void Aa::LowerConstantExpr::process(Instruction *inst)
{
if(isa<llvm::CallInst>(inst))
{
CallInst* ci = dyn_cast<CallInst>(inst);
for(int idx = 0; idx < ci->getNumArgOperands(); idx++)
{
llvm::Value *opnd = ci->getArgOperand(idx);
if (ConstantExpr *ce = dyn_cast<ConstantExpr>(opnd)) {
lower(ce, inst);
}
}
}
else
{
for (unsigned oi = 0, oe = inst->getNumOperands();
oi != oe; oi++) {
llvm::Value *opnd = inst->getOperand(oi);
if (ConstantExpr *ce = dyn_cast<ConstantExpr>(opnd)) {
lower(ce, inst);
}
}
}
}
/// \brief Check call has a unary float signature
/// It checks following:
/// a) call should have a single argument
/// b) argument type should be floating point type
/// c) call instruction type and argument type should be same
/// d) call should only reads memory.
/// If all these condition is met then return ValidIntrinsicID
/// else return not_intrinsic.
Intrinsic::ID
llvm::checkUnaryFloatSignature(const CallInst &I,
Intrinsic::ID ValidIntrinsicID) {
if (I.getNumArgOperands() != 1 ||
!I.getArgOperand(0)->getType()->isFloatingPointTy() ||
I.getType() != I.getArgOperand(0)->getType() || !I.onlyReadsMemory())
return Intrinsic::not_intrinsic;
return ValidIntrinsicID;
}
bool EncoderPass::isAtoiFunction(Instruction *I) {
if(I->getOpcode()==Instruction::Add && I->getName().substr(0, 1) == "_") {
CallInst *call = dyn_cast<CallInst>(I->getOperand(0));
if(call && call->getNumArgOperands()==1) {
Function *F = call->getCalledFunction();
if (F && F->getName()=="atoi") {
return true;
}
}
}
return false;
}
/// AddCatchInfo - Extract the personality and type infos from an eh.selector
/// call, and add them to the specified machine basic block.
void llvm::AddCatchInfo(const CallInst &I, MachineModuleInfo *MMI,
MachineBasicBlock *MBB) {
// Inform the MachineModuleInfo of the personality for this landing pad.
const ConstantExpr *CE = cast<ConstantExpr>(I.getArgOperand(1));
assert(CE->getOpcode() == Instruction::BitCast &&
isa<Function>(CE->getOperand(0)) &&
"Personality should be a function");
MMI->addPersonality(MBB, cast<Function>(CE->getOperand(0)));
// Gather all the type infos for this landing pad and pass them along to
// MachineModuleInfo.
std::vector<const GlobalVariable *> TyInfo;
unsigned N = I.getNumArgOperands();
for (unsigned i = N - 1; i > 1; --i) {
if (const ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(i))) {
unsigned FilterLength = CI->getZExtValue();
unsigned FirstCatch = i + FilterLength + !FilterLength;
assert(FirstCatch <= N && "Invalid filter length");
if (FirstCatch < N) {
TyInfo.reserve(N - FirstCatch);
for (unsigned j = FirstCatch; j < N; ++j)
TyInfo.push_back(ExtractTypeInfo(I.getArgOperand(j)));
MMI->addCatchTypeInfo(MBB, TyInfo);
TyInfo.clear();
}
if (!FilterLength) {
// Cleanup.
MMI->addCleanup(MBB);
} else {
// Filter.
TyInfo.reserve(FilterLength - 1);
for (unsigned j = i + 1; j < FirstCatch; ++j)
TyInfo.push_back(ExtractTypeInfo(I.getArgOperand(j)));
MMI->addFilterTypeInfo(MBB, TyInfo);
TyInfo.clear();
}
N = i;
}
}
if (N > 2) {
TyInfo.reserve(N - 2);
for (unsigned j = 2; j < N; ++j)
TyInfo.push_back(ExtractTypeInfo(I.getArgOperand(j)));
MMI->addCatchTypeInfo(MBB, TyInfo);
}
}
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;
}
}
}
}
}
/// 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 (po_iterator<Type*> i = po_begin(T), e = po_end(T);
i != e; ++i) {
if (i->isFloatingPointTy()) {
MMI->setUsesVAFloatArgument(true);
return;
}
}
}
}
}
void ExtractContracts::validateAnnotation(CallInst &I) {
assert(I.getCalledFunction() && "Unexpected virtual function.");
assert(I.getNumArgOperands() == 1 && "Unexpected operands.");
auto B = I.getParent();
auto F = B->getParent();
auto& DT = getAnalysis<DominatorTreeWrapperPass>(*F).getDomTree();
auto& LI = getAnalysis<LoopInfoWrapperPass>(*F).getLoopInfo();
if (I.getCalledFunction()->getName() == Naming::CONTRACT_INVARIANT) {
auto L = LI[B];
if (!L) {
llvm_unreachable("Loop invariants must occur inside loops.");
}
} else {
for (auto L : LI) {
auto& J = L->getHeader()->getInstList().front();
if (DT.dominates(&J, &I)) {
llvm_unreachable("Procedure specifications must occur before loops.");
}
}
}
}
void RuntimeHelperFixupPass::FixStringConstructor(Function *old_construct, Function *new_construct)
{
for (auto it = old_construct->use_begin(), end = old_construct->use_end(); it != end; ++it)
{
CallInst *CI = dyn_cast<CallInst>(*it);
assert (CI);
auto num_ops = CI->getNumArgOperands();
SmallVector<Value*, 4> ops;
for (size_t i = 1; i < num_ops; ++i)
ops.push_back(CI->getArgOperand(i));
IRBuilder<> builder(CI);
auto new_call = builder.CreateCall(new_construct, ops);
BitCastInst *this_ptr = cast<BitCastInst>(CI->getArgOperand(0));
auto alloc = cast<Instruction>(this_ptr->llvm::User::getOperand(0));
this_ptr->replaceAllUsesWith(new_call);
CI->eraseFromParent();
this_ptr->eraseFromParent();
alloc->eraseFromParent();
}
}
bool FuncAddrTaken::runOnModule(Module &M) {
bool Changed = false;
// add declaration of function __patch_at
// declare void @__patch_at(void)
FunctionType *FT = FunctionType::get(Type::getVoidTy(M.getContext()),
false);
Function* PatchAt = Function::Create(FT, Function::ExternalLinkage, "__patch_at", &M);
if (PatchAt->getName() != "__patch_at") {
PatchAt->eraseFromParent();
return false;
}
Changed = true;
// Simple optimization so that no function address will be taken twice
// in the same basic block.
