// Convert the given function to use normalized argument/return types.
static bool ConvertFunction(Function *Func) {
FunctionType *FTy = Func->getFunctionType();
FunctionType *NFTy = NormalizeFunctionType(FTy);
if (NFTy == FTy)
return false; // No change needed.
Function *NewFunc = RecreateFunction(Func, NFTy);
// Move the arguments across to the new function.
for (Function::arg_iterator Arg = Func->arg_begin(), E = Func->arg_end(),
NewArg = NewFunc->arg_begin();
Arg != E; ++Arg, ++NewArg) {
NewArg->takeName(Arg);
if (Arg->getType() == NewArg->getType()) {
Arg->replaceAllUsesWith(NewArg);
} else {
Instruction *Trunc = new TruncInst(
NewArg, Arg->getType(), NewArg->getName() + ".arg_trunc",
NewFunc->getEntryBlock().getFirstInsertionPt());
Arg->replaceAllUsesWith(Trunc);
}
}
if (FTy->getReturnType() != NFTy->getReturnType()) {
// Fix up return instructions.
Instruction::CastOps CastType =
Func->getAttributes().hasAttribute(0, Attribute::SExt) ?
Instruction::SExt : Instruction::ZExt;
for (Function::iterator BB = NewFunc->begin(), E = NewFunc->end();
BB != E;
++BB) {
for (BasicBlock::iterator Iter = BB->begin(), E = BB->end();
Iter != E; ) {
Instruction *Inst = Iter++;
if (ReturnInst *Ret = dyn_cast<ReturnInst>(Inst)) {
Value *Ext = CopyDebug(
CastInst::Create(CastType, Ret->getReturnValue(),
NFTy->getReturnType(),
Ret->getReturnValue()->getName() + ".ret_ext",
Ret),
Ret);
CopyDebug(ReturnInst::Create(Ret->getContext(), Ext, Ret), Ret);
Ret->eraseFromParent();
}
}
}
}
Func->eraseFromParent();
return true;
}
/// cloneFunctionBody - Create a new function based on F and
/// insert it into module. Remove first argument. Use STy as
/// the return type for new function.
Function *SRETPromotion::cloneFunctionBody(Function *F,
const StructType *STy) {
const FunctionType *FTy = F->getFunctionType();
std::vector<const Type*> Params;
// Attributes - Keep track of the parameter attributes for the arguments.
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));
// Skip first argument.
Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
++I;
// 0th parameter attribute is reserved for return type.
// 1th parameter attribute is for first 1st sret argument.
unsigned ParamIndex = 2;
while (I != E) {
Params.push_back(I->getType());
if (Attributes Attrs = PAL.getParamAttributes(ParamIndex))
AttributesVec.push_back(AttributeWithIndex::get(ParamIndex - 1, Attrs));
++I;
++ParamIndex;
}
// Add any fn attributes.
if (Attributes attrs = PAL.getFnAttributes())
AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
FunctionType *NFTy = FunctionType::get(STy, Params, FTy->isVarArg());
Function *NF = Function::Create(NFTy, F->getLinkage());
NF->takeName(F);
NF->copyAttributesFrom(F);
NF->setAttributes(AttrListPtr::get(AttributesVec.begin(), AttributesVec.end()));
F->getParent()->getFunctionList().insert(F, NF);
NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
// Replace arguments
I = F->arg_begin();
E = F->arg_end();
Function::arg_iterator NI = NF->arg_begin();
++I;
while (I != E) {
I->replaceAllUsesWith(NI);
NI->takeName(I);
++I;
++NI;
}
return NF;
}
/// lowerIncomingArguments - To avoid having to handle incoming arguments
/// specially, we lower each arg to a copy instruction in the entry block. This
/// ensures that the argument value itself cannot be live out of the entry
/// block.
void SjLjEHPass::lowerIncomingArguments(Function &F) {
BasicBlock::iterator AfterAllocaInsPt = F.begin()->begin();
while (isa<AllocaInst>(AfterAllocaInsPt) &&
isa<ConstantInt>(cast<AllocaInst>(AfterAllocaInsPt)->getArraySize()))
++AfterAllocaInsPt;
for (Function::arg_iterator
AI = F.arg_begin(), AE = F.arg_end(); AI != AE; ++AI) {
Type *Ty = AI->getType();
// Aggregate types can't be cast, but are legal argument types, so we have
// to handle them differently. We use an extract/insert pair as a
// lightweight method to achieve the same goal.
if (isa<StructType>(Ty) || isa<ArrayType>(Ty) || isa<VectorType>(Ty)) {
Instruction *EI = ExtractValueInst::Create(AI, 0, "", AfterAllocaInsPt);
Instruction *NI = InsertValueInst::Create(AI, EI, 0);
NI->insertAfter(EI);
AI->replaceAllUsesWith(NI);
// Set the operand of the instructions back to the AllocaInst.
EI->setOperand(0, AI);
NI->setOperand(0, AI);
} else {
// This is always a no-op cast because we're casting AI to AI->getType()
// so src and destination types are identical. BitCast is the only
// possibility.
CastInst *NC =
new BitCastInst(AI, AI->getType(), AI->getName() + ".tmp",
AfterAllocaInsPt);
AI->replaceAllUsesWith(NC);
// Set the operand of the cast instruction back to the AllocaInst.
// Normally it's forbidden to replace a CastInst's operand because it
// could cause the opcode to reflect an illegal conversion. However, we're
// replacing it here with the same value it was constructed with. We do
// this because the above replaceAllUsesWith() clobbered the operand, but
// we want this one to remain.
NC->setOperand(0, AI);
}
}
}
/// lowerIncomingArguments - To avoid having to handle incoming arguments
/// specially, we lower each arg to a copy instruction in the entry block. This
/// ensures that the argument value itself cannot be live out of the entry
/// block.
void SjLjEHPrepare::lowerIncomingArguments(Function &F) {
BasicBlock::iterator AfterAllocaInsPt = F.begin()->begin();
while (isa<AllocaInst>(AfterAllocaInsPt) &&
isa<ConstantInt>(cast<AllocaInst>(AfterAllocaInsPt)->getArraySize()))
++AfterAllocaInsPt;
for (Function::arg_iterator AI = F.arg_begin(), AE = F.arg_end(); AI != AE;
++AI) {
Type *Ty = AI->getType();
if (isa<StructType>(Ty) || isa<ArrayType>(Ty)) {
// Aggregate types can't be cast, but are legal argument types,
// so we have to handle them differently. We use
// select i8 true, %arg, undef to achieve the same goal
Value *TrueValue = ConstantInt::getTrue(F.getContext());
Value *UndefValue = UndefValue::get(Ty);
Instruction *SI = SelectInst::Create(TrueValue, AI, UndefValue,
AI->getName() + ".tmp",
AfterAllocaInsPt);
AI->replaceAllUsesWith(SI);
SI->setOperand(1, AI);
} else {
// This is always a no-op cast because we're casting AI to AI->getType()
// so src and destination types are identical. BitCast is the only
// possibility.
CastInst *NC = new BitCastInst(AI, AI->getType(), AI->getName() + ".tmp",
AfterAllocaInsPt);
AI->replaceAllUsesWith(NC);
// Set the operand of the cast instruction back to the AllocaInst.
// Normally it's forbidden to replace a CastInst's operand because it
// could cause the opcode to reflect an illegal conversion. However, we're
// replacing it here with the same value it was constructed with. We do
// this because the above replaceAllUsesWith() clobbered the operand, but
// we want this one to remain.
NC->setOperand(0, AI);
}
}
}
/// lowerIncomingArguments - To avoid having to handle incoming arguments
/// specially, we lower each arg to a copy instruction in the entry block. This
/// ensures that the argument value itself cannot be live out of the entry
/// block.
void SjLjEHPrepare::lowerIncomingArguments(Function &F) {
BasicBlock::iterator AfterAllocaInsPt = F.begin()->begin();
while (isa<AllocaInst>(AfterAllocaInsPt) &&
isa<ConstantInt>(cast<AllocaInst>(AfterAllocaInsPt)->getArraySize()))
++AfterAllocaInsPt;
for (Function::arg_iterator AI = F.arg_begin(), AE = F.arg_end(); AI != AE;
++AI) {
Type *Ty = AI->getType();
// Use 'select i8 true, %arg, undef' to simulate a 'no-op' instruction.
Value *TrueValue = ConstantInt::getTrue(F.getContext());
Value *UndefValue = UndefValue::get(Ty);
Instruction *SI = SelectInst::Create(TrueValue, AI, UndefValue,
AI->getName() + ".tmp",
AfterAllocaInsPt);
AI->replaceAllUsesWith(SI);
// Reset the operand, because it was clobbered by the RAUW above.
