void Preparer::expandCallSite(CallSite CS) {
// Skip the callsites that are not calling a va function.
Value *Callee = CS.getCalledValue();
FunctionType *CalleeType = cast<FunctionType>(
cast<PointerType>(Callee->getType())->getElementType());
if (!CalleeType->isVarArg()) {
return;
}
vector<Value *> Args;
for (CallSite::arg_iterator ArgI = CS.arg_begin();
ArgI != CS.arg_end(); ArgI++) {
Args.push_back(*ArgI);
}
Args.push_back(ConstantInt::get(
IntegerType::get(CS.getInstruction()->getContext(), 8), 0));
string InstName = "";
if (CS.getInstruction()->getName() != "")
InstName = CS.getInstruction()->getName().str() + ".padded";
if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
CallInst *NewCI = CallInst::Create(Callee, Args, InstName, CI);
NewCI->setAttributes(CI->getAttributes());
CI->replaceAllUsesWith(NewCI);
CI->eraseFromParent();
} else if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
InvokeInst *NewII = InvokeInst::Create(Callee,
II->getNormalDest(),
II->getUnwindDest(),
Args,
InstName,
II);
NewII->setAttributes(II->getAttributes());
II->replaceAllUsesWith(NewII);
II->eraseFromParent();
}
}
bool LowerEmExceptions::runOnModule(Module &M) {
TheModule = &M;
// Add functions
Type *i32 = Type::getInt32Ty(M.getContext());
Type *i8 = Type::getInt8Ty(M.getContext());
Type *i1 = Type::getInt1Ty(M.getContext());
Type *i8P = i8->getPointerTo();
Type *Void = Type::getVoidTy(M.getContext());
if (!(GetHigh = TheModule->getFunction("getHigh32"))) {
FunctionType *GetHighFunc = FunctionType::get(i32, false);
GetHigh = Function::Create(GetHighFunc, GlobalValue::ExternalLinkage,
"getHigh32", TheModule);
}
if (!(PreInvoke = TheModule->getFunction("emscripten_preinvoke"))) {
FunctionType *VoidFunc = FunctionType::get(Void, false);
PreInvoke = Function::Create(VoidFunc, GlobalValue::ExternalLinkage, "emscripten_preinvoke", TheModule);
}
if (!(PostInvoke = TheModule->getFunction("emscripten_postinvoke"))) {
FunctionType *IntFunc = FunctionType::get(i32, false);
PostInvoke = Function::Create(IntFunc, GlobalValue::ExternalLinkage, "emscripten_postinvoke", TheModule);
}
FunctionType *LandingPadFunc = FunctionType::get(i8P, true);
LandingPad = Function::Create(LandingPadFunc, GlobalValue::ExternalLinkage, "emscripten_landingpad", TheModule);
FunctionType *ResumeFunc = FunctionType::get(Void, true);
Resume = Function::Create(ResumeFunc, GlobalValue::ExternalLinkage, "emscripten_resume", TheModule);
// Process
bool HasWhitelist = Whitelist.size() > 0;
std::string WhitelistChecker;
if (HasWhitelist) WhitelistChecker = "," + Whitelist + ",";
bool Changed = false;
for (Module::iterator Iter = M.begin(), E = M.end(); Iter != E; ) {
Function *F = Iter++;
std::vector<Instruction*> ToErase;
std::set<LandingPadInst*> LandingPads;
bool AllowExceptionsInFunc = !HasWhitelist || int(WhitelistChecker.find(F->getName())) > 0;
for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
// check terminator for invokes
if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator())) {
LandingPads.insert(II->getLandingPadInst());
bool NeedInvoke = AllowExceptionsInFunc && canThrow(II->getCalledValue());
if (NeedInvoke) {
// Insert a normal call instruction folded in between pre- and post-invoke
CallInst::Create(PreInvoke, "", II);
SmallVector<Value*,16> CallArgs(II->op_begin(), II->op_end() - 3);
CallInst *NewCall = CallInst::Create(II->getCalledValue(),
CallArgs, "", II);
NewCall->takeName(II);
NewCall->setCallingConv(II->getCallingConv());
NewCall->setAttributes(II->getAttributes());
NewCall->setDebugLoc(II->getDebugLoc());
II->replaceAllUsesWith(NewCall);
ToErase.push_back(II);
CallInst *Post = CallInst::Create(PostInvoke, "", II);
Instruction *Post1 = new TruncInst(Post, i1, "", II);
// Insert a branch based on the postInvoke
BranchInst::Create(II->getUnwindDest(), II->getNormalDest(), Post1, II);
} else {
// This can't throw, and we don't need this invoke, just replace it with a call+branch
SmallVector<Value*,16> CallArgs(II->op_begin(), II->op_end() - 3);
CallInst *NewCall = CallInst::Create(II->getCalledValue(),
CallArgs, "", II);
NewCall->takeName(II);
NewCall->setCallingConv(II->getCallingConv());
NewCall->setAttributes(II->getAttributes());
NewCall->setDebugLoc(II->getDebugLoc());
II->replaceAllUsesWith(NewCall);
ToErase.push_back(II);
BranchInst::Create(II->getNormalDest(), II);
// Remove any PHI node entries from the exception destination.
II->getUnwindDest()->removePredecessor(BB);
}
Changed = true;
}
// scan the body of the basic block for resumes
for (BasicBlock::iterator Iter = BB->begin(), E = BB->end();
Iter != E; ) {
Instruction *I = Iter++;
if (ResumeInst *R = dyn_cast<ResumeInst>(I)) {
// split the input into legal values
Value *Input = R->getValue();
ExtractValueInst *Low = ExtractValueInst::Create(Input, 0, "", R);
ExtractValueInst *High = ExtractValueInst::Create(Input, 1, "", R);
// create a resume call
SmallVector<Value*,2> CallArgs;
CallArgs.push_back(Low);
CallArgs.push_back(High);
CallInst::Create(Resume, CallArgs, "", R);
new UnreachableInst(TheModule->getContext(), R); // add a terminator to the block
ToErase.push_back(R);
}
}
}
// Look for orphan landingpads, can occur in blocks with no predecesors
for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
Instruction *I = BB->getFirstNonPHI();
if (LandingPadInst *LP = dyn_cast<LandingPadInst>(I)) {
LandingPads.insert(LP);
}
}
// Handle all the landingpad for this function together, as multiple invokes may share a single lp
for (std::set<LandingPadInst*>::iterator I = LandingPads.begin(); I != LandingPads.end(); I++) {
// Replace the landingpad with a landingpad call to get the low part, and a getHigh for the high
LandingPadInst *LP = *I;
unsigned Num = LP->getNumClauses();
SmallVector<Value*,16> NewLPArgs;
NewLPArgs.push_back(LP->getPersonalityFn());
for (unsigned i = 0; i < Num; i++) {
Value *Arg = LP->getClause(i);
// As a temporary workaround for the lack of aggregate varargs support
// in the varargs lowering code, break out filter operands into their
// component elements.
if (LP->isFilter(i)) {
ArrayType *ATy = cast<ArrayType>(Arg->getType());
for (unsigned elem = 0, elemEnd = ATy->getNumElements(); elem != elemEnd; ++elem) {
Instruction *EE = ExtractValueInst::Create(Arg, makeArrayRef(elem), "", LP);
NewLPArgs.push_back(EE);
}
} else {
NewLPArgs.push_back(Arg);
}
}
NewLPArgs.push_back(LP->isCleanup() ? ConstantInt::getTrue(i1) : ConstantInt::getFalse(i1));
CallInst *NewLP = CallInst::Create(LandingPad, NewLPArgs, "", LP);
Instruction *High = CallInst::Create(GetHigh, "", LP);
// New recreate an aggregate for them, which will be all simplified later (simplification cannot handle landingpad, hence all this)
InsertValueInst *IVA = InsertValueInst::Create(UndefValue::get(LP->getType()), NewLP, 0, "", LP);
InsertValueInst *IVB = InsertValueInst::Create(IVA, High, 1, "", LP);
LP->replaceAllUsesWith(IVB);
ToErase.push_back(LP);
}
// erase everything we no longer need in this function
for (unsigned i = 0; i < ToErase.size(); i++) ToErase[i]->eraseFromParent();
}
return Changed;
}
/// Replaces the given call site (Call or Invoke) with a gc.statepoint
/// intrinsic with an empty deoptimization arguments list. This does
/// NOT do explicit relocation for GC support.
static Value *ReplaceWithStatepoint(const CallSite &CS, /* to replace */
Pass *P) {
assert(CS.getInstruction()->getParent()->getParent()->getParent() &&
"must be set");
// TODO: technically, a pass is not allowed to get functions from within a
// function pass since it might trigger a new function addition. Refactor
// this logic out to the initialization of the pass. Doesn't appear to
// matter in practice.
// Then go ahead and use the builder do actually do the inserts. We insert
// immediately before the previous instruction under the assumption that all
// arguments will be available here. We can't insert afterwards since we may
// be replacing a terminator.
IRBuilder<> Builder(CS.getInstruction());
// Note: The gc args are not filled in at this time, that's handled by
// RewriteStatepointsForGC (which is currently under review).
// Create the statepoint given all the arguments
Instruction *Token = nullptr;
AttributeSet OriginalAttrs;
if (CS.isCall()) {
CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
CallInst *Call = Builder.CreateGCStatepointCall(
CS.getCalledValue(), makeArrayRef(CS.arg_begin(), CS.arg_end()), None,
None, "safepoint_token");
Call->setTailCall(ToReplace->isTailCall());
Call->setCallingConv(ToReplace->getCallingConv());
// Before we have to worry about GC semantics, all attributes are legal
// TODO: handle param attributes
OriginalAttrs = ToReplace->getAttributes();
// In case if we can handle this set of attributes - set up function
// attributes directly on statepoint and return attributes later for
// gc_result intrinsic.