std::map<BasicBlock*, std::set<std::string> > UniqPatchAt;
// before each store instruction that manipulates a function, create a call
// to __patch_at
for (auto F = M.getFunctionList().begin(); F != M.getFunctionList().end(); F++) {
for (auto BB = F->begin(); BB != F->end(); BB++) {
for (auto MI = BB->begin(); MI != BB->end(); MI++) {
if (isa<StoreInst>(MI)) {
// check if the store inst moves a function to a variable
Value *V = InnerMost(cast<StoreInst>(MI)->getValueOperand());
addPatchAt(M, FT, V, MI, UniqPatchAt);
if (isa<ConstantVector>(V)) {
std::set<Value*> patched;
for (unsigned i = 0; i < cast<ConstantVector>(V)->getNumOperands(); i++) {
Value *VV = InnerMost(cast<ConstantVector>(V)->getOperand(i));
if (patched.find(VV) == patched.end()) {
addPatchAt(M, FT, VV, MI, UniqPatchAt);
patched.insert(VV);
}
}
} else if (isa<ConstantStruct>(V)) {
std::set<Value*> patched;
for (unsigned i = 0; i < cast<ConstantStruct>(V)->getNumOperands(); i++) {
Value *VV = InnerMost(cast<ConstantStruct>(V)->getOperand(i));
if (patched.find(VV) == patched.end()) {
addPatchAt(M, FT, VV, MI, UniqPatchAt);
patched.insert(VV);
}
}
} else if (isa<ConstantArray>(V)) {
std::set<Value*> patched;
for (unsigned i = 0; i < cast<ConstantArray>(V)->getNumOperands(); i++) {
Value *VV = InnerMost(cast<ConstantArray>(V)->getOperand(i));
if (patched.find(VV) == patched.end()) {
addPatchAt(M, FT, VV, MI, UniqPatchAt);
patched.insert(VV);
}
}
}
} else if (isa<SelectInst>(MI)) {
Value *V = InnerMost(cast<SelectInst>(MI)->getTrueValue());
addPatchAt(M, FT, V, MI, UniqPatchAt);
V = InnerMost(cast<SelectInst>(MI)->getFalseValue());
addPatchAt(M, FT, V, MI, UniqPatchAt);
} else if (isa<CallInst>(MI)) {
CallInst* CI = cast<CallInst>(MI);
for (unsigned i = 0; i < CI->getNumArgOperands(); i++) {
Value *V = InnerMost(CI->getArgOperand(i));
addPatchAt(M, FT, V, MI, UniqPatchAt);
}
} else if (isa<InvokeInst>(MI)) {
InvokeInst* CI = cast<InvokeInst>(MI);
for (unsigned i = 0; i < CI->getNumArgOperands(); i++) {
Value *V = InnerMost(CI->getArgOperand(i));
addPatchAt(M, FT, V, MI, UniqPatchAt);
}
} else if (isa<ReturnInst>(MI)) {
Value *V = cast<ReturnInst>(MI)->getReturnValue();
if (V) {
V = InnerMost(V);
addPatchAt(M, FT, V, MI, UniqPatchAt);
}
} else if (isa<PHINode>(MI)) {
for (unsigned i = 0; i < cast<PHINode>(MI)->getNumIncomingValues(); i++) {
Value *V = InnerMost(cast<PHINode>(MI)->getIncomingValue(i));
BasicBlock* BB = cast<PHINode>(MI)->getIncomingBlock(i);
// right before the last (maybe terminator) instruction.
addPatchAt(M, FT, V, &(BB->back()), UniqPatchAt);
}
}
}
}
}
// TODO: separate the following virtual table traversal code into another pass
for (auto G = M.getGlobalList().begin(); G != M.getGlobalList().end(); G++) {
if (isa<GlobalVariable>(G) &&
cast<GlobalVariable>(G)->hasInitializer()) {
const GlobalVariable *GV = cast<GlobalVariable>(G);
const Constant *C = GV->getInitializer();
if (GV->hasName() && isa<ConstantArray>(C) && GV->getName().startswith("_ZTV")) {
std::string VTName = CXXDemangledName(GV->getName().data());
if (VTName.size() && VTName.find("vtable") == 0) {
//llvm::errs() << VTName << "\n";
VTName = VTName.substr(11); // get pass "vtable for "
llvm::NamedMDNode* MD = M.getOrInsertNamedMetadata("MCFIVtable");
std::string info = VTName;
for (unsigned i = 0; i < cast<ConstantArray>(C)->getNumOperands(); i++) {
Value *V = InnerMost(cast<ConstantArray>(C)->getOperand(i));
if (isa<Function>(V) && cast<Function>(V)->hasName()) {
//llvm::errs() << cast<Function>(V)->getName() << "\n";
info += std::string("#") + cast<Function>(V)->getName().str();
}
}
MD->addOperand(llvm::MDNode::get(M.getContext(),
llvm::MDString::get(
M.getContext(), info.c_str())));
}
}
}
}
return Changed;
}
void AMDGPUPromoteAlloca::handleAlloca(AllocaInst &I) {
// Array allocations are probably not worth handling, since an allocation of
// the array type is the canonical form.
