Loads.cpp

Go to the documentation of this file.

//===- Loads.cpp - Local load analysis ------------------------------------===//
//
// 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 file defines simple local analyses for load instructions.
//
//===----------------------------------------------------------------------===//
 
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AssumeBundleQueries.h"
#include "llvm/Analysis/LoopAccessAnalysis.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/MemoryBuiltins.h"
#include "llvm/Analysis/MemoryLocation.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/GetElementPtrTypeIterator.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Operator.h"
 
using namespace llvm;
 
static bool isAligned(const Value *Base, Align Alignment,
                      const DataLayout &DL) {
  return Base->getPointerAlignment(DL) >= Alignment;
}
 
static bool isDereferenceableAndAlignedPointerViaAssumption(
    const Value *Ptr, Align Alignment, const SimplifyQuery &SQ, bool IgnoreFree,
    function_ref<bool(const RetainedKnowledge &RK)> CheckSize) {
  if (!SQ.CxtI)
    return false;
  // Look through assumes to see if both dereferenceability and alignment can
  // be proven by an assume if needed.
  bool PtrCanBeFreed = Ptr->canBeFreed() && !IgnoreFree;
  bool IsAligned = Ptr->getPointerAlignment(SQ.DL) >= Alignment;
  bool IsDerefable = false;
  return getKnowledgeForValue(
      Ptr, {Attribute::Dereferenceable, Attribute::Alignment}, *SQ.AC,
      [&](RetainedKnowledge RK, Instruction *Assume, auto) {
        if (!isValidAssumeForContext(Assume, SQ.CxtI, SQ.DT))
          return false;
        if (RK.AttrKind == Attribute::Alignment) {
          IsAligned |= RK.ArgValue >= Alignment.value();
        } else {
          assert(RK.AttrKind == Attribute::Dereferenceable);
          // Dereferenceable information from assumptions is only valid if the
          // value cannot be freed between the assumption and use.
          if (!IsDerefable &&
              (!PtrCanBeFreed || willNotFreeBetween(Assume, SQ.CxtI)) &&
              CheckSize(RK))
            IsDerefable = true;
        }
        // Stop looking if we have proven both necessary facts.
        return IsAligned && IsDerefable;
      });
}
 
/// Test if V is always a pointer to allocated and suitably aligned memory for
/// a simple load or store.
static bool isDereferenceableAndAlignedPointer(
    const Value *V, Align Alignment, const APInt &Size, const SimplifyQuery &SQ,
    bool IgnoreFree, SmallPtrSetImpl<const Value *> &Visited,
    unsigned MaxDepth) {
  assert(V->getType()->isPointerTy() && "Base must be pointer");
 
  // Recursion limit.
  if (MaxDepth-- == 0)
    return false;
 
  // Already visited?  Bail out, we've likely hit unreachable code.
  if (!Visited.insert(V).second)
    return false;
 
  // Note that it is not safe to speculate into a malloc'd region because
  // malloc may return null.
 
  // For GEPs, determine if the indexing lands within the allocated object.
  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
    const Value *Base = GEP->getPointerOperand();
 
    APInt Offset(SQ.DL.getIndexTypeSizeInBits(GEP->getType()), 0);
    if (!GEP->accumulateConstantOffset(SQ.DL, Offset) || Offset.isNegative() ||
        !Offset.urem(APInt(Offset.getBitWidth(), Alignment.value()))
             .isMinValue())
      return false;
 
    // If the base pointer is dereferenceable for Offset+Size bytes, then the
    // GEP (== Base + Offset) is dereferenceable for Size bytes.  If the base
    // pointer is aligned to Align bytes, and the Offset is divisible by Align
    // then the GEP (== Base + Offset == k_0 * Align + k_1 * Align) is also
    // aligned to Align bytes.
 
    // Offset and Size may have different bit widths if we have visited an
    // addrspacecast, so we can't do arithmetic directly on the APInt values.
    return isDereferenceableAndAlignedPointer(
        Base, Alignment, Offset + Size.sextOrTrunc(Offset.getBitWidth()), SQ,
        IgnoreFree, Visited, MaxDepth);
  }
 
  // bitcast instructions are no-ops as far as dereferenceability is concerned.
  if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
    if (BC->getSrcTy()->isPointerTy())
      return isDereferenceableAndAlignedPointer(BC->getOperand(0), Alignment,
                                                Size, SQ, IgnoreFree, Visited,
                                                MaxDepth);
  }
 
  // Recurse into both hands of select.
  if (const SelectInst *Sel = dyn_cast<SelectInst>(V)) {
    return isDereferenceableAndAlignedPointer(Sel->getTrueValue(), Alignment,
                                              Size, SQ, IgnoreFree, Visited,
                                              MaxDepth) &&
           isDereferenceableAndAlignedPointer(Sel->getFalseValue(), Alignment,
                                              Size, SQ, IgnoreFree, Visited,
                                              MaxDepth);
  }
 
  auto IsKnownDeref = [&]() {
    bool CheckForNonNull, CheckForFreed;
    if (!Size.ule(V->getPointerDereferenceableBytes(SQ.DL, CheckForNonNull,
                                                    &CheckForFreed)))
      return false;
    if (CheckForNonNull && !isKnownNonZero(V, SQ))
      return false;
 