Call->setAttributes(OriginalAttrs.getFnAttributes());
Token = Call;
// Put the following gc_result and gc_relocate calls immediately after the
// the old call (which we're about to delete).
assert(ToReplace->getNextNode() && "not a terminator, must have next");
Builder.SetInsertPoint(ToReplace->getNextNode());
Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
} else if (CS.isInvoke()) {
InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
// Insert the new invoke into the old block. We'll remove the old one in a
// moment at which point this will become the new terminator for the
// original block.
Builder.SetInsertPoint(ToReplace->getParent());
InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
CS.getCalledValue(), ToReplace->getNormalDest(),
ToReplace->getUnwindDest(), makeArrayRef(CS.arg_begin(), CS.arg_end()),
Builder.getInt32(0), None, "safepoint_token");
// Currently we will fail on parameter attributes and on certain
// function attributes.
OriginalAttrs = ToReplace->getAttributes();
// In case if we can handle this set of attributes - set up function
// attributes directly on statepoint and return attributes later for
// gc_result intrinsic.
Invoke->setAttributes(OriginalAttrs.getFnAttributes());
Token = Invoke;
// We'll insert the gc.result into the normal block
BasicBlock *NormalDest = normalizeBBForInvokeSafepoint(
ToReplace->getNormalDest(), Invoke->getParent());
Builder.SetInsertPoint(NormalDest->getFirstInsertionPt());
} else {
llvm_unreachable("unexpect type of CallSite");
}
assert(Token);
// Handle the return value of the original call - update all uses to use a
// gc_result hanging off the statepoint node we just inserted
// Only add the gc_result iff there is actually a used result
if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
std::string TakenName =
CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), TakenName);
GCResult->setAttributes(OriginalAttrs.getRetAttributes());
return GCResult;
} else {
// No return value for the call.
return nullptr;
}
}
bool WebAssemblyLowerEmscriptenEHSjLj::runEHOnFunction(Function &F) {
Module &M = *F.getParent();
LLVMContext &C = F.getContext();
IRBuilder<> IRB(C);
bool Changed = false;
SmallVector<Instruction *, 64> ToErase;
SmallPtrSet<LandingPadInst *, 32> LandingPads;
bool AllowExceptions =
areAllExceptionsAllowed() || EHWhitelistSet.count(F.getName());
for (BasicBlock &BB : F) {
auto *II = dyn_cast<InvokeInst>(BB.getTerminator());
if (!II)
continue;
Changed = true;
LandingPads.insert(II->getLandingPadInst());
IRB.SetInsertPoint(II);
bool NeedInvoke = AllowExceptions && canThrow(II->getCalledValue());
if (NeedInvoke) {
// Wrap invoke with invoke wrapper and generate preamble/postamble
Value *Threw = wrapInvoke(II);
ToErase.push_back(II);
// Insert a branch based on __THREW__ variable
Value *Cmp = IRB.CreateICmpEQ(Threw, IRB.getInt32(1), "cmp");
IRB.CreateCondBr(Cmp, II->getUnwindDest(), II->getNormalDest());
} else {
// This can't throw, and we don't need this invoke, just replace it with a
// call+branch
SmallVector<Value *, 16> Args(II->arg_begin(), II->arg_end());
CallInst *NewCall = IRB.CreateCall(II->getCalledValue(), Args);
NewCall->takeName(II);
NewCall->setCallingConv(II->getCallingConv());
NewCall->setDebugLoc(II->getDebugLoc());
NewCall->setAttributes(II->getAttributes());
II->replaceAllUsesWith(NewCall);
ToErase.push_back(II);
IRB.CreateBr(II->getNormalDest());
// Remove any PHI node entries from the exception destination
II->getUnwindDest()->removePredecessor(&BB);
}
}
// Process resume instructions
for (BasicBlock &BB : F) {
// Scan the body of the basic block for resumes
for (Instruction &I : BB) {
auto *RI = dyn_cast<ResumeInst>(&I);
if (!RI)
continue;
// Split the input into legal values
Value *Input = RI->getValue();
IRB.SetInsertPoint(RI);
Value *Low = IRB.CreateExtractValue(Input, 0, "low");
// Create a call to __resumeException function
IRB.CreateCall(ResumeF, {Low});
// Add a terminator to the block
IRB.CreateUnreachable();
ToErase.push_back(RI);
}
}
// Process llvm.eh.typeid.for intrinsics
for (BasicBlock &BB : F) {
for (Instruction &I : BB) {
auto *CI = dyn_cast<CallInst>(&I);
if (!CI)
continue;
const Function *Callee = CI->getCalledFunction();
if (!Callee)
continue;
if (Callee->getIntrinsicID() != Intrinsic::eh_typeid_for)
continue;
IRB.SetInsertPoint(CI);
CallInst *NewCI =
IRB.CreateCall(EHTypeIDF, CI->getArgOperand(0), "typeid");
CI->replaceAllUsesWith(NewCI);
ToErase.push_back(CI);
}
}
// Look for orphan landingpads, can occur in blocks with no predecessors
for (BasicBlock &BB : F) {
Instruction *I = BB.getFirstNonPHI();
if (auto *LPI = dyn_cast<LandingPadInst>(I))
LandingPads.insert(LPI);
}
// Handle all the landingpad for this function together, as multiple invokes
// may share a single lp
for (LandingPadInst *LPI : LandingPads) {
IRB.SetInsertPoint(LPI);
SmallVector<Value *, 16> FMCArgs;
for (unsigned i = 0, e = LPI->getNumClauses(); i < e; ++i) {
Constant *Clause = LPI->getClause(i);
// As a temporary workaround for the lack of aggregate varargs support
// in the interface between JS and wasm, break out filter operands into
// their component elements.
if (LPI->isFilter(i)) {
auto *ATy = cast<ArrayType>(Clause->getType());
for (unsigned j = 0, e = ATy->getNumElements(); j < e; ++j) {
Value *EV = IRB.CreateExtractValue(Clause, makeArrayRef(j), "filter");
FMCArgs.push_back(EV);
}
} else
FMCArgs.push_back(Clause);
}
// Create a call to __cxa_find_matching_catch_N function
Function *FMCF = getFindMatchingCatch(M, FMCArgs.size());
CallInst *FMCI = IRB.CreateCall(FMCF, FMCArgs, "fmc");
Value *Undef = UndefValue::get(LPI->getType());
Value *Pair0 = IRB.CreateInsertValue(Undef, FMCI, 0, "pair0");
Value *TempRet0 =
IRB.CreateLoad(TempRet0GV, TempRet0GV->getName() + ".val");
Value *Pair1 = IRB.CreateInsertValue(Pair0, TempRet0, 1, "pair1");
LPI->replaceAllUsesWith(Pair1);
ToErase.push_back(LPI);
}
// Erase everything we no longer need in this function
for (Instruction *I : ToErase)
I->eraseFromParent();
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;
}
/// Replace direct callers of Old with New. Also add parameters to the call to
/// \p New, which are defined by the FuncIdx's value in \p Params.
bool SwiftMergeFunctions::replaceDirectCallers(Function *Old, Function *New,
const ParamInfos &Params, unsigned FuncIdx) {
bool AllReplaced = true;
SmallVector<CallInst *, 8> Callers;
for (Use &U : Old->uses()) {
auto *I = dyn_cast<Instruction>(U.getUser());
if (!I) {
AllReplaced = false;
continue;
}
FunctionEntry *FE = getEntry(I->getFunction());
if (FE)
removeEquivalenceClassFromTree(FE);
auto *CI = dyn_cast<CallInst>(I);
if (!CI || CI->getCalledValue() != Old) {
AllReplaced = false;
continue;
}
Callers.push_back(CI);
}
if (!AllReplaced)
return false;
for (CallInst *CI : Callers) {
auto &Context = New->getContext();
auto NewPAL = New->getAttributes();
SmallVector<Type *, 8> OldParamTypes;
SmallVector<Value *, 16> NewArgs;
SmallVector<AttributeSet, 8> NewArgAttrs;
IRBuilder<> Builder(CI);
FunctionType *NewFuncTy = New->getFunctionType();
(void) NewFuncTy;
unsigned ParamIdx = 0;
// Add the existing parameters.
for (Value *OldArg : CI->arg_operands()) {
NewArgAttrs.push_back(NewPAL.getParamAttributes(ParamIdx));
NewArgs.push_back(OldArg);
OldParamTypes.push_back(OldArg->getType());
++ParamIdx;
}
// Add the new parameters.
for (const ParamInfo &PI : Params) {
assert(ParamIdx < NewFuncTy->getNumParams());
Constant *ArgValue = PI.Values[FuncIdx];
assert(ArgValue != Old &&
"should not try to replace all callers of self referencing functions");
NewArgs.push_back(ArgValue);
OldParamTypes.push_back(ArgValue->getType());
++ParamIdx;
}
auto *FType = FunctionType::get(Old->getFunctionType()->getReturnType(),
OldParamTypes, false);
auto *FPtrType = PointerType::get(FType,
cast<PointerType>(New->getType())->getAddressSpace());
Value *Callee = ConstantExpr::getBitCast(New, FPtrType);
CallInst *NewCI = Builder.CreateCall(Callee, NewArgs);
NewCI->setCallingConv(CI->getCallingConv());
// Don't transfer attributes from the function to the callee. Function
// attributes typically aren't relevant to the calling convention or ABI.
