bool InlineMalloc::runOnFunction(Function& F) {
Function* Malloc = F.getParent()->getFunction("gcmalloc");
if (!Malloc || Malloc->isDeclaration()) return false;
bool Changed = false;
for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; BI++) {
BasicBlock *Cur = BI;
for (BasicBlock::iterator II = Cur->begin(), IE = Cur->end(); II != IE;) {
Instruction *I = II;
II++;
CallSite Call = CallSite::get(I);
Instruction* CI = Call.getInstruction();
if (CI) {
Function* Temp = Call.getCalledFunction();
if (Temp == Malloc) {
if (dyn_cast<Constant>(Call.getArgument(0))) {
InlineFunctionInfo IFI(NULL, mvm::MvmModule::TheTargetData);
Changed |= InlineFunction(Call, IFI);
break;
}
}
}
}
}
return Changed;
}
void inlineFunctionCalls(Function* f, TargetData* targetData) {
//WFVOPENCL_DEBUG( outs() << " inlining function calls... "; );
bool functionChanged = true;
while (functionChanged) {
functionChanged = false;
for (Function::iterator BB=f->begin(); BB!=f->end(); ++BB) {
bool blockChanged = false;
for (BasicBlock::iterator I=BB->begin(); !blockChanged && I!=BB->end();) {
if (!isa<CallInst>(I)) { ++I; continue; }
CallInst* call = cast<CallInst>(I++);
Function* callee = call->getCalledFunction();
if (!callee) {
//errs() << "ERROR: could not inline function call: " << *call;
continue;
}
if (callee->getAttributes().hasAttrSomewhere(Attribute::NoInline)) {
//WFVOPENCL_DEBUG( outs() << " function '" << callee->getNameStr() << "' has attribute 'no inline', ignored call!\n"; );
continue;
}
const std::string calleeName = callee->getNameStr(); //possibly deleted by InlineFunction()
InlineFunctionInfo IFI(NULL, targetData);
blockChanged = InlineFunction(call, IFI);
//WFVOPENCL_DEBUG(
//if (blockChanged) outs() << " inlined call to function '" << calleeName << "'\n";
//else errs() << " inlining of call to function '" << calleeName << "' FAILED!\n";
//);
functionChanged |= blockChanged;
}
}
}
//WFVOPENCL_DEBUG( outs() << "done.\n"; );
}
// =============================================================================
// andOOPIsGone (formerly: createProcess)
//
// Formerly, OOP permitted the same SC_{METHOD,THREAD} functions to apply
// to each copy of a SC_MODULE. Aaaaand it's gone !
// (but OTOH we enable better optimizations)
// Creates a new C-style function that calls the old member function with the
// given sc_module. The call is then inlined.
// FIXME: assumes the method is non-virtual and that sc_module is the first
// inherited class of the SC_MODULE
// =============================================================================
Function *TwetoPassImpl::andOOPIsGone(Function * oldProc,
sc_core::sc_module * initiatorMod)
{
if (!oldProc)
return NULL;
// can't statically optimize if the address of the module isn't predictible
// TODO: also handle already-static variables, which also have
// fixed $pc-relative addresses
if (staticopt == optlevel && !permalloc::is_from (initiatorMod))
return NULL;
LLVMContext & context = getGlobalContext();
FunctionType *funType = oldProc->getFunctionType();
Type *type = funType->getParamType(0);
FunctionType *newProcType =
FunctionType::get(oldProc->getReturnType(),
ArrayRef < Type * >(), false);
// Create the new function
std::ostringstream id;
id << proc_counter++;
std::string name =
oldProc->getName().str() + std::string("_clone_") + id.str();
Function *newProc =
Function::Create(newProcType, Function::ExternalLinkage, name,
this->llvmMod);
assert(newProc->empty());
newProc->addFnAttr(Attribute::InlineHint);
// Create call to old function
BasicBlock *bb = BasicBlock::Create(context, "entry", newProc);
IRBuilder <> *irb = new IRBuilder <> (context);
irb->SetInsertPoint(bb);
