bool LLPE::runOnModule(Module &M) {
vector<int> argv = readInputFile();
for (Module::iterator F = M.begin(), F_end = M.end(); F != F_end; ++F) {
for (Function::arg_iterator A = F->arg_begin(), A_end = F->arg_end(); A != A_end; ++A) {
//Search for variables referencing argv
if (A->getName() == "argv") {
//Iterate through uses of argv
for (Value::use_iterator U = A->use_begin(), U_end = A->use_end(); U != U_end; ++U) {
Instruction *User = dyn_cast<Instruction>(*U);
StoreInst *SI = dyn_cast<StoreInst>(User);
AllocaInst *OrigAlloca = dyn_cast<AllocaInst>(SI->getOperand(1));
for (Value::use_iterator U2 = OrigAlloca->use_begin(), U2_end = OrigAlloca->use_end(); U2 != U2_end; ++U2) {
Instruction *User2 = dyn_cast<Instruction>(*U2);
for (Value::use_iterator U3 = User2->use_begin(), U3_end = OrigAlloca->use_end(); U3 != U3_end; ++U3) {
searchForStoreInstruction(dyn_cast<Instruction>(*U3)->getParent(), argv);
}
}
}
}
}
}
return true;
}
void ExecutionState::dumpStack(llvm::raw_ostream &out) const {
unsigned idx = 0;
const KInstruction *target = prevPC;
for (ExecutionState::stack_ty::const_reverse_iterator
it = stack.rbegin(), ie = stack.rend();
it != ie; ++it) {
const StackFrame &sf = *it;
Function *f = sf.kf->function;
const InstructionInfo &ii = *target->info;
out << "\t#" << idx++;
std::stringstream AssStream;
AssStream << std::setw(8) << std::setfill('0') << ii.assemblyLine;
out << AssStream.str();
out << " in " << f->getName().str() << " (";
// Yawn, we could go up and print varargs if we wanted to.
unsigned index = 0;
for (Function::arg_iterator ai = f->arg_begin(), ae = f->arg_end();
ai != ae; ++ai) {
if (ai!=f->arg_begin()) out << ", ";
out << ai->getName().str();
// XXX should go through function
ref<Expr> value = sf.locals[sf.kf->getArgRegister(index++)].value;
if (value.get() && isa<ConstantExpr>(value))
out << "=" << value;
}
out << ")";
if (ii.file != "")
out << " at " << ii.file << ":" << ii.line;
out << "\n";
target = sf.caller;
}
}
void StackTrace::dump(std::ostream &out) const {
unsigned idx = 0;
for (stack_t::const_iterator it = contents.begin(); it != contents.end(); it++) {
Function *f = it->first.first->function;
const InstructionInfo &ii = *it->first.second->info;
out << "\t#" << idx++
<< " " << std::setw(8) << std::setfill('0') << ii.assemblyLine
<< " in " << f->getName().str() << " (";
unsigned index = 0;
for (Function::arg_iterator ai = f->arg_begin(), ae = f->arg_end();
ai != ae; ++ai) {
if (ai!=f->arg_begin()) out << ", ";
out << ai->getName().str();
// XXX should go through function
ref<Expr> value = it->second[index++];
if (isa<ConstantExpr>(value))
out << "=" << value;
}
out << ")";
if (ii.file != "")
out << " at " << ii.file << ":" << ii.line;
out << "\n";
}
}
int get_arg_count(const char *name, Function *op)
{
uint8_t args_present[MAX_ARGS];
int nb_args, i, n;
const char *p;
for(i = 0;i < MAX_ARGS; i++)
args_present[i] = 0;
// compute the number of arguments by looking at
// the uses of the op parameters
for (Function::arg_iterator i = op->arg_begin(), e = op->arg_end(); i != e; ++i) {
const char *tmpArgName = i->getName().c_str();
char *argName = (char *) malloc(strlen(tmpArgName + 1));
strcpy(argName, tmpArgName);
if (strstart(argName, "__op_param", &p)) {
if (i->hasNUsesOrMore(1)) {
n = strtoul(p, NULL, 10);
if (n > MAX_ARGS)
error("too many arguments in %s", name);
args_present[n - 1] = 1;
}
}
}
for (i = 1; i <= MAX_ARGS; i++) {
char *funcName = (char *) malloc(20);
sprintf(funcName, "__op_gen_label%d", i);
Function *f = ops->getFunction(std::string(funcName));
if (f != NULL) {
for (Function::use_iterator j = f->use_begin(), e = f->use_end(); j != e; ++j) {
if (Instruction *inst = dyn_cast<Instruction>(*j)) {
if (inst->getParent()->getParent() == op) {
args_present[i - 1] = 1;
}
}
}
} else {
std::cout << "symbol not found\n";
}
}
nb_args = 0;
while (nb_args < MAX_ARGS && args_present[nb_args])
nb_args++;
for(i = nb_args; i < MAX_ARGS; i++) {
if (args_present[i])
error("inconsistent argument numbering in %s", name);
}
return nb_args;
}
/// 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);
}
}
}
/* for the vi and lsup, find their corresponding argument*/
Value *FindFunctionArgumentOfInstr(Instruction *I, Function *F) {
int notFound = 1;
Value *via = 0;
Function::arg_iterator FI = F->arg_begin(), FE = F->arg_end();
while (FI != FE && notFound) {
if (FI->getName() == I->getName()) {
via = &*FI;
notFound = 0;
}
++FI;
}
return via;
}
// Maybe make a clone, if a clone is made, return a pointer to it, if a clone
// was not made return nullptr.