NewCI->setAttributes(AttributeList::get(Context, /*FnAttrs=*/AttributeSet(),
NewPAL.getRetAttributes(),
NewArgAttrs));
CI->replaceAllUsesWith(NewCI);
CI->eraseFromParent();
}
assert(Old->use_empty() && "should have replaced all uses of old function");
return Old->hasLocalLinkage();
}
// Convert the given call to use normalized argument/return types.
template <class T> static bool ConvertCall(T *Call, Pass *P) {
// Don't try to change calls to intrinsics.
if (isa<IntrinsicInst>(Call))
return false;
FunctionType *FTy = cast<FunctionType>(
Call->getCalledValue()->getType()->getPointerElementType());
FunctionType *NFTy = NormalizeFunctionType(FTy);
if (NFTy == FTy)
return false; // No change needed.
// Convert arguments.
SmallVector<Value *, 8> Args;
for (unsigned I = 0; I < Call->getNumArgOperands(); ++I) {
Value *Arg = Call->getArgOperand(I);
if (NFTy->getParamType(I) != FTy->getParamType(I)) {
Instruction::CastOps CastType =
Call->getAttributes().hasAttribute(I + 1, Attribute::SExt) ?
Instruction::SExt : Instruction::ZExt;
Arg = CopyDebug(CastInst::Create(CastType, Arg, NFTy->getParamType(I),
"arg_ext", Call), Call);
}
Args.push_back(Arg);
}
Value *CastFunc =
CopyDebug(new BitCastInst(Call->getCalledValue(), NFTy->getPointerTo(),
Call->getName() + ".arg_cast", Call), Call);
Value *Result = NULL;
if (CallInst *OldCall = dyn_cast<CallInst>(Call)) {
CallInst *NewCall = CopyDebug(CallInst::Create(CastFunc, Args, "", OldCall),
OldCall);
NewCall->takeName(OldCall);
NewCall->setAttributes(OldCall->getAttributes());
NewCall->setCallingConv(OldCall->getCallingConv());
NewCall->setTailCall(OldCall->isTailCall());
Result = NewCall;
if (FTy->getReturnType() != NFTy->getReturnType()) {
Result = CopyDebug(new TruncInst(NewCall, FTy->getReturnType(),
NewCall->getName() + ".ret_trunc", Call),
Call);
}
} else if (InvokeInst *OldInvoke = dyn_cast<InvokeInst>(Call)) {
BasicBlock *Parent = OldInvoke->getParent();
BasicBlock *NormalDest = OldInvoke->getNormalDest();
BasicBlock *UnwindDest = OldInvoke->getUnwindDest();
if (FTy->getReturnType() != NFTy->getReturnType()) {
if (BasicBlock *SplitDest = SplitCriticalEdge(Parent, NormalDest)) {
NormalDest = SplitDest;
}
}
InvokeInst *New = CopyDebug(InvokeInst::Create(CastFunc, NormalDest,
UnwindDest, Args,
"", OldInvoke),
OldInvoke);
New->takeName(OldInvoke);
if (FTy->getReturnType() != NFTy->getReturnType()) {
Result = CopyDebug(new TruncInst(New, FTy->getReturnType(),
New->getName() + ".ret_trunc",
NormalDest->getTerminator()),
OldInvoke);
} else {
Result = New;
}
New->setAttributes(OldInvoke->getAttributes());
New->setCallingConv(OldInvoke->getCallingConv());
}
Call->replaceAllUsesWith(Result);
Call->eraseFromParent();
return true;
}
/// InsertStackProtectors - Insert code into the prologue and epilogue of the
/// function.
///
/// - The prologue code loads and stores the stack guard onto the stack.
/// - The epilogue checks the value stored in the prologue against the original
/// value. It calls __stack_chk_fail if they differ.
bool StackProtector::InsertStackProtectors() {
bool SupportsSelectionDAGSP =
EnableSelectionDAGSP && !TM->Options.EnableFastISel;
AllocaInst *AI = nullptr; // Place on stack that stores the stack guard.
for (Function::iterator I = F->begin(), E = F->end(); I != E;) {
BasicBlock *BB = &*I++;
ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
if (!RI)
continue;
// Generate prologue instrumentation if not already generated.
if (!HasPrologue) {
HasPrologue = true;
SupportsSelectionDAGSP &= CreatePrologue(F, M, RI, TLI, AI);
}
// SelectionDAG based code generation. Nothing else needs to be done here.
// The epilogue instrumentation is postponed to SelectionDAG.
if (SupportsSelectionDAGSP)
break;
// Set HasIRCheck to true, so that SelectionDAG will not generate its own
// version. SelectionDAG called 'shouldEmitSDCheck' to check whether
// instrumentation has already been generated.
HasIRCheck = true;
// Generate epilogue instrumentation. The epilogue intrumentation can be
// function-based or inlined depending on which mechanism the target is
// providing.
if (Value* GuardCheck = TLI->getSSPStackGuardCheck(*M)) {
// Generate the function-based epilogue instrumentation.
// The target provides a guard check function, generate a call to it.
IRBuilder<> B(RI);
LoadInst *Guard = B.CreateLoad(AI, true, "Guard");
CallInst *Call = B.CreateCall(GuardCheck, {Guard});
llvm::Function *Function = cast<llvm::Function>(GuardCheck);
Call->setAttributes(Function->getAttributes());
Call->setCallingConv(Function->getCallingConv());
} else {
// Generate the epilogue with inline instrumentation.
// If we do not support SelectionDAG based tail calls, generate IR level
// tail calls.
//
// For each block with a return instruction, convert this:
//
// return:
// ...
// ret ...
//
// into this:
//
// return:
// ...
// %1 = <stack guard>
// %2 = load StackGuardSlot
// %3 = cmp i1 %1, %2
// br i1 %3, label %SP_return, label %CallStackCheckFailBlk
//
// SP_return:
// ret ...
//
// CallStackCheckFailBlk:
// call void @__stack_chk_fail()
// unreachable
// Create the FailBB. We duplicate the BB every time since the MI tail
// merge pass will merge together all of the various BB into one including
// fail BB generated by the stack protector pseudo instruction.
BasicBlock *FailBB = CreateFailBB();
// Split the basic block before the return instruction.
BasicBlock *NewBB = BB->splitBasicBlock(RI->getIterator(), "SP_return");
// Update the dominator tree if we need to.
if (DT && DT->isReachableFromEntry(BB)) {
DT->addNewBlock(NewBB, BB);
DT->addNewBlock(FailBB, BB);
}
// Remove default branch instruction to the new BB.
BB->getTerminator()->eraseFromParent();
// Move the newly created basic block to the point right after the old
// basic block so that it's in the "fall through" position.
NewBB->moveAfter(BB);
// Generate the stack protector instructions in the old basic block.
IRBuilder<> B(BB);
Value *Guard = getStackGuard(TLI, M, B);
LoadInst *LI2 = B.CreateLoad(AI, true);
Value *Cmp = B.CreateICmpEQ(Guard, LI2);
auto SuccessProb =
BranchProbabilityInfo::getBranchProbStackProtector(true);
auto FailureProb =
BranchProbabilityInfo::getBranchProbStackProtector(false);
MDNode *Weights = MDBuilder(F->getContext())
.createBranchWeights(SuccessProb.getNumerator(),
FailureProb.getNumerator());
B.CreateCondBr(Cmp, NewBB, FailBB, Weights);
}
}
// Return if we didn't modify any basic blocks. i.e., there are no return
// statements in the function.
return HasPrologue;
}
//
// 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;
}
bool LowerEmSetjmp::runOnModule(Module &M) {
TheModule = &M;
Function *Setjmp = TheModule->getFunction("setjmp");
Function *Longjmp = TheModule->getFunction("longjmp");
if (!Setjmp && !Longjmp) return false;
Type *i32 = Type::getInt32Ty(M.getContext());
Type *Void = Type::getVoidTy(M.getContext());
// Add functions
Function *EmSetjmp = NULL;
if (Setjmp) {
SmallVector<Type*, 2> EmSetjmpTypes;
EmSetjmpTypes.push_back(Setjmp->getFunctionType()->getParamType(0));
EmSetjmpTypes.push_back(i32); // extra param that says which setjmp in the function it is
FunctionType *EmSetjmpFunc = FunctionType::get(i32, EmSetjmpTypes, false);
EmSetjmp = Function::Create(EmSetjmpFunc, GlobalValue::ExternalLinkage, "emscripten_setjmp", TheModule);
}
Function *EmLongjmp = Longjmp ? Function::Create(Longjmp->getFunctionType(), GlobalValue::ExternalLinkage, "emscripten_longjmp", TheModule) : NULL;
SmallVector<Type*, 1> IntArgTypes;
IntArgTypes.push_back(i32);
FunctionType *IntIntFunc = FunctionType::get(i32, IntArgTypes, false);
Function *CheckLongjmp = Function::Create(IntIntFunc, GlobalValue::ExternalLinkage, "emscripten_check_longjmp", TheModule); // gets control flow
Function *GetLongjmpResult = Function::Create(IntIntFunc, GlobalValue::ExternalLinkage, "emscripten_get_longjmp_result", TheModule); // gets int value longjmp'd
FunctionType *VoidFunc = FunctionType::get(Void, false);
Function *PrepSetjmp = Function::Create(VoidFunc, GlobalValue::ExternalLinkage, "emscripten_prep_setjmp", TheModule);
Function *CleanupSetjmp = Function::Create(VoidFunc, GlobalValue::ExternalLinkage, "emscripten_cleanup_setjmp", TheModule);
Function *PreInvoke = TheModule->getFunction("emscripten_preinvoke");
if (!PreInvoke) PreInvoke = Function::Create(VoidFunc, GlobalValue::ExternalLinkage, "emscripten_preinvoke", TheModule);
FunctionType *IntFunc = FunctionType::get(i32, false);
Function *PostInvoke = TheModule->getFunction("emscripten_postinvoke");
if (!PostInvoke) PostInvoke = Function::Create(IntFunc, GlobalValue::ExternalLinkage, "emscripten_postinvoke", TheModule);
// Process all callers of setjmp and longjmp. Start with setjmp.
