bool AMDGPUCodeGenPrepare::visitLoadInst(LoadInst &I) {
  if (!WidenLoads)
    return false;

  if ((I.getPointerAddressSpace() == AMDGPUASI.CONSTANT_ADDRESS ||
       I.getPointerAddressSpace() == AMDGPUASI.CONSTANT_ADDRESS_32BIT) &&
      canWidenScalarExtLoad(I)) {
    IRBuilder<> Builder(&I);
    Builder.SetCurrentDebugLocation(I.getDebugLoc());

    Type *I32Ty = Builder.getInt32Ty();
    Type *PT = PointerType::get(I32Ty, I.getPointerAddressSpace());
    Value *BitCast= Builder.CreateBitCast(I.getPointerOperand(), PT);
    LoadInst *WidenLoad = Builder.CreateLoad(BitCast);
    WidenLoad->copyMetadata(I);

    // If we have range metadata, we need to convert the type, and not make
    // assumptions about the high bits.
    if (auto *Range = WidenLoad->getMetadata(LLVMContext::MD_range)) {
      ConstantInt *Lower =
        mdconst::extract<ConstantInt>(Range->getOperand(0));

      if (Lower->getValue().isNullValue()) {
        WidenLoad->setMetadata(LLVMContext::MD_range, nullptr);
      } else {
        Metadata *LowAndHigh[] = {
          ConstantAsMetadata::get(ConstantInt::get(I32Ty, Lower->getValue().zext(32))),
          // Don't make assumptions about the high bits.
          ConstantAsMetadata::get(ConstantInt::get(I32Ty, 0))
        };

        WidenLoad->setMetadata(LLVMContext::MD_range,
                               MDNode::get(Mod->getContext(), LowAndHigh));
      }
    }

    int TySize = Mod->getDataLayout().getTypeSizeInBits(I.getType());
    Type *IntNTy = Builder.getIntNTy(TySize);
    Value *ValTrunc = Builder.CreateTrunc(WidenLoad, IntNTy);
    Value *ValOrig = Builder.CreateBitCast(ValTrunc, I.getType());
    I.replaceAllUsesWith(ValOrig);
    I.eraseFromParent();
    return true;
  }

  return false;
}
Пример #2
0
/// DoPromotion - This method actually performs the promotion of the specified
/// arguments, and returns the new function.  At this point, we know that it's
/// safe to do so.
CallGraphNode *ArgPromotion::DoPromotion(Function *F,
                               SmallPtrSet<Argument*, 8> &ArgsToPromote,
                              SmallPtrSet<Argument*, 8> &ByValArgsToTransform) {

  // Start by computing a new prototype for the function, which is the same as
  // the old function, but has modified arguments.
  const FunctionType *FTy = F->getFunctionType();
  std::vector<const Type*> Params;

  typedef std::set<IndicesVector> ScalarizeTable;

  // ScalarizedElements - If we are promoting a pointer that has elements
  // accessed out of it, keep track of which elements are accessed so that we
  // can add one argument for each.
  //
  // Arguments that are directly loaded will have a zero element value here, to
  // handle cases where there are both a direct load and GEP accesses.
  //
  std::map<Argument*, ScalarizeTable> ScalarizedElements;

  // OriginalLoads - Keep track of a representative load instruction from the
  // original function so that we can tell the alias analysis implementation
  // what the new GEP/Load instructions we are inserting look like.
  std::map<IndicesVector, LoadInst*> OriginalLoads;

  // Attributes - Keep track of the parameter attributes for the arguments
  // that we are *not* promoting. For the ones that we do promote, the parameter
  // attributes are lost
  SmallVector<AttributeWithIndex, 8> AttributesVec;
  const AttrListPtr &PAL = F->getAttributes();

  // Add any return attributes.
  if (Attributes attrs = PAL.getRetAttributes())
    AttributesVec.push_back(AttributeWithIndex::get(0, attrs));