    auto *I = dyn_cast<Instruction>(V);
    if (CheckForFreed && !IgnoreFree) {
      const Instruction *DefI;
      if (I) {
        // We don't want to consider frees by the instruction producing the
        // pointer, so skip it if we can.
        if (auto *II = dyn_cast<InvokeInst>(V)) {
          DefI = &II->getNormalDest()->front();
        } else if (!I->isTerminator()) {
          DefI = I->getNextNode();
        } else {
          DefI = I;
        }
      } else {
        // For arguments, check frees from the start of the entry block.
        DefI = &cast<Argument>(V)->getParent()->getEntryBlock().front();
      }
 
      if (!SQ.CxtI || !willNotFreeBetween(DefI, SQ.CxtI))
        return false;
    }
 
    // When using something like !dereferenceable on a load, the
    // dereferenceability may only be valid on a specific control-flow path.
    // If the instruction doesn't dominate the context instruction, we're
    // asking about dereferenceability under the assumption that the
    // instruction has been speculated to the point of the context instruction,
    // in which case we don't know if the dereferenceability info still holds.
    // We don't bother handling allocas here, as they aren't speculatable
    // anyway.
    if (I && !isa<AllocaInst>(I))
      return SQ.CxtI && isValidAssumeForContext(I, SQ.CxtI, SQ.DT);
    return true;
  };
  if (IsKnownDeref()) {
    // As we recursed through GEPs to get here, we've incrementally checked
    // that each step advanced by a multiple of the alignment. If our base is
    // properly aligned, then the original offset accessed must also be.
    return isAligned(V, Alignment, SQ.DL);
  }
 
  /// TODO refactor this function to be able to search independently for
  /// Dereferencability and Alignment requirements.
 
 
  if (const auto *Call = dyn_cast<CallBase>(V)) {
    if (auto *RP = getArgumentAliasingToReturnedPointer(
            Call, /*MustPreserveOffset=*/true))
      return isDereferenceableAndAlignedPointer(RP, Alignment, Size, SQ,
                                                IgnoreFree, Visited, MaxDepth);
 
    // If we have a call we can't recurse through, check to see if this is an
    // allocation function for which we can establish an minimum object size.
    // Such a minimum object size is analogous to a deref_or_null attribute in
    // that we still need to prove the result non-null at point of use.
    // NOTE: We can only use the object size as a base fact as we a) need to
    // prove alignment too, and b) don't want the compile time impact of a
    // separate recursive walk.
    ObjectSizeOpts Opts;
    // TODO: It may be okay to round to align, but that would imply that
    // accessing slightly out of bounds was legal, and we're currently
    // inconsistent about that.  For the moment, be conservative.
    Opts.RoundToAlign = false;
    Opts.NullIsUnknownSize = true;
    uint64_t ObjSize;
    if (getObjectSize(V, ObjSize, SQ.DL, SQ.TLI, Opts)) {
      APInt KnownDerefBytes(Size.getBitWidth(), ObjSize);
      if (KnownDerefBytes.getBoolValue() && KnownDerefBytes.uge(Size) &&
          isKnownNonZero(V, SQ) && !V->canBeFreed()) {
        // As we recursed through GEPs to get here, we've incrementally
        // checked that each step advanced by a multiple of the alignment. If
        // our base is properly aligned, then the original offset accessed
        // must also be.
        return isAligned(V, Alignment, SQ.DL);
      }
    }
  }
 
  // For gc.relocate, look through relocations
  if (const GCRelocateInst *RelocateInst = dyn_cast<GCRelocateInst>(V))
    return isDereferenceableAndAlignedPointer(RelocateInst->getDerivedPtr(),
                                              Alignment, Size, SQ, IgnoreFree,
                                              Visited, MaxDepth);
 
  if (const AddrSpaceCastOperator *ASC = dyn_cast<AddrSpaceCastOperator>(V))
    return isDereferenceableAndAlignedPointer(
        ASC->getOperand(0), Alignment, Size, SQ, IgnoreFree, Visited, MaxDepth);
 
  return SQ.AC &&
         isDereferenceableAndAlignedPointerViaAssumption(
             V, Alignment, SQ, IgnoreFree, [Size](const RetainedKnowledge &RK) {
               return RK.ArgValue >= Size.getZExtValue();
             });
}
 
bool llvm::isDereferenceableAndAlignedPointer(const Value *V, Align Alignment,
                                              const APInt &Size,
                                              const SimplifyQuery &SQ,
                                              bool IgnoreFree) {
  // Note: At the moment, Size can be zero.  This ends up being interpreted as
  // a query of whether [Base, V] is dereferenceable and V is aligned (since
  // that's what the implementation happened to do).  It's unclear if this is
  // the desired semantic, but at least SelectionDAG does exercise this case.
 
  SmallPtrSet<const Value *, 32> Visited;
  return ::isDereferenceableAndAlignedPointer(V, Alignment, Size, SQ,
                                              IgnoreFree, Visited,
                                              /*MaxDepth=*/16);
}
 
bool llvm::isDereferenceableAndAlignedPointer(const Value *V, Type *Ty,
                                              Align Alignment,
                                              const SimplifyQuery &SQ,
                                              bool IgnoreFree) {
  // For unsized types or scalable vectors we don't know exactly how many bytes
  // are dereferenced, so bail out.
  if (!Ty->isSized() || Ty->isScalableTy())
    return false;
 
  // When dereferenceability information is provided by a dereferenceable
  // attribute, we know exactly how many bytes are dereferenceable. If we can
  // determine the exact offset to the attributed variable, we can use that
  // information here.
 