typedef std::vector<PHINode*> Phis;
typedef std::map<Function*, Phis> FunctionPhisMap;
FunctionPhisMap SetjmpOutputPhis;
std::vector<Instruction*> ToErase;
if (Setjmp) {
for (Instruction::user_iterator UI = Setjmp->user_begin(), UE = Setjmp->user_end(); UI != UE; ++UI) {
User *U = *UI;
if (CallInst *CI = dyn_cast<CallInst>(U)) {
BasicBlock *SJBB = CI->getParent();
// The tail is everything right after the call, and will be reached once when setjmp is
// called, and later when longjmp returns to the setjmp
BasicBlock *Tail = SplitBlock(SJBB, CI->getNextNode());
// Add a phi to the tail, which will be the output of setjmp, which indicates if this is the
// first call or a longjmp back. The phi directly uses the right value based on where we
// arrive from
PHINode *SetjmpOutput = PHINode::Create(i32, 2, "", Tail->getFirstNonPHI());
SetjmpOutput->addIncoming(ConstantInt::get(i32, 0), SJBB); // setjmp initial call returns 0
CI->replaceAllUsesWith(SetjmpOutput); // The proper output is now this, not the setjmp call itself
// longjmp returns to the setjmp will add themselves to this phi
Phis& P = SetjmpOutputPhis[SJBB->getParent()];
P.push_back(SetjmpOutput);
// fix call target
SmallVector<Value *, 2> Args;
Args.push_back(CI->getArgOperand(0));
Args.push_back(ConstantInt::get(i32, P.size())); // our index in the function is our place in the array + 1
CallInst::Create(EmSetjmp, Args, "", CI);
ToErase.push_back(CI);
} else {
errs() << **UI << "\n";
report_fatal_error("bad use of setjmp, should only call it");
}
}
}
// Update longjmp FIXME: we could avoid throwing in longjmp as an optimization when longjmping back into the current function perhaps?
if (Longjmp) Longjmp->replaceAllUsesWith(EmLongjmp);
// Update all setjmping functions
for (FunctionPhisMap::iterator I = SetjmpOutputPhis.begin(); I != SetjmpOutputPhis.end(); I++) {
Function *F = I->first;
Phis& P = I->second;
CallInst::Create(PrepSetjmp, "", F->begin()->begin());
// Update each call that can longjmp so it can return to a setjmp where relevant
for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ) {
BasicBlock *BB = BBI++;
for (BasicBlock::iterator Iter = BB->begin(), E = BB->end(); Iter != E; ) {
Instruction *I = Iter++;
CallInst *CI;
if ((CI = dyn_cast<CallInst>(I))) {
Value *V = CI->getCalledValue();
if (V == PrepSetjmp || V == EmSetjmp || V == CheckLongjmp || V == GetLongjmpResult || V == PreInvoke || V == PostInvoke) continue;
if (Function *CF = dyn_cast<Function>(V)) if (CF->isIntrinsic()) continue;
// TODO: proper analysis of what can actually longjmp. Currently we assume anything but setjmp can.
// This may longjmp, so we need to check if it did. Split at that point, and
// envelop the call in pre/post invoke, if we need to
CallInst *After;
Instruction *Check = NULL;
if (Iter != E && (After = dyn_cast<CallInst>(Iter)) && After->getCalledValue() == PostInvoke) {
// use the pre|postinvoke that exceptions lowering already made
Check = Iter++;
}
BasicBlock *Tail = SplitBlock(BB, Iter); // Iter already points to the next instruction, as we need
TerminatorInst *TI = BB->getTerminator();
if (!Check) {
// no existing pre|postinvoke, create our own
CallInst::Create(PreInvoke, "", CI);
Check = CallInst::Create(PostInvoke, "", TI); // CI is at the end of the block
// If we are calling a function that is noreturn, we must remove that attribute. The code we
// insert here does expect it to return, after we catch the exception.
if (CI->doesNotReturn()) {
if (Function *F = dyn_cast<Function>(CI->getCalledValue())) {
F->removeFnAttr(Attribute::NoReturn);
}
CI->setAttributes(CI->getAttributes().removeAttribute(TheModule->getContext(), AttributeSet::FunctionIndex, Attribute::NoReturn));
assert(!CI->doesNotReturn());
}
}
// We need to replace the terminator in Tail - SplitBlock makes BB go straight to Tail, we need to check if a longjmp occurred, and
// go to the right setjmp-tail if so
SmallVector<Value *, 1> Args;
Args.push_back(Check);
Instruction *LongjmpCheck = CallInst::Create(CheckLongjmp, Args, "", BB);
Instruction *LongjmpResult = CallInst::Create(GetLongjmpResult, Args, "", BB);
SwitchInst *SI = SwitchInst::Create(LongjmpCheck, Tail, 2, BB);
// -1 means no longjmp happened, continue normally (will hit the default switch case). 0 means a longjmp that is not ours to handle, needs a rethrow. Otherwise
// the index mean is the same as the index in P+1 (to avoid 0).
for (unsigned i = 0; i < P.size(); i++) {
SI->addCase(cast<ConstantInt>(ConstantInt::get(i32, i+1)), P[i]->getParent());
P[i]->addIncoming(LongjmpResult, BB);
}
ToErase.push_back(TI); // new terminator is now the switch
// we are splitting the block here, and must continue to find other calls in the block - which is now split. so continue
// to traverse in the Tail
BB = Tail;
Iter = BB->begin();
E = BB->end();
} else if (InvokeInst *CI = dyn_cast<InvokeInst>(I)) { // XXX check if target is setjmp
(void)CI;
report_fatal_error("TODO: invoke inside setjmping functions");
}
}
}
// add a cleanup before each return
for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ) {
BasicBlock *BB = BBI++;
TerminatorInst *TI = BB->getTerminator();
if (isa<ReturnInst>(TI)) {
CallInst::Create(CleanupSetjmp, "", TI);
}
}
}
for (unsigned i = 0; i < ToErase.size(); i++) {
ToErase[i]->eraseFromParent();
}
// Finally, our modifications to the cfg can break dominance of SSA variables. For example,
// if (x()) { .. setjmp() .. }
// if (y()) { .. longjmp() .. }
// We must split the longjmp block, and it can jump into the setjmp one. But that means that when
// we split the setjmp block, it's first part no longer dominates its second part - there is
// a theoretically possible control flow path where x() is false, then y() is true and we
// reach the second part of the setjmp block, without ever reaching the first part. So,
// we recalculate regs vs. mem
for (FunctionPhisMap::iterator I = SetjmpOutputPhis.begin(); I != SetjmpOutputPhis.end(); I++) {
Function *F = I->first;
doRegToMem(*F);
doMemToReg(*F);
}
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;
}
/// HandleURoRInvokes - Handle invokes of "_Unwind_Resume_or_Rethrow" calls. The
/// "unwind" part of these invokes jump to a landing pad within the current
/// function. This is a candidate to merge the selector associated with the URoR
/// invoke with the one from the URoR's landing pad.
bool DwarfEHPrepare::HandleURoRInvokes() {
if (!EHCatchAllValue) {
EHCatchAllValue =
F->getParent()->getNamedGlobal("llvm.eh.catch.all.value");
if (!EHCatchAllValue) return false;
}
if (!SelectorIntrinsic) {
SelectorIntrinsic =
Intrinsic::getDeclaration(F->getParent(), Intrinsic::eh_selector);
if (!SelectorIntrinsic) return false;
}
SmallPtrSet<IntrinsicInst*, 32> Sels;
SmallPtrSet<IntrinsicInst*, 32> CatchAllSels;
FindAllCleanupSelectors(Sels, CatchAllSels);
if (!DT)
// We require DominatorTree information.
return CleanupSelectors(CatchAllSels);
if (!URoR) {
URoR = F->getParent()->getFunction("_Unwind_Resume_or_Rethrow");
if (!URoR) return CleanupSelectors(CatchAllSels);
}
SmallPtrSet<InvokeInst*, 32> URoRInvokes;
FindAllURoRInvokes(URoRInvokes);
SmallPtrSet<IntrinsicInst*, 32> SelsToConvert;
for (SmallPtrSet<IntrinsicInst*, 32>::iterator
SI = Sels.begin(), SE = Sels.end(); SI != SE; ++SI) {
const BasicBlock *SelBB = (*SI)->getParent();
for (SmallPtrSet<InvokeInst*, 32>::iterator
UI = URoRInvokes.begin(), UE = URoRInvokes.end(); UI != UE; ++UI) {
const BasicBlock *URoRBB = (*UI)->getParent();
if (DT->dominates(SelBB, URoRBB)) {
SelsToConvert.insert(*SI);
break;
}
}
}
bool Changed = false;
if (Sels.size() != SelsToConvert.size()) {
// If we haven't been able to convert all of the clean-up selectors, then
// loop through the slow way to see if they still need to be converted.
if (!ExceptionValueIntrinsic) {
ExceptionValueIntrinsic =
Intrinsic::getDeclaration(F->getParent(), Intrinsic::eh_exception);
if (!ExceptionValueIntrinsic)
return CleanupSelectors(CatchAllSels);
}
for (Value::use_iterator
I = ExceptionValueIntrinsic->use_begin(),
E = ExceptionValueIntrinsic->use_end(); I != E; ++I) {
IntrinsicInst *EHPtr = dyn_cast<IntrinsicInst>(*I);
if (!EHPtr || EHPtr->getParent()->getParent() != F) continue;
Changed |= PromoteEHPtrStore(EHPtr);
bool URoRInvoke = false;
SmallPtrSet<IntrinsicInst*, 8> SelCalls;
Changed |= FindSelectorAndURoR(EHPtr, URoRInvoke, SelCalls);
if (URoRInvoke) {
// This EH pointer is being used by an invoke of an URoR instruction and
// an eh.selector intrinsic call. If the eh.selector is a 'clean-up', we
// need to convert it to a 'catch-all'.
for (SmallPtrSet<IntrinsicInst*, 8>::iterator
SI = SelCalls.begin(), SE = SelCalls.end(); SI != SE; ++SI)
if (!HasCatchAllInSelector(*SI))
SelsToConvert.insert(*SI);
}
}
}
if (!SelsToConvert.empty()) {
// Convert all clean-up eh.selectors, which are associated with "invokes" of
// URoR calls, into catch-all eh.selectors.