if (!I.isStaticAlloca() || I.isArrayAllocation())
return;
IRBuilder<> Builder(&I);
// First try to replace the alloca with a vector
Type *AllocaTy = I.getAllocatedType();
DEBUG(dbgs() << "Trying to promote " << I << '\n');
if (tryPromoteAllocaToVector(&I))
return;
DEBUG(dbgs() << " alloca is not a candidate for vectorization.\n");
const Function &ContainingFunction = *I.getParent()->getParent();
// FIXME: We should also try to get this value from the reqd_work_group_size
// function attribute if it is available.
unsigned WorkGroupSize = AMDGPU::getMaximumWorkGroupSize(ContainingFunction);
int AllocaSize =
WorkGroupSize * Mod->getDataLayout().getTypeAllocSize(AllocaTy);
if (AllocaSize > LocalMemAvailable) {
DEBUG(dbgs() << " Not enough local memory to promote alloca.\n");
return;
}
std::vector<Value*> WorkList;
if (!collectUsesWithPtrTypes(&I, WorkList)) {
DEBUG(dbgs() << " Do not know how to convert all uses\n");
return;
}
DEBUG(dbgs() << "Promoting alloca to local memory\n");
LocalMemAvailable -= AllocaSize;
Function *F = I.getParent()->getParent();
Type *GVTy = ArrayType::get(I.getAllocatedType(), WorkGroupSize);
GlobalVariable *GV = new GlobalVariable(
*Mod, GVTy, false, GlobalValue::InternalLinkage,
UndefValue::get(GVTy),
Twine(F->getName()) + Twine('.') + I.getName(),
nullptr,
GlobalVariable::NotThreadLocal,
AMDGPUAS::LOCAL_ADDRESS);
GV->setUnnamedAddr(true);
GV->setAlignment(I.getAlignment());
Value *TCntY, *TCntZ;
std::tie(TCntY, TCntZ) = getLocalSizeYZ(Builder);
Value *TIdX = getWorkitemID(Builder, 0);
Value *TIdY = getWorkitemID(Builder, 1);
Value *TIdZ = getWorkitemID(Builder, 2);
Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ, "", true, true);
Tmp0 = Builder.CreateMul(Tmp0, TIdX);
Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ, "", true, true);
Value *TID = Builder.CreateAdd(Tmp0, Tmp1);
TID = Builder.CreateAdd(TID, TIdZ);
Value *Indices[] = {
Constant::getNullValue(Type::getInt32Ty(Mod->getContext())),
TID
};
Value *Offset = Builder.CreateInBoundsGEP(GVTy, GV, Indices);
I.mutateType(Offset->getType());
I.replaceAllUsesWith(Offset);
I.eraseFromParent();
for (Value *V : WorkList) {
CallInst *Call = dyn_cast<CallInst>(V);
if (!Call) {
Type *EltTy = V->getType()->getPointerElementType();
PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
// The operand's value should be corrected on its own.
if (isa<AddrSpaceCastInst>(V))
continue;
// FIXME: It doesn't really make sense to try to do this for all
// instructions.
V->mutateType(NewTy);
continue;
}
IntrinsicInst *Intr = dyn_cast<IntrinsicInst>(Call);
if (!Intr) {
// FIXME: What is this for? It doesn't make sense to promote arbitrary
// function calls. If the call is to a defined function that can also be
// promoted, we should be able to do this once that function is also
// rewritten.
std::vector<Type*> ArgTypes;
for (unsigned ArgIdx = 0, ArgEnd = Call->getNumArgOperands();
ArgIdx != ArgEnd; ++ArgIdx) {
ArgTypes.push_back(Call->getArgOperand(ArgIdx)->getType());
}
Function *F = Call->getCalledFunction();
FunctionType *NewType = FunctionType::get(Call->getType(), ArgTypes,
F->isVarArg());
Constant *C = Mod->getOrInsertFunction((F->getName() + ".local").str(),
NewType, F->getAttributes());
Function *NewF = cast<Function>(C);
Call->setCalledFunction(NewF);
continue;
}
Builder.SetInsertPoint(Intr);
switch (Intr->getIntrinsicID()) {
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
// These intrinsics are for address space 0 only
Intr->eraseFromParent();
continue;
case Intrinsic::memcpy: {
MemCpyInst *MemCpy = cast<MemCpyInst>(Intr);
Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getRawSource(),
MemCpy->getLength(), MemCpy->getAlignment(),
MemCpy->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memmove: {
MemMoveInst *MemMove = cast<MemMoveInst>(Intr);
Builder.CreateMemMove(MemMove->getRawDest(), MemMove->getRawSource(),
MemMove->getLength(), MemMove->getAlignment(),
MemMove->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memset: {
MemSetInst *MemSet = cast<MemSetInst>(Intr);
Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(),
MemSet->getLength(), MemSet->getAlignment(),
MemSet->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::invariant_start:
case Intrinsic::invariant_end:
case Intrinsic::invariant_group_barrier:
Intr->eraseFromParent();
// FIXME: I think the invariant marker should still theoretically apply,
// but the intrinsics need to be changed to accept pointers with any
// address space.