  APInt AccessSize(SQ.DL.getPointerTypeSizeInBits(V->getType()),
                   SQ.DL.getTypeStoreSize(Ty));
  return isDereferenceableAndAlignedPointer(V, Alignment, AccessSize, SQ,
                                            IgnoreFree);
}
 
bool llvm::isDereferenceablePointer(const Value *V, Type *Ty,
                                    const SimplifyQuery &SQ, bool IgnoreFree) {
  return isDereferenceableAndAlignedPointer(V, Ty, Align(1), SQ, IgnoreFree);
}
 
bool llvm::isDereferenceablePointer(const Value *V, const APInt &Size,
                                    const SimplifyQuery &Q, bool IgnoreFree) {
  return isDereferenceableAndAlignedPointer(V, Align(1), Size, Q, IgnoreFree);
}
 
/// Test if A and B will obviously have the same value.
///
/// This includes recognizing that %t0 and %t1 will have the same
/// value in code like this:
/// \code
///   %t0 = getelementptr \@a, 0, 3
///   store i32 0, i32* %t0
///   %t1 = getelementptr \@a, 0, 3
///   %t2 = load i32* %t1
/// \endcode
///
static bool AreEquivalentAddressValues(const Value *A, const Value *B) {
  // Test if the values are trivially equivalent.
  if (A == B)
    return true;
 
  // Test if the values come from identical arithmetic instructions.
  // Use isIdenticalToWhenDefined instead of isIdenticalTo because
  // this function is only used when one address use dominates the
  // other, which means that they'll always either have the same
  // value or one of them will have an undefined value.
  if (isa<CastInst>(A) || isa<PHINode>(A) || isa<GetElementPtrInst>(A))
    if (const Instruction *BI = dyn_cast<Instruction>(B))
      if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
        return true;
 
  // Otherwise they may not be equivalent.
  return false;
}
 
bool llvm::isDereferenceableAndAlignedInLoop(
    LoadInst *LI, Loop *L, ScalarEvolution &SE, DominatorTree &DT,
    AssumptionCache *AC, SmallVectorImpl<const SCEVPredicate *> *Predicates) {
  auto &DL = LI->getDataLayout();
  Value *Ptr = LI->getPointerOperand();
  const SCEV *PtrSCEV = SE.getSCEV(Ptr);
  APInt EltSize(DL.getIndexTypeSizeInBits(Ptr->getType()),
                DL.getTypeStoreSize(LI->getType()).getFixedValue());
 
  // If given a uniform (i.e. non-varying) address, see if we can prove the
  // access is safe within the loop w/o needing predication.
  if (L->isLoopInvariant(Ptr))
    return isDereferenceableAndAlignedPointer(
        Ptr, LI->getAlign(), EltSize,
        SimplifyQuery(DL, &DT, AC, &*L->getHeader()->getFirstNonPHIIt()));
 
  const SCEV *EltSizeSCEV = SE.getConstant(EltSize);
  return isDereferenceableAndAlignedInLoop(PtrSCEV, LI->getAlign(), EltSizeSCEV,
                                           L, SE, DT, AC, Predicates);
}
 
bool llvm::isDereferenceableAndAlignedInLoop(
    const SCEV *PtrSCEV, Align Alignment, const SCEV *EltSizeSCEV, Loop *L,
    ScalarEvolution &SE, DominatorTree &DT, AssumptionCache *AC,
    SmallVectorImpl<const SCEVPredicate *> *Predicates) {
  auto *AddRec = dyn_cast<SCEVAddRecExpr>(PtrSCEV);
 
  // Check to see if we have a repeating access pattern and it's possible
  // to prove all accesses are well aligned.
  if (!AddRec || AddRec->getLoop() != L || !AddRec->isAffine())
    return false;
 
  auto *Step = dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(SE));
  if (!Step)
    return false;
 
  const APInt &EltSize = cast<SCEVConstant>(EltSizeSCEV)->getAPInt();
  // For the moment, restrict ourselves to the case where the access size is a
  // multiple of the requested alignment and the base is aligned.
  // TODO: generalize if a case found which warrants
  if (EltSize.urem(Alignment.value()) != 0)
    return false;
 
  // TODO: Handle overlapping accesses.
  if (EltSize.ugt(Step->getAPInt().abs()))
    return false;
 
  const SCEV *MaxBECount =
      Predicates ? SE.getPredicatedSymbolicMaxBackedgeTakenCount(L, *Predicates)
                 : SE.getSymbolicMaxBackedgeTakenCount(L);
  const SCEV *BECount = Predicates
                            ? SE.getPredicatedBackedgeTakenCount(L, *Predicates)
                            : SE.getBackedgeTakenCount(L);
  if (isa<SCEVCouldNotCompute>(MaxBECount))
    return false;
  std::optional<ScalarEvolution::LoopGuards> LoopGuards;
 
  auto &DL = L->getHeader()->getDataLayout();
  const auto &[AccessStart, AccessEnd] =
      getStartAndEndForAccess(L, PtrSCEV, EltSizeSCEV, BECount, MaxBECount, &SE,
                              nullptr, &DT, AC, LoopGuards);
  if (isa<SCEVCouldNotCompute>(AccessStart) ||
      isa<SCEVCouldNotCompute>(AccessEnd))
    return false;
 
  // Try to get the access size.
  const SCEV *PtrDiff = SE.getMinusSCEV(AccessEnd, AccessStart);
  if (isa<SCEVCouldNotCompute>(PtrDiff))
    return false;
 
  if (!LoopGuards)
    LoopGuards.emplace(
        ScalarEvolution::LoopGuards::collect(AddRec->getLoop(), SE));
 
  APInt MaxPtrDiff =
      SE.getUnsignedRangeMax(SE.applyLoopGuards(PtrDiff, *LoopGuards));
 
