bool AMDGPUCodeGenPrepare::canWidenScalarExtLoad(LoadInst &I) const {
Type *Ty = I.getType();
const DataLayout &DL = Mod->getDataLayout();
int TySize = DL.getTypeSizeInBits(Ty);
unsigned Align = I.getAlignment() ?
I.getAlignment() : DL.getABITypeAlignment(Ty);
return I.isSimple() && TySize < 32 && Align >= 4 && DA->isUniform(&I);
}
/// %res = load {atomic|volatile} T* %ptr memory_order, align sizeof(T)
/// becomes:
/// %res = call T @llvm.nacl.atomic.load.i<size>(%ptr, memory_order)
void AtomicVisitor::visitLoadInst(LoadInst &I) {
return; // XXX EMSCRIPTEN
if (I.isSimple())
return;
PointerHelper<LoadInst> PH(*this, I);
const NaCl::AtomicIntrinsics::AtomicIntrinsic *Intrinsic =
findAtomicIntrinsic(I, Intrinsic::nacl_atomic_load, PH.PET);
checkAlignment(I, I.getAlignment(), PH.BitSize / CHAR_BIT);
Value *Args[] = {PH.P, freezeMemoryOrder(I, I.getOrdering())};
replaceInstructionWithIntrinsicCall(I, Intrinsic, PH.OriginalPET, PH.PET,
Args);
}
void X86InterleavedAccessGroup::decompose(
Instruction *VecInst, unsigned NumSubVectors, VectorType *SubVecTy,
SmallVectorImpl<Instruction *> &DecomposedVectors) {
assert((isa<LoadInst>(VecInst) || isa<ShuffleVectorInst>(VecInst)) &&
"Expected Load or Shuffle");
Type *VecTy = VecInst->getType();
(void)VecTy;
assert(VecTy->isVectorTy() &&
DL.getTypeSizeInBits(VecTy) >=
DL.getTypeSizeInBits(SubVecTy) * NumSubVectors &&
"Invalid Inst-size!!!");
if (auto *SVI = dyn_cast<ShuffleVectorInst>(VecInst)) {
Value *Op0 = SVI->getOperand(0);
Value *Op1 = SVI->getOperand(1);
// Generate N(= NumSubVectors) shuffles of T(= SubVecTy) type.
for (unsigned i = 0; i < NumSubVectors; ++i)
DecomposedVectors.push_back(
cast<ShuffleVectorInst>(Builder.CreateShuffleVector(
Op0, Op1,
createSequentialMask(Builder, Indices[i],
SubVecTy->getVectorNumElements(), 0))));
return;
}
// Decompose the load instruction.
LoadInst *LI = cast<LoadInst>(VecInst);
Type *VecBasePtrTy = SubVecTy->getPointerTo(LI->getPointerAddressSpace());
Value *VecBasePtr;
unsigned int NumLoads = NumSubVectors;
// In the case of stride 3 with a vector of 32 elements load the information
// in the following way:
// [0,1...,VF/2-1,VF/2+VF,VF/2+VF+1,...,2VF-1]
if (DL.getTypeSizeInBits(VecTy) == 768) {
Type *VecTran =
VectorType::get(Type::getInt8Ty(LI->getContext()), 16)->getPointerTo();
VecBasePtr = Builder.CreateBitCast(LI->getPointerOperand(), VecTran);
NumLoads = NumSubVectors * 2;
} else
VecBasePtr = Builder.CreateBitCast(LI->getPointerOperand(), VecBasePtrTy);
// Generate N loads of T type.
for (unsigned i = 0; i < NumLoads; i++) {
// TODO: Support inbounds GEP.
Value *NewBasePtr = Builder.CreateGEP(VecBasePtr, Builder.getInt32(i));
Instruction *NewLoad =
Builder.CreateAlignedLoad(NewBasePtr, LI->getAlignment());
DecomposedVectors.push_back(NewLoad);
}
}
/// \brief Try to aggregate loads from a sorted list of loads to be combined.