Value* thisAddr = createRelocatablePointer (type, initiatorMod, irb);
CallInst *ci = irb->CreateCall(oldProc,
ArrayRef < Value * >(std::vector<Value*>(1,thisAddr)));
//bb->getInstList().insert(ci, thisAddr);
if (ci->getType()->isVoidTy())
irb->CreateRetVoid();
else
irb->CreateRet(ci);
// The function should be valid now
verifyFunction(*newProc);
{ // Inline the call
DataLayout *td = new DataLayout(this->llvmMod);
InlineFunctionInfo i(NULL, td);
bool success = InlineFunction(ci, i);
assert(success);
verifyFunction(*newProc);
}
// further optimize the function
inlineBasicIO (initiatorMod, newProc);
newProc->dump();
return newProc;
}
void GNUstep::IMPCacher::SpeculativelyInline(Instruction *call, Function
*function) {
BasicBlock *beforeCallBB = call->getParent();
BasicBlock *callBB = SplitBlock(beforeCallBB, call, Owner);
BasicBlock *inlineBB = BasicBlock::Create(Context, "inline",
callBB->getParent());
BasicBlock::iterator iter = call;
iter++;
BasicBlock *afterCallBB = SplitBlock(iter->getParent(), iter, Owner);
removeTerminator(beforeCallBB);
// Put a branch before the call, testing whether the callee really is the
// function
IRBuilder<> B = IRBuilder<>(beforeCallBB);
Value *callee = isa<CallInst>(call) ? cast<CallInst>(call)->getCalledValue()
: cast<InvokeInst>(call)->getCalledValue();
const FunctionType *FTy = function->getFunctionType();
const FunctionType *calleeTy = cast<FunctionType>(
cast<PointerType>(callee->getType())->getElementType());
if (calleeTy != FTy) {
callee = B.CreateBitCast(callee, function->getType());
}
Value *isInlineValid = B.CreateICmpEQ(callee, function);
B.CreateCondBr(isInlineValid, inlineBB, callBB);
// In the inline BB, add a copy of the call, but this time calling the real
// version.
Instruction *inlineCall = call->clone();
Value *inlineResult= inlineCall;
inlineBB->getInstList().push_back(inlineCall);
B.SetInsertPoint(inlineBB);
if (calleeTy != FTy) {
for (unsigned i=0 ; i<FTy->getNumParams() ; i++) {
LLVMType *callType = calleeTy->getParamType(i);
LLVMType *argType = FTy->getParamType(i);
if (callType != argType) {
inlineCall->setOperand(i, new
BitCastInst(inlineCall->getOperand(i), argType, "", inlineCall));
}
}
if (FTy->getReturnType() != calleeTy->getReturnType()) {
if (FTy->getReturnType() == Type::getVoidTy(Context)) {
inlineResult = Constant::getNullValue(calleeTy->getReturnType());
} else {
inlineResult =
new BitCastInst(inlineCall, calleeTy->getReturnType(), "", inlineBB);
}
}
}
B.CreateBr(afterCallBB);
// Unify the return values
if (call->getType() != Type::getVoidTy(Context)) {
PHINode *phi = CreatePHI(call->getType(), 2, "", afterCallBB->begin());
call->replaceAllUsesWith(phi);
phi->addIncoming(call, callBB);
phi->addIncoming(inlineResult, inlineBB);
}
// Really do the real inlining
InlineFunctionInfo IFI(0, 0);
if (CallInst *c = dyn_cast<CallInst>(inlineCall)) {
c->setCalledFunction(function);
InlineFunction(c, IFI);
} else if (InvokeInst *c = dyn_cast<InvokeInst>(inlineCall)) {
c->setCalledFunction(function);
InlineFunction(c, IFI);
}
}
Function* generateFunctionWrapperWithParams(const std::string& wrapper_name, Function* f, Module* mod, std::vector<const Type*>& additionalParams, const bool inlineCall) {
assert (f && mod);
assert (f->getParent());
if (f->getParent() != mod) {
errs() << "WARNING: generateFunctionWrapper(): module '"
<< mod->getModuleIdentifier()
<< "' is not the parent of function '"
<< f->getNameStr() << "' (parent: '"
<< f->getParent()->getModuleIdentifier() << "')!\n";
}
// first make sure there is no function with that name in mod
// TODO: Implement logic from LLVM tutorial that checks for matching extern
// declaration without body and, in this case, goes on.
