// Helper to find the set of values described by a TSpecifier
std::set<const Value*>
getValues(const ImmutableCallSite cs, TSpecifier TS) {
std::set<const Value*> Values;
switch (TS) {
case Ret:
assert(!cs.getInstruction()->getType()->isVoidTy());
Values.insert(cs.getInstruction());
break;
case Arg0:
assert(0 < cs.arg_size());
Values.insert(cs.getArgument(0));
break;
case Arg1:
assert(1 < cs.arg_size());
Values.insert(cs.getArgument(1));
break;
case Arg2:
assert(2 < cs.arg_size());
Values.insert(cs.getArgument(2));
break;
case Arg3:
assert(3 < cs.arg_size());
Values.insert(cs.getArgument(3));
break;
case Arg4:
assert(4 < cs.arg_size());
Values.insert(cs.getArgument(4));
break;
case AllArgs:
assert(!cs.arg_empty());
for (unsigned i = 0; i < cs.arg_size(); ++i)
Values.insert(cs.getArgument(i));
break;
case VarArgs: {
const Value *Callee = cs.getCalledValue()->stripPointerCasts();
FunctionType *CalleeType =
dyn_cast<FunctionType>(
dyn_cast<PointerType>(Callee->getType())->getElementType()
);
for (unsigned i = CalleeType->getNumParams(); i < cs.arg_size(); ++i)
Values.insert(cs.getArgument(i));
break;
}
}
return Values;
}
/// ComputeUsesVAFloatArgument - Determine if any floating-point values are
/// being passed to this variadic function, and set the MachineModuleInfo's
/// usesVAFloatArgument flag if so. This flag is used to emit an undefined
/// reference to _fltused on Windows, which will link in MSVCRT's
/// floating-point support.
void llvm::ComputeUsesVAFloatArgument(const CallInst &I,
MachineModuleInfo *MMI)
{
FunctionType *FT = cast<FunctionType>(
I.getCalledValue()->getType()->getContainedType(0));
if (FT->isVarArg() && !MMI->usesVAFloatArgument()) {
for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
Type* T = I.getArgOperand(i)->getType();
for (auto i : post_order(T)) {
if (i->isFloatingPointTy()) {
MMI->setUsesVAFloatArgument(true);
return;
}
}
}
}
}
static void
lCreateSymbol(const std::string &name, const Type *returnType,
const std::vector<const Type *> &argTypes,
const llvm::FunctionType *ftype, llvm::Function *func,
SymbolTable *symbolTable) {
SourcePos noPos;
noPos.name = "__stdlib";
FunctionType *funcType = new FunctionType(returnType, argTypes, noPos);
Debug(noPos, "Created builtin symbol \"%s\" [%s]\n", name.c_str(),
funcType->GetString().c_str());
Symbol *sym = new Symbol(name, noPos, funcType);
sym->function = func;
symbolTable->AddFunction(sym);
}
static bool ExpandVarArgFunc(Module *M, Function *Func) {
if (isEmscriptenJSArgsFunc(M, Func->getName()))
return false;
Type *PtrType = Type::getInt8PtrTy(Func->getContext());
FunctionType *FTy = Func->getFunctionType();
SmallVector<Type *, 8> Params(FTy->param_begin(), FTy->param_end());
Params.push_back(PtrType);
FunctionType *NFTy =
FunctionType::get(FTy->getReturnType(), Params, /*isVarArg=*/false);
Function *NewFunc = RecreateFunction(Func, NFTy);
// Declare the new argument as "noalias".
NewFunc->setAttributes(Func->getAttributes().addAttribute(
Func->getContext(), FTy->getNumParams() + 1, Attribute::NoAlias));
// Move the arguments across to the new function.
auto NewArg = NewFunc->arg_begin();
for (Argument &Arg : Func->args()) {
Arg.replaceAllUsesWith(NewArg);
NewArg->takeName(&Arg);
++NewArg;
}
// The last argument is the new `i8 * noalias %varargs`.
NewArg->setName("varargs");
Func->eraseFromParent();
// Expand out uses of llvm.va_start in this function.
for (BasicBlock &BB : *NewFunc) {
for (auto BI = BB.begin(), BE = BB.end(); BI != BE;) {
Instruction *I = BI++;
if (auto *VAS = dyn_cast<VAStartInst>(I)) {
IRBuilder<> IRB(VAS);
Value *Cast = IRB.CreateBitCast(VAS->getArgList(),
PtrType->getPointerTo(), "arglist");
IRB.CreateStore(NewArg, Cast);
VAS->eraseFromParent();
}
}
}
return true;
}
// Replace G with a simple tail call to bitcast(F). Also replace direct uses
// of G with bitcast(F). Deletes G.
void MergeFunctions::writeThunk(Function *F, Function *G) {
if (!G->mayBeOverridden()) {
// Redirect direct callers of G to F.
replaceDirectCallers(G, F);
}
// If G was internal then we may have replaced all uses of G with F. If so,
// stop here and delete G. There's no need for a thunk.
if (G->hasLocalLinkage() && G->use_empty()) {
G->eraseFromParent();
return;
}
Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
G->getParent());
BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
IRBuilder<false> Builder(BB);
SmallVector<Value *, 16> Args;
unsigned i = 0;
FunctionType *FFTy = F->getFunctionType();
for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
AI != AE; ++AI) {
Args.push_back(createCast(Builder, (Value*)AI, FFTy->getParamType(i)));
++i;
}
CallInst *CI = Builder.CreateCall(F, Args);
CI->setTailCall();
CI->setCallingConv(F->getCallingConv());
if (NewG->getReturnType()->isVoidTy()) {
Builder.CreateRetVoid();
} else {
Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType()));
}
NewG->copyAttributesFrom(G);
NewG->takeName(G);
removeUsers(G);
G->replaceAllUsesWith(NewG);
G->eraseFromParent();
DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
++NumThunksWritten;
}
void TypeDeduction::Visit(CallExpression* node)
{
node->GetTarget()->Accept(this);
std::vector<Expression *> *arguments = node->GetArgumentList();
for (std::vector<Expression *>::iterator it = arguments->begin(); it != arguments->end(); ++it)
{
Expression *arg = *it;
arg->Accept(this);
//TODO: Check the parameter types, and do implicit conversion if needed
}
FunctionType *funcType = dynamic_cast<FunctionType *>(node->GetTarget()->GetTag<Type>("Type"));
if (funcType == NULL)
{
CompilationContext::GetInstance()->ReportError(node->SourceLocation, false, "Requires a function type.");
}
node->SetTag<Type>("Type", funcType->GetReturnType());
}
/// isFreeCall - Returns non-null if the value is a call to the builtin free()
const CallInst *llvm::isFreeCall(const Value *I, const TargetLibraryInfo *TLI) {
const CallInst *CI = dyn_cast<CallInst>(I);
if (!CI || isa<IntrinsicInst>(CI))
return nullptr;
Function *Callee = CI->getCalledFunction();
if (Callee == nullptr || !Callee->isDeclaration())
return nullptr;
StringRef FnName = Callee->getName();
LibFunc::Func TLIFn;
if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn))
return nullptr;
unsigned ExpectedNumParams;
if (TLIFn == LibFunc::free ||
TLIFn == LibFunc::ZdlPv || // operator delete(void*)
TLIFn == LibFunc::ZdaPv) // operator delete[](void*)
ExpectedNumParams = 1;
else if (TLIFn == LibFunc::ZdlPvj || // delete(void*, uint)
TLIFn == LibFunc::ZdlPvm || // delete(void*, ulong)
TLIFn == LibFunc::ZdlPvRKSt9nothrow_t || // delete(void*, nothrow)
TLIFn == LibFunc::ZdaPvj || // delete[](void*, uint)
TLIFn == LibFunc::ZdaPvm || // delete[](void*, ulong)
TLIFn == LibFunc::ZdaPvRKSt9nothrow_t) // delete[](void*, nothrow)
ExpectedNumParams = 2;
else
return nullptr;
// Check free prototype.
// FIXME: workaround for PR5130, this will be obsolete when a nobuiltin
// attribute will exist.
FunctionType *FTy = Callee->getFunctionType();
if (!FTy->getReturnType()->isVoidTy())
return nullptr;
if (FTy->getNumParams() != ExpectedNumParams)
return nullptr;
if (FTy->getParamType(0) != Type::getInt8PtrTy(Callee->getContext()))
return nullptr;
return CI;
}
// Try to find address of external function given a Function object.
// Please note, that interpreter doesn't know how to assemble a
// real call in general case (this is JIT job), that's why it assumes,
// that all external functions has the same (and pretty "general") signature.
// The typical example of such functions are "lle_X_" ones.
static ExFunc lookupFunction(const Function *F) {
// Function not found, look it up... start by figuring out what the
// composite function name should be.
std::string ExtName = "lle_";
FunctionType *FT = F->getFunctionType();
ExtName += getTypeID(FT->getReturnType());
for (Type *T : FT->params())
ExtName += getTypeID(T);
ExtName += ("_" + F->getName()).str();
sys::ScopedLock Writer(*FunctionsLock);
ExFunc FnPtr = (*FuncNames)[ExtName];
if (!FnPtr)
FnPtr = (*FuncNames)[("lle_X_" + F->getName()).str()];
if (!FnPtr) // Try calling a generic function... if it exists...
FnPtr = (ExFunc)(intptr_t)sys::DynamicLibrary::SearchForAddressOfSymbol(
("lle_X_" + F->getName()).str());
if (FnPtr)
ExportedFunctions->insert(std::make_pair(F, FnPtr)); // Cache for later
return FnPtr;
}
/// Replace \p Thunk with a simple tail call to \p ToFunc. Also add parameters
/// to the call to \p ToFunc, which are defined by the FuncIdx's value in
/// \p Params.
void SwiftMergeFunctions::writeThunk(Function *ToFunc, Function *Thunk,
const ParamInfos &Params,
unsigned FuncIdx) {
// Delete the existing content of Thunk.