SI->setOperand(1, AI);
}
}
void llvm::InsertProfilingInitCall(Function *MainFn, const char *FnName,
GlobalValue *Array) {
const Type *ArgVTy =
PointerType::getUnqual(PointerType::getUnqual(Type::Int8Ty));
const PointerType *UIntPtr = PointerType::getUnqual(Type::Int32Ty);
Module &M = *MainFn->getParent();
Constant *InitFn = M.getOrInsertFunction(FnName, Type::Int32Ty, Type::Int32Ty,
ArgVTy, UIntPtr, Type::Int32Ty,
(Type *)0);
// This could force argc and argv into programs that wouldn't otherwise have
// them, but instead we just pass null values in.
std::vector<Value*> Args(4);
Args[0] = Constant::getNullValue(Type::Int32Ty);
Args[1] = Constant::getNullValue(ArgVTy);
// Skip over any allocas in the entry block.
BasicBlock *Entry = MainFn->begin();
BasicBlock::iterator InsertPos = Entry->begin();
while (isa<AllocaInst>(InsertPos)) ++InsertPos;
std::vector<Constant*> GEPIndices(2, Constant::getNullValue(Type::Int32Ty));
unsigned NumElements = 0;
if (Array) {
Args[2] = ConstantExpr::getGetElementPtr(Array, &GEPIndices[0],
GEPIndices.size());
NumElements =
cast<ArrayType>(Array->getType()->getElementType())->getNumElements();
} else {
// If this profiling instrumentation doesn't have a constant array, just
// pass null.
Args[2] = ConstantPointerNull::get(UIntPtr);
}
Args[3] = ConstantInt::get(Type::Int32Ty, NumElements);
Instruction *InitCall = CallInst::Create(InitFn, Args.begin(), Args.end(),
"newargc", InsertPos);
// If argc or argv are not available in main, just pass null values in.
Function::arg_iterator AI;
switch (MainFn->arg_size()) {
default:
case 2:
AI = MainFn->arg_begin(); ++AI;
if (AI->getType() != ArgVTy) {
Instruction::CastOps opcode = CastInst::getCastOpcode(AI, false, ArgVTy,
false);
InitCall->setOperand(2,
CastInst::create(opcode, AI, ArgVTy, "argv.cast", InitCall));
} else {
InitCall->setOperand(2, AI);
}
/* FALL THROUGH */
case 1:
AI = MainFn->arg_begin();
// If the program looked at argc, have it look at the return value of the
// init call instead.
if (AI->getType() != Type::Int32Ty) {
Instruction::CastOps opcode;
if (!AI->use_empty()) {
opcode = CastInst::getCastOpcode(InitCall, true, AI->getType(), true);
AI->replaceAllUsesWith(
CastInst::create(opcode, InitCall, AI->getType(), "", InsertPos));
}
opcode = CastInst::getCastOpcode(AI, true, Type::Int32Ty, true);
InitCall->setOperand(1,
CastInst::create(opcode, AI, Type::Int32Ty, "argc.cast", InitCall));
} else {
AI->replaceAllUsesWith(InitCall);
InitCall->setOperand(1, AI);
}
case 0: break;
}
}
// RemoveDeadStuffFromFunction - Remove any arguments and return values from F
// that are not in LiveValues. Transform the function and all of the callees of
// the function to not have these arguments and return values.
//
bool DAE::RemoveDeadStuffFromFunction(Function *F) {
// Don't modify fully live functions
if (LiveFunctions.count(F))
return false;
// Start by computing a new prototype for the function, which is the same as
// the old function, but has fewer arguments and a different return type.
FunctionType *FTy = F->getFunctionType();
std::vector<Type*> Params;
// Set up to build a new list of parameter attributes.
SmallVector<AttributeWithIndex, 8> AttributesVec;
const AttributeSet &PAL = F->getAttributes();
// Find out the new return value.
Type *RetTy = FTy->getReturnType();
Type *NRetTy = NULL;
unsigned RetCount = NumRetVals(F);
// -1 means unused, other numbers are the new index
SmallVector<int, 5> NewRetIdxs(RetCount, -1);
std::vector<Type*> RetTypes;
if (RetTy->isVoidTy()) {
NRetTy = RetTy;
} else {
StructType *STy = dyn_cast<StructType>(RetTy);
if (STy)
// Look at each of the original return values individually.
for (unsigned i = 0; i != RetCount; ++i) {
RetOrArg Ret = CreateRet(F, i);
if (LiveValues.erase(Ret)) {
RetTypes.push_back(STy->getElementType(i));
NewRetIdxs[i] = RetTypes.size() - 1;
} else {
++NumRetValsEliminated;
DEBUG(dbgs() << "DAE - Removing return value " << i << " from "
<< F->getName() << "\n");
}
}
else
// We used to return a single value.
if (LiveValues.erase(CreateRet(F, 0))) {
RetTypes.push_back(RetTy);
NewRetIdxs[0] = 0;
} else {
DEBUG(dbgs() << "DAE - Removing return value from " << F->getName()
<< "\n");
++NumRetValsEliminated;
}
if (RetTypes.size() > 1)
// More than one return type? Return a struct with them. Also, if we used
// to return a struct and didn't change the number of return values,
// return a struct again. This prevents changing {something} into
// something and {} into void.
// Make the new struct packed if we used to return a packed struct
// already.
NRetTy = StructType::get(STy->getContext(), RetTypes, STy->isPacked());
else if (RetTypes.size() == 1)
// One return type? Just a simple value then, but only if we didn't use to
// return a struct with that simple value before.
NRetTy = RetTypes.front();
else if (RetTypes.size() == 0)
// No return types? Make it void, but only if we didn't use to return {}.
NRetTy = Type::getVoidTy(F->getContext());
}
assert(NRetTy && "No new return type found?");
// The existing function return attributes.
AttributeSet RAttrs = PAL.getRetAttributes();
// Remove any incompatible attributes, but only if we removed all return
// values. Otherwise, ensure that we don't have any conflicting attributes
// here. Currently, this should not be possible, but special handling might be
// required when new return value attributes are added.
if (NRetTy->isVoidTy())
RAttrs =
AttributeSet::get(NRetTy->getContext(), AttributeSet::ReturnIndex,
AttrBuilder(RAttrs, AttributeSet::ReturnIndex).
removeAttributes(AttributeFuncs::typeIncompatible(NRetTy)));
else
assert(!AttrBuilder(RAttrs, AttributeSet::ReturnIndex).
hasAttributes(AttributeFuncs::typeIncompatible(NRetTy)) &&
"Return attributes no longer compatible?");
if (RAttrs.hasAttributes(AttributeSet::ReturnIndex))
AttributesVec.push_back(AttributeWithIndex::get(NRetTy->getContext(),
AttributeSet::ReturnIndex,
RAttrs));
// Remember which arguments are still alive.
SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false);
// Construct the new parameter list from non-dead arguments. Also construct
// a new set of parameter attributes to correspond. Skip the first parameter
// attribute, since that belongs to the return value.
unsigned i = 0;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
I != E; ++I, ++i) {
RetOrArg Arg = CreateArg(F, i);
if (LiveValues.erase(Arg)) {
Params.push_back(I->getType());
ArgAlive[i] = true;
// Get the original parameter attributes (skipping the first one, that is
// for the return value.
if (PAL.hasAttributes(i + 1)) {
AttributesVec.
push_back(AttributeWithIndex::get(F->getContext(), i + 1,
PAL.getParamAttributes(i + 1)));
AttributesVec.back().Index = Params.size();
}
} else {
++NumArgumentsEliminated;
DEBUG(dbgs() << "DAE - Removing argument " << i << " (" << I->getName()
<< ") from " << F->getName() << "\n");
}
}
if (PAL.hasAttributes(AttributeSet::FunctionIndex))
AttributesVec.push_back(AttributeWithIndex::get(F->getContext(),
AttributeSet::FunctionIndex,
PAL.getFnAttributes()));
// Reconstruct the AttributesList based on the vector we constructed.
AttributeSet NewPAL = AttributeSet::get(F->getContext(), AttributesVec);
// Create the new function type based on the recomputed parameters.
FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg());
// No change?
if (NFTy == FTy)
return false;
// Create the new function body and insert it into the module...
Function *NF = Function::Create(NFTy, F->getLinkage());
NF->copyAttributesFrom(F);
NF->setAttributes(NewPAL);
// Insert the new function before the old function, so we won't be processing
// it again.
F->getParent()->getFunctionList().insert(F, NF);
NF->takeName(F);
// Loop over all of the callers of the function, transforming the call sites
// to pass in a smaller number of arguments into the new function.
//
std::vector<Value*> Args;
while (!F->use_empty()) {
CallSite CS(F->use_back());
Instruction *Call = CS.getInstruction();
AttributesVec.clear();
const AttributeSet &CallPAL = CS.getAttributes();
// The call return attributes.
AttributeSet RAttrs = CallPAL.getRetAttributes();
// Adjust in case the function was changed to return void.
