsOutsideSet, 0> collectPromotionCandidates(
    MemorySSA *MSSA, AliasAnalysis *AA, DominatorTree *DT,
    ICFLoopSafetyInfo *SafetyInfo,
    const SmallPtrSetImpl<const MDNode *> &LoopLocalAliasScopes, Loop *L) {
  BatchAAResults BatchAA(*AA);
  AliasSetTracker AST(BatchAA);
 
  // Only conditionally executed stores need this, so compute it on demand to
  // keep the common case free.
  std::optional<SmallPtrSet<const StoreInst *, 8>> StoresWithInvariantAATags;
  auto HasInvariantAATags = [&](const StoreInst *SI) {
    if (!StoresWithInvariantAATags)
      StoresWithInvariantAATags =
          collectStoresWithInvariantAATags(MSSA, DT, LoopLocalAliasScopes, L);
    return StoresWithInvariantAATags->contains(SI);
  };
 
  // Populate AST with potentially promotable accesses.
  SmallPtrSet<Value *, 16> AttemptingPromotion;
  foreachMemoryAccess(MSSA, L, [&](Instruction *I) {
    if (isPotentiallyPromotable(I, L)) {
      AttemptingPromotion.insert(I);
      if (StoreInst *SI = dyn_cast<StoreInst>(I);
          SI && SI->getAAMetadata() &&
          !SafetyInfo->isGuaranteedToExecute(*SI, DT) &&
          !HasInvariantAATags(SI)) {
        // Promotion requires inserting a new store at the loop exits; we need
        // to prove that store doesn't alias anything, in addition to proving
        // aliasing for the stores we're removing. The new store is executed
        // unconditionally, so when we're proving aliasing for that store, we
        // can only rely on AA tags that likewise hold unconditionally.
        AST.addWithoutAATags(SI);
      } else {
        AST.add(I);
      }
    }
  });
 
  // We're only interested in must-alias sets that contain a mod.
  SmallVector<PointerIntPair<const AliasSet *, 1, bool>, 8> Sets;
  for (AliasSet &AS : AST)
    if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias())
      Sets.push_back({&AS, false});
 
  if (Sets.empty())
    return {}; // Nothing to promote...
 
  // Discard any sets for which there is an aliasing non-promotable access.
  foreachMemoryAccess(MSSA, L, [&](Instruction *I) {
    if (AttemptingPromotion.contains(I))
      return;
 
    llvm::erase_if(Sets, [&](PointerIntPair<const AliasSet *, 1, bool> &Pair) {
      ModRefInfo MR = Pair.getPointer()->aliasesUnknownInst(I, BatchAA);
      // Cannot promote if there are writes outside the set.
      if (isModSet(MR))
        return true;
      if (isRefSet(MR)) {
        // Remember reads outside the set.
        Pair.setInt(true);
        // If this is a mod-only set and there are reads outside the set,
        // we will not be able to promote, so bail out early.
        return !Pair.getPointer()->isRef();
      }
      return false;
    });
  });
 
  SmallVector<std::pair<SmallSetVector<Value *, 8>, bool>, 0> Result;
  for (auto [Set, HasReadsOutsideSet] : Sets) {
    SmallSetVector<Value *, 8> PointerMustAliases;
    for (const auto &MemLoc : *Set)
      PointerMustAliases.insert(const_cast<Value *>(MemLoc.Ptr));
    Result.emplace_back(std::move(PointerMustAliases), HasReadsOutsideSet);
  }
 
  return Result;
}
 
// For a given store instruction or writeonly call instruction, this function
// checks that there are no read or writes that conflict with the memory
// access in the instruction
static bool noConflictingReadWrites(Instruction *I, MemorySSA *MSSA,
                                    AAResults *AA, Loop *CurLoop,
                                    SinkAndHoistLICMFlags &Flags) {
  assert(isa<CallInst>(*I) || isa<StoreInst>(*I));
  // If there are more accesses than the Promotion cap, then give up as we're
  // not walking a list that long.
  if (Flags.tooManyMemoryAccesses())
    return false;
 
  auto *IMD = MSSA->getMemoryAccess(I);
  BatchAAResults BAA(*AA);
  auto *Source = getClobberingMemoryAccess(*MSSA, BAA, Flags, IMD);
  // Make sure there are no clobbers inside the loop.
  if (!MSSA->isLiveOnEntryDef(Source) && CurLoop->contains(Source->getBlock()))
    return false;
 