Thunk->dropAllReferences();
BasicBlock *BB = BasicBlock::Create(Thunk->getContext(), "", Thunk);
IRBuilder<> Builder(BB);
SmallVector<Value *, 16> Args;
unsigned ParamIdx = 0;
FunctionType *ToFuncTy = ToFunc->getFunctionType();
// Add arguments which are passed through Thunk.
for (Argument & AI : Thunk->args()) {
Args.push_back(createCast(Builder, &AI, ToFuncTy->getParamType(ParamIdx)));
++ParamIdx;
}
// Add new arguments defined by Params.
for (const ParamInfo &PI : Params) {
assert(ParamIdx < ToFuncTy->getNumParams());
Args.push_back(createCast(Builder, PI.Values[FuncIdx],
ToFuncTy->getParamType(ParamIdx)));
++ParamIdx;
}
CallInst *CI = Builder.CreateCall(ToFunc, Args);
CI->setTailCall();
CI->setCallingConv(ToFunc->getCallingConv());
CI->setAttributes(ToFunc->getAttributes());
if (Thunk->getReturnType()->isVoidTy()) {
Builder.CreateRetVoid();
} else {
Builder.CreateRet(createCast(Builder, CI, Thunk->getReturnType()));
}
LLVM_DEBUG(dbgs() << " writeThunk: " << Thunk->getName() << '\n');
++NumSwiftThunksWritten;
}
static bool isMallocCall(const CallInst *CI) {
if (!CI)
return false;
Function *Callee = CI->getCalledFunction();
if (Callee == 0 || !Callee->isDeclaration())
return false;
if (Callee->getName() != "malloc" &&
Callee->getName() != "_Znwj" && // operator new(unsigned int)
Callee->getName() != "_Znwm" && // operator new(unsigned long)
Callee->getName() != "_Znaj" && // operator new[](unsigned int)
Callee->getName() != "_Znam") // operator new[](unsigned long)
return false;
// Check malloc prototype.
// FIXME: workaround for PR5130, this will be obsolete when a nobuiltin
// attribute will exist.
FunctionType *FTy = Callee->getFunctionType();
return FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) &&
FTy->getNumParams() == 1 &&
(FTy->getParamType(0)->isIntegerTy(32) ||
FTy->getParamType(0)->isIntegerTy(64));
}
bool CallingConvention_x86_64_systemv::analyzeFunctionType(ParameterRegistry& registry, CallInformation& fillOut, FunctionType& type)
{
TargetInfo& targetInfo = registry.getTargetInfo();
auto iter = begin(returnRegisters);
auto addReturn = &CallInformation::addReturn<ValueInformation>;
if (!addEntriesForType(targetInfo, fillOut, addReturn, type.getReturnType(), iter, end(returnRegisters)))
{
return false;
}
size_t spOffset = 0;
iter = begin(parameterRegisters);
auto addParam = &CallInformation::addParameter<ValueInformation>;
for (Type* t : type.params())
{
if (!addEntriesForType(targetInfo, fillOut, addParam, t, iter, end(parameterRegisters), &spOffset))
{
return false;
}
}
return true;
}
void Preparer::expandCallSite(CallSite CS) {
// Skip the callsites that are not calling a va function.
Value *Callee = CS.getCalledValue();
FunctionType *CalleeType = cast<FunctionType>(
cast<PointerType>(Callee->getType())->getElementType());
if (!CalleeType->isVarArg()) {
return;
}
vector<Value *> Args;
for (CallSite::arg_iterator ArgI = CS.arg_begin();
ArgI != CS.arg_end(); ArgI++) {
Args.push_back(*ArgI);
}
Args.push_back(ConstantInt::get(
IntegerType::get(CS.getInstruction()->getContext(), 8), 0));
string InstName = "";
if (CS.getInstruction()->getName() != "")
InstName = CS.getInstruction()->getName().str() + ".padded";
if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
CallInst *NewCI = CallInst::Create(Callee, Args, InstName, CI);
NewCI->setAttributes(CI->getAttributes());
CI->replaceAllUsesWith(NewCI);
CI->eraseFromParent();
} else if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
InvokeInst *NewII = InvokeInst::Create(Callee,
II->getNormalDest(),
II->getUnwindDest(),
Args,
InstName,
II);
NewII->setAttributes(II->getAttributes());
II->replaceAllUsesWith(NewII);
II->eraseFromParent();
}
}
double L_BFGS::Evaluate(FunctionType& function,
const arma::mat& iterate,
std::pair<arma::mat, double>& minPointIterate)
{
// Evaluate the function and keep track of the minimum function
// value encountered during the optimization.
const double functionValue = function.Evaluate(iterate);
if (functionValue < minPointIterate.second)
{
minPointIterate.first = iterate;
minPointIterate.second = functionValue;
}
return functionValue;
}
/// \brief Returns the allocation data for the given value if it is a call to a
/// known allocation function, and NULL otherwise.
static const AllocFnsTy *getAllocationData(const Value *V, AllocType AllocTy,
const TargetLibraryInfo *TLI,
bool LookThroughBitCast = false) {
// Skip all intrinsics but duetto.allocate
if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(V))
{
if (II->getIntrinsicID() == Intrinsic::duetto_allocate)
return &AllocationFnData[0];
return 0;
}
Function *Callee = getCalledFunction(V, LookThroughBitCast);
if (!Callee)
return 0;
// Make sure that the function is available.
StringRef FnName = Callee->getName();
LibFunc::Func TLIFn;
if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn))
return 0;
unsigned i = 0;
bool found = false;
for ( ; i < array_lengthof(AllocationFnData); ++i) {
if (AllocationFnData[i].Func == TLIFn) {
found = true;
break;
}
}
if (!found)
return 0;
const AllocFnsTy *FnData = &AllocationFnData[i];
if ((FnData->AllocTy & AllocTy) != FnData->AllocTy)
return 0;
// Check function prototype.
int FstParam = FnData->FstParam;
int SndParam = FnData->SndParam;
FunctionType *FTy = Callee->getFunctionType();
if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) &&
FTy->getNumParams() == FnData->NumParams &&
(FstParam < 0 ||
(FTy->getParamType(FstParam)->isIntegerTy(32) ||
FTy->getParamType(FstParam)->isIntegerTy(64))) &&
(SndParam < 0 ||
FTy->getParamType(SndParam)->isIntegerTy(32) ||
FTy->getParamType(SndParam)->isIntegerTy(64)))
return FnData;
return 0;
}
static void check(Value *Func, ArrayRef<Value *> Args) {
FunctionType *FTy =
cast<FunctionType>(cast<PointerType>(Func->getType())->getElementType());
assert((Args.size() == FTy->getNumParams() ||
(FTy->isVarArg() && Args.size() > FTy->getNumParams())) &&
"XXCalling a function with bad signature!");
for (unsigned i = 0; i != Args.size(); ++i) {
if (!(FTy->getParamType(i) == Args[i]->getType())) {
errs() << "types:\n ";
FTy->getParamType(i)->dump();
errs() << "\n ";
Args[i]->getType()->dump();
errs() << "\n";
}
assert((i >= FTy->getNumParams() ||
FTy->getParamType(i) == Args[i]->getType()) &&
"YYCalling a function with a bad signature!");
}
}
/// \brief Returns the allocation data for the given value if it is a call to a
/// known allocation function, and NULL otherwise.
static const AllocFnsTy *getAllocationData(const Value *V, AllocType AllocTy,
const TargetLibraryInfo *TLI,
bool LookThroughBitCast = false) {
// Skip intrinsics
if (isa<IntrinsicInst>(V))
return nullptr;
Function *Callee = getCalledFunction(V, LookThroughBitCast);
if (!Callee)
return nullptr;
// Make sure that the function is available.
StringRef FnName = Callee->getName();
LibFunc::Func TLIFn;
if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn))
return nullptr;
const AllocFnsTy *FnData =
std::find_if(std::begin(AllocationFnData), std::end(AllocationFnData),
[TLIFn](const AllocFnsTy &Fn) { return Fn.Func == TLIFn; });
if (FnData == std::end(AllocationFnData))
return nullptr;
if ((FnData->AllocTy & AllocTy) != FnData->AllocTy)
return nullptr;
// Check function prototype.
int FstParam = FnData->FstParam;
int SndParam = FnData->SndParam;
FunctionType *FTy = Callee->getFunctionType();
if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) &&
FTy->getNumParams() == FnData->NumParams &&
(FstParam < 0 ||
(FTy->getParamType(FstParam)->isIntegerTy(32) ||
FTy->getParamType(FstParam)->isIntegerTy(64))) &&
(SndParam < 0 ||
FTy->getParamType(SndParam)->isIntegerTy(32) ||
FTy->getParamType(SndParam)->isIntegerTy(64)))
return FnData;
return nullptr;
}
static bool isCallocCall(const CallInst *CI) {
if (!CI)
return false;
Function *Callee = CI->getCalledFunction();
if (Callee == 0 || !Callee->isDeclaration())
return false;
if (Callee->getName() != "calloc")
return false;
// Check malloc prototype.
// FIXME: workaround for PR5130, this will be obsolete when a nobuiltin
// attribute exists.