if (mod->getFunction(wrapper_name)) {
errs() << "ERROR: generateFunctionWrapper(): Function with name '"
<< wrapper_name << "' already exists in module '"
<< mod->getModuleIdentifier() << "'!\n";
return NULL;
}
// warn if f has a return value
if (!f->getReturnType()->isVoidTy()) {
errs() << "WARNING: generateFunctionWrapper(): target function '"
<< f->getNameStr() << "' must not have a return type (ignored)!\n";
}
LLVMContext& context = mod->getContext();
IRBuilder<> builder(context);
// determine all arguments of f
std::vector<const Argument*> oldArgs;
std::vector<const Type*> oldArgTypes;
for (Function::const_arg_iterator A=f->arg_begin(), AE=f->arg_end(); A!=AE; ++A) {
oldArgs.push_back(A);
oldArgTypes.push_back(A->getType());
}
// create a struct type with a member for each argument
const StructType* argStructType = StructType::get(context, oldArgTypes, false);
// create function
//const FunctionType* fType = TypeBuilder<void(void*), true>::get(context);
std::vector<const Type*> params;
params.push_back(PointerType::getUnqual(argStructType));
for (std::vector<const Type*>::const_iterator it=additionalParams.begin(), E=additionalParams.end(); it!=E; ++it) {
params.push_back(*it);
}
const FunctionType* fType = FunctionType::get(Type::getVoidTy(context), params, false);
Function* wrapper = Function::Create(fType, Function::ExternalLinkage, wrapper_name, mod);
// set name of argument
Argument* arg_str = wrapper->arg_begin();
arg_str->setName("arg_struct");
// create entry block
BasicBlock* entryBB = BasicBlock::Create(context, "entry", wrapper);
builder.SetInsertPoint(entryBB);
// create extractions of arguments out of the struct
SmallVector<Value*, 8> extractedArgs;
for (unsigned i=0, e=oldArgTypes.size(); i<e; ++i) {
// create GEP
std::vector<Value*> indices;
indices.push_back(Constant::getNullValue(Type::getInt32Ty(context))); // step through pointer
indices.push_back(ConstantInt::get(context, APInt(32, i))); // index of argument
Value* gep = builder.CreateGEP(arg_str, indices.begin(), indices.end(), "");
// create load
LoadInst* load = builder.CreateLoad(gep, false, "");
// store as argument for call to f
extractedArgs.push_back(load);
}
// create the call to f
CallInst* call = builder.CreateCall(f, extractedArgs.begin(), extractedArgs.end(), "");
// the function returns void
builder.CreateRetVoid();
//wrapper->addAttribute(0, Attribute::NoUnwind); // function does not unwind stack -> why is there an index required ???
wrapper->setDoesNotCapture(1, true); // arg ptr does not capture
wrapper->setDoesNotAlias(1, true); // arg ptr does not alias
// inline call if required
if (inlineCall) {
InlineFunctionInfo IFI(NULL, new TargetData(mod));
const bool success = InlineFunction(call, IFI);
if (!success) {
errs() << "WARNING: could not inline function call inside wrapper: " << *call << "\n";
}
assert (success);
}
//verifyFunction(*wrapper, NULL);
return wrapper;
}
// =============================================================================
// createProcess
//
// Create a new function that contains a call to the old function.
// We inline the call in order to clone the old function's implementation.