continue;
case Intrinsic::objectsize: {
Value *Src = Intr->getOperand(0);
Type *SrcTy = Src->getType()->getPointerElementType();
Function *ObjectSize = Intrinsic::getDeclaration(Mod,
Intrinsic::objectsize,
{ Intr->getType(), PointerType::get(SrcTy, AMDGPUAS::LOCAL_ADDRESS) }
);
CallInst *NewCall
= Builder.CreateCall(ObjectSize, { Src, Intr->getOperand(1) });
Intr->replaceAllUsesWith(NewCall);
Intr->eraseFromParent();
continue;
}
default:
Intr->dump();
llvm_unreachable("Don't know how to promote alloca intrinsic use.");
}
}
}
void DuettoNativeRewriter::rewriteConstructorImplementation(Module& M, Function& F)
{
//Copy the code in a function with the right signature
Function* newFunc=getReturningConstructor(M, &F);
if(!newFunc->empty())
return;
//Visit each instruction and take note of the ones that needs to be replaced
Function::const_iterator B=F.begin();
Function::const_iterator BE=F.end();
ValueToValueMapTy valueMap;
CallInst* lowerConstructor = NULL;
const CallInst* oldLowerConstructor = NULL;
for(;B!=BE;++B)
{
BasicBlock::const_iterator I=B->begin();
BasicBlock::const_iterator IE=B->end();
for(;I!=IE;++I)
{
if(I->getOpcode()!=Instruction::Call)
continue;
const CallInst* callInst=cast<CallInst>(&(*I));
Function* f=callInst->getCalledFunction();
if(!f)
continue;
const char* startOfType;
const char* endOfType;
if(!DuettoNativeRewriter::isBuiltinConstructor(f->getName().data(), startOfType, endOfType))
continue;
//Check that the constructor is for 'this'
if(callInst->getOperand(0)!=F.arg_begin())
continue;
//If this is another constructor for the same type, change it to a
//returning constructor and use it as the 'this' argument
Function* newFunc = getReturningConstructor(M, f);
llvm::SmallVector<Value*, 4> newArgs;
for(unsigned i=1;i<callInst->getNumArgOperands();i++)
newArgs.push_back(callInst->getArgOperand(i));
lowerConstructor = CallInst::Create(newFunc, newArgs);
oldLowerConstructor = callInst;
break;
}
if(lowerConstructor)
break;
}
//Clone the linkage first
newFunc->setLinkage(F.getLinkage());
Function::arg_iterator origArg=++F.arg_begin();
Function::arg_iterator newArg=newFunc->arg_begin();
valueMap.insert(make_pair(F.arg_begin(), lowerConstructor));
for(unsigned i=1;i<F.arg_size();i++)
{
valueMap.insert(make_pair(&(*origArg), &(*newArg)));
++origArg;
++newArg;
}
SmallVector<ReturnInst*, 4> returns;
CloneFunctionInto(newFunc, &F, valueMap, false, returns);
//Find the right place to add the base construtor call
assert(lowerConstructor->getNumArgOperands()<=1 && "Native constructors with multiple args are not supported");
Instruction* callPred = NULL;
if (lowerConstructor->getNumArgOperands()==1 && Instruction::classof(lowerConstructor->getArgOperand(0)))
{
//Switch the argument to the one in the new func
lowerConstructor->setArgOperand(0, valueMap[lowerConstructor->getArgOperand(0)]);
callPred = cast<Instruction>(lowerConstructor->getArgOperand(0));
}
else
callPred = &newFunc->getEntryBlock().front();
//Add add it
lowerConstructor->insertAfter(callPred);
//Override the returs values
for(unsigned i=0;i<returns.size();i++)
{
Instruction* newInst = ReturnInst::Create(M.getContext(),lowerConstructor);
newInst->insertBefore(returns[i]);
returns[i]->removeFromParent();
}
//Recursively move all the users of the lower constructor after the call itself
Instruction* insertPoint = lowerConstructor->getNextNode();
SmallVector<Value*, 4> usersQueue(lowerConstructor->getNumUses());
unsigned int i;
Value::use_iterator it;
for(i=usersQueue.size()-1,it=lowerConstructor->use_begin();it!=lowerConstructor->use_end();++it,i--)
usersQueue[i]=it->getUser();
SmallSet<Instruction*, 4> movedInstructions;
while(!usersQueue.empty())
{
Instruction* cur=dyn_cast<Instruction>(usersQueue.pop_back_val());
if(!cur)
continue;
if(movedInstructions.count(cur))
continue;
movedInstructions.insert(cur);
cur->moveBefore(insertPoint);
//Add users of this instrucution as well
usersQueue.resize(usersQueue.size()+cur->getNumUses());
for(i=usersQueue.size()-1,it=cur->use_begin();it!=cur->use_end();++it,i--)
usersQueue[i]=it->getUser();
}
cast<Instruction>(valueMap[oldLowerConstructor])->eraseFromParent();
}
void AMDGPUPromoteAlloca::visitAlloca(AllocaInst &I) {
IRBuilder<> Builder(&I);
// First try to replace the alloca with a vector
Type *AllocaTy = I.getAllocatedType();
DEBUG(dbgs() << "Trying to promote " << I << '\n');
if (tryPromoteAllocaToVector(&I))
return;
DEBUG(dbgs() << " alloca is not a candidate for vectorization.\n");
// FIXME: This is the maximum work group size. We should try to get
// value from the reqd_work_group_size function attribute if it is
// available.