FunctionType *FTy = Callee->getFunctionType();
return FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) &&
FTy->getNumParams() == 2 &&
((FTy->getParamType(0)->isIntegerTy(32) &&
FTy->getParamType(1)->isIntegerTy(32)) ||
(FTy->getParamType(0)->isIntegerTy(64) &&
FTy->getParamType(1)->isIntegerTy(64)));
}
/// \brief Returns the allocation data for the given value if it is a call to a
/// known allocation function, and NULL otherwise.
static const AllocFnsTy *getAllocationData(const Value *V, AllocType AllocTy,
bool LookThroughBitCast = false) {
Function *Callee = getCalledFunction(V, LookThroughBitCast);
if (!Callee)
return 0;
unsigned i = 0;
bool found = false;
for ( ; i < array_lengthof(AllocationFnData); ++i) {
if (Callee->getName() == AllocationFnData[i].Name) {
found = true;
break;
}
}
if (!found)
return 0;
const AllocFnsTy *FnData = &AllocationFnData[i];
if ((FnData->AllocTy & AllocTy) == 0)
return 0;
// Check function prototype.
// FIXME: Check the nobuiltin metadata?? (PR5130)
int FstParam = FnData->FstParam;
int SndParam = FnData->SndParam;
FunctionType *FTy = Callee->getFunctionType();
if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) &&
FTy->getNumParams() == FnData->NumParams &&
(FstParam < 0 ||
(FTy->getParamType(FstParam)->isIntegerTy(32) ||
FTy->getParamType(FstParam)->isIntegerTy(64))) &&
(SndParam < 0 ||
FTy->getParamType(SndParam)->isIntegerTy(32) ||
FTy->getParamType(SndParam)->isIntegerTy(64)))
return FnData;
return 0;
}
/// Returns the allocation data for the given value if it's either a call to a
/// known allocation function, or a call to a function with the allocsize
/// attribute.
static Optional<AllocFnsTy>
getAllocationDataForFunction(const Function *Callee, AllocType AllocTy,
const TargetLibraryInfo *TLI) {
// Make sure that the function is available.
StringRef FnName = Callee->getName();
LibFunc TLIFn;
if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn))
return None;
const auto *Iter = find_if(
AllocationFnData, [TLIFn](const std::pair<LibFunc, AllocFnsTy> &P) {
return P.first == TLIFn;
});
if (Iter == std::end(AllocationFnData))
return None;
const AllocFnsTy *FnData = &Iter->second;
if ((FnData->AllocTy & AllocTy) != FnData->AllocTy)
return None;
// Check function prototype.
int FstParam = FnData->FstParam;
int SndParam = FnData->SndParam;
FunctionType *FTy = Callee->getFunctionType();
if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) &&
FTy->getNumParams() == FnData->NumParams &&
(FstParam < 0 ||
(FTy->getParamType(FstParam)->isIntegerTy(32) ||
FTy->getParamType(FstParam)->isIntegerTy(64))) &&
(SndParam < 0 ||
FTy->getParamType(SndParam)->isIntegerTy(32) ||
FTy->getParamType(SndParam)->isIntegerTy(64)))
return *FnData;
return None;
}
// 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;
}
/// DeleteDeadVarargs - If this is an function that takes a ... list, and if
/// llvm.vastart is never called, the varargs list is dead for the function.
bool DAE::DeleteDeadVarargs(Function &Fn) {
assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!");
if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false;
// Ensure that the function is only directly called.
if (Fn.hasAddressTaken())
return false;
// Don't touch naked functions. The assembly might be using an argument, or
// otherwise rely on the frame layout in a way that this analysis will not
// see.
if (Fn.hasFnAttribute(Attribute::Naked)) {
return false;
}
// Okay, we know we can transform this function if safe. Scan its body
// looking for calls marked musttail or calls to llvm.vastart.
for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
CallInst *CI = dyn_cast<CallInst>(I);
if (!CI)
continue;
if (CI->isMustTailCall())
return false;
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
if (II->getIntrinsicID() == Intrinsic::vastart)
return false;
}
}
}
// If we get here, there are no calls to llvm.vastart in the function body,
// remove the "..." and adjust all the calls.
// Start by computing a new prototype for the function, which is the same as
// the old function, but doesn't have isVarArg set.
FunctionType *FTy = Fn.getFunctionType();
std::vector<Type*> Params(FTy->param_begin(), FTy->param_end());
FunctionType *NFTy = FunctionType::get(FTy->getReturnType(),
Params, false);
unsigned NumArgs = Params.size();
// Create the new function body and insert it into the module...
Function *NF = Function::Create(NFTy, Fn.getLinkage());
NF->copyAttributesFrom(&Fn);
Fn.getParent()->getFunctionList().insert(Fn.getIterator(), NF);
NF->takeName(&Fn);
// Loop over all of the callers of the function, transforming the call sites
// to pass in a smaller number of arguments into the new function.
//
std::vector<Value*> Args;
for (Value::user_iterator I = Fn.user_begin(), E = Fn.user_end(); I != E; ) {
CallSite CS(*I++);
if (!CS)
continue;
Instruction *Call = CS.getInstruction();
// Pass all the same arguments.
Args.assign(CS.arg_begin(), CS.arg_begin() + NumArgs);
// Drop any attributes that were on the vararg arguments.
AttributeSet PAL = CS.getAttributes();
if (!PAL.isEmpty() && PAL.getSlotIndex(PAL.getNumSlots() - 1) > NumArgs) {
SmallVector<AttributeSet, 8> AttributesVec;
for (unsigned i = 0; PAL.getSlotIndex(i) <= NumArgs; ++i)
AttributesVec.push_back(PAL.getSlotAttributes(i));
if (PAL.hasAttributes(AttributeSet::FunctionIndex))
AttributesVec.push_back(AttributeSet::get(Fn.getContext(),
PAL.getFnAttributes()));
PAL = AttributeSet::get(Fn.getContext(), AttributesVec);
}
Instruction *New;
if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
Args, "", Call);
cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
cast<InvokeInst>(New)->setAttributes(PAL);
} else {
New = CallInst::Create(NF, Args, "", Call);
cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
cast<CallInst>(New)->setAttributes(PAL);
if (cast<CallInst>(Call)->isTailCall())
cast<CallInst>(New)->setTailCall();
}
New->setDebugLoc(Call->getDebugLoc());
Args.clear();
if (!Call->use_empty())
Call->replaceAllUsesWith(New);
New->takeName(Call);
// Finally, remove the old call from the program, reducing the use-count of
// F.
Call->eraseFromParent();
}
// Since we have now created the new function, splice the body of the old
// function right into the new function, leaving the old rotting hulk of the
// function empty.
NF->getBasicBlockList().splice(NF->begin(), Fn.getBasicBlockList());
// Loop over the argument list, transferring uses of the old arguments over to
// the new arguments, also transferring over the names as well. While we're at
// it, remove the dead arguments from the DeadArguments list.
//
for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(),
I2 = NF->arg_begin(); I != E; ++I, ++I2) {
// Move the name and users over to the new version.
I->replaceAllUsesWith(&*I2);
I2->takeName(&*I);
}
// Patch the pointer to LLVM function in debug info descriptor.
NF->setSubprogram(Fn.getSubprogram());
// Fix up any BlockAddresses that refer to the function.
Fn.replaceAllUsesWith(ConstantExpr::getBitCast(NF, Fn.getType()));
// Delete the bitcast that we just created, so that NF does not
// appear to be address-taken.
NF->removeDeadConstantUsers();
// Finally, nuke the old function.
Fn.eraseFromParent();
return true;
}
/// cmpType - compares two types,
/// defines total ordering among the types set.
/// See method declaration comments for more details.
int FunctionComparator::cmpType(Type *TyL, Type *TyR) const {
PointerType *PTyL = dyn_cast<PointerType>(TyL);
PointerType *PTyR = dyn_cast<PointerType>(TyR);
if (DL) {
if (PTyL && PTyL->getAddressSpace() == 0) TyL = DL->getIntPtrType(TyL);
if (PTyR && PTyR->getAddressSpace() == 0) TyR = DL->getIntPtrType(TyR);
}
if (TyL == TyR)
return 0;
if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
return Res;
switch (TyL->getTypeID()) {
default:
llvm_unreachable("Unknown type!");
// Fall through in Release mode.
case Type::IntegerTyID:
case Type::VectorTyID:
// TyL == TyR would have returned true earlier.
return cmpNumbers((uint64_t)TyL, (uint64_t)TyR);
case Type::VoidTyID:
case Type::FloatTyID:
case Type::DoubleTyID:
case Type::X86_FP80TyID:
case Type::FP128TyID:
case Type::PPC_FP128TyID:
case Type::LabelTyID:
case Type::MetadataTyID:
return 0;
case Type::PointerTyID: {
assert(PTyL && PTyR && "Both types must be pointers here.");
return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
}
case Type::StructTyID: {
StructType *STyL = cast<StructType>(TyL);
StructType *STyR = cast<StructType>(TyR);
if (STyL->getNumElements() != STyR->getNumElements())
return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
if (STyL->isPacked() != STyR->isPacked())
return cmpNumbers(STyL->isPacked(), STyR->isPacked());
for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
if (int Res = cmpType(STyL->getElementType(i),
STyR->getElementType(i)))
return Res;
}
return 0;
}
case Type::FunctionTyID: {
FunctionType *FTyL = cast<FunctionType>(TyL);
FunctionType *FTyR = cast<FunctionType>(TyR);
if (FTyL->getNumParams() != FTyR->getNumParams())
return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
if (FTyL->isVarArg() != FTyR->isVarArg())
return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
if (int Res = cmpType(FTyL->getReturnType(), FTyR->getReturnType()))
return Res;
for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
if (int Res = cmpType(FTyL->getParamType(i), FTyR->getParamType(i)))
return Res;
}
return 0;
}
case Type::ArrayTyID: {
ArrayType *ATyL = cast<ArrayType>(TyL);
ArrayType *ATyR = cast<ArrayType>(TyR);
if (ATyL->getNumElements() != ATyR->getNumElements())
return cmpNumbers(ATyL->getNumElements(), ATyR->getNumElements());
return cmpType(ATyL->getElementType(), ATyR->getElementType());
}
}
}
//
// Method: visitCallSite()
//
// Description:
// This method transforms a call site. A call site may either be a call
// instruction or an invoke instruction.