// =============================================================================
Function *TLMBasicPassImpl::createProcess(Function *oldProc,
sc_core::sc_module *initiatorMod) {
LLVMContext &context = getGlobalContext();
IntegerType *intType;
if (this->is64Bit) {
intType = Type::getInt64Ty(context);
} else {
intType = Type::getInt32Ty(context);
}
// Retrieve a pointer to the initiator module
ConstantInt *initiatorModVal =
ConstantInt::getSigned(intType,reinterpret_cast<intptr_t>(initiatorMod));
FunctionType *funType = oldProc->getFunctionType();
Type *type = funType->getParamType(0);
IntToPtrInst *thisAddr =
new IntToPtrInst(initiatorModVal, type, "");
// Compute the type of the new function
FunctionType *oldProcType = oldProc->getFunctionType();
Value **argsBegin = new Value*[1];
Value **argsEnd = argsBegin;
*argsEnd++ = thisAddr;
const unsigned argsSize = argsEnd-argsBegin;
Value **args = argsBegin;
assert(oldProcType->getNumParams()==argsSize);
assert(!oldProc->isDeclaration());
std::vector<Type*> argTypes;
for (unsigned i = 0; i!=argsSize; ++i)
argTypes.push_back(oldProcType->getParamType(i));
FunctionType *newProcType =
FunctionType::get(oldProc->getReturnType(), ArrayRef<Type*>(argTypes), false);
// Create the new function
std::ostringstream id;
id << proc_counter++;
std::string name = oldProc->getName().str()+std::string("_clone_")+id.str();
Function *newProc =
Function::Create(newProcType, Function::ExternalLinkage, name, this->llvmMod);
assert(newProc->empty());
newProc->addFnAttr(Attributes::InlineHint);
{ // Set name of newfunc arguments and complete args
Function::arg_iterator nai = newProc->arg_begin();
Function::arg_iterator oai = oldProc->arg_begin();
for (unsigned i = 0; i!=argsSize; ++i, ++oai) {
nai->setName(oai->getName());
args[i] = nai;
++nai;
}
assert(nai==newProc->arg_end());
assert(oai==oldProc->arg_end());
}
// Create call to old function
BasicBlock *bb = BasicBlock::Create(context, "entry", newProc);
IRBuilder<> *irb = new IRBuilder<>(context);
irb->SetInsertPoint(bb);
CallInst *ci = irb->CreateCall(oldProc, ArrayRef<Value*>(argsBegin, argsEnd));
bb->getInstList().insert(ci, thisAddr);
if (ci->getType()->isVoidTy())
irb->CreateRetVoid();
else
irb->CreateRet(ci);
// The function should be valid now
verifyFunction(*newProc);
{ // Inline the call
DataLayout *td = new DataLayout(this->llvmMod);
InlineFunctionInfo i(NULL, td);
bool success = InlineFunction(ci, i);
assert(success);
verifyFunction(*newProc);
}
//newProc->dump();
return newProc;
}
// =============================================================================
// replaceCallsInProcess
//
// Replace indirect calls to write() or read() by direct calls
// in the given process.
// =============================================================================
void TLMBasicPassImpl::replaceCallsInProcess(sc_core::sc_module *initiatorMod,
sc_core::sc_process_b *proc) {
// Get associate function
std::string fctName = proc->func_process;
std::string modType = typeid(*initiatorMod).name();
std::string mainFctName = "_ZN" + modType +
utostr(fctName.size()) + fctName + "Ev";
Function *oldProcf = this->llvmMod->getFunction(mainFctName);
if (oldProcf==NULL)
return;
// We do not modifie the original function
// Instead, we create a clone.
Function *procf = createProcess(oldProcf, initiatorMod);
void *funPtr = this->engine->getPointerToFunction(procf);
sc_core::SC_ENTRY_FUNC_OPT scfun =
reinterpret_cast<sc_core::SC_ENTRY_FUNC_OPT>(funPtr);
proc->m_semantics_p = scfun;
std::string procfName = procf->getName();
MSG(" Replace in the process's function : "+procfName+"\n");
std::ostringstream oss;
sc_core::sc_module *targetMod;
std::vector<CallInfo*> *work = new std::vector<CallInfo*>;
inst_iterator ii;
for (ii = inst_begin(procf); ii!=inst_end(procf); ii++) {
Instruction &i = *ii;
CallSite cs(&i);
if (cs.getInstruction()) {
// Candidate for a replacement
Function *oldfun = cs.getCalledFunction();
if (oldfun!=NULL && !oldfun->isDeclaration()) {
std::string name = oldfun->getName();
// === Write ===
if (!strcmp(name.c_str(), wFunName.c_str())) {
CallInfo *info = new CallInfo();
info->oldcall = dyn_cast<CallInst>(cs.getInstruction());
MSG(" Checking adress : ");
// Retrieve the adress argument by executing
// the appropriated piece of code
SCJit *scjit = new SCJit(this->llvmMod, this->elab);
Process *irProc = this->elab->getProcess(proc);
scjit->setCurrentProcess(irProc);
bool jitErr = false;
info->addrArg = cs.getArgument(1);
int value =
scjit->jitInt(procf, info->oldcall, info->addrArg, &jitErr);
if(jitErr) {
std::cout << " cannot get the address value!"