  // If there are interfering Uses don't move this store.
  // TODO: Cache set of Uses on the first walk in runOnLoop, update when
  // moving accesses. Can also extend to dominating uses.
  for (auto *BB : CurLoop->getBlocks()) {
    auto *Accesses = MSSA->getBlockAccesses(BB);
    if (!Accesses)
      continue;
    for (const auto &MA : *Accesses) {
      // Accesses are ordered. If we find one that I dominates we can stop.
      if (!Flags.getIsSink() && MSSA->dominates(IMD, &MA))
        break;
 
      if (const auto *MemUseOrDef = dyn_cast<MemoryUseOrDef>(&MA)) {
        // Skip unrelated accesses.
        if (isNoModRef(BAA.getModRefInfo(MemUseOrDef->getMemoryInst(), I)))
          continue;
 
        return false;
      }
    }
  }
  return true;
}
 
static bool pointerInvalidatedByLoop(const SinkAndHoistDependencies &Deps,
                                     MemoryUse *MU, Loop *CurLoop,
                                     Instruction &I,
                                     SinkAndHoistLICMFlags &Flags,
                                     bool InvariantGroup) {
  MemorySSA *MSSA = Deps.MSSAU.getMemorySSA();
  // For hoisting, use the walker to determine safety
  if (!Flags.getIsSink()) {
    // If hoisting an invariant group, we only need to check that there
    // is no store to the loaded pointer between the start of the loop,
    // and the load (since all values must be the same).
 
    // This can be checked in two conditions:
    // 1) if the memoryaccess is outside the loop
    // 2) the earliest access is at the loop header,
    // if the memory loaded is the phi node
 
    BatchAAResults BAA(MSSA->getAA());
    MemoryAccess *Source = getClobberingMemoryAccess(*MSSA, BAA, Flags, MU);
    return !MSSA->isLiveOnEntryDef(Source) &&
           CurLoop->contains(Source->getBlock()) &&
           !(InvariantGroup && Source->getBlock() == CurLoop->getHeader() && isa<MemoryPhi>(Source));
  }
 
  // For sinking, we'd need to check all Defs below this use. The getClobbering
  // call will look on the backedge of the loop, but will check aliasing with
  // the instructions on the previous iteration.
  // For example:
  // for (i ... )
  //   load a[i] ( Use (LoE)
  //   store a[i] ( 1 = Def (2), with 2 = Phi for the loop.
  //   i++;
  // The load sees no clobbering inside the loop, as the backedge alias check
  // does phi translation, and will check aliasing against store a[i-1].
  // However sinking the load outside the loop, below the store is incorrect.
 
  // For now, only sink if there are no Defs in the loop, and the existing ones
  // precede the use and are in the same block.
  // FIXME: Increase precision: Safe to sink if Use post dominates the Def;
  // needs PostDominatorTreeAnalysis.
  // FIXME: More precise: no Defs that alias this Use.
  if (Flags.tooManyMemoryAccesses())
    return true;
  for (auto *BB : CurLoop->getBlocks())
    if (pointerInvalidatedByBlock(*BB, *MSSA, *MU))
      return true;
  // When sinking, the source block may not be part of the loop so check it.
  if (!CurLoop->contains(&I))
    return pointerInvalidatedByBlock(*I.getParent(), *MSSA, *MU);
 
  return false;
}
 
bool pointerInvalidatedByBlock(BasicBlock &BB, MemorySSA &MSSA, MemoryUse &MU) {
  if (const auto *Accesses = MSSA.getBlockDefs(&BB))
    for (const auto &MA : *Accesses)
      if (const auto *MD = dyn_cast<MemoryDef>(&MA))
        if (MU.getBlock() != MD->getBlock() || !MSSA.locallyDominates(MD, &MU))
          return true;
  return false;
}
 
/// Try to simplify things like (A < INV_1 AND icmp A < INV_2) into (A <
/// min(INV_1, INV_2)), if INV_1 and INV_2 are both loop invariants and their
/// minimun can be computed outside of loop, and X is not a loop-invariant.
static bool hoistMinMax(Instruction &I, Loop &L, const LICMDependencies &Deps) {
  bool Inverse = false;
  using namespace PatternMatch;
  Value *Cond1, *Cond2;
  if (match(&I, m_LogicalOr(m_Value(Cond1), m_Value(Cond2)))) {
    Inverse = true;
  } else if (match(&I, m_LogicalAnd(m_Value(Cond1), m_Value(Cond2)))) {
    // Do nothing
  } else
    return false;
 
  auto MatchICmpAgainstInvariant = [&](Value *C, CmpPredicate &P, Value *&LHS,
                                       Value *&RHS) {
    if (!match(C, m_OneUse(m_ICmp(P, m_Value(LHS), m_Value(RHS)))))
      return false;
    if (!LHS->getType()->isIntegerTy())
      return false;
    if (!ICmpInst::isRelational(P))
      return false;
    if (L.isLoopInvariant(LHS)) {
      std::swap(LHS, RHS);
      P = ICmpInst::getSwappedPredicate(P);
    }
    if (L.isLoopInvariant(LHS) || !L.isLoopInvariant(RHS))
      return false;
    if (Inverse)
      P = ICmpInst::getInversePredicate(P);
    return true;
  };
  CmpPredicate P1, P2;
  Value *LHS1, *LHS2, *RHS1, *RHS2;
  if (!MatchICmpAgainstInvariant(Cond1, P1, LHS1, RHS1) ||
      !MatchICmpAgainstInvariant(Cond2, P2, LHS2, RHS2))
    return false;
  auto MatchingPred = CmpPredicate::getMatching(P1, P2);
  if (!MatchingPred || LHS1 != LHS2)
    return false;
 