  // First, determine the new argument list
  unsigned ArgIndex = 1;
  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
       ++I, ++ArgIndex) {
    if (ByValArgsToTransform.count(I)) {
      // Simple byval argument? Just add all the struct element types.
      const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
      const StructType *STy = cast<StructType>(AgTy);
      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
        Params.push_back(STy->getElementType(i));
      ++NumByValArgsPromoted;
    } else if (!ArgsToPromote.count(I)) {
      // Unchanged argument
      Params.push_back(I->getType());
      if (Attributes attrs = PAL.getParamAttributes(ArgIndex))
        AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs));
    } else if (I->use_empty()) {
      // Dead argument (which are always marked as promotable)
      ++NumArgumentsDead;
    } else {
      // Okay, this is being promoted. This means that the only uses are loads
      // or GEPs which are only used by loads

      // In this table, we will track which indices are loaded from the argument
      // (where direct loads are tracked as no indices).
      ScalarizeTable &ArgIndices = ScalarizedElements[I];
      for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
           ++UI) {
        Instruction *User = cast<Instruction>(*UI);
        assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User));
        IndicesVector Indices;
        Indices.reserve(User->getNumOperands() - 1);
        // Since loads will only have a single operand, and GEPs only a single
        // non-index operand, this will record direct loads without any indices,
        // and gep+loads with the GEP indices.
        for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end();
             II != IE; ++II)
          Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
        // GEPs with a single 0 index can be merged with direct loads
        if (Indices.size() == 1 && Indices.front() == 0)
          Indices.clear();
        ArgIndices.insert(Indices);
        LoadInst *OrigLoad;
        if (LoadInst *L = dyn_cast<LoadInst>(User))
          OrigLoad = L;
        else
          // Take any load, we will use it only to update Alias Analysis
          OrigLoad = cast<LoadInst>(User->use_back());
        OriginalLoads[Indices] = OrigLoad;
      }

      // Add a parameter to the function for each element passed in.
      for (ScalarizeTable::iterator SI = ArgIndices.begin(),
             E = ArgIndices.end(); SI != E; ++SI) {
        // not allowed to dereference ->begin() if size() is 0
        Params.push_back(GetElementPtrInst::getIndexedType(I->getType(),
                                                           SI->begin(),
                                                           SI->end()));
        assert(Params.back());
      }

      if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
        ++NumArgumentsPromoted;
      else
        ++NumAggregatesPromoted;
    }
  }

  // Add any function attributes.
  if (Attributes attrs = PAL.getFnAttributes())
    AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));

  const Type *RetTy = FTy->getReturnType();

  // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which
  // have zero fixed arguments.
  bool ExtraArgHack = false;
  if (Params.empty() && FTy->isVarArg()) {
    ExtraArgHack = true;
    Params.push_back(Type::getInt32Ty(F->getContext()));
  }

  // Construct the new function type using the new arguments.
  FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());

  // Create the new function body and insert it into the module.
  Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
  NF->copyAttributesFrom(F);

  
  DEBUG(dbgs() << "ARG PROMOTION:  Promoting to:" << *NF << "\n"
        << "From: " << *F);
  
  // Recompute the parameter attributes list based on the new arguments for
  // the function.
  NF->setAttributes(AttrListPtr::get(AttributesVec.begin(),
                                     AttributesVec.end()));
  AttributesVec.clear();

  F->getParent()->getFunctionList().insert(F, NF);
  NF->takeName(F);

  // Get the alias analysis information that we need to update to reflect our
  // changes.
  AliasAnalysis &AA = getAnalysis<AliasAnalysis>();

  // Get the callgraph information that we need to update to reflect our
  // changes.
  CallGraph &CG = getAnalysis<CallGraph>();
  
  // Get a new callgraph node for NF.
  CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);