<< std::endl;
} else {
oss.str(""); oss << std::hex << value;
MSG("0x"+oss.str()+"\n");
basic::addr_t a = static_cast<basic::addr_t>(value);
// Checking address alignment
if(value % sizeof(basic::data_t)) {
std::cerr << " unaligned write : " <<
std::hex << value << std::endl;
abort();
}
// Retreive the target module using the address
targetMod = getTargetModule(initiatorMod, a);
// Save informations to build a new call later
FunctionType *writeFunType =
this->basicWriteFun->getFunctionType();
info->targetType = writeFunType->getParamType(0);
LLVMContext &context = getGlobalContext();
IntegerType *intType;
if (this->is64Bit) {
intType = Type::getInt64Ty(context);
} else {
intType = Type::getInt32Ty(context);
}
info->targetModVal = ConstantInt::getSigned(intType,
reinterpret_cast<intptr_t>(targetMod));
info->dataArg = cs.getArgument(2);
work->push_back(info);
}
} else
// === Read ===
if (!strcmp(name.c_str(), rFunName.c_str())) {
// Not yet supported
}
}
}
}
// Before
//procf->dump();
// Replace calls
std::vector<CallInfo*>::iterator it;
for (it = work->begin(); it!=work->end(); ++it) {
CallInfo *i = *it;
LLVMContext &context = getGlobalContext();
FunctionType *writeFunType =
this->writeFun->getFunctionType();
IntegerType *i64 = Type::getInt64Ty(context);
// Get a pointer to the target module
basic::target_module_base *tmb =
dynamic_cast<basic::target_module_base*>(targetMod);
Value *ptr =
ConstantInt::getSigned(i64, reinterpret_cast<intptr_t>(tmb));
IntToPtrInst *modPtr = new IntToPtrInst(ptr,
writeFunType->getParamType(0),
"myitp", i->oldcall);
// Get a the address value
LoadInst *addr = new LoadInst(i->addrArg, "", i->oldcall);
// Create the new call
Value *args[] = {modPtr, addr, i->dataArg};
i->newcall = CallInst::Create(this->writeFun, ArrayRef<Value*>(args, 3));
// Replace the old call
BasicBlock::iterator it(i->oldcall);
ReplaceInstWithInst(i->oldcall->getParent()->getInstList(), it, i->newcall);
i->oldcall->replaceAllUsesWith(i->newcall);
// Inline the new call
DataLayout *td = new DataLayout(this->llvmMod);
InlineFunctionInfo ifi(NULL, td);
bool success = InlineFunction(i->newcall, ifi);
if(!success) {
MSG(" The call cannot be inlined (it's not an error :D)");
}
MSG(" Call optimized (^_-)\n");
callOptCounter++;
}
//std::cout << "==================================\n";
// Run preloaded passes on the function to propagate constants
funPassManager->run(*procf);
// After
//procf->dump();
// Check if the function is corrupt
verifyFunction(*procf);
this->engine->recompileAndRelinkFunction(procf);
}
/// inlineFuctions - Walk all call sites in all functions supplied by
/// client. Inline as many call sites as possible. Delete completely
/// inlined functions.
void BasicInlinerImpl::inlineFunctions() {
// Scan through and identify all call sites ahead of time so that we only
// inline call sites in the original functions, not call sites that result
// from inlining other functions.
std::vector<CallSite> CallSites;
for (std::vector<Function *>::iterator FI = Functions.begin(),
FE = Functions.end(); FI != FE; ++FI) {
Function *F = *FI;
for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
CallSite CS(cast<Value>(I));
if (CS && CS.getCalledFunction()
&& !CS.getCalledFunction()->isDeclaration())
CallSites.push_back(CS);
}
}
DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
// Inline call sites.
bool Changed = false;
do {
Changed = false;
for (unsigned index = 0; index != CallSites.size() && !CallSites.empty();
++index) {
CallSite CS = CallSites[index];
if (Function *Callee = CS.getCalledFunction()) {
// Eliminate calls that are never inlinable.