  // Everything is fine, we can do the transform.
  bool UseMin = ICmpInst::isLT(*MatchingPred) || ICmpInst::isLE(*MatchingPred);
  assert(
      (UseMin || ICmpInst::isGT(*MatchingPred) ||
       ICmpInst::isGE(*MatchingPred)) &&
      "Relational predicate is either less (or equal) or greater (or equal)!");
  Intrinsic::ID id = ICmpInst::isSigned(*MatchingPred)
                         ? (UseMin ? Intrinsic::smin : Intrinsic::smax)
                         : (UseMin ? Intrinsic::umin : Intrinsic::umax);
  auto *Preheader = L.getLoopPreheader();
  assert(Preheader && "Loop is not in simplify form?");
  IRBuilder<> Builder(Preheader->getTerminator());
  // We are about to create a new guaranteed use for RHS2 which might not exist
  // before (if it was a non-taken input of logical and/or instruction). If it
  // was poison, we need to freeze it. Note that no new use for LHS and RHS1 are
  // introduced, so they don't need this.
  if (isa<SelectInst>(I))
    RHS2 = Builder.CreateFreeze(RHS2, RHS2->getName() + ".fr");
  Value *NewRHS = Builder.CreateBinaryIntrinsic(
      id, RHS1, RHS2, nullptr,
      StringRef("invariant.") +
          (ICmpInst::isSigned(*MatchingPred) ? "s" : "u") +
          (UseMin ? "min" : "max"));
  Builder.SetInsertPoint(&I);
  ICmpInst::Predicate P = *MatchingPred;
  if (Inverse)
    P = ICmpInst::getInversePredicate(P);
  Value *NewCond = Builder.CreateICmp(P, LHS1, NewRHS);
  NewCond->takeName(&I);
  I.replaceAllUsesWith(NewCond);
  eraseInstruction(I, Deps);
  Instruction &CondI1 = *cast<Instruction>(Cond1);
  Instruction &CondI2 = *cast<Instruction>(Cond2);
  salvageDebugInfo(CondI1);
  salvageDebugInfo(CondI2);
  eraseInstruction(CondI1, Deps);
  eraseInstruction(CondI2, Deps);
  return true;
}
 
/// Reassociate gep (gep ptr, idx1), idx2 to gep (gep ptr, idx2), idx1 if
/// this allows hoisting the inner GEP.
static bool hoistGEP(Instruction &I, Loop &L, const LICMDependencies &Deps) {
  auto *GEP = dyn_cast<GetElementPtrInst>(&I);
  if (!GEP)
    return false;
 
  // Do not try to hoist a constant GEP out of the loop via reassociation.
  // Constant GEPs can often be folded into addressing modes, and reassociating
  // them may inhibit CSE of a common base.
  if (GEP->hasAllConstantIndices())
    return false;
 
  auto *Src = dyn_cast<GetElementPtrInst>(GEP->getPointerOperand());
  if (!Src || !Src->hasOneUse() || !L.contains(Src))
    return false;
 
  Value *SrcPtr = Src->getPointerOperand();
  auto LoopInvariant = [&](Value *V) { return L.isLoopInvariant(V); };
  if (!L.isLoopInvariant(SrcPtr) || !all_of(GEP->indices(), LoopInvariant))
    return false;
 
  // This can only happen if !AllowSpeculation, otherwise this would already be
  // handled.
  // FIXME: Should we respect AllowSpeculation in these reassociation folds?
  // The flag exists to prevent metadata dropping, which is not relevant here.
  if (all_of(Src->indices(), LoopInvariant))
    return false;
 
  // The swapped GEPs are inbounds if both original GEPs are inbounds
  // and the sign of the offsets is the same. For simplicity, only
  // handle both offsets being non-negative.
  const DataLayout &DL = GEP->getDataLayout();
  auto NonNegative = [&](Value *V) {
    return isKnownNonNegative(V, SimplifyQuery(DL, &Deps.DT, &Deps.AC, GEP));
  };
  bool IsInBounds = Src->isInBounds() && GEP->isInBounds() &&
                    all_of(Src->indices(), NonNegative) &&
                    all_of(GEP->indices(), NonNegative);
 
