void visitCallInst(CallInst &I) {
string intrinsic = I.getCalledFunction()->getName().str();
if(intrinsic.find("modmul") != -1) {
CallInst *enterMontpro1 = enterMontgomery(I.getOperand(0), I.getOperand(2), &I);
CallInst *enterMontpro2 = enterMontgomery(I.getOperand(1), I.getOperand(2), &I);
CallInst *mulMontpro = mulMontgomery(I.getName().str(), enterMontpro1, enterMontpro2, I.getOperand(2), &I);
CallInst *exitMontpro = leaveMontgomery(mulMontpro, I.getOperand(2), &I);
I.replaceAllUsesWith(exitMontpro);
I.removeFromParent();
} else if(intrinsic.find("modexp") != -1) {
CallInst *enterMontpro1 = enterMontgomery(I.getOperand(0), I.getOperand(2), &I);
CallInst *expMontpro = expMontgomery(I.getName().str(), enterMontpro1, I.getOperand(1), I.getOperand(2), &I);
CallInst *exitMontpro = leaveMontgomery(expMontpro, I.getOperand(2), &I);
I.replaceAllUsesWith(exitMontpro);
I.eraseFromParent();
}
}
bool LowerExpectIntrinsic::HandleSwitchExpect(SwitchInst *SI) {
CallInst *CI = dyn_cast<CallInst>(SI->getCondition());
if (!CI)
return false;
Function *Fn = CI->getCalledFunction();
if (!Fn || Fn->getIntrinsicID() != Intrinsic::expect)
return false;
Value *ArgValue = CI->getArgOperand(0);
ConstantInt *ExpectedValue = dyn_cast<ConstantInt>(CI->getArgOperand(1));
if (!ExpectedValue)
return false;
LLVMContext &Context = CI->getContext();
Type *Int32Ty = Type::getInt32Ty(Context);
SwitchInst::CaseIt Case = SI->findCaseValue(ExpectedValue);
std::vector<Value *> Vec;
unsigned n = SI->getNumCases();
Vec.resize(n + 1 + 1); // +1 for MDString and +1 for default case
Vec[0] = MDString::get(Context, "branch_weights");
Vec[1] = ConstantInt::get(Int32Ty, Case == SI->case_default() ?
LikelyBranchWeight : UnlikelyBranchWeight);
for (unsigned i = 0; i < n; ++i) {
Vec[i + 1 + 1] = ConstantInt::get(Int32Ty, i == Case.getCaseIndex() ?
LikelyBranchWeight : UnlikelyBranchWeight);
}
MDNode *WeightsNode = llvm::MDNode::get(Context, Vec);
SI->setMetadata(LLVMContext::MD_prof, WeightsNode);
SI->setCondition(ArgValue);
return true;
}
/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
/// in the body of the inlined function into invokes and turn unwind
/// instructions into branches to the invoke unwind dest.
///
/// II is the invoke instruction begin inlined. FirstNewBlock is the first
/// block of the inlined code (the last block is the end of the function),
/// and InlineCodeInfo is information about the code that got inlined.
static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
ClonedCodeInfo &InlinedCodeInfo) {
BasicBlock *InvokeDest = II->getUnwindDest();
std::vector<Value*> InvokeDestPHIValues;
// If there are PHI nodes in the unwind destination block, we need to
// keep track of which values came into them from this invoke, then remove
// the entry for this block.
BasicBlock *InvokeBlock = II->getParent();
for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) {
PHINode *PN = cast<PHINode>(I);
// Save the value to use for this edge.
InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock));
}
Function *Caller = FirstNewBlock->getParent();
// The inlined code is currently at the end of the function, scan from the
// start of the inlined code to its end, checking for stuff we need to
// rewrite.
if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) {
for (Function::iterator BB = FirstNewBlock, E = Caller->end();
BB != E; ++BB) {
if (InlinedCodeInfo.ContainsCalls) {
for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){
Instruction *I = BBI++;
// We only need to check for function calls: inlined invoke
// instructions require no special handling.
if (!isa<CallInst>(I)) continue;
CallInst *CI = cast<CallInst>(I);
// If this is an intrinsic function call or an inline asm, don't
// convert it to an invoke.
if ((CI->getCalledFunction() &&
CI->getCalledFunction()->getIntrinsicID()) ||
isa<InlineAsm>(CI->getCalledValue()))
continue;
// Convert this function call into an invoke instruction.
// First, split the basic block.
BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
// Next, create the new invoke instruction, inserting it at the end
// of the old basic block.
SmallVector<Value*, 8> InvokeArgs(CI->op_begin()+1, CI->op_end());
InvokeInst *II =
new InvokeInst(CI->getCalledValue(), Split, InvokeDest,
&InvokeArgs[0], InvokeArgs.size(),
CI->getName(), BB->getTerminator());
II->setCallingConv(CI->getCallingConv());
// Make sure that anything using the call now uses the invoke!
CI->replaceAllUsesWith(II);
// Delete the unconditional branch inserted by splitBasicBlock
BB->getInstList().pop_back();
Split->getInstList().pop_front(); // Delete the original call
// Update any PHI nodes in the exceptional block to indicate that
// there is now a new entry in them.
unsigned i = 0;
for (BasicBlock::iterator I = InvokeDest->begin();
isa<PHINode>(I); ++I, ++i) {
PHINode *PN = cast<PHINode>(I);
PN->addIncoming(InvokeDestPHIValues[i], BB);
}
// This basic block is now complete, start scanning the next one.
break;
}
}
if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
// An UnwindInst requires special handling when it gets inlined into an
// invoke site. Once this happens, we know that the unwind would cause
// a control transfer to the invoke exception destination, so we can
// transform it into a direct branch to the exception destination.
new BranchInst(InvokeDest, UI);
// Delete the unwind instruction!
UI->getParent()->getInstList().pop_back();
// Update any PHI nodes in the exceptional block to indicate that
// there is now a new entry in them.
unsigned i = 0;
for (BasicBlock::iterator I = InvokeDest->begin();
isa<PHINode>(I); ++I, ++i) {
PHINode *PN = cast<PHINode>(I);
PN->addIncoming(InvokeDestPHIValues[i], BB);
}
}
}
}
// Now that everything is happy, we have one final detail. The PHI nodes in
// the exception destination block still have entries due to the original
// invoke instruction. Eliminate these entries (which might even delete the
// PHI node) now.
InvokeDest->removePredecessor(II->getParent());
}
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.");
}
}
}
bool GambasPass::runOnFunction(Function &F){
IRBuilder<> Builder(F.getContext());
bool changed = false;
for(Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
for(BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ){
ICmpInst* ICI = dyn_cast<ICmpInst>(I);
CallInst* CI = dyn_cast<CallInst>(I++);
if (ICI && ICI->hasMetadata() && ICI->getMetadata("unref_slt") && dyn_cast<LoadInst>(ICI->getOperand(0))){
ICI->replaceAllUsesWith(ConstantInt::get(ICI->getType(), false));
ICI->eraseFromParent();
changed = true;
continue;
}
if (!CI)
continue;
Function* callee = CI->getCalledFunction();
if (callee == NULL || !callee->isDeclaration())
continue;
StringRef name = callee->getName();
if (name == "JR_release_variant" || name == "JR_borrow_variant"){
ConstantInt* vtype_int = dyn_cast<ConstantInt>(CI->getArgOperand(0));
if (!vtype_int)
continue;
uint64_t vtype = vtype_int->getZExtValue();
if (TYPE_is_string(vtype) || TYPE_is_object(vtype))
continue;
CI->eraseFromParent();
changed = true;
} else if (name == FUNCTION_NAME(__finite)){
ConstantFP* op = dyn_cast<ConstantFP>(CI->getArgOperand(0));
if (!op)
continue;
int val = __finite(op->getValueAPF().convertToDouble());
Constant* res = ConstantInt::get(CI->getType(), val);
CI->replaceAllUsesWith(res);
CI->eraseFromParent();
changed = true;
} else if (name == FUNCTION_NAME(__isnan)){
ConstantFP* op = dyn_cast<ConstantFP>(CI->getArgOperand(0));
if (!op)
continue;
int val = __isnan(op->getValueAPF().convertToDouble());
Constant* res = ConstantInt::get(CI->getType(), val);
CI->replaceAllUsesWith(res);
CI->eraseFromParent();
changed = true;
} else if (name == FUNCTION_NAME(__isinf)){
ConstantFP* op = dyn_cast<ConstantFP>(CI->getArgOperand(0));
if (!op)
continue;
int val = __isinf(op->getValueAPF().convertToDouble());
Constant* res = ConstantInt::get(CI->getType(), val);
CI->replaceAllUsesWith(res);
CI->eraseFromParent();
changed = true;
}
}
}
return changed;
}
void TaskDebugBranchCheck::addFunctionSummaries(BasicBlock* from_bb, BasicBlock* to_bb, Instruction* first_inst) {
//errs() << "++++++++++++++++++++DETECTED BRANCHES++++++++++++++++++++++\n";
//errs() << "BB 1 First inst: " << *(from_bb->getFirstNonPHI()) << "\n";
//TerminatorInst *TInst = from_bb->getTerminator();
//errs() << "BB 1 Last inst: " << *TInst << "\n\n";
//errs() << "BB 2 First inst: " << *(to_bb->getFirstNonPHI()) << "\n";
//TInst = to_bb->getTerminator();
//errs() << "BB 2 Last inst: " << *TInst << "\n\n";
//errs() << "++++++++++++++++++++DETECTED BRANCHES++++++++++++++++++++++\n";
std::vector<Value*> locks_acq;
std::vector<Value*> locks_rel;
bool startInst = false;
for (BasicBlock::iterator i = from_bb->begin(); i != from_bb->end(); ++i) {
if (startInst == false) {
if (first_inst == dyn_cast<Instruction>(i)) {
startInst = true;
} else {
continue;
}
}
switch (i->getOpcode()) {
case Instruction::Call:
{
CallInst* callInst = dyn_cast<CallInst>(i);
if(callInst->getCalledFunction() != NULL) {
if(callInst->getCalledFunction() == lockAcquire) {
locks_acq.push_back(callInst->getArgOperand(1));
} else if (callInst->getCalledFunction() == lockRelease) {
locks_rel.push_back(callInst->getArgOperand(1));
}
}
break;
}
case Instruction::Load:
{
//errs() << "LOAD INST " << *i << "\n";
Value* op_l = i->getOperand(0);
if (hasAnnotation(i, op_l, "check_av", 1)) {
Constant* read =
ConstantInt::get(Type::getInt32Ty(to_bb->getContext()), 0);
instrument_access(to_bb->getFirstNonPHI(), op_l, read, locks_acq, locks_rel);
}
break;
}
case Instruction::Store:
{
//errs() << "STR INST " << *i << "\n";
Value* op_s = i->getOperand(1);
if (hasAnnotation(i, op_s, "check_av", 1)) {
Constant* write =
ConstantInt::get(Type::getInt32Ty(to_bb->getContext()), 1);
instrument_access(to_bb->getFirstNonPHI(), op_s, write, locks_acq, locks_rel);
}
break;
}
}
}
}
void MutationGen::genMutationFile(Function & F){
int index = 0;
for(Function::iterator FI = F.begin(); FI != F.end(); ++FI){