///
/// It is guaranteed that no writes occur between any of the loads. All loads
/// have the same base pointer. There are at least two loads.
bool LoadCombine::aggregateLoads(SmallVectorImpl<LoadPOPPair> &Loads) {
assert(Loads.size() >= 2 && "Insufficient loads!");
LoadInst *BaseLoad = nullptr;
SmallVector<LoadPOPPair, 8> AggregateLoads;
bool Combined = false;
bool ValidPrevOffset = false;
APInt PrevOffset;
uint64_t PrevSize = 0;
for (auto &L : Loads) {
if (ValidPrevOffset == false) {
BaseLoad = L.Load;
PrevOffset = L.POP.Offset;
PrevSize = L.Load->getModule()->getDataLayout().getTypeStoreSize(
L.Load->getType());
AggregateLoads.push_back(L);
ValidPrevOffset = true;
continue;
}
if (L.Load->getAlignment() > BaseLoad->getAlignment())
continue;
APInt PrevEnd = PrevOffset + PrevSize;
if (L.POP.Offset.sgt(PrevEnd)) {
// No other load will be combinable
if (combineLoads(AggregateLoads))
Combined = true;
AggregateLoads.clear();
ValidPrevOffset = false;
continue;
}
if (L.POP.Offset != PrevEnd)
// This load is offset less than the size of the last load.
// FIXME: We may want to handle this case.
continue;
PrevOffset = L.POP.Offset;
PrevSize = L.Load->getModule()->getDataLayout().getTypeStoreSize(
L.Load->getType());
AggregateLoads.push_back(L);
}
if (combineLoads(AggregateLoads))
Combined = true;
return Combined;
}
bool Scalarizer::visitLoadInst(LoadInst &LI) {
if (!ScalarizeLoadStore)
return false;
if (!LI.isSimple())
return false;
VectorLayout Layout;
if (!getVectorLayout(LI.getType(), LI.getAlignment(), Layout))
return false;
unsigned NumElems = Layout.VecTy->getNumElements();
IRBuilder<> Builder(LI.getParent(), &LI);
Scatterer Ptr = scatter(&LI, LI.getPointerOperand());
ValueVector Res;
Res.resize(NumElems);
for (unsigned I = 0; I < NumElems; ++I)
Res[I] = Builder.CreateAlignedLoad(Ptr[I], Layout.getElemAlign(I),
LI.getName() + ".i" + Twine(I));
gather(&LI, Res);
return true;
}
void Lint::visitLoadInst(LoadInst &I) {
visitMemoryReference(I, I.getPointerOperand(),
DL->getTypeStoreSize(I.getType()), I.getAlignment(),
I.getType(), MemRef::Read);
}
unsigned CostModelAnalysis::getInstructionCost(Instruction *I) const {
if (!VTTI)
return -1;
switch (I->getOpcode()) {
case Instruction::Ret:
case Instruction::PHI:
case Instruction::Br: {
return VTTI->getCFInstrCost(I->getOpcode());
}
case Instruction::Add:
case Instruction::FAdd:
case Instruction::Sub:
case Instruction::FSub:
case Instruction::Mul:
case Instruction::FMul:
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::FDiv:
case Instruction::URem:
case Instruction::SRem:
case Instruction::FRem:
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor: {
return VTTI->getArithmeticInstrCost(I->getOpcode(), I->getType());
}
case Instruction::Select: {
SelectInst *SI = cast<SelectInst>(I);
Type *CondTy = SI->getCondition()->getType();
return VTTI->getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy);
}
case Instruction::ICmp:
case Instruction::FCmp: {
Type *ValTy = I->getOperand(0)->getType();
return VTTI->getCmpSelInstrCost(I->getOpcode(), ValTy);
}
case Instruction::Store: {
StoreInst *SI = cast<StoreInst>(I);
Type *ValTy = SI->getValueOperand()->getType();
return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
SI->getAlignment(),
SI->getPointerAddressSpace());
}
case Instruction::Load: {
LoadInst *LI = cast<LoadInst>(I);
return VTTI->getMemoryOpCost(I->getOpcode(), I->getType(),
LI->getAlignment(),
LI->getPointerAddressSpace());
}
case Instruction::ZExt:
case Instruction::SExt:
case Instruction::FPToUI:
case Instruction::FPToSI:
case Instruction::FPExt:
case Instruction::PtrToInt:
case Instruction::IntToPtr:
case Instruction::SIToFP:
case Instruction::UIToFP:
case Instruction::Trunc:
case Instruction::FPTrunc:
case Instruction::BitCast: {
Type *SrcTy = I->getOperand(0)->getType();
return VTTI->getCastInstrCost(I->getOpcode(), I->getType(), SrcTy);
}
case Instruction::ExtractElement: {
ExtractElementInst * EEI = cast<ExtractElementInst>(I);
ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
unsigned Idx = -1;
if (CI)
Idx = CI->getZExtValue();
return VTTI->getVectorInstrCost(I->getOpcode(),
EEI->getOperand(0)->getType(), Idx);
}
case Instruction::InsertElement: {
InsertElementInst * IE = cast<InsertElementInst>(I);
ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2));
unsigned Idx = -1;
if (CI)
Idx = CI->getZExtValue();
return VTTI->getVectorInstrCost(I->getOpcode(),
IE->getType(), Idx);
}
default:
// We don't have any information on this instruction.