  BasicBlock *Preheader = L.getLoopPreheader();
  IRBuilder<> Builder(Preheader->getTerminator());
  Value *NewSrc = Builder.CreateGEP(GEP->getSourceElementType(), SrcPtr,
                                    SmallVector<Value *>(GEP->indices()),
                                    "invariant.gep", IsInBounds);
  Builder.SetInsertPoint(GEP);
  Value *NewGEP = Builder.CreateGEP(Src->getSourceElementType(), NewSrc,
                                    SmallVector<Value *>(Src->indices()), "gep",
                                    IsInBounds);
  GEP->replaceAllUsesWith(NewGEP);
  eraseInstruction(*GEP, Deps);
  salvageDebugInfo(*Src);
  eraseInstruction(*Src, Deps);
  return true;
}
 
/// Try to turn things like "LV + C1 < C2" into "LV < C2 - C1". Here
/// C1 and C2 are loop invariants and LV is a loop-variant.
static bool hoistAdd(ICmpInst::Predicate Pred, Value *VariantLHS,
                     Value *InvariantRHS, ICmpInst &ICmp, Loop &L,
                     const LICMDependencies &Deps) {
  assert(!L.isLoopInvariant(VariantLHS) && "Precondition.");
  assert(L.isLoopInvariant(InvariantRHS) && "Precondition.");
 
  bool IsSigned = ICmpInst::isSigned(Pred);
 
  // Try to represent VariantLHS as sum of invariant and variant operands.
  using namespace PatternMatch;
  Value *VariantOp, *InvariantOp;
  if (IsSigned && !match(VariantLHS, m_NSWAddLike(m_Value(VariantOp),
                                                  m_Value(InvariantOp))))
    return false;
  if (!IsSigned && !match(VariantLHS, m_NUWAddLike(m_Value(VariantOp),
                                                   m_Value(InvariantOp))))
    return false;
 
  // LHS itself is a loop-variant, try to represent it in the form:
  // "VariantOp + InvariantOp". If it is possible, then we can reassociate.
  if (L.isLoopInvariant(VariantOp))
    std::swap(VariantOp, InvariantOp);
  if (L.isLoopInvariant(VariantOp) || !L.isLoopInvariant(InvariantOp))
    return false;
 
  // In order to turn "LV + C1 < C2" into "LV < C2 - C1", we need to be able to
  // freely move values from left side of inequality to right side (just as in
  // normal linear arithmetics). Overflows make things much more complicated, so
  // we want to avoid this.
  auto &DL = L.getHeader()->getDataLayout();
  SimplifyQuery SQ(DL, &Deps.DT, &Deps.AC, &ICmp);
  if (IsSigned && computeOverflowForSignedSub(InvariantRHS, InvariantOp, SQ) !=
                      llvm::OverflowResult::NeverOverflows)
    return false;
  if (!IsSigned &&
      computeOverflowForUnsignedSub(InvariantRHS, InvariantOp, SQ) !=
          llvm::OverflowResult::NeverOverflows)
    return false;
  auto *Preheader = L.getLoopPreheader();
  assert(Preheader && "Loop is not in simplify form?");
  IRBuilder<> Builder(Preheader->getTerminator());
  Value *NewCmpOp =
      Builder.CreateSub(InvariantRHS, InvariantOp, "invariant.op",
                        /*HasNUW*/ !IsSigned, /*HasNSW*/ IsSigned);
  ICmp.setPredicate(Pred);
  ICmp.setOperand(0, VariantOp);
  ICmp.setOperand(1, NewCmpOp);
  // The new LHS is a different value, so a samesign (or any other
  // poison-generating) flag asserted about the old operands may no longer hold.
  ICmp.dropPoisonGeneratingFlags();
 
  Instruction &DeadI = cast<Instruction>(*VariantLHS);
  salvageDebugInfo(DeadI);
  eraseInstruction(DeadI, Deps);
  return true;
}
 
/// Try to reassociate and hoist the following two patterns:
/// LV - C1 < C2 --> LV < C1 + C2,
/// C1 - LV < C2 --> LV > C1 - C2.
static bool hoistSub(ICmpInst::Predicate Pred, Value *VariantLHS,
                     Value *InvariantRHS, ICmpInst &ICmp, Loop &L,
                     const LICMDependencies &Deps) {
  assert(!L.isLoopInvariant(VariantLHS) && "Precondition.");
  assert(L.isLoopInvariant(InvariantRHS) && "Precondition.");
 
  bool IsSigned = ICmpInst::isSigned(Pred);
 