Function *CSDataRando::makeFunctionClone(Function *F) {
// Now we know how many arguments need to be passed, so we make the clones
FuncInfo &FI = FunctionInfo[F];
if (FI.ArgNodes.size() == 0) {
// No additional args to pass, no need to clone.
return nullptr;
}
// Determine the type of the new function, we insert the new parameters for
// the masks after the normal arguments, but before any va_args
Type *MaskTy = TypeBuilder<mask_t, false>::get(F->getContext());
FunctionType *OldFuncTy = F->getFunctionType();
std::vector<Type*> ArgTys;
ArgTys.insert(ArgTys.end(), OldFuncTy->param_begin(), OldFuncTy->param_end());
ArgTys.insert(ArgTys.end(), FI.ArgNodes.size(), MaskTy);
FunctionType *CloneFuncTy = FunctionType::get(OldFuncTy->getReturnType(), ArgTys, OldFuncTy->isVarArg());
Function *Clone = Function::Create(CloneFuncTy, Function::InternalLinkage, F->getName() + "_CONTEXT_SENSITIVE");
F->getParent()->getFunctionList().insert(F->getIterator(), Clone);
Function::arg_iterator CI = Clone->arg_begin(), CE = Clone->arg_end();
// Map the old arguments to the clone arguments and set the name of the
// clone arguments the same as the original.
for (Function::arg_iterator i = F->arg_begin(), e = F->arg_end(); i != e && CI != CE; i++, CI++) {
FI.OldToNewMap[&*i] = &*CI;
CI->setName(i->getName());
}
// Set the name of the arg masks and associate them with the nodes they are
// the masks for.
for (unsigned i = 0, e = FI.ArgNodes.size(); i != e; ++i, ++CI) {
CI->setName("arg_mask");
FI.ArgMaskMap[FI.ArgNodes[i]] = &*CI;
}
SmallVector<ReturnInst*, 8> Returns;
CloneFunctionInto(Clone, F, FI.OldToNewMap, false, Returns);
Clone->setCallingConv(F->getCallingConv());
// Invert OldToNewMap
for (auto I : FI.OldToNewMap) {
FI.NewToOldMap[I.second] = I.first;
}
NumClones++;
return Clone;
}
void HeterotbbTransform::copy_function (Function* NF, Function* F) {
DenseMap<const Value*, Value *> ValueMap;
// Get the names of the parameters for old function
for(Function::arg_iterator FI = F->arg_begin(), FE=F->arg_end(), DI=NF->arg_begin(); FE!=FI; ++FI,++DI) {
DI->setName(FI->getName());
ValueMap[FI]=DI;
}
for (Function::const_iterator BI=F->begin(),BE = F->end(); BI != BE; ++BI) {
const BasicBlock &FBB = *BI;
BasicBlock *NFBB = BasicBlock::Create(FBB.getContext(), "", NF);
ValueMap[&FBB] = NFBB;
if (FBB.hasName()) {
NFBB->setName(FBB.getName());
//DEBUG(dbgs()<<NFBB->getName()<<"\n");
}
for (BasicBlock::const_iterator II = FBB.begin(), IE = FBB.end(); II != IE; ++II) {
Instruction *NFInst = II->clone(/*F->getContext()*/);
if (II->hasName()) NFInst->setName(II->getName());
const Instruction *FInst = &(*II);
rewrite_instruction((Instruction *)FInst, NFInst, ValueMap);
NFBB->getInstList().push_back(NFInst);
ValueMap[II] = NFInst;
}
}
// Remap the instructions again to take care of forward jumps
for (Function::iterator BB = NF->begin(), BE=NF->end(); BB != BE; ++ BB) {
for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II) {
int opIdx = 0;
//DEBUG(dbgs()<<*II<<"\n");
for (User::op_iterator i = II->op_begin(), e = II->op_end(); i != e; ++i, opIdx++) {
Value *V = *i;
if (ValueMap[V] != NULL) {
II->setOperand(opIdx, ValueMap[V]);
}
}
}
}
//NF->dump();
}
/// 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);
}
}
}
/////////////////////////
//dupFuncArgu() //
/////////////////////////
void InsDuplica::dupFuncArgu(Function &F) {
BasicBlock *firstBB = F.begin();
Instruction *firstI = firstBB->begin();
for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end();
AI != E; ++AI) {
arguSet.insert(AI);
//if argument is used only once or all on the same block,
//do not dup it.
if (AI->hasOneUse()) {
valueMap[AI] = AI;
} else {
//make copy of cast
Type *arguType = AI->getType();
//FIXME unknow signed or not
CastInst* newCast = CastInst::CreateIntegerCast(AI, arguType, 1, AI->getName()+"_dup", firstI);
//CastInst *newCast = new CastInst(AI, arguType, AI->getName()+"_dup", firstI);
valueMap[AI] = newCast;
}
}
}
/// 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);
}
}
// linkFunctionBody - Copy the source function over into the dest function and
// fix up references to values. At this point we know that Dest is an external
// function, and that Src is not.
void ModuleLinker::linkFunctionBody(Function *Dst, Function *Src) {
assert(Src && Dst && Dst->isDeclaration() && !Src->isDeclaration());
// Go through and convert function arguments over, remembering the mapping.
Function::arg_iterator DI = Dst->arg_begin();
for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end();
I != E; ++I, ++DI) {
DI->setName(I->getName()); // Copy the name over.
// Add a mapping to our mapping.
ValueMap[I] = DI;
}
if (Mode == Linker::DestroySource) {
// Splice the body of the source function into the dest function.
Dst->getBasicBlockList().splice(Dst->end(), Src->getBasicBlockList());
// At this point, all of the instructions and values of the function are now
// copied over. The only problem is that they are still referencing values in
// the Source function as operands. Loop through all of the operands of the
// functions and patch them up to point to the local versions.
for (Function::iterator BB = Dst->begin(), BE = Dst->end(); BB != BE; ++BB)
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
RemapInstruction(I, ValueMap, RF_IgnoreMissingEntries, &TypeMap);
} else {
// Clone the body of the function into the dest function.
SmallVector<ReturnInst*, 8> Returns; // Ignore returns.
CloneFunctionInto(Dst, Src, ValueMap, false, Returns, "", NULL, &TypeMap);
}
// There is no need to map the arguments anymore.
for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end();
I != E; ++I)
ValueMap.erase(I);
}
// 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);
}
}
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Clone functions that take GEPs as arguments
//
// Inputs:
// M - A reference to the LLVM module to transform
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool GEPExprArgs::runOnModule(Module& M) {
bool changed;
do {
changed = false;
for (Module::iterator F = M.begin(); F != M.end(); ++F){
for (Function::iterator B = F->begin(), FE = F->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
CallInst *CI = dyn_cast<CallInst>(I++);
if(!CI)
continue;
if(CI->hasByValArgument())
continue;
// if the GEP calls a function, that is externally defined,
// or might be changed, ignore this call site.