  Value *Base = nullptr;
  APInt AccessSize;
  const SCEV *AccessSizeSCEV = nullptr;
  if (const SCEVUnknown *NewBase = dyn_cast<SCEVUnknown>(AccessStart)) {
    Base = NewBase->getValue();
    AccessSize = std::move(MaxPtrDiff);
    AccessSizeSCEV = PtrDiff;
  } else if (auto *MinAdd = dyn_cast<SCEVAddExpr>(AccessStart)) {
    if (MinAdd->getNumOperands() != 2)
      return false;
 
    const auto *Offset = dyn_cast<SCEVConstant>(MinAdd->getOperand(0));
    const auto *NewBase = dyn_cast<SCEVUnknown>(MinAdd->getOperand(1));
    if (!Offset || !NewBase)
      return false;
 
    // The following code below assumes the offset is unsigned, but GEP
    // offsets are treated as signed so we can end up with a signed value
    // here too. For example, suppose the initial PHI value is (i8 255),
    // the offset will be treated as (i8 -1) and sign-extended to (i64 -1).
    if (Offset->getAPInt().isNegative())
      return false;
 
    // For the moment, restrict ourselves to the case where the offset is a
    // multiple of the requested alignment and the base is aligned.
    // TODO: generalize if a case found which warrants
    if (Offset->getAPInt().urem(Alignment.value()) != 0)
      return false;
 
    bool Overflow = false;
    AccessSize = MaxPtrDiff.uadd_ov(Offset->getAPInt(), Overflow);
    if (Overflow)
      return false;
    AccessSizeSCEV = SE.getAddExpr(PtrDiff, Offset);
    Base = NewBase->getValue();
  } else
    return false;
 
  Instruction *CtxI = &*L->getHeader()->getFirstNonPHIIt();
  if (BasicBlock *LoopPred = L->getLoopPredecessor()) {
    if (isa<UncondBrInst, CondBrInst>(LoopPred->getTerminator()))
      CtxI = LoopPred->getTerminator();
  }
  SimplifyQuery SQ(DL, &DT, AC, CtxI);
  return isDereferenceableAndAlignedPointerViaAssumption(
             Base, Alignment, SQ, /*IgnoreFree=*/false,
             [&SE, AccessSizeSCEV, &LoopGuards](const RetainedKnowledge &RK) {
               const SCEV *DerefBytesSCEV = SE.getSCEV(RK.IRArgValue);
               Type *WiderTy = SE.getWiderType(AccessSizeSCEV->getType(),
                                               DerefBytesSCEV->getType());
               const SCEV *AccessSizeExt =
                   SE.getNoopOrZeroExtend(AccessSizeSCEV, WiderTy);
               const SCEV *DerefBytesExt =
                   SE.getNoopOrZeroExtend(DerefBytesSCEV, WiderTy);
               return SE.isKnownPredicate(
                   CmpInst::ICMP_ULE,
                   SE.applyLoopGuards(AccessSizeExt, *LoopGuards),
                   SE.applyLoopGuards(DerefBytesExt, *LoopGuards));
             }) ||
         isDereferenceableAndAlignedPointer(Base, Alignment, AccessSize, SQ);
}
 
static bool suppressSpeculativeLoadForSanitizers(const Instruction &CtxI) {
  const Function &F = *CtxI.getFunction();
  // Speculative load may create a race that did not exist in the source.
  return F.hasFnAttribute(Attribute::SanitizeThread) ||
         // Speculative load may load data from dirty regions.
         F.hasFnAttribute(Attribute::SanitizeAddress) ||
         F.hasFnAttribute(Attribute::SanitizeHWAddress);
}
 
bool llvm::mustSuppressSpeculation(const LoadInst &LI) {
  return !LI.isUnordered() || suppressSpeculativeLoadForSanitizers(LI);
}
 
bool llvm::isSafeToLoadUnconditionally(Value *V, Align Alignment,
                                       const APInt &Size,
                                       const SimplifyQuery &SQ) {
  if (isDereferenceableAndAlignedPointer(V, Alignment, Size, SQ)) {
    // With sanitizers `Dereferenceable` is not always enough for unconditional
    // load.
    if (!SQ.CxtI || !suppressSpeculativeLoadForSanitizers(*SQ.CxtI))
      return true;
  }
 
  if (!SQ.CxtI)
    return false;
 
  if (Size.getBitWidth() > 64)
    return false;
  const TypeSize LoadSize = TypeSize::getFixed(Size.getZExtValue());
 
  // Otherwise, be a little bit aggressive by scanning the local block where we
  // want to check to see if the pointer is already being loaded or stored
  // from/to.  If so, the previous load or store would have already trapped,
  // so there is no harm doing an extra load (also, CSE will later eliminate
  // the load entirely).
  auto BBI = SQ.CxtI->getIterator(), E = SQ.CxtI->getParent()->begin();
 
  // We can at least always strip pointer casts even though we can't use the
  // base here.
  V = V->stripPointerCasts();
 
  while (BBI != E) {
    --BBI;
 
