コード例 #1
0
ファイル: ArgumentPromotion.cpp プロジェクト: Drup/llvm
/// 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,
                             SmallPtrSetImpl<Argument*> &ArgsToPromote,
                             SmallPtrSetImpl<Argument*> &ByValArgsToTransform) {

  // Start by computing a new prototype for the function, which is the same as
  // the old function, but has modified arguments.
  FunctionType *FTy = F->getFunctionType();
  std::vector<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.
  // We need to keep the original loads for each argument and the elements
  // of the argument that are accessed.
  std::map<std::pair<Argument*, IndicesVector>, LoadInst*> OriginalLoads;

  // Attribute - 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<AttributeSet, 8> AttributesVec;
  const AttributeSet &PAL = F->getAttributes();

  // Add any return attributes.
  if (PAL.hasAttributes(AttributeSet::ReturnIndex))
    AttributesVec.push_back(AttributeSet::get(F->getContext(),
                                              PAL.getRetAttributes()));

  // 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.
      Type *AgTy = cast<PointerType>(I->getType())->getElementType();
      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());
      AttributeSet attrs = PAL.getParamAttributes(ArgIndex);
      if (attrs.hasAttributes(ArgIndex)) {
        AttrBuilder B(attrs, ArgIndex);
        AttributesVec.
          push_back(AttributeSet::get(F->getContext(), Params.size(), B));
      }
    } 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 (User *U : I->users()) {
        Instruction *UI = cast<Instruction>(U);
        assert(isa<LoadInst>(UI) || isa<GetElementPtrInst>(UI));
        IndicesVector Indices;
        Indices.reserve(UI->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 = UI->op_begin() + 1, IE = UI->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>(UI))
          OrigLoad = L;
        else
          // Take any load, we will use it only to update Alias Analysis
          OrigLoad = cast<LoadInst>(UI->user_back());
        OriginalLoads[std::make_pair(I, 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));
        assert(Params.back());
      }

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

  // Add any function attributes.
  if (PAL.hasAttributes(AttributeSet::FunctionIndex))
    AttributesVec.push_back(AttributeSet::get(FTy->getContext(),
                                              PAL.getFnAttributes()));

  Type *RetTy = FTy->getReturnType();

  // 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);

  // Patch the pointer to LLVM function in debug info descriptor.
  auto DI = FunctionDIs.find(F);
  if (DI != FunctionDIs.end()) {
    DISubprogram SP = DI->second;
    SP.replaceFunction(NF);
    // Ensure the map is updated so it can be reused on subsequent argument
    // promotions of the same function.
    FunctionDIs.erase(DI);
    FunctionDIs[NF] = SP;
  }

  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(AttributeSet::get(F->getContext(), AttributesVec));
  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<CallGraphWrapperPass>().getCallGraph();

  // 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->user_back());
    assert(CS.getCalledFunction() == F);
    Instruction *Call = CS.getInstruction();
    const AttributeSet &CallPAL = CS.getAttributes();

    // Add any return attributes.
    if (CallPAL.hasAttributes(AttributeSet::ReturnIndex))
      AttributesVec.push_back(AttributeSet::get(F->getContext(),
                                                CallPAL.getRetAttributes()));

    // 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 (CallPAL.hasAttributes(ArgIndex)) {
          AttrBuilder B(CallPAL, ArgIndex);
          AttributesVec.
            push_back(AttributeSet::get(F->getContext(), Args.size(), B));
        }
      } else if (ByValArgsToTransform.count(I)) {
        // Emit a GEP and load for each element of the struct.
        Type *AgTy = cast<PointerType>(I->getType())->getElementType();
        StructType *STy = cast<StructType>(AgTy);
        Value *Idxs[2] = {
              ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
        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,
                                                 (*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[std::make_pair(I, *SI)];
          if (!SI->empty()) {
            Ops.reserve(SI->size());
            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.
              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, 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 AA info too.
          AAMDNodes AAInfo;
          OrigLoad->getAAMetadata(AAInfo);
          newLoad->setAAMetadata(AAInfo);

          Args.push_back(newLoad);
          AA.copyValue(OrigLoad, Args.back());
        }
      }

    // Push any varargs arguments on the list.
    for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
      Args.push_back(*AI);
      if (CallPAL.hasAttributes(ArgIndex)) {
        AttrBuilder B(CallPAL, ArgIndex);
        AttributesVec.
          push_back(AttributeSet::get(F->getContext(), Args.size(), B));
      }
    }

    // Add any function attributes.
    if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
      AttributesVec.push_back(AttributeSet::get(Call->getContext(),
                                                CallPAL.getFnAttributes()));