Function *F = CI->getCalledFunction();
if (!F || (F->isDeclaration() || F->mayBeOverridden()))
continue;
if(F->hasStructRetAttr())
continue;
if(F->isVarArg())
continue;
// find the argument we must replace
Function::arg_iterator ai = F->arg_begin(), ae = F->arg_end();
unsigned argNum = 1;
for(; argNum < CI->getNumOperands();argNum++, ++ai) {
if(ai->use_empty())
continue;
if (isa<GEPOperator>(CI->getOperand(argNum)))
break;
}
// if no argument was a GEP operator to be changed
if(ai == ae)
continue;
GEPOperator *GEP = dyn_cast<GEPOperator>(CI->getOperand(argNum));
if(!GEP->hasAllConstantIndices())
continue;
// Construct the new Type
// Appends the struct Type at the beginning
std::vector<Type*>TP;
TP.push_back(GEP->getPointerOperand()->getType());
for(unsigned c = 1; c < CI->getNumOperands();c++) {
TP.push_back(CI->getOperand(c)->getType());
}
//return type is same as that of original instruction
FunctionType *NewFTy = FunctionType::get(CI->getType(), TP, false);
Function *NewF;
numSimplified++;
if(numSimplified > 800)
return true;
NewF = Function::Create(NewFTy,
GlobalValue::InternalLinkage,
F->getName().str() + ".TEST",
&M);
Function::arg_iterator NI = NewF->arg_begin();
NI->setName("GEParg");
++NI;
ValueToValueMapTy ValueMap;
for (Function::arg_iterator II = F->arg_begin(); NI != NewF->arg_end(); ++II, ++NI) {
ValueMap[II] = NI;
NI->setName(II->getName());
NI->addAttr(F->getAttributes().getParamAttributes(II->getArgNo() + 1));
}
NewF->setAttributes(NewF->getAttributes().addAttr(
0, F->getAttributes().getRetAttributes()));
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
CloneFunctionInto(NewF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NewF->arg_begin(),
ae= NewF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NewF->setAttributes(NewF->getAttributes().addAttr(
~0, F->getAttributes().getFnAttributes()));
//Get the point to insert the GEP instr.
SmallVector<Value*, 8> Ops(CI->op_begin()+1, CI->op_end());
Instruction *InsertPoint;
for (BasicBlock::iterator insrt = NewF->front().begin();
isa<AllocaInst>(InsertPoint = insrt); ++insrt) {;}
NI = NewF->arg_begin();
SmallVector<Value*, 8> Indices;
Indices.append(GEP->op_begin()+1, GEP->op_end());
GetElementPtrInst *GEP_new = GetElementPtrInst::Create(cast<Value>(NI),
Indices,
"", InsertPoint);
fargs.at(argNum)->replaceAllUsesWith(GEP_new);
unsigned j = argNum + 1;
for(; j < CI->getNumOperands();j++) {
if(CI->getOperand(j) == GEP)
fargs.at(j)->replaceAllUsesWith(GEP_new);
}
SmallVector<AttributeWithIndex, 8> AttributesVec;
// Get the initial attributes of the call
AttrListPtr CallPAL = CI->getAttributes();
Attributes RAttrs = CallPAL.getRetAttributes();
Attributes FnAttrs = CallPAL.getFnAttributes();
if (RAttrs)
AttributesVec.push_back(AttributeWithIndex::get(0, RAttrs));
SmallVector<Value*, 8> Args;
Args.push_back(GEP->getPointerOperand());
for(unsigned j =1;j<CI->getNumOperands();j++) {
Args.push_back(CI->getOperand(j));
// position in the AttributesVec
if (Attributes Attrs = CallPAL.getParamAttributes(j))
AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
}
// Create the new attributes vec.
if (FnAttrs != Attribute::None)
AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
AttrListPtr NewCallPAL = AttrListPtr::get(AttributesVec.begin(),
AttributesVec.end());
CallInst *CallI = CallInst::Create(NewF,Args,"", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
CI->replaceAllUsesWith(CallI);
CI->eraseFromParent();
changed = true;
}
}
}
} while(changed);
return true;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// If a function returns a struct, make it return
// a pointer to the struct.
//
// Inputs:
// M - A reference to the LLVM module to transform
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool StructRet::runOnModule(Module& M) {
const llvm::DataLayout targetData(&M);
std::vector<Function*> worklist;
for (Module::iterator I = M.begin(); I != M.end(); ++I)
if (!I->mayBeOverridden()) {
if(I->hasAddressTaken())
continue;
if(I->getReturnType()->isStructTy()) {
worklist.push_back(I);
}
}
while(!worklist.empty()) {
Function *F = worklist.back();
worklist.pop_back();
Type *NewArgType = F->getReturnType()->getPointerTo();
// Construct the new Type
std::vector<Type*>TP;
TP.push_back(NewArgType);
for (Function::arg_iterator ii = F->arg_begin(), ee = F->arg_end();
ii != ee; ++ii) {
TP.push_back(ii->getType());
}
FunctionType *NFTy = FunctionType::get(F->getReturnType(), TP, F->isVarArg());
// Create the new function body and insert it into the module.
