// visitCallInst - This converts all LLVM call instructions into invoke
// instructions. The except part of the invoke goes to the "LongJmpBlkPre"
// that grabs the exception and proceeds to determine if it's a longjmp
// exception or not.
void LowerSetJmp::visitCallInst(CallInst& CI)
{
if (CI.getCalledFunction())
if (!IsTransformableFunction(CI.getCalledFunction()->getName()) ||
CI.getCalledFunction()->isIntrinsic()) return;
BasicBlock* OldBB = CI.getParent();
// If not reachable from a setjmp call, don't transform.
if (!DFSBlocks.count(OldBB)) return;
BasicBlock* NewBB = OldBB->splitBasicBlock(CI);
assert(NewBB && "Couldn't split BB of \"call\" instruction!!");
DFSBlocks.insert(NewBB);
NewBB->setName("Call2Invoke");
Function* Func = OldBB->getParent();
// Construct the new "invoke" instruction.
TerminatorInst* Term = OldBB->getTerminator();
std::vector<Value*> Params(CI.op_begin() + 1, CI.op_end());
InvokeInst* II =
InvokeInst::Create(CI.getCalledValue(), NewBB, PrelimBBMap[Func],
Params.begin(), Params.end(), CI.getName(), Term);
II->setCallingConv(CI.getCallingConv());
II->setParamAttrs(CI.getParamAttrs());
// Replace the old call inst with the invoke inst and remove the call.
CI.replaceAllUsesWith(II);
CI.getParent()->getInstList().erase(&CI);
// The old terminator is useless now that we have the invoke inst.
Term->getParent()->getInstList().erase(Term);
++CallsTransformed;
}
bool LowerExpectIntrinsic::runOnFunction(Function &F) {
for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
BasicBlock *BB = I++;
// Create "block_weights" metadata.
if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
if (HandleIfExpect(BI))
IfHandled++;
} else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
if (HandleSwitchExpect(SI))
IfHandled++;
}
// remove llvm.expect intrinsics.
for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
BI != BE; ) {
CallInst *CI = dyn_cast<CallInst>(BI++);
if (!CI)
continue;
Function *Fn = CI->getCalledFunction();
if (Fn && Fn->getIntrinsicID() == Intrinsic::expect) {
Value *Exp = CI->getArgOperand(0);
CI->replaceAllUsesWith(Exp);
CI->eraseFromParent();
}
}
}
return false;
}
/**
* removeUndefCalls -- remove calls with undef function
*
* These are irrelevant to the code, so may be removed completely.
*/
void FunctionStaticSlicer::removeUndefCalls(ModulePass *MP, Function &F) {
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E;) {
CallInst *CI = dyn_cast<CallInst>(&*I);
++I;
if (CI && isa<UndefValue>(CI->getCalledValue())) {
CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
CI->eraseFromParent();
}
}
}
static bool ExpandOpForIntSize(Module *M, unsigned Bits, bool Mul) {
IntegerType *IntTy = IntegerType::get(M->getContext(), Bits);
SmallVector<Type *, 1> Types;
Types.push_back(IntTy);
Intrinsic::ID ID = (Mul ? Intrinsic::umul_with_overflow
: Intrinsic::uadd_with_overflow);
std::string Name = Intrinsic::getName(ID, Types);
Function *Intrinsic = M->getFunction(Name);
if (!Intrinsic)
return false;
for (Value::use_iterator CallIter = Intrinsic->use_begin(),
E = Intrinsic->use_end(); CallIter != E; ) {
CallInst *Call = dyn_cast<CallInst>(*CallIter++);
if (!Call) {
report_fatal_error("ExpandArithWithOverflow: Taking the address of a "
"*.with.overflow intrinsic is not allowed");
}
Value *VariableArg;
ConstantInt *ConstantArg;
if (ConstantInt *C = dyn_cast<ConstantInt>(Call->getArgOperand(0))) {
VariableArg = Call->getArgOperand(1);
ConstantArg = C;
} else if (ConstantInt *C = dyn_cast<ConstantInt>(Call->getArgOperand(1))) {
VariableArg = Call->getArgOperand(0);
ConstantArg = C;
} else {
errs() << "Use: " << *Call << "\n";
report_fatal_error("ExpandArithWithOverflow: At least one argument of "
"*.with.overflow must be a constant");
}
Value *ArithResult = BinaryOperator::Create(
(Mul ? Instruction::Mul : Instruction::Add), VariableArg, ConstantArg,
Call->getName() + ".arith", Call);
uint64_t ArgMax;
if (Mul) {
ArgMax = UintTypeMax(Bits) / ConstantArg->getZExtValue();
} else {
ArgMax = UintTypeMax(Bits) - ConstantArg->getZExtValue();
}
Value *OverflowResult = new ICmpInst(
Call, CmpInst::ICMP_UGT, VariableArg, ConstantInt::get(IntTy, ArgMax),
Call->getName() + ".overflow");
// Construct the struct result.
Value *NewStruct = UndefValue::get(Call->getType());
NewStruct = CreateInsertValue(NewStruct, 0, ArithResult, Call);
NewStruct = CreateInsertValue(NewStruct, 1, OverflowResult, Call);
Call->replaceAllUsesWith(NewStruct);
Call->eraseFromParent();
}
Intrinsic->eraseFromParent();
return true;
}
/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
/// an invoke, we have to turn all of the calls that can throw into
/// invokes. This function analyze BB to see if there are any calls, and if so,
/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
/// nodes in that block with the values specified in InvokeDestPHIValues.
///
/// Returns true to indicate that the next block should be skipped.
static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
InvokeInliningInfo &Invoke) {
LandingPadInst *LPI = Invoke.getLandingPadInst();
for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
Instruction *I = BBI++;
if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) {
unsigned NumClauses = LPI->getNumClauses();
L->reserveClauses(NumClauses);
for (unsigned i = 0; i != NumClauses; ++i)
L->addClause(LPI->getClause(i));
}
// We only need to check for function calls: inlined invoke
// instructions require no special handling.
CallInst *CI = dyn_cast<CallInst>(I);
// If this call cannot unwind, don't convert it to an invoke.
// Inline asm calls cannot throw.
if (!CI || CI->doesNotThrow() || 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");
// Delete the unconditional branch inserted by splitBasicBlock
BB->getInstList().pop_back();
// Create the new invoke instruction.
ImmutableCallSite CS(CI);
SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
Invoke.getOuterResumeDest(),
InvokeArgs, CI->getName(), BB);
II->setCallingConv(CI->getCallingConv());
II->setAttributes(CI->getAttributes());
// Make sure that anything using the call now uses the invoke! This also
// updates the CallGraph if present, because it uses a WeakVH.
CI->replaceAllUsesWith(II);
// Delete the original call
Split->getInstList().pop_front();
// Update any PHI nodes in the exceptional block to indicate that there is
// now a new entry in them.
Invoke.addIncomingPHIValuesFor(BB);
return false;
}
return false;
}
/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
/// an invoke, we have to turn all of the calls that can throw into
/// invokes. This function analyze BB to see if there are any calls, and if so,
/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
/// nodes in that block with the values specified in InvokeDestPHIValues.
///
static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
BasicBlock *InvokeDest,
const SmallVectorImpl<Value*> &InvokeDestPHIValues) {
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.
CallInst *CI = dyn_cast<CallInst>(I);
if (CI == 0) continue;
// If this call cannot unwind, don't convert it to an invoke.
if (CI->doesNotThrow())
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.
ImmutableCallSite CS(CI);
SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
InvokeInst *II =
InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest,
InvokeArgs.begin(), InvokeArgs.end(),
CI->getName(), BB->getTerminator());
II->setCallingConv(CI->getCallingConv());
II->setAttributes(CI->getAttributes());
// Make sure that anything using the call now uses the invoke! This also
// updates the CallGraph if present, because it uses a WeakVH.