    // If we see a free or a call which may write to memory (i.e. which might do
    // a free) the pointer could be marked invalid.
    if (isa<CallInst>(BBI) && BBI->mayWriteToMemory() &&
        !isa<LifetimeIntrinsic>(BBI))
      return false;
 
    const Value *AccessedPtr;
    Type *AccessedTy;
    Align AccessedAlign;
    if (const auto *LI = dyn_cast<LoadInst>(BBI)) {
      // Ignore volatile loads. The execution of a volatile load cannot
      // be used to prove an address is backed by regular memory; it can,
      // for example, point to an MMIO register.
      if (LI->isVolatile())
        continue;
      AccessedPtr = LI->getPointerOperand();
      AccessedTy = LI->getType();
      AccessedAlign = LI->getAlign();
    } else if (const auto *SI = dyn_cast<StoreInst>(BBI)) {
      // Ignore volatile stores (see comment for loads).
      if (SI->isVolatile())
        continue;
      AccessedPtr = SI->getPointerOperand();
      AccessedTy = SI->getValueOperand()->getType();
      AccessedAlign = SI->getAlign();
    } else
      continue;
 
    if (AccessedAlign < Alignment)
      continue;
 
    // Handle trivial cases.
    if (AccessedPtr == V &&
        TypeSize::isKnownLE(LoadSize, SQ.DL.getTypeStoreSize(AccessedTy)))
      return true;
 
    if (AreEquivalentAddressValues(AccessedPtr->stripPointerCasts(), V) &&
        TypeSize::isKnownLE(LoadSize, SQ.DL.getTypeStoreSize(AccessedTy)))
      return true;
  }
  return false;
}
 
bool llvm::isSafeToLoadUnconditionally(Value *V, Type *Ty, Align Alignment,
                                       const SimplifyQuery &SQ) {
  TypeSize TySize = SQ.DL.getTypeStoreSize(Ty);
  if (TySize.isScalable())
    return false;
  APInt Size(SQ.DL.getIndexTypeSizeInBits(V->getType()),
             TySize.getFixedValue());
  return isSafeToLoadUnconditionally(V, Alignment, Size, SQ);
}
 
/// DefMaxInstsToScan - the default number of maximum instructions
/// to scan in the block, used by FindAvailableLoadedValue().
/// FindAvailableLoadedValue() was introduced in r60148, to improve jump
/// threading in part by eliminating partially redundant loads.
/// At that point, the value of MaxInstsToScan was already set to '6'
/// without documented explanation.
cl::opt<unsigned>
llvm::DefMaxInstsToScan("available-load-scan-limit", cl::init(6), cl::Hidden,
  cl::desc("Use this to specify the default maximum number of instructions "
           "to scan backward from a given instruction, when searching for "
           "available loaded value"));
 
Value *llvm::FindAvailableLoadedValue(LoadInst *Load, BasicBlock *ScanBB,
                                      BasicBlock::iterator &ScanFrom,
                                      unsigned MaxInstsToScan,
                                      BatchAAResults *AA, bool *IsLoad,
                                      unsigned *NumScanedInst) {
  // Don't CSE load that is volatile or anything stronger than unordered.
  if (!Load->isUnordered())
    return nullptr;
 
  MemoryLocation Loc = MemoryLocation::get(Load);
  return findAvailablePtrLoadStore(Loc, Load->getType(), Load->getProperties(),
                                   ScanBB, ScanFrom, MaxInstsToScan, AA, IsLoad,
                                   NumScanedInst);
}
 
// Check if the load and the store have the same base, constant offsets and
// non-overlapping access ranges.
static bool areNonOverlapSameBaseLoadAndStore(const Value *LoadPtr,
                                              Type *LoadTy,
                                              const Value *StorePtr,
                                              Type *StoreTy,
                                              const DataLayout &DL) {
  APInt LoadOffset(DL.getIndexTypeSizeInBits(LoadPtr->getType()), 0);
  APInt StoreOffset(DL.getIndexTypeSizeInBits(StorePtr->getType()), 0);
  if (LoadOffset.getBitWidth() != StoreOffset.getBitWidth())
    return false;
  const Value *LoadBase = LoadPtr->stripAndAccumulateConstantOffsets(
      DL, LoadOffset, /* AllowNonInbounds */ false);
  const Value *StoreBase = StorePtr->stripAndAccumulateConstantOffsets(
      DL, StoreOffset, /* AllowNonInbounds */ false);
  if (LoadBase != StoreBase)
    return false;
  auto LoadAccessSize = LocationSize::precise(DL.getTypeStoreSize(LoadTy));
  auto StoreAccessSize = LocationSize::precise(DL.getTypeStoreSize(StoreTy));
  ConstantRange LoadRange(LoadOffset,
                          LoadOffset + LoadAccessSize.toRaw());
  ConstantRange StoreRange(StoreOffset,
                           StoreOffset + StoreAccessSize.toRaw());
  return LoadRange.intersectWith(StoreRange).isEmptySet();
}
 
// For elementwise atomics, each vector element is a separate atomic access.
// Reusing an operation with a different access size would change atomicity.
static bool hasCompatibleAtomicAccessSize(
    Type *AccessTy, const LoadStoreInstProperties &AccessProps,
    Type *OtherAccessTy, const LoadStoreInstProperties &OtherProps,
    const DataLayout &DL) {
  if (AccessProps.Ordering == AtomicOrdering::NotAtomic)
    return true;
 
  Type *AtomicAccessTy =
      AccessProps.IsElementwise ? AccessTy->getScalarType() : AccessTy;
  Type *OtherAtomicAccessTy =
      OtherProps.IsElementwise ? OtherAccessTy->getScalarType() : OtherAccessTy;
  return DL.getTypeStoreSize(AtomicAccessTy) ==
         DL.getTypeStoreSize(OtherAtomicAccessTy);
}
 
static Value *getAvailableLoadStore(Instruction *Inst, const Value *Ptr,
                                    Type *AccessTy,
                                    const LoadStoreInstProperties &AccessProps,
                                    const DataLayout &DL, bool *IsLoadCSE) {
  const bool AtLeastAtomic = AccessProps.Ordering != AtomicOrdering::NotAtomic;
 