  // Loop over all of the callers of the function, transforming the call sites
  // to pass in the loaded pointers.
  //
  SmallVector<Value*, 16> Args;
  while (!F->use_empty()) {
    CallSite CS(F->use_back());
    assert(CS.getCalledFunction() == F);
    Instruction *Call = CS.getInstruction();
    const AttrListPtr &CallPAL = CS.getAttributes();

    // Add any return attributes.
    if (Attributes attrs = CallPAL.getRetAttributes())
      AttributesVec.push_back(AttributeWithIndex::get(0, attrs));

    // Loop over the operands, inserting GEP and loads in the caller as
    // appropriate.
    CallSite::arg_iterator AI = CS.arg_begin();
    ArgIndex = 1;
    for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
         I != E; ++I, ++AI, ++ArgIndex)
      if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
        Args.push_back(*AI);          // Unmodified argument

        if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
          AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));

      } else if (ByValArgsToTransform.count(I)) {
        // Emit a GEP and load for each element of the struct.
        const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
        const StructType *STy = cast<StructType>(AgTy);
        Value *Idxs[2] = {
              ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
        for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
          Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
          Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2,
                                                 (*AI)->getName()+"."+utostr(i),
                                                 Call);
          // TODO: Tell AA about the new values?
          Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
        }
      } else if (!I->use_empty()) {
        // Non-dead argument: insert GEPs and loads as appropriate.
        ScalarizeTable &ArgIndices = ScalarizedElements[I];
        // Store the Value* version of the indices in here, but declare it now
        // for reuse.
        std::vector<Value*> Ops;
        for (ScalarizeTable::iterator SI = ArgIndices.begin(),
               E = ArgIndices.end(); SI != E; ++SI) {
          Value *V = *AI;
          LoadInst *OrigLoad = OriginalLoads[*SI];
          if (!SI->empty()) {
            Ops.reserve(SI->size());
            const Type *ElTy = V->getType();
            for (IndicesVector::const_iterator II = SI->begin(),
                 IE = SI->end(); II != IE; ++II) {
              // Use i32 to index structs, and i64 for others (pointers/arrays).
              // This satisfies GEP constraints.
              const Type *IdxTy = (ElTy->isStructTy() ?
                    Type::getInt32Ty(F->getContext()) : 
                    Type::getInt64Ty(F->getContext()));
              Ops.push_back(ConstantInt::get(IdxTy, *II));
              // Keep track of the type we're currently indexing.
              ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
            }
            // And create a GEP to extract those indices.
            V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(),
                                          V->getName()+".idx", Call);
            Ops.clear();
            AA.copyValue(OrigLoad->getOperand(0), V);
          }
          // Since we're replacing a load make sure we take the alignment
          // of the previous load.
          LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
          newLoad->setAlignment(OrigLoad->getAlignment());
          // Transfer the TBAA info too.
          newLoad->setMetadata(LLVMContext::MD_tbaa,
                               OrigLoad->getMetadata(LLVMContext::MD_tbaa));
          Args.push_back(newLoad);
          AA.copyValue(OrigLoad, Args.back());
        }
      }

    if (ExtraArgHack)
      Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext())));

    // Push any varargs arguments on the list.
    for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
      Args.push_back(*AI);
      if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
        AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
    }

    // Add any function attributes.
    if (Attributes attrs = CallPAL.getFnAttributes())
      AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));

    Instruction *New;
    if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
      New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
                               Args.begin(), Args.end(), "", Call);
      cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
      cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
                                                          AttributesVec.end()));
    } else {
      New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call);
      cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
      cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
                                                        AttributesVec.end()));
      if (cast<CallInst>(Call)->isTailCall())
        cast<CallInst>(New)->setTailCall();
    }
    Args.clear();
    AttributesVec.clear();

    // Update the alias analysis implementation to know that we are replacing
    // the old call with a new one.
    AA.replaceWithNewValue(Call, New);