unsigned WorkGroupSize = 256;
int AllocaSize = WorkGroupSize *
Mod->getDataLayout()->getTypeAllocSize(AllocaTy);
if (AllocaSize > LocalMemAvailable) {
DEBUG(dbgs() << " Not enough local memory to promote alloca.\n");
return;
}
std::vector<Value*> WorkList;
if (!collectUsesWithPtrTypes(&I, WorkList)) {
DEBUG(dbgs() << " Do not know how to convert all uses\n");
return;
}
DEBUG(dbgs() << "Promoting alloca to local memory\n");
LocalMemAvailable -= AllocaSize;
GlobalVariable *GV = new GlobalVariable(
*Mod, ArrayType::get(I.getAllocatedType(), 256), false,
GlobalValue::ExternalLinkage, 0, I.getName(), 0,
GlobalVariable::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS);
FunctionType *FTy = FunctionType::get(
Type::getInt32Ty(Mod->getContext()), false);
AttributeSet AttrSet;
AttrSet.addAttribute(Mod->getContext(), 0, Attribute::ReadNone);
Value *ReadLocalSizeY = Mod->getOrInsertFunction(
"llvm.r600.read.local.size.y", FTy, AttrSet);
Value *ReadLocalSizeZ = Mod->getOrInsertFunction(
"llvm.r600.read.local.size.z", FTy, AttrSet);
Value *ReadTIDIGX = Mod->getOrInsertFunction(
"llvm.r600.read.tidig.x", FTy, AttrSet);
Value *ReadTIDIGY = Mod->getOrInsertFunction(
"llvm.r600.read.tidig.y", FTy, AttrSet);
Value *ReadTIDIGZ = Mod->getOrInsertFunction(
"llvm.r600.read.tidig.z", FTy, AttrSet);
Value *TCntY = Builder.CreateCall(ReadLocalSizeY);
Value *TCntZ = Builder.CreateCall(ReadLocalSizeZ);
Value *TIdX = Builder.CreateCall(ReadTIDIGX);
Value *TIdY = Builder.CreateCall(ReadTIDIGY);
Value *TIdZ = Builder.CreateCall(ReadTIDIGZ);
Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ);
Tmp0 = Builder.CreateMul(Tmp0, TIdX);
Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ);
Value *TID = Builder.CreateAdd(Tmp0, Tmp1);
TID = Builder.CreateAdd(TID, TIdZ);
std::vector<Value*> Indices;
Indices.push_back(Constant::getNullValue(Type::getInt32Ty(Mod->getContext())));
Indices.push_back(TID);
Value *Offset = Builder.CreateGEP(GV, Indices);
I.mutateType(Offset->getType());
I.replaceAllUsesWith(Offset);
I.eraseFromParent();
for (std::vector<Value*>::iterator i = WorkList.begin(),
e = WorkList.end(); i != e; ++i) {
Value *V = *i;
CallInst *Call = dyn_cast<CallInst>(V);
if (!Call) {
Type *EltTy = V->getType()->getPointerElementType();
PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
// The operand's value should be corrected on its own.
if (isa<AddrSpaceCastInst>(V))
continue;
// FIXME: It doesn't really make sense to try to do this for all
// instructions.
V->mutateType(NewTy);
continue;
}
IntrinsicInst *Intr = dyn_cast<IntrinsicInst>(Call);
if (!Intr) {
std::vector<Type*> ArgTypes;
for (unsigned ArgIdx = 0, ArgEnd = Call->getNumArgOperands();
ArgIdx != ArgEnd; ++ArgIdx) {
ArgTypes.push_back(Call->getArgOperand(ArgIdx)->getType());
}
Function *F = Call->getCalledFunction();
FunctionType *NewType = FunctionType::get(Call->getType(), ArgTypes,
F->isVarArg());
Constant *C = Mod->getOrInsertFunction(StringRef(F->getName().str() + ".local"), NewType,
F->getAttributes());
Function *NewF = cast<Function>(C);
Call->setCalledFunction(NewF);
continue;
}
Builder.SetInsertPoint(Intr);
switch (Intr->getIntrinsicID()) {
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
// These intrinsics are for address space 0 only
Intr->eraseFromParent();
continue;
case Intrinsic::memcpy: {
MemCpyInst *MemCpy = cast<MemCpyInst>(Intr);
Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getRawSource(),
MemCpy->getLength(), MemCpy->getAlignment(),
MemCpy->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memset: {
MemSetInst *MemSet = cast<MemSetInst>(Intr);
Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(),
MemSet->getLength(), MemSet->getAlignment(),
MemSet->isVolatile());
Intr->eraseFromParent();
continue;
}
default:
Intr->dump();
llvm_unreachable("Don't know how to promote alloca intrinsic use.");
}
}
}
void insertCallToAccessFunction(Function *F, Function *cF) {
CallInst *I;
Instruction *bI;
std::vector<Value *> Args;
std::vector<Type *> ArgsTy;
Module *M = F->getParent();
std::string name;
Function *nF, *tF;
FunctionType *FTy;
std::stringstream out;
Value::user_iterator i = F->user_begin(), e = F->user_end();
while (i != e) {
Args.clear();
ArgsTy.clear();
/************* C codes ***********/
if (isa<CallInst>(*i)) {
I = dyn_cast<CallInst>(*i);
// call to the access function F
Args.push_back(I->getArgOperand(0));
ArgsTy.push_back(I->getArgOperand(0)->getType());
// call to the execute function cF
Args.push_back(cF);
ArgsTy.push_back(PointerType::get(cF->getFunctionType(), 0));
unsigned int t;
for (t = 1; t < I->getNumArgOperands(); t++) {
Args.push_back(I->getArgOperand(t));
ArgsTy.push_back(I->getArgOperand(t)->getType());
// errs() << *(I->getArgOperand(t)) << " is or not " <<