  // Try to represent VariantLHS as sum of invariant and variant operands.
  using namespace PatternMatch;
  Value *VariantOp, *InvariantOp;
  if (IsSigned &&
      !match(VariantLHS, m_NSWSub(m_Value(VariantOp), m_Value(InvariantOp))))
    return false;
  if (!IsSigned &&
      !match(VariantLHS, m_NUWSub(m_Value(VariantOp), m_Value(InvariantOp))))
    return false;
 
  bool VariantSubtracted = false;
  // LHS itself is a loop-variant, try to represent it in the form:
  // "VariantOp + InvariantOp". If it is possible, then we can reassociate. If
  // the variant operand goes with minus, we use a slightly different scheme.
  if (L.isLoopInvariant(VariantOp)) {
    std::swap(VariantOp, InvariantOp);
    VariantSubtracted = true;
    Pred = ICmpInst::getSwappedPredicate(Pred);
  }
  if (L.isLoopInvariant(VariantOp) || !L.isLoopInvariant(InvariantOp))
    return false;
 
  // In order to turn "LV - C1 < C2" into "LV < C2 + C1", we need to be able to
  // freely move values from left side of inequality to right side (just as in
  // normal linear arithmetics). Overflows make things much more complicated, so
  // we want to avoid this. Likewise, for "C1 - LV < C2" we need to prove that
  // "C1 - C2" does not overflow.
  auto &DL = L.getHeader()->getDataLayout();
  SimplifyQuery SQ(DL, &Deps.DT, &Deps.AC, &ICmp);
  if (VariantSubtracted && IsSigned) {
    // C1 - LV < C2 --> LV > C1 - C2
    if (computeOverflowForSignedSub(InvariantOp, InvariantRHS, SQ) !=
        llvm::OverflowResult::NeverOverflows)
      return false;
  } else if (VariantSubtracted && !IsSigned) {
    // C1 - LV < C2 --> LV > C1 - C2
    if (computeOverflowForUnsignedSub(InvariantOp, InvariantRHS, SQ) !=
        llvm::OverflowResult::NeverOverflows)
      return false;
  } else if (!VariantSubtracted && IsSigned) {
    // LV - C1 < C2 --> LV < C1 + C2
    if (computeOverflowForSignedAdd(InvariantOp, InvariantRHS, SQ) !=
        llvm::OverflowResult::NeverOverflows)
      return false;
  } else { // !VariantSubtracted && !IsSigned
    // LV - C1 < C2 --> LV < C1 + C2
    if (computeOverflowForUnsignedAdd(InvariantOp, InvariantRHS, SQ) !=
        llvm::OverflowResult::NeverOverflows)
      return false;
  }
  auto *Preheader = L.getLoopPreheader();
  assert(Preheader && "Loop is not in simplify form?");
  IRBuilder<> Builder(Preheader->getTerminator());
  Value *NewCmpOp =
      VariantSubtracted
          ? Builder.CreateSub(InvariantOp, InvariantRHS, "invariant.op",
                              /*HasNUW*/ !IsSigned, /*HasNSW*/ IsSigned)
          : Builder.CreateAdd(InvariantOp, InvariantRHS, "invariant.op",
                              /*HasNUW*/ !IsSigned, /*HasNSW*/ IsSigned);
  ICmp.setPredicate(Pred);
  ICmp.setOperand(0, VariantOp);
  ICmp.setOperand(1, NewCmpOp);
  // The new LHS is a different value, so a samesign (or any other
  // poison-generating) flag asserted about the old operands may no longer hold.
  ICmp.dropPoisonGeneratingFlags();
 
  Instruction &DeadI = cast<Instruction>(*VariantLHS);
  salvageDebugInfo(DeadI);
  eraseInstruction(DeadI, Deps);
  return true;
}
 
/// Reassociate and hoist add/sub expressions.
static bool hoistAddSub(Instruction &I, Loop &L, const LICMDependencies &Deps) {
  using namespace PatternMatch;
  CmpPredicate Pred;
  Value *LHS, *RHS;
  if (!match(&I, m_ICmp(Pred, m_Value(LHS), m_Value(RHS))))
    return false;
 
  // Put variant operand to LHS position.
  if (L.isLoopInvariant(LHS)) {
    std::swap(LHS, RHS);
    Pred = ICmpInst::getSwappedPredicate(Pred);
  }
  // We want to delete the initial operation after reassociation, so only do it
  // if it has no other uses.
  if (L.isLoopInvariant(LHS) || !L.isLoopInvariant(RHS) || !LHS->hasOneUse())
    return false;
 