Function *NF = Function::Create(NFTy,
F->getLinkage(),
F->getName(), &M);
ValueToValueMapTy ValueMap;
Function::arg_iterator NI = NF->arg_begin();
NI->setName("ret");
++NI;
for (Function::arg_iterator II = F->arg_begin(); II != F->arg_end(); ++II, ++NI) {
ValueMap[II] = NI;
NI->setName(II->getName());
AttributeSet attrs = F->getAttributes().getParamAttributes(II->getArgNo() + 1);
if (!attrs.isEmpty())
NI->addAttr(attrs);
}
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
if (!F->isDeclaration())
CloneFunctionInto(NF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NF->arg_begin(),
ae= NF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NF->setAttributes(NF->getAttributes().addAttributes(
M.getContext(), 0, F->getAttributes().getRetAttributes()));
NF->setAttributes(NF->getAttributes().addAttributes(
M.getContext(), ~0, F->getAttributes().getFnAttributes()));
for (Function::iterator B = NF->begin(), FE = NF->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
ReturnInst * RI = dyn_cast<ReturnInst>(I++);
if(!RI)
continue;
LoadInst *LI = dyn_cast<LoadInst>(RI->getOperand(0));
assert(LI && "Return should be preceded by a load instruction");
IRBuilder<> Builder(RI);
Builder.CreateMemCpy(fargs.at(0),
LI->getPointerOperand(),
targetData.getTypeStoreSize(LI->getType()),
targetData.getPrefTypeAlignment(LI->getType()));
}
}
for(Value::use_iterator ui = F->use_begin(), ue = F->use_end();
ui != ue; ) {
CallInst *CI = dyn_cast<CallInst>(*ui++);
if(!CI)
continue;
if(CI->getCalledFunction() != F)
continue;
if(CI->hasByValArgument())
continue;
AllocaInst *AllocaNew = new AllocaInst(F->getReturnType(), 0, "", CI);
SmallVector<Value*, 8> Args;
//this should probably be done in a different manner
AttributeSet NewCallPAL=AttributeSet();
// Get the initial attributes of the call
AttributeSet CallPAL = CI->getAttributes();
AttributeSet RAttrs = CallPAL.getRetAttributes();
AttributeSet FnAttrs = CallPAL.getFnAttributes();
if (!RAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),0, RAttrs);
Args.push_back(AllocaNew);
for(unsigned j = 0; j < CI->getNumOperands()-1; j++) {
Args.push_back(CI->getOperand(j));
// position in the NewCallPAL
AttributeSet Attrs = CallPAL.getParamAttributes(j);
if (!Attrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),Args.size(), Attrs);
}
// Create the new attributes vec.
if (!FnAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),~0, FnAttrs);
CallInst *CallI = CallInst::Create(NF, Args, "", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
LoadInst *LI = new LoadInst(AllocaNew, "", CI);
CI->replaceAllUsesWith(LI);
CI->eraseFromParent();
}
if(F->use_empty())
F->eraseFromParent();
}
return true;
}
// generate code for ops which return a void value
void gen_code_void_op(const char *name,
FILE *outfile, Function *op)
{
uint8_t args_present[MAX_ARGS];
int nb_args, i, n;
const char *p;
for(i = 0;i < MAX_ARGS; i++)
args_present[i] = 0;
// compute the number of arguments by looking at
// the uses of the op parameters
for (Function::arg_iterator i = op->arg_begin(), e = op->arg_end(); i != e; ++i) {
const char *tmpArgName = i->getName().c_str();
char *argName = (char *) malloc(strlen(tmpArgName + 1));
strcpy(argName, tmpArgName);
if (strstart(argName, "__op_param", &p)) {
if (i->hasNUsesOrMore(1)) {
n = strtoul(p, NULL, 10);
if (n > MAX_ARGS)
error("too many arguments in %s", name);
args_present[n - 1] = 1;
}
}
}
nb_args = 0;
while (nb_args < MAX_ARGS && args_present[nb_args])
nb_args++;
for(i = nb_args; i < MAX_ARGS; i++) {
if (args_present[i])
error("inconsistent argument numbering in %s", name);
}
// add local variables for op parameters
if (nb_args > 0) {
fprintf(outfile, " long ");
for(i = 0; i < nb_args; i++) {
if (i != 0)
fprintf(outfile, ", ");
fprintf(outfile, "param%d", i + 1);
}
fprintf(outfile, ";\n");
}
// load parameters in variables
for(i = 0; i < nb_args; i++) {
fprintf(outfile, " param%d = *opparam_ptr++;\n", i + 1);
}
// load op parameters into the arguments of the call
// fprintf(outfile, "Value * args[MAX_ARGS];\n");
// for (i = 0; i < nb_args; i++) {
// fprintf(outfile, "args[%d] = ConstantInt::get(Type::Int32Ty, param%d);\n", i, i + 1);
// }
// for (i = nb_args; i < MAX_ARGS; i++) {
// fprintf(outfile, "args[%d] = zero;\n", i);
// }
// add call to micro op
//fprintf(outfile, " currCall = new CallInst(M->getFunction(\"%s\"), (Value **)&args, %d, \"\", currInst);\n", name, MAX_ARGS);
// load op parameters into the arguments of the call
fprintf(outfile, "std::vector<Value *> args;\n");
for (i = 0; i < nb_args; i++) {
fprintf(outfile, "args.push_back(ConstantInt::get(Type::Int32Ty, param%d));\n", i + 1);
}
for (i = nb_args; i < MAX_ARGS; i++) {
fprintf(outfile, "args.push_back(zero);\n");
}
// add call to micro op
fprintf(outfile, " currCall = new CallInst(M->getFunction(\"%s\"), args.begin(), args.end(), \"\", currInst);\n", name);
if (strcmp(name, "op_exit_tb") == 0) {
fprintf(outfile,
" currBB = new BasicBlock(\"\", tb);\n"
" currInst = new ReturnInst(fallThrough, currBB);\n"
);
} else if (strcmp(name, "op_goto_tb0")) {
// fprintf(outfile,
// " currBB = new BasicBlock(\"\", tb);\n"
// " currInst = new ReturnInst(tb0, currBB);\n"
// );
} else if (strcmp(name, "op_goto_tb1")) {
// fprintf(outfile,
// " currBB = new BasicBlock(\"\", tb);\n"
// " currInst = new ReturnInst(tb1, currBB);\n"
// );
}
fprintf(outfile, " InlineFunction(currCall);");