  // If this is a load of Ptr, the loaded value is available.
  // (This is true even if the load is volatile or atomic, although
  // those cases are unlikely.)
  if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
    // We can value forward from an atomic to a non-atomic, but not the
    // other way around.
    if (LI->isAtomic() < AtLeastAtomic)
      return nullptr;
 
    Value *LoadPtr = LI->getPointerOperand()->stripPointerCasts();
    if (!AreEquivalentAddressValues(LoadPtr, Ptr))
      return nullptr;
 
    if (hasCompatibleAtomicAccessSize(AccessTy, AccessProps, LI->getType(),
                                      LI->getProperties(), DL) &&
        CastInst::isBitOrNoopPointerCastable(LI->getType(), AccessTy, DL)) {
      if (IsLoadCSE)
        *IsLoadCSE = true;
      return LI;
    }
  }
 
  // If this is a store through Ptr, the value is available!
  // (This is true even if the store is volatile or atomic, although
  // those cases are unlikely.)
  if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
    // We can value forward from an atomic to a non-atomic, but not the
    // other way around.
    if (SI->isAtomic() < AtLeastAtomic)
      return nullptr;
 
    Value *StorePtr = SI->getPointerOperand()->stripPointerCasts();
    if (!AreEquivalentAddressValues(StorePtr, Ptr))
      return nullptr;
 
    if (IsLoadCSE)
      *IsLoadCSE = false;
 
    Value *Val = SI->getValueOperand();
    if (!hasCompatibleAtomicAccessSize(AccessTy, AccessProps, Val->getType(),
                                       SI->getProperties(), DL))
      return nullptr;
 
    if (CastInst::isBitOrNoopPointerCastable(Val->getType(), AccessTy, DL))
      return Val;
 
    TypeSize StoreSize = DL.getTypeSizeInBits(Val->getType());
    TypeSize LoadSize = DL.getTypeSizeInBits(AccessTy);
    if (TypeSize::isKnownLE(LoadSize, StoreSize))
      if (auto *C = dyn_cast<Constant>(Val))
        return ConstantFoldLoadFromConst(C, AccessTy, DL);
  }
 
  if (auto *MSI = dyn_cast<MemSetInst>(Inst)) {
    // Don't forward from (non-atomic) memset to atomic load.
    if (AtLeastAtomic)
      return nullptr;
 
    // Only handle constant memsets.
    auto *Val = dyn_cast<ConstantInt>(MSI->getValue());
    auto *Len = dyn_cast<ConstantInt>(MSI->getLength());
    if (!Val || !Len)
      return nullptr;
 
    // Handle offsets.
    int64_t StoreOffset = 0, LoadOffset = 0;
    const Value *StoreBase =
        GetPointerBaseWithConstantOffset(MSI->getDest(), StoreOffset, DL);
    const Value *LoadBase =
        GetPointerBaseWithConstantOffset(Ptr, LoadOffset, DL);
    if (StoreBase != LoadBase || LoadOffset < StoreOffset)
      return nullptr;
 
    if (IsLoadCSE)
      *IsLoadCSE = false;
 
    TypeSize LoadTypeSize = DL.getTypeSizeInBits(AccessTy);
    if (LoadTypeSize.isScalable())
      return nullptr;
 
    // Make sure the read bytes are contained in the memset.
    uint64_t LoadSize = LoadTypeSize.getFixedValue();
    if ((Len->getValue() * 8).ult(LoadSize + (LoadOffset - StoreOffset) * 8))
      return nullptr;
 
    APInt Splat = LoadSize >= 8 ? APInt::getSplat(LoadSize, Val->getValue())
                                : Val->getValue().trunc(LoadSize);
    ConstantInt *SplatC = ConstantInt::get(MSI->getContext(), Splat);
    if (CastInst::isBitOrNoopPointerCastable(SplatC->getType(), AccessTy, DL))
      return SplatC;
 
    return nullptr;
  }
 
  return nullptr;
}
 
Value *llvm::findAvailablePtrLoadStore(
    const MemoryLocation &Loc, Type *AccessTy,
    const LoadStoreInstProperties &AccessProps, BasicBlock *ScanBB,
    BasicBlock::iterator &ScanFrom, unsigned MaxInstsToScan, BatchAAResults *AA,
    bool *IsLoadCSE, unsigned *NumScanedInst) {
  if (MaxInstsToScan == 0)
    MaxInstsToScan = ~0U;
 
  const DataLayout &DL = ScanBB->getDataLayout();
  const Value *StrippedPtr = Loc.Ptr->stripPointerCasts();
 
  while (ScanFrom != ScanBB->begin()) {
    // We must ignore debug info directives when counting (otherwise they
    // would affect codegen).
    Instruction *Inst = &*--ScanFrom;
    if (Inst->isDebugOrPseudoInst())
      continue;
 
    // Restore ScanFrom to expected value in case next test succeeds
    ScanFrom++;
 
    if (NumScanedInst)
      ++(*NumScanedInst);
 
    // Don't scan huge blocks.
    if (MaxInstsToScan-- == 0)
      return nullptr;
 
    --ScanFrom;
 
    if (Value *Available = getAvailableLoadStore(Inst, StrippedPtr, AccessTy,
                                                 AccessProps, DL, IsLoadCSE))
      return Available;
 
    // Try to get the store size for the type.
    if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
      Value *StorePtr = SI->getPointerOperand()->stripPointerCasts();
 