BasicBlock *BB = FI;
#if NEED_LOOP_INFO
bool isLoop = LI->getLoopFor(BB);
#endif
for(BasicBlock::iterator BI = BB->begin(); BI != BB->end(); ++BI, index++){
unsigned opc = BI->getOpcode();
if( !((opc >= 14 && opc <= 31) || opc == 34 || opc == 52 || opc == 55) ){// omit alloca and getelementptr
continue;
}
int idxtmp = index;
#if NEED_LOOP_INFO
if(isLoop){
assert(idxtmp != 0);
idxtmp = 0 - idxtmp;
}
#endif
switch(opc){
case Instruction::Add:
case Instruction::Sub:
case Instruction::Mul:
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::URem:
case Instruction::SRem:{
// TODO: add for i1, i8. Support i32 and i64 first
if(! (BI->getType()->isIntegerTy(32) || BI->getType()->isIntegerTy(64))){
continue;
}
genLVR(BI, F.getName(), idxtmp);
genUOI(BI, F.getName(), idxtmp);
genROV(BI, F.getName(), idxtmp);
genABV(BI, F.getName(), idxtmp);
genAOR(BI, F.getName(), idxtmp);
break;
}
case Instruction::ICmp:{
if(! (BI->getOperand(0)->getType()->isIntegerTy(32) ||
BI->getOperand(0)->getType()->isIntegerTy(64)) ){
continue;
}
genLVR(BI, F.getName(), idxtmp);
genUOI(BI, F.getName(), idxtmp);
genROV(BI, F.getName(), idxtmp);
genABV(BI, F.getName(), idxtmp);
genROR(BI, F.getName(), idxtmp);
break;
}
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor:{
// TODO: add for i1, i8. Support i32 and i64 first
if(! (BI->getType()->isIntegerTy(32) || BI->getType()->isIntegerTy(64))){
continue;
}
genLVR(BI, F.getName(), idxtmp);
genUOI(BI, F.getName(), idxtmp);
genROV(BI, F.getName(), idxtmp);
genABV(BI, F.getName(), idxtmp);
genLOR(BI, F.getName(), idxtmp);
break;
}
case Instruction::Call:
{
CallInst* call = cast<CallInst>(BI);
// TODO: omit function-pointer
if(call->getCalledFunction() == NULL){
continue;
}
/*Value* callee = dyn_cast<Value>(&*(call->op_end() - 1));
if(callee->getType()->isPointerTy()){
continue;
}*/
StringRef name = call->getCalledFunction()->getName();
if(name.startswith("llvm")){//omit llvm inside functions
continue;
}
// TODO: add for ommiting i8. Support i32 and i64 first
if(! ( isSupportedType(BI->getType())|| BI->getType()->isVoidTy() ) ){
continue;
}
genLVR(BI, F.getName(), idxtmp);
genUOI(BI, F.getName(), idxtmp);
genROV(BI, F.getName(), idxtmp);
genABV(BI, F.getName(), idxtmp);
genSTDCall(BI, F.getName(), idxtmp);
break;
}
case Instruction::Store:{
auto addr = BI->op_begin() + 1;// the pointer of the storeinst
if( ! (dyn_cast<LoadInst>(&*addr) ||
dyn_cast<AllocaInst>(&*addr) ||
dyn_cast<Constant>(&*addr) ||
dyn_cast<GetElementPtrInst>(&*addr)
)
){
continue;
}
// TODO:: add for i8
Value* tobestore = dyn_cast<Value>(BI->op_begin());
if(! isSupportedType(tobestore->getType())){
continue;
}
genLVR(BI, F.getName(), idxtmp);
genUOI(BI, F.getName(), idxtmp);
genABV(BI, F.getName(), idxtmp);
genSTDStore(BI, F.getName(), idxtmp);
break;
}
case Instruction::GetElementPtr:{
// TODO:
break;
}
default:{
}
}
}
}
ofresult.flush();
}
/// \brief Recursively handle the condition leading to a loop
Value *SIAnnotateControlFlow::handleLoopCondition(
Value *Cond, PHINode *Broken, llvm::Loop *L, BranchInst *Term,
SmallVectorImpl<WeakTrackingVH> &LoopPhiConditions) {
// Only search through PHI nodes which are inside the loop. If we try this
// with PHI nodes that are outside of the loop, we end up inserting new PHI
// nodes outside of the loop which depend on values defined inside the loop.
// This will break the module with
// 'Instruction does not dominate all users!' errors.
PHINode *Phi = nullptr;
if ((Phi = dyn_cast<PHINode>(Cond)) && L->contains(Phi)) {
BasicBlock *Parent = Phi->getParent();
PHINode *NewPhi = PHINode::Create(Int64, 0, "loop.phi", &Parent->front());
Value *Ret = NewPhi;
// Handle all non-constant incoming values first
for (unsigned i = 0, e = Phi->getNumIncomingValues(); i != e; ++i) {
Value *Incoming = Phi->getIncomingValue(i);
BasicBlock *From = Phi->getIncomingBlock(i);
if (isa<ConstantInt>(Incoming)) {
NewPhi->addIncoming(Broken, From);
continue;
}
Phi->setIncomingValue(i, BoolFalse);
Value *PhiArg = handleLoopCondition(Incoming, Broken, L,
Term, LoopPhiConditions);
NewPhi->addIncoming(PhiArg, From);
}
BasicBlock *IDom = DT->getNode(Parent)->getIDom()->getBlock();
for (unsigned i = 0, e = Phi->getNumIncomingValues(); i != e; ++i) {
Value *Incoming = Phi->getIncomingValue(i);
if (Incoming != BoolTrue)
continue;
BasicBlock *From = Phi->getIncomingBlock(i);
if (From == IDom) {
// We're in the following situation:
// IDom/From
// | \
// | If-block
// | /
// Parent
// where we want to break out of the loop if the If-block is not taken.
// Due to the depth-first traversal, there should be an end.cf
// intrinsic in Parent, and we insert an else.break before it.
//
// Note that the end.cf need not be the first non-phi instruction
// of parent, particularly when we're dealing with a multi-level
// break, but it should occur within a group of intrinsic calls
// at the beginning of the block.
CallInst *OldEnd = dyn_cast<CallInst>(Parent->getFirstInsertionPt());
while (OldEnd && OldEnd->getCalledFunction() != EndCf)
OldEnd = dyn_cast<CallInst>(OldEnd->getNextNode());
if (OldEnd && OldEnd->getCalledFunction() == EndCf) {
Value *Args[] = { OldEnd->getArgOperand(0), NewPhi };
Ret = CallInst::Create(ElseBreak, Args, "", OldEnd);
continue;
}
}
TerminatorInst *Insert = From->getTerminator();
Value *PhiArg = CallInst::Create(Break, Broken, "", Insert);
NewPhi->setIncomingValue(i, PhiArg);
}
LoopPhiConditions.push_back(WeakTrackingVH(Phi));
return Ret;
}
if (Instruction *Inst = dyn_cast<Instruction>(Cond)) {
BasicBlock *Parent = Inst->getParent();
Instruction *Insert;
if (L->contains(Inst)) {
Insert = Parent->getTerminator();
} else {
Insert = L->getHeader()->getFirstNonPHIOrDbgOrLifetime();
}
Value *Args[] = { Cond, Broken };
return CallInst::Create(IfBreak, Args, "", Insert);
}
// Insert IfBreak in the loop header TERM for constant COND other than true.
if (isa<Constant>(Cond)) {
Instruction *Insert = Cond == BoolTrue ?
Term : L->getHeader()->getTerminator();
Value *Args[] = { Cond, Broken };
return CallInst::Create(IfBreak, Args, "", Insert);
}
llvm_unreachable("Unhandled loop condition!");
}
// addEdgesFor
// Creates a node for I and inserts edges from the created node to the
// appropriate node of other values.
void IneqGraph::addEdgesFor(Instruction *I) {
if (I->getType()->isPointerTy())
return;
Range RI = RA->getRange(I);
if (!RI.getLower().isMinSignedValue())
addMayEdge(AlfaConst, I, -RI.getLower().getSExtValue());
if (!RI.getUpper().isMaxSignedValue())
addMayEdge(I, AlfaConst, RI.getUpper().getSExtValue());
// TODO: Handle multiplication, remainder and division instructions.
switch (I->getOpcode()) {
case Instruction::SExt:
case Instruction::ZExt:
case Instruction::Trunc:
case Instruction::BitCast:
addMayEdge(I, I->getOperand(0), 0);
addMayEdge(I->getOperand(0), I, 0);
break;
case Instruction::Add:
// a = b + c
// ==> a <= b + sup(c)
// ==> a <= c + sup(b)
// ==> b <= a - inf(c)
// ==> c <= a - inf(b)
{
Value *A = I->getOperand(0);
Value *B = I->getOperand(1);
Range AR = RA->getRange(A);
Range BR = RA->getRange(B);
if (!isa<ConstantInt>(B) && !AR.getUpper().isMaxSignedValue())
addMayEdge(I, B, AR.getUpper());
if (!isa<ConstantInt>(A) && !BR.getUpper().isMaxSignedValue())
addMayEdge(I, A, BR.getUpper());
if (!isa<ConstantInt>(A) && !BR.getLower().isMinSignedValue())
addMayEdge(A, I, -BR.getUpper());
if (!isa<ConstantInt>(B) && !AR.getLower().isMinSignedValue())
addMayEdge(B, I, -AR.getUpper());
break;
}
case Instruction::Sub:
// a = b - c
// ==> a <= b - inf(c)
{
Value *A = I->getOperand(0);
Value *B = I->getOperand(1);
Range AR = RA->getRange(A);
Range BR = RA->getRange(B);
if (!isa<ConstantInt>(A) && !BR.getLower().isMinSignedValue())
addMayEdge(I, A, -BR.getLower());
break;
}
case Instruction::Br:
// if (a > b) {
// a1 = sigma(a)
// b1 = sigma(b)
{
BranchInst *BI = cast<BranchInst>(I);
ICmpInst *Cmp = dyn_cast<ICmpInst>(I->getOperand(0));
if (!Cmp)
break;
Value *L = Cmp->getOperand(0);
DEBUG(dbgs() << "IneqGraph: L: " << *L << "\n");
Value *R = Cmp->getOperand(1);
DEBUG(dbgs() << "IneqGraph: R: " << *R << "\n");
Value *LSigma = VS->findSigma(L, BI->getSuccessor(0), BI->getParent());
DEBUG(dbgs() << "IneqGraph: LSigma: " << *LSigma << "\n");
Value *RSigma = VS->findSigma(R, BI->getSuccessor(0), BI->getParent());
DEBUG(dbgs() << "IneqGraph: RSigma: " << *RSigma << "\n");
Value *LSExtSigma = VS->findSExtSigma(L, BI->getSuccessor(0), BI->getParent());
DEBUG(dbgs() << "IneqGraph: LSExtSigma: " << *LSExtSigma << "\n");
Value *RSExtSigma = VS->findSExtSigma(R, BI->getSuccessor(0), BI->getParent());
DEBUG(dbgs() << "IneqGraph: RSExtSigma: " << *RSExtSigma << "\n");
switch (Cmp->getPredicate()) {
case ICmpInst::ICMP_SLT:
DEBUG(dbgs() << "IneqGraph: SLT:\n");
if (!isa<ConstantInt>(R) && LSigma) {
if (RSigma)
addMayEdge(LSigma, RSigma, -1);
if (RSExtSigma)
addMayEdge(LSigma, RSExtSigma, -1);
}
if (!isa<ConstantInt>(R) && LSExtSigma && LSExtSigma != LSigma) {
if (RSigma)
addMayEdge(LSExtSigma, RSigma, -1);
if (RSExtSigma)
addMayEdge(LSExtSigma, RSExtSigma, -1);
}
break;
case ICmpInst::ICMP_SLE:
DEBUG(dbgs() << "IneqGraph: SLE:\n");
if (!isa<ConstantInt>(R) && LSigma && RSigma)
addMayEdge(LSigma, RSigma, 0);
if (!isa<ConstantInt>(R) && (LSExtSigma != LSigma || RSExtSigma != RSigma))
addMayEdge(LSExtSigma, RSExtSigma, 0);
break;
default:
break;
}
break;
}
case Instruction::PHI:
{
PHINode *Phi = cast<PHINode>(I);
for (unsigned Idx = 0; Idx < Phi->getNumIncomingValues(); ++Idx) {
addMustEdge(Phi, Phi->getIncomingValue(Idx), 0);
}
break;
}
case Instruction::Call:
{
CallInst *CI = cast<CallInst>(I);
if (Function *F = CI->getCalledFunction()) {
unsigned Idx = 0;
for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
AI != AE; ++AI, ++Idx) {
addMustEdge(&(*AI), CI->getArgOperand(Idx), 0);
}
}
break;
}
}
}
//
// 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;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// If a function returns a struct, make it return
// a pointer to the struct.