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)
cast<PHINode>(I)->addIncoming(InvokeDestPHIValues[i], BB);
// This basic block is now complete, the caller will continue scanning the
// next one.
return;
}
}
CallInst* FunctionCalls::changeFunctionCall(Module &module, Change* change) {
FunctionChange *funChange = (FunctionChange*)change;
CallInst *oldCallInst = dyn_cast<CallInst>(funChange->getValue());
CallInst *newCallInst = NULL;
string oldFunction = oldCallInst->getCalledFunction()->getName();
string newFunction = funChange->getSwitch();
// TODO: use the types vector to not assume signature
Function *newCallee = module.getFunction(newFunction);
if (newCallee) {
errs() << "Changing function call from " << oldFunction << " to " << newFunction << "\n";
if (oldFunction != newFunction) {
// retrieving original operand
Value *oldOperand = oldCallInst->getOperand(0);
// downcasting operand
Type *fType = Type::getFloatTy(module.getContext());
FPTruncInst *newOperand = new FPTruncInst(oldOperand, fType, "", oldCallInst);
// populating array of operands
vector<Value*> operands;
operands.push_back(newOperand);
ArrayRef<Value*> *arrayRefOperands = new ArrayRef<Value*>(operands);
// creating the new CallInst
newCallInst = CallInst::Create(newCallee, *arrayRefOperands, "newCall", oldCallInst);
// casting result to double
Type *dType = Type::getDoubleTy(module.getContext());
FPExtInst *result = new FPExtInst(newCallInst, dType, "", oldCallInst);
// replacing all uses of call instruction
oldCallInst->replaceAllUsesWith(result);
// deleting old callInst
oldCallInst->eraseFromParent();
errs() << "\tChange was successful\n";
}
else {
errs() << "\tNo change required\n";
}
}
else {
errs() << "\tDid not find function " << newFunction << "\n";
}
return newCallInst;
}
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 OptimizeGEPChecks::runOnFunction(Function &F) {
ABC = &getAnalysis<ArrayBoundsCheckLocal>();
MSCI = &getAnalysis<MSCInfo>();
// Visit all call instructions in the function.
visit(F);
for (size_t i = 0, N = ToRemove.size(); i < N; ++i) {
CallInst *CI = ToRemove[i];
CheckInfoType *Info = MSCI->getCheckInfo(CI->getCalledFunction());
CI->replaceAllUsesWith(CI->getArgOperand(Info->DestPtrArgNo));
CI->eraseFromParent();
++SafeGEPs;
}
bool ChangedAnything = !ToRemove.empty();
ToRemove.clear();
return ChangedAnything;
}
void insertCallToAccessFunctionSequential(Function *F, Function *cF) {
CallInst *I;
BasicBlock *b;
Value::user_iterator i = F->user_begin(), e = F->user_end();
while (i != e) {
if (isa<CallInst>(*i)) {
I = dyn_cast<CallInst>(*i);
b = I->getParent();
BasicBlock::iterator helper(I);
CallInst *ci = dyn_cast<CallInst>(I->clone());
ci->setCalledFunction(cF);
b->getInstList().insertAfter(helper, ci);
i++;
I->replaceAllUsesWith(ci);
insertCallToPAPI(I, ci);
}
}
}
bool InlineSpecials::runOnFunction(Function &F) {
std::set<CallInst*> worklist;
for(inst_iterator i = inst_begin(F), e = inst_end(F); i != e; ++i) {
Instruction *I = &*i;
if(CallInst *C = dyn_cast<CallInst>(I)) {
//get the target of the call, if known
Function *invoked = C->getCalledFunction();
if(invoked != NULL) {
worklist.insert(C);
}
}
}
for(std::set<CallInst*>::iterator itr = worklist.begin();
itr != worklist.end();
itr++)
{
CallInst *C = *itr;
Function *invoked = C->getCalledFunction();
string invokedFuncName = invoked->getName();
map<string,ReplaceFunctionPt>::iterator it = specialMap.find(invokedFuncName);
if(it != specialMap.end()) {
ReplaceFunctionPt func = it->second;
Value *newv = func(F.getParent(), C, this);
C->replaceAllUsesWith(newv);
C->eraseFromParent();
}
}
return false;
}
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;
}
/// 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 being 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 call cannot unwind, don't convert it to an invoke.
if (CI->doesNotThrow())
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 =
InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest,
InvokeArgs.begin(), InvokeArgs.end(),
CI->getName(), BB->getTerminator());
II->setCallingConv(CI->getCallingConv());
II->setAttributes(CI->getAttributes());
// 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.
BranchInst::Create(InvokeDest, UI);
// Delete the unwind instruction!
UI->eraseFromParent();
// 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());
}
//
// 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;
}
bool LowerExcHandlers::runOnFunction(Function &F) {
if (!except_enter_func)
return false; // No EH frames in this module
/* Step 1: EH Depth Numbering */
std::map<llvm::CallInst *, int> EnterDepth;
std::map<llvm::CallInst *, int> LeaveDepth;
std::map<BasicBlock *, int> ExitDepth;
int MaxDepth = 0;
// Compute EH Depth at each basic block using a DFS traversal.
for (df_iterator<BasicBlock *> I = df_begin(&F.getEntryBlock()),
E = df_end(&F.getEntryBlock()); I != E; ++I) {
auto *BB = *I;
int Depth = 0;
/* Here we use the assumption that all incoming edges have the same
* EH depth.
*/
for (auto *Pred : predecessors(BB)) {
auto it = ExitDepth.find(Pred);
if (it != ExitDepth.end()) {
Depth = it->second;
break;
}
}
/* Compute the depth within the basic block */
for (auto &I : *BB) {
auto *CI = dyn_cast<CallInst>(&I);
if (!CI)
continue;
Function *Callee = CI->getCalledFunction();
if (!Callee)
continue;
if (Callee == except_enter_func)
EnterDepth[CI] = Depth++;
else if (Callee == leave_func) {
LeaveDepth[CI] = Depth;
Depth -= cast<ConstantInt>(CI->getArgOperand(0))->getLimitedValue();
}
assert(Depth >= 0);
if (Depth > MaxDepth)
MaxDepth = Depth;
}
/* Remember the depth at the BB boundary */
ExitDepth[BB] = Depth;
}
/* Step 2: EH Frame lowering */
// Allocate stack space for each handler. We allocate these as separate
// allocas so the optimizer can later merge and reaarange them if it wants
// to.
Value *handler_sz = ConstantInt::get(Type::getInt32Ty(F.getContext()),
sizeof(jl_handler_t));
Value *handler_sz64 = ConstantInt::get(Type::getInt64Ty(F.getContext()),
sizeof(jl_handler_t));
Instruction *firstInst = &F.getEntryBlock().front();
std::vector<AllocaInst *> buffs;
for (int i = 0; i < MaxDepth; ++i) {
auto *buff = new AllocaInst(Type::getInt8Ty(F.getContext()),
0,
handler_sz, "", firstInst);
buff->setAlignment(16);
buffs.push_back(buff);
}
// Lower enter funcs
for (auto it : EnterDepth) {
assert(it.second >= 0);
AllocaInst *buff = buffs[it.second];
CallInst *enter = it.first;
auto new_enter = CallInst::Create(jlenter_func, buff, "", enter);
Value *lifetime_args[] = {
handler_sz64,
buff
};
CallInst::Create(lifetime_start, lifetime_args, "", new_enter);
#ifndef _OS_WINDOWS_
// For LLVM 3.3 compatibility
Value *args[] = {buff,
ConstantInt::get(Type::getInt32Ty(F.getContext()), 0)};
auto sj = CallInst::Create(setjmp_func, args, "", enter);
#else
auto sj = CallInst::Create(setjmp_func, buff, "", enter);
#endif
// We need to mark this on the call site as well. See issue #6757
sj->setCanReturnTwice();
if (auto dbg = enter->getMetadata(LLVMContext::MD_dbg)) {
new_enter->setMetadata(LLVMContext::MD_dbg, dbg);
sj->setMetadata(LLVMContext::MD_dbg, dbg);
}
enter->replaceAllUsesWith(sj);
enter->eraseFromParent();
}
// Insert lifetime end intrinsics after every leave.