    Instruction *New;
    if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
      New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
                               Args, "", Call);
      cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
      cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(),
                                                            AttributesVec));
    } else {
      New = CallInst::Create(NF, Args, "", Call);
      cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
      cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(),
                                                          AttributesVec));
      if (cast<CallInst>(Call)->isTailCall())
        cast<CallInst>(New)->setTailCall();
    }
    New->setDebugLoc(Call->getDebugLoc());
    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.
      Type *AgTy = cast<PointerType>(I->getType())->getElementType();
      Value *TheAlloca = new AllocaInst(AgTy, nullptr, "", InsertPt);
      StructType *STy = cast<StructType>(AgTy);
      Value *Idxs[2] = {
            ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };

      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,
                                    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);

      // If the alloca is used in a call, we must clear the tail flag since
      // the callee now uses an alloca from the caller.
      for (User *U : TheAlloca->users()) {
        CallInst *Call = dyn_cast<CallInst>(U);
        if (!Call)
          continue;
        Call->setTailCall(false);
      }
      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->user_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->user_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->user_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.
    std::advance(I2, ArgIndices.size());
  }

  // 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;
}
コード例 #2
0
ファイル: ArgumentPromotion.cpp プロジェクト: mkurdej/llvm
/// 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.
static Function *
doPromotion(Function *F, SmallPtrSetImpl<Argument *> &ArgsToPromote,
            SmallPtrSetImpl<Argument *> &ByValArgsToTransform,
            Optional<function_ref<void(CallSite OldCS, CallSite NewCS)>>
                ReplaceCallSite) {
  // Start by computing a new prototype for the function, which is the same as
  // the old function, but has modified arguments.
  FunctionType *FTy = F->getFunctionType();
  std::vector<Type *> Params;

  using ScalarizeTable = std::set<std::pair<Type *, IndicesVector>>;

  // 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.
  // We need to keep the original loads for each argument and the elements
  // of the argument that are accessed.
  std::map<std::pair<Argument *, IndicesVector>, LoadInst *> OriginalLoads;

  // Attribute - 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<AttributeSet, 8> ArgAttrVec;
  AttributeList PAL = F->getAttributes();

  // First, determine the new argument list
  unsigned ArgNo = 0;
  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
       ++I, ++ArgNo) {
    if (ByValArgsToTransform.count(&*I)) {
      // Simple byval argument? Just add all the struct element types.
      Type *AgTy = cast<PointerType>(I->getType())->getElementType();
      StructType *STy = cast<StructType>(AgTy);
      Params.insert(Params.end(), STy->element_begin(), STy->element_end());
      ArgAttrVec.insert(ArgAttrVec.end(), STy->getNumElements(),
                        AttributeSet());
      ++NumByValArgsPromoted;
    } else if (!ArgsToPromote.count(&*I)) {
      // Unchanged argument
      Params.push_back(I->getType());
      ArgAttrVec.push_back(PAL.getParamAttributes(ArgNo));
    } else if (I->use_empty()) {
      // Dead argument (which are always marked as promotable)
      ++NumArgumentsDead;

      // There may be remaining metadata uses of the argument for things like
      // llvm.dbg.value. Replace them with undef.
      I->replaceAllUsesWith(UndefValue::get(I->getType()));
    } 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 (User *U : I->users()) {
        Instruction *UI = cast<Instruction>(U);
        Type *SrcTy;
        if (LoadInst *L = dyn_cast<LoadInst>(UI))
          SrcTy = L->getType();
        else
          SrcTy = cast<GetElementPtrInst>(UI)->getSourceElementType();
        IndicesVector Indices;
        Indices.reserve(UI->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 = UI->op_begin() + 1, IE = UI->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(std::make_pair(SrcTy, Indices));
        LoadInst *OrigLoad;
        if (LoadInst *L = dyn_cast<LoadInst>(UI))
          OrigLoad = L;
        else
          // Take any load, we will use it only to update Alias Analysis
          OrigLoad = cast<LoadInst>(UI->user_back());
        OriginalLoads[std::make_pair(&*I, Indices)] = OrigLoad;
      }

      // Add a parameter to the function for each element passed in.
      for (const auto &ArgIndex : ArgIndices) {
        // not allowed to dereference ->begin() if size() is 0
        Params.push_back(GetElementPtrInst::getIndexedType(
            cast<PointerType>(I->getType()->getScalarType())->getElementType(),
            ArgIndex.second));
        ArgAttrVec.push_back(AttributeSet());
        assert(Params.back());
      }

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

  Type *RetTy = FTy->getReturnType();

  // 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);

  // Patch the pointer to LLVM function in debug info descriptor.
  NF->setSubprogram(F->getSubprogram());
  F->setSubprogram(nullptr);

  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(AttributeList::get(F->getContext(), PAL.getFnAttributes(),
                                       PAL.getRetAttributes(), ArgAttrVec));
  ArgAttrVec.clear();

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

  // 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->user_back());
    assert(CS.getCalledFunction() == F);
    Instruction *Call = CS.getInstruction();
    const AttributeList &CallPAL = CS.getAttributes();