// inlining can change the basic block the return instruction belongs to
fprintf(outfile, " currBB = currInst->getParent();");
}
/// 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);
}
}
}
Function* HeterotbbTransform::create_new_join(Module &M,Function *join) {
//reduce->dump();
if(!templat) {
DEBUG(dbgs()<<"NO Template Found\n");
return NULL;
}
//join_main->dump();
DEBUG(dbgs()<<"Objext size array"<<object_size_hetero<<"\n");
//create a global with 64*object/4 size
GlobalVariable *gb = M.getGlobalVariable("opencl_kernel_join_name_local_arr",true);
//gb->dump();
Value *val=gb->getOperand(0);
//if(isa<ArrayType>(val->getType()))DEBUG(dbgs()<<"YES\n");
//since we are creating an integer array, the size gets divided by 4
// Do not make it a global variable -- make it a local variable with annotation for local
int local_size = 64*object_size_hetero;
/*const*/ ArrayType *arr= ArrayType::get(Type::getInt32Ty(M.getContext()),(64*object_size_hetero)/4);
/*vector<Constant *> Initializer;
APInt zero(32,0);
for(int i=0;i<(16*object_size_hetero);i++){
Initializer.push_back(ConstantInt::get(M.getContext(),zero));
}
Constant *init = ConstantArray::get(arr, Initializer);
GlobalVariable *new_gb = new GlobalVariable(M, arr, false, GlobalVariable::InternalLinkage,init, "__hetero_local_"+join->getName()+"__local__arr",gb,false,3);
new_gb->setAlignment(gb->getAlignment());
DEBUG(dbgs()<<"Global Created\n");
new_gb->dump();
*/
vector</*const*/ Type *> params;
int temp_size=0;
object_size_hetero=0;
// void join(class.name *,class.name *)
//re-write join
const FunctionType *FTy = join->getFunctionType();
Function::arg_iterator ArgI = join->arg_begin();
// class.name *
params.push_back(PointerType::get((dyn_cast<PointerType>(ArgI->getType())->getElementType()),3));
params.push_back(PointerType::get((dyn_cast<PointerType>(ArgI->getType())->getElementType()),3));
/*const*/ Type *RetTy = FTy->getReturnType();
FunctionType *NFty = FunctionType::get(RetTy,params, false);
Function *NF=Function::Create(NFty, join->getLinkage(), join->getName()+"_inline");
NF->copyAttributesFrom(join);
#if EXPLICIT_REWRITE
copy_function(NF,join);
#else
ValueToValueMapTy VMap;
for(Function::arg_iterator FI = join->arg_begin(), FE=join->arg_end(), DI=NF->arg_begin(); FE!=FI; ++FI,++DI) {
DI->setName(FI->getName());
VMap[FI]=DI;
}
CloneFunctionWithExistingBBInto(NF, NULL, join, VMap);
#endif
//NF->removeFnAttr(Attributes::get(NF->getContext(), Attribute::NoInline));
NF->addFnAttr(Attribute::AlwaysInline);
join->getParent()->getFunctionList().insert(join, NF);
params.clear();
const FunctionType *FTemp = templat->getFunctionType();
//create a new template
for(Function::arg_iterator FI = templat->arg_begin(), FE=templat->arg_end(); FE!=FI; ++FI) {
params.push_back(FI->getType());
}
// templat->replaceUsesOfWith(reduce,NF);
RetTy = FTy->getReturnType();
NFty = FunctionType::get(RetTy,params, false);
Function *templat_copy =Function::Create(NFty, join->getLinkage(), join->getName()+"_hetero");
templat_copy->copyAttributesFrom(templat);
#if EXPLICIT_REWRITE
copy_function(templat_copy,templat);
#else
ValueToValueMapTy VMapp;
for(Function::arg_iterator FI = templat->arg_begin(), FE=templat->arg_end(), DI=templat_copy->arg_begin(); FE!=FI; ++FI,++DI) {
DI->setName(FI->getName());
VMapp[FI]=DI;
}
CloneFunctionWithExistingBBInto(templat_copy, NULL, templat, VMapp);
#endif
/* create a local variable with the following type */
Function::iterator BI = templat_copy->begin();
BasicBlock::iterator II = BI->begin();
Instruction *insn = &(*II);
Constant *l_size = ConstantInt::get(Type::getInt32Ty(M.getContext()), local_size);
Instruction *new_gb_ = new AllocaInst(arr, l_size, gb->getAlignment(), "hetero_local", insn);
//new_gb_->dump();
Value *Elts[] = {MDString::get(M.getContext(), new_gb_->getName())};
MDNode *Node = MDNode::get(M.getContext(), Elts);
new_gb_->setMetadata("local",Node);
Instruction *new_gb= CastInst::Create(Instruction::BitCast, new_gb_,
PointerType::get(arr,3), "hetero_local_cast", insn);
//new_gb->dump();
Value *Elts1[] = {MDString::get(M.getContext(), new_gb->getName())};
MDNode *Node1 = MDNode::get(M.getContext(), Elts1);
new_gb->setMetadata("local_cast",Node1);
edit_template_function(M,templat_copy,NF,gb,new_gb);
templat->getParent()->getFunctionList().insert(templat, templat_copy);
return templat_copy;
}
/// 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;
}
void
RocketShip::processFunction(Function &F) {
// Everything is reset per function. Ideally we would just make
// this a per-function pass, but extended feature plans make
// applying this at the module level a better idea.
_nodeId = 0;
_blockId = 0;
_blocks.clear();
_pnodes.clear();
std::vector<BasicBlock*> blockList;
std::string functionLabel = F.getName();
std::string demangledLabel = getDemangledName(functionLabel);
if (demangledLabel == functionLabel ||
demangledLabel.length() == 0) {
functionLabel = F.getReturnType()->getDescription() + " " + functionLabel;
functionLabel = functionLabel + "(";
for (Function::arg_iterator arg = F.arg_begin();
arg != F.arg_end();
arg++) {
if (arg != F.arg_begin()) {
functionLabel = functionLabel + ", ";
}
functionLabel = functionLabel + arg->getType()->getDescription() + " " + std::string(arg->getName());
}
functionLabel = functionLabel + ")";
} else {
functionLabel = demangledLabel;
}
// Each block in the function needs to be processed and added to
// the mapping.