    // Update the callgraph to know that the callsite has been transformed.
    CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
    CalleeNode->replaceCallEdge(Call, New, NF_CGN);

    if (!Call->use_empty()) {
      Call->replaceAllUsesWith(New);
      New->takeName(Call);
    }

    // Finally, remove the old call from the program, reducing the use-count of
    // F.
    Call->eraseFromParent();
  }

  // Since we have now created the new function, splice the body of the old
  // function right into the new function, leaving the old rotting hulk of the
  // function empty.
  NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());

  // Loop over the argument list, transferring uses of the old arguments over to
  // the new arguments, also transferring over the names as well.
  //
  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
       I2 = NF->arg_begin(); I != E; ++I) {
    if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
      // If this is an unmodified argument, move the name and users over to the
      // new version.
      I->replaceAllUsesWith(I2);
      I2->takeName(I);
      AA.replaceWithNewValue(I, I2);
      ++I2;
      continue;
    }

    if (ByValArgsToTransform.count(I)) {
      // In the callee, we create an alloca, and store each of the new incoming
      // arguments into the alloca.
      Instruction *InsertPt = NF->begin()->begin();

      // Just add all the struct element types.
      const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
      Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt);
      const StructType *STy = cast<StructType>(AgTy);
      Value *Idxs[2] = {
            ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };

      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
        Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
        Value *Idx = 
          GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2,
                                    TheAlloca->getName()+"."+Twine(i), 
                                    InsertPt);
        I2->setName(I->getName()+"."+Twine(i));
        new StoreInst(I2++, Idx, InsertPt);
      }

      // Anything that used the arg should now use the alloca.
      I->replaceAllUsesWith(TheAlloca);
      TheAlloca->takeName(I);
      AA.replaceWithNewValue(I, TheAlloca);
      continue;
    }

    if (I->use_empty()) {
      AA.deleteValue(I);
      continue;
    }

    // Otherwise, if we promoted this argument, then all users are load
    // instructions (or GEPs with only load users), and all loads should be
    // using the new argument that we added.
    ScalarizeTable &ArgIndices = ScalarizedElements[I];

    while (!I->use_empty()) {
      if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
        assert(ArgIndices.begin()->empty() &&
               "Load element should sort to front!");
        I2->setName(I->getName()+".val");
        LI->replaceAllUsesWith(I2);
        AA.replaceWithNewValue(LI, I2);
        LI->eraseFromParent();
        DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
              << "' in function '" << F->getName() << "'\n");
      } else {
        GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
        IndicesVector Operands;
        Operands.reserve(GEP->getNumIndices());
        for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
             II != IE; ++II)
          Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());

        // GEPs with a single 0 index can be merged with direct loads
        if (Operands.size() == 1 && Operands.front() == 0)
          Operands.clear();

        Function::arg_iterator TheArg = I2;
        for (ScalarizeTable::iterator It = ArgIndices.begin();
             *It != Operands; ++It, ++TheArg) {
          assert(It != ArgIndices.end() && "GEP not handled??");
        }

        std::string NewName = I->getName();
        for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
            NewName += "." + utostr(Operands[i]);
        }
        NewName += ".val";
        TheArg->setName(NewName);

        DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
              << "' of function '" << NF->getName() << "'\n");

        // All of the uses must be load instructions.  Replace them all with
        // the argument specified by ArgNo.
        while (!GEP->use_empty()) {
          LoadInst *L = cast<LoadInst>(GEP->use_back());
          L->replaceAllUsesWith(TheArg);
          AA.replaceWithNewValue(L, TheArg);
          L->eraseFromParent();
        }
        AA.deleteValue(GEP);
        GEP->eraseFromParent();
      }
    }

    // Increment I2 past all of the arguments added for this promoted pointer.
    for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i)
      ++I2;
  }

  // Notify the alias analysis implementation that we inserted a new argument.
  if (ExtraArgHack)
    AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())), 
                 NF->arg_begin());


  // Tell the alias analysis that the old function is about to disappear.
  AA.replaceWithNewValue(F, NF);