//
// Inputs:
// M - A reference to the LLVM module to transform
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool StructRet::runOnModule(Module& M) {
const llvm::DataLayout targetData(&M);
std::vector<Function*> worklist;
for (Module::iterator I = M.begin(); I != M.end(); ++I)
if (!I->mayBeOverridden()) {
if(I->hasAddressTaken())
continue;
if(I->getReturnType()->isStructTy()) {
worklist.push_back(I);
}
}
while(!worklist.empty()) {
Function *F = worklist.back();
worklist.pop_back();
Type *NewArgType = F->getReturnType()->getPointerTo();
// Construct the new Type
std::vector<Type*>TP;
TP.push_back(NewArgType);
for (Function::arg_iterator ii = F->arg_begin(), ee = F->arg_end();
ii != ee; ++ii) {
TP.push_back(ii->getType());
}
FunctionType *NFTy = FunctionType::get(F->getReturnType(), TP, F->isVarArg());
// Create the new function body and insert it into the module.
Function *NF = Function::Create(NFTy,
F->getLinkage(),
F->getName(), &M);
ValueToValueMapTy ValueMap;
Function::arg_iterator NI = NF->arg_begin();
NI->setName("ret");
++NI;
for (Function::arg_iterator II = F->arg_begin(); II != F->arg_end(); ++II, ++NI) {
ValueMap[II] = NI;
NI->setName(II->getName());
AttributeSet attrs = F->getAttributes().getParamAttributes(II->getArgNo() + 1);
if (!attrs.isEmpty())
NI->addAttr(attrs);
}
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
if (!F->isDeclaration())
CloneFunctionInto(NF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NF->arg_begin(),
ae= NF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NF->setAttributes(NF->getAttributes().addAttributes(
M.getContext(), 0, F->getAttributes().getRetAttributes()));
NF->setAttributes(NF->getAttributes().addAttributes(
M.getContext(), ~0, F->getAttributes().getFnAttributes()));
for (Function::iterator B = NF->begin(), FE = NF->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
ReturnInst * RI = dyn_cast<ReturnInst>(I++);
if(!RI)
continue;
LoadInst *LI = dyn_cast<LoadInst>(RI->getOperand(0));
assert(LI && "Return should be preceded by a load instruction");
IRBuilder<> Builder(RI);
Builder.CreateMemCpy(fargs.at(0),
LI->getPointerOperand(),
targetData.getTypeStoreSize(LI->getType()),
targetData.getPrefTypeAlignment(LI->getType()));
}
}
for(Value::use_iterator ui = F->use_begin(), ue = F->use_end();
ui != ue; ) {
CallInst *CI = dyn_cast<CallInst>(*ui++);
if(!CI)
continue;
if(CI->getCalledFunction() != F)
continue;
if(CI->hasByValArgument())
continue;
AllocaInst *AllocaNew = new AllocaInst(F->getReturnType(), 0, "", CI);
SmallVector<Value*, 8> Args;
//this should probably be done in a different manner
AttributeSet NewCallPAL=AttributeSet();
// Get the initial attributes of the call
AttributeSet CallPAL = CI->getAttributes();
AttributeSet RAttrs = CallPAL.getRetAttributes();
AttributeSet FnAttrs = CallPAL.getFnAttributes();
if (!RAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),0, RAttrs);
Args.push_back(AllocaNew);
for(unsigned j = 0; j < CI->getNumOperands()-1; j++) {
Args.push_back(CI->getOperand(j));
// position in the NewCallPAL
AttributeSet Attrs = CallPAL.getParamAttributes(j);
if (!Attrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),Args.size(), Attrs);
}
// Create the new attributes vec.
if (!FnAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),~0, FnAttrs);
CallInst *CallI = CallInst::Create(NF, Args, "", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
LoadInst *LI = new LoadInst(AllocaNew, "", CI);
CI->replaceAllUsesWith(LI);
CI->eraseFromParent();
}
if(F->use_empty())
F->eraseFromParent();
}
return true;
}
bool PathList::runOnModule(Module &M) {
module = &M;
llvm::dbgs() << "[runOnModule]: Moduel M has " << M.getFunctionList().size() << " Functions in all.\n";
// for test
Function *f1 = M.getFunction("fprintf");
if (!f1)
dbgs() << "[Test]: can not find function fprintf.\n";
else
dbgs() << "[Test]: find function fprintf.\n";
CallGraph &CG = getAnalysis<CallGraph>();
// CG.dump();
CallGraphNode *cgNode = CG.getRoot();
cgNode->dump();
// errs()<<node->getFunction()->getName()<<'\n';
Function *startFunc;
Function *endFunc;
startFunc = M.getFunction("__user_main");
//std::string fileName("/home/xqx/data/xqx/projects/benckmarks-klee/texinfo-4.8/build-shit/makeinfo/../../makeinfo/insertion.c");
//int lineNo = 407;
BB = getBB(fileName, lineNo);
*targetBbpp = getBB(fileName, lineNo);
if (BB) {
endFunc = BB->getParent();
if (!endFunc) {
errs()<<"Error: get endFunc failed.\n";
return false;
}
if (!startFunc) {
errs()<<"Error: get startFunc failed.\n";
return false;
}
errs()<<startFunc->getName()<<'\n';
}
else {
errs()<<"Error: get BB failed.\n";
return false;
}
//read start and end from xml files
// defectList enStart, enEnd;
// getEntryList("/tmp/entrys.xml", &enStart, "start");
// getEntryList("/tmp/entrys.xml", &enEnd, "end");
// getEntryList("/tmp/entrys.xml", &dl, "end");
// dumpEntryList(&enStart);
// dumpEntryList(&enEnd);
// dumpEntryList(&dl);
//read bug information from xml file
/* for (defectList::iterator dit = dl.begin(); dit != dl.end(); dit++) {
StringRef file(dit->first.c_str());
std::vector<int> lines = dit->second;
BasicBlock *BB = getBB(file, *(lines.begin()));
if (BB) {
endFunc = BB->getParent();
}
}
*/
//to store temporary path
std::vector<BasicBlock*> p;
// a counter
int map_count = 0;
for (Module::iterator i = M.begin(), e = M.end(); i != e; ++i) {
Function *F = i;
if (!F) {
llvm::errs() << "***NULL Function***\n";
continue;
}
cgNode = CG.getOrInsertFunction(F);
F = cgNode->getFunction();
//
for (CallGraphNode::iterator I = cgNode->begin(), E = cgNode->end();
I != E; ++I){
CallGraphNode::CallRecord *cr = &*I;
// llvm::errs() << "\tCS<" << cr->first << "> calls";
// check if the CallInst is existed
if(cr->first){
Instruction *TmpIns = dyn_cast<Instruction>(cr->first);
if(TmpIns) {
// errs() << "\t" << *TmpIns << "\n";
//unsigned int l, c;
//std::string cfi_path = getInstPath(TmpIns, l, c);
//if (!cfi_path.empty()) {
// if (cfi_path.find("uclibc") != std::string::npos) {
// dbgs() << "[Filter Uclib]: find an instruction from uclibc.\n";
// continue;
// } else if (cfi_path.find("POSIX") != std::string::npos) {
// dbgs() << "[Filter Uclib]: find an instruction from POSIX.\n";
// continue;
// }
//}
} else
continue;
}
// get the funciton pointer which is called by current CallRecord cr
Function *FI = cr->second->getFunction();
if (!FI)
continue;
// create a new CalledFunctions element and push it into calledFunctionMap.
calledFunctionMap[FI].push_back(std::make_pair(F, dyn_cast<Instruction>(cr->first)));
// for debuging
map_count++;
}
}
dbgs() << "[Count Number of calledFunctionMap]: "<< calledFunctionMap.size() <<'\n';
// analyze the global function pointer table
if(function_pointer_analysis()) {
errs() << "[Analyze global function pointer table success]\n";
} else {
errs() << "[Analyze global function pointer table failed]\n";
}
dbgs() << "[Count Number of calledFunctionMap]: "<< calledFunctionMap.size() <<'\n';
// filter the instructions from uclibc
//filter_uclibc();
llvm::errs() << "=================================hh\n";
llvm::errs() << "get Function Path: " << endFunc->getName()
<< " to " << startFunc->getName() << " \n";
// printCalledFuncAndCFGPath(endFunc, startFunc, BB, p);
// modification by wh
evo_paths = new entire_path;
//filter_paths = new func_bbs_type;
//BB_paths_map = new std::map<std::pair<Function*, BasicBlock*>, std::vector<BasicBlock*> >;
std::vector<std::pair< Function*, Instruction*> > tmp_func_path;
// std::vector<BasicBlock*> tmp_bb_path;
// explore_function_paths(endFunc, startFunc, bug_Inst, &tmp_func_path);
collect_funcitons(endFunc, startFunc, bug_Inst, &tmp_func_path);
// dbgs() << "++++++Found " << evo_paths->size() << " function paths.\n";
// for (entire_path::iterator ep_it = evo_paths->begin(); ep_it != evo_paths->end(); ep_it++) {
// for (std::vector<std::pair< Function*, Instruction*> >::iterator pair_it = ep_it->begin(); pair_it != ep_it->end(); pair_it++) {
// if (filter_paths->size() != 0) {
// std::vector<Instruction*>::iterator inst_it = std::find((*filter_paths)[pair_it->first].begin(), (*filter_paths)[pair_it->first].end(), pair_it->second);
// if (inst_it != (*filter_paths)[pair_it->first].end()) {
// continue;
// }
// }
// (*filter_paths)[pair_it->first].push_back(pair_it->second);
// }
// }
dbgs() << "[filter_paths]: contain " << filter_paths->size() << " functions in all.\n";
for (func_bbs_type::iterator fbs_it = filter_paths->begin(); fbs_it != filter_paths->end(); fbs_it++) {
for (std::vector<Instruction*>::iterator bb_it2 = fbs_it->second.begin(); bb_it2 != fbs_it->second.end(); bb_it2++) {
dbgs() << "^^^^^^ " << fbs_it->first->getName() << ": " << (*bb_it2)->getParent()->getName() << '\n';
// to expand functions
call_insts.push_back((*bb_it2));
explore_basicblock_paths(fbs_it->first, (*bb_it2)->getParent(), &(*BB_paths_map)[std::make_pair(fbs_it->first, *bb_it2)]);
dbgs() << "^^^^^^ found " << (*BB_paths_map)[std::make_pair(fbs_it->first, *bb_it2)].size() << " basicblocks.\n";
}
}
llvm::dbgs() << "!!!!!!!! Found " << call_insts.size() << " call instructions.\n";
llvm::dbgs() << "!!!!!!!! Found " << path_basicblocks.size() << " path basicblocks.\n";
// expand functions
for (std::vector<Instruction*>::iterator ci_it = call_insts.begin(); ci_it != call_insts.end(); ci_it++) {
BasicBlock *call_bb = (*ci_it)->getParent();
if (!call_bb) {
continue;
}
for (BasicBlock::iterator inst = call_bb->begin(); inst != call_bb->end(); inst++) {
if (&*inst == *ci_it) {
break;
}
if (isa<CallInst>(&*inst)) {
std::vector<Instruction*>::iterator ci = std::find(path_call_insts.begin(), path_call_insts.end(), &*inst);
if (ci != path_call_insts.end())
continue;
path_call_insts.push_back(&*inst);
}
}
}
llvm::dbgs() << "@@@@@@@@ After search call_insts, found " << path_call_insts.size() << " call instructions.\n";
for (std::vector<BasicBlock*>::iterator p_bb_it = path_basicblocks.begin(); p_bb_it != path_basicblocks.end(); p_bb_it++) {
for (BasicBlock::iterator inst = (*p_bb_it)->begin(); inst != (*p_bb_it)->end(); inst++) {
if (isa<CallInst>(&*inst)) {
std::vector<Instruction*>::iterator ci = std::find(path_call_insts.begin(), path_call_insts.end(), &*inst);
if (ci != path_call_insts.end())
continue;
path_call_insts.push_back(&*inst);
}
}
}
llvm::dbgs() << "@@@@@@@@ After search path_basicblocks, found " << path_call_insts.size() << " call instructions.\n";
for (std::vector<Instruction*>::iterator iit = path_call_insts.begin(); iit != path_call_insts.end(); iit++) {
CallInst *ci = dyn_cast<CallInst>(*iit);
if (!ci)
continue;
Function *ff = ci->getCalledFunction();
if (!ff) {
//ci->dump();
//dbgs() << "\t[called value] " << ci->getOperand(0)->getName() << '\n';
continue;
}
std::vector<Function*>::iterator fit = std::find(otherCalledFuncs->begin(), otherCalledFuncs->end(), ff);
if (fit == otherCalledFuncs->end())
otherCalledFuncs->push_back(ff);
}
llvm::dbgs() << "((((((((( Found " << otherCalledFuncs->size() << " functions.\n";
for (int index = 0; index < otherCalledFuncs->size(); index++) {
Function *f = otherCalledFuncs->at(index);
/* if (!f) {
//f->dump();
llvm::dbgs() << "?????? index = " << index << " size = " << otherCalledFuncs->size()<< '\n';
continue;
}
*/ for (inst_iterator f_it = inst_begin(f); f_it != inst_end(f); f_it++) {
CallInst *ci = dyn_cast<CallInst>(&*f_it);
if (!ci)
continue;
if (!ci->getCalledFunction()) {
//ci->dump();
continue;
}
std::vector<Function*>::iterator fit = std::find(otherCalledFuncs->begin(), otherCalledFuncs->end(), ci->getCalledFunction());
if (fit == otherCalledFuncs->end())
otherCalledFuncs->push_back(ci->getCalledFunction());
}
}
llvm::dbgs() << "((((((((( Found " << otherCalledFuncs->size() << " functions.\n";
//This should be just for statistic.