for (Function::iterator bblock = F.begin();
bblock != F.end();
bblock++) {
pBlock block(new Block(_nodeId++, bblock->getName()));
_blocks.insert(std::pair<BasicBlock*, pBlock>(bblock, block));
blockList.push_back(bblock);
if (bblock == F.begin()) {
pNode node(new Node(_nodeId++));
block->appendNode(node);
node->setNodeLabel(functionLabel);
node->setNodeType(Node::START);
}
processBlock(bblock, block);
}
// Each block needs to process its contained nodes and we need to
// keep a local copy of each node for later processing.
for (std::map<BasicBlock*, pBlock>::iterator it = _blocks.begin();
it != _blocks.end();
it++) {
it->second->processNodes(_blocks);
Nodes nodes = it->second->getNodes();
for (Nodes::iterator node = nodes.begin();
node != nodes.end();
node++) {
_pnodes.push_back(*node);
}
}
// Generates the function name and filename/output stream.
std::string functionIdentifier = F.getName();
char* result = cplus_demangle(functionIdentifier.c_str(), DMGL_ANSI|DMGL_PARAMS);
std::replace(functionIdentifier.begin(), functionIdentifier.end(), '.', '_');
_outputFile.open(std::string(functionIdentifier + ".dot").c_str());
_outputFile << "digraph " << functionIdentifier << " {\n";
// Emit each node to the output stream. This should be modified
// to have the node print itself out by passing in the output
// stream rather than calling a separate function
for (Nodes::iterator it = _pnodes.begin();
it != _pnodes.end();
it++) {
if ((*it) == NULL) {
continue;
}
// We only care about nodes with labels since they are what is
// actually presented.
if ((*it)->getNodeLabel().length() > 0) {
emitNode(&(*(*it)));
}
}
_outputFile << "}";
_outputFile.close();
}
// generate code for ops which return a void value
void gen_code_void_op(const char *name,
FILE *outfile, Function *op)
{
uint8_t args_present[MAX_ARGS];
int nb_args, i, n;
const char *p;
for(i = 0;i < MAX_ARGS; i++)
args_present[i] = 0;
// compute the number of arguments by looking at
// the uses of the op parameters
for (Function::arg_iterator i = op->arg_begin(), e = op->arg_end(); i != e; ++i) {
const char *tmpArgName = i->getName().c_str();
char *argName = (char *) malloc(strlen(tmpArgName + 1));
strcpy(argName, tmpArgName);
if (strstart(argName, "__op_param", &p)) {
if (i->hasNUsesOrMore(1)) {
n = strtoul(p, NULL, 10);
if (n > MAX_ARGS)
error("too many arguments in %s", name);
args_present[n - 1] = 1;
}
}
}
nb_args = 0;
while (nb_args < MAX_ARGS && args_present[nb_args])
nb_args++;
for(i = nb_args; i < MAX_ARGS; i++) {
if (args_present[i])
error("inconsistent argument numbering in %s", name);
}
// add local variables for op parameters
if (nb_args > 0) {
fprintf(outfile, " long ");
for(i = 0; i < nb_args; i++) {
if (i != 0)
fprintf(outfile, ", ");
fprintf(outfile, "param%d", i + 1);
}
fprintf(outfile, ";\n");
}
// load parameres in variables
for(i = 0; i < nb_args; i++) {
fprintf(outfile, " param%d = *opparam_ptr++;\n", i + 1);
}
// load op parameters into the arguments of the call
fprintf(outfile, "Value * args[MAX_ARGS];\n");
for (i = 0; i < nb_args; i++) {
fprintf(outfile, "args[%d] = ConstantInt::get(Type::Int32Ty, param%d);\n", i, i + 1);
}
for (i = nb_args; i < MAX_ARGS; i++) {
fprintf(outfile, "args[%d] = zero;\n", i);
}
// add call to micro op
fprintf(outfile, " currCall = new CallInst(M->getFunction(\"%s\"), (Value **)&args, %d, \"\", currBB);\n", name, MAX_ARGS);
fprintf(outfile, " InlineFunction(currCall);");
}
/*
* Clone a given function removing dead stores
*/
Function* DeadStoreEliminationPass::cloneFunctionWithoutDeadStore(Function *Fn,
Instruction* caller, std::string suffix) {
Function *NF = Function::Create(Fn->getFunctionType(), Fn->getLinkage());
NF->copyAttributesFrom(Fn);
// Copy the parameter names, to ease function inspection afterwards.
Function::arg_iterator NFArg = NF->arg_begin();
for (Function::arg_iterator Arg = Fn->arg_begin(), ArgEnd = Fn->arg_end();
Arg != ArgEnd; ++Arg, ++NFArg) {
NFArg->setName(Arg->getName());
}
// To avoid name collision, we should select another name.
NF->setName(Fn->getName() + suffix);
// Fill clone content
ValueToValueMapTy VMap;
SmallVector<ReturnInst*, 8> Returns;
Function::arg_iterator NI = NF->arg_begin();
for (Function::arg_iterator I = Fn->arg_begin();
NI != NF->arg_end(); ++I, ++NI) {
VMap[I] = NI;
}
CloneAndPruneFunctionInto(NF, Fn, VMap, false, Returns);
// Remove dead stores
std::set<Value*> deadArgs = deadArguments[caller];
std::set<Value*> removeStoresTo;
Function::arg_iterator NFArgIter = NF->arg_begin();
for (Function::arg_iterator FnArgIter = Fn->arg_begin(); FnArgIter !=
Fn->arg_end(); ++FnArgIter, ++NFArgIter) {
Value *FnArg = FnArgIter;
if (deadArgs.count(FnArg)) {
removeStoresTo.insert(NFArgIter);
}
}
std::vector<Instruction*> toRemove;
for (Function::iterator BB = NF->begin(); BB != NF->end(); ++BB) {
for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
Instruction *inst = I;
if (!isa<StoreInst>(inst)) continue;
StoreInst *SI = dyn_cast<StoreInst>(inst);
Value *ptrOp = SI->getPointerOperand();
if (removeStoresTo.count(ptrOp)) {
DEBUG(errs() << "will remove this store: " << *inst << "\n");
toRemove.push_back(inst);
}
}
}
for (std::vector<Instruction*>::iterator it = toRemove.begin();
it != toRemove.end(); ++it) {
Instruction* inst = *it;
inst->eraseFromParent();
RemovedStores++;
}
// Insert the clone function before the original
Fn->getParent()->getFunctionList().insert(Fn, NF);
return NF;
}
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;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Clone functions that take LoadInsts as arguments
//
// Inputs:
// M - A reference to the LLVM module to transform
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool LoadArgs::runOnModule(Module& M) {
std::map<std::pair<Function*, const Type * > , Function* > fnCache;
bool changed;
do {
changed = false;
for (Module::iterator Func = M.begin(); Func != M.end(); ++Func) {
for (Function::iterator B = Func->begin(), FE = Func->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
CallInst *CI = dyn_cast<CallInst>(I++);
if(!CI)
continue;
if(CI->hasByValArgument())
continue;
// if the CallInst calls a function, that is externally defined,
// or might be changed, ignore this call site.