//
// Inputs:
// CS - The call site representing the instruction that should be transformed.
//
void FuncTransform::visitCallSite(CallSite& CS) {
const Function *CF = CS.getCalledFunction();
Instruction *TheCall = CS.getInstruction();
bool thread_creation_point = false;
//
// Get the value that is called at this call site. Strip away any pointer
// casts that do not change the representation of the data (i.e., are
// lossless casts).
//
Value * CalledValue = CS.getCalledValue()->stripPointerCasts();
//
// The CallSite::getCalledFunction() method is not guaranteed to strip off
// pointer casts. If no called function was found, manually strip pointer
// casts off of the called value and see if we get a function. If so, this
// is a direct call, and we want to update CF accordingly.
//
if (!CF) CF = dyn_cast<Function>(CalledValue);
//
// Do not change any inline assembly code.
//
if (isa<InlineAsm>(TheCall->getOperand(0))) {
errs() << "INLINE ASM: ignoring. Hoping that's safe.\n";
return;
}
//
// Ignore calls to NULL pointers or undefined values.
//
if ((isa<ConstantPointerNull>(CalledValue)) ||
(isa<UndefValue>(CalledValue))) {
errs() << "WARNING: Ignoring call using NULL/Undef function pointer.\n";
return;
}
// If this function is one of the memory manipulating functions built into
// libc, emulate it with pool calls as appropriate.
if (CF && CF->isDeclaration()) {
std::string Name = CF->getName();
if (Name == "free" || Name == "cfree") {
visitFreeCall(CS);
return;
} else if (Name == "malloc") {
visitMallocCall(CS);
return;
} else if (Name == "calloc") {
visitCallocCall(CS);
return;
} else if (Name == "realloc") {
visitReallocCall(CS);
return;
} else if (Name == "memalign" || Name == "posix_memalign") {
visitMemAlignCall(CS);
return;
} else if (Name == "strdup") {
visitStrdupCall(CS);
return;
} else if (Name == "valloc") {
errs() << "VALLOC USED BUT NOT HANDLED!\n";
abort();
} else if (unsigned PoolArgc = PAInfo.getNumInitialPoolArguments(Name)) {
visitRuntimeCheck(CS, PoolArgc);
return;
} else if (Name == "pthread_create") {
thread_creation_point = true;
//
// Get DSNode representing the DSNode of the function pointer Value of
// the pthread_create call
//
DSNode* thread_callee_node = G->getNodeForValue(CS.getArgument(2)).getNode();
if (!thread_callee_node) {
assert(0 && "apparently you need this code");
FuncInfo *CFI = PAInfo.getFuncInfo(*CF);
thread_callee_node = G->getNodeForValue(CFI->MapValueToOriginal(CS.getArgument(2))).getNode();
}
// Fill in CF with the name of one of the functions in thread_callee_node
CF = const_cast<Function*>(dyn_cast<Function>(*thread_callee_node->globals_begin()));
}
}
//
// We need to figure out which local pool descriptors correspond to the pool
// descriptor arguments passed into the function call. Calculate a mapping
// from callee DSNodes to caller DSNodes. We construct a partial isomophism
// between the graphs to figure out which pool descriptors need to be passed
// in. The roots of this mapping is found from arguments and return values.
//
DataStructures& Graphs = PAInfo.getGraphs();
DSGraph::NodeMapTy NodeMapping;
Instruction *NewCall;
Value *NewCallee;
std::vector<const DSNode*> ArgNodes;
DSGraph *CalleeGraph; // The callee graph
// For indirect callees, find any callee since all DS graphs have been
// merged.
if (CF) { // Direct calls are nice and simple.
DEBUG(errs() << " Handling direct call: " << *TheCall << "\n");
//
// Do not try to add pool handles to the function if it:
// a) Already calls a cloned function; or
// b) Calls a function which was never cloned.
//
// For such a call, just replace any arguments that take original functions
// with their cloned function poiner values.
//
FuncInfo *CFI = PAInfo.getFuncInfo(*CF);
if (CFI == 0 || CFI->Clone == 0) { // Nothing to transform...
visitInstruction(*TheCall);
return;
}
//
// Oh, dear. We must add pool descriptors to this direct call.
//
NewCallee = CFI->Clone;
ArgNodes = CFI->ArgNodes;
assert ((Graphs.hasDSGraph (*CF)) && "Function has no ECGraph!\n");
CalleeGraph = Graphs.getDSGraph(*CF);
} else {
DEBUG(errs() << " Handling indirect call: " << *TheCall << "\n");
DSGraph *G = Graphs.getGlobalsGraph();
DSGraph::ScalarMapTy& SM = G->getScalarMap();
// Here we fill in CF with one of the possible called functions. Because we
// merged together all of the arguments to all of the functions in the
// equivalence set, it doesn't really matter which one we pick.
// (If the function was cloned, we have to map the cloned call instruction
// in CS back to the original call instruction.)
Instruction *OrigInst =
cast<Instruction>(getOldValueIfAvailable(CS.getInstruction()));
//
// Attempt to get one of the function targets of this indirect call site by
// looking at the call graph constructed by the points-to analysis. Be
// sure to use the original call site from the original function; the
// points-to analysis has no information on the clones we've created.
//
// Also, look for the target that has the greatest number of arguments that
// have associated DSNodes. This ensures that we pass the maximum number
// of pools possible and prevents us from eliding a pool because we're
// examining a target that doesn't need it.
//
const DSCallGraph & callGraph = Graphs.getCallGraph();
DSCallGraph::callee_iterator I = callGraph.callee_begin(OrigInst);
for (; I != callGraph.callee_end(OrigInst); ++I) {
for(DSCallGraph::scc_iterator sccii = callGraph.scc_begin(*I),
sccee = callGraph.scc_end(*I); sccii != sccee; ++sccii){
if(SM.find(SM.getLeaderForGlobal(*sccii)) == SM.end())
continue;
//
// Get the information for this function. Since this is coming from
// DSA, it should be an original function.
//
// This call site calls a function, that is not defined in this module
if (!(Graphs.hasDSGraph(**sccii))) return;
// For all other cases Func Info must exist.
PAInfo.getFuncInfo(**sccii);
//
// If this target takes more DSNodes than the last one we found, then
// make *this* target our canonical target.
//
CF = *sccii;
break;
}
}
if(!CF){
const Function *F1 = OrigInst->getParent()->getParent();
F1 = callGraph.sccLeader(&*F1);
for(DSCallGraph::scc_iterator sccii = callGraph.scc_begin(F1),
sccee = callGraph.scc_end(F1); sccii != sccee; ++sccii){
if(SM.find(SM.getLeaderForGlobal(*sccii)) == SM.end())
continue;
//
// Get the information for this function. Since this is coming from DSA,
// it should be an original function.
//
// This call site calls a function, that is not defined in this module
if (!(Graphs.hasDSGraph(**sccii))) return;
// For all other cases Func Info must exist.
PAInfo.getFuncInfo(**sccii);
//
// If this target takes more DSNodes than the last one we found, then
// make *this* target our canonical target.
//
CF = *sccii;
}
}
// Assuming the call graph is always correct. And if the call graph reports,
// no callees, we can assume that it is right.
//
// If we didn't find the callee in the constructed call graph, try
// checking in the DSNode itself.
// This isn't ideal as it means that this call site didn't have inlining
// happen.
//
//
// If we still haven't been able to find a target function of the call site
// to transform, do nothing.
//
// One may be tempted to think that we should always have at least one
// target, but this is not true. There are perfectly acceptable (but
// strange) programs for which no function targets exist. Function
// pointers loaded from undef values, for example, will have no targets.
//
if (!CF) return;
//
// It's possible that this program has indirect call targets that are
// not defined in this module. Do not transformation for such functions.
//
if (!(Graphs.hasDSGraph(*CF))) return;
//
// Get the common graph for the set of functions this call may invoke.
//
assert ((Graphs.hasDSGraph(*CF)) && "Function has no DSGraph!\n");
CalleeGraph = Graphs.getDSGraph(*CF);
#ifndef NDEBUG
// Verify that all potential callees at call site have the same DS graph.
DSCallGraph::callee_iterator E = Graphs.getCallGraph().callee_end(OrigInst);
for (; I != E; ++I) {
const Function * F = *I;
assert (F);
if (!(F)->isDeclaration())
assert(CalleeGraph == Graphs.getDSGraph(**I) &&
"Callees at call site do not have a common graph!");
}
#endif
// Find the DS nodes for the arguments that need to be added, if any.
FuncInfo *CFI = PAInfo.getFuncInfo(*CF);
assert(CFI && "No function info for callee at indirect call?");
ArgNodes = CFI->ArgNodes;
if (ArgNodes.empty())
return; // No arguments to add? Transformation is a noop!
// Cast the function pointer to an appropriate type!
std::vector<Type*> ArgTys(ArgNodes.size(),
PoolAllocate::PoolDescPtrTy);
for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
I != E; ++I)
ArgTys.push_back((*I)->getType());
FunctionType *FTy = FunctionType::get(TheCall->getType(), ArgTys, false);
PointerType *PFTy = PointerType::getUnqual(FTy);
// If there are any pool arguments cast the func ptr to the right type.