  
  NF_CGN->stealCalledFunctionsFrom(CG[F]);
  
  // Now that the old function is dead, delete it.  If there is a dangling
  // reference to the CallgraphNode, just leave the dead function around for
  // someone else to nuke.
  CallGraphNode *CGN = CG[F];
  if (CGN->getNumReferences() == 0)
    delete CG.removeFunctionFromModule(CGN);
  else
    F->setLinkage(Function::ExternalLinkage);
  
  return NF_CGN;
}
Пример #3
0
/// Many allocas are only used within a single basic block.  If this is the
/// case, avoid traversing the CFG and inserting a lot of potentially useless
/// PHI nodes by just performing a single linear pass over the basic block
/// using the Alloca.
///
/// If we cannot promote this alloca (because it is read before it is written),
/// return false.  This is necessary in cases where, due to control flow, the
/// alloca is undefined only on some control flow paths.  e.g. code like
/// this is correct in LLVM IR:
///  // A is an alloca with no stores so far
///  for (...) {
///    int t = *A;
///    if (!first_iteration)
///      use(t);
///    *A = 42;
///  }
static bool promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
                                     LargeBlockInfo &LBI,
                                     const DataLayout &DL,
                                     DominatorTree &DT,
                                     AssumptionCache *AC) {
  // The trickiest case to handle is when we have large blocks. Because of this,
  // this code is optimized assuming that large blocks happen.  This does not
  // significantly pessimize the small block case.  This uses LargeBlockInfo to
  // make it efficient to get the index of various operations in the block.

  // Walk the use-def list of the alloca, getting the locations of all stores.
  using StoresByIndexTy = SmallVector<std::pair<unsigned, StoreInst *>, 64>;
  StoresByIndexTy StoresByIndex;

  for (User *U : AI->users())
    if (StoreInst *SI = dyn_cast<StoreInst>(U))
      StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));

  // Sort the stores by their index, making it efficient to do a lookup with a
  // binary search.
  llvm::sort(StoresByIndex, less_first());

  // Walk all of the loads from this alloca, replacing them with the nearest
  // store above them, if any.
  for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
    LoadInst *LI = dyn_cast<LoadInst>(*UI++);
    if (!LI)
      continue;

    unsigned LoadIdx = LBI.getInstructionIndex(LI);

    // Find the nearest store that has a lower index than this load.
    StoresByIndexTy::iterator I =
        std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
                         std::make_pair(LoadIdx,
                                        static_cast<StoreInst *>(nullptr)),
                         less_first());
    if (I == StoresByIndex.begin()) {
      if (StoresByIndex.empty())
        // If there are no stores, the load takes the undef value.
        LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
      else
        // There is no store before this load, bail out (load may be affected
        // by the following stores - see main comment).
        return false;
    } else {
      // Otherwise, there was a store before this load, the load takes its value.
      // Note, if the load was marked as nonnull we don't want to lose that
      // information when we erase it. So we preserve it with an assume.
      Value *ReplVal = std::prev(I)->second->getOperand(0);
      if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
          !isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT))
        addAssumeNonNull(AC, LI);

      // If the replacement value is the load, this must occur in unreachable
      // code.
      if (ReplVal == LI)
        ReplVal = UndefValue::get(LI->getType());

      LI->replaceAllUsesWith(ReplVal);
    }

    LI->eraseFromParent();
    LBI.deleteValue(LI);
  }

  // Remove the (now dead) stores and alloca.
  while (!AI->use_empty()) {
    StoreInst *SI = cast<StoreInst>(AI->user_back());
    // Record debuginfo for the store before removing it.
    for (DbgVariableIntrinsic *DII : Info.DbgDeclares) {
      DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false);
      ConvertDebugDeclareToDebugValue(DII, SI, DIB);
    }
    SI->eraseFromParent();
    LBI.deleteValue(SI);
  }