return -1;
}
}
/// tryAggregating - When scanning forward over instructions, we look for
/// other loads or stores that could be aggregated with this one.
/// Returns the last instruction added (if one was added) since we might have
/// removed some loads or stores and that might invalidate an iterator.
Instruction *AggregateGlobalOpsOpt::tryAggregating(Instruction *StartInst, Value *StartPtr,
bool DebugThis) {
if (TD == 0) return 0;
Module* M = StartInst->getParent()->getParent()->getParent();
LLVMContext& Context = StartInst->getContext();
Type* int8Ty = Type::getInt8Ty(Context);
Type* sizeTy = Type::getInt64Ty(Context);
Type* globalInt8PtrTy = int8Ty->getPointerTo(globalSpace);
bool isLoad = isa<LoadInst>(StartInst);
bool isStore = isa<StoreInst>(StartInst);
Instruction *lastAddedInsn = NULL;
Instruction *LastLoadOrStore = NULL;
SmallVector<Instruction*, 8> toRemove;
// Okay, so we now have a single global load/store. Scan to find
// all subsequent stores of the same value to offset from the same pointer.
// Join these together into ranges, so we can decide whether contiguous blocks
// are stored.
MemOpRanges Ranges(*TD);
// Put the first store in since we want to preserve the order.
Ranges.addInst(0, StartInst);
BasicBlock::iterator BI = StartInst;
for (++BI; !isa<TerminatorInst>(BI); ++BI) {
if( isGlobalLoadOrStore(BI, globalSpace, isLoad, isStore) ) {
// OK!
} else {
// If the instruction is readnone, ignore it, otherwise bail out. We
// don't even allow readonly here because we don't want something like:
// A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
if (BI->mayWriteToMemory())
break;
if (isStore && BI->mayReadFromMemory())
break;
continue;
}
if ( isStore && isa<StoreInst>(BI) ) {
StoreInst *NextStore = cast<StoreInst>(BI);
// If this is a store, see if we can merge it in.
if (!NextStore->isSimple()) break;
// Check to see if this store is to a constant offset from the start ptr.
int64_t Offset;
if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD))
break;
Ranges.addStore(Offset, NextStore);
LastLoadOrStore = NextStore;
} else {
LoadInst *NextLoad = cast<LoadInst>(BI);
if (!NextLoad->isSimple()) break;
// Check to see if this load is to a constant offset from the start ptr.
int64_t Offset;
if (!IsPointerOffset(StartPtr, NextLoad->getPointerOperand(), Offset, *TD))
break;
Ranges.addLoad(Offset, NextLoad);
LastLoadOrStore = NextLoad;
}
}
// If we have no ranges, then we just had a single store with nothing that
// could be merged in. This is a very common case of course.
if (!Ranges.moreThanOneOp())
return 0;
// Divide the instructions between StartInst and LastLoadOrStore into
// addressing, memops, and uses of memops (uses of loads)
reorderAddressingMemopsUses(StartInst, LastLoadOrStore, DebugThis);
Instruction* insertBefore = StartInst;
IRBuilder<> builder(insertBefore);
// Now that we have full information about ranges, loop over the ranges and
// emit memcpy's for anything big enough to be worthwhile.
for (MemOpRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
I != E; ++I) {
const MemOpRange &Range = *I;
Value* oldBaseI = NULL;
Value* newBaseI = NULL;
if (Range.TheStores.size() == 1) continue; // Don't bother if there's only one thing...
builder.SetInsertPoint(insertBefore);
// Otherwise, we do want to transform this! Create a new memcpy.
// Get the starting pointer of the block.