NewCallee = CastInst::CreatePointerCast(CS.getCalledValue(), PFTy, "tmp", TheCall);
}
//
// FIXME: Why do we disable strict checking when calling the
// DSGraph::computeNodeMapping() method?
//
Function::const_arg_iterator FAI = CF->arg_begin(), E = CF->arg_end();
CallSite::arg_iterator AI = CS.arg_begin() + (thread_creation_point ? 3 : 0);
CallSite::arg_iterator AE = CS.arg_end();
for ( ; FAI != E && AI != AE; ++FAI, ++AI)
if (!isa<Constant>(*AI)) {
DSGraph::computeNodeMapping(CalleeGraph->getNodeForValue(FAI),
getDSNodeHFor(*AI), NodeMapping, false);
}
//assert(AI == AE && "Varargs calls not handled yet!");
// Map the return value as well...
if (isa<PointerType>(TheCall->getType()))
DSGraph::computeNodeMapping(CalleeGraph->getReturnNodeFor(*CF),
getDSNodeHFor(TheCall), NodeMapping, false);
// This code seems redundant (and crashes occasionally)
// There is no reason to map globals here, since they are not passed as
// arguments
// // Map the nodes that are pointed to by globals.
// DSScalarMap &CalleeSM = CalleeGraph->getScalarMap();
// for (DSScalarMap::global_iterator GI = G.getScalarMap().global_begin(),
// E = G.getScalarMap().global_end(); GI != E; ++GI)
// if (CalleeSM.count(*GI))
// DSGraph::computeNodeMapping(CalleeGraph->getNodeForValue(*GI),
// getDSNodeHFor(*GI),
// NodeMapping, false);
//
// Okay, now that we have established our mapping, we can figure out which
// pool descriptors to pass in...
//
// Note:
// There used to be code here that would create a new pool before the
// function call and destroy it after the function call. This could would
// get triggered if bounds checking was disbled or the DSNode for the
// argument was an array value.
//
// I believe that code was incorrect; an argument may have a NULL pool handle
// (i.e., no pool handle) because the pool allocation heuristic used simply
// decided not to assign that value a pool. The argument may alias data
// that should not be freed after the function call is complete, so calling
// pooldestroy() after the call would free data, causing dangling pointer
// dereference errors.
//
std::vector<Value*> Args;
for (unsigned i = 0, e = ArgNodes.size(); i != e; ++i) {
Value *ArgVal = Constant::getNullValue(PoolAllocate::PoolDescPtrTy);
if (NodeMapping.count(ArgNodes[i])) {
if (DSNode *LocalNode = NodeMapping[ArgNodes[i]].getNode())
if (FI.PoolDescriptors.count(LocalNode))
ArgVal = FI.PoolDescriptors.find(LocalNode)->second;
}
Args.push_back(ArgVal);
}
// Add the rest of the arguments unless we're a thread creation point, in which case we only need the pools
if(!thread_creation_point)
Args.insert(Args.end(), CS.arg_begin(), CS.arg_end());
//
// There are circumstances where a function is casted to another type and
// then called (que horible). We need to perform a similar cast if the
// type doesn't match the number of arguments.
//
if (Function * NewFunction = dyn_cast<Function>(NewCallee)) {
FunctionType * NewCalleeType = NewFunction->getFunctionType();
if (NewCalleeType->getNumParams() != Args.size()) {
std::vector<Type *> Types;
Type * FuncTy = FunctionType::get (NewCalleeType->getReturnType(),
Types,
true);
FuncTy = PointerType::getUnqual (FuncTy);
NewCallee = new BitCastInst (NewCallee, FuncTy, "", TheCall);
}
}
std::string Name = TheCall->getName(); TheCall->setName("");
if(thread_creation_point) {
Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
Value* pthread_replacement = M->getFunction("poolalloc_pthread_create");
std::vector<Value*> thread_args;
//Push back original thread arguments through the callee
thread_args.push_back(CS.getArgument(0));
thread_args.push_back(CS.getArgument(1));
thread_args.push_back(CS.getArgument(2));
//Push back the integer argument saying how many uses there are
thread_args.push_back(Constant::getIntegerValue(llvm::Type::getInt32Ty(M->getContext()),APInt(32,Args.size())));
thread_args.insert(thread_args.end(),Args.begin(),Args.end());
thread_args.push_back(CS.getArgument(3));
//Make the thread creation call
NewCall = CallInst::Create(pthread_replacement,
thread_args,
Name,TheCall);
}
else if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
NewCall = InvokeInst::Create (NewCallee, II->getNormalDest(),
II->getUnwindDest(),
Args, Name, TheCall);
} else {
NewCall = CallInst::Create (NewCallee, Args, Name,
TheCall);
}
// Add all of the uses of the pool descriptor
for (unsigned i = 0, e = ArgNodes.size(); i != e; ++i)
AddPoolUse(*NewCall, Args[i], PoolUses);
TheCall->replaceAllUsesWith(NewCall);
DEBUG(errs() << " Result Call: " << *NewCall << "\n");
if (!TheCall->getType()->isVoidTy()) {
// If we are modifying the original function, update the DSGraph...
DSGraph::ScalarMapTy &SM = G->getScalarMap();
DSGraph::ScalarMapTy::iterator CII = SM.find(TheCall);
if (CII != SM.end()) {
SM[NewCall] = CII->second;
SM.erase(CII); // Destroy the CallInst
} else if (!FI.NewToOldValueMap.empty()) {
// Otherwise, if this is a clone, update the NewToOldValueMap with the new
// CI return value.
UpdateNewToOldValueMap(TheCall, NewCall);
}
} else if (!FI.NewToOldValueMap.empty()) {
UpdateNewToOldValueMap(TheCall, NewCall);
}
//
// Copy over the calling convention and attributes of the original call
// instruction to the new call instruction.
//
CallSite(NewCall).setCallingConv(CallSite(TheCall).getCallingConv());
TheCall->eraseFromParent();
visitInstruction(*NewCall);
}
void Lint::visitCallSite(CallSite CS) {
Instruction &I = *CS.getInstruction();
Value *Callee = CS.getCalledValue();
visitMemoryReference(I, Callee, MemoryLocation::UnknownSize, 0, nullptr,
MemRef::Callee);
if (Function *F = dyn_cast<Function>(findValue(Callee,
/*OffsetOk=*/false))) {
Assert(CS.getCallingConv() == F->getCallingConv(),
"Undefined behavior: Caller and callee calling convention differ",
&I);
FunctionType *FT = F->getFunctionType();
unsigned NumActualArgs = CS.arg_size();
Assert(FT->isVarArg() ? FT->getNumParams() <= NumActualArgs
: FT->getNumParams() == NumActualArgs,
"Undefined behavior: Call argument count mismatches callee "
"argument count",
&I);
Assert(FT->getReturnType() == I.getType(),
"Undefined behavior: Call return type mismatches "
"callee return type",
&I);
// Check argument types (in case the callee was casted) and attributes.
// TODO: Verify that caller and callee attributes are compatible.
Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
for (; AI != AE; ++AI) {
Value *Actual = *AI;
if (PI != PE) {
Argument *Formal = &*PI++;
Assert(Formal->getType() == Actual->getType(),
"Undefined behavior: Call argument type mismatches "
"callee parameter type",
&I);
// Check that noalias arguments don't alias other arguments. This is
// not fully precise because we don't know the sizes of the dereferenced
// memory regions.
if (Formal->hasNoAliasAttr() && Actual->getType()->isPointerTy())
for (CallSite::arg_iterator BI = CS.arg_begin(); BI != AE; ++BI)
if (AI != BI && (*BI)->getType()->isPointerTy()) {
AliasResult Result = AA->alias(*AI, *BI);
Assert(Result != MustAlias && Result != PartialAlias,
"Unusual: noalias argument aliases another argument", &I);
}
// Check that an sret argument points to valid memory.
if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
Type *Ty =
cast<PointerType>(Formal->getType())->getElementType();
visitMemoryReference(I, Actual, DL->getTypeStoreSize(Ty),
DL->getABITypeAlignment(Ty), Ty,
MemRef::Read | MemRef::Write);
}
}
}
}
if (CS.isCall() && cast<CallInst>(CS.getInstruction())->isTailCall())
for (CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
AI != AE; ++AI) {
Value *Obj = findValue(*AI, /*OffsetOk=*/true);
Assert(!isa<AllocaInst>(Obj),
"Undefined behavior: Call with \"tail\" keyword references "
"alloca",
&I);
}
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I))
switch (II->getIntrinsicID()) {
default: break;
// TODO: Check more intrinsics
case Intrinsic::memcpy: {
MemCpyInst *MCI = cast<MemCpyInst>(&I);
// TODO: If the size is known, use it.
visitMemoryReference(I, MCI->getDest(), MemoryLocation::UnknownSize,
MCI->getAlignment(), nullptr, MemRef::Write);
visitMemoryReference(I, MCI->getSource(), MemoryLocation::UnknownSize,
MCI->getAlignment(), nullptr, MemRef::Read);
// Check that the memcpy arguments don't overlap. The AliasAnalysis API
// isn't expressive enough for what we really want to do. Known partial
// overlap is not distinguished from the case where nothing is known.
uint64_t Size = 0;
if (const ConstantInt *Len =
dyn_cast<ConstantInt>(findValue(MCI->getLength(),
/*OffsetOk=*/false)))
if (Len->getValue().isIntN(32))
Size = Len->getValue().getZExtValue();
Assert(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
MustAlias,
"Undefined behavior: memcpy source and destination overlap", &I);
break;
}
case Intrinsic::memmove: {
MemMoveInst *MMI = cast<MemMoveInst>(&I);
// TODO: If the size is known, use it.