for (auto it : LeaveDepth) {
int StartDepth = it.second - 1;
int npops = cast<ConstantInt>(it.first->getArgOperand(0))->getLimitedValue();
for (int i = 0; i < npops; ++i) {
assert(StartDepth-i >= 0);
Value *lifetime_args[] = {
handler_sz64,
buffs[StartDepth-i]
};
auto LifetimeEnd = CallInst::Create(lifetime_end, lifetime_args);
LifetimeEnd->insertAfter(it.first);
}
}
return true;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Clone functions that take LoadInsts as arguments
//
// Inputs:
// M - A reference to the LLVM module to transform
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool LoadArgs::runOnModule(Module& M) {
std::map<std::pair<Function*, const Type * > , Function* > fnCache;
bool changed;
do {
changed = false;
for (Module::iterator Func = M.begin(); Func != M.end(); ++Func) {
for (Function::iterator B = Func->begin(), FE = Func->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
CallInst *CI = dyn_cast<CallInst>(I++);
if(!CI)
continue;
if(CI->hasByValArgument())
continue;
// if the CallInst calls a function, that is externally defined,
// or might be changed, ignore this call site.
Function *F = CI->getCalledFunction();
if (!F || (F->isDeclaration() || F->mayBeOverridden()))
continue;
if(F->hasStructRetAttr())
continue;
if(F->isVarArg())
continue;
// find the argument we must replace
Function::arg_iterator ai = F->arg_begin(), ae = F->arg_end();
unsigned argNum = 0;
for(; argNum < CI->getNumArgOperands();argNum++, ++ai) {
// do not care about dead arguments
if(ai->use_empty())
continue;
if(F->getAttributes().getParamAttributes(argNum).hasAttrSomewhere(Attribute::SExt) ||
F->getAttributes().getParamAttributes(argNum).hasAttrSomewhere(Attribute::ZExt))
continue;
if (isa<LoadInst>(CI->getArgOperand(argNum)))
break;
}
// if no argument was a GEP operator to be changed
if(ai == ae)
continue;
LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(argNum));
Instruction * InsertPt = &(Func->getEntryBlock().front());
AllocaInst *NewVal = new AllocaInst(LI->getType(), "",InsertPt);
StoreInst *Copy = new StoreInst(LI, NewVal);
Copy->insertAfter(LI);
/*if(LI->getParent() != CI->getParent())
continue;
// Also check that there is no store after the load.
// TODO: Check if the load/store do not alias.
BasicBlock::iterator bii = LI->getParent()->begin();
Instruction *BII = bii;
while(BII != LI) {
++bii;
BII = bii;
}
while(BII != CI) {
if(isa<StoreInst>(BII))
break;
++bii;
BII = bii;
}
if(isa<StoreInst>(bii)){
continue;
}*/
// Construct the new Type
// Appends the struct Type at the beginning
std::vector<Type*>TP;
for(unsigned c = 0; c < CI->getNumArgOperands();c++) {
if(c == argNum)
TP.push_back(LI->getPointerOperand()->getType());
TP.push_back(CI->getArgOperand(c)->getType());
}
//return type is same as that of original instruction
FunctionType *NewFTy = FunctionType::get(CI->getType(), TP, false);
numSimplified++;
//if(numSimplified > 1000)
//return true;
Function *NewF;
std::map<std::pair<Function*, const Type* > , Function* >::iterator Test;
Test = fnCache.find(std::make_pair(F, NewFTy));
if(Test != fnCache.end()) {
NewF = Test->second;
} else {
NewF = Function::Create(NewFTy,
GlobalValue::InternalLinkage,
F->getName().str() + ".TEST",
&M);
fnCache[std::make_pair(F, NewFTy)] = NewF;
Function::arg_iterator NI = NewF->arg_begin();
ValueToValueMapTy ValueMap;
unsigned count = 0;
for (Function::arg_iterator II = F->arg_begin(); NI != NewF->arg_end(); ++count, ++NI) {
if(count == argNum) {
NI->setName("LDarg");
continue;
}
ValueMap[II] = NI;
NI->setName(II->getName());
NI->addAttr(F->getAttributes().getParamAttributes(II->getArgNo() + 1));
++II;
}
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
CloneFunctionInto(NewF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NewF->arg_begin(),
ae= NewF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NewF->setAttributes(NewF->getAttributes().addAttributes(
F->getContext(), 0, F->getAttributes().getRetAttributes()));
NewF->setAttributes(NewF->getAttributes().addAttributes(
F->getContext(), ~0, F->getAttributes().getFnAttributes()));
//Get the point to insert the GEP instr.
Instruction *InsertPoint;
for (BasicBlock::iterator insrt = NewF->front().begin(); isa<AllocaInst>(InsertPoint = insrt); ++insrt) {;}
LoadInst *LI_new = new LoadInst(fargs.at(argNum), "", InsertPoint);
fargs.at(argNum+1)->replaceAllUsesWith(LI_new);
}
//this does not seem to be a good idea
AttributeSet NewCallPAL=AttributeSet();
// Get the initial attributes of the call
AttributeSet CallPAL = CI->getAttributes();
AttributeSet RAttrs = CallPAL.getRetAttributes();
AttributeSet FnAttrs = CallPAL.getFnAttributes();
if (!RAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),0, RAttrs);
SmallVector<Value*, 8> Args;
for(unsigned j =0;j<CI->getNumArgOperands();j++) {
if(j == argNum) {
Args.push_back(NewVal);
}
Args.push_back(CI->getArgOperand(j));
// position in the NewCallPAL
AttributeSet Attrs = CallPAL.getParamAttributes(j+1);
if (!Attrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),Args.size(), Attrs);
}
// Create the new attributes vec.
if (!FnAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),~0, FnAttrs);
CallInst *CallI = CallInst::Create(NewF,Args,"", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
CI->replaceAllUsesWith(CallI);
CI->eraseFromParent();
changed = true;
}
}
}
} while(changed);
return true;
}
bool ObjCARCContract::tryToPeepholeInstruction(
Function &F, Instruction *Inst, inst_iterator &Iter,
SmallPtrSetImpl<Instruction *> &DependingInsts,
SmallPtrSetImpl<const BasicBlock *> &Visited,
bool &TailOkForStoreStrongs) {
// Only these library routines return their argument. In particular,
// objc_retainBlock does not necessarily return its argument.
ARCInstKind Class = GetBasicARCInstKind(Inst);
switch (Class) {
case ARCInstKind::FusedRetainAutorelease:
case ARCInstKind::FusedRetainAutoreleaseRV:
return false;
case ARCInstKind::Autorelease:
case ARCInstKind::AutoreleaseRV:
return contractAutorelease(F, Inst, Class, DependingInsts, Visited);
case ARCInstKind::Retain:
// Attempt to convert retains to retainrvs if they are next to function
// calls.
if (!optimizeRetainCall(F, Inst))
return false;
// If we succeed in our optimization, fall through.
// FALLTHROUGH
case ARCInstKind::RetainRV: {
// If we're compiling for a target which needs a special inline-asm
// marker to do the retainAutoreleasedReturnValue optimization,
// insert it now.
if (!RetainRVMarker)
return false;
BasicBlock::iterator BBI = Inst;
BasicBlock *InstParent = Inst->getParent();
// Step up to see if the call immediately precedes the RetainRV call.