Function *F = CI->getCalledFunction();
if (!F || (F->isDeclaration() || F->mayBeOverridden()))
continue;
if(F->hasStructRetAttr())
continue;
if(F->isVarArg())
continue;
// find the argument we must replace
Function::arg_iterator ai = F->arg_begin(), ae = F->arg_end();
unsigned argNum = 0;
for(; argNum < CI->getNumArgOperands();argNum++, ++ai) {
// do not care about dead arguments
if(ai->use_empty())
continue;
if(F->getAttributes().getParamAttributes(argNum).hasAttrSomewhere(Attribute::SExt) ||
F->getAttributes().getParamAttributes(argNum).hasAttrSomewhere(Attribute::ZExt))
continue;
if (isa<LoadInst>(CI->getArgOperand(argNum)))
break;
}
// if no argument was a GEP operator to be changed
if(ai == ae)
continue;
LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(argNum));
Instruction * InsertPt = &(Func->getEntryBlock().front());
AllocaInst *NewVal = new AllocaInst(LI->getType(), "",InsertPt);
StoreInst *Copy = new StoreInst(LI, NewVal);
Copy->insertAfter(LI);
/*if(LI->getParent() != CI->getParent())
continue;
// Also check that there is no store after the load.
// TODO: Check if the load/store do not alias.
BasicBlock::iterator bii = LI->getParent()->begin();
Instruction *BII = bii;
while(BII != LI) {
++bii;
BII = bii;
}
while(BII != CI) {
if(isa<StoreInst>(BII))
break;
++bii;
BII = bii;
}
if(isa<StoreInst>(bii)){
continue;
}*/
// Construct the new Type
// Appends the struct Type at the beginning
std::vector<Type*>TP;
for(unsigned c = 0; c < CI->getNumArgOperands();c++) {
if(c == argNum)
TP.push_back(LI->getPointerOperand()->getType());
TP.push_back(CI->getArgOperand(c)->getType());
}
//return type is same as that of original instruction
FunctionType *NewFTy = FunctionType::get(CI->getType(), TP, false);
numSimplified++;
//if(numSimplified > 1000)
//return true;
Function *NewF;
std::map<std::pair<Function*, const Type* > , Function* >::iterator Test;
Test = fnCache.find(std::make_pair(F, NewFTy));
if(Test != fnCache.end()) {
NewF = Test->second;
} else {
NewF = Function::Create(NewFTy,
GlobalValue::InternalLinkage,
F->getName().str() + ".TEST",
&M);
fnCache[std::make_pair(F, NewFTy)] = NewF;
Function::arg_iterator NI = NewF->arg_begin();
ValueToValueMapTy ValueMap;
unsigned count = 0;
for (Function::arg_iterator II = F->arg_begin(); NI != NewF->arg_end(); ++count, ++NI) {
if(count == argNum) {
NI->setName("LDarg");
continue;
}
ValueMap[II] = NI;
NI->setName(II->getName());
NI->addAttr(F->getAttributes().getParamAttributes(II->getArgNo() + 1));
++II;
}
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
CloneFunctionInto(NewF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NewF->arg_begin(),
ae= NewF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NewF->setAttributes(NewF->getAttributes().addAttributes(
F->getContext(), 0, F->getAttributes().getRetAttributes()));
NewF->setAttributes(NewF->getAttributes().addAttributes(
F->getContext(), ~0, F->getAttributes().getFnAttributes()));
//Get the point to insert the GEP instr.
Instruction *InsertPoint;
for (BasicBlock::iterator insrt = NewF->front().begin(); isa<AllocaInst>(InsertPoint = insrt); ++insrt) {;}
LoadInst *LI_new = new LoadInst(fargs.at(argNum), "", InsertPoint);
fargs.at(argNum+1)->replaceAllUsesWith(LI_new);
}
//this does not seem to be a good idea
AttributeSet NewCallPAL=AttributeSet();
// Get the initial attributes of the call
AttributeSet CallPAL = CI->getAttributes();
AttributeSet RAttrs = CallPAL.getRetAttributes();
AttributeSet FnAttrs = CallPAL.getFnAttributes();
if (!RAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),0, RAttrs);
SmallVector<Value*, 8> Args;
for(unsigned j =0;j<CI->getNumArgOperands();j++) {
if(j == argNum) {
Args.push_back(NewVal);
}
Args.push_back(CI->getArgOperand(j));
// position in the NewCallPAL
AttributeSet Attrs = CallPAL.getParamAttributes(j+1);
if (!Attrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),Args.size(), Attrs);
}
// Create the new attributes vec.
if (!FnAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),~0, FnAttrs);
CallInst *CallI = CallInst::Create(NewF,Args,"", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
CI->replaceAllUsesWith(CallI);
CI->eraseFromParent();
changed = true;
}
}
}
} while(changed);
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;
}
// =============================================================================
// createProcess
//
// Create a new function that contains a call to the old function.
// We inline the call in order to clone the old function's implementation.