StartPtr = Range.StartPtr;
if( DebugThis ) {
errs() << "base is:";
StartPtr->dump();
}
// Determine alignment
unsigned Alignment = Range.Alignment;
if (Alignment == 0) {
Type *EltType =
cast<PointerType>(StartPtr->getType())->getElementType();
Alignment = TD->getABITypeAlignment(EltType);
}
Instruction *alloc = NULL;
Value *globalPtr = NULL;
// create temporary alloca space to communicate to/from.
alloc = makeAlloca(int8Ty, "agg.tmp", insertBefore,
Range.End-Range.Start, Alignment);
// Generate the old and new base pointers before we output
// anything else.
{
Type* iPtrTy = TD->getIntPtrType(alloc->getType());
Type* iNewBaseTy = TD->getIntPtrType(alloc->getType());
oldBaseI = builder.CreatePtrToInt(StartPtr, iPtrTy, "agg.tmp.oldb.i");
newBaseI = builder.CreatePtrToInt(alloc, iNewBaseTy, "agg.tmp.newb.i");
}
// If storing, do the stores we had into our alloca'd region.
if( isStore ) {
for (SmallVector<Instruction*, 16>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI) {
StoreInst* oldStore = cast<StoreInst>(*SI);
if( DebugThis ) {
errs() << "have store in range:";
oldStore->dump();
}
Value* ptrToAlloc = rebasePointer(oldStore->getPointerOperand(),
StartPtr, alloc, "agg.tmp",
&builder, *TD, oldBaseI, newBaseI);
// Old load must not be volatile or atomic... or we shouldn't have put
// it in ranges
assert(!(oldStore->isVolatile() || oldStore->isAtomic()));
StoreInst* newStore =
builder.CreateStore(oldStore->getValueOperand(), ptrToAlloc);
newStore->setAlignment(oldStore->getAlignment());
newStore->takeName(oldStore);
}
}
// cast the pointer that was load/stored to i8 if necessary.
if( StartPtr->getType()->getPointerElementType() == int8Ty ) {
globalPtr = StartPtr;
} else {
globalPtr = builder.CreatePointerCast(StartPtr, globalInt8PtrTy, "agg.cast");
}
// Get a Constant* for the length.
Constant* len = ConstantInt::get(sizeTy, Range.End-Range.Start, false);
// Now add the memcpy instruction
unsigned addrSpaceDst,addrSpaceSrc;
addrSpaceDst = addrSpaceSrc = 0;
if( isStore ) addrSpaceDst = globalSpace;
if( isLoad ) addrSpaceSrc = globalSpace;
Type *types[3];
types[0] = PointerType::get(int8Ty, addrSpaceDst);
types[1] = PointerType::get(int8Ty, addrSpaceSrc);
types[2] = sizeTy;
Function *func = Intrinsic::getDeclaration(M, Intrinsic::memcpy, types);
Value* args[5]; // dst src len alignment isvolatile
if( isStore ) {
// it's a store (ie put)
args[0] = globalPtr;
args[1] = alloc;
} else {
// it's a load (ie get)
args[0] = alloc;
args[1] = globalPtr;
}
args[2] = len;
// alignment
args[3] = ConstantInt::get(Type::getInt32Ty(Context), 0, false);
// isvolatile
args[4] = ConstantInt::get(Type::getInt1Ty(Context), 0, false);
Instruction* aMemCpy = builder.CreateCall(func, args);
/*
DEBUG(dbgs() << "Replace ops:\n";
for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
dbgs() << *Range.TheStores[i] << '\n';
dbgs() << "With: " << *AMemSet << '\n');
*/
if (!Range.TheStores.empty())
aMemCpy->setDebugLoc(Range.TheStores[0]->getDebugLoc());
lastAddedInsn = aMemCpy;
// If loading, load from the memcpy'd region
if( isLoad ) {
for (SmallVector<Instruction*, 16>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI) {
LoadInst* oldLoad = cast<LoadInst>(*SI);
if( DebugThis ) {
errs() << "have load in range:";
oldLoad->dump();
}
Value* ptrToAlloc = rebasePointer(oldLoad->getPointerOperand(),
StartPtr, alloc, "agg.tmp",
&builder, *TD, oldBaseI, newBaseI);
// Old load must not be volatile or atomic... or we shouldn't have put
// it in ranges
assert(!(oldLoad->isVolatile() || oldLoad->isAtomic()));
LoadInst* newLoad = builder.CreateLoad(ptrToAlloc);
newLoad->setAlignment(oldLoad->getAlignment());
oldLoad->replaceAllUsesWith(newLoad);
newLoad->takeName(oldLoad);
lastAddedInsn = newLoad;
}
}
// Save old loads/stores for removal
for (SmallVector<Instruction*, 16>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI) {
Instruction* insn = *SI;
toRemove.push_back(insn);
}
}
// Zap all the old loads/stores
for (SmallVector<Instruction*, 16>::const_iterator
SI = toRemove.begin(),
SE = toRemove.end(); SI != SE; ++SI) {
(*SI)->eraseFromParent();
}
return lastAddedInsn;
}
/// 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 = CallSite::get(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());
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, transfering uses of the old arguments over to
// the new arguments, also transfering 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;
}
Value *BoUpSLP::vectorizeTree(ValueList &VL, int VF) {
Type *ScalarTy = VL[0]->getType();
if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
ScalarTy = SI->getValueOperand()->getType();
VectorType *VecTy = VectorType::get(ScalarTy, VF);
// Check if all of the operands are constants or identical.