RAttrs =
AttributeSet::get(NF->getContext(), AttributeSet::ReturnIndex,
AttrBuilder(RAttrs, AttributeSet::ReturnIndex).
removeAttributes(AttributeFuncs::typeIncompatible(NF->getReturnType())));
if (RAttrs.hasAttributes(AttributeSet::ReturnIndex))
AttributesVec.push_back(AttributeWithIndex::get(NF->getContext(),
AttributeSet::ReturnIndex,
RAttrs));
// Declare these outside of the loops, so we can reuse them for the second
// loop, which loops the varargs.
CallSite::arg_iterator I = CS.arg_begin();
unsigned i = 0;
// Loop over those operands, corresponding to the normal arguments to the
// original function, and add those that are still alive.
for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i)
if (ArgAlive[i]) {
Args.push_back(*I);
// Get original parameter attributes, but skip return attributes.
if (CallPAL.hasAttributes(i + 1)) {
AttributesVec.
push_back(AttributeWithIndex::get(F->getContext(), i + 1,
CallPAL.getParamAttributes(i + 1)));
AttributesVec.back().Index = Args.size();
}
}
// Push any varargs arguments on the list. Don't forget their attributes.
for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) {
Args.push_back(*I);
if (CallPAL.hasAttributes(i + 1)) {
AttributesVec.
push_back(AttributeWithIndex::get(F->getContext(), i + 1,
CallPAL.getParamAttributes(i + 1)));
AttributesVec.back().Index = Args.size();
}
}
if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
AttributesVec.push_back(AttributeWithIndex::get(Call->getContext(),
AttributeSet::FunctionIndex,
CallPAL.getFnAttributes()));
// Reconstruct the AttributesList based on the vector we constructed.
AttributeSet NewCallPAL = AttributeSet::get(F->getContext(), AttributesVec);
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(NewCallPAL);
} else {
New = CallInst::Create(NF, Args, "", Call);
cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
cast<CallInst>(New)->setAttributes(NewCallPAL);
if (cast<CallInst>(Call)->isTailCall())
cast<CallInst>(New)->setTailCall();
}
New->setDebugLoc(Call->getDebugLoc());
Args.clear();
if (!Call->use_empty()) {
if (New->getType() == Call->getType()) {
// Return type not changed? Just replace users then.
Call->replaceAllUsesWith(New);
New->takeName(Call);
} else if (New->getType()->isVoidTy()) {
// Our return value has uses, but they will get removed later on.
// Replace by null for now.
if (!Call->getType()->isX86_MMXTy())
Call->replaceAllUsesWith(Constant::getNullValue(Call->getType()));
} else {
assert(RetTy->isStructTy() &&
"Return type changed, but not into a void. The old return type"
" must have been a struct!");
Instruction *InsertPt = Call;
if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
BasicBlock::iterator IP = II->getNormalDest()->begin();
while (isa<PHINode>(IP)) ++IP;
InsertPt = IP;
}
// We used to return a struct. Instead of doing smart stuff with all the
// uses of this struct, we will just rebuild it using
// extract/insertvalue chaining and let instcombine clean that up.
//
// Start out building up our return value from undef
Value *RetVal = UndefValue::get(RetTy);
for (unsigned i = 0; i != RetCount; ++i)
if (NewRetIdxs[i] != -1) {
Value *V;
if (RetTypes.size() > 1)
// We are still returning a struct, so extract the value from our
// return value
V = ExtractValueInst::Create(New, NewRetIdxs[i], "newret",
InsertPt);
else
// We are now returning a single element, so just insert that
V = New;
// Insert the value at the old position
RetVal = InsertValueInst::Create(RetVal, V, i, "oldret", InsertPt);
}
// Now, replace all uses of the old call instruction with the return
// struct we built
Call->replaceAllUsesWith(RetVal);
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.
i = 0;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
I2 = NF->arg_begin(); I != E; ++I, ++i)
if (ArgAlive[i]) {
// If this is a live argument, move the name and users over to the new
// version.
I->replaceAllUsesWith(I2);
I2->takeName(I);
++I2;
} else {
// If this argument is dead, replace any uses of it with null constants
// (these are guaranteed to become unused later on).
if (!I->getType()->isX86_MMXTy())
I->replaceAllUsesWith(Constant::getNullValue(I->getType()));
}
// If we change the return value of the function we must rewrite any return
// instructions. Check this now.
if (F->getReturnType() != NF->getReturnType())
for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB)
if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
Value *RetVal;
if (NFTy->getReturnType()->isVoidTy()) {
RetVal = 0;
} else {
assert (RetTy->isStructTy());
// The original return value was a struct, insert
// extractvalue/insertvalue chains to extract only the values we need
// to return and insert them into our new result.
// This does generate messy code, but we'll let it to instcombine to
// clean that up.
Value *OldRet = RI->getOperand(0);
// Start out building up our return value from undef
RetVal = UndefValue::get(NRetTy);
for (unsigned i = 0; i != RetCount; ++i)
if (NewRetIdxs[i] != -1) {
ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i,
"oldret", RI);
if (RetTypes.size() > 1) {
// We're still returning a struct, so reinsert the value into
// our new return value at the new index
RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i],
"newret", RI);
} else {
// We are now only returning a simple value, so just return the
// extracted value.
RetVal = EV;
}
}
}
// Replace the return instruction with one returning the new return
// value (possibly 0 if we became void).
ReturnInst::Create(F->getContext(), RetVal, RI);
BB->getInstList().erase(RI);
}
// Patch the pointer to LLVM function in debug info descriptor.
FunctionDIMap::iterator DI = FunctionDIs.find(F);
if (DI != FunctionDIs.end())
DI->second.replaceFunction(NF);
// Now that the old function is dead, delete it.
F->eraseFromParent();
return true;
}
/// 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;
}
static bool eliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
BasicBlock *&OldEntry,
bool &TailCallsAreMarkedTail,
SmallVectorImpl<PHINode *> &ArgumentPHIs,
bool CannotTailCallElimCallsMarkedTail) {
// If we are introducing accumulator recursion to eliminate operations after
// the call instruction that are both associative and commutative, the initial
// value for the accumulator is placed in this variable. If this value is set
// then we actually perform accumulator recursion elimination instead of
// simple tail recursion elimination. If the operation is an LLVM instruction
// (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then
// we are handling the case when the return instruction returns a constant C
// which is different to the constant returned by other return instructions
// (which is recorded in AccumulatorRecursionEliminationInitVal). This is a
// special case of accumulator recursion, the operation being "return C".
Value *AccumulatorRecursionEliminationInitVal = nullptr;
Instruction *AccumulatorRecursionInstr = nullptr;
// Ok, we found a potential tail call. We can currently only transform the
// tail call if all of the instructions between the call and the return are
// movable to above the call itself, leaving the call next to the return.
// Check that this is the case now.
BasicBlock::iterator BBI(CI);
for (++BBI; &*BBI != Ret; ++BBI) {
if (canMoveAboveCall(&*BBI, CI)) continue;
// If we can't move the instruction above the call, it might be because it
// is an associative and commutative operation that could be transformed
// using accumulator recursion elimination. Check to see if this is the
// case, and if so, remember the initial accumulator value for later.
if ((AccumulatorRecursionEliminationInitVal =
canTransformAccumulatorRecursion(&*BBI, CI))) {
// Yes, this is accumulator recursion. Remember which instruction
// accumulates.
AccumulatorRecursionInstr = &*BBI;
} else {
return false; // Otherwise, we cannot eliminate the tail recursion!
}
}
// We can only transform call/return pairs that either ignore the return value
// of the call and return void, ignore the value of the call and return a
// constant, return the value returned by the tail call, or that are being
// accumulator recursion variable eliminated.
if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
!isa<UndefValue>(Ret->getReturnValue()) &&
AccumulatorRecursionEliminationInitVal == nullptr &&
!getCommonReturnValue(nullptr, CI)) {
// One case remains that we are able to handle: the current return
// instruction returns a constant, and all other return instructions
// return a different constant.
if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret))
return false; // Current return instruction does not return a constant.
// Check that all other return instructions return a common constant. If
// so, record it in AccumulatorRecursionEliminationInitVal.
AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI);
if (!AccumulatorRecursionEliminationInitVal)
return false;
}
BasicBlock *BB = Ret->getParent();
Function *F = BB->getParent();
emitOptimizationRemark(F->getContext(), "tailcallelim", *F, CI->getDebugLoc(),
"transforming tail recursion to loop");
// OK! We can transform this tail call. If this is the first one found,
// create the new entry block, allowing us to branch back to the old entry.
if (!OldEntry) {
OldEntry = &F->getEntryBlock();
BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
NewEntry->takeName(OldEntry);
OldEntry->setName("tailrecurse");
BranchInst::Create(OldEntry, NewEntry);
// If this tail call is marked 'tail' and if there are any allocas in the
// entry block, move them up to the new entry block.