// =============================================================================
Function *TLMBasicPassImpl::createProcess(Function *oldProc,
sc_core::sc_module *initiatorMod) {
LLVMContext &context = getGlobalContext();
IntegerType *intType;
if (this->is64Bit) {
intType = Type::getInt64Ty(context);
} else {
intType = Type::getInt32Ty(context);
}
// Retrieve a pointer to the initiator module
ConstantInt *initiatorModVal =
ConstantInt::getSigned(intType,reinterpret_cast<intptr_t>(initiatorMod));
FunctionType *funType = oldProc->getFunctionType();
Type *type = funType->getParamType(0);
IntToPtrInst *thisAddr =
new IntToPtrInst(initiatorModVal, type, "");
// Compute the type of the new function
FunctionType *oldProcType = oldProc->getFunctionType();
Value **argsBegin = new Value*[1];
Value **argsEnd = argsBegin;
*argsEnd++ = thisAddr;
const unsigned argsSize = argsEnd-argsBegin;
Value **args = argsBegin;
assert(oldProcType->getNumParams()==argsSize);
assert(!oldProc->isDeclaration());
std::vector<Type*> argTypes;
for (unsigned i = 0; i!=argsSize; ++i)
argTypes.push_back(oldProcType->getParamType(i));
FunctionType *newProcType =
FunctionType::get(oldProc->getReturnType(), ArrayRef<Type*>(argTypes), false);
// Create the new function
std::ostringstream id;
id << proc_counter++;
std::string name = oldProc->getName().str()+std::string("_clone_")+id.str();
Function *newProc =
Function::Create(newProcType, Function::ExternalLinkage, name, this->llvmMod);
assert(newProc->empty());
newProc->addFnAttr(Attributes::InlineHint);
{ // Set name of newfunc arguments and complete args
Function::arg_iterator nai = newProc->arg_begin();
Function::arg_iterator oai = oldProc->arg_begin();
for (unsigned i = 0; i!=argsSize; ++i, ++oai) {
nai->setName(oai->getName());
args[i] = nai;
++nai;
}
assert(nai==newProc->arg_end());
assert(oai==oldProc->arg_end());
}
// Create call to old function
BasicBlock *bb = BasicBlock::Create(context, "entry", newProc);
IRBuilder<> *irb = new IRBuilder<>(context);
irb->SetInsertPoint(bb);
CallInst *ci = irb->CreateCall(oldProc, ArrayRef<Value*>(argsBegin, argsEnd));
bb->getInstList().insert(ci, thisAddr);
if (ci->getType()->isVoidTy())
irb->CreateRetVoid();
else
irb->CreateRet(ci);
// The function should be valid now
verifyFunction(*newProc);
{ // Inline the call
DataLayout *td = new DataLayout(this->llvmMod);
InlineFunctionInfo i(NULL, td);
bool success = InlineFunction(ci, i);
assert(success);
verifyFunction(*newProc);
}
//newProc->dump();
return newProc;
}
// 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;
}
// MakeFunctionClone - If the specified function needs to be modified for pool
// allocation support, make a clone of it, adding additional arguments as
// necessary, and return it. If not, just return null.
//
Function* RTAssociate::MakeFunctionClone(Function &F, FuncInfo& FI, DSGraph* G) {
if (G->node_begin() == G->node_end()) return 0;
if (FI.ArgNodes.empty())
return 0; // No need to clone if no pools need to be passed in!
// Update statistics..
NumArgsAdded += FI.ArgNodes.size();
if (MaxArgsAdded < FI.ArgNodes.size()) MaxArgsAdded = FI.ArgNodes.size();
++NumCloned;
// Figure out what the arguments are to be for the new version of the
// function
FunctionType *OldFuncTy = F.getFunctionType();
std::vector<Type*> ArgTys(FI.ArgNodes.size(), PoolDescPtrTy);
ArgTys.reserve(OldFuncTy->getNumParams() + FI.ArgNodes.size());
ArgTys.insert(ArgTys.end(), OldFuncTy->param_begin(), OldFuncTy->param_end());
// Create the new function prototype
FunctionType *FuncTy = FunctionType::get(OldFuncTy->getReturnType(), ArgTys,
OldFuncTy->isVarArg());
// Create the new function...
Function *New = Function::Create(FuncTy, Function::InternalLinkage, F.getName());
New->copyAttributesFrom(&F);
F.getParent()->getFunctionList().insert(&F, New);
// Set the rest of the new arguments names to be PDa<n> and add entries to the
// pool descriptors map
Function::arg_iterator NI = New->arg_begin();
for (unsigned i = 0, e = FI.ArgNodes.size(); i != e; ++i, ++NI) {
FI.PoolDescriptors[FI.ArgNodes[i]] = CreateArgPool(FI.ArgNodes[i], NI);
NI->setName("PDa");
}
// Map the existing arguments of the old function to the corresponding
// arguments of the new function, and copy over the names.
ValueToValueMapTy ValueMap;
for (Function::arg_iterator I = F.arg_begin();
NI != New->arg_end(); ++I, ++NI) {
ValueMap[I] = NI;
NI->setName(I->getName());
}
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
// TODO: review the boolean flag here
CloneFunctionInto(New, &F, ValueMap, true, Returns);
//
// The CloneFunctionInto() function will copy the parameter attributes
// verbatim. This is incorrect; each attribute should be shifted one so
// that the pool descriptor has no attributes.
//
const AttributeSet OldAttrs = New->getAttributes();
if (!OldAttrs.isEmpty()) {
AttributeSet NewAttrs;
for (unsigned index = 0; index < OldAttrs.getNumSlots(); ++index) {
const AttributeSet & PAWI = OldAttrs.getSlotAttributes(index);
unsigned argIndex = OldAttrs.getSlotIndex(index);
// If it's not the return value, move the attribute to the next
// parameter.
if (argIndex) ++argIndex;
// Add the parameter to the new list.
NewAttrs = NewAttrs.addAttributes(F.getContext(), argIndex, PAWI);
}
// Assign the new attributes to the function clone
New->setAttributes(NewAttrs);
}
for (ValueToValueMapTy::iterator I = ValueMap.begin(),
E = ValueMap.end(); I != E; ++I)
FI.NewToOldValueMap.insert(std::make_pair(I->second, const_cast<Value*>(I->first)));
return FI.Clone = New;
}
// 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;
}