int tmp_funcNum_in_filter_notIn_other = 0;
for (func_bbs_type::iterator fbs_it = filter_paths->begin(); fbs_it != filter_paths->end(); fbs_it++) {
if (!fbs_it->first) {
llvm::dbgs() << "[Warning]: Found a null Function pointer in filter_paths.\n";
continue;
}
std::vector<Function*>::iterator fit = std::find(otherCalledFuncs->begin(), otherCalledFuncs->end(), fbs_it->first);
if (fit == otherCalledFuncs->end())
//otherCalledFuncs->push_back(fbs_it->first);
tmp_funcNum_in_filter_notIn_other ++;
}
llvm::dbgs() << "<><><><> After searching filter_paths, found " << otherCalledFuncs->size() + tmp_funcNum_in_filter_notIn_other << " functions.\n";
/* for (entire_path::iterator ep_it = evo_paths->begin(); ep_it != evo_paths->end(); ep_it++) {
dbgs() << "Path length is: " << ep_it->size() << '\n';
for (std::vector<std::pair< Function*, BasicBlock*> >::iterator pair_it = ep_it->begin(); pair_it != ep_it->end(); pair_it++) {
dbgs() << "^^^^^^ " << pair_it->first->getName() << ": " << pair_it->second->getName() << '\n';
explore_basicblock_paths(pair_it->first, pair_it->second, &(*BB_paths_map)[*pair_it]);
dbgs() << "^^^^^^ found " << (*BB_paths_map)[*pair_it].size() << " basicblocks.\n";
}
}
*/
llvm::errs() << "on-end\n";
llvm::errs() << "=================================\n";
// output all of the paths
/* errs()<<"Find "<<paths_found->size()<<" paths in all.\n";
for(paths::iterator ips = paths_found->begin();ips != paths_found->end();ips++) {
// std::vector<BasicBlock*> *tmpP = dyn_cast<std::vector<BasicBlock*>*>(&*ips);
dbgs() << "=========A Path Start============\n";
for(std::vector<BasicBlock*>::iterator ps = ips->begin(), pe = ips->end(); ps != pe; ps++) {
BasicBlock *tmpStr = *ps;
errs()<<"\t"<<tmpStr->getParent()->getName()<<": "<<tmpStr->getName()<<" -> \n";
}
errs()<<"=================================\n";
}
*/
return false;
}
void OptimizeFastMemoryChecks::visitCallInst(CallInst &CI) {
CheckInfoType *Info = MSCI->getCheckInfo(CI.getCalledFunction());
if (Info && Info->isFastMemoryCheck())
FastCheckCalls.push_back(&CI);
}
bool TracingNoGiri::visitSpecialCall(CallInst &CI) {
Function *CalledFunc = CI.getCalledFunction();
// We do not support indirect calls to special functions.
if (CalledFunc == nullptr)
return false;
// Do not consider a function special if it has a function body; in this
// case, the programmer has supplied his or her version of the function, and
// we will instrument it.
if (!CalledFunc->isDeclaration())
return false;
// Check the name of the function against a list of known special functions.
std::string name = CalledFunc->getName().str();
if (name.substr(0,12) == "llvm.memset.") {
instrumentLock(&CI);
// Get the destination pointer and cast it to a void pointer.
Value *dstPointer = CI.getOperand(0);
dstPointer = castTo(dstPointer, VoidPtrType, dstPointer->getName(), &CI);
// Get the number of bytes that will be written into the buffer.
Value *NumElts = CI.getOperand(2);
// Get the ID of the external funtion call instruction.
Value *CallID = ConstantInt::get(Int32Type, lsNumPass->getID(&CI));
// Create the call to the run-time to record the external call instruction.
std::vector<Value *> args = make_vector(CallID, dstPointer, NumElts, 0);
CallInst::Create(RecordStore, args, "", &CI);
instrumentUnlock(&CI);
++NumExtFuns; // Update statistics
return true;
} else if (name.substr(0,12) == "llvm.memcpy." ||
name.substr(0,13) == "llvm.memmove." ||
name == "strcpy") {
instrumentLock(&CI);
/* Record Load src, [CI] Load dst [CI] */
// Get the destination and source pointers and cast them to void pointers.
Value *dstPointer = CI.getOperand(0);
Value *srcPointer = CI.getOperand(1);
dstPointer = castTo(dstPointer, VoidPtrType, dstPointer->getName(), &CI);
srcPointer = castTo(srcPointer, VoidPtrType, srcPointer->getName(), &CI);
// Get the ID of the ext fun call instruction.
Value *CallID = ConstantInt::get(Int32Type, lsNumPass->getID(&CI));
// Create the call to the run-time to record the loads and stores of
// external call instruction.
if(name == "strcpy") {
// FIXME: If the tracer function should be inserted before or after????
std::vector<Value *> args = make_vector(CallID, srcPointer, 0);
CallInst::Create(RecordStrLoad, args, "", &CI);
args = make_vector(CallID, dstPointer, 0);
CallInst *recStore = CallInst::Create(RecordStrStore, args, "", &CI);
CI.moveBefore(recStore);
} else {
// get the num elements to be transfered
Value *NumElts = CI.getOperand(2);
std::vector<Value *> args = make_vector(CallID, srcPointer, NumElts, 0);
CallInst::Create(RecordLoad, args, "", &CI);
args = make_vector(CallID, dstPointer, NumElts, 0);
CallInst::Create(RecordStore, args, "", &CI);
}
instrumentUnlock(&CI);
++NumExtFuns; // Update statistics
return true;
} else if (name == "strcat") { /* Record Load dst, Load Src, Store dst-end before call inst */
instrumentLock(&CI);
// Get the destination and source pointers and cast them to void pointers.
Value *dstPointer = CI.getOperand(0);
Value *srcPointer = CI.getOperand(1);
dstPointer = castTo(dstPointer, VoidPtrType, dstPointer->getName(), &CI);
srcPointer = castTo(srcPointer, VoidPtrType, srcPointer->getName(), &CI);
// Get the ID of the ext fun call instruction.
Value *CallID = ConstantInt::get(Int32Type, lsNumPass->getID(&CI));
// Create the call to the run-time to record the loads and stores of
// external call instruction.
// CHECK: If the tracer function should be inserted before or after????
std::vector<Value *> args = make_vector(CallID, dstPointer, 0);
CallInst::Create(RecordStrLoad, args, "", &CI);
args = make_vector(CallID, srcPointer, 0);
CallInst::Create(RecordStrLoad, args, "", &CI);
// Record the addresses before concat as they will be lost after concat
args = make_vector(CallID, dstPointer, srcPointer, 0);
CallInst::Create(RecordStrcatStore, args, "", &CI);
instrumentUnlock(&CI);
++NumExtFuns; // Update statistics
return true;
} else if (name == "strlen") { /* Record Load */
instrumentLock(&CI);
// Get the destination and source pointers and cast them to void pointers.
Value *srcPointer = CI.getOperand(0);
srcPointer = castTo(srcPointer, VoidPtrType, srcPointer->getName(), &CI);
// Get the ID of the ext fun call instruction.
Value *CallID = ConstantInt::get(Int32Type, lsNumPass->getID(&CI));
std::vector<Value *> args = make_vector(CallID, srcPointer, 0);
CallInst::Create(RecordStrLoad, args, "", &CI);
instrumentUnlock(&CI);
++NumExtFuns; // Update statistics
return true;
} else if (name == "calloc") {
instrumentLock(&CI);
// Get the number of bytes that will be written into the buffer.
Value *NumElts = BinaryOperator::Create(BinaryOperator::Mul,
CI.getOperand(0),
CI.getOperand(1),
"calloc par1 * par2",
&CI);
// Get the destination pointer and cast it to a void pointer.
// Instruction * dstPointerInst;
Value *dstPointer = castTo(&CI, VoidPtrType, CI.getName(), &CI);
/* // To move after call inst, we need to know if cast is a constant expr or inst
if ((dstPointerInst = dyn_cast<Instruction>(dstPointer))) {
CI.moveBefore(dstPointerInst); // dstPointerInst->insertAfter(&CI);
// ((Instruction *)NumElts)->insertAfter(dstPointerInst);
}
else {
CI.moveBefore((Instruction *)NumElts); // ((Instruction *)NumElts)->insertAfter(&CI);
}
dstPointer = dstPointerInst; // Assign to dstPointer for instrn or non-instrn values
*/
// Get the ID of the external funtion call instruction.
Value *CallID = ConstantInt::get(Int32Type, lsNumPass->getID(&CI));
//
// Create the call to the run-time to record the external call instruction.