      // If both StrippedPtr and StorePtr reach all the way to an alloca or
      // global and they are different, ignore the store. This is a trivial form
      // of alias analysis that is important for reg2mem'd code.
      if ((isa<AllocaInst>(StrippedPtr) || isa<GlobalVariable>(StrippedPtr)) &&
          (isa<AllocaInst>(StorePtr) || isa<GlobalVariable>(StorePtr)) &&
          StrippedPtr != StorePtr)
        continue;
 
      if (!AA) {
        // When AA isn't available, but if the load and the store have the same
        // base, constant offsets and non-overlapping access ranges, ignore the
        // store. This is a simple form of alias analysis that is used by the
        // inliner. FIXME: use BasicAA if possible.
        if (areNonOverlapSameBaseLoadAndStore(
                Loc.Ptr, AccessTy, SI->getPointerOperand(),
                SI->getValueOperand()->getType(), DL))
          continue;
      } else {
        // If we have alias analysis and it says the store won't modify the
        // loaded value, ignore the store.
        if (!isModSet(AA->getModRefInfo(SI, Loc)))
          continue;
      }
 
      // Otherwise the store that may or may not alias the pointer, bail out.
      ++ScanFrom;
      return nullptr;
    }
 
    // If this is some other instruction that may clobber Ptr, bail out.
    if (Inst->mayWriteToMemory()) {
      // If alias analysis claims that it really won't modify the load,
      // ignore it.
      if (AA && !isModSet(AA->getModRefInfo(Inst, Loc)))
        continue;
 
      // May modify the pointer, bail out.
      ++ScanFrom;
      return nullptr;
    }
  }
 
  // Got to the start of the block, we didn't find it, but are done for this
  // block.
  return nullptr;
}
 
Value *llvm::FindAvailableLoadedValue(LoadInst *Load, BatchAAResults &AA,
                                      bool *IsLoadCSE,
                                      unsigned MaxInstsToScan) {
  const DataLayout &DL = Load->getDataLayout();
  Value *StrippedPtr = Load->getPointerOperand()->stripPointerCasts();
  BasicBlock *ScanBB = Load->getParent();
  Type *AccessTy = Load->getType();
  LoadStoreInstProperties AccessProps = Load->getProperties();
 
  if (!Load->isUnordered())
    return nullptr;
 
  // Try to find an available value first, and delay expensive alias analysis
  // queries until later.
  Value *Available = nullptr;
  SmallVector<Instruction *> MustNotAliasInsts;
  for (Instruction &Inst : make_range(++Load->getReverseIterator(),
                                      ScanBB->rend())) {
    if (Inst.isDebugOrPseudoInst())
      continue;
 
    if (MaxInstsToScan-- == 0)
      return nullptr;
 
    Available = getAvailableLoadStore(&Inst, StrippedPtr, AccessTy, AccessProps,
                                      DL, IsLoadCSE);
    if (Available)
      break;
 
    if (Inst.mayWriteToMemory())
      MustNotAliasInsts.push_back(&Inst);
  }
 
  // If we found an available value, ensure that the instructions in between
  // did not modify the memory location.
  if (Available) {
    MemoryLocation Loc = MemoryLocation::get(Load);
    for (Instruction *Inst : MustNotAliasInsts)
      if (isModSet(AA.getModRefInfo(Inst, Loc)))
        return nullptr;
  }
 
  return Available;
}
 
bool llvm::isStorePreservingMemoryLocation(const StoreInst *SI,
                                           const MemoryLocation &MemLoc,
                                           Align MemLocAlign,
                                           BatchAAResults &AA,
                                           unsigned ScanLimit) {
  // Ensure no partial overlap is possible, and that the stored value is the
  // current content of MemLoc.
  if (!MemLoc.Size.hasValue() || MemLoc.Size.isScalable())
    return false;
  if (MemoryLocation::get(SI).Size != MemLoc.Size)
    return false;
  if (std::min(MemLocAlign, SI->getAlign()).value() <
      MemLoc.Size.getValue().getFixedValue())
    return false;
 
  auto *LI = dyn_cast<LoadInst>(SI->getValueOperand());
  if (!LI || LI->getParent() != SI->getParent())
    return false;
  if (AA.alias(MemoryLocation::get(LI), MemLoc) != AliasResult::MustAlias)
    return false;
 
  // No memory operation in between may modify MemLoc.
  unsigned NumVisited = 0;
  for (const Instruction *I = LI; I != SI; I = I->getNextNode())
    if (++NumVisited > ScanLimit || isModSet(AA.getModRefInfo(I, MemLoc)))
      return false;
 
  return true;
}
 
// Returns true if a use is either in an ICmp/PtrToInt or a Phi/Select that only
// feeds into them.
static bool isPointerUseReplaceable(const Use &U, bool HasNonAddressBits) {
  unsigned Limit = 40;
  SmallVector<const User *> Worklist({U.getUser()});
  SmallPtrSet<const User *, 8> Visited;
 
  while (!Worklist.empty() && --Limit) {
    auto *User = Worklist.pop_back_val();
    if (!Visited.insert(User).second)
      continue;
    if (isa<ICmpInst, PtrToAddrInst>(User))
      continue;
    // FIXME: The PtrToIntInst case here is not strictly correct, as it
    // changes which provenance is exposed.
    if (!HasNonAddressBits && isa<PtrToIntInst>(User))
      continue;
    if (isa<PHINode, SelectInst>(User))
      Worklist.append(User->user_begin(), User->user_end());
    else
      return false;
  }
 