// isa<GlobalVariable>(I->getArgOperand(t)) << "\n";
}
tF = dyn_cast<Function>(I->getCalledFunction());
FTy = FunctionType::get(tF->getReturnType(), ArgsTy, 0);
out.str(std::string());
out << "task_DAE_" << I->getNumArgOperands() - 1;
nF = (Function *)M->getOrInsertFunction(out.str(), FTy);
CallInst *ci = CallInst::Create(nF, Args, I->getName(), I);
i++;
I->replaceAllUsesWith(ci);
I->eraseFromParent();
}
/************* C++ codes ***********/
else {
Value::user_iterator bit = (*i)->user_begin(), bite = (*i)->user_end();
Type *iTy = (*i)->getType();
i++;
while (bit != bite) {
Args.clear();
ArgsTy.clear();
I = dyn_cast<CallInst>(*bit);
bit++;
// call to the access function F
Args.push_back(I->getArgOperand(0));
ArgsTy.push_back(I->getArgOperand(0)->getType());
// call to the execute function cF
bI = new BitCastInst(cF, (iTy), "_TPR", I);
Args.push_back(bI);
ArgsTy.push_back(bI->getType());
unsigned int t;
for (t = 1; t < I->getNumArgOperands(); t++) {
Args.push_back(I->getArgOperand(t));
ArgsTy.push_back(I->getArgOperand(t)->getType());
}
tF = dyn_cast<Function>(I->getCalledFunction());
FTy = FunctionType::get(tF->getReturnType(), ArgsTy, 0);
out.str(std::string());
out << "task_DAE_" << I->getNumArgOperands() - 1;
nF = (Function *)M->getOrInsertFunction(out.str(), FTy);
CallInst *ci = CallInst::Create(nF, Args, I->getName(), I);
I->replaceAllUsesWith(ci);
I->eraseFromParent();
}
}
}
}
void mapArgumentsToParams(Function *F, ValueToValueMapTy *VMap) {
CallInst *I;
Instruction *aux = 0;
Value *vaux;
Value::user_iterator i = F->user_begin(), e = F->user_end();
Function::arg_iterator a = F->arg_begin();
while (i != e) {
a = F->arg_begin();
/************* C codes ***********/
if (isa<CallInst>(*i)) {
I = dyn_cast<CallInst>(*i);
unsigned int t;
for (t = 1; t < I->getNumArgOperands(); t++) {
aux = 0;
vaux = I->getArgOperand(t);
while (isa<Instruction>(vaux)) {
aux = dyn_cast<Instruction>(vaux);
vaux = aux->getOperand(0);
}
// errs() << "ARG "<< *(a) << " ----------------- " << *vaux << "\n";
Value *argument = &*a;
VMap->insert(std::pair<Value *, Value *>(argument, &*vaux));
a++;
}
i++;
}
/************* C++ codes ***********/
else {
Value::user_iterator bit = (*i)->user_begin(), bite = (*i)->user_end();
i++;
while (bit != bite) {
I = dyn_cast<CallInst>(*bit);
a = F->arg_begin();
bit++;
unsigned int t;
for (t = 1; t < I->getNumArgOperands(); t++) {
aux = 0;
vaux = I->getArgOperand(t);
errs() << "vaux=" << *vaux << " a=" << *a << "\n";
while ((isa<Instruction>(vaux)) && (!isa<PHINode>(vaux))) {
aux = dyn_cast<Instruction>(vaux);
if (!isa<PHINode>(aux))
vaux = aux->getOperand(0);
// errs() << "vaux="<<*vaux << " a="<<*a <<"\n";
}
errs() << "ARG " << *(a) << " ----------------- " << *vaux << "\n";
Value *argument = &*a;
VMap->insert(std::pair<Value *, Value *>(argument, vaux));
a++;
}
}
}
}
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Clone functions that take LoadInsts as arguments
//
// Inputs:
// M - A reference to the LLVM module to transform
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool LoadArgs::runOnModule(Module& M) {
std::map<std::pair<Function*, const Type * > , Function* > fnCache;
bool changed;
do {
changed = false;
for (Module::iterator Func = M.begin(); Func != M.end(); ++Func) {
for (Function::iterator B = Func->begin(), FE = Func->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
CallInst *CI = dyn_cast<CallInst>(I++);
if(!CI)
continue;
if(CI->hasByValArgument())
continue;
// if the CallInst calls a function, that is externally defined,
// or might be changed, ignore this call site.
Function *F = CI->getCalledFunction();
if (!F || (F->isDeclaration() || F->mayBeOverridden()))
continue;
if(F->hasStructRetAttr())
continue;
if(F->isVarArg())
continue;
// find the argument we must replace
Function::arg_iterator ai = F->arg_begin(), ae = F->arg_end();
unsigned argNum = 0;
for(; argNum < CI->getNumArgOperands();argNum++, ++ai) {
// do not care about dead arguments
if(ai->use_empty())
continue;
if(F->getAttributes().getParamAttributes(argNum).hasAttrSomewhere(Attribute::SExt) ||
F->getAttributes().getParamAttributes(argNum).hasAttrSomewhere(Attribute::ZExt))
continue;
if (isa<LoadInst>(CI->getArgOperand(argNum)))
break;
}
// if no argument was a GEP operator to be changed
if(ai == ae)
continue;
LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(argNum));
Instruction * InsertPt = &(Func->getEntryBlock().front());
AllocaInst *NewVal = new AllocaInst(LI->getType(), "",InsertPt);
StoreInst *Copy = new StoreInst(LI, NewVal);
Copy->insertAfter(LI);
/*if(LI->getParent() != CI->getParent())
continue;
// Also check that there is no store after the load.
// TODO: Check if the load/store do not alias.