  // TODO: We could go with smarter context, taking common dominator of all I's
  // users instead of I itself.
  if (hoistAdd(Pred, LHS, RHS, cast<ICmpInst>(I), L, Deps))
    return true;
 
  if (hoistSub(Pred, LHS, RHS, cast<ICmpInst>(I), L, Deps))
    return true;
 
  return false;
}
 
static bool isReassociableOp(Instruction *I, unsigned IntOpcode,
                             unsigned FPOpcode) {
  if (I->getOpcode() == IntOpcode)
    return true;
  if (I->getOpcode() == FPOpcode && I->hasAllowReassoc() &&
      I->hasNoSignedZeros())
    return true;
  return false;
}
 
/// Try to reassociate expressions like ((A1 * B1) + (A2 * B2) + ...) * C where
/// A1, A2, ... and C are loop invariants into expressions like
/// ((A1 * C * B1) + (A2 * C * B2) + ...) and hoist the (A1 * C), (A2 * C), ...
/// invariant expressions. This functions returns true only if any hoisting has
/// actually occurred.
static bool hoistMulAddAssociation(const ScalarOptions &Opts, Instruction &I,
                                   Loop &L, const LICMDependencies &Deps) {
  if (!isReassociableOp(&I, Instruction::Mul, Instruction::FMul))
    return false;
  Value *VariantOp = I.getOperand(0);
  Value *InvariantOp = I.getOperand(1);
  if (L.isLoopInvariant(VariantOp))
    std::swap(VariantOp, InvariantOp);
  if (L.isLoopInvariant(VariantOp) || !L.isLoopInvariant(InvariantOp))
    return false;
  Value *Factor = InvariantOp;
 
  // First, we need to make sure we should do the transformation.
  SmallVector<Use *> Changes;
  SmallVector<BinaryOperator *> Adds;
  SmallVector<BinaryOperator *> Worklist;
  if (BinaryOperator *VariantBinOp = dyn_cast<BinaryOperator>(VariantOp))
    Worklist.push_back(VariantBinOp);
  while (!Worklist.empty()) {
    BinaryOperator *BO = Worklist.pop_back_val();
    if (!BO->hasOneUse())
      return false;
    if (isReassociableOp(BO, Instruction::Add, Instruction::FAdd) &&
        isa<BinaryOperator>(BO->getOperand(0)) &&
        isa<BinaryOperator>(BO->getOperand(1))) {
      Worklist.push_back(cast<BinaryOperator>(BO->getOperand(0)));
      Worklist.push_back(cast<BinaryOperator>(BO->getOperand(1)));
      Adds.push_back(BO);
      continue;
    }
    if (!isReassociableOp(BO, Instruction::Mul, Instruction::FMul) ||
        L.isLoopInvariant(BO))
      return false;
    Use &U0 = BO->getOperandUse(0);
    Use &U1 = BO->getOperandUse(1);
    if (L.isLoopInvariant(U0))
      Changes.push_back(&U0);
    else if (L.isLoopInvariant(U1))
      Changes.push_back(&U1);
    else
      return false;
    unsigned Limit = I.getType()->isIntOrIntVectorTy()
                         ? Opts.licm_max_num_int_reassociations
                         : Opts.licm_max_num_fp_reassociations;
    if (Changes.size() > Limit)
      return false;
  }
  if (Changes.empty())
    return false;
 
  // Drop the poison flags for any adds we looked through.
  if (I.getType()->isIntOrIntVectorTy()) {
    for (auto *Add : Adds)
      Add->dropPoisonGeneratingFlags();
  }
 
  // We know we should do it so let's do the transformation.
  auto *Preheader = L.getLoopPreheader();
  assert(Preheader && "Loop is not in simplify form?");
  IRBuilder<> Builder(Preheader->getTerminator());
  for (auto *U : Changes) {
    assert(L.isLoopInvariant(U->get()));
    auto *Ins = cast<BinaryOperator>(U->getUser());
    Value *Mul;
    if (I.getType()->isIntOrIntVectorTy()) {
      Mul = Builder.CreateMul(U->get(), Factor, "factor.op.mul");
      // Drop the poison flags on the original multiply.
      Ins->dropPoisonGeneratingFlags();
    } else
      Mul = Builder.CreateFMulFMF(U->get(), Factor, Ins, "factor.op.fmul");
 
    // Rewrite the reassociable instruction.
    unsigned OpIdx = U->getOperandNo();
    auto *LHS = OpIdx == 0 ? Mul : Ins->getOperand(0);
    auto *RHS = OpIdx == 1 ? Mul : Ins->getOperand(1);
    auto *NewBO =
        BinaryOperator::Create(Ins->getOpcode(), LHS, RHS,
                               Ins->getName() + ".reass", Ins->getIterator());
    NewBO->setDebugLoc(DebugLoc::getDropped());
    NewBO->copyIRFlags(Ins);
    if (VariantOp == Ins)
      VariantOp = NewBO;
    Ins->replaceAllUsesWith(NewBO);
    eraseInstruction(*Ins, Deps);
  }
 