// If it's an invoke, we have to cross a block boundary. And we have
// to carefully dodge no-op instructions.
do {
if (&*BBI == InstParent->begin()) {
BasicBlock *Pred = InstParent->getSinglePredecessor();
if (!Pred)
goto decline_rv_optimization;
BBI = Pred->getTerminator();
break;
}
--BBI;
} while (IsNoopInstruction(BBI));
if (&*BBI == GetArgRCIdentityRoot(Inst)) {
DEBUG(dbgs() << "Adding inline asm marker for "
"retainAutoreleasedReturnValue optimization.\n");
Changed = true;
InlineAsm *IA =
InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
/*isVarArg=*/false),
RetainRVMarker->getString(),
/*Constraints=*/"", /*hasSideEffects=*/true);
CallInst::Create(IA, "", Inst);
}
decline_rv_optimization:
return false;
}
case ARCInstKind::InitWeak: {
// objc_initWeak(p, null) => *p = null
CallInst *CI = cast<CallInst>(Inst);
if (IsNullOrUndef(CI->getArgOperand(1))) {
Value *Null =
ConstantPointerNull::get(cast<PointerType>(CI->getType()));
Changed = true;
new StoreInst(Null, CI->getArgOperand(0), CI);
DEBUG(dbgs() << "OBJCARCContract: Old = " << *CI << "\n"
<< " New = " << *Null << "\n");
CI->replaceAllUsesWith(Null);
CI->eraseFromParent();
}
return true;
}
case ARCInstKind::Release:
// Try to form an objc store strong from our release. If we fail, there is
// nothing further to do below, so continue.
tryToContractReleaseIntoStoreStrong(Inst, Iter);
return true;
case ARCInstKind::User:
// Be conservative if the function has any alloca instructions.
// Technically we only care about escaping alloca instructions,
// but this is sufficient to handle some interesting cases.
if (isa<AllocaInst>(Inst))
TailOkForStoreStrongs = false;
return true;
case ARCInstKind::IntrinsicUser:
// Remove calls to @clang.arc.use(...).
Inst->eraseFromParent();
return true;
default:
return true;
}
}
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 ObjCARCContract::runOnFunction(Function &F) {
if (!EnableARCOpts)
return false;
// If nothing in the Module uses ARC, don't do anything.
if (!Run)
return false;
Changed = false;
AA = &getAnalysis<AliasAnalysis>();
DT = &getAnalysis<DominatorTree>();
PA.setAA(&getAnalysis<AliasAnalysis>());
// Track whether it's ok to mark objc_storeStrong calls with the "tail"
// keyword. Be conservative if the function has variadic arguments.
// It seems that functions which "return twice" are also unsafe for the
// "tail" argument, because they are setjmp, which could need to
// return to an earlier stack state.
bool TailOkForStoreStrongs = !F.isVarArg() &&
!F.callsFunctionThatReturnsTwice();
// For ObjC library calls which return their argument, replace uses of the
// argument with uses of the call return value, if it dominates the use. This
// reduces register pressure.
SmallPtrSet<Instruction *, 4> DependingInstructions;
SmallPtrSet<const BasicBlock *, 4> Visited;
for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
Instruction *Inst = &*I++;
DEBUG(dbgs() << "ObjCARCContract: Visiting: " << *Inst << "\n");
// Only these library routines return their argument. In particular,
// objc_retainBlock does not necessarily return its argument.
InstructionClass Class = GetBasicInstructionClass(Inst);
switch (Class) {
case IC_Retain:
case IC_FusedRetainAutorelease:
case IC_FusedRetainAutoreleaseRV:
break;
case IC_Autorelease:
case IC_AutoreleaseRV:
if (ContractAutorelease(F, Inst, Class, DependingInstructions, Visited))
continue;
break;
case IC_RetainRV: {
// If we're compiling for a target which needs a special inline-asm
// marker to do the retainAutoreleasedReturnValue optimization,
// insert it now.
if (!RetainRVMarker)
break;
BasicBlock::iterator BBI = Inst;
BasicBlock *InstParent = Inst->getParent();
// Step up to see if the call immediately precedes the RetainRV call.
// If it's an invoke, we have to cross a block boundary. And we have
// to carefully dodge no-op instructions.
do {
if (&*BBI == InstParent->begin()) {
BasicBlock *Pred = InstParent->getSinglePredecessor();
if (!Pred)
goto decline_rv_optimization;
BBI = Pred->getTerminator();
break;
}
--BBI;
} while (IsNoopInstruction(BBI));
if (&*BBI == GetObjCArg(Inst)) {
DEBUG(dbgs() << "ObjCARCContract: Adding inline asm marker for "
"retainAutoreleasedReturnValue optimization.\n");
Changed = true;
InlineAsm *IA =
InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
/*isVarArg=*/false),
RetainRVMarker->getString(),
/*Constraints=*/"", /*hasSideEffects=*/true);
CallInst::Create(IA, "", Inst);
}
decline_rv_optimization:
break;
}
case IC_InitWeak: {
// objc_initWeak(p, null) => *p = null
CallInst *CI = cast<CallInst>(Inst);
if (IsNullOrUndef(CI->getArgOperand(1))) {
Value *Null =
ConstantPointerNull::get(cast<PointerType>(CI->getType()));
Changed = true;
new StoreInst(Null, CI->getArgOperand(0), CI);
DEBUG(dbgs() << "OBJCARCContract: Old = " << *CI << "\n"
<< " New = " << *Null << "\n");
CI->replaceAllUsesWith(Null);
CI->eraseFromParent();
}
continue;
}
case IC_Release:
ContractRelease(Inst, I);
continue;
case IC_User:
// Be conservative if the function has any alloca instructions.
// Technically we only care about escaping alloca instructions,
// but this is sufficient to handle some interesting cases.
if (isa<AllocaInst>(Inst))
TailOkForStoreStrongs = false;
continue;
case IC_IntrinsicUser:
// Remove calls to @clang.arc.use(...).
Inst->eraseFromParent();
continue;
default:
continue;
}
DEBUG(dbgs() << "ObjCARCContract: Finished List.\n\n");
// Don't use GetObjCArg because we don't want to look through bitcasts
// and such; to do the replacement, the argument must have type i8*.
const Value *Arg = cast<CallInst>(Inst)->getArgOperand(0);
for (;;) {
// If we're compiling bugpointed code, don't get in trouble.
if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))
break;
// Look through the uses of the pointer.
for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
UI != UE; ) {
Use &U = UI.getUse();
unsigned OperandNo = UI.getOperandNo();
++UI; // Increment UI now, because we may unlink its element.
// If the call's return value dominates a use of the call's argument
// value, rewrite the use to use the return value. We check for
// reachability here because an unreachable call is considered to
// trivially dominate itself, which would lead us to rewriting its
// argument in terms of its return value, which would lead to
// infinite loops in GetObjCArg.
if (DT->isReachableFromEntry(U) && DT->dominates(Inst, U)) {
Changed = true;
Instruction *Replacement = Inst;
Type *UseTy = U.get()->getType();
if (PHINode *PHI = dyn_cast<PHINode>(U.getUser())) {
// For PHI nodes, insert the bitcast in the predecessor block.
unsigned ValNo = PHINode::getIncomingValueNumForOperand(OperandNo);
BasicBlock *BB = PHI->getIncomingBlock(ValNo);
if (Replacement->getType() != UseTy)
Replacement = new BitCastInst(Replacement, UseTy, "",
&BB->back());
// While we're here, rewrite all edges for this PHI, rather
// than just one use at a time, to minimize the number of
// bitcasts we emit.
for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
if (PHI->getIncomingBlock(i) == BB) {
// Keep the UI iterator valid.