Changed = true;
for (SmallPtrSet<IntrinsicInst*, 8>::iterator
SI = SelsToConvert.begin(), SE = SelsToConvert.end();
SI != SE; ++SI) {
IntrinsicInst *II = *SI;
// Use the exception object pointer and the personality function
// from the original selector.
CallSite CS(II);
IntrinsicInst::op_iterator I = CS.arg_begin();
IntrinsicInst::op_iterator E = CS.arg_end();
IntrinsicInst::op_iterator B = prior(E);
// Exclude last argument if it is an integer.
if (isa<ConstantInt>(B)) E = B;
// Add exception object pointer (front).
// Add personality function (next).
// Add in any filter IDs (rest).
SmallVector<Value*, 8> Args(I, E);
Args.push_back(EHCatchAllValue->getInitializer()); // Catch-all indicator.
CallInst *NewSelector =
CallInst::Create(SelectorIntrinsic, Args.begin(), Args.end(),
"eh.sel.catch.all", II);
NewSelector->setTailCall(II->isTailCall());
NewSelector->setAttributes(II->getAttributes());
NewSelector->setCallingConv(II->getCallingConv());
II->replaceAllUsesWith(NewSelector);
II->eraseFromParent();
}
}
Changed |= CleanupSelectors(CatchAllSels);
return Changed;
}
void WorklessInstrument::InstrumentWorkless0Or1Star(Module * pModule, Loop * pLoop, set<string> & setWorkingBlocks)
{
LoadInst * pLoad0 = NULL;
LoadInst * pLoad1 = NULL;
//BinaryOperator* pAdd = NULL;
StoreInst * pStore = NULL;
CallInst * pCall = NULL;
Function * pMain = NULL;
if(strMainName != "" )
{
pMain = pModule->getFunction(strMainName.c_str());
}
else
{
pMain = pModule->getFunction("main");
}
for (Function::iterator BB = pMain->begin(); BB != pMain->end(); ++BB)
{
if(BB->getName().equals("entry"))
{
CallInst * pCall;
StoreInst * pStore;
Instruction * II = BB->begin();
pCall = CallInst::Create(this->InitHooks, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet emptySet;
pCall->setAttributes(emptySet);
pCall = CallInst::Create(this->getenv, this->SAMPLE_RATE_ptr, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet AS;
{
SmallVector<AttributeSet, 4> Attrs;
AttributeSet PAS;
{
AttrBuilder B;
B.addAttribute(Attribute::NoUnwind);
PAS = AttributeSet::get(pModule->getContext(), ~0U, B);
}
Attrs.push_back(PAS);
AS = AttributeSet::get(pModule->getContext(), Attrs);
}
pCall->setAttributes(AS);
pCall = CallInst::Create(this->function_atoi, pCall, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
{
SmallVector<AttributeSet, 4> Attrs;
AttributeSet PAS;
{
AttrBuilder B;
B.addAttribute(Attribute::NoUnwind);
B.addAttribute(Attribute::ReadOnly);
PAS = AttributeSet::get(pModule->getContext(), ~0U, B);
}
Attrs.push_back(PAS);
AS = AttributeSet::get(pModule->getContext(), Attrs);
}
pCall->setAttributes(AS);
pStore = new StoreInst(pCall, this->SAMPLE_RATE, false, II);
pStore->setAlignment(4);
pCall = CallInst::Create(this->geo, pCall, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
pCall->setAttributes(emptySet);
CastInst * pCast = CastInst::CreateIntegerCast(pCall, this->LongType, true, "", II);
pStore = new StoreInst(pCast, this->CURRENT_SAMPLE, false, II);
pStore->setAlignment(8);
vector<Value *> vecParam;
vecParam.push_back(this->Output_Format_String);
vecParam.push_back(pCall);
pCall = CallInst::Create(this->printf, vecParam, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
pCall->setAttributes(emptySet);
break;
}
}
for (Function::iterator BB = pMain->begin(); BB != pMain->end(); ++BB)
{
for (BasicBlock::iterator Ins = BB->begin(); Ins != BB->end(); ++Ins)
{
if (isa<ReturnInst>(Ins) || isa<ResumeInst>(Ins))
{
vector<Value*> vecParams;
pLoad0 = new LoadInst(numIterations, "", false, Ins);
pLoad0->setAlignment(8);
vecParams.push_back(pLoad0);
pLoad1 = new LoadInst(numInstances, "", false, Ins);
pLoad1->setAlignment(8);
vecParams.push_back(pLoad1);
pCall = CallInst::Create(this->PrintLoopInfo, vecParams, "", Ins);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet aSet;
pCall->setAttributes(aSet);
vecParams.clear();
pLoad0 = new LoadInst(numIterations, "", false, Ins);
pLoad0->setAlignment(8);
vecParams.push_back(pLoad0);
pLoad1 = new LoadInst(numWorkingIterations, "", false, Ins);
pLoad1->setAlignment(8);
vecParams.push_back(pLoad1);
pCall = CallInst::Create(PrintWorkingRatio, vecParams, "", Ins);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
pCall->setAttributes(aSet);
}
else if(isa<CallInst>(Ins) || isa<InvokeInst>(Ins))
{
CallSite cs(Ins);
Function * pCalled = cs.getCalledFunction();
if(pCalled == NULL)
{
continue;
}
if(pCalled->getName() == "exit" || pCalled->getName() == "_ZL9mysql_endi")
{
vector<Value*> vecParams;
pLoad0 = new LoadInst(numIterations, "", false, Ins);
pLoad0->setAlignment(8);
vecParams.push_back(pLoad0);
pLoad1 = new LoadInst(numInstances, "", false, Ins);
pLoad1->setAlignment(8);
vecParams.push_back(pLoad1);
pCall = CallInst::Create(this->PrintLoopInfo, vecParams, "", Ins);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet aSet;
pCall->setAttributes(aSet);
vecParams.clear();
pLoad0 = new LoadInst(numIterations, "", false, Ins);
pLoad0->setAlignment(8);
vecParams.push_back(pLoad0);
pLoad1 = new LoadInst(numWorkingIterations, "", false, Ins);
pLoad1->setAlignment(8);
vecParams.push_back(pLoad1);
pCall = CallInst::Create(PrintWorkingRatio, vecParams, "", Ins);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
pCall->setAttributes(aSet);
}
}
}
}
Function * pFunction = pLoop->getHeader()->getParent();
BasicBlock * pEntry = &(pFunction->getEntryBlock());
AllocaInst * pAlloc = new AllocaInst(this->LongType, "bWorkingIteration.local", pEntry->getFirstInsertionPt());
vector<BasicBlock *> vecWorkingBlock;
for(Function::iterator BB = pFunction->begin(); BB != pFunction->end(); ++ BB)
{
if(setWorkingBlocks.find(BB->getName()) != setWorkingBlocks.end() )
{
vecWorkingBlock.push_back(BB);
}
}
errs() << "working block number: " << vecWorkingBlock.size() << "\n";
BasicBlock * pHeader = pLoop->getHeader();
set<BasicBlock *> setExitBlock;
CollectExitBlock(pLoop, setExitBlock);
vector<BasicBlock *> vecAdded;
CreateIfElseBlock(pLoop, vecAdded);
ValueToValueMapTy VMap;
set<BasicBlock *> setCloned;
CloneInnerLoop(pLoop, vecAdded, VMap, setCloned);
//BasicBlock * pPreHeader = vecAdded[0];
BasicBlock * pElseBody = vecAdded[1];
vector<BasicBlock *>::iterator itVecBegin = vecWorkingBlock.begin();
vector<BasicBlock *>::iterator itVecEnd = vecWorkingBlock.end();
for(; itVecBegin != itVecEnd; itVecBegin++ )
{
BasicBlock * pClonedBlock = cast<BasicBlock>(VMap[*itVecBegin]);
pStore = new StoreInst(this->ConstantLong1, pAlloc, false, pClonedBlock->getFirstInsertionPt());
pStore->setAlignment(8);
pClonedBlock->dump();
}
pStore = new StoreInst(this->ConstantLong0, pAlloc, false, pElseBody->getTerminator());
pStore->setAlignment(8);
pLoad0 = new LoadInst(this->numIterations, "", false, pElseBody->getTerminator());
pLoad0->setAlignment(8);
pLoad1 = new LoadInst(this->numWorkingIterations, "", false, pElseBody->getTerminator());
pLoad1->setAlignment(8);
BasicBlock * pClonedHeader = cast<BasicBlock>(VMap[pHeader]);
set<BasicBlock *> setPredBlocks;
for(pred_iterator PI = pred_begin(pClonedHeader), E = pred_end(pClonedHeader); PI != E; ++PI)
{
setPredBlocks.insert(*PI);
}
BasicBlock::iterator itInsert = pClonedHeader->getFirstInsertionPt();
PHINode * pNewIterations = PHINode::Create(pLoad0->getType(), setPredBlocks.size(), "numIterations.2", itInsert);
PHINode * pNewWorkingIterations = PHINode::Create(pLoad1->getType(), setPredBlocks.size(), "WorkingIterations.2", itInsert);
BinaryOperator * pIterationAdd = BinaryOperator::Create(Instruction::Add, pNewIterations, this->ConstantLong1, "Iterations.add.2", itInsert);