  I.replaceAllUsesWith(VariantOp);
  eraseInstruction(I, Deps);
  return true;
}
 
/// Reassociate associative binary expressions of the form
///
/// 1. "(LV op C1) op C2" ==> "LV op (C1 op C2)"
/// 2. "(C1 op LV) op C2" ==> "LV op (C1 op C2)"
/// 3. "C2 op (C1 op LV)" ==> "LV op (C1 op C2)"
/// 4. "C2 op (LV op C1)" ==> "LV op (C1 op C2)"
///
/// where op is an associative BinOp, LV is a loop variant, and C1 and C2 are
/// loop invariants that we want to hoist, noting that associativity implies
/// commutativity.
static bool hoistBOAssociation(Instruction &I, Loop &L,
                               const LICMDependencies &Deps) {
  auto *BO = dyn_cast<BinaryOperator>(&I);
  if (!BO || !BO->isAssociative())
    return false;
 
  Instruction::BinaryOps Opcode = BO->getOpcode();
  bool LVInRHS = L.isLoopInvariant(BO->getOperand(0));
  auto *BO0 = dyn_cast<BinaryOperator>(BO->getOperand(LVInRHS));
  if (!BO0 || BO0->getOpcode() != Opcode || !BO0->isAssociative() ||
      BO0->hasNUsesOrMore(BO0->getType()->isIntegerTy() ? 2 : 3))
    return false;
 
  Value *LV = BO0->getOperand(0);
  Value *C1 = BO0->getOperand(1);
  Value *C2 = BO->getOperand(!LVInRHS);
 
  assert(BO->isCommutative() && BO0->isCommutative() &&
         "Associativity implies commutativity");
  if (L.isLoopInvariant(LV) && !L.isLoopInvariant(C1))
    std::swap(LV, C1);
  if (L.isLoopInvariant(LV) || !L.isLoopInvariant(C1) || !L.isLoopInvariant(C2))
    return false;
 
  auto *Preheader = L.getLoopPreheader();
  assert(Preheader && "Loop is not in simplify form?");
 
  IRBuilder<> Builder(Preheader->getTerminator());
  auto *Inv = Builder.CreateBinOp(Opcode, C1, C2, "invariant.op");
 
  auto *NewBO = BinaryOperator::Create(
      Opcode, LV, Inv, BO->getName() + ".reass", BO->getIterator());
  NewBO->setDebugLoc(DebugLoc::getDropped());
 
  if (Opcode == Instruction::FAdd || Opcode == Instruction::FMul) {
    // Intersect FMF flags for FADD and FMUL.
    FastMathFlags Intersect = BO->getFastMathFlags() & BO0->getFastMathFlags();
    if (auto *I = dyn_cast<Instruction>(Inv))
      I->setFastMathFlags(Intersect);
    NewBO->setFastMathFlags(Intersect);
  } else {
    OverflowTracking Flags;
    Flags.AllKnownNonNegative = false;
    Flags.AllKnownNonZero = false;
    Flags.mergeFlags(*BO);
    Flags.mergeFlags(*BO0);
    // If `Inv` was not constant-folded, a new Instruction has been created.
    auto *InvI = dyn_cast<Instruction>(Inv);
    if (InvI)
      Flags.applyFlags(*InvI);
    Flags.applyFlags(*NewBO);
 
    // The original nsw flags guarantee that LV + C1 + C2 is representable.
    // If C1 + C2 is representable too, both reassociated adds keep nsw.
    SimplifyQuery SQ(L.getHeader()->getDataLayout(), &Deps.DT, &Deps.AC,
                     Preheader->getTerminator());
    if (Opcode == Instruction::Add && Flags.HasNSW && !Flags.HasNUW &&
        computeOverflowForSignedAdd(C1, C2, SQ) ==
            OverflowResult::NeverOverflows) {
      if (InvI)
        InvI->setHasNoSignedWrap();
      NewBO->setHasNoSignedWrap();
    }
  }
 
  BO->replaceAllUsesWith(NewBO);
  eraseInstruction(*BO, Deps);
 
  // (LV op C1) might not be erased if it has more uses than the one we just
  // replaced.
  if (BO0->use_empty()) {
    salvageDebugInfo(*BO0);
    eraseInstruction(*BO0, Deps);
  }
 
  return true;
}
 
/// Reassociate add/sub expressions of the form:
///
/// 1. "(LV + C1) - C2" ==> "LV + (C1 - C2)"
/// 2. "(LV - C1) - C2" ==> "LV - (C1 + C2)"
/// 3. "(LV - C1) + C2" ==> "LV + (C2 - C1)"
///
/// where LV is a loop variant, and C1 and C2 are loop invariants.
/// Sub is not associative, but these algebraic identities allow hoisting
/// invariant computations out of the loop.
static bool hoistSubAddAssociation(Instruction &I, Loop &L,
                                   const LICMDependencies &Deps) {
  using namespace PatternMatch;
 