if (&PHI->getOperandUse(
PHINode::getOperandNumForIncomingValue(i)) ==
&UI.getUse())
++UI;
PHI->setIncomingValue(i, Replacement);
}
} else {
if (Replacement->getType() != UseTy)
Replacement = new BitCastInst(Replacement, UseTy, "",
cast<Instruction>(U.getUser()));
U.set(Replacement);
}
}
}
// If Arg is a no-op casted pointer, strip one level of casts and iterate.
if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg))
Arg = BI->getOperand(0);
else if (isa<GEPOperator>(Arg) &&
cast<GEPOperator>(Arg)->hasAllZeroIndices())
Arg = cast<GEPOperator>(Arg)->getPointerOperand();
else if (isa<GlobalAlias>(Arg) &&
!cast<GlobalAlias>(Arg)->mayBeOverridden())
Arg = cast<GlobalAlias>(Arg)->getAliasee();
else
break;
}
}
// If this function has no escaping allocas or suspicious vararg usage,
// objc_storeStrong calls can be marked with the "tail" keyword.
if (TailOkForStoreStrongs)
for (SmallPtrSet<CallInst *, 8>::iterator I = StoreStrongCalls.begin(),
E = StoreStrongCalls.end(); I != E; ++I)
(*I)->setTailCall();
StoreStrongCalls.clear();
return Changed;
}
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;
}
bool SjLjEHPass::insertSjLjEHSupport(Function &F) {
SmallVector<ReturnInst*,16> Returns;
SmallVector<UnwindInst*,16> Unwinds;
SmallVector<InvokeInst*,16> Invokes;
// Look through the terminators of the basic blocks to find invokes, returns
// and unwinds.
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
// Remember all return instructions in case we insert an invoke into this
// function.
Returns.push_back(RI);
} else if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator())) {
Invokes.push_back(II);
} else if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
Unwinds.push_back(UI);
}
}
NumInvokes += Invokes.size();
NumUnwinds += Unwinds.size();
// If we don't have any invokes, there's nothing to do.
if (Invokes.empty()) return false;
// Find the eh.selector.*, eh.exception and alloca calls.
//
// Remember any allocas() that aren't in the entry block, as the
// jmpbuf saved SP will need to be updated for them.
//
// We'll use the first eh.selector to determine the right personality
// function to use. For SJLJ, we always use the same personality for the
// whole function, not on a per-selector basis.
// FIXME: That's a bit ugly. Better way?
SmallVector<CallInst*,16> EH_Selectors;
SmallVector<CallInst*,16> EH_Exceptions;
SmallVector<Instruction*,16> JmpbufUpdatePoints;
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
// Note: Skip the entry block since there's nothing there that interests
// us. eh.selector and eh.exception shouldn't ever be there, and we
// want to disregard any allocas that are there.
//
// FIXME: This is awkward. The new EH scheme won't need to skip the entry
// block.
if (BB == F.begin()) {
if (InvokeInst *II = dyn_cast<InvokeInst>(F.begin()->getTerminator())) {
// FIXME: This will be always non-NULL in the new EH.
if (LandingPadInst *LPI = II->getUnwindDest()->getLandingPadInst())
if (!PersonalityFn) PersonalityFn = LPI->getPersonalityFn();
}
continue;
}
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
if (CallInst *CI = dyn_cast<CallInst>(I)) {
if (CI->getCalledFunction() == SelectorFn) {
if (!PersonalityFn) PersonalityFn = CI->getArgOperand(1);
EH_Selectors.push_back(CI);
} else if (CI->getCalledFunction() == ExceptionFn) {
EH_Exceptions.push_back(CI);
} else if (CI->getCalledFunction() == StackRestoreFn) {
JmpbufUpdatePoints.push_back(CI);
}
} else if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
JmpbufUpdatePoints.push_back(AI);
} else if (InvokeInst *II = dyn_cast<InvokeInst>(I)) {
// FIXME: This will be always non-NULL in the new EH.
if (LandingPadInst *LPI = II->getUnwindDest()->getLandingPadInst())
if (!PersonalityFn) PersonalityFn = LPI->getPersonalityFn();
}
}
}
// If we don't have any eh.selector calls, we can't determine the personality
// function. Without a personality function, we can't process exceptions.
if (!PersonalityFn) return false;
// We have invokes, so we need to add register/unregister calls to get this
// function onto the global unwind stack.
//
// First thing we need to do is scan the whole function for values that are
// live across unwind edges. Each value that is live across an unwind edge we
// spill into a stack location, guaranteeing that there is nothing live across
// the unwind edge. This process also splits all critical edges coming out of
// invoke's.
splitLiveRangesAcrossInvokes(Invokes);
SmallVector<LandingPadInst*, 16> LandingPads;
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator()))
// FIXME: This will be always non-NULL in the new EH.
if (LandingPadInst *LPI = II->getUnwindDest()->getLandingPadInst())
LandingPads.push_back(LPI);
}
BasicBlock *EntryBB = F.begin();
// Create an alloca for the incoming jump buffer ptr and the new jump buffer
// that needs to be restored on all exits from the function. This is an
// alloca because the value needs to be added to the global context list.
unsigned Align = 4; // FIXME: Should be a TLI check?
AllocaInst *FunctionContext =
new AllocaInst(FunctionContextTy, 0, Align,
"fcn_context", F.begin()->begin());
Value *Idxs[2];
Type *Int32Ty = Type::getInt32Ty(F.getContext());
Value *Zero = ConstantInt::get(Int32Ty, 0);
// We need to also keep around a reference to the call_site field
Idxs[0] = Zero;
Idxs[1] = ConstantInt::get(Int32Ty, 1);
CallSite = GetElementPtrInst::Create(FunctionContext, Idxs, "call_site",
EntryBB->getTerminator());
// The exception selector comes back in context->data[1]
Idxs[1] = ConstantInt::get(Int32Ty, 2);
Value *FCData = GetElementPtrInst::Create(FunctionContext, Idxs, "fc_data",
EntryBB->getTerminator());
Idxs[1] = ConstantInt::get(Int32Ty, 1);
Value *SelectorAddr = GetElementPtrInst::Create(FCData, Idxs,
"exc_selector_gep",
EntryBB->getTerminator());
// The exception value comes back in context->data[0]
Idxs[1] = Zero;
Value *ExceptionAddr = GetElementPtrInst::Create(FCData, Idxs,
"exception_gep",
EntryBB->getTerminator());
// The result of the eh.selector call will be replaced with a a reference to
// the selector value returned in the function context. We leave the selector
// itself so the EH analysis later can use it.
for (int i = 0, e = EH_Selectors.size(); i < e; ++i) {
CallInst *I = EH_Selectors[i];
Value *SelectorVal = new LoadInst(SelectorAddr, "select_val", true, I);
I->replaceAllUsesWith(SelectorVal);
}
// eh.exception calls are replaced with references to the proper location in
// the context. Unlike eh.selector, the eh.exception calls are removed
// entirely.
for (int i = 0, e = EH_Exceptions.size(); i < e; ++i) {
CallInst *I = EH_Exceptions[i];
// Possible for there to be duplicates, so check to make sure the
// instruction hasn't already been removed.
if (!I->getParent()) continue;
Value *Val = new LoadInst(ExceptionAddr, "exception", true, I);
Type *Ty = Type::getInt8PtrTy(F.getContext());
Val = CastInst::Create(Instruction::IntToPtr, Val, Ty, "", I);
I->replaceAllUsesWith(Val);
I->eraseFromParent();
}
for (unsigned i = 0, e = LandingPads.size(); i != e; ++i)
ReplaceLandingPadVal(F, LandingPads[i], ExceptionAddr, SelectorAddr);
// The entry block changes to have the eh.sjlj.setjmp, with a conditional
// branch to a dispatch block for non-zero returns. If we return normally,
// we're not handling an exception and just register the function context and
// continue.
// Create the dispatch block. The dispatch block is basically a big switch
// statement that goes to all of the invoke landing pads.