if (Callee->isDeclaration() ||
CS.getInstruction()->getParent()->getParent() == Callee) {
CallSites.erase(CallSites.begin() + index);
--index;
continue;
}
InlineCost IC = CA.getInlineCost(CS, NeverInline);
if (IC.isAlways()) {
DEBUG(dbgs() << " Inlining: cost=always"
<<", call: " << *CS.getInstruction());
} else if (IC.isNever()) {
DEBUG(dbgs() << " NOT Inlining: cost=never"
<<", call: " << *CS.getInstruction());
continue;
} else {
int Cost = IC.getValue();
if (Cost >= (int) BasicInlineThreshold) {
DEBUG(dbgs() << " NOT Inlining: cost = " << Cost
<< ", call: " << *CS.getInstruction());
continue;
} else {
DEBUG(dbgs() << " Inlining: cost = " << Cost
<< ", call: " << *CS.getInstruction());
}
}
// Inline
InlineFunctionInfo IFI(0, TD);
if (InlineFunction(CS, IFI)) {
Callee->removeDeadConstantUsers();
if (Callee->isDefTriviallyDead())
DeadFunctions.insert(Callee);
Changed = true;
CallSites.erase(CallSites.begin() + index);
--index;
}
}
}
} while (Changed);
// Remove completely inlined functions from module.
for(SmallPtrSet<Function *, 8>::iterator I = DeadFunctions.begin(),
E = DeadFunctions.end(); I != E; ++I) {
Function *D = *I;
Module *M = D->getParent();
M->getFunctionList().remove(D);
}
}
void JitCodeSpecializer::_inline_function_calls(SpecializationContext& context) const {
// This method implements the main code fusion functionality.
// It works as follows:
// Throughout the fusion process, a working list of call sites (i.e., function calls) is maintained.
// This queue is initialized with all call instructions from the initial root function.
// Each call site is then handled in one of the following ways:
// - Direct Calls:
// Direct calls explicitly encode the name of the called function. If a function implementation can be located in
// the JitRepository repository by that name the function is inlined (i.e., the call instruction is replaced with
// the function body from the repository. Calls to functions not available in the repository (e.g., calls into the
// C++ standard library) require a corresponding function declaration to be inserted into the module.
// This is necessary to avoid any unresolved function calls, which would invalidate the module and cause the
// just-in-time compiler to reject it. With an appropriate declaration, however, the just-in-time compiler is able
// to locate the machine code of these functions in the Hyrise binary when compiling and linking the module.
// - Indirect Calls
// Indirect calls do not reference their called function by name. Instead, the address of the target function is
// either loaded from memory or computed. The code specialization unit analyzes the instructions surrounding the
// call and tries to infer the name of the called function. If this succeeds, the call is handled like a regular
// direct call. If not, no specialization of the call can be performed.
// In this case, the target address computed for the call at runtime will point to the correct machine code in the
// Hyrise binary and the pipeline will still execute successfully.
//
// Whenever a call instruction is replaced by the corresponding function body, the inlining algorithm reports back
// any new call sites encountered in the process. These are pushed to the end of the working queue. Once this queue
// is empty, the operator fusion process is completed.
std::queue<llvm::CallSite> call_sites;
// Initialize the call queue with all call and invoke instructions from the root function.
_visit<llvm::CallInst>(*context.root_function, [&](llvm::CallInst& inst) { call_sites.push(llvm::CallSite(&inst)); });
_visit<llvm::InvokeInst>(*context.root_function,
[&](llvm::InvokeInst& inst) { call_sites.push(llvm::CallSite(&inst)); });
while (!call_sites.empty()) {
auto& call_site = call_sites.front();
// Resolve indirect (virtual) function calls
if (call_site.isIndirectCall()) {
const auto called_value = call_site.getCalledValue();
// Get the runtime location of the called function (i.e., the compiled machine code of the function)
const auto called_runtime_value =
std::dynamic_pointer_cast<const JitKnownRuntimePointer>(GetRuntimePointerForValue(called_value, context));
if (called_runtime_value && called_runtime_value->is_valid()) {
// LLVM implements virtual function calls via virtual function tables
// (see https://en.wikipedia.org/wiki/Virtual_method_table).
// The resolution of virtual calls starts with a pointer to the object that the virtual call should be performed
// on. This object contains a pointer to its vtable (usually at offset 0). The vtable contains a list of
// function pointers (one for each virtual function).
// In the LLVM bitcode, a virtual call is resolved through a number of LLVM pointer operations:
// - a load instruction to dereference the vtable pointer in the object
// - a getelementptr instruction to select the correct virtual function in the vtable
// - another load instruction to dereference the virtual function pointer from the table
// When analyzing a virtual call, the code specializer works backwards starting from the pointer to the called
// function (called_runtime_value).