  return Limit != 0;
}
 
static bool isPointerAlwaysReplaceable(const Value *From, const Value *To,
                                       const DataLayout &DL) {
  // This is not strictly correct, but we do it for now to retain important
  // optimizations.
  if (isa<ConstantPointerNull>(To))
    return true;
  // Conversely, replacing null in the default address space with destination
  // pointer is always valid.
  if (isa<ConstantPointerNull>(From) &&
      From->getType()->getPointerAddressSpace() == 0)
    return true;
  // Allow replacement with dereferenceable constants. This is not strictly
  // correct, but required for vtable assumptions.
  auto IsBasedOnConstantGlobal = [](const Value *V) {
    auto *GV = dyn_cast<GlobalVariable>(getUnderlyingObject(V));
    return GV && GV->isConstant();
  };
  if (isa<Constant>(To) && To->getType()->isPointerTy() &&
      isDereferenceablePointer(To, Type::getInt8Ty(To->getContext()), DL) &&
      IsBasedOnConstantGlobal(To))
    return true;
  return getUnderlyingObjectAggressive(From, /*MustPreserveProvenance=*/true) ==
         getUnderlyingObjectAggressive(To, /*MustPreserveProvenance=*/true);
}
 
bool llvm::canReplacePointersInUseIfEqual(const Use &U, const Value *To,
                                          const DataLayout &DL) {
  Type *Ty = To->getType();
  assert(U->getType() == Ty && "values must have matching types");
  // Not a pointer, just return true.
  if (!Ty->isPtrOrPtrVectorTy())
    return true;
 
  // Do not perform replacements in lifetime intrinsic arguments.
  if (isa<LifetimeIntrinsic>(U.getUser()))
    return false;
 
  if (isPointerAlwaysReplaceable(&*U, To, DL))
    return true;
 
  bool HasNonAddressBits =
      DL.getAddressSizeInBits(Ty) != DL.getPointerTypeSizeInBits(Ty);
  return isPointerUseReplaceable(U, HasNonAddressBits);
}
 
bool llvm::canReplacePointersIfEqual(const Value *From, const Value *To,
                                     const DataLayout &DL) {
  assert(From->getType() == To->getType() && "values must have matching types");
  // Not a pointer, just return true.
  if (!From->getType()->isPtrOrPtrVectorTy())
    return true;
 
  return isPointerAlwaysReplaceable(From, To, DL);
}
 
bool llvm::isReadOnlyLoop(
    Loop *L, ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC,
    SmallVectorImpl<LoadInst *> &NonDereferenceableAndAlignedLoads,
    SmallVectorImpl<const SCEVPredicate *> *Predicates) {
  for (BasicBlock *BB : L->blocks()) {
    for (Instruction &I : *BB) {
      if (auto *LI = dyn_cast<LoadInst>(&I)) {
        if (!isDereferenceableAndAlignedInLoop(LI, L, *SE, *DT, AC, Predicates))
          NonDereferenceableAndAlignedLoads.push_back(LI);
      } else if (I.mayReadFromMemory() || I.mayWriteToMemory() ||
                 I.mayThrow()) {
        return false;
      }
    }
  }
  return true;
}
 
LinearExpression llvm::decomposeLinearExpression(const DataLayout &DL,
                                                 Value *Ptr) {
  assert(Ptr->getType()->isPointerTy() && "Must be called with pointer arg");
 
  unsigned BitWidth = DL.getIndexTypeSizeInBits(Ptr->getType());
  LinearExpression Expr(Ptr, BitWidth);
 
  while (true) {
    auto *GEP = dyn_cast<GEPOperator>(Expr.BasePtr);
    if (!GEP || GEP->getSourceElementType()->isScalableTy())
      return Expr;
 
    Value *VarIndex = nullptr;
    for (Value *Index : GEP->indices()) {
      if (isa<ConstantInt>(Index))
        continue;
      // Only allow a single variable index. We do not bother to handle the
      // case of the same variable index appearing multiple times.
      if (Expr.Index || VarIndex)
        return Expr;
      VarIndex = Index;
    }
 
    // Don't return non-canonical indexes.
    if (VarIndex && !VarIndex->getType()->isIntegerTy(BitWidth))
      return Expr;
 
    // We have verified that we can fully handle this GEP, so we can update Expr
    // members past this point.
    Expr.BasePtr = GEP->getPointerOperand();
    Expr.Flags = Expr.Flags.intersectForOffsetAdd(GEP->getNoWrapFlags());
    for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
         GTI != GTE; ++GTI) {
      Value *Index = GTI.getOperand();
      if (auto *ConstOffset = dyn_cast<ConstantInt>(Index)) {
        if (ConstOffset->isZero())
          continue;
        if (StructType *STy = GTI.getStructTypeOrNull()) {
          unsigned ElementIdx = ConstOffset->getZExtValue();
          const StructLayout *SL = DL.getStructLayout(STy);
          Expr.Offset += SL->getElementOffset(ElementIdx);
          continue;
        }
        // Truncate if type size exceeds index space.
        APInt IndexedSize(BitWidth, GTI.getSequentialElementStride(DL),
                          /*isSigned=*/false,
                          /*implcitTrunc=*/true);
        Expr.Offset += ConstOffset->getValue() * IndexedSize;
        continue;
      }
 
      // FIXME: Also look through a mul/shl in the index.
      assert(Expr.Index == nullptr && "Shouldn't have index yet");
      Expr.Index = Index;
      // Truncate if type size exceeds index space.
      Expr.Scale = APInt(BitWidth, GTI.getSequentialElementStride(DL),
                         /*isSigned=*/false, /*implicitTrunc=*/true);
    }
  }
 
  return Expr;
}