TailCallsAreMarkedTail = CI->isTailCall();
if (TailCallsAreMarkedTail)
// Move all fixed sized allocas from OldEntry to NewEntry.
for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
NEBI = NewEntry->begin(); OEBI != E; )
if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
if (isa<ConstantInt>(AI->getArraySize()))
AI->moveBefore(&*NEBI);
// Now that we have created a new block, which jumps to the entry
// block, insert a PHI node for each argument of the function.
// For now, we initialize each PHI to only have the real arguments
// which are passed in.
Instruction *InsertPos = &OldEntry->front();
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
I != E; ++I) {
PHINode *PN = PHINode::Create(I->getType(), 2,
I->getName() + ".tr", InsertPos);
I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
PN->addIncoming(&*I, NewEntry);
ArgumentPHIs.push_back(PN);
}
}
// If this function has self recursive calls in the tail position where some
// are marked tail and some are not, only transform one flavor or another. We
// have to choose whether we move allocas in the entry block to the new entry
// block or not, so we can't make a good choice for both. NOTE: We could do
// slightly better here in the case that the function has no entry block
// allocas.
if (TailCallsAreMarkedTail && !CI->isTailCall())
return false;
// Ok, now that we know we have a pseudo-entry block WITH all of the
// required PHI nodes, add entries into the PHI node for the actual
// parameters passed into the tail-recursive call.
for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB);
// If we are introducing an accumulator variable to eliminate the recursion,
// do so now. Note that we _know_ that no subsequent tail recursion
// eliminations will happen on this function because of the way the
// accumulator recursion predicate is set up.
//
if (AccumulatorRecursionEliminationInitVal) {
Instruction *AccRecInstr = AccumulatorRecursionInstr;
// Start by inserting a new PHI node for the accumulator.
pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry);
PHINode *AccPN = PHINode::Create(
AccumulatorRecursionEliminationInitVal->getType(),
std::distance(PB, PE) + 1, "accumulator.tr", &OldEntry->front());
// Loop over all of the predecessors of the tail recursion block. For the
// real entry into the function we seed the PHI with the initial value,
// computed earlier. For any other existing branches to this block (due to
// other tail recursions eliminated) the accumulator is not modified.
// Because we haven't added the branch in the current block to OldEntry yet,
// it will not show up as a predecessor.
for (pred_iterator PI = PB; PI != PE; ++PI) {
BasicBlock *P = *PI;
if (P == &F->getEntryBlock())
AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P);
else
AccPN->addIncoming(AccPN, P);
}
if (AccRecInstr) {
// Add an incoming argument for the current block, which is computed by
// our associative and commutative accumulator instruction.
AccPN->addIncoming(AccRecInstr, BB);
// Next, rewrite the accumulator recursion instruction so that it does not
// use the result of the call anymore, instead, use the PHI node we just
// inserted.
AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
} else {
// Add an incoming argument for the current block, which is just the
// constant returned by the current return instruction.
AccPN->addIncoming(Ret->getReturnValue(), BB);
}
// Finally, rewrite any return instructions in the program to return the PHI
// node instead of the "initval" that they do currently. This loop will
// actually rewrite the return value we are destroying, but that's ok.
for (BasicBlock &BBI : *F)
if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI.getTerminator()))
RI->setOperand(0, AccPN);
++NumAccumAdded;
}
// Now that all of the PHI nodes are in place, remove the call and
// ret instructions, replacing them with an unconditional branch.
BranchInst *NewBI = BranchInst::Create(OldEntry, Ret);
NewBI->setDebugLoc(CI->getDebugLoc());
BB->getInstList().erase(Ret); // Remove return.
BB->getInstList().erase(CI); // Remove call.
++NumEliminated;
return true;
}
// First thing we need to do is scan the whole function for values that are
// live across unwind edges. Each value that is live across an unwind edge
// we spill into a stack location, guaranteeing that there is nothing live
// across the unwind edge. This process also splits all critical edges
// coming out of invoke's.
void LowerInvoke::
splitLiveRangesLiveAcrossInvokes(std::vector<InvokeInst*> &Invokes) {
// First step, split all critical edges from invoke instructions.
for (unsigned i = 0, e = Invokes.size(); i != e; ++i) {
InvokeInst *II = Invokes[i];
SplitCriticalEdge(II, 0, this);
SplitCriticalEdge(II, 1, this);
assert(!isa<PHINode>(II->getNormalDest()) &&
!isa<PHINode>(II->getUnwindDest()) &&
"critical edge splitting left single entry phi nodes?");
}
Function *F = Invokes.back()->getParent()->getParent();
// To avoid having to handle incoming arguments specially, we lower each arg
// to a copy instruction in the entry block. This ensures that the argument
// value itself cannot be live across the entry block.
BasicBlock::iterator AfterAllocaInsertPt = F->begin()->begin();
while (isa<AllocaInst>(AfterAllocaInsertPt) &&
isa<ConstantInt>(cast<AllocaInst>(AfterAllocaInsertPt)->getArraySize()))
++AfterAllocaInsertPt;
for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
AI != E; ++AI) {
// This is always a no-op cast because we're casting AI to AI->getType() so
// src and destination types are identical. BitCast is the only possibility.
CastInst *NC = new BitCastInst(
AI, AI->getType(), AI->getName()+".tmp", AfterAllocaInsertPt);
AI->replaceAllUsesWith(NC);
// Normally its is forbidden to replace a CastInst's operand because it
// could cause the opcode to reflect an illegal conversion. However, we're
// replacing it here with the same value it was constructed with to simply
// make NC its user.
NC->setOperand(0, AI);
}
// Finally, scan the code looking for instructions with bad live ranges.
for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
// Ignore obvious cases we don't have to handle. In particular, most
// instructions either have no uses or only have a single use inside the
// current block. Ignore them quickly.
Instruction *Inst = II;
if (Inst->use_empty()) continue;
if (Inst->hasOneUse() &&
cast<Instruction>(Inst->use_back())->getParent() == BB &&
!isa<PHINode>(Inst->use_back())) continue;
// If this is an alloca in the entry block, it's not a real register
// value.
if (AllocaInst *AI = dyn_cast<AllocaInst>(Inst))
if (isa<ConstantInt>(AI->getArraySize()) && BB == F->begin())
continue;
// Avoid iterator invalidation by copying users to a temporary vector.
std::vector<Instruction*> Users;
for (Value::use_iterator UI = Inst->use_begin(), E = Inst->use_end();
UI != E; ++UI) {
Instruction *User = cast<Instruction>(*UI);
if (User->getParent() != BB || isa<PHINode>(User))
Users.push_back(User);
}
// Scan all of the uses and see if the live range is live across an unwind
// edge. If we find a use live across an invoke edge, create an alloca
// and spill the value.
std::set<InvokeInst*> InvokesWithStoreInserted;
// Find all of the blocks that this value is live in.
std::set<BasicBlock*> LiveBBs;
LiveBBs.insert(Inst->getParent());
while (!Users.empty()) {
Instruction *U = Users.back();
Users.pop_back();
if (!isa<PHINode>(U)) {
MarkBlocksLiveIn(U->getParent(), LiveBBs);
} else {
// Uses for a PHI node occur in their predecessor block.
PHINode *PN = cast<PHINode>(U);
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
if (PN->getIncomingValue(i) == Inst)
MarkBlocksLiveIn(PN->getIncomingBlock(i), LiveBBs);
}
}
// Now that we know all of the blocks that this thing is live in, see if
// it includes any of the unwind locations.
bool NeedsSpill = false;
for (unsigned i = 0, e = Invokes.size(); i != e; ++i) {
BasicBlock *UnwindBlock = Invokes[i]->getUnwindDest();
if (UnwindBlock != BB && LiveBBs.count(UnwindBlock)) {
NeedsSpill = true;
}
}
// If we decided we need a spill, do it.
if (NeedsSpill) {
++NumSpilled;
DemoteRegToStack(*Inst, true);
}
}
}
/// 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;
}
/// DeleteDeadVarargs - If this is an function that takes a ... list, and if
/// llvm.vastart is never called, the varargs list is dead for the function.
bool DAE::DeleteDeadVarargs(Function &Fn) {
assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!");
if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false;
// Ensure that the function is only directly called.
if (Fn.hasAddressTaken())
return false;
// Okay, we know we can transform this function if safe. Scan its body
// looking for calls to llvm.vastart.
for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
if (II->getIntrinsicID() == Intrinsic::vastart)
return false;
}
}
}
// If we get here, there are no calls to llvm.vastart in the function body,
// remove the "..." and adjust all the calls.
// Start by computing a new prototype for the function, which is the same as
// the old function, but doesn't have isVarArg set.
const FunctionType *FTy = Fn.getFunctionType();
std::vector<const Type*> Params(FTy->param_begin(), FTy->param_end());
FunctionType *NFTy = FunctionType::get(FTy->getReturnType(),
Params, false);
unsigned NumArgs = Params.size();
// Create the new function body and insert it into the module...