BasicBlock::iterator bii = LI->getParent()->begin();
Instruction *BII = bii;
while(BII != LI) {
++bii;
BII = bii;
}
while(BII != CI) {
if(isa<StoreInst>(BII))
break;
++bii;
BII = bii;
}
if(isa<StoreInst>(bii)){
continue;
}*/
// Construct the new Type
// Appends the struct Type at the beginning
std::vector<Type*>TP;
for(unsigned c = 0; c < CI->getNumArgOperands();c++) {
if(c == argNum)
TP.push_back(LI->getPointerOperand()->getType());
TP.push_back(CI->getArgOperand(c)->getType());
}
//return type is same as that of original instruction
FunctionType *NewFTy = FunctionType::get(CI->getType(), TP, false);
numSimplified++;
//if(numSimplified > 1000)
//return true;
Function *NewF;
std::map<std::pair<Function*, const Type* > , Function* >::iterator Test;
Test = fnCache.find(std::make_pair(F, NewFTy));
if(Test != fnCache.end()) {
NewF = Test->second;
} else {
NewF = Function::Create(NewFTy,
GlobalValue::InternalLinkage,
F->getName().str() + ".TEST",
&M);
fnCache[std::make_pair(F, NewFTy)] = NewF;
Function::arg_iterator NI = NewF->arg_begin();
ValueToValueMapTy ValueMap;
unsigned count = 0;
for (Function::arg_iterator II = F->arg_begin(); NI != NewF->arg_end(); ++count, ++NI) {
if(count == argNum) {
NI->setName("LDarg");
continue;
}
ValueMap[II] = NI;
NI->setName(II->getName());
NI->addAttr(F->getAttributes().getParamAttributes(II->getArgNo() + 1));
++II;
}
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
CloneFunctionInto(NewF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NewF->arg_begin(),
ae= NewF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NewF->setAttributes(NewF->getAttributes().addAttributes(
F->getContext(), 0, F->getAttributes().getRetAttributes()));
NewF->setAttributes(NewF->getAttributes().addAttributes(
F->getContext(), ~0, F->getAttributes().getFnAttributes()));
//Get the point to insert the GEP instr.
Instruction *InsertPoint;
for (BasicBlock::iterator insrt = NewF->front().begin(); isa<AllocaInst>(InsertPoint = insrt); ++insrt) {;}
LoadInst *LI_new = new LoadInst(fargs.at(argNum), "", InsertPoint);
fargs.at(argNum+1)->replaceAllUsesWith(LI_new);
}
//this does not seem to be a good idea
AttributeSet NewCallPAL=AttributeSet();
// Get the initial attributes of the call
AttributeSet CallPAL = CI->getAttributes();
AttributeSet RAttrs = CallPAL.getRetAttributes();
AttributeSet FnAttrs = CallPAL.getFnAttributes();
if (!RAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),0, RAttrs);
SmallVector<Value*, 8> Args;
for(unsigned j =0;j<CI->getNumArgOperands();j++) {
if(j == argNum) {
Args.push_back(NewVal);
}
Args.push_back(CI->getArgOperand(j));
// position in the NewCallPAL
AttributeSet Attrs = CallPAL.getParamAttributes(j+1);
if (!Attrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),Args.size(), Attrs);
}
// Create the new attributes vec.
if (!FnAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),~0, FnAttrs);
CallInst *CallI = CallInst::Create(NewF,Args,"", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
CI->replaceAllUsesWith(CallI);
CI->eraseFromParent();
changed = true;
}
}
}
} while(changed);
return true;
}
void AAAnalyzer::handle_inst(Instruction *inst, FunctionWrapper * parent_func) {
//outs()<<*inst<<"\n"; outs().flush();
switch (inst->getOpcode()) {
// common/bitwise binary operations
// Terminator instructions
case Instruction::Ret:
{
ReturnInst* retInst = ((ReturnInst*) inst);
if (retInst->getNumOperands() > 0 && !retInst->getOperandUse(0)->getType()->isVoidTy()) {
parent_func->addRet(retInst->getOperandUse(0));
}
}
break;
case Instruction::Resume:
{
Value* resume = ((ResumeInst*) inst)->getOperand(0);
parent_func->addResume(resume);
}
break;
case Instruction::Switch:
case Instruction::Br:
case Instruction::IndirectBr:
case Instruction::Unreachable:
break;
// vector operations
case Instruction::ExtractElement:
{
}
break;
case Instruction::InsertElement:
{
}
break;
case Instruction::ShuffleVector:
{
}
break;
// aggregate operations
case Instruction::ExtractValue:
{
Value * agg = ((ExtractValueInst*) inst)->getAggregateOperand();
DyckVertex* aggV = wrapValue(agg);
Type* aggTy = agg->getType();
ArrayRef<unsigned> indices = ((ExtractValueInst*) inst)->getIndices();
DyckVertex* currentStruct = aggV;
for (unsigned int i = 0; i < indices.size(); i++) {
if (isa<CompositeType>(aggTy) && aggTy->isSized()) {
if (!aggTy->isStructTy()) {
aggTy = ((CompositeType*) aggTy)->getTypeAtIndex(indices[i]);
#ifndef ARRAY_SIMPLIFIED
current = addPtrOffset(current, (int) indices[i] * dl.getTypeAllocSize(aggTy), dgraph);
#endif
if (i == indices.size() - 1) {
this->makeAlias(currentStruct, wrapValue(inst));
}
} else {
aggTy = ((CompositeType*) aggTy)->getTypeAtIndex(indices[i]);