//
std::vector<Value *> args = make_vector(CallID, dstPointer, NumElts, 0);
CallInst *recStore = CallInst::Create(RecordStore, args, "", &CI);
CI.moveBefore(recStore); //recStore->insertAfter((Instruction *)NumElts);
// Moove cast, #byte computation and store to after call inst
CI.moveBefore(cast<Instruction>(NumElts));
instrumentUnlock(&CI);
++NumExtFuns; // Update statistics
return true;
} else if (name == "tolower" || name == "toupper") {
// Not needed as there are no loads and stores
/* } else if (name == "strncpy/itoa/stdarg/scanf/fscanf/sscanf/fread/complex/strftime/strptime/asctime/ctime") { */
} else if (name == "fscanf") {
// TODO
// In stead of parsing format string, can we use the type of the arguments??
} else if (name == "sscanf") {
// TODO
} else if (name == "sprintf") {
instrumentLock(&CI);
// Get the pointer to the destination buffer.
Value *dstPointer = CI.getOperand(0);
dstPointer = castTo(dstPointer, VoidPtrType, dstPointer->getName(), &CI);
// Get the ID of the call instruction.
Value *CallID = ConstantInt::get(Int32Type, lsNumPass->getID(&CI));
// Scan through the arguments looking for what appears to be a character
// string. Generate load records for each of these strings.
for (unsigned index = 2; index < CI.getNumOperands(); ++index) {
if (CI.getOperand(index)->getType() == VoidPtrType) {
// Create the call to the run-time to record the load from the string.
// What about other loads??
Value *Ptr = CI.getOperand(index);
std::vector<Value *> args = make_vector(CallID, Ptr, 0);
CallInst::Create(RecordStrLoad, args, "", &CI);
++NumLoadStrings; // Update statistics
}
}
// Create the call to the run-time to record the external call instruction.
std::vector<Value *> args = make_vector(CallID, dstPointer, 0);
CallInst *recStore = CallInst::Create(RecordStrStore, args, "", &CI);
CI.moveBefore(recStore);
instrumentUnlock(&CI);
++NumStoreStrings; // Update statistics
return true;
} else if (name == "fgets") {
instrumentLock(&CI);
// Get the pointer to the destination buffer.
Value * dstPointer = CI.getOperand(0);
dstPointer = castTo(dstPointer, VoidPtrType, dstPointer->getName(), &CI);
// Get the ID of the ext fun call instruction.
Value * CallID = ConstantInt::get(Int32Type, lsNumPass->getID(&CI));
// Create the call to the run-time to record the external call instruction.
std::vector<Value *> args = make_vector(CallID, dstPointer, 0);
CallInst *recStore = CallInst::Create(RecordStrStore, args, "", &CI);
CI.moveBefore(recStore);
instrumentUnlock(&CI);
// Update statistics
++NumStoreStrings;
return true;
}
return false;
}
bool InlineModule::runOnModule( Module & M ) {
std::vector<std::string> leafNames;
//File *file = fopen("inline_info.txt", "r");
//if (!file) {
// errs() << "Error: Could not open inline_info file.\n";
// retrun true;
//}
std::string line;
std::ifstream file ("inline_info.txt");
if(file.is_open()) {
while(std::getline(file, line))
leafNames.push_back(line);
file.close();
}
else
errs() << "Error: Could not open inline_info file.\n";
//makeLeaf.push_back(M.getFunction("ORACLE_0"));
//makeLeaf.push_back(M.getFunction("ORACLE_1"));
//makeLeaf.push_back(M.getFunction("ORACLE_2"));
//makeLeaf.push_back(M.getFunction("ORACLE_3"));
for (std::vector<std::string>::iterator i = leafNames.begin(), e = leafNames.end();
i!=e; ++i) {
if (debugInlining)
errs() << "inline_info: " << *i << "\n";
makeLeaf.push_back(M.getFunction(*i));
}
// First, get a pointer to previous analysis results
CallGraph & CG = getAnalysis<CallGraph>();
CallGraphNode * entry = CG.getRoot();
if( entry && entry->getFunction() && debugInlining)
errs() << "Entry is function: " << entry->getFunction()->getName() << "\n";
// Iterate over all SCCs in the module in bottom-up order
for( scc_iterator<CallGraph*>
si=scc_begin( &CG ), se=scc_end( &CG ); si != se; ++si ) {
runOnSCC( *si );
}
//reverse the vector for preorder
std::reverse(vectPostOrder.begin(),vectPostOrder.end());
for(std::vector<Function*>::iterator vit = vectPostOrder.begin(), vitE = vectPostOrder.end();
vit!=vitE; ++vit) {
Function *f = *vit;
runOnFunction(*f);
}
// now we have all the call sites which need to be inlined
// inline from the leaves all the way up
const TargetData *TD = getAnalysisIfAvailable<TargetData>();
InlineFunctionInfo InlineInfo(&CG, TD);
std::reverse(inlineCallInsts.begin(),inlineCallInsts.end());
for (std::vector<CallInst*>::iterator i = inlineCallInsts.begin(), e = inlineCallInsts.end();
i!=e; ++i) {
CallInst* CI = *i;
bool success = InlineFunction(CI, InlineInfo, false);
if(!success) {
if (debugInlining)
errs() << "Error: Could not inline callee function " << CI->getCalledFunction()->getName()
<< " into caller function " << "\n";
continue;
}
if (debugInlining)
errs() << "Successfully inlined callee function " << CI->getCalledFunction()->getName()
<< "into caller function " << "\n";
}
return false;
}
void TracingNoGiri::visitCallInst(CallInst &CI) {
// Attempt to get the called function.
Function *CalledFunc = CI.getCalledFunction();
if (!CalledFunc)
return;
// Do not instrument calls to tracing run-time functions or debug functions.
if (isTracerFunction(CalledFunc))
return;
if (!CalledFunc->getName().str().compare(0,9,"llvm.dbg."))
return;
// Instrument external calls which can have invariants on its return value
if (CalledFunc->isDeclaration() && CalledFunc->isIntrinsic()) {
// Instrument special external calls which loads/stores
// e.g. strlen(), strcpy(), memcpy() etc.
visitSpecialCall(CI);
return;
}
// If the called value is inline assembly code, then don't instrument it.
if (isa<InlineAsm>(CI.getCalledValue()->stripPointerCasts()))
return;
instrumentLock(&CI);
// Get the ID of the store instruction.
Value *CallID = ConstantInt::get(Int32Type, lsNumPass->getID(&CI));
// Get the called function value and cast it to a void pointer.
Value *FP = castTo(CI.getCalledValue(), VoidPtrType, "", &CI);
// Create the call to the run-time to record the call instruction.
std::vector<Value *> args = make_vector<Value *>(CallID, FP, 0);
// Do not add calls to function call stack for external functions
// as return records won't be used/needed for them, so call a special record function
// FIXME!!!! Do we still need it after adding separate return records????
Instruction *RC;
if (CalledFunc->isDeclaration())
RC = CallInst::Create(RecordExtCall, args, "", &CI);
else
RC = CallInst::Create(RecordCall, args, "", &CI);
instrumentUnlock(RC);
// Create the call to the run-time to record the return of call instruction.
CallInst *CallInst = CallInst::Create(RecordReturn, args, "", &CI);
CI.moveBefore(CallInst);
instrumentLock(CallInst);
instrumentUnlock(CallInst);
++NumCalls; // Update statistics
// The best way to handle external call is to set a flag before calling ext fn and
// use that to determine if an internal function is called from ext fn. It flag can be
// reset afterwards and restored to its original value before returning to ext code.
// FIXME!!!! LATER
#if 0
if (CalledFunc->isDeclaration() &&
CalledFunc->getName().str() == "pthread_create") {
// If pthread_create is called then handle it specially as it calls
// functions externally and add an extra call for the externally
// called functions with the same id so that returns can match with it.
// In addition to a function call to pthread_create.
// Get the external function pointer operand and cast it to a void pointer
Value *FP = castTo(CI.getOperand(2), VoidPtrType, "", &CI);
// Create the call to the run-time to record the call instruction.
std::vector<Value *> argsExt = make_vector<Value *>(CallID, FP, 0);
CallInst = CallInst::Create(RecordCall, argsExt, "", &CI);
CI.moveBefore(CallInst);
// Update statistics
++Calls;
// For, both external functions and internal/ext functions called from
// external functions, return records are not useful as they won't be used.
// Since, we won't create return records for them, simply update the call
// stack to mark the end of function call.
//args = make_vector<Value *>(CallID, FP, 0);
//CallInst::Create(RecordExtCallRet, args.begin(), args.end(), "", &CI);
// Create the call to the run-time to record the return of call instruction.
CallInst::Create(RecordReturn, argsExt, "", &CI);
}
#endif
// Instrument special external calls which loads/stores
// like strlen, strcpy, memcpy etc.
visitSpecialCall(CI);
}
int compile(list<string> args, list<string> kgen_args,
string merge, list<string> merge_args,
string input, string output, int arch,
string host_compiler, string fileprefix)
{
//
// The LLVM compiler to emit IR.
//
const char* llvm_compiler = "kernelgen-gfortran";
//
// Interpret kernelgen compile options.
//
for (list<string>::iterator iarg = kgen_args.begin(),
iearg = kgen_args.end(); iarg != iearg; iarg++)
{
const char* arg = (*iarg).c_str();
if (!strncmp(arg, "-Wk,--llvm-compiler=", 20))
llvm_compiler = arg + 20;
}
//
// Generate temporary output file.
// Check if output file is specified in the command line.
// Replace or add output to the temporary file.
//
cfiledesc tmp_output = cfiledesc::mktemp(fileprefix);
bool output_specified = false;
for (list<string>::iterator iarg = args.begin(),
iearg = args.end(); iarg != iearg; iarg++)
{
const char* arg = (*iarg).c_str();
if (!strcmp(arg, "-o"))
{
iarg++;
*iarg = tmp_output.getFilename();
output_specified = true;
break;
}
}
if (!output_specified)
{
args.push_back("-o");
args.push_back(tmp_output.getFilename());
}
//
// 1) Compile source code using regular host compiler.
//
{
if (verbose)
{
cout << host_compiler;
for (list<string>::iterator iarg = args.begin(),
iearg = args.end(); iarg != iearg; iarg++)
cout << " " << *iarg;
cout << endl;
}
int status = execute(host_compiler, args, "", NULL, NULL);
if (status) return status;
}
//
// 2) Emit LLVM IR.
//
string out = "";
{
list<string> emit_ir_args;
for (list<string>::iterator iarg = args.begin(),
iearg = args.end(); iarg != iearg; iarg++)
{
const char* arg = (*iarg).c_str();
if (!strcmp(arg, "-c") || !strcmp(arg, "-o"))
{
iarg++;
continue;
}
if (!strcmp(arg, "-g"))
{
continue;
}
emit_ir_args.push_back(*iarg);
}
emit_ir_args.push_back("-fplugin=/opt/kernelgen/lib/dragonegg.so");
emit_ir_args.push_back("-fplugin-arg-dragonegg-emit-ir");
emit_ir_args.push_back("-S");
emit_ir_args.push_back(input);
emit_ir_args.push_back("-o");
emit_ir_args.push_back("-");
if (verbose)
{
cout << llvm_compiler;
for (list<string>::iterator iarg = emit_ir_args.begin(),
iearg = emit_ir_args.end(); iarg != iearg; iarg++)
cout << " " << *iarg;
cout << endl;
}
int status = execute(llvm_compiler, emit_ir_args, "", &out, NULL);
if (status) return status;
}
//
// 3) Record existing module functions.