Function *NF = Function::Create(NFTy, Fn.getLinkage());
NF->copyAttributesFrom(&Fn);
Fn.getParent()->getFunctionList().insert(&Fn, NF);
NF->takeName(&Fn);
// Loop over all of the callers of the function, transforming the call sites
// to pass in a smaller number of arguments into the new function.
//
std::vector<Value*> Args;
while (!Fn.use_empty()) {
CallSite CS = CallSite::get(Fn.use_back());
Instruction *Call = CS.getInstruction();
// Pass all the same arguments.
Args.assign(CS.arg_begin(), CS.arg_begin()+NumArgs);
// Drop any attributes that were on the vararg arguments.
AttrListPtr PAL = CS.getAttributes();
if (!PAL.isEmpty() && PAL.getSlot(PAL.getNumSlots() - 1).Index > NumArgs) {
SmallVector<AttributeWithIndex, 8> AttributesVec;
for (unsigned i = 0; PAL.getSlot(i).Index <= NumArgs; ++i)
AttributesVec.push_back(PAL.getSlot(i));
if (Attributes FnAttrs = PAL.getFnAttributes())
AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
PAL = AttrListPtr::get(AttributesVec.begin(), AttributesVec.end());
}
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(PAL);
} else {
New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call);
cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
cast<CallInst>(New)->setAttributes(PAL);
if (cast<CallInst>(Call)->isTailCall())
cast<CallInst>(New)->setTailCall();
}
if (MDNode *N = Call->getDbgMetadata())
New->setDbgMetadata(N);
Args.clear();
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(), Fn.getBasicBlockList());
// Loop over the argument list, transfering uses of the old arguments over to
// the new arguments, also transfering over the names as well. While we're at
// it, remove the dead arguments from the DeadArguments list.
//
for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(),
I2 = NF->arg_begin(); I != E; ++I, ++I2) {
// Move the name and users over to the new version.
I->replaceAllUsesWith(I2);
I2->takeName(I);
}
// Finally, nuke the old function.
Fn.eraseFromParent();
return true;
}
/// splitLiveRangesAcrossInvokes - Each value that is live across an unwind edge
/// we spill into a stack location, guaranteeing that there is nothing live
/// across the unwind edge. This process also splits all critical edges
/// coming out of invoke's.
/// FIXME: Move this function to a common utility file (Local.cpp?) so
/// both SjLj and LowerInvoke can use it.
void SjLjEHPass::
splitLiveRangesAcrossInvokes(SmallVector<InvokeInst*,16> &Invokes) {
// First step, split all critical edges from invoke instructions.
for (unsigned i = 0, e = Invokes.size(); i != e; ++i) {
InvokeInst *II = Invokes[i];
SplitCriticalEdge(II, 0, this);
// FIXME: New EH - This if-condition will be always true in the new scheme.
if (II->getUnwindDest()->isLandingPad()) {
SmallVector<BasicBlock*, 2> NewBBs;
SplitLandingPadPredecessors(II->getUnwindDest(), II->getParent(),
".1", ".2", this, NewBBs);
LPadSuccMap[II] = *succ_begin(NewBBs[0]);
} else {
SplitCriticalEdge(II, 1, this);
}
assert(!isa<PHINode>(II->getNormalDest()) &&
!isa<PHINode>(II->getUnwindDest()) &&
"Critical edge splitting left single entry phi nodes?");
}
Function *F = Invokes.back()->getParent()->getParent();
// To avoid having to handle incoming arguments specially, we lower each arg
// to a copy instruction in the entry block. This ensures that the argument
// value itself cannot be live across the entry block.
BasicBlock::iterator AfterAllocaInsertPt = F->begin()->begin();
while (isa<AllocaInst>(AfterAllocaInsertPt) &&
isa<ConstantInt>(cast<AllocaInst>(AfterAllocaInsertPt)->getArraySize()))
++AfterAllocaInsertPt;
for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
AI != E; ++AI) {
Type *Ty = AI->getType();
// Aggregate types can't be cast, but are legal argument types, so we have
// to handle them differently. We use an extract/insert pair as a
// lightweight method to achieve the same goal.
if (isa<StructType>(Ty) || isa<ArrayType>(Ty) || isa<VectorType>(Ty)) {
Instruction *EI = ExtractValueInst::Create(AI, 0, "",AfterAllocaInsertPt);
Instruction *NI = InsertValueInst::Create(AI, EI, 0);
NI->insertAfter(EI);
AI->replaceAllUsesWith(NI);
// Set the operand of the instructions back to the AllocaInst.
EI->setOperand(0, AI);
NI->setOperand(0, AI);
} else {
// This is always a no-op cast because we're casting AI to AI->getType()
// so src and destination types are identical. BitCast is the only
// possibility.
CastInst *NC = new BitCastInst(
AI, AI->getType(), AI->getName()+".tmp", AfterAllocaInsertPt);
AI->replaceAllUsesWith(NC);
// Set the operand of the cast instruction back to the AllocaInst.
// Normally it's forbidden to replace a CastInst's operand because it
// could cause the opcode to reflect an illegal conversion. However,
// we're replacing it here with the same value it was constructed with.
// We do this because the above replaceAllUsesWith() clobbered the
// operand, but we want this one to remain.
NC->setOperand(0, AI);
}
}
// Finally, scan the code looking for instructions with bad live ranges.
for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
// Ignore obvious cases we don't have to handle. In particular, most
// instructions either have no uses or only have a single use inside the
// current block. Ignore them quickly.
Instruction *Inst = II;
if (Inst->use_empty()) continue;
if (Inst->hasOneUse() &&
cast<Instruction>(Inst->use_back())->getParent() == BB &&
!isa<PHINode>(Inst->use_back())) continue;
// If this is an alloca in the entry block, it's not a real register
// value.
if (AllocaInst *AI = dyn_cast<AllocaInst>(Inst))
if (isa<ConstantInt>(AI->getArraySize()) && BB == F->begin())
continue;
// Avoid iterator invalidation by copying users to a temporary vector.
SmallVector<Instruction*,16> Users;
for (Value::use_iterator UI = Inst->use_begin(), E = Inst->use_end();
UI != E; ++UI) {
Instruction *User = cast<Instruction>(*UI);
if (User->getParent() != BB || isa<PHINode>(User))
Users.push_back(User);
}
// Find all of the blocks that this value is live in.
std::set<BasicBlock*> LiveBBs;
LiveBBs.insert(Inst->getParent());
while (!Users.empty()) {
Instruction *U = Users.back();
Users.pop_back();
if (!isa<PHINode>(U)) {
MarkBlocksLiveIn(U->getParent(), LiveBBs);
} else {
// Uses for a PHI node occur in their predecessor block.
PHINode *PN = cast<PHINode>(U);
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
if (PN->getIncomingValue(i) == Inst)
MarkBlocksLiveIn(PN->getIncomingBlock(i), LiveBBs);
}
}
// Now that we know all of the blocks that this thing is live in, see if
// it includes any of the unwind locations.
bool NeedsSpill = false;
for (unsigned i = 0, e = Invokes.size(); i != e; ++i) {
BasicBlock *UnwindBlock = Invokes[i]->getUnwindDest();
if (UnwindBlock != BB && LiveBBs.count(UnwindBlock)) {
NeedsSpill = true;
}
}
// If we decided we need a spill, do it.
// FIXME: Spilling this way is overkill, as it forces all uses of
// the value to be reloaded from the stack slot, even those that aren't
// in the unwind blocks. We should be more selective.
if (NeedsSpill) {
++NumSpilled;
DemoteRegToStack(*Inst, true);
}
}
}
bool RemovePhantomArg::removePhantomArgument(Function* F) {
assert(!F->getFunctionType()->isVarArg() && "Function is varargs!");
// if this is a callee function that get inlined, we don't
// need to change it because it will be dropped before
// codegen anyway
if (F->hasFnAttribute("is-inlined-callee")) return false;
// if a function should get safepoint then it needs vmstate
// and should have the phantom arg
if (!shouldGetSafepoints(*F)) return false;
// Start by computing a new prototype for the function, which is the same as
// the old function, but has one less argument.
FunctionType *FTy = F->getFunctionType();
std::vector<Type*> Params;
// Set up to build a new list of parameter attributes.
SmallVector<AttributeSet, 8> AttributesVec;
const AttributeSet &PAL = F->getAttributes();
// Construct the new parameter list from non-phantom arguments. Also
// construct a new set of parameter attributes to correspond. Skip the
// first parameter attribute, since that belongs to the return value.
unsigned i = 1;
Function::arg_iterator I = F->arg_begin();
llvm::Argument* phantomArg = I++;
assert(phantomArg->getType()->isIntegerTy(32) && "First arg must be int32");
for (Function::arg_iterator E = F->arg_end();
I != E; ++I, ++i) {
Params.push_back(I->getType());
// Get the original parameter attributes (skipping the first one, that is
// for the return value.
if (PAL.hasAttributes(i + 1)) {
AttrBuilder B(PAL, i + 1);
AttributesVec.
push_back(AttributeSet::get(F->getContext(), Params.size(), B));
}
}
// Find out the new return value.