BasicBlock *DispatchBlock =
BasicBlock::Create(F.getContext(), "eh.sjlj.setjmp.catch", &F);
// Insert a load of the callsite in the dispatch block, and a switch on its
// value. By default, we issue a trap statement.
BasicBlock *TrapBlock =
BasicBlock::Create(F.getContext(), "trapbb", &F);
CallInst::Create(Intrinsic::getDeclaration(F.getParent(), Intrinsic::trap),
"", TrapBlock);
new UnreachableInst(F.getContext(), TrapBlock);
Value *DispatchLoad = new LoadInst(CallSite, "invoke.num", true,
DispatchBlock);
SwitchInst *DispatchSwitch =
SwitchInst::Create(DispatchLoad, TrapBlock, Invokes.size(),
DispatchBlock);
// Split the entry block to insert the conditional branch for the setjmp.
BasicBlock *ContBlock = EntryBB->splitBasicBlock(EntryBB->getTerminator(),
"eh.sjlj.setjmp.cont");
// Populate the Function Context
// 1. LSDA address
// 2. Personality function address
// 3. jmpbuf (save SP, FP and call eh.sjlj.setjmp)
// LSDA address
Idxs[0] = Zero;
Idxs[1] = ConstantInt::get(Int32Ty, 4);
Value *LSDAFieldPtr =
GetElementPtrInst::Create(FunctionContext, Idxs, "lsda_gep",
EntryBB->getTerminator());
Value *LSDA = CallInst::Create(LSDAAddrFn, "lsda_addr",
EntryBB->getTerminator());
new StoreInst(LSDA, LSDAFieldPtr, true, EntryBB->getTerminator());
Idxs[1] = ConstantInt::get(Int32Ty, 3);
Value *PersonalityFieldPtr =
GetElementPtrInst::Create(FunctionContext, Idxs, "lsda_gep",
EntryBB->getTerminator());
new StoreInst(PersonalityFn, PersonalityFieldPtr, true,
EntryBB->getTerminator());
// Save the frame pointer.
Idxs[1] = ConstantInt::get(Int32Ty, 5);
Value *JBufPtr
= GetElementPtrInst::Create(FunctionContext, Idxs, "jbuf_gep",
EntryBB->getTerminator());
Idxs[1] = ConstantInt::get(Int32Ty, 0);
Value *FramePtr =
GetElementPtrInst::Create(JBufPtr, Idxs, "jbuf_fp_gep",
EntryBB->getTerminator());
Value *Val = CallInst::Create(FrameAddrFn,
ConstantInt::get(Int32Ty, 0),
"fp",
EntryBB->getTerminator());
new StoreInst(Val, FramePtr, true, EntryBB->getTerminator());
// Save the stack pointer.
Idxs[1] = ConstantInt::get(Int32Ty, 2);
Value *StackPtr =
GetElementPtrInst::Create(JBufPtr, Idxs, "jbuf_sp_gep",
EntryBB->getTerminator());
Val = CallInst::Create(StackAddrFn, "sp", EntryBB->getTerminator());
new StoreInst(Val, StackPtr, true, EntryBB->getTerminator());
// Call the setjmp instrinsic. It fills in the rest of the jmpbuf.
Value *SetjmpArg =
CastInst::Create(Instruction::BitCast, JBufPtr,
Type::getInt8PtrTy(F.getContext()), "",
EntryBB->getTerminator());
Value *DispatchVal = CallInst::Create(BuiltinSetjmpFn, SetjmpArg,
"dispatch",
EntryBB->getTerminator());
// Add a call to dispatch_setup after the setjmp call. This is expanded to any
// target-specific setup that needs to be done.
CallInst::Create(DispatchSetupFn, DispatchVal, "", EntryBB->getTerminator());
// check the return value of the setjmp. non-zero goes to dispatcher.
Value *IsNormal = new ICmpInst(EntryBB->getTerminator(),
ICmpInst::ICMP_EQ, DispatchVal, Zero,
"notunwind");
// Nuke the uncond branch.
EntryBB->getTerminator()->eraseFromParent();
// Put in a new condbranch in its place.
BranchInst::Create(ContBlock, DispatchBlock, IsNormal, EntryBB);
// Register the function context and make sure it's known to not throw
CallInst *Register =
CallInst::Create(RegisterFn, FunctionContext, "",
ContBlock->getTerminator());
Register->setDoesNotThrow();
// At this point, we are all set up, update the invoke instructions to mark
// their call_site values, and fill in the dispatch switch accordingly.
for (unsigned i = 0, e = Invokes.size(); i != e; ++i)
markInvokeCallSite(Invokes[i], i+1, CallSite, DispatchSwitch);
// Mark call instructions that aren't nounwind as no-action (call_site ==
// -1). Skip the entry block, as prior to then, no function context has been
// created for this function and any unexpected exceptions thrown will go
// directly to the caller's context, which is what we want anyway, so no need
// to do anything here.
for (Function::iterator BB = F.begin(), E = F.end(); ++BB != E;) {
for (BasicBlock::iterator I = BB->begin(), end = BB->end(); I != end; ++I)
if (CallInst *CI = dyn_cast<CallInst>(I)) {
// Ignore calls to the EH builtins (eh.selector, eh.exception)
Constant *Callee = CI->getCalledFunction();
if (Callee != SelectorFn && Callee != ExceptionFn
&& !CI->doesNotThrow())
insertCallSiteStore(CI, -1, CallSite);
} else if (ResumeInst *RI = dyn_cast<ResumeInst>(I)) {
insertCallSiteStore(RI, -1, CallSite);
}
}
// Replace all unwinds with a branch to the unwind handler.
// ??? Should this ever happen with sjlj exceptions?
for (unsigned i = 0, e = Unwinds.size(); i != e; ++i) {
BranchInst::Create(TrapBlock, Unwinds[i]);
Unwinds[i]->eraseFromParent();
}
// Following any allocas not in the entry block, update the saved SP in the
// jmpbuf to the new value.
for (unsigned i = 0, e = JmpbufUpdatePoints.size(); i != e; ++i) {
Instruction *AI = JmpbufUpdatePoints[i];
Instruction *StackAddr = CallInst::Create(StackAddrFn, "sp");
StackAddr->insertAfter(AI);
Instruction *StoreStackAddr = new StoreInst(StackAddr, StackPtr, true);
StoreStackAddr->insertAfter(StackAddr);
}
// Finally, for any returns from this function, if this function contains an
// invoke, add a call to unregister the function context.
for (unsigned i = 0, e = Returns.size(); i != e; ++i)
CallInst::Create(UnregisterFn, FunctionContext, "", Returns[i]);
return true;
}
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;
}
/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
/// an invoke, we have to turn all of the calls that can throw into
/// invokes. This function analyze BB to see if there are any calls, and if so,
/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
/// nodes in that block with the values specified in InvokeDestPHIValues.
///
/// Returns true to indicate that the next block should be skipped.
static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
InvokeInliningInfo &Invoke) {
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.
CallInst *CI = dyn_cast<CallInst>(I);
if (CI == 0) continue;
// LIBUNWIND: merge selector instructions.
if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) {
EHSelectorInst *Outer = Invoke.getOuterSelector();
if (!Outer) continue;
bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner);
bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer);
// If both selectors contain only cleanups, we don't need to do
// anything. TODO: this is really just a very specific instance
// of a much more general optimization.
if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue;
// Otherwise, we just append the outer selector to the inner selector.
SmallVector<Value*, 16> NewSelector;
for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i)
NewSelector.push_back(Inner->getArgOperand(i));
for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i)
NewSelector.push_back(Outer->getArgOperand(i));
CallInst *NewInner =
IRBuilder<>(Inner).CreateCall(Inner->getCalledValue(), NewSelector);
// No need to copy attributes, calling convention, etc.