visitMemoryReference(I, MMI->getDest(), MemoryLocation::UnknownSize,
MMI->getAlignment(), nullptr, MemRef::Write);
visitMemoryReference(I, MMI->getSource(), MemoryLocation::UnknownSize,
MMI->getAlignment(), nullptr, MemRef::Read);
break;
}
case Intrinsic::memset: {
MemSetInst *MSI = cast<MemSetInst>(&I);
// TODO: If the size is known, use it.
visitMemoryReference(I, MSI->getDest(), MemoryLocation::UnknownSize,
MSI->getAlignment(), nullptr, MemRef::Write);
break;
}
case Intrinsic::vastart:
Assert(I.getParent()->getParent()->isVarArg(),
"Undefined behavior: va_start called in a non-varargs function",
&I);
visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Read | MemRef::Write);
break;
case Intrinsic::vacopy:
visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Write);
visitMemoryReference(I, CS.getArgument(1), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Read);
break;
case Intrinsic::vaend:
visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Read | MemRef::Write);
break;
case Intrinsic::stackrestore:
// Stackrestore doesn't read or write memory, but it sets the
// stack pointer, which the compiler may read from or write to
// at any time, so check it for both readability and writeability.
visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Read | MemRef::Write);
break;
}
}
GenericValue MCJIT::runFunction(Function *F,
const std::vector<GenericValue> &ArgValues) {
assert(F && "Function *F was null at entry to run()");
void *FPtr = getPointerToFunction(F);
assert(FPtr && "Pointer to fn's code was null after getPointerToFunction");
FunctionType *FTy = F->getFunctionType();
Type *RetTy = FTy->getReturnType();
assert((FTy->getNumParams() == ArgValues.size() ||
(FTy->isVarArg() && FTy->getNumParams() <= ArgValues.size())) &&
"Wrong number of arguments passed into function!");
assert(FTy->getNumParams() == ArgValues.size() &&
"This doesn't support passing arguments through varargs (yet)!");
// Handle some common cases first. These cases correspond to common `main'
// prototypes.
if (RetTy->isIntegerTy(32) || RetTy->isVoidTy()) {
switch (ArgValues.size()) {
case 3:
if (FTy->getParamType(0)->isIntegerTy(32) &&
FTy->getParamType(1)->isPointerTy() &&
FTy->getParamType(2)->isPointerTy()) {
int (*PF)(int, char **, const char **) =
(int(*)(int, char **, const char **))(intptr_t)FPtr;
// Call the function.
GenericValue rv;
rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(),
(char **)GVTOP(ArgValues[1]),
(const char **)GVTOP(ArgValues[2])));
return rv;
}
break;
case 2:
if (FTy->getParamType(0)->isIntegerTy(32) &&
FTy->getParamType(1)->isPointerTy()) {
int (*PF)(int, char **) = (int(*)(int, char **))(intptr_t)FPtr;
// Call the function.
GenericValue rv;
rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(),
(char **)GVTOP(ArgValues[1])));
return rv;
}
break;
case 1:
if (FTy->getNumParams() == 1 &&
FTy->getParamType(0)->isIntegerTy(32)) {
GenericValue rv;
int (*PF)(int) = (int(*)(int))(intptr_t)FPtr;
rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue()));
return rv;
}
break;
}
}
// Handle cases where no arguments are passed first.
if (ArgValues.empty()) {
GenericValue rv;
switch (RetTy->getTypeID()) {
default: llvm_unreachable("Unknown return type for function call!");
case Type::IntegerTyID: {
unsigned BitWidth = cast<IntegerType>(RetTy)->getBitWidth();
if (BitWidth == 1)
rv.IntVal = APInt(BitWidth, ((bool(*)())(intptr_t)FPtr)());
else if (BitWidth <= 8)
rv.IntVal = APInt(BitWidth, ((char(*)())(intptr_t)FPtr)());
else if (BitWidth <= 16)
rv.IntVal = APInt(BitWidth, ((short(*)())(intptr_t)FPtr)());
else if (BitWidth <= 32)
rv.IntVal = APInt(BitWidth, ((int(*)())(intptr_t)FPtr)());
else if (BitWidth <= 64)
rv.IntVal = APInt(BitWidth, ((int64_t(*)())(intptr_t)FPtr)());
else
llvm_unreachable("Integer types > 64 bits not supported");
return rv;
}
case Type::VoidTyID:
rv.IntVal = APInt(32, ((int(*)())(intptr_t)FPtr)());
return rv;
case Type::FloatTyID:
rv.FloatVal = ((float(*)())(intptr_t)FPtr)();
return rv;
case Type::DoubleTyID:
rv.DoubleVal = ((double(*)())(intptr_t)FPtr)();
return rv;
case Type::X86_FP80TyID:
case Type::FP128TyID:
case Type::PPC_FP128TyID:
llvm_unreachable("long double not supported yet");
case Type::PointerTyID:
return PTOGV(((void*(*)())(intptr_t)FPtr)());
}
}
llvm_unreachable("Full-featured argument passing not supported yet!");
}
// 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;
}
/// PromotePrivates Implementation - Used to promote pointers from new/delete operations to global segment in HSA
///
bool PromotePrivates::runOnFunction(Function &F) {
if (F.getName().find("cxxamp_trampoline") == StringRef::npos)
return false;
// Need refactor!
Module *M = F.getParent();
if ((M->getFunction(/*"_Znwj"*/ "_Znwm") == NULL) && (M->getFunction(/*"_Znaj"*/ "_Znam") == NULL))
return false;
//errs() << "Execute PromotePrivates::runOnFunction: " << F.getName() << "\n";
LLVMContext& C = F.getContext();
std::vector<Instruction*> NeedPromoted;
for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I) {
NewedMemoryAnalyzer NMA(F);
NMA.analyze(*I);
if (NMA.isNewed())
NeedPromoted.push_back(&*I);
}
#if 0
for (unsigned i = 0; i < NeedPromoted.size(); i++)
errs () << *NeedPromoted[i] << "\n";
#endif
while (!NeedPromoted.empty()) {
Instruction *I = NeedPromoted.back();
#if 0
llvm::errs() << "NeedPromoted:" << *I << "\n";
#endif
if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
IRBuilder<> Builder(SI);
//Value *StoreAddr = Builder.CreatePointerCast(SI->getPointerOperand(), Type::getInt64PtrTy(C, 1));
PointerType *DestTy = SI->getPointerOperand()->getType()->getPointerElementType()->getPointerTo(1);
Value *StoreAddr = Builder.CreatePointerCast(SI->getPointerOperand(), DestTy);
StoreInst* nSI = new StoreInst(SI->getValueOperand(), StoreAddr);
ReplaceInstWithInst(SI, nSI);
}
if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
IRBuilder<> Builder(LI);
Value *LoadAddr = Builder.CreatePointerCast(LI->getPointerOperand(), Type::getInt32PtrTy(C, 1));
LoadInst* nLI = new LoadInst(LoadAddr);
ReplaceInstWithInst(LI, nLI);
}
if (CallInst *CI = dyn_cast<CallInst>(I)) {
IRBuilder<> Builder(CI);
PointerType *DestTy = CI->getArgOperand(0)->getType()->getPointerElementType()->getPointerTo(1);
Value *MemsetAddr = Builder.CreatePointerCast(CI->getArgOperand(0), DestTy);
std::vector<Value*> ArgsVec;
ArgsVec.push_back(MemsetAddr);
for (int i = 1, e = CI->getNumArgOperands(); i < e; i++) {
ArgsVec.push_back(CI->getArgOperand(i));
}
ArrayRef<Value*> Args(ArgsVec);
FunctionType *MemsetFuncType = CI->getCalledFunction()->getFunctionType();
Type *MemsetRetType = MemsetFuncType->getReturnType();
std::vector<Type*> ArgsTypeVec;
ArgsTypeVec.push_back(DestTy);
for (int i = 1, e = MemsetFuncType->getNumParams(); i < e; i++) {
ArgsTypeVec.push_back(MemsetFuncType->getParamType(i));
}
ArrayRef<Type*> ArgsType(ArgsTypeVec);
FunctionType *nMemsetFuncType = FunctionType::get(MemsetRetType, ArgsType, false);
M->getOrInsertFunction("llvm.memset.p1i8.i64", nMemsetFuncType);
Function *MemsetFunc = M->getFunction("llvm.memset.p1i8.i64");
CallInst* nCI = CallInst::Create(MemsetFunc, Args);
ReplaceInstWithInst(CI, nCI);
}
#if 0
if (StoreInst *SI = dyn_cast<StoreInst>(NeedPromoted.back()) ) {
IRBuilder<> Builder(SI);
Value *StoreAddr = Builder.CreatePointerCast(SI->getPointerOperand(), Type::getInt32PtrTy(C, 1));
StoreInst* nSI = new StoreInst(SI->getValueOperand(), StoreAddr);
ReplaceInstWithInst(SI, nSI);
}
else if (LoadInst *LI = dyn_cast<LoadInst>(NeedPromoted.back()) ) {
IRBuilder<> Builder(LI);
Value *StoreAddr = Builder.CreatePointerCast(LI->getPointerOperand(), Type::getInt32PtrTy(C, 1));
LoadInst* nLI = new LoadInst(/*SI->getValueOperand(), */StoreAddr);
ReplaceInstWithInst(LI, nLI);
}
#endif
NeedPromoted.pop_back();
}
//F.dump();
//errs() << "Finished PromotePrivates::runOnFunction\n";
return true;
}
// =============================================================================
// andOOPIsGone (formerly: createProcess)
//
// Formerly, OOP permitted the same SC_{METHOD,THREAD} functions to apply
// to each copy of a SC_MODULE. Aaaaand it's gone !