Type *NRetTy = FTy->getReturnType();
// The existing function return attributes.
AttributeSet RAttrs = PAL.getRetAttributes();
if (RAttrs.hasAttributes(AttributeSet::ReturnIndex))
AttributesVec.push_back(AttributeSet::get(NRetTy->getContext(), RAttrs));
if (PAL.hasAttributes(AttributeSet::FunctionIndex))
AttributesVec.push_back(AttributeSet::get(F->getContext(),
PAL.getFnAttributes()));
// Reconstruct the AttributesList based on the vector we constructed.
AttributeSet NewPAL = AttributeSet::get(F->getContext(), AttributesVec);
// Create the new function type based on the recomputed parameters.
FunctionType *NFTy = FunctionType::get(NRetTy, Params, false);
assert(NFTy != FTy && "They should not be the same");
// Create the new function body and insert it into the module...
Function *NF = Function::Create(NFTy, F->getLinkage());
NF->copyAttributesFrom(F);
NF->setAttributes(NewPAL);
// Insert the new function before the old function, so we won't be processing
// it again.
F->getParent()->getFunctionList().insert(F, NF);
NF->takeName(F);
// 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, ++I2) {
I->replaceAllUsesWith(I2);
I2->takeName(I);
}
// Patch the pointer to LLVM function in debug info descriptor.
auto DI = FunctionDIs.find(F);
if (DI != FunctionDIs.end())
DI->second.replaceFunction(NF);
assert(F->use_empty() && "Function should have no directly use");
// Now that the old function is dead, delete it.
F->eraseFromParent();
return true;
}
/// DeleteDeadVarargs - If this is an function that takes a ... list, and if
/// llvm.vastart is never called, the varargs list is dead for the function.
bool DAE::DeleteDeadVarargs(Function &Fn) {
assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!");
if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false;
// Ensure that the function is only directly called.
if (Fn.hasAddressTaken())
return false;
// Don't touch naked functions. The assembly might be using an argument, or
// otherwise rely on the frame layout in a way that this analysis will not
// see.
if (Fn.hasFnAttribute(Attribute::Naked)) {
return false;
}
// Okay, we know we can transform this function if safe. Scan its body
// looking for calls marked musttail or calls to llvm.vastart.
for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
CallInst *CI = dyn_cast<CallInst>(I);
if (!CI)
continue;
if (CI->isMustTailCall())
return false;
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
if (II->getIntrinsicID() == Intrinsic::vastart)
return false;
}
}
}
// If we get here, there are no calls to llvm.vastart in the function body,
// remove the "..." and adjust all the calls.
// Start by computing a new prototype for the function, which is the same as
// the old function, but doesn't have isVarArg set.
FunctionType *FTy = Fn.getFunctionType();
std::vector<Type*> Params(FTy->param_begin(), FTy->param_end());
FunctionType *NFTy = FunctionType::get(FTy->getReturnType(),
Params, false);
unsigned NumArgs = Params.size();
// Create the new function body and insert it into the module...
Function *NF = Function::Create(NFTy, Fn.getLinkage());
NF->copyAttributesFrom(&Fn);
Fn.getParent()->getFunctionList().insert(Fn.getIterator(), NF);
NF->takeName(&Fn);
// Loop over all of the callers of the function, transforming the call sites
// to pass in a smaller number of arguments into the new function.
//
std::vector<Value*> Args;
for (Value::user_iterator I = Fn.user_begin(), E = Fn.user_end(); I != E; ) {
CallSite CS(*I++);
if (!CS)
continue;
Instruction *Call = CS.getInstruction();
// Pass all the same arguments.
Args.assign(CS.arg_begin(), CS.arg_begin() + NumArgs);
// Drop any attributes that were on the vararg arguments.
AttributeSet PAL = CS.getAttributes();
if (!PAL.isEmpty() && PAL.getSlotIndex(PAL.getNumSlots() - 1) > NumArgs) {
SmallVector<AttributeSet, 8> AttributesVec;
for (unsigned i = 0; PAL.getSlotIndex(i) <= NumArgs; ++i)
AttributesVec.push_back(PAL.getSlotAttributes(i));
if (PAL.hasAttributes(AttributeSet::FunctionIndex))
AttributesVec.push_back(AttributeSet::get(Fn.getContext(),
PAL.getFnAttributes()));
PAL = AttributeSet::get(Fn.getContext(), AttributesVec);
}
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(PAL);
} else {
New = CallInst::Create(NF, Args, "", Call);
cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
cast<CallInst>(New)->setAttributes(PAL);
if (cast<CallInst>(Call)->isTailCall())
cast<CallInst>(New)->setTailCall();
}
New->setDebugLoc(Call->getDebugLoc());
Args.clear();
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(), Fn.getBasicBlockList());
// Loop over the argument list, transferring uses of the old arguments over to
// the new arguments, also transferring over the names as well. While we're at
// it, remove the dead arguments from the DeadArguments list.
//
for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(),
I2 = NF->arg_begin(); I != E; ++I, ++I2) {
// Move the name and users over to the new version.
I->replaceAllUsesWith(&*I2);
I2->takeName(&*I);
}
// Patch the pointer to LLVM function in debug info descriptor.
NF->setSubprogram(Fn.getSubprogram());
// Fix up any BlockAddresses that refer to the function.
Fn.replaceAllUsesWith(ConstantExpr::getBitCast(NF, Fn.getType()));
// Delete the bitcast that we just created, so that NF does not
// appear to be address-taken.
NF->removeDeadConstantUsers();
// Finally, nuke the old function.
Fn.eraseFromParent();
return true;
}
/// PropagateConstantsIntoArguments - Look at all uses of the specified
/// function. If all uses are direct call sites, and all pass a particular
/// constant in for an argument, propagate that constant in as the argument.
///
bool IPCP::PropagateConstantsIntoArguments(Function &F) {
if (F.arg_empty() || F.use_empty()) return false; // No arguments? Early exit.
// For each argument, keep track of its constant value and whether it is a
// constant or not. The bool is driven to true when found to be non-constant.
SmallVector<std::pair<Constant*, bool>, 16> ArgumentConstants;
ArgumentConstants.resize(F.arg_size());
unsigned NumNonconstant = 0;
for (Value::use_iterator UI = F.use_begin(), E = F.use_end(); UI != E; ++UI) {
User *U = *UI;
// Ignore blockaddress uses.
if (isa<BlockAddress>(U)) continue;
// Used by a non-instruction, or not the callee of a function, do not
// transform.
if (!isa<CallInst>(U) && !isa<InvokeInst>(U))
return false;
CallSite CS(cast<Instruction>(U));
if (!CS.isCallee(UI))
return false;
// Check out all of the potentially constant arguments. Note that we don't
// inspect varargs here.
CallSite::arg_iterator AI = CS.arg_begin();
Function::arg_iterator Arg = F.arg_begin();
for (unsigned i = 0, e = ArgumentConstants.size(); i != e;
++i, ++AI, ++Arg) {
// If this argument is known non-constant, ignore it.
if (ArgumentConstants[i].second)
continue;
Constant *C = dyn_cast<Constant>(*AI);
if (C && ArgumentConstants[i].first == 0) {
ArgumentConstants[i].first = C; // First constant seen.
} else if (C && ArgumentConstants[i].first == C) {
// Still the constant value we think it is.
} else if (*AI == &*Arg) {
// Ignore recursive calls passing argument down.
} else {
// Argument became non-constant. If all arguments are non-constant now,
// give up on this function.
if (++NumNonconstant == ArgumentConstants.size())
return false;
ArgumentConstants[i].second = true;
}
}
}
// If we got to this point, there is a constant argument!
assert(NumNonconstant != ArgumentConstants.size());
bool MadeChange = false;
Function::arg_iterator AI = F.arg_begin();
for (unsigned i = 0, e = ArgumentConstants.size(); i != e; ++i, ++AI) {
// Do we have a constant argument?
if (ArgumentConstants[i].second || AI->use_empty() ||
(AI->hasByValAttr() && !F.onlyReadsMemory()))
continue;
Value *V = ArgumentConstants[i].first;
if (V == 0) V = UndefValue::get(AI->getType());
AI->replaceAllUsesWith(V);
++NumArgumentsProped;
MadeChange = true;
}
return MadeChange;
}
// RemoveDeadStuffFromFunction - Remove any arguments and return values from F
// that are not in LiveValues. Transform the function and all of the callees of
// the function to not have these arguments and return values.
//
bool DAE::RemoveDeadStuffFromFunction(Function *F) {
// Don't modify fully live functions
if (LiveFunctions.count(F))
return false;
// Start by computing a new prototype for the function, which is the same as
// the old function, but has fewer arguments and a different return type.
FunctionType *FTy = F->getFunctionType();
std::vector<Type*> Params;
// Keep track of if we have a live 'returned' argument
bool HasLiveReturnedArg = false;
// Set up to build a new list of parameter attributes.