if (i != indices.size() - 1) {
currentStruct = this->addField(currentStruct, -2 - (int) indices[i], NULL);
} else {
currentStruct = this->addField(currentStruct, -2 - (int) indices[i], wrapValue(inst));
}
}
} else {
break;
}
}
}
break;
case Instruction::InsertValue:
{
DyckVertex* resultV = wrapValue(inst);
Value * agg = ((InsertValueInst*) inst)->getAggregateOperand();
if (!isa<UndefValue>(agg)) {
makeAlias(resultV, wrapValue(agg));
}
Value * val = ((InsertValueInst*) inst)->getInsertedValueOperand();
DyckVertex* insertedVal = wrapValue(val);
Type *aggTy = inst->getType();
ArrayRef<unsigned> indices = ((InsertValueInst*) inst)->getIndices();
DyckVertex* currentStruct = resultV;
for (unsigned int i = 0; i < indices.size(); i++) {
if (isa<CompositeType>(aggTy) && aggTy->isSized()) {
if (!aggTy->isStructTy()) {
aggTy = ((CompositeType*) aggTy)->getTypeAtIndex(indices[i]);
#ifndef ARRAY_SIMPLIFIED
current = addPtrOffset(current, (int) indices[i] * dl.getTypeAllocSize(aggTy), dgraph);
#endif
if (i == indices.size() - 1) {
this->makeAlias(currentStruct, insertedVal);
}
} else {
aggTy = ((CompositeType*) aggTy)->getTypeAtIndex(indices[i]);
if (i != indices.size() - 1) {
currentStruct = this->addField(currentStruct, -2 - (int) indices[i], NULL);
} else {
currentStruct = this->addField(currentStruct, -2 - (int) indices[i], insertedVal);
}
}
} else {
break;
}
}
}
break;
// memory accessing and addressing operations
case Instruction::Alloca:
{
}
break;
case Instruction::Fence:
{
}
break;
case Instruction::AtomicCmpXchg:
{
Value * retXchg = inst;
Value * ptrXchg = inst->getOperand(0);
Value * newXchg = inst->getOperand(2);
addPtrTo(wrapValue(ptrXchg), wrapValue(retXchg));
addPtrTo(wrapValue(ptrXchg), wrapValue(newXchg));
}
break;
case Instruction::AtomicRMW:
{
Value * retRmw = inst;
Value * ptrRmw = ((AtomicRMWInst*) inst)->getPointerOperand();
addPtrTo(wrapValue(ptrRmw), wrapValue(retRmw));
switch (((AtomicRMWInst*) inst)->getOperation()) {
case AtomicRMWInst::Max:
case AtomicRMWInst::Min:
case AtomicRMWInst::UMax:
case AtomicRMWInst::UMin:
case AtomicRMWInst::Xchg:
{
Value * newRmw = ((AtomicRMWInst*) inst)->getValOperand();
addPtrTo(wrapValue(ptrRmw), wrapValue(newRmw));
}
break;
default:
//others are binary ops like add/sub/...
///@TODO
break;
}
}
break;
case Instruction::Load:
{
Value *lval = inst;
Value *ladd = inst->getOperand(0);
addPtrTo(wrapValue(ladd), wrapValue(lval));
}
break;
case Instruction::Store:
{
Value * sval = inst->getOperand(0);
Value * sadd = inst->getOperand(1);
addPtrTo(wrapValue(sadd), wrapValue(sval));
}
break;
case Instruction::GetElementPtr:
{
makeAlias(wrapValue(inst), handle_gep((GEPOperator*) inst));
}
break;
// conversion operations
case Instruction::Trunc:
case Instruction::ZExt:
case Instruction::SExt:
case Instruction::FPTrunc:
case Instruction::FPExt:
case Instruction::FPToUI:
case Instruction::FPToSI:
case Instruction::UIToFP:
case Instruction::SIToFP:
case Instruction::BitCast:
case Instruction::PtrToInt:
case Instruction::IntToPtr:
{
Value * itpv = inst->getOperand(0);
makeAlias(wrapValue(inst), wrapValue(itpv));
}
break;
// other operations
case Instruction::Invoke: // invoke is a terminal operation
{
InvokeInst * invoke = (InvokeInst*) inst;
LandingPadInst* lpd = invoke->getLandingPadInst();
parent_func->addLandingPad(invoke, lpd);
Value * cv = invoke->getCalledValue();
vector<Value*> args;
for (unsigned i = 0; i < invoke->getNumArgOperands(); i++) {
args.push_back(invoke->getArgOperand(i));
}
this->handle_invoke_call_inst(invoke, cv, &args, parent_func);
}
break;
case Instruction::Call:
{
CallInst * callinst = (CallInst*) inst;
if (callinst->isInlineAsm()) {
break;
}
Value * cv = callinst->getCalledValue();
vector<Value*> args;
for (unsigned i = 0; i < callinst->getNumArgOperands(); i++) {
args.push_back(callinst->getArgOperand(i));
}
this->handle_invoke_call_inst(callinst, cv, &args, parent_func);
}
break;
case Instruction::PHI:
{
PHINode *phi = (PHINode *) inst;
int nums = phi->getNumIncomingValues();
for (int i = 0; i < nums; i++) {
Value * p = phi->getIncomingValue(i);
makeAlias(wrapValue(inst), wrapValue(p));
}
}
break;
case Instruction::Select:
{
Value *first = ((SelectInst*) inst)->getTrueValue();
Value *second = ((SelectInst*) inst)->getFalseValue();
makeAlias(wrapValue(inst), wrapValue(first));
makeAlias(wrapValue(inst), wrapValue(second));
}
break;
case Instruction::VAArg:
{
parent_func->addVAArg(inst);
DyckVertex* vaarg = wrapValue(inst);
Value * ptrVaarg = inst->getOperand(0);
addPtrTo(wrapValue(ptrVaarg), vaarg);
}
break;
case Instruction::LandingPad: // handled with invoke inst
case Instruction::ICmp:
case Instruction::FCmp:
default:
break;
}
}