NewInner->takeName(Inner);
Inner->replaceAllUsesWith(NewInner);
Inner->eraseFromParent();
continue;
}
// If this call cannot unwind, don't convert it to an invoke.
if (CI->doesNotThrow())
continue;
// Convert this function call into an invoke instruction.
// First, split the basic block.
BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
// Delete the unconditional branch inserted by splitBasicBlock
BB->getInstList().pop_back();
// LIBUNWIND: If this is a call to @llvm.eh.resume, just branch
// directly to the new landing pad.
if (Invoke.forwardEHResume(CI, BB)) {
// TODO: 'Split' is now unreachable; clean it up.
// We want to leave the original call intact so that the call
// graph and other structures won't get misled. We also have to
// avoid processing the next block, or we'll iterate here forever.
return true;
}
// Otherwise, create the new invoke instruction.
ImmutableCallSite CS(CI);
SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
InvokeInst *II =
InvokeInst::Create(CI->getCalledValue(), Split,
Invoke.getOuterUnwindDest(),
InvokeArgs, CI->getName(), BB);
II->setCallingConv(CI->getCallingConv());
II->setAttributes(CI->getAttributes());
// Make sure that anything using the call now uses the invoke! This also
// updates the CallGraph if present, because it uses a WeakVH.
CI->replaceAllUsesWith(II);
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.
Invoke.addIncomingPHIValuesFor(BB);
return false;
}
return false;
}
//
// 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 NVVMReflect::runOnModule(Module &M) {
if (!NVVMReflectEnabled)
return false;
setVarMap();
ReflectFunction = M.getFunction(NVVM_REFLECT_FUNCTION);
// If reflect function is not used, then there will be
// no entry in the module.
if (ReflectFunction == 0)
return false;
// Validate _reflect function
assert(ReflectFunction->isDeclaration() &&
"_reflect function should not have a body");
assert(ReflectFunction->getReturnType()->isIntegerTy() &&
"_reflect's return type should be integer");
std::vector<Instruction *> ToRemove;
// Go through the uses of ReflectFunction in this Function.
// Each of them should 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.
for (User *U : ReflectFunction->users()) {
assert(isa<CallInst>(U) && "Only a call instruction can use _reflect");
CallInst *Reflect = cast<CallInst>(U);
assert((Reflect->getNumOperands() == 2) &&
"Only one operand expect for _reflect function");
// In cuda, we will have an extra constant-to-generic conversion of
// the string.
const Value *conv = Reflect->getArgOperand(0);
assert(isa<CallInst>(conv) && "Expected a const-to-gen conversion");
const CallInst *ConvCall = cast<CallInst>(conv);
const Value *str = ConvCall->getArgOperand(0);
assert(isa<ConstantExpr>(str) &&
"Format of _reflect function not recognized");
const ConstantExpr *GEP = cast<ConstantExpr>(str);
const Value *Sym = GEP->getOperand(0);
assert(isa<Constant>(Sym) && "Format of _reflect function not recognized");
const Constant *SymStr = cast<Constant>(Sym);
assert(isa<ConstantDataSequential>(SymStr->getOperand(0)) &&
"Format of _reflect function not recognized");
assert(cast<ConstantDataSequential>(SymStr->getOperand(0))->isCString() &&
"Format of _reflect function not recognized");
std::string ReflectArg =
cast<ConstantDataSequential>(SymStr->getOperand(0))->getAsString();
ReflectArg = ReflectArg.substr(0, ReflectArg.size() - 1);
DEBUG(dbgs() << "Arg of _reflect : " << ReflectArg << "\n");
int ReflectVal = 0; // The default value is 0
if (VarMap.find(ReflectArg) != VarMap.end()) {
ReflectVal = VarMap[ReflectArg];
}
Reflect->replaceAllUsesWith(
ConstantInt::get(Reflect->getType(), ReflectVal));
ToRemove.push_back(Reflect);
}
if (ToRemove.size() == 0)
return false;
for (unsigned i = 0, e = ToRemove.size(); i != e; ++i)
ToRemove[i]->eraseFromParent();
return true;
}
// Replace "packW" and "unpackW" intrinsics by insert/extract operations and
// update the uses accordingly.
void
FunctionVectorizer::generatePackUnpackCode(Function* f,
const WFVInfo& info)
{
assert (f);
SmallVector<CallInst*, 16> eraseVec;
for (auto &BB : *f)
{
Instruction* allocPos = BB.getFirstInsertionPt();
for (auto &I : BB)
{
Instruction* inst = &I;
if (isUnpackWFunctionCall(inst))
{
DEBUG_WFV( outs() << "generateUnpackCode(" << *inst << " )\n"; );
CallInst* unpackCall = cast<CallInst>(inst);
Value* value = unpackCall->getArgOperand(0);
Value* indexVal = unpackCall->getArgOperand(1);
// Extract scalar values.
Value* extract = generateHorizontalExtract(value,
indexVal,
unpackCall->getName(),
allocPos,
unpackCall,
info);
// If the type only matches structurally, create an additional bitcast.
Type* oldType = unpackCall->getType();
Type* newType = extract->getType();
if (oldType != newType)
{
assert (newType->canLosslesslyBitCastTo(oldType) || WFV::typesMatch(oldType, newType));
Instruction* bc = new BitCastInst(extract, oldType, "", unpackCall);
// Copy properties from unpackCall.
WFV::copyMetadata(bc, *unpackCall);
extract = bc;
}
// Rewire the use.
assert (unpackCall->getNumUses() == 1);
Value* use = *unpackCall->use_begin();
assert (isa<Instruction>(use));
Instruction* scalarUse = cast<Instruction>(use);
scalarUse->replaceUsesOfWith(unpackCall, extract);
// Erase now unused unpack call.
eraseVec.push_back(unpackCall);
// If the returned extract operation is an alloca, we have to
// make sure that all changes to that memory location are
// correctly written back to the original memory from which
// the sub-element was extracted.
// This means we have to insert merge and store operations
// after every use of this value (including "forwarded" uses
// via casts, phis, and GEPs).
// However, we must only merge back those values that were
// modified. This is not only for efficiency, but also for
// correctness, since there may be uninitialized pointers in
// a structure, which we must not load/store from/to (see
// test_struct_extra05 with all analyses disabled).
if (isa<AllocaInst>(extract) ||
(isa<BitCastInst>(extract) &&
isa<AllocaInst>(cast<BitCastInst>(extract)->getOperand(0))))
{
generateWriteBackOperations(cast<Instruction>(extract),
cast<Instruction>(extract),
value,
indexVal,
info);
}
}
else if (isPackWFunctionCall(inst))
{
DEBUG_WFV( outs() << "generatePackCode(" << *inst << " )\n"; );
CallInst* packCall = cast<CallInst>(inst);
assert (WFV::isVectorizedType(*packCall->getType()) &&
"packCall should have vector return type after inst vectorization!");
SmallVector<Value*, 8> scalarVals(info.mVectorizationFactor);
// Get scalar results for merge.
for (unsigned i=0; i<info.mVectorizationFactor; ++i)
{
scalarVals[i] = packCall->getArgOperand(i);
}
// Merge scalar results.
Instruction* merge = generateHorizontalMerge(scalarVals,
packCall->getType(),
"",
packCall,
info);
// Rewire the uses.
packCall->replaceAllUsesWith(merge);
// Copy properties from packCall.
WFV::copyMetadata(merge, *packCall);
// Erase now unused pack call.
eraseVec.push_back(packCall);
}
bool NVVMReflect::handleFunction(Function *ReflectFunction) {
// Validate _reflect function
assert(ReflectFunction->isDeclaration() &&
"_reflect function should not have a body");
assert(ReflectFunction->getReturnType()->isIntegerTy() &&
"_reflect's return type should be integer");
std::vector<Instruction *> ToRemove;
// Go through the uses of ReflectFunction in this Function.