SmallVector<AttributeSet, 8> AttributesVec;
const AttributeSet &PAL = F->getAttributes();
// Remember which arguments are still alive.
SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false);
// Construct the new parameter list from non-dead arguments. Also construct
// a new set of parameter attributes to correspond. Skip the first parameter
// attribute, since that belongs to the return value.
unsigned i = 0;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
I != E; ++I, ++i) {
RetOrArg Arg = CreateArg(F, i);
if (LiveValues.erase(Arg)) {
Params.push_back(I->getType());
ArgAlive[i] = true;
// Get the original parameter attributes (skipping the first one, that is
// for the return value.
if (PAL.hasAttributes(i + 1)) {
AttrBuilder B(PAL, i + 1);
if (B.contains(Attribute::Returned))
HasLiveReturnedArg = true;
AttributesVec.
push_back(AttributeSet::get(F->getContext(), Params.size(), B));
}
} else {
++NumArgumentsEliminated;
DEBUG(dbgs() << "DAE - Removing argument " << i << " (" << I->getName()
<< ") from " << F->getName() << "\n");
}
}
// Find out the new return value.
Type *RetTy = FTy->getReturnType();
Type *NRetTy = nullptr;
unsigned RetCount = NumRetVals(F);
// -1 means unused, other numbers are the new index
SmallVector<int, 5> NewRetIdxs(RetCount, -1);
std::vector<Type*> RetTypes;
// If there is a function with a live 'returned' argument but a dead return
// value, then there are two possible actions:
// 1) Eliminate the return value and take off the 'returned' attribute on the
// argument.
// 2) Retain the 'returned' attribute and treat the return value (but not the
// entire function) as live so that it is not eliminated.
//
// It's not clear in the general case which option is more profitable because,
// even in the absence of explicit uses of the return value, code generation
// is free to use the 'returned' attribute to do things like eliding
// save/restores of registers across calls. Whether or not this happens is
// target and ABI-specific as well as depending on the amount of register
// pressure, so there's no good way for an IR-level pass to figure this out.
//
// Fortunately, the only places where 'returned' is currently generated by
// the FE are places where 'returned' is basically free and almost always a
// performance win, so the second option can just be used always for now.
//
// This should be revisited if 'returned' is ever applied more liberally.
if (RetTy->isVoidTy() || HasLiveReturnedArg) {
NRetTy = RetTy;
} else {
// Look at each of the original return values individually.
for (unsigned i = 0; i != RetCount; ++i) {
RetOrArg Ret = CreateRet(F, i);
if (LiveValues.erase(Ret)) {
RetTypes.push_back(getRetComponentType(F, i));
NewRetIdxs[i] = RetTypes.size() - 1;
} else {
++NumRetValsEliminated;
DEBUG(dbgs() << "DAE - Removing return value " << i << " from "
<< F->getName() << "\n");
}
}
if (RetTypes.size() > 1) {
// More than one return type? Reduce it down to size.
if (StructType *STy = dyn_cast<StructType>(RetTy)) {
// Make the new struct packed if we used to return a packed struct
// already.
NRetTy = StructType::get(STy->getContext(), RetTypes, STy->isPacked());
} else {
assert(isa<ArrayType>(RetTy) && "unexpected multi-value return");
NRetTy = ArrayType::get(RetTypes[0], RetTypes.size());
}
} else if (RetTypes.size() == 1)
// One return type? Just a simple value then, but only if we didn't use to
// return a struct with that simple value before.
NRetTy = RetTypes.front();
else if (RetTypes.size() == 0)
// No return types? Make it void, but only if we didn't use to return {}.
NRetTy = Type::getVoidTy(F->getContext());
}
assert(NRetTy && "No new return type found?");
// The existing function return attributes.
AttributeSet RAttrs = PAL.getRetAttributes();
// Remove any incompatible attributes, but only if we removed all return
// values. Otherwise, ensure that we don't have any conflicting attributes
// here. Currently, this should not be possible, but special handling might be
// required when new return value attributes are added.
if (NRetTy->isVoidTy())
RAttrs = RAttrs.removeAttributes(NRetTy->getContext(),
AttributeSet::ReturnIndex,
AttributeFuncs::typeIncompatible(NRetTy));
else
assert(!AttrBuilder(RAttrs, AttributeSet::ReturnIndex).
overlaps(AttributeFuncs::typeIncompatible(NRetTy)) &&
"Return attributes no longer compatible?");
if (RAttrs.hasAttributes(AttributeSet::ReturnIndex))
AttributesVec.push_back(AttributeSet::get(NRetTy->getContext(), RAttrs));
if (PAL.hasAttributes(AttributeSet::FunctionIndex))
AttributesVec.push_back(AttributeSet::get(F->getContext(),
PAL.getFnAttributes()));
// Reconstruct the AttributesList based on the vector we constructed.
AttributeSet NewPAL = AttributeSet::get(F->getContext(), AttributesVec);
// Create the new function type based on the recomputed parameters.
FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg());
// No change?
if (NFTy == FTy)
return false;
// Create the new function body and insert it into the module...
Function *NF = Function::Create(NFTy, F->getLinkage());
NF->copyAttributesFrom(F);
NF->setAttributes(NewPAL);
// Insert the new function before the old function, so we won't be processing
// it again.
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 a smaller number of arguments into the new function.
//
std::vector<Value*> Args;
while (!F->use_empty()) {
CallSite CS(F->user_back());
Instruction *Call = CS.getInstruction();
AttributesVec.clear();
const AttributeSet &CallPAL = CS.getAttributes();
// The call return attributes.
AttributeSet RAttrs = CallPAL.getRetAttributes();
// Adjust in case the function was changed to return void.
RAttrs = RAttrs.removeAttributes(NRetTy->getContext(),
AttributeSet::ReturnIndex,
AttributeFuncs::typeIncompatible(NF->getReturnType()));
if (RAttrs.hasAttributes(AttributeSet::ReturnIndex))
AttributesVec.push_back(AttributeSet::get(NF->getContext(), RAttrs));
// Declare these outside of the loops, so we can reuse them for the second
// loop, which loops the varargs.
CallSite::arg_iterator I = CS.arg_begin();
unsigned i = 0;
// Loop over those operands, corresponding to the normal arguments to the
// original function, and add those that are still alive.
for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i)
if (ArgAlive[i]) {
Args.push_back(*I);
// Get original parameter attributes, but skip return attributes.
if (CallPAL.hasAttributes(i + 1)) {
AttrBuilder B(CallPAL, i + 1);
// If the return type has changed, then get rid of 'returned' on the
// call site. The alternative is to make all 'returned' attributes on
// call sites keep the return value alive just like 'returned'
// attributes on function declaration but it's less clearly a win
// and this is not an expected case anyway
if (NRetTy != RetTy && B.contains(Attribute::Returned))
B.removeAttribute(Attribute::Returned);
AttributesVec.
push_back(AttributeSet::get(F->getContext(), Args.size(), B));
}
}
// Push any varargs arguments on the list. Don't forget their attributes.
for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) {
Args.push_back(*I);
if (CallPAL.hasAttributes(i + 1)) {
AttrBuilder B(CallPAL, i + 1);
AttributesVec.
push_back(AttributeSet::get(F->getContext(), Args.size(), B));
}
}
if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
AttributesVec.push_back(AttributeSet::get(Call->getContext(),
CallPAL.getFnAttributes()));
// Reconstruct the AttributesList based on the vector we constructed.
AttributeSet NewCallPAL = AttributeSet::get(F->getContext(), AttributesVec);
Instruction *New;
if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
Args, "", Call->getParent());
cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
cast<InvokeInst>(New)->setAttributes(NewCallPAL);
} else {
New = CallInst::Create(NF, Args, "", Call);
cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
cast<CallInst>(New)->setAttributes(NewCallPAL);
if (cast<CallInst>(Call)->isTailCall())
cast<CallInst>(New)->setTailCall();
}
New->setDebugLoc(Call->getDebugLoc());
Args.clear();
if (!Call->use_empty()) {
if (New->getType() == Call->getType()) {
// Return type not changed? Just replace users then.
Call->replaceAllUsesWith(New);
New->takeName(Call);
} else if (New->getType()->isVoidTy()) {
// Our return value has uses, but they will get removed later on.
// Replace by null for now.
if (!Call->getType()->isX86_MMXTy())
Call->replaceAllUsesWith(Constant::getNullValue(Call->getType()));
} else {
assert((RetTy->isStructTy() || RetTy->isArrayTy()) &&
"Return type changed, but not into a void. The old return type"
" must have been a struct or an array!");
Instruction *InsertPt = Call;
if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
BasicBlock *NewEdge = SplitEdge(New->getParent(), II->getNormalDest());
InsertPt = &*NewEdge->getFirstInsertionPt();
}
// We used to return a struct or array. Instead of doing smart stuff
// with all the uses, we will just rebuild it using extract/insertvalue
// chaining and let instcombine clean that up.