  Instruction *BO;
  Value *LV, *C1, *C2;
  Instruction::BinaryOps InvOp, ResultOp;
 
  // Try to match one of three reassociation patterns involving sub.
  //
  //   1. (LV + C1) - C2 ==> LV + (C1 - C2)
  //   2. (LV - C1) - C2 ==> LV - (C1 + C2)
  //   3. (LV - C1) + C2 ==> LV + (C2 - C1)
  //                            ^     ^
  //                             \     \___ InvOp
  //                              \
  //                               \____ ResultOp
  //
  if (match(&I,
            m_Sub(m_OneUse(m_Instruction(BO, m_Add(m_Value(LV), m_Value(C1)))),
                  m_Value(C2)))) {
    // Case 1.
    //
    // Depending on which of the addition is invariant, we might need to swap
    // the arguments
    if (L.isLoopInvariant(LV) && !L.isLoopInvariant(C1))
      std::swap(LV, C1);
    InvOp = Instruction::Sub;
    ResultOp = Instruction::Add;
  } else if (match(&I, m_Sub(m_OneUse(m_Instruction(
                                 BO, m_Sub(m_Value(LV), m_Value(C1)))),
                             m_Value(C2)))) {
    // Case 2.
    InvOp = Instruction::Add;
    ResultOp = Instruction::Sub;
  } else if (match(&I, m_c_Add(m_OneUse(m_Instruction(
                                   BO, m_Sub(m_Value(LV), m_Value(C1)))),
                               m_Value(C2)))) {
    // Case 3.
    //
    // We use (C2 - C1) as the invariant as opposed to case 1, but instead of
    // adding a special case in invariant creation, we can just swap the
    // operands here.
    std::swap(C1, C2);
    InvOp = Instruction::Sub;
    ResultOp = Instruction::Add;
  } else {
    return false;
  }
 
  if (L.isLoopInvariant(LV) || !L.isLoopInvariant(C1) || !L.isLoopInvariant(C2))
    return false;
 
  auto *Preheader = L.getLoopPreheader();
  assert(Preheader && "Loop is not in simplify form?");
 
  IRBuilder<> Builder(Preheader->getTerminator());
  auto *Inv = Builder.CreateBinOp(InvOp, C1, C2, "invariant.op");
 
  auto *NewBO = BinaryOperator::Create(ResultOp, LV, Inv,
                                       I.getName() + ".reass", I.getIterator());
  NewBO->setDebugLoc(DebugLoc::getDropped());
 
  // No overflow flags are set on the new instructions -- reassociation
  // involving sub does not preserve nsw/nuw in general.
 
  I.replaceAllUsesWith(NewBO);
  eraseInstruction(I, Deps);
 
  salvageDebugInfo(*BO);
  eraseInstruction(*BO, Deps);
 
  return true;
}
 
static bool hoistArithmetics(const ScalarOptions &Opts, Instruction &I, Loop &L,
                             const LICMDependencies &Deps) {
  // Optimize complex patterns, such as (x < INV1 && x < INV2), turning them
  // into (x < min(INV1, INV2)), and hoisting the invariant part of this
  // expression out of the loop.
  if (hoistMinMax(I, L, Deps)) {
    ++NumHoisted;
    ++NumMinMaxHoisted;
    return true;
  }
 
  // Try to hoist GEPs by reassociation.
  if (hoistGEP(I, L, Deps)) {
    ++NumHoisted;
    ++NumGEPsHoisted;
    return true;
  }
 
  // Try to hoist add/sub's by reassociation.
  if (hoistAddSub(I, L, Deps)) {
    ++NumHoisted;
    ++NumAddSubHoisted;
    return true;
  }
 
  bool IsInt = I.getType()->isIntOrIntVectorTy();
  if (hoistMulAddAssociation(Opts, I, L, Deps)) {
    ++NumHoisted;
    if (IsInt)
      ++NumIntAssociationsHoisted;
    else
      ++NumFPAssociationsHoisted;
    return true;
  }
 
  if (hoistBOAssociation(I, L, Deps)) {
    ++NumHoisted;
    ++NumBOAssociationsHoisted;
    return true;
  }
 
  if (hoistSubAddAssociation(I, L, Deps)) {
    ++NumHoisted;
    ++NumBOAssociationsHoisted;
    return true;
  }
 
  return false;
}
 
/// Little predicate that returns true if the specified basic block is in
/// a subloop of the current one, not the current one itself.
///
static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
  assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
  return LI->getLoopFor(BB) != CurLoop;
}