set<BasicBlock *>::iterator itSetBegin = setPredBlocks.begin();
set<BasicBlock *>::iterator itSetEnd = setPredBlocks.end();
for(; itSetBegin != itSetEnd; itSetBegin ++ )
{
if((*itSetBegin) == pElseBody)
{
pNewIterations->addIncoming(pLoad0, pElseBody);
}
else
{
pNewIterations->addIncoming(pIterationAdd, *itSetBegin);
}
}
pLoad0 = new LoadInst(pAlloc, "", false, itInsert);
BinaryOperator * pWorkingAdd = BinaryOperator::Create(Instruction::Add, pNewWorkingIterations, pLoad0, "Working.add.2", itInsert);
itSetBegin = setPredBlocks.begin();
itSetEnd = setPredBlocks.end();
for(; itSetBegin != itSetEnd; itSetBegin ++ )
{
if((*itSetBegin) == pElseBody)
{
pNewWorkingIterations->addIncoming(pLoad1, pElseBody);
}
else
{
pNewWorkingIterations->addIncoming(pWorkingAdd, *itSetBegin);
}
}
pStore = new StoreInst(this->ConstantLong0, pAlloc, false, itInsert);
pStore->setAlignment(8);
itSetBegin = setExitBlock.begin();
itSetEnd = setExitBlock.end();
for(; itSetBegin != itSetEnd; itSetBegin ++ )
{
SmallVector<BasicBlock*, 8> LoopBlocks;
for(pred_iterator PI = pred_begin(*itSetBegin), E = pred_end(*itSetBegin); PI != E; ++PI)
{
if(setCloned.find(*PI) != setCloned.end())
{
LoopBlocks.push_back(*PI);
}
}
BasicBlock * NewExitBB = SplitBlockPredecessors(*itSetBegin, LoopBlocks, ".WL.loopexit", this);
pStore = new StoreInst(pIterationAdd, this->numIterations, false, NewExitBB->getFirstInsertionPt());
pStore->setAlignment(8);
pStore = new StoreInst(pWorkingAdd, this->numWorkingIterations, false, NewExitBB->getFirstInsertionPt());
pStore->setAlignment(8);
}
//pFunction->dump();
DominatorTree * DT = &(getAnalysis<DominatorTree>(*pFunction));
vector<AllocaInst *> vecAlloc;
vecAlloc.push_back(pAlloc);
PromoteMemToReg(vecAlloc, *DT);
pFunction->dump();
}
void WorklessInstrument::InstrumentWorkless0Star1(Module * pModule, Loop * pLoop)
{
Function * pMain = NULL;
if(strMainName != "" )
{
pMain = pModule->getFunction(strMainName.c_str());
}
else
{
pMain = pModule->getFunction("main");
}
LoadInst * pLoad;
BinaryOperator* pAdd = NULL;
StoreInst * pStore = NULL;
for (Function::iterator BB = pMain->begin(); BB != pMain->end(); ++BB)
{
if(BB->getName().equals("entry"))
{
CallInst * pCall;
StoreInst * pStore;
Instruction * II = BB->begin();
pCall = CallInst::Create(this->InitHooks, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet emptySet;
pCall->setAttributes(emptySet);
pCall = CallInst::Create(this->getenv, this->SAMPLE_RATE_ptr, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet AS;
{
SmallVector<AttributeSet, 4> Attrs;
AttributeSet PAS;
{
AttrBuilder B;
B.addAttribute(Attribute::NoUnwind);
PAS = AttributeSet::get(pModule->getContext(), ~0U, B);
}
Attrs.push_back(PAS);
AS = AttributeSet::get(pModule->getContext(), Attrs);
}
pCall->setAttributes(AS);
pCall = CallInst::Create(this->function_atoi, pCall, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
{
SmallVector<AttributeSet, 4> Attrs;
AttributeSet PAS;
{
AttrBuilder B;
B.addAttribute(Attribute::NoUnwind);
B.addAttribute(Attribute::ReadOnly);
PAS = AttributeSet::get(pModule->getContext(), ~0U, B);
}
Attrs.push_back(PAS);
AS = AttributeSet::get(pModule->getContext(), Attrs);
}
pCall->setAttributes(AS);
pStore = new StoreInst(pCall, this->SAMPLE_RATE, false, II);
pStore->setAlignment(4);
pCall = CallInst::Create(this->geo, pCall, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
pCall->setAttributes(emptySet);
CastInst * pCast = CastInst::CreateIntegerCast(pCall, this->LongType, true, "", II);
pStore = new StoreInst(pCast, this->CURRENT_SAMPLE, false, II);
pStore->setAlignment(8);
vector<Value *> vecParam;
vecParam.push_back(this->Output_Format_String);
vecParam.push_back(pCall);
pCall = CallInst::Create(this->printf, vecParam, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
pCall->setAttributes(emptySet);
break;
}
}
for (Function::iterator BB = pMain->begin(); BB != pMain->end(); ++BB)
{
for (BasicBlock::iterator Ins = BB->begin(); Ins != BB->end(); ++Ins)
{
if (isa<ReturnInst>(Ins) || isa<ResumeInst>(Ins))
{
vector<Value*> vecParams;
pLoad = new LoadInst(numIterations, "", false, Ins);
pLoad->setAlignment(8);
vecParams.push_back(pLoad);
pLoad = new LoadInst(numInstances, "", false, Ins);
pLoad->setAlignment(8);
vecParams.push_back(pLoad);
CallInst* pCall = CallInst::Create(this->PrintLoopInfo, vecParams, "", Ins);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet aSet;
pCall->setAttributes(aSet);
}
else if(isa<CallInst>(Ins) || isa<InvokeInst>(Ins))
{
CallSite cs(Ins);
Function * pCalled = cs.getCalledFunction();
if(pCalled == NULL)
{
continue;
}
if(pCalled->getName() == "exit" || pCalled->getName() == "_ZL9mysql_endi")
{
vector<Value*> vecParams;
pLoad = new LoadInst(numIterations, "", false, Ins);
pLoad->setAlignment(8);
vecParams.push_back(pLoad);
pLoad = new LoadInst(numInstances, "", false, Ins);
pLoad->setAlignment(8);
vecParams.push_back(pLoad);
CallInst* pCall = CallInst::Create(this->PrintLoopInfo, vecParams, "", Ins);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet aSet;
pCall->setAttributes(aSet);
}
}
}
}
BasicBlock * pHeader = pLoop->getHeader();
set<BasicBlock *> setExitBlock;
CollectExitBlock(pLoop, setExitBlock);
vector<BasicBlock *> vecAdded;
CreateIfElseBlock(pLoop, vecAdded);
ValueToValueMapTy VMap;
set<BasicBlock *> setCloned;
CloneInnerLoop(pLoop, vecAdded, VMap, setCloned);
BasicBlock * pPreHeader = vecAdded[1];
pLoad = new LoadInst(this->numIterations, "", false, pPreHeader->getTerminator());
pLoad->setAlignment(8);
BasicBlock * pClonedHeader = cast<BasicBlock>(VMap[pHeader]);
set<BasicBlock *> setPredBlocks;
for(pred_iterator PI = pred_begin(pClonedHeader), E = pred_end(pClonedHeader); PI != E; ++PI)
{
setPredBlocks.insert(*PI);
}
PHINode * pNew = PHINode::Create(pLoad->getType(), setPredBlocks.size(), "numIterations", pClonedHeader->getFirstInsertionPt());
pAdd = BinaryOperator::Create(Instruction::Add, pNew, this->ConstantLong1, "add", pClonedHeader->getFirstInsertionPt());
set<BasicBlock *>::iterator itSetBegin = setPredBlocks.begin();
set<BasicBlock *>::iterator itSetEnd = setPredBlocks.end();
for(; itSetBegin != itSetEnd; itSetBegin ++ )
{
if((*itSetBegin) == pPreHeader)
{
pNew->addIncoming(pLoad, pPreHeader);
}
else
{
pNew->addIncoming(pAdd, *itSetBegin);
}
}
itSetBegin = setExitBlock.begin();
itSetEnd = setExitBlock.end();
for(; itSetBegin != itSetEnd; itSetBegin ++ )
{
SmallVector<BasicBlock*, 8> LoopBlocks;
for(pred_iterator PI = pred_begin(*itSetBegin), E = pred_end(*itSetBegin); PI != E; ++PI)
{
if(setCloned.find(*PI) != setCloned.end())
{
LoopBlocks.push_back(*PI);
}
}
BasicBlock * NewExitBB = SplitBlockPredecessors(*itSetBegin, LoopBlocks, ".WL.loopexit", this);
pStore = new StoreInst(pAdd, this->numIterations, false, NewExitBB->getFirstInsertionPt());
pStore->setAlignment(8);
}
pPreHeader->getParent()->dump();
}
/// Replaces the given call site (Call or Invoke) with a gc.statepoint
/// intrinsic with an empty deoptimization arguments list. This does
/// NOT do explicit relocation for GC support.
static Value *ReplaceWithStatepoint(const CallSite &CS, /* to replace */
Pass *P) {
BasicBlock *BB = CS.getInstruction()->getParent();
Function *F = BB->getParent();
Module *M = F->getParent();
assert(M && "must be set");
// TODO: technically, a pass is not allowed to get functions from within a
// function pass since it might trigger a new function addition. Refactor
// this logic out to the initialization of the pass. Doesn't appear to
// matter in practice.