//
LLVMContext &context = getGlobalContext();
SMDiagnostic diag;
MemoryBuffer* buffer1 = MemoryBuffer::getMemBuffer(out);
auto_ptr<Module> m1;
m1.reset(ParseIR(buffer1, diag, context));
//m1.get()->dump();
//
// 4) Inline calls and extract loops into new functions.
//
MemoryBuffer* buffer2 = MemoryBuffer::getMemBuffer(out);
auto_ptr<Module> m2;
m2.reset(ParseIR(buffer2, diag, context));
{
PassManager manager;
manager.add(createInstructionCombiningPass());
manager.run(*m2.get());
}
std::vector<CallInst *> LoopFuctionCalls;
{
PassManager manager;
manager.add(createBranchedLoopExtractorPass(LoopFuctionCalls));
manager.run(*m2.get());
}
//m2.get()->dump();
//
// 5) Replace call to loop functions with call to launcher.
// Append "always inline" attribute to all other functions.
//
Type* int32Ty = Type::getInt32Ty(context);
Function* launch = Function::Create(
TypeBuilder<types::i<32>(types::i<8>*, types::i<64>, types::i<32>*), true>::get(context),
GlobalValue::ExternalLinkage, "kernelgen_launch", m2.get());
for (Module::iterator f1 = m2.get()->begin(), fe1 = m2.get()->end(); f1 != fe1; f1++)
{
Function* func = f1;
if (func->isDeclaration()) continue;
// Search for the current function in original module
// functions list.
// If function is not in list of original module, then
// it is generated by the loop extractor.
// Append "always inline" attribute to all other functions.
if (m1.get()->getFunction(func->getName()))
{
const AttrListPtr attr = func->getAttributes();
const AttrListPtr attr_new = attr.addAttr(~0U, Attribute::AlwaysInline);
func->setAttributes(attr_new);
continue;
}
// Each such function must be extracted to the
// standalone module and packed into resulting
// object file data section.
if (verbose)
cout << "Preparing loop function " << func->getName().data() <<
" ..." << endl;
// Reset to default visibility.
func->setVisibility(GlobalValue::DefaultVisibility);
// Reset to default linkage.
func->setLinkage(GlobalValue::ExternalLinkage);
// Replace call to this function in module with call to launcher.
bool found = false;
for (Module::iterator f2 = m2->begin(), fe2 = m2->end(); (f2 != fe2) && !found; f2++)
for (Function::iterator bb = f2->begin(); (bb != f2->end()) && !found; bb++)
for (BasicBlock::iterator i = bb->begin(); i != bb->end(); i++)
{
// Check if instruction in focus is a call.
CallInst* call = dyn_cast<CallInst>(cast<Value>(i));
if (!call) continue;
// Check if function is called (needs -instcombine pass).
Function* callee = call->getCalledFunction();
if (!callee) continue;
if (callee->isDeclaration()) continue;
if (callee->getName() != func->getName()) continue;
// Create a constant array holding original called
// function name.
Constant* name = ConstantArray::get(
context, callee->getName(), true);
// Create and initialize the memory buffer for name.
ArrayType* nameTy = cast<ArrayType>(name->getType());
AllocaInst* nameAlloc = new AllocaInst(nameTy, "", call);
StoreInst* nameInit = new StoreInst(name, nameAlloc, "", call);
Value* Idx[2];
Idx[0] = Constant::getNullValue(Type::getInt32Ty(context));
Idx[1] = ConstantInt::get(Type::getInt32Ty(context), 0);
GetElementPtrInst* namePtr = GetElementPtrInst::Create(nameAlloc, Idx, "", call);
// Add pointer to the original function string name.
SmallVector<Value*, 16> call_args;
call_args.push_back(namePtr);
// Add size of the aggregated arguments structure.
{
BitCastInst* BC = new BitCastInst(
call->getArgOperand(0), Type::getInt64PtrTy(context),
"", call);
LoadInst* LI = new LoadInst(BC, "", call);
call_args.push_back(LI);
}
// Add original aggregated structure argument.
call_args.push_back(call->getArgOperand(0));
// Create new function call with new call arguments
// and copy old call properties.
CallInst* newcall = CallInst::Create(launch, call_args, "", call);
//newcall->takeName(call);
newcall->setCallingConv(call->getCallingConv());
newcall->setAttributes(call->getAttributes());
newcall->setDebugLoc(call->getDebugLoc());
// Replace old call with new one.
call->replaceAllUsesWith(newcall);
call->eraseFromParent();
found = true;
break;
}
}
//m2.get()->dump();
//
// 6) Apply optimization passes to the resulting common
// module.
//
{
PassManager manager;
manager.add(createLowerSetJmpPass());
PassManagerBuilder builder;
builder.Inliner = createFunctionInliningPass();
builder.OptLevel = 3;
builder.DisableSimplifyLibCalls = true;
builder.populateModulePassManager(manager);
manager.run(*m2.get());
}
//m2.get()->dump();
//
// 7) Embed the resulting module into object file.
//
{
string ir_string;
raw_string_ostream ir(ir_string);
ir << (*m2.get());
celf e(tmp_output.getFilename(), output);
e.getSection(".data")->addSymbol(
"__kernelgen_" + string(input),
ir_string.c_str(), ir_string.size() + 1);
}
return 0;
}
void MutationGen::genSTDCall(Instruction * inst, StringRef fname, int index){
CallInst *call = cast<CallInst>(inst);
Function *fun = call->getCalledFunction();
if(fun->getName().startswith("llvm")){
return;
}
Type *t = call->getCalledValue()->getType();
FunctionType* ft = cast<FunctionType>(cast<PointerType>(t)->getElementType());
Type *tt = ft->getReturnType();
//tt->dump();
if(tt->isIntegerTy(32)){
//1. if the func returns a int32 val, let it be 0, 1 or a random number
//errs()<<"IT IS A 32 !!\n";
std::stringstream ss;
ss<<"STD:"<<std::string(fname)<<":"<<index<< ":"<<inst->getOpcode()
<< ":"<<32<<":0\n";
muts_num++;
ss<<"STD:"<<std::string(fname)<<":"<<index<< ":"<<inst->getOpcode()
<< ":"<<32<<":1\n";
muts_num++;
//srand((int)time(0));
//int random = rand();
ss<<"STD:"<<std::string(fname)<<":"<<index<< ":"<<inst->getOpcode()
<< ":"<<32<<":"<<-1<<"\n";
muts_num++;
ofresult<<ss.str();
ofresult.flush();
}else if(tt->isVoidTy()){
//2. if the func returns void, subsitute @llvm.donothing for the func
//errs()<<"IT IS A VOID !!\n";
std::stringstream ss;
ss<<"STD:"<<std::string(fname)<<":"<<index<< ":"<<inst->getOpcode()
<< ":"<<0<<'\n';
ofresult<<ss.str();
ofresult.flush();
muts_num++;
}else if(tt->isIntegerTy(64)){
//1. if the func returns a int64 val, let it be 0, 1 or a random number
//errs()<<"IT IS A 64 !!\n";
std::stringstream ss;
ss<<"STD:"<<std::string(fname)<<":"<<index<< ":"<<inst->getOpcode()
<< ":"<<64<<":0\n";
muts_num++;
ss<<"STD:"<<std::string(fname)<<":"<<index<< ":"<<inst->getOpcode()
<< ":"<<64<<":1\n";
muts_num++;
//srand((int)time(0));
//int random = rand();
ss<<"STD:"<<std::string(fname)<<":"<<index<< ":"<<inst->getOpcode()
<< ":"<<64<<":"<<-1<<"\n";
muts_num++;
ofresult<<ss.str();
ofresult.flush();
}
}
/// InlineHalfPowrs - Inline a sequence of adjacent half_powr calls, rearranging
/// their control flow to better facilitate subsequent optimization.
Instruction *
SimplifyHalfPowrLibCalls::
InlineHalfPowrs(const std::vector<Instruction *> &HalfPowrs,
Instruction *InsertPt) {
std::vector<BasicBlock *> Bodies;
BasicBlock *NewBlock = 0;
for (unsigned i = 0, e = HalfPowrs.size(); i != e; ++i) {
CallInst *Call = cast<CallInst>(HalfPowrs[i]);
Function *Callee = Call->getCalledFunction();
// Minimally sanity-check the CFG of half_powr to ensure that it contains
// the kind of code we expect. If we're running this pass, we have
// reason to believe it will be what we expect.
Function::iterator I = Callee->begin();
BasicBlock *Prologue = I++;
if (I == Callee->end()) break;
BasicBlock *SubnormalHandling = I++;
if (I == Callee->end()) break;
BasicBlock *Body = I++;
if (I != Callee->end()) break;
if (SubnormalHandling->getSinglePredecessor() != Prologue)
break;
BranchInst *PBI = dyn_cast<BranchInst>(Prologue->getTerminator());
if (!PBI || !PBI->isConditional())
break;
BranchInst *SNBI = dyn_cast<BranchInst>(SubnormalHandling->getTerminator());
if (!SNBI || SNBI->isConditional())
break;
if (!isa<ReturnInst>(Body->getTerminator()))
break;
Instruction *NextInst = llvm::next(BasicBlock::iterator(Call));
// Inline the call, taking care of what code ends up where.
NewBlock = SplitBlock(NextInst->getParent(), NextInst, this);
InlineFunctionInfo IFI(0, TD);
bool B = InlineFunction(Call, IFI);
assert(B && "half_powr didn't inline?");
(void)B;
BasicBlock *NewBody = NewBlock->getSinglePredecessor();
assert(NewBody);
Bodies.push_back(NewBody);
}
if (!NewBlock)
return InsertPt;
// Put the code for all the bodies into one block, to facilitate
// subsequent optimization.
(void)SplitEdge(NewBlock->getSinglePredecessor(), NewBlock, this);
for (unsigned i = 0, e = Bodies.size(); i != e; ++i) {
BasicBlock *Body = Bodies[i];
Instruction *FNP = Body->getFirstNonPHI();
// Splice the insts from body into NewBlock.
NewBlock->getInstList().splice(NewBlock->begin(), Body->getInstList(),
FNP, Body->getTerminator());
}
return NewBlock->begin();
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Clone functions that take GEPs as arguments
//
// Inputs:
// M - A reference to the LLVM module to transform
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool GEPExprArgs::runOnModule(Module& M) {
bool changed;
do {
changed = false;
for (Module::iterator F = M.begin(); F != M.end(); ++F){
for (Function::iterator B = F->begin(), FE = F->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
CallInst *CI = dyn_cast<CallInst>(I++);
if(!CI)
continue;
if(CI->hasByValArgument())
continue;
// if the GEP calls a function, that is externally defined,
// or might be changed, ignore this call site.
Function *F = CI->getCalledFunction();
if (!F || (F->isDeclaration() || F->mayBeOverridden()))
continue;
if(F->hasStructRetAttr())
continue;
if(F->isVarArg())
continue;
// find the argument we must replace
Function::arg_iterator ai = F->arg_begin(), ae = F->arg_end();
unsigned argNum = 1;
for(; argNum < CI->getNumOperands();argNum++, ++ai) {
if(ai->use_empty())
continue;
if (isa<GEPOperator>(CI->getOperand(argNum)))
break;
}
// if no argument was a GEP operator to be changed
if(ai == ae)
continue;
GEPOperator *GEP = dyn_cast<GEPOperator>(CI->getOperand(argNum));
if(!GEP->hasAllConstantIndices())
continue;
// Construct the new Type
// Appends the struct Type at the beginning
std::vector<Type*>TP;
TP.push_back(GEP->getPointerOperand()->getType());
for(unsigned c = 1; c < CI->getNumOperands();c++) {
TP.push_back(CI->getOperand(c)->getType());
}
//return type is same as that of original instruction
FunctionType *NewFTy = FunctionType::get(CI->getType(), TP, false);
Function *NewF;
numSimplified++;
if(numSimplified > 800)
return true;
NewF = Function::Create(NewFTy,
GlobalValue::InternalLinkage,
F->getName().str() + ".TEST",
&M);
Function::arg_iterator NI = NewF->arg_begin();
NI->setName("GEParg");
++NI;
ValueToValueMapTy ValueMap;
for (Function::arg_iterator II = F->arg_begin(); NI != NewF->arg_end(); ++II, ++NI) {
ValueMap[II] = NI;
NI->setName(II->getName());
NI->addAttr(F->getAttributes().getParamAttributes(II->getArgNo() + 1));
}
NewF->setAttributes(NewF->getAttributes().addAttr(
0, F->getAttributes().getRetAttributes()));
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
CloneFunctionInto(NewF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NewF->arg_begin(),
ae= NewF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NewF->setAttributes(NewF->getAttributes().addAttr(
~0, F->getAttributes().getFnAttributes()));
//Get the point to insert the GEP instr.