bool AllConst = true;
bool AllSameScalar = true;
for (unsigned i = 0, e = VF; i < e; ++i) {
AllConst &= !!dyn_cast<Constant>(VL[i]);
AllSameScalar &= (VL[0] == VL[i]);
// Must have a single use.
Instruction *I = dyn_cast<Instruction>(VL[i]);
if (I && (I->getNumUses() > 1 || I->getParent() != BB))
return Scalarize(VL, VecTy);
}
// Is this a simple vector constant.
if (AllConst || AllSameScalar) return Scalarize(VL, VecTy);
// Scalarize unknown structures.
Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
if (!VL0) return Scalarize(VL, VecTy);
unsigned Opcode = VL0->getOpcode();
for (unsigned i = 0, e = VF; i < e; ++i) {
Instruction *I = dyn_cast<Instruction>(VL[i]);
// If not all of the instructions are identical then we have to scalarize.
if (!I || Opcode != I->getOpcode()) return Scalarize(VL, VecTy);
}
switch (Opcode) {
case Instruction::Add:
case Instruction::FAdd:
case Instruction::Sub:
case Instruction::FSub:
case Instruction::Mul:
case Instruction::FMul:
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::FDiv:
case Instruction::URem:
case Instruction::SRem:
case Instruction::FRem:
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor: {
ValueList LHSVL, RHSVL;
for (int i = 0; i < VF; ++i) {
RHSVL.push_back(cast<Instruction>(VL[i])->getOperand(0));
LHSVL.push_back(cast<Instruction>(VL[i])->getOperand(1));
}
Value *RHS = vectorizeTree(RHSVL, VF);
Value *LHS = vectorizeTree(LHSVL, VF);
IRBuilder<> Builder(GetLastInstr(VL, VF));
BinaryOperator *BinOp = dyn_cast<BinaryOperator>(VL0);
return Builder.CreateBinOp(BinOp->getOpcode(), RHS,LHS);
}
case Instruction::Load: {
LoadInst *LI = dyn_cast<LoadInst>(VL0);
unsigned Alignment = LI->getAlignment();
// Check if all of the loads are consecutive.
for (unsigned i = 1, e = VF; i < e; ++i)
if (!isConsecutiveAccess(VL[i-1], VL[i]))
return Scalarize(VL, VecTy);
IRBuilder<> Builder(GetLastInstr(VL, VF));
Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
VecTy->getPointerTo());
LI = Builder.CreateLoad(VecPtr);
LI->setAlignment(Alignment);
return LI;
}
case Instruction::Store: {
StoreInst *SI = dyn_cast<StoreInst>(VL0);
unsigned Alignment = SI->getAlignment();
ValueList ValueOp;
for (int i = 0; i < VF; ++i)
ValueOp.push_back(cast<StoreInst>(VL[i])->getValueOperand());
Value *VecValue = vectorizeTree(ValueOp, VF);
IRBuilder<> Builder(GetLastInstr(VL, VF));
Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
VecTy->getPointerTo());
Builder.CreateStore(VecValue, VecPtr)->setAlignment(Alignment);
for (int i = 0; i < VF; ++i)
cast<Instruction>(VL[i])->eraseFromParent();
return 0;
}
default:
return Scalarize(VL, VecTy);
}
}
void Lint::visitLoadInst(LoadInst &I) {
visitMemoryReference(I, I.getPointerOperand(), I.getAlignment(), I.getType());
}
/// 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;
}