// (but OTOH we enable better optimizations)
// Creates a new C-style function that calls the old member function with the
// given sc_module. The call is then inlined.
// FIXME: assumes the method is non-virtual and that sc_module is the first
// inherited class of the SC_MODULE
// =============================================================================
Function *TwetoPassImpl::andOOPIsGone(Function * oldProc,
sc_core::sc_module * initiatorMod)
{
if (!oldProc)
return NULL;
// can't statically optimize if the address of the module isn't predictible
// TODO: also handle already-static variables, which also have
// fixed $pc-relative addresses
if (staticopt == optlevel && !permalloc::is_from (initiatorMod))
return NULL;
LLVMContext & context = getGlobalContext();
FunctionType *funType = oldProc->getFunctionType();
Type *type = funType->getParamType(0);
FunctionType *newProcType =
FunctionType::get(oldProc->getReturnType(),
ArrayRef < Type * >(), 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(Attribute::InlineHint);
// Create call to old function
BasicBlock *bb = BasicBlock::Create(context, "entry", newProc);
IRBuilder <> *irb = new IRBuilder <> (context);
irb->SetInsertPoint(bb);
Value* thisAddr = createRelocatablePointer (type, initiatorMod, irb);
CallInst *ci = irb->CreateCall(oldProc,
ArrayRef < Value * >(std::vector<Value*>(1,thisAddr)));
//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);
}
// further optimize the function
inlineBasicIO (initiatorMod, newProc);
newProc->dump();
return newProc;
}
bool SimplifyFortifiedLibCalls::fold(CallInst *CI, const DataLayout *TD,
const TargetLibraryInfo *TLI) {
// We really need DataLayout for later.
if (!TD) return false;
this->CI = CI;
Function *Callee = CI->getCalledFunction();
StringRef Name = Callee->getName();
FunctionType *FT = Callee->getFunctionType();
LLVMContext &Context = CI->getParent()->getContext();
IRBuilder<> B(CI);
if (Name == "__memcpy_chk") {
// Check if this has the right signature.
if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) ||
!FT->getParamType(0)->isPointerTy() ||
!FT->getParamType(1)->isPointerTy() ||
FT->getParamType(2) != TD->getIntPtrType(Context) ||
FT->getParamType(3) != TD->getIntPtrType(Context))
return false;
if (isFoldable(3, 2, false)) {
B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
CI->getArgOperand(2), 1);
replaceCall(CI->getArgOperand(0));
return true;
}
return false;
}
// Should be similar to memcpy.
if (Name == "__mempcpy_chk") {
return false;
}
if (Name == "__memmove_chk") {
// Check if this has the right signature.
if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) ||
!FT->getParamType(0)->isPointerTy() ||
!FT->getParamType(1)->isPointerTy() ||
FT->getParamType(2) != TD->getIntPtrType(Context) ||
FT->getParamType(3) != TD->getIntPtrType(Context))
return false;
if (isFoldable(3, 2, false)) {
B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
CI->getArgOperand(2), 1);
replaceCall(CI->getArgOperand(0));
return true;
}
return false;
}
if (Name == "__memset_chk") {
// Check if this has the right signature.
if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) ||
!FT->getParamType(0)->isPointerTy() ||
!FT->getParamType(1)->isIntegerTy() ||
FT->getParamType(2) != TD->getIntPtrType(Context) ||
FT->getParamType(3) != TD->getIntPtrType(Context))
return false;
if (isFoldable(3, 2, false)) {
Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(),
false);
B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
replaceCall(CI->getArgOperand(0));
return true;
}
return false;
}
if (Name == "__strcpy_chk" || Name == "__stpcpy_chk") {
// Check if this has the right signature.
if (FT->getNumParams() != 3 ||
FT->getReturnType() != FT->getParamType(0) ||
FT->getParamType(0) != FT->getParamType(1) ||
FT->getParamType(0) != Type::getInt8PtrTy(Context) ||
FT->getParamType(2) != TD->getIntPtrType(Context))
return 0;
// If a) we don't have any length information, or b) we know this will
// fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
// st[rp]cpy_chk call which may fail at runtime if the size is too long.
// TODO: It might be nice to get a maximum length out of the possible
// string lengths for varying.
if (isFoldable(2, 1, true)) {
Value *Ret = EmitStrCpy(CI->getArgOperand(0), CI->getArgOperand(1), B, TD,
TLI, Name.substr(2, 6));
if (!Ret)
return false;
replaceCall(Ret);
return true;
}
return false;
}
if (Name == "__strncpy_chk" || Name == "__stpncpy_chk") {
// Check if this has the right signature.
if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) ||
FT->getParamType(0) != FT->getParamType(1) ||
FT->getParamType(0) != Type::getInt8PtrTy(Context) ||
!FT->getParamType(2)->isIntegerTy() ||
FT->getParamType(3) != TD->getIntPtrType(Context))
return false;
if (isFoldable(3, 2, false)) {
Value *Ret = EmitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
CI->getArgOperand(2), B, TD, TLI,
Name.substr(2, 7));
if (!Ret)
return false;
replaceCall(Ret);
return true;
}
return false;
}
if (Name == "__strcat_chk") {
return false;
}
if (Name == "__strncat_chk") {
return false;
}
return false;
}
/// DoPromotion - This method actually performs the promotion of the specified
/// arguments, and returns the new function. At this point, we know that it's
/// safe to do so.
CallGraphNode *ArgPromotion::DoPromotion(Function *F,
SmallPtrSetImpl<Argument*> &ArgsToPromote,
SmallPtrSetImpl<Argument*> &ByValArgsToTransform) {
// Start by computing a new prototype for the function, which is the same as
// the old function, but has modified arguments.
FunctionType *FTy = F->getFunctionType();
std::vector<Type*> Params;
typedef std::set<IndicesVector> ScalarizeTable;
// ScalarizedElements - If we are promoting a pointer that has elements
// accessed out of it, keep track of which elements are accessed so that we
// can add one argument for each.
//
// Arguments that are directly loaded will have a zero element value here, to
// handle cases where there are both a direct load and GEP accesses.
//
std::map<Argument*, ScalarizeTable> ScalarizedElements;
// OriginalLoads - Keep track of a representative load instruction from the
// original function so that we can tell the alias analysis implementation
// what the new GEP/Load instructions we are inserting look like.
// We need to keep the original loads for each argument and the elements
// of the argument that are accessed.
std::map<std::pair<Argument*, IndicesVector>, LoadInst*> OriginalLoads;
// Attribute - Keep track of the parameter attributes for the arguments
// that we are *not* promoting. For the ones that we do promote, the parameter
// attributes are lost
SmallVector<AttributeSet, 8> AttributesVec;
const AttributeSet &PAL = F->getAttributes();
// Add any return attributes.
if (PAL.hasAttributes(AttributeSet::ReturnIndex))
AttributesVec.push_back(AttributeSet::get(F->getContext(),
PAL.getRetAttributes()));
// First, determine the new argument list
unsigned ArgIndex = 1;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
++I, ++ArgIndex) {
if (ByValArgsToTransform.count(I)) {
// Simple byval argument? Just add all the struct element types.
Type *AgTy = cast<PointerType>(I->getType())->getElementType();
StructType *STy = cast<StructType>(AgTy);
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
Params.push_back(STy->getElementType(i));
++NumByValArgsPromoted;
} else if (!ArgsToPromote.count(I)) {
// Unchanged argument
Params.push_back(I->getType());
AttributeSet attrs = PAL.getParamAttributes(ArgIndex);
if (attrs.hasAttributes(ArgIndex)) {
AttrBuilder B(attrs, ArgIndex);
AttributesVec.
push_back(AttributeSet::get(F->getContext(), Params.size(), B));
}
} else if (I->use_empty()) {
// Dead argument (which are always marked as promotable)
++NumArgumentsDead;
} else {
// Okay, this is being promoted. This means that the only uses are loads
// or GEPs which are only used by loads
// In this table, we will track which indices are loaded from the argument
// (where direct loads are tracked as no indices).
ScalarizeTable &ArgIndices = ScalarizedElements[I];
for (User *U : I->users()) {
Instruction *UI = cast<Instruction>(U);
assert(isa<LoadInst>(UI) || isa<GetElementPtrInst>(UI));
IndicesVector Indices;
Indices.reserve(UI->getNumOperands() - 1);
// Since loads will only have a single operand, and GEPs only a single
// non-index operand, this will record direct loads without any indices,
// and gep+loads with the GEP indices.
for (User::op_iterator II = UI->op_begin() + 1, IE = UI->op_end();
II != IE; ++II)
Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
// GEPs with a single 0 index can be merged with direct loads
if (Indices.size() == 1 && Indices.front() == 0)
Indices.clear();
ArgIndices.insert(Indices);
LoadInst *OrigLoad;
if (LoadInst *L = dyn_cast<LoadInst>(UI))
OrigLoad = L;
else
// Take any load, we will use it only to update Alias Analysis
OrigLoad = cast<LoadInst>(UI->user_back());
OriginalLoads[std::make_pair(I, Indices)] = OrigLoad;
}
// Add a parameter to the function for each element passed in.
for (ScalarizeTable::iterator SI = ArgIndices.begin(),
E = ArgIndices.end(); SI != E; ++SI) {
// not allowed to dereference ->begin() if size() is 0
Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI));
assert(Params.back());
}
if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
++NumArgumentsPromoted;
else
++NumAggregatesPromoted;
}
}
// Add any function attributes.
if (PAL.hasAttributes(AttributeSet::FunctionIndex))
AttributesVec.push_back(AttributeSet::get(FTy->getContext(),
PAL.getFnAttributes()));
Type *RetTy = FTy->getReturnType();
// Construct the new function type using the new arguments.
FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
// Create the new function body and insert it into the module.
Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
NF->copyAttributesFrom(F);
// Patch the pointer to LLVM function in debug info descriptor.
auto DI = FunctionDIs.find(F);
if (DI != FunctionDIs.end()) {
DISubprogram SP = DI->second;
SP.replaceFunction(NF);
// Ensure the map is updated so it can be reused on subsequent argument
// promotions of the same function.
FunctionDIs.erase(DI);
FunctionDIs[NF] = SP;
}
DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
<< "From: " << *F);
// Recompute the parameter attributes list based on the new arguments for
// the function.
NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec));
AttributesVec.clear();
F->getParent()->getFunctionList().insert(F, NF);
NF->takeName(F);
// Get the alias analysis information that we need to update to reflect our
// changes.
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
// Get the callgraph information that we need to update to reflect our
// changes.
CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
// Get a new callgraph node for NF.
CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
// Loop over all of the callers of the function, transforming the call sites
// to pass in the loaded pointers.
//
SmallVector<Value*, 16> Args;
while (!F->use_empty()) {
CallSite CS(F->user_back());
assert(CS.getCalledFunction() == F);
Instruction *Call = CS.getInstruction();
const AttributeSet &CallPAL = CS.getAttributes();
// Add any return attributes.
if (CallPAL.hasAttributes(AttributeSet::ReturnIndex))
AttributesVec.push_back(AttributeSet::get(F->getContext(),
CallPAL.getRetAttributes()));
// Loop over the operands, inserting GEP and loads in the caller as
// appropriate.
CallSite::arg_iterator AI = CS.arg_begin();
ArgIndex = 1;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
I != E; ++I, ++AI, ++ArgIndex)
if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
Args.push_back(*AI); // Unmodified argument
if (CallPAL.hasAttributes(ArgIndex)) {
AttrBuilder B(CallPAL, ArgIndex);
AttributesVec.
push_back(AttributeSet::get(F->getContext(), Args.size(), B));
}
} else if (ByValArgsToTransform.count(I)) {
// Emit a GEP and load for each element of the struct.
Type *AgTy = cast<PointerType>(I->getType())->getElementType();
StructType *STy = cast<StructType>(AgTy);
Value *Idxs[2] = {
ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
Value *Idx = GetElementPtrInst::Create(*AI, Idxs,
(*AI)->getName()+"."+utostr(i),
Call);
// TODO: Tell AA about the new values?
Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
}
} else if (!I->use_empty()) {
// Non-dead argument: insert GEPs and loads as appropriate.
ScalarizeTable &ArgIndices = ScalarizedElements[I];
// Store the Value* version of the indices in here, but declare it now
// for reuse.
std::vector<Value*> Ops;
for (ScalarizeTable::iterator SI = ArgIndices.begin(),
E = ArgIndices.end(); SI != E; ++SI) {
Value *V = *AI;
LoadInst *OrigLoad = OriginalLoads[std::make_pair(I, *SI)];
if (!SI->empty()) {
Ops.reserve(SI->size());
Type *ElTy = V->getType();
for (IndicesVector::const_iterator II = SI->begin(),
IE = SI->end(); II != IE; ++II) {
// Use i32 to index structs, and i64 for others (pointers/arrays).
// This satisfies GEP constraints.
Type *IdxTy = (ElTy->isStructTy() ?
Type::getInt32Ty(F->getContext()) :
Type::getInt64Ty(F->getContext()));
Ops.push_back(ConstantInt::get(IdxTy, *II));
// Keep track of the type we're currently indexing.
ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
}
// And create a GEP to extract those indices.
V = GetElementPtrInst::Create(V, Ops, V->getName()+".idx", Call);
Ops.clear();
AA.copyValue(OrigLoad->getOperand(0), V);
}
// Since we're replacing a load make sure we take the alignment
// of the previous load.
LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
newLoad->setAlignment(OrigLoad->getAlignment());
// Transfer the AA info too.
AAMDNodes AAInfo;
OrigLoad->getAAMetadata(AAInfo);
newLoad->setAAMetadata(AAInfo);
Args.push_back(newLoad);
AA.copyValue(OrigLoad, Args.back());
}
}
// Push any varargs arguments on the list.
for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
Args.push_back(*AI);
if (CallPAL.hasAttributes(ArgIndex)) {
AttrBuilder B(CallPAL, ArgIndex);
AttributesVec.
push_back(AttributeSet::get(F->getContext(), Args.size(), B));
}
}
// Add any function attributes.
if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
AttributesVec.push_back(AttributeSet::get(Call->getContext(),
CallPAL.getFnAttributes()));
Instruction *New;
if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
Args, "", Call);
cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(),
AttributesVec));
} else {
New = CallInst::Create(NF, Args, "", Call);
cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(),
AttributesVec));
if (cast<CallInst>(Call)->isTailCall())
cast<CallInst>(New)->setTailCall();
}
New->setDebugLoc(Call->getDebugLoc());
Args.clear();
AttributesVec.clear();
// Update the alias analysis implementation to know that we are replacing
// the old call with a new one.
AA.replaceWithNewValue(Call, New);
// Update the callgraph to know that the callsite has been transformed.
CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
CalleeNode->replaceCallEdge(Call, New, NF_CGN);
if (!Call->use_empty()) {
Call->replaceAllUsesWith(New);
New->takeName(Call);
}
// Finally, remove the old call from the program, reducing the use-count of
// F.
Call->eraseFromParent();
}
// Since we have now created the new function, splice the body of the old
// function right into the new function, leaving the old rotting hulk of the
// function empty.
NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
// Loop over the argument list, transferring uses of the old arguments over to
// the new arguments, also transferring over the names as well.
//
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
I2 = NF->arg_begin(); I != E; ++I) {
if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
// If this is an unmodified argument, move the name and users over to the
// new version.
I->replaceAllUsesWith(I2);
I2->takeName(I);
AA.replaceWithNewValue(I, I2);
++I2;
continue;
}
if (ByValArgsToTransform.count(I)) {
// In the callee, we create an alloca, and store each of the new incoming
// arguments into the alloca.
Instruction *InsertPt = NF->begin()->begin();
// Just add all the struct element types.
Type *AgTy = cast<PointerType>(I->getType())->getElementType();
Value *TheAlloca = new AllocaInst(AgTy, nullptr, "", InsertPt);
StructType *STy = cast<StructType>(AgTy);
Value *Idxs[2] = {
ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
Value *Idx =
GetElementPtrInst::Create(TheAlloca, Idxs,
TheAlloca->getName()+"."+Twine(i),
InsertPt);
I2->setName(I->getName()+"."+Twine(i));
new StoreInst(I2++, Idx, InsertPt);
}
// Anything that used the arg should now use the alloca.
I->replaceAllUsesWith(TheAlloca);
TheAlloca->takeName(I);
AA.replaceWithNewValue(I, TheAlloca);
// If the alloca is used in a call, we must clear the tail flag since
// the callee now uses an alloca from the caller.
for (User *U : TheAlloca->users()) {
CallInst *Call = dyn_cast<CallInst>(U);
if (!Call)
continue;
Call->setTailCall(false);
}
continue;
}
if (I->use_empty()) {
AA.deleteValue(I);
continue;
}
// Otherwise, if we promoted this argument, then all users are load
// instructions (or GEPs with only load users), and all loads should be
// using the new argument that we added.
ScalarizeTable &ArgIndices = ScalarizedElements[I];
while (!I->use_empty()) {
if (LoadInst *LI = dyn_cast<LoadInst>(I->user_back())) {
assert(ArgIndices.begin()->empty() &&
"Load element should sort to front!");
I2->setName(I->getName()+".val");
LI->replaceAllUsesWith(I2);
AA.replaceWithNewValue(LI, I2);
LI->eraseFromParent();
DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
<< "' in function '" << F->getName() << "'\n");
} else {
GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->user_back());
IndicesVector Operands;
Operands.reserve(GEP->getNumIndices());
for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
II != IE; ++II)
Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
// GEPs with a single 0 index can be merged with direct loads
if (Operands.size() == 1 && Operands.front() == 0)
Operands.clear();
Function::arg_iterator TheArg = I2;
for (ScalarizeTable::iterator It = ArgIndices.begin();
*It != Operands; ++It, ++TheArg) {
assert(It != ArgIndices.end() && "GEP not handled??");
}
std::string NewName = I->getName();
for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
NewName += "." + utostr(Operands[i]);
}
NewName += ".val";
TheArg->setName(NewName);
DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
<< "' of function '" << NF->getName() << "'\n");
// All of the uses must be load instructions. Replace them all with
// the argument specified by ArgNo.
while (!GEP->use_empty()) {
LoadInst *L = cast<LoadInst>(GEP->user_back());
L->replaceAllUsesWith(TheArg);
AA.replaceWithNewValue(L, TheArg);
L->eraseFromParent();
}
AA.deleteValue(GEP);
GEP->eraseFromParent();
}
}
// Increment I2 past all of the arguments added for this promoted pointer.
std::advance(I2, ArgIndices.size());
}
// Tell the alias analysis that the old function is about to disappear.
AA.replaceWithNewValue(F, NF);
NF_CGN->stealCalledFunctionsFrom(CG[F]);
// Now that the old function is dead, delete it. If there is a dangling
// reference to the CallgraphNode, just leave the dead function around for
// someone else to nuke.
CallGraphNode *CGN = CG[F];
if (CGN->getNumReferences() == 0)
delete CG.removeFunctionFromModule(CGN);
else
F->setLinkage(Function::ExternalLinkage);
return NF_CGN;
}