SmallVector<Value*, 8> Ops(CI->op_begin()+1, CI->op_end());
Instruction *InsertPoint;
for (BasicBlock::iterator insrt = NewF->front().begin();
isa<AllocaInst>(InsertPoint = insrt); ++insrt) {;}
NI = NewF->arg_begin();
SmallVector<Value*, 8> Indices;
Indices.append(GEP->op_begin()+1, GEP->op_end());
GetElementPtrInst *GEP_new = GetElementPtrInst::Create(cast<Value>(NI),
Indices,
"", InsertPoint);
fargs.at(argNum)->replaceAllUsesWith(GEP_new);
unsigned j = argNum + 1;
for(; j < CI->getNumOperands();j++) {
if(CI->getOperand(j) == GEP)
fargs.at(j)->replaceAllUsesWith(GEP_new);
}
SmallVector<AttributeWithIndex, 8> AttributesVec;
// Get the initial attributes of the call
AttrListPtr CallPAL = CI->getAttributes();
Attributes RAttrs = CallPAL.getRetAttributes();
Attributes FnAttrs = CallPAL.getFnAttributes();
if (RAttrs)
AttributesVec.push_back(AttributeWithIndex::get(0, RAttrs));
SmallVector<Value*, 8> Args;
Args.push_back(GEP->getPointerOperand());
for(unsigned j =1;j<CI->getNumOperands();j++) {
Args.push_back(CI->getOperand(j));
// position in the AttributesVec
if (Attributes Attrs = CallPAL.getParamAttributes(j))
AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
}
// Create the new attributes vec.
if (FnAttrs != Attribute::None)
AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
AttrListPtr NewCallPAL = AttrListPtr::get(AttributesVec.begin(),
AttributesVec.end());
CallInst *CallI = CallInst::Create(NewF,Args,"", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
CI->replaceAllUsesWith(CallI);
CI->eraseFromParent();
changed = true;
}
}
}
} while(changed);
return true;
}
/// performLocalRetainMotion - Scan forward from the specified retain, moving it
/// later in the function if possible, over instructions that provably can't
/// release the object. If we get to a release of the object, zap both.
///
/// NOTE: this handles both objc_retain and swift_retain.
///
static bool performLocalRetainMotion(CallInst &Retain, BasicBlock &BB,
SwiftRCIdentity *RC) {
// FIXME: Call classifier should identify the object for us. Too bad C++
// doesn't have nice Swift-style enums.
Value *RetainedObject = RC->getSwiftRCIdentityRoot(Retain.getArgOperand(0));
BasicBlock::iterator BBI = Retain.getIterator(),
BBE = BB.getTerminator()->getIterator();
bool isObjCRetain = Retain.getCalledFunction()->getName() == "objc_retain";
bool MadeProgress = false;
// Scan until we get to the end of the block.
for (++BBI; BBI != BBE; ++BBI) {
Instruction &CurInst = *BBI;
// Classify the instruction. This switch does a "break" when the instruction
// can be skipped and is interesting, and a "continue" when it is a retain
// of the same pointer.
switch (classifyInstruction(CurInst)) {
// These instructions should not reach here based on the pass ordering.
// i.e. LLVMARCOpt -> LLVMContractOpt.
case RT_RetainN:
case RT_UnknownRetainN:
case RT_BridgeRetainN:
case RT_ReleaseN:
case RT_UnknownReleaseN:
case RT_BridgeReleaseN:
llvm_unreachable("These are only created by LLVMARCContract !");
case RT_NoMemoryAccessed:
case RT_AllocObject:
case RT_CheckUnowned:
// Skip over random instructions that don't touch memory. They don't need
// protection by retain/release.
break;
case RT_FixLifetime: // This only stops release motion. Retains can move over it.
break;
case RT_Retain:
case RT_UnknownRetain:
case RT_BridgeRetain:
case RT_RetainUnowned:
case RT_ObjCRetain: { // swift_retain(obj)
//CallInst &ThisRetain = cast<CallInst>(CurInst);
//Value *ThisRetainedObject = ThisRetain.getArgOperand(0);
// If we see a retain of the same object, we can skip over it, but we
// can't count it as progress. Just pushing a retain(x) past a retain(y)
// doesn't change the program.
continue;
}
case RT_UnknownRelease:
case RT_BridgeRelease:
case RT_ObjCRelease:
case RT_Release: {
// If we get to a release that is provably to this object, then we can zap
// it and the retain.
CallInst &ThisRelease = cast<CallInst>(CurInst);
Value *ThisReleasedObject = ThisRelease.getArgOperand(0);
ThisReleasedObject = RC->getSwiftRCIdentityRoot(ThisReleasedObject);
if (ThisReleasedObject == RetainedObject) {
Retain.eraseFromParent();
ThisRelease.eraseFromParent();
if (isObjCRetain) {
++NumObjCRetainReleasePairs;
} else {
++NumRetainReleasePairs;
}
return true;
}
// Otherwise, if this is some other pointer, we can only ignore it if we
// can prove that the two objects don't alias.
// Retain.dump(); ThisRelease.dump(); BB.getParent()->dump();
goto OutOfLoop;
}
case RT_Unknown:
// Loads cannot affect the retain.
if (isa<LoadInst>(CurInst))
continue;
// Load, store, memcpy etc can't do a release.
if (isa<LoadInst>(CurInst) || isa<StoreInst>(CurInst) ||
isa<MemIntrinsic>(CurInst))
break;
// CurInst->dump(); BBI->dump();
// Otherwise, we get to something unknown/unhandled. Bail out for now.
goto OutOfLoop;
}
// If the switch did a break, we made some progress moving this retain.
MadeProgress = true;
}
OutOfLoop:
// If we were able to move the retain down, move it now.
// TODO: This is where we'd plug in some global algorithms someday.
if (MadeProgress) {
Retain.moveBefore(&*BBI);
return true;
}
return false;
}
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();
}
}
}
}
bool NVVMReflect::runOnFunction(Function &F) {
if (!NVVMReflectEnabled)
return false;
if (F.getName() == NVVM_REFLECT_FUNCTION) {
assert(F.isDeclaration() && "_reflect function should not have a body");
assert(F.getReturnType()->isIntegerTy() &&
"_reflect's return type should be integer");
return false;
}
SmallVector<Instruction *, 4> ToRemove;
// Go through the calls in this function. Each call to __nvvm_reflect or
// llvm.nvvm.reflect should be a CallInst with a ConstantArray argument.
// First validate that. If the c-string corresponding to the ConstantArray can
// be found successfully, see if it can be found in VarMap. If so, replace the
// uses of CallInst with the value found in VarMap. If not, replace the use
// with value 0.
// The IR for __nvvm_reflect calls differs between CUDA versions.
//
// CUDA 6.5 and earlier uses this sequence:
// %ptr = tail call i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8
// (i8 addrspace(4)* getelementptr inbounds
// ([8 x i8], [8 x i8] addrspace(4)* @str, i32 0, i32 0))
// %reflect = tail call i32 @__nvvm_reflect(i8* %ptr)
//
// The value returned by Sym->getOperand(0) is a Constant with a
// ConstantDataSequential operand which can be converted to string and used
// for lookup.
//
// CUDA 7.0 does it slightly differently:
// %reflect = call i32 @__nvvm_reflect(i8* addrspacecast
// (i8 addrspace(1)* getelementptr inbounds
// ([8 x i8], [8 x i8] addrspace(1)* @str, i32 0, i32 0) to i8*))
//
// In this case, we get a Constant with a GlobalVariable operand and we need
// to dig deeper to find its initializer with the string we'll use for lookup.
for (Instruction &I : instructions(F)) {
CallInst *Call = dyn_cast<CallInst>(&I);
if (!Call)
continue;
Function *Callee = Call->getCalledFunction();
if (!Callee || (Callee->getName() != NVVM_REFLECT_FUNCTION &&
Callee->getIntrinsicID() != Intrinsic::nvvm_reflect))
continue;
// FIXME: Improve error handling here and elsewhere in this pass.
assert(Call->getNumOperands() == 2 &&
"Wrong number of operands to __nvvm_reflect function");
// In cuda 6.5 and earlier, we will have an extra constant-to-generic
// conversion of the string.
const Value *Str = Call->getArgOperand(0);
if (const CallInst *ConvCall = dyn_cast<CallInst>(Str)) {
// FIXME: Add assertions about ConvCall.
Str = ConvCall->getArgOperand(0);
}
assert(isa<ConstantExpr>(Str) &&
"Format of __nvvm__reflect function not recognized");
const ConstantExpr *GEP = cast<ConstantExpr>(Str);
const Value *Sym = GEP->getOperand(0);
assert(isa<Constant>(Sym) &&
"Format of __nvvm_reflect function not recognized");
const Value *Operand = cast<Constant>(Sym)->getOperand(0);
if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) {
// For CUDA-7.0 style __nvvm_reflect calls, we need to find the operand's
// initializer.
assert(GV->hasInitializer() &&
"Format of _reflect function not recognized");
const Constant *Initializer = GV->getInitializer();
Operand = Initializer;
}
assert(isa<ConstantDataSequential>(Operand) &&
"Format of _reflect function not recognized");
assert(cast<ConstantDataSequential>(Operand)->isCString() &&
"Format of _reflect function not recognized");
StringRef ReflectArg = cast<ConstantDataSequential>(Operand)->getAsString();
ReflectArg = ReflectArg.substr(0, ReflectArg.size() - 1);
LLVM_DEBUG(dbgs() << "Arg of _reflect : " << ReflectArg << "\n");
int ReflectVal = 0; // The default value is 0
if (ReflectArg == "__CUDA_FTZ") {
// Try to pull __CUDA_FTZ from the nvvm-reflect-ftz module flag. Our
// choice here must be kept in sync with AutoUpgrade, which uses the same
// technique to detect whether ftz is enabled.
if (auto *Flag = mdconst::extract_or_null<ConstantInt>(
F.getParent()->getModuleFlag("nvvm-reflect-ftz")))
ReflectVal = Flag->getSExtValue();
} else if (ReflectArg == "__CUDA_ARCH") {
ReflectVal = SmVersion * 10;
}
Call->replaceAllUsesWith(ConstantInt::get(Call->getType(), ReflectVal));
ToRemove.push_back(Call);
}
for (Instruction *I : ToRemove)
I->eraseFromParent();
return ToRemove.size() > 0;
}
/// \brief Recursively handle the condition leading to a loop
Value *SIAnnotateControlFlow::handleLoopCondition(Value *Cond, PHINode *Broken,
llvm::Loop *L) {
// Only search through PHI nodes which are inside the loop. If we try this
// with PHI nodes that are outside of the loop, we end up inserting new PHI
// nodes outside of the loop which depend on values defined inside the loop.