// Fill in the one generic type'd argument (the function is also vararg)
std::vector<Type *> argTypes;
argTypes.push_back(CS.getCalledValue()->getType());
Function *gc_statepoint_decl = Intrinsic::getDeclaration(
M, Intrinsic::experimental_gc_statepoint, argTypes);
// Then go ahead and use the builder do actually do the inserts. We insert
// immediately before the previous instruction under the assumption that all
// arguments will be available here. We can't insert afterwards since we may
// be replacing a terminator.
Instruction *insertBefore = CS.getInstruction();
IRBuilder<> Builder(insertBefore);
// First, create the statepoint (with all live ptrs as arguments).
std::vector<llvm::Value *> args;
// target, #args, unused, args
Value *Target = CS.getCalledValue();
args.push_back(Target);
int callArgSize = CS.arg_size();
args.push_back(
ConstantInt::get(Type::getInt32Ty(M->getContext()), callArgSize));
// TODO: add a 'Needs GC-rewrite' later flag
args.push_back(ConstantInt::get(Type::getInt32Ty(M->getContext()), 0));
// Copy all the arguments of the original call
args.insert(args.end(), CS.arg_begin(), CS.arg_end());
// Create the statepoint given all the arguments
Instruction *token = nullptr;
AttributeSet return_attributes;
if (CS.isCall()) {
CallInst *toReplace = cast<CallInst>(CS.getInstruction());
CallInst *call =
Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
call->setTailCall(toReplace->isTailCall());
call->setCallingConv(toReplace->getCallingConv());
// Before we have to worry about GC semantics, all attributes are legal
AttributeSet new_attrs = toReplace->getAttributes();
// In case if we can handle this set of sttributes - set up function attrs
// directly on statepoint and return attrs later for gc_result intrinsic.
call->setAttributes(new_attrs.getFnAttributes());
return_attributes = new_attrs.getRetAttributes();
// TODO: handle param attributes
token = call;
// Put the following gc_result and gc_relocate calls immediately after the
// the old call (which we're about to delete)
BasicBlock::iterator next(toReplace);
assert(BB->end() != next && "not a terminator, must have next");
next++;
Instruction *IP = &*(next);
Builder.SetInsertPoint(IP);
Builder.SetCurrentDebugLocation(IP->getDebugLoc());
} else if (CS.isInvoke()) {
InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
// Insert the new invoke into the old block. We'll remove the old one in a
// moment at which point this will become the new terminator for the
// original block.
InvokeInst *invoke = InvokeInst::Create(
gc_statepoint_decl, toReplace->getNormalDest(),
toReplace->getUnwindDest(), args, "", toReplace->getParent());
invoke->setCallingConv(toReplace->getCallingConv());
// Currently we will fail on parameter attributes and on certain
// function attributes.
AttributeSet new_attrs = toReplace->getAttributes();
// In case if we can handle this set of sttributes - set up function attrs
// directly on statepoint and return attrs later for gc_result intrinsic.
invoke->setAttributes(new_attrs.getFnAttributes());
return_attributes = new_attrs.getRetAttributes();
token = invoke;
// We'll insert the gc.result into the normal block
BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
toReplace->getNormalDest(), invoke->getParent());
Instruction *IP = &*(normalDest->getFirstInsertionPt());
Builder.SetInsertPoint(IP);
} else {
llvm_unreachable("unexpect type of CallSite");
}
assert(token);
// Handle the return value of the original call - update all uses to use a
// gc_result hanging off the statepoint node we just inserted
// Only add the gc_result iff there is actually a used result
if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
Instruction *gc_result = nullptr;
std::vector<Type *> types; // one per 'any' type
types.push_back(CS.getType()); // result type
auto get_gc_result_id = [&](Type &Ty) {
if (Ty.isIntegerTy()) {
return Intrinsic::experimental_gc_result_int;
} else if (Ty.isFloatingPointTy()) {
return Intrinsic::experimental_gc_result_float;
} else if (Ty.isPointerTy()) {
return Intrinsic::experimental_gc_result_ptr;
} else {
llvm_unreachable("non java type encountered");
}
};
Intrinsic::ID Id = get_gc_result_id(*CS.getType());
Value *gc_result_func = Intrinsic::getDeclaration(M, Id, types);
std::vector<Value *> args;
args.push_back(token);
gc_result = Builder.CreateCall(
gc_result_func, args,
CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "");
cast<CallInst>(gc_result)->setAttributes(return_attributes);
return gc_result;
} else {
// No return value for the call.
return nullptr;
}
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Clone functions that take GEPs as arguments
//
// Inputs:
// M - A reference to the LLVM module to transform
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool GEPExprArgs::runOnModule(Module& M) {
bool changed;
do {
changed = false;
for (Module::iterator F = M.begin(); F != M.end(); ++F){
for (Function::iterator B = F->begin(), FE = F->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
CallInst *CI = dyn_cast<CallInst>(I++);
if(!CI)
continue;
if(CI->hasByValArgument())
continue;
// if the GEP calls a function, that is externally defined,
// or might be changed, ignore this call site.
Function *F = CI->getCalledFunction();
if (!F || (F->isDeclaration() || F->mayBeOverridden()))
continue;
if(F->hasStructRetAttr())
continue;
if(F->isVarArg())
continue;
// find the argument we must replace
Function::arg_iterator ai = F->arg_begin(), ae = F->arg_end();
unsigned argNum = 1;
for(; argNum < CI->getNumOperands();argNum++, ++ai) {
if(ai->use_empty())
continue;
if (isa<GEPOperator>(CI->getOperand(argNum)))
break;
}
// if no argument was a GEP operator to be changed
if(ai == ae)
continue;
GEPOperator *GEP = dyn_cast<GEPOperator>(CI->getOperand(argNum));
if(!GEP->hasAllConstantIndices())
continue;
// Construct the new Type
// Appends the struct Type at the beginning
std::vector<Type*>TP;
TP.push_back(GEP->getPointerOperand()->getType());
for(unsigned c = 1; c < CI->getNumOperands();c++) {
TP.push_back(CI->getOperand(c)->getType());
}
//return type is same as that of original instruction
FunctionType *NewFTy = FunctionType::get(CI->getType(), TP, false);
Function *NewF;
numSimplified++;
if(numSimplified > 800)
return true;
NewF = Function::Create(NewFTy,
GlobalValue::InternalLinkage,
F->getName().str() + ".TEST",
&M);
Function::arg_iterator NI = NewF->arg_begin();
NI->setName("GEParg");
++NI;
ValueToValueMapTy ValueMap;
for (Function::arg_iterator II = F->arg_begin(); NI != NewF->arg_end(); ++II, ++NI) {
ValueMap[II] = NI;
NI->setName(II->getName());
NI->addAttr(F->getAttributes().getParamAttributes(II->getArgNo() + 1));
}
NewF->setAttributes(NewF->getAttributes().addAttr(
0, F->getAttributes().getRetAttributes()));
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
CloneFunctionInto(NewF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NewF->arg_begin(),
ae= NewF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NewF->setAttributes(NewF->getAttributes().addAttr(
~0, F->getAttributes().getFnAttributes()));
//Get the point to insert the GEP instr.
SmallVector<Value*, 8> Ops(CI->op_begin()+1, CI->op_end());
Instruction *InsertPoint;
for (BasicBlock::iterator insrt = NewF->front().begin();
isa<AllocaInst>(InsertPoint = insrt); ++insrt) {;}
NI = NewF->arg_begin();
SmallVector<Value*, 8> Indices;
Indices.append(GEP->op_begin()+1, GEP->op_end());
GetElementPtrInst *GEP_new = GetElementPtrInst::Create(cast<Value>(NI),
Indices,
"", InsertPoint);
fargs.at(argNum)->replaceAllUsesWith(GEP_new);
unsigned j = argNum + 1;
for(; j < CI->getNumOperands();j++) {
if(CI->getOperand(j) == GEP)
fargs.at(j)->replaceAllUsesWith(GEP_new);
}
SmallVector<AttributeWithIndex, 8> AttributesVec;
// Get the initial attributes of the call
AttrListPtr CallPAL = CI->getAttributes();
Attributes RAttrs = CallPAL.getRetAttributes();
Attributes FnAttrs = CallPAL.getFnAttributes();
if (RAttrs)
AttributesVec.push_back(AttributeWithIndex::get(0, RAttrs));
SmallVector<Value*, 8> Args;
Args.push_back(GEP->getPointerOperand());
for(unsigned j =1;j<CI->getNumOperands();j++) {
Args.push_back(CI->getOperand(j));
// position in the AttributesVec
if (Attributes Attrs = CallPAL.getParamAttributes(j))
AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
}
// Create the new attributes vec.
if (FnAttrs != Attribute::None)
AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
AttrListPtr NewCallPAL = AttrListPtr::get(AttributesVec.begin(),
AttributesVec.end());
CallInst *CallI = CallInst::Create(NewF,Args,"", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
CI->replaceAllUsesWith(CallI);
CI->eraseFromParent();
changed = true;
}
}
}
} while(changed);
return true;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// 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;
}
/// Replaces the given call site (Call or Invoke) with a gc.statepoint
/// intrinsic with an empty deoptimization arguments list. This does
/// NOT do explicit relocation for GC support.
static Value *ReplaceWithStatepoint(const CallSite &CS /* to replace */) {
assert(CS.getInstruction()->getModule() && "must be set");
// TODO: technically, a pass is not allowed to get functions from within a
// function pass since it might trigger a new function addition. Refactor
// this logic out to the initialization of the pass. Doesn't appear to
// matter in practice.