    // Loop over the operands, inserting GEP and loads in the caller as
    // appropriate.
    CallSite::arg_iterator AI = CS.arg_begin();
    ArgNo = 0;
    for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
         ++I, ++AI, ++ArgNo)
      if (!ArgsToPromote.count(&*I) && !ByValArgsToTransform.count(&*I)) {
        Args.push_back(*AI); // Unmodified argument
        ArgAttrVec.push_back(CallPAL.getParamAttributes(ArgNo));
      } else if (ByValArgsToTransform.count(&*I)) {
        // Emit a GEP and load for each element of the struct.
        Type *AgTy = cast<PointerType>(I->getType())->getElementType();
        StructType *STy = cast<StructType>(AgTy);
        Value *Idxs[2] = {
            ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr};
        for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
          Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
          Value *Idx = GetElementPtrInst::Create(
              STy, *AI, Idxs, (*AI)->getName() + "." + Twine(i), Call);
          // TODO: Tell AA about the new values?
          Args.push_back(new LoadInst(Idx, Idx->getName() + ".val", Call));
          ArgAttrVec.push_back(AttributeSet());
        }
      } 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 (const auto &ArgIndex : ArgIndices) {
          Value *V = *AI;
          LoadInst *OrigLoad =
              OriginalLoads[std::make_pair(&*I, ArgIndex.second)];
          if (!ArgIndex.second.empty()) {
            Ops.reserve(ArgIndex.second.size());
            Type *ElTy = V->getType();
            for (auto II : ArgIndex.second) {
              // Use i32 to index structs, and i64 for others (pointers/arrays).
              // This satisfies GEP constraints.
              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.
              if (auto *ElPTy = dyn_cast<PointerType>(ElTy))
                ElTy = ElPTy->getElementType();
              else
                ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(II);
            }
            // And create a GEP to extract those indices.
            V = GetElementPtrInst::Create(ArgIndex.first, V, Ops,
                                          V->getName() + ".idx", Call);
            Ops.clear();
          }
          // 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 AA info too.
          AAMDNodes AAInfo;
          OrigLoad->getAAMetadata(AAInfo);
          newLoad->setAAMetadata(AAInfo);

          Args.push_back(newLoad);
          ArgAttrVec.push_back(AttributeSet());
        }
      }

    // Push any varargs arguments on the list.
    for (; AI != CS.arg_end(); ++AI, ++ArgNo) {
      Args.push_back(*AI);
      ArgAttrVec.push_back(CallPAL.getParamAttributes(ArgNo));
    }

    SmallVector<OperandBundleDef, 1> OpBundles;
    CS.getOperandBundlesAsDefs(OpBundles);

    CallSite NewCS;
    if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
      NewCS = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
                                 Args, OpBundles, "", Call);
    } else {
      auto *NewCall = CallInst::Create(NF, Args, OpBundles, "", Call);
      NewCall->setTailCallKind(cast<CallInst>(Call)->getTailCallKind());
      NewCS = NewCall;
    }
    NewCS.setCallingConv(CS.getCallingConv());
    NewCS.setAttributes(
        AttributeList::get(F->getContext(), CallPAL.getFnAttributes(),
                           CallPAL.getRetAttributes(), ArgAttrVec));
    NewCS->setDebugLoc(Call->getDebugLoc());
    uint64_t W;
    if (Call->extractProfTotalWeight(W))
      NewCS->setProfWeight(W);
    Args.clear();
    ArgAttrVec.clear();

    // Update the callgraph to know that the callsite has been transformed.
    if (ReplaceCallSite)
      (*ReplaceCallSite)(CS, NewCS);

    if (!Call->use_empty()) {
      Call->replaceAllUsesWith(NewCS.getInstruction());
      NewCS->takeName(Call);
    }

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

  const DataLayout &DL = F->getParent()->getDataLayout();

  // 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);
      ++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()->front();

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

      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
        Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
        Value *Idx = GetElementPtrInst::Create(
            AgTy, TheAlloca, Idxs, 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);

      // If the alloca is used in a call, we must clear the tail flag since
      // the callee now uses an alloca from the caller.
      for (User *U : TheAlloca->users()) {
        CallInst *Call = dyn_cast<CallInst>(U);
        if (!Call)
          continue;
        Call->setTailCall(false);
      }
      continue;
    }

    if (I->use_empty())
      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->user_back())) {
        assert(ArgIndices.begin()->second.empty() &&
               "Load element should sort to front!");
        I2->setName(I->getName() + ".val");
        LI->replaceAllUsesWith(&*I2);
        LI->eraseFromParent();
        DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
                     << "' in function '" << F->getName() << "'\n");
      } else {
        GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->user_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->second != 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->user_back());
          L->replaceAllUsesWith(&*TheArg);
          L->eraseFromParent();
        }
        GEP->eraseFromParent();
      }
    }

    // Increment I2 past all of the arguments added for this promoted pointer.
    std::advance(I2, ArgIndices.size());
  }

  return NF;
}