//
// Start out building up our return value from undef
Value *RetVal = UndefValue::get(RetTy);
for (unsigned i = 0; i != RetCount; ++i)
if (NewRetIdxs[i] != -1) {
Value *V;
if (RetTypes.size() > 1)
// We are still returning a struct, so extract the value from our
// return value
V = ExtractValueInst::Create(New, NewRetIdxs[i], "newret",
InsertPt);
else
// We are now returning a single element, so just insert that
V = New;
// Insert the value at the old position
RetVal = InsertValueInst::Create(RetVal, V, i, "oldret", InsertPt);
}
// Now, replace all uses of the old call instruction with the return
// struct we built
Call->replaceAllUsesWith(RetVal);
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.
i = 0;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
I2 = NF->arg_begin(); I != E; ++I, ++i)
if (ArgAlive[i]) {
// If this is a live argument, move the name and users over to the new
// version.
I->replaceAllUsesWith(&*I2);
I2->takeName(&*I);
++I2;
} else {
// If this argument is dead, replace any uses of it with null constants
// (these are guaranteed to become unused later on).
if (!I->getType()->isX86_MMXTy())
I->replaceAllUsesWith(Constant::getNullValue(I->getType()));
}
// If we change the return value of the function we must rewrite any return
// instructions. Check this now.
if (F->getReturnType() != NF->getReturnType())
for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB)
if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
Value *RetVal;
if (NFTy->getReturnType()->isVoidTy()) {
RetVal = nullptr;
} else {
assert(RetTy->isStructTy() || RetTy->isArrayTy());
// The original return value was a struct or array, insert
// extractvalue/insertvalue chains to extract only the values we need
// to return and insert them into our new result.
// This does generate messy code, but we'll let it to instcombine to
// clean that up.
Value *OldRet = RI->getOperand(0);
// Start out building up our return value from undef
RetVal = UndefValue::get(NRetTy);
for (unsigned i = 0; i != RetCount; ++i)
if (NewRetIdxs[i] != -1) {
ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i,
"oldret", RI);
if (RetTypes.size() > 1) {
// We're still returning a struct, so reinsert the value into
// our new return value at the new index
RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i],
"newret", RI);
} else {
// We are now only returning a simple value, so just return the
// extracted value.
RetVal = EV;
}
}
}
// Replace the return instruction with one returning the new return
// value (possibly 0 if we became void).
ReturnInst::Create(F->getContext(), RetVal, RI);
BB->getInstList().erase(RI);
}
// Patch the pointer to LLVM function in debug info descriptor.
NF->setSubprogram(F->getSubprogram());
// Now that the old function is dead, delete it.
F->eraseFromParent();
return true;
}
bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
bool &TailCallsAreMarkedTail,
SmallVector<PHINode*, 8> &ArgumentPHIs,
bool CannotTailCallElimCallsMarkedTail) {
BasicBlock *BB = Ret->getParent();
Function *F = BB->getParent();
if (&BB->front() == Ret) // Make sure there is something before the ret...
return false;
// If the return is in the entry block, then making this transformation would
// turn infinite recursion into an infinite loop. This transformation is ok
// in theory, but breaks some code like:
// double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
// disable this xform in this case, because the code generator will lower the
// call to fabs into inline code.
if (BB == &F->getEntryBlock())
return false;
// Scan backwards from the return, checking to see if there is a tail call in
// this block. If so, set CI to it.
CallInst *CI;
BasicBlock::iterator BBI = Ret;
while (1) {
CI = dyn_cast<CallInst>(BBI);
if (CI && CI->getCalledFunction() == F)
break;
if (BBI == BB->begin())
return false; // Didn't find a potential tail call.
--BBI;
}
// If this call is marked as a tail call, and if there are dynamic allocas in
// the function, we cannot perform this optimization.
if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
return false;
// If we are introducing accumulator recursion to eliminate associative
// operations after the call instruction, this variable contains the initial
// value for the accumulator. If this value is set, we actually perform
// accumulator recursion elimination instead of simple tail recursion
// elimination.
Value *AccumulatorRecursionEliminationInitVal = 0;
Instruction *AccumulatorRecursionInstr = 0;
// Ok, we found a potential tail call. We can currently only transform the
// tail call if all of the instructions between the call and the return are
// movable to above the call itself, leaving the call next to the return.
// Check that this is the case now.
for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI)
if (!CanMoveAboveCall(BBI, CI)) {
// If we can't move the instruction above the call, it might be because it
// is an associative operation that could be tranformed using accumulator
// recursion elimination. Check to see if this is the case, and if so,
// remember the initial accumulator value for later.
if ((AccumulatorRecursionEliminationInitVal =
CanTransformAccumulatorRecursion(BBI, CI))) {
// Yes, this is accumulator recursion. Remember which instruction
// accumulates.
AccumulatorRecursionInstr = BBI;
} else {
return false; // Otherwise, we cannot eliminate the tail recursion!
}
}
// We can only transform call/return pairs that either ignore the return value
// of the call and return void, ignore the value of the call and return a
// constant, return the value returned by the tail call, or that are being
// accumulator recursion variable eliminated.
if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
!isa<UndefValue>(Ret->getReturnValue()) &&
AccumulatorRecursionEliminationInitVal == 0 &&
!getCommonReturnValue(Ret, CI))
return false;
// OK! We can transform this tail call. If this is the first one found,
// create the new entry block, allowing us to branch back to the old entry.
if (OldEntry == 0) {
OldEntry = &F->getEntryBlock();
BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
NewEntry->takeName(OldEntry);
OldEntry->setName("tailrecurse");
BranchInst::Create(OldEntry, NewEntry);
// If this tail call is marked 'tail' and if there are any allocas in the
// entry block, move them up to the new entry block.
TailCallsAreMarkedTail = CI->isTailCall();
if (TailCallsAreMarkedTail)
// Move all fixed sized allocas from OldEntry to NewEntry.
for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
NEBI = NewEntry->begin(); OEBI != E; )
if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
if (isa<ConstantInt>(AI->getArraySize()))
AI->moveBefore(NEBI);
// Now that we have created a new block, which jumps to the entry
// block, insert a PHI node for each argument of the function.
// For now, we initialize each PHI to only have the real arguments
// which are passed in.
Instruction *InsertPos = OldEntry->begin();
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
I != E; ++I) {
PHINode *PN = PHINode::Create(I->getType(),
I->getName() + ".tr", InsertPos);
I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
PN->addIncoming(I, NewEntry);
ArgumentPHIs.push_back(PN);
}
}
// If this function has self recursive calls in the tail position where some
// are marked tail and some are not, only transform one flavor or another. We
// have to choose whether we move allocas in the entry block to the new entry
// block or not, so we can't make a good choice for both. NOTE: We could do
// slightly better here in the case that the function has no entry block
// allocas.
if (TailCallsAreMarkedTail && !CI->isTailCall())
return false;
// Ok, now that we know we have a pseudo-entry block WITH all of the
// required PHI nodes, add entries into the PHI node for the actual
// parameters passed into the tail-recursive call.
for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i)
ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB);
// If we are introducing an accumulator variable to eliminate the recursion,
// do so now. Note that we _know_ that no subsequent tail recursion
// eliminations will happen on this function because of the way the
// accumulator recursion predicate is set up.
//
if (AccumulatorRecursionEliminationInitVal) {
Instruction *AccRecInstr = AccumulatorRecursionInstr;
// Start by inserting a new PHI node for the accumulator.
PHINode *AccPN = PHINode::Create(AccRecInstr->getType(), "accumulator.tr",
OldEntry->begin());
// Loop over all of the predecessors of the tail recursion block. For the
// real entry into the function we seed the PHI with the initial value,
// computed earlier. For any other existing branches to this block (due to
// other tail recursions eliminated) the accumulator is not modified.
// Because we haven't added the branch in the current block to OldEntry yet,
// it will not show up as a predecessor.
for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry);
PI != PE; ++PI) {
if (*PI == &F->getEntryBlock())
AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI);
else
AccPN->addIncoming(AccPN, *PI);
}
// Add an incoming argument for the current block, which is computed by our
// associative accumulator instruction.
AccPN->addIncoming(AccRecInstr, BB);
// Next, rewrite the accumulator recursion instruction so that it does not
// use the result of the call anymore, instead, use the PHI node we just
// inserted.
AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
// Finally, rewrite any return instructions in the program to return the PHI
// node instead of the "initval" that they do currently. This loop will
// actually rewrite the return value we are destroying, but that's ok.
for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
RI->setOperand(0, AccPN);
++NumAccumAdded;
}
// Now that all of the PHI nodes are in place, remove the call and
// ret instructions, replacing them with an unconditional branch.
BranchInst::Create(OldEntry, Ret);
BB->getInstList().erase(Ret); // Remove return.
BB->getInstList().erase(CI); // Remove call.
++NumEliminated;
return true;
}