  AI->eraseFromParent();
  LBI.deleteValue(AI);

  // The alloca's debuginfo can be removed as well.
  for (DbgVariableIntrinsic *DII : Info.DbgDeclares) {
    DII->eraseFromParent();
    LBI.deleteValue(DII);
  }

  ++NumLocalPromoted;
  return true;
}
Пример #4
0
/// Rewrite as many loads as possible given a single store.
///
/// When there is only a single store, we can use the domtree to trivially
/// replace all of the dominated loads with the stored value. Do so, and return
/// true if this has successfully promoted the alloca entirely. If this returns
/// false there were some loads which were not dominated by the single store
/// and thus must be phi-ed with undef. We fall back to the standard alloca
/// promotion algorithm in that case.
static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
                                     LargeBlockInfo &LBI, const DataLayout &DL,
                                     DominatorTree &DT, AssumptionCache *AC) {
  StoreInst *OnlyStore = Info.OnlyStore;
  bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
  BasicBlock *StoreBB = OnlyStore->getParent();
  int StoreIndex = -1;

  // Clear out UsingBlocks.  We will reconstruct it here if needed.
  Info.UsingBlocks.clear();

  for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
    Instruction *UserInst = cast<Instruction>(*UI++);
    if (!isa<LoadInst>(UserInst)) {
      assert(UserInst == OnlyStore && "Should only have load/stores");
      continue;
    }
    LoadInst *LI = cast<LoadInst>(UserInst);

    // Okay, if we have a load from the alloca, we want to replace it with the
    // only value stored to the alloca.  We can do this if the value is
    // dominated by the store.  If not, we use the rest of the mem2reg machinery
    // to insert the phi nodes as needed.
    if (!StoringGlobalVal) { // Non-instructions are always dominated.
      if (LI->getParent() == StoreBB) {
        // If we have a use that is in the same block as the store, compare the
        // indices of the two instructions to see which one came first.  If the
        // load came before the store, we can't handle it.
        if (StoreIndex == -1)
          StoreIndex = LBI.getInstructionIndex(OnlyStore);

        if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
          // Can't handle this load, bail out.
          Info.UsingBlocks.push_back(StoreBB);
          continue;
        }
      } else if (LI->getParent() != StoreBB &&
                 !DT.dominates(StoreBB, LI->getParent())) {
        // If the load and store are in different blocks, use BB dominance to
        // check their relationships.  If the store doesn't dom the use, bail
        // out.
        Info.UsingBlocks.push_back(LI->getParent());
        continue;
      }
    }

    // Otherwise, we *can* safely rewrite this load.
    Value *ReplVal = OnlyStore->getOperand(0);
    // If the replacement value is the load, this must occur in unreachable
    // code.
    if (ReplVal == LI)
      ReplVal = UndefValue::get(LI->getType());

    // If the load was marked as nonnull we don't want to lose
    // that information when we erase this Load. So we preserve
    // it with an assume.
    if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
        !isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT))
      addAssumeNonNull(AC, LI);

    LI->replaceAllUsesWith(ReplVal);
    LI->eraseFromParent();
    LBI.deleteValue(LI);
  }

  // Finally, after the scan, check to see if the store is all that is left.
  if (!Info.UsingBlocks.empty())
    return false; // If not, we'll have to fall back for the remainder.

  // Record debuginfo for the store and remove the declaration's
  // debuginfo.
  for (DbgVariableIntrinsic *DII : Info.DbgDeclares) {
    DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false);
    ConvertDebugDeclareToDebugValue(DII, Info.OnlyStore, DIB);
    DII->eraseFromParent();
    LBI.deleteValue(DII);
  }
  // Remove the (now dead) store and alloca.
  Info.OnlyStore->eraseFromParent();
  LBI.deleteValue(Info.OnlyStore);

  AI->eraseFromParent();
  LBI.deleteValue(AI);
  return true;
}