// Each of them should 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.
// 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)
//
// 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 (User *U : ReflectFunction->users()) {
assert(isa<CallInst>(U) && "Only a call instruction can use _reflect");
CallInst *Reflect = cast<CallInst>(U);
assert((Reflect->getNumOperands() == 2) &&
"Only one operand expect for _reflect function");
// In cuda, we will have an extra constant-to-generic conversion of
// the string.
const Value *Str = Reflect->getArgOperand(0);
if (isa<CallInst>(Str)) {
// CUDA path
const CallInst *ConvCall = cast<CallInst>(Str);
Str = ConvCall->getArgOperand(0);
}
assert(isa<ConstantExpr>(Str) &&
"Format of _reflect function not recognized");
const ConstantExpr *GEP = cast<ConstantExpr>(Str);
const Value *Sym = GEP->getOperand(0);
assert(isa<Constant>(Sym) && "Format of _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 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");
std::string ReflectArg =
cast<ConstantDataSequential>(Operand)->getAsString();
ReflectArg = ReflectArg.substr(0, ReflectArg.size() - 1);
DEBUG(dbgs() << "Arg of _reflect : " << ReflectArg << "\n");
int ReflectVal = 0; // The default value is 0
if (VarMap.find(ReflectArg) != VarMap.end()) {
ReflectVal = VarMap[ReflectArg];
}
Reflect->replaceAllUsesWith(
ConstantInt::get(Reflect->getType(), ReflectVal));
ToRemove.push_back(Reflect);
}
if (ToRemove.size() == 0)
return false;
for (unsigned i = 0, e = ToRemove.size(); i != e; ++i)
ToRemove[i]->eraseFromParent();
return true;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Search for all call sites to casted functions.
// Check if they only differ in an argument type
// Cast the argument, and call the original function
//
// 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 ArgCast::runOnModule(Module& M) {
std::vector<CallInst*> worklist;
for (Module::iterator I = M.begin(); I != M.end(); ++I) {
if (I->mayBeOverridden())
continue;
// Find all uses of this function
for(Value::user_iterator ui = I->user_begin(), ue = I->user_end(); ui != ue; ) {
// check if is ever casted to a different function type
ConstantExpr *CE = dyn_cast<ConstantExpr>(*ui++);
if(!CE)
continue;
if (CE->getOpcode() != Instruction::BitCast)
continue;
if(CE->getOperand(0) != I)
continue;
const PointerType *PTy = dyn_cast<PointerType>(CE->getType());
if (!PTy)
continue;
const Type *ETy = PTy->getElementType();
const FunctionType *FTy = dyn_cast<FunctionType>(ETy);
if(!FTy)
continue;
// casting to a varargs funtion
// or function with same number of arguments
// possibly varying types of arguments
if(FTy->getNumParams() != I->arg_size() && !FTy->isVarArg())
continue;
for(Value::user_iterator uii = CE->user_begin(),
uee = CE->user_end(); uii != uee; ++uii) {
// Find all uses of the casted value, and check if it is
// used in a Call Instruction
if (CallInst* CI = dyn_cast<CallInst>(*uii)) {
// Check that it is the called value, and not an argument
if(CI->getCalledValue() != CE)
continue;
// Check that the number of arguments passed, and expected
// by the function are the same.
if(!I->isVarArg()) {
if(CI->getNumOperands() != I->arg_size() + 1)
continue;
} else {
if(CI->getNumOperands() < I->arg_size() + 1)
continue;
}
// If so, add to worklist
worklist.push_back(CI);
}
}
}
}
// Proces the worklist of potential call sites to transform
while(!worklist.empty()) {
CallInst *CI = worklist.back();
worklist.pop_back();
// Get the called Function
Function *F = cast<Function>(CI->getCalledValue()->stripPointerCasts());
const FunctionType *FTy = F->getFunctionType();
SmallVector<Value*, 8> Args;
unsigned i =0;
for(i =0; i< FTy->getNumParams(); ++i) {
Type *ArgType = CI->getOperand(i+1)->getType();
Type *FormalType = FTy->getParamType(i);
// If the types for this argument match, just add it to the
// parameter list. No cast needs to be inserted.
if(ArgType == FormalType) {
Args.push_back(CI->getOperand(i+1));
}
else if(ArgType->isPointerTy() && FormalType->isPointerTy()) {
CastInst *CastI = CastInst::CreatePointerCast(CI->getOperand(i+1),
FormalType, "", CI);
Args.push_back(CastI);
} else if (ArgType->isIntegerTy() && FormalType->isIntegerTy()) {
unsigned SrcBits = ArgType->getScalarSizeInBits();
unsigned DstBits = FormalType->getScalarSizeInBits();
if(SrcBits > DstBits) {
CastInst *CastI = CastInst::CreateIntegerCast(CI->getOperand(i+1),
FormalType, true, "", CI);
Args.push_back(CastI);
} else {
if (F->getAttributes().hasAttribute(i+1, Attribute::SExt)) {
CastInst *CastI = CastInst::CreateIntegerCast(CI->getOperand(i+1),
FormalType, true, "", CI);
Args.push_back(CastI);
} else if (F->getAttributes().hasAttribute(i+1, Attribute::ZExt)) {
CastInst *CastI = CastInst::CreateIntegerCast(CI->getOperand(i+1),
FormalType, false, "", CI);
Args.push_back(CastI);
} else {
// Use ZExt in default case.
// Derived from InstCombine. Also, the only reason this should happen
// is mismatched prototypes.
// Seen in case of integer constants which get interpreted as i32,
// even if being used as i64.
// TODO: is this correct?
CastInst *CastI = CastInst::CreateIntegerCast(CI->getOperand(i+1),
FormalType, false, "", CI);
Args.push_back(CastI);
}
}
} else {
DEBUG(ArgType->dump());
DEBUG(FormalType->dump());
break;
}
}
// If we found an argument we could not cast, try the next instruction
if(i != FTy->getNumParams()) {
continue;
}
if(FTy->isVarArg()) {
for(; i< CI->getNumOperands() - 1 ;i++) {
Args.push_back(CI->getOperand(i+1));
}
}
// else replace the call instruction
CallInst *CINew = CallInst::Create(F, Args, "", CI);
CINew->setCallingConv(CI->getCallingConv());
CINew->setAttributes(CI->getAttributes());
if(!CI->use_empty()) {
CastInst *RetCast;
if(CI->getType() != CINew->getType()) {
if(CI->getType()->isPointerTy() && CINew->getType()->isPointerTy())
RetCast = CastInst::CreatePointerCast(CINew, CI->getType(), "", CI);
else if(CI->getType()->isIntOrIntVectorTy() && CINew->getType()->isIntOrIntVectorTy())
RetCast = CastInst::CreateIntegerCast(CINew, CI->getType(), false, "", CI);
else if(CI->getType()->isIntOrIntVectorTy() && CINew->getType()->isPointerTy())
RetCast = CastInst::CreatePointerCast(CINew, CI->getType(), "", CI);
else if(CI->getType()->isPointerTy() && CINew->getType()->isIntOrIntVectorTy())
RetCast = new IntToPtrInst(CINew, CI->getType(), "", CI);
else {
// TODO: I'm not sure what right behavior is here, but this case should be handled.
llvm_unreachable("Unexpected type conversion in call!");
abort();
}
CI->replaceAllUsesWith(RetCast);
} else {
CI->replaceAllUsesWith(CINew);
}
}
// Debug printing
DEBUG(errs() << "ARGCAST:");
DEBUG(errs() << "ERASE:");
DEBUG(CI->dump());
DEBUG(errs() << "ARGCAST:");
DEBUG(errs() << "ADDED:");
DEBUG(CINew->dump());
CI->eraseFromParent();
numChanged++;
}
return true;
}