// Moving "up" one level (undoing one load operation) yields the pointer to the function pointer in the
// vtable. The total_offset of this pointer corresponds to the index in the vtable that is used for the virtual
// call.
// Moving "up" another level yields the pointer to the object that the virtual function is called on. Using RTTI
// the class name of the object can be determined.
// These two pieces of information (the class name and the vtable index) are sufficient to unambiguously
// identify the called virtual function and locate it in the bitcode repository.
// Determine the vtable index and class name of the virtual call
const auto vtable_index =
called_runtime_value->up().total_offset() / context.module->getDataLayout().getPointerSize();
const auto instance = reinterpret_cast<JitRTTIHelper*>(called_runtime_value->up().up().base().address());
const auto class_name = typeid(*instance).name();
// If the called function can be located in the repository, the virtual call is replaced by a direct call to
// that function.
if (const auto repo_function = _repository.get_vtable_entry(class_name, vtable_index)) {
call_site.setCalledFunction(repo_function);
}
} else {
// The virtual call could not be resolved. There is nothing we can inline so we move on.
call_sites.pop();
continue;
}
}
auto function = call_site.getCalledFunction();
// ignore invalid functions
if (!function) {
call_sites.pop();
continue;
}
const auto function_name = function->getName().str();
const auto function_has_opossum_namespace =
boost::starts_with(function_name, "_ZNK7opossum") || boost::starts_with(function_name, "_ZN7opossum");
// A note about "__clang_call_terminate":
// __clang_call_terminate is generated / used internally by clang to call the std::terminate function when exception
// handling fails. For some unknown reason this function cannot be resolved in the Hyrise binary when jit-compiling
// bitcode that uses the function. The function is, however, present in the bitcode repository.
// We thus always inline this function from the repository.
// All function that are not in the opossum:: namespace are not considered for inlining. Instead, a function
// declaration (without a function body) is created.
if (!function_has_opossum_namespace && function_name != "__clang_call_terminate") {
context.llvm_value_map[function] = _create_function_declaration(context, *function, function->getName());
call_sites.pop();
continue;
}
// Determine whether the first function argument is a pointer/reference to an object, but the runtime location
// for this object cannot be determined.
// This is the case for member functions that are called within a loop body. These functions may be called on
// different objects in different loop iterations.
// If two specialization passes are performed, these functions should be inlined after loop unrolling has been
// performed (i.e., during the second pass).
auto first_argument = call_site.arg_begin();
auto first_argument_cannot_be_resolved = first_argument->get()->getType()->isPointerTy() &&
!GetRuntimePointerForValue(first_argument->get(), context)->is_valid();
if (first_argument_cannot_be_resolved && function_name != "__clang_call_terminate") {
call_sites.pop();
continue;
}
// We create a mapping from values in the source module to values in the target module.
// This mapping is passed to LLVM's cloning function and ensures that all references to other functions, global
// variables, and function arguments refer to values in the target module and NOT in the source module.
// This way the module stays self-contained and valid.
context.llvm_value_map.clear();
// Map called functions
_visit<const llvm::Function>(*function, [&](const auto& fn) {
if (fn.isDeclaration() && !context.llvm_value_map.count(&fn)) {
context.llvm_value_map[&fn] = _create_function_declaration(context, fn, fn.getName());
}
});
// Map global variables
_visit<const llvm::GlobalVariable>(*function, [&](auto& global) {
if (!context.llvm_value_map.count(&global)) {
context.llvm_value_map[&global] = _clone_global_variable(context, global);
}
});
// Map function arguments
auto function_arg = function->arg_begin();
auto call_arg = call_site.arg_begin();
for (; function_arg != function->arg_end() && call_arg != call_site.arg_end(); ++function_arg, ++call_arg) {
context.llvm_value_map[function_arg] = call_arg->get();
}
// Instruct LLVM to perform the function inlining and push all new call sites to the working queue
llvm::InlineFunctionInfo info;
if (InlineFunction(call_site, info, nullptr, false, nullptr, context)) {
for (const auto& new_call_site : info.InlinedCallSites) {
call_sites.push(new_call_site);
}
}
// clear runtime_value_map to allow multiple inlining of same function
auto runtime_this = context.runtime_value_map[context.root_function->arg_begin()];
context.runtime_value_map.clear();
context.runtime_value_map[context.root_function->arg_begin()] = runtime_this;
call_sites.pop();
}
}