// Then go ahead and use the builder do actually do the inserts. We insert
// immediately before the previous instruction under the assumption that all
// arguments will be available here. We can't insert afterwards since we may
// be replacing a terminator.
IRBuilder<> Builder(CS.getInstruction());
// Note: The gc args are not filled in at this time, that's handled by
// RewriteStatepointsForGC (which is currently under review).
// Create the statepoint given all the arguments
Instruction *Token = nullptr;
uint64_t ID;
uint32_t NumPatchBytes;
AttributeSet OriginalAttrs = CS.getAttributes();
Attribute AttrID =
OriginalAttrs.getAttribute(AttributeSet::FunctionIndex, "statepoint-id");
Attribute AttrNumPatchBytes = OriginalAttrs.getAttribute(
AttributeSet::FunctionIndex, "statepoint-num-patch-bytes");
AttrBuilder AttrsToRemove;
bool HasID = AttrID.isStringAttribute() &&
!AttrID.getValueAsString().getAsInteger(10, ID);
if (HasID)
AttrsToRemove.addAttribute("statepoint-id");
else
ID = 0xABCDEF00;
bool HasNumPatchBytes =
AttrNumPatchBytes.isStringAttribute() &&
!AttrNumPatchBytes.getValueAsString().getAsInteger(10, NumPatchBytes);
if (HasNumPatchBytes)
AttrsToRemove.addAttribute("statepoint-num-patch-bytes");
else
NumPatchBytes = 0;
OriginalAttrs = OriginalAttrs.removeAttributes(
CS.getInstruction()->getContext(), AttributeSet::FunctionIndex,
AttrsToRemove);
if (CS.isCall()) {
CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
CallInst *Call = Builder.CreateGCStatepointCall(
ID, NumPatchBytes, CS.getCalledValue(),
makeArrayRef(CS.arg_begin(), CS.arg_end()), None, None,
"safepoint_token");
Call->setTailCall(ToReplace->isTailCall());
Call->setCallingConv(ToReplace->getCallingConv());
// In case if we can handle this set of attributes - set up function
// attributes directly on statepoint and return attributes later for
// gc_result intrinsic.
Call->setAttributes(OriginalAttrs.getFnAttributes());
Token = Call;
// Put the following gc_result and gc_relocate calls immediately after
// the old call (which we're about to delete).
assert(ToReplace->getNextNode() && "not a terminator, must have next");
Builder.SetInsertPoint(ToReplace->getNextNode());
Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
} else if (CS.isInvoke()) {
InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
// Insert the new invoke into the old block. We'll remove the old one in a
// moment at which point this will become the new terminator for the
// original block.
Builder.SetInsertPoint(ToReplace->getParent());
InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
ID, NumPatchBytes, CS.getCalledValue(), ToReplace->getNormalDest(),
ToReplace->getUnwindDest(), makeArrayRef(CS.arg_begin(), CS.arg_end()),
None, None, "safepoint_token");
Invoke->setCallingConv(ToReplace->getCallingConv());
// In case if we can handle this set of attributes - set up function
// attributes directly on statepoint and return attributes later for
// gc_result intrinsic.
Invoke->setAttributes(OriginalAttrs.getFnAttributes());
Token = Invoke;
// We'll insert the gc.result into the normal block
BasicBlock *NormalDest = ToReplace->getNormalDest();
// Can not insert gc.result in case of phi nodes preset.
// Should have removed this cases prior to running this function
assert(!isa<PHINode>(NormalDest->begin()));
Instruction *IP = &*(NormalDest->getFirstInsertionPt());
Builder.SetInsertPoint(IP);
} else {
llvm_unreachable("unexpect type of CallSite");
}
assert(Token);
// Handle the return value of the original call - update all uses to use a
// gc_result hanging off the statepoint node we just inserted
// Only add the gc_result iff there is actually a used result
if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
std::string TakenName =
CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), TakenName);
GCResult->setAttributes(OriginalAttrs.getRetAttributes());
return GCResult;
} else {
// No return value for the call.
return nullptr;
}
}
// Creates the helper function that will do the setjmp() call and
// function call for implementing Invoke. Creates the call to the
// helper function. Returns a Value which is zero on the normal
// execution path and non-zero if the landingpad block should be
// entered.
Value *FuncRewriter::createSetjmpWrappedCall(InvokeInst *Invoke) {
Type *I32 = Type::getInt32Ty(Func->getContext());
// Allocate space for storing the invoke's result temporarily (so
// that the helper function can return multiple values). We don't
// need to do this if the result is unused, and we can't if its type
// is void.
Instruction *ResultAlloca = NULL;
if (!Invoke->use_empty()) {
ResultAlloca = new AllocaInst(Invoke->getType(), "invoke_result_ptr");
Func->getEntryBlock().getInstList().push_front(ResultAlloca);
}
// Create type for the helper function.
SmallVector<Type *, 10> ArgTypes;
for (unsigned I = 0, E = Invoke->getNumArgOperands(); I < E; ++I)
ArgTypes.push_back(Invoke->getArgOperand(I)->getType());
ArgTypes.push_back(Invoke->getCalledValue()->getType());
ArgTypes.push_back(FrameJmpBuf->getType());
if (ResultAlloca)
ArgTypes.push_back(Invoke->getType()->getPointerTo());
FunctionType *FTy = FunctionType::get(I32, ArgTypes, false);
// Create the helper function.
Function *HelperFunc = Function::Create(
FTy, GlobalValue::InternalLinkage, Func->getName() + "_setjmp_caller");
Func->getParent()->getFunctionList().insertAfter(Func, HelperFunc);
BasicBlock *EntryBB = BasicBlock::Create(Func->getContext(), "", HelperFunc);
BasicBlock *NormalBB = BasicBlock::Create(Func->getContext(), "normal",
HelperFunc);
BasicBlock *ExceptionBB = BasicBlock::Create(Func->getContext(), "exception",
HelperFunc);
// Unpack the helper function's arguments.
Function::arg_iterator ArgIter = HelperFunc->arg_begin();
SmallVector<Value *, 10> InnerCallArgs;
for (unsigned I = 0, E = Invoke->getNumArgOperands(); I < E; ++I) {
ArgIter->setName("arg");
InnerCallArgs.push_back(ArgIter++);
}
Argument *CalleeArg = ArgIter++;
Argument *JmpBufArg = ArgIter++;
CalleeArg->setName("func_ptr");
JmpBufArg->setName("jmp_buf");
// Create setjmp() call.
Value *SetjmpArgs[] = { JmpBufArg };
CallInst *SetjmpCall = CallInst::Create(SetjmpIntrinsic, SetjmpArgs,
"invoke_sj", EntryBB);
CopyDebug(SetjmpCall, Invoke);
// Setting the "returns_twice" attribute here prevents optimization
// passes from inlining HelperFunc into its caller.
SetjmpCall->setCanReturnTwice();
// Check setjmp()'s result.
Value *IsZero = CopyDebug(new ICmpInst(*EntryBB, CmpInst::ICMP_EQ, SetjmpCall,
ConstantInt::get(I32, 0),
"invoke_sj_is_zero"), Invoke);
CopyDebug(BranchInst::Create(NormalBB, ExceptionBB, IsZero, EntryBB), Invoke);
// Handle the normal, non-exceptional code path.
CallInst *InnerCall = CallInst::Create(CalleeArg, InnerCallArgs, "",
NormalBB);
CopyDebug(InnerCall, Invoke);
InnerCall->setAttributes(Invoke->getAttributes());
InnerCall->setCallingConv(Invoke->getCallingConv());
if (ResultAlloca) {
InnerCall->setName("result");
Argument *ResultArg = ArgIter++;
ResultArg->setName("result_ptr");
CopyDebug(new StoreInst(InnerCall, ResultArg, NormalBB), Invoke);
}
ReturnInst::Create(Func->getContext(), ConstantInt::get(I32, 0), NormalBB);
// Handle the exceptional code path.
ReturnInst::Create(Func->getContext(), ConstantInt::get(I32, 1), ExceptionBB);
// Create the outer call to the helper function.
SmallVector<Value *, 10> OuterCallArgs;
for (unsigned I = 0, E = Invoke->getNumArgOperands(); I < E; ++I)
OuterCallArgs.push_back(Invoke->getArgOperand(I));
OuterCallArgs.push_back(Invoke->getCalledValue());
OuterCallArgs.push_back(FrameJmpBuf);
if (ResultAlloca)
OuterCallArgs.push_back(ResultAlloca);
CallInst *OuterCall = CallInst::Create(HelperFunc, OuterCallArgs,
"invoke_is_exc", Invoke);
CopyDebug(OuterCall, Invoke);
// Retrieve the function return value stored in the alloca. We only
// need to do this on the non-exceptional path, but we currently do
// it unconditionally because that is simpler.
if (ResultAlloca) {
Value *Result = new LoadInst(ResultAlloca, "", Invoke);
Result->takeName(Invoke);
Invoke->replaceAllUsesWith(Result);
}
return OuterCall;
}