// This will break the module with
// 'Instruction does not dominate all users!' errors.
PHINode *Phi = nullptr;
if ((Phi = dyn_cast<PHINode>(Cond)) && L->contains(Phi)) {
BasicBlock *Parent = Phi->getParent();
PHINode *NewPhi = PHINode::Create(Int64, 0, "", &Parent->front());
Value *Ret = NewPhi;
// Handle all non-constant incoming values first
for (unsigned i = 0, e = Phi->getNumIncomingValues(); i != e; ++i) {
Value *Incoming = Phi->getIncomingValue(i);
BasicBlock *From = Phi->getIncomingBlock(i);
if (isa<ConstantInt>(Incoming)) {
NewPhi->addIncoming(Broken, From);
continue;
}
Phi->setIncomingValue(i, BoolFalse);
Value *PhiArg = handleLoopCondition(Incoming, Broken, L);
NewPhi->addIncoming(PhiArg, From);
}
BasicBlock *IDom = DT->getNode(Parent)->getIDom()->getBlock();
for (unsigned i = 0, e = Phi->getNumIncomingValues(); i != e; ++i) {
Value *Incoming = Phi->getIncomingValue(i);
if (Incoming != BoolTrue)
continue;
BasicBlock *From = Phi->getIncomingBlock(i);
if (From == IDom) {
CallInst *OldEnd = dyn_cast<CallInst>(Parent->getFirstInsertionPt());
if (OldEnd && OldEnd->getCalledFunction() == EndCf) {
Value *Args[] = { OldEnd->getArgOperand(0), NewPhi };
Ret = CallInst::Create(ElseBreak, Args, "", OldEnd);
continue;
}
}
TerminatorInst *Insert = From->getTerminator();
Value *PhiArg = CallInst::Create(Break, Broken, "", Insert);
NewPhi->setIncomingValue(i, PhiArg);
}
eraseIfUnused(Phi);
return Ret;
} else if (Instruction *Inst = dyn_cast<Instruction>(Cond)) {
BasicBlock *Parent = Inst->getParent();
Instruction *Insert;
if (L->contains(Inst)) {
Insert = Parent->getTerminator();
} else {
Insert = L->getHeader()->getFirstNonPHIOrDbgOrLifetime();
}
Value *Args[] = { Cond, Broken };
return CallInst::Create(IfBreak, Args, "", Insert);
} else {
llvm_unreachable("Unhandled loop condition!");
}
return 0;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass. Search for functions which could be called
// indirectly and create clones for them which are only called by direct
// calls.
//
// 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
IndClone::runOnModule(Module& M) {
// Set of functions to clone
std::vector<Function*> toClone;
//
// Check all of the functions in the module. If the function could be called
// by an indirect function call, add it to our worklist of functions to
// clone.
//
for (Module::iterator I = M.begin(); I != M.end(); ++I) {
// Flag whether the function should be cloned
bool pleaseCloneTheFunction = false;
//
// Only clone functions which are defined and cannot be replaced by another
// function by the linker.
//
if (!I->isDeclaration() && !I->mayBeOverridden()) {
for (Value::use_iterator ui = I->use_begin(), ue = I->use_end();
ui != ue; ++ui) {
if (!isa<CallInst>(*ui) && !isa<InvokeInst>(*ui)) {
if(!ui->use_empty())
//
// If this function is used for anything other than a direct function
// call, then we want to clone it.
//
pleaseCloneTheFunction = true;
} else {
//
// This is a call instruction, but hold up ranger! We need to make
// sure that the function isn't passed as an argument to *another*
// function. That would make the function usable in an indirect
// function call.
//
for (unsigned index = 1; index < ui->getNumOperands(); ++index) {
if (ui->getOperand(index)->stripPointerCasts() == I) {
pleaseCloneTheFunction = true;
break;
}
}
}
//
// If we've discovered that the function could be used by an indirect
// call site, schedule it for cloning.
//
if (pleaseCloneTheFunction) {
toClone.push_back(I);
break;
}
}
}
}
//
// Update the statistics on the number of functions we'll be cloning.
// We only update the statistic if we want to clone one or more functions;
// due to the magic of how statistics work, avoiding assignment prevents it
// from needlessly showing up.
//
if (toClone.size())
numCloned += toClone.size();
//
// Go through the worklist and clone each function. After cloning a
// function, change all direct calls to use the clone instead of using the
// original function.
//
for (unsigned index = 0; index < toClone.size(); ++index) {
//
// Clone the function and give it a name indicating that it is a clone to
// be used for direct function calls.
//
Function * Original = toClone[index];
Function* DirectF = CloneFunction(Original);
DirectF->setName(Original->getName() + "_DIRECT");
//
// Make the clone internal; external code can use the original function.
//
DirectF->setLinkage(GlobalValue::InternalLinkage);
//
// Link the cloned function into the set of functions belonging to the
// module.
//
Original->getParent()->getFunctionList().push_back(DirectF);
//
// Find all uses of the function that use it as a direct call. Change
// them to use the clone.
//
for (Value::use_iterator ui = Original->use_begin(),
ue = Original->use_end();
ui != ue; ) {
CallInst *CI = dyn_cast<CallInst>(*ui);
ui++;
if (CI) {
if (CI->getCalledFunction() == Original) {
++numReplaced;
CI->setCalledFunction(DirectF);
}
}
}
}
//
// Assume that we've cloned at least one function.
//
return true;
}
bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
bool &TailCallsAreMarkedTail,
SmallVector<PHINode*, 8> &ArgumentPHIs,
bool CannotTailCallElimCallsMarkedTail) {
BasicBlock *BB = Ret->getParent();
Function *F = BB->getParent();
if (&BB->front() == Ret) // Make sure there is something before the ret...
return false;
// If the return is in the entry block, then making this transformation would
// turn infinite recursion into an infinite loop. This transformation is ok
// in theory, but breaks some code like:
// double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
// disable this xform in this case, because the code generator will lower the
// call to fabs into inline code.
if (BB == &F->getEntryBlock())
return false;
// Scan backwards from the return, checking to see if there is a tail call in
// this block. If so, set CI to it.
CallInst *CI;
BasicBlock::iterator BBI = Ret;
while (1) {
CI = dyn_cast<CallInst>(BBI);
if (CI && CI->getCalledFunction() == F)
break;
if (BBI == BB->begin())
return false; // Didn't find a potential tail call.
--BBI;
}
// If this call is marked as a tail call, and if there are dynamic allocas in
// the function, we cannot perform this optimization.
if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
return false;
// If we are introducing accumulator recursion to eliminate associative
// operations after the call instruction, this variable contains the initial
// value for the accumulator. If this value is set, we actually perform
// accumulator recursion elimination instead of simple tail recursion
// elimination.
Value *AccumulatorRecursionEliminationInitVal = 0;
Instruction *AccumulatorRecursionInstr = 0;
// Ok, we found a potential tail call. We can currently only transform the
// tail call if all of the instructions between the call and the return are
// movable to above the call itself, leaving the call next to the return.
// Check that this is the case now.
for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI)
if (!CanMoveAboveCall(BBI, CI)) {
// If we can't move the instruction above the call, it might be because it
// is an associative operation that could be tranformed using accumulator
// recursion elimination. Check to see if this is the case, and if so,
// remember the initial accumulator value for later.
if ((AccumulatorRecursionEliminationInitVal =
CanTransformAccumulatorRecursion(BBI, CI))) {
// Yes, this is accumulator recursion. Remember which instruction
// accumulates.
AccumulatorRecursionInstr = BBI;
} else {
return false; // Otherwise, we cannot eliminate the tail recursion!
}
}
// We can only transform call/return pairs that either ignore the return value
// of the call and return void, ignore the value of the call and return a
// constant, return the value returned by the tail call, or that are being
// accumulator recursion variable eliminated.
if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
!isa<UndefValue>(Ret->getReturnValue()) &&
AccumulatorRecursionEliminationInitVal == 0 &&
!getCommonReturnValue(Ret, CI))
return false;
// OK! We can transform this tail call. If this is the first one found,
// create the new entry block, allowing us to branch back to the old entry.
if (OldEntry == 0) {
OldEntry = &F->getEntryBlock();
BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
NewEntry->takeName(OldEntry);
OldEntry->setName("tailrecurse");
BranchInst::Create(OldEntry, NewEntry);
// If this tail call is marked 'tail' and if there are any allocas in the
// entry block, move them up to the new entry block.
TailCallsAreMarkedTail = CI->isTailCall();
if (TailCallsAreMarkedTail)
// Move all fixed sized allocas from OldEntry to NewEntry.
for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
NEBI = NewEntry->begin(); OEBI != E; )
if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
if (isa<ConstantInt>(AI->getArraySize()))
AI->moveBefore(NEBI);
// Now that we have created a new block, which jumps to the entry
// block, insert a PHI node for each argument of the function.
// For now, we initialize each PHI to only have the real arguments
// which are passed in.
Instruction *InsertPos = OldEntry->begin();
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
I != E; ++I) {
PHINode *PN = PHINode::Create(I->getType(),
I->getName() + ".tr", InsertPos);
I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
PN->addIncoming(I, NewEntry);
ArgumentPHIs.push_back(PN);
}
}
// If this function has self recursive calls in the tail position where some
// are marked tail and some are not, only transform one flavor or another. We
// have to choose whether we move allocas in the entry block to the new entry
// block or not, so we can't make a good choice for both. NOTE: We could do
// slightly better here in the case that the function has no entry block
// allocas.
if (TailCallsAreMarkedTail && !CI->isTailCall())
return false;
// Ok, now that we know we have a pseudo-entry block WITH all of the
// required PHI nodes, add entries into the PHI node for the actual
// parameters passed into the tail-recursive call.
for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i)
ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB);
// If we are introducing an accumulator variable to eliminate the recursion,
// do so now. Note that we _know_ that no subsequent tail recursion
// eliminations will happen on this function because of the way the
// accumulator recursion predicate is set up.
//
if (AccumulatorRecursionEliminationInitVal) {
Instruction *AccRecInstr = AccumulatorRecursionInstr;
// Start by inserting a new PHI node for the accumulator.
PHINode *AccPN = PHINode::Create(AccRecInstr->getType(), "accumulator.tr",
OldEntry->begin());
// Loop over all of the predecessors of the tail recursion block. For the
// real entry into the function we seed the PHI with the initial value,
// computed earlier. For any other existing branches to this block (due to
// other tail recursions eliminated) the accumulator is not modified.
// Because we haven't added the branch in the current block to OldEntry yet,
// it will not show up as a predecessor.
for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry);
PI != PE; ++PI) {
if (*PI == &F->getEntryBlock())
AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI);
else
AccPN->addIncoming(AccPN, *PI);
}
// Add an incoming argument for the current block, which is computed by our
// associative accumulator instruction.
AccPN->addIncoming(AccRecInstr, BB);
// Next, rewrite the accumulator recursion instruction so that it does not
// use the result of the call anymore, instead, use the PHI node we just
// inserted.
AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
// Finally, rewrite any return instructions in the program to return the PHI
// node instead of the "initval" that they do currently. This loop will
// actually rewrite the return value we are destroying, but that's ok.
for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
RI->setOperand(0, AccPN);
++NumAccumAdded;
}
// Now that all of the PHI nodes are in place, remove the call and
// ret instructions, replacing them with an unconditional branch.
BranchInst::Create(OldEntry, Ret);
BB->getInstList().erase(Ret); // Remove return.
BB->getInstList().erase(CI); // Remove call.
++NumEliminated;
return true;
}
