void HeterotbbTransform::rewrite_invoke_site(Module &M, CallSite &CS,Function *NF,int type) { // create_hetero_clone_void(f); Instruction *OldCall = CS.getInstruction(); Instruction *NewCall; // New Call Instruction created SmallVector<Value*, 16> Args; // Argument lists to the new call //DEBUG(dbgs() << "Old Call Instruction:"); //OldCall->dump(); // Any attributes (parameter attribute list PAL) of the // parallel_for_hetero is #if defined(LLVM_3_2) SmallVector<AttributeWithIndex, 8> AttrVec; // Attributes list to the new call const AttrListPtr &OldCallPAL = CS.getAttributes(); // Add any return attributes. Attributes attrs = OldCallPAL.getRetAttributes(); if (attrs.hasAttributes()) AttrVec.push_back(AttributeWithIndex::get(0, attrs)); #endif SmallVector<AttributeSet, 8> AttrVec; const AttributeSet &OldCallPAL = CS.getAttributes(); // Add any return attributes. if (OldCallPAL.hasAttributes(AttributeSet::ReturnIndex)) AttrVec.push_back(AttributeSet::get(NF->getContext(), OldCallPAL.getRetAttributes())); CallSite::arg_iterator AI = CS.arg_begin(); Args.push_back(CS.getArgument(0)); // num_iters //Args.push_back(CS.getArgument(1)); //params.push_back(CS.getArgument(1)->getType()); //create a new cast from class_name to i8* before the old instruction site CastInst *StrucCast = CastInst::Create(Instruction::BitCast, CS.getArgument(1), PointerType::get(Type::getInt8Ty(M.getContext()), 0), "temp_cast", OldCall); //push the type into the argument list Args.push_back(StrucCast); // struct //push the function as third argument Args.push_back(NF); //NF->dump(); //NF->getType()->dump(); //Args.push_back(CS.getArgument(2)); vector</*const*/ Type *> params; const FunctionType *FTy = NF->getFunctionType(); //#ifndef IVB_64 params.push_back(Type::getInt32Ty(M.getContext())); /*#else params.push_back(Type::getInt64Ty(M.getContext())); #endif*/ params.push_back(PointerType::get(Type::getInt8Ty(M.getContext()),0)); params.push_back(NF->getType()); //params.push_back(Type::getInt32Ty(M.getContext())); /*const*/ Type *RetTy = FTy->getReturnType(); FunctionType *NFty = FunctionType::get(RetTy,params, false); //NF->getType()->dump(); //NFty->dump(); Constant *hetero_f_const; //if (hetero_f_const == NULL) { hetero_f_const = /*cast<Function>*/(M.getOrInsertFunction("offload", NFty)); //} //hetero_f_const->dump(); NewCall = InvokeInst::Create(hetero_f_const,cast<InvokeInst>(OldCall)->getNormalDest(),cast<InvokeInst>(OldCall)->getUnwindDest(), Args, "", OldCall); cast<InvokeInst>(NewCall)->setCallingConv(CS.getCallingConv()); //cast<InvokeInst>(NewCall)->setAttributes(AttrListPtr::get(NF->getContext(), AttrVec)); cast<InvokeInst>(NewCall)->setAttributes(AttributeSet::get(NF->getContext(), AttrVec)); //NewCall->dump(); //NewCall = CallInst::Create(hetero_f_const, Args.begin(), Args.end(), "", OldCall); //NewCall->dump(); //cast<CallInst>(NewCall)->setCallingConv(CS.getCallingConv()); //cast<CallInst>(NewCall)->setAttributes(AttrListPtr::get(AttrVec.begin(), AttrVec.end())); //if (CallInst *c=dyn_cast<CallInst>(OldCall)){ // if(c->isTailCall()) cast<CallInst>(NewCall)->setTailCall(); //} char buf[32]; ConstantInt *ci; //DEBUG(dbgs() << "original scheduler_hint="); //CS.getArgument(2)->dump(); if (ci = dyn_cast<ConstantInt>(CS.getArgument(2))) { DEBUG(dbgs() << "scheduler_hint=" << ci->getZExtValue()); sprintf(buf,"%d",ci->getZExtValue()); } else { DEBUG(dbgs() << "scheduler_hint is not supplied and assumed 0"); sprintf(buf,"%d",0); } Value *e2[] = {MDString::get(M.getContext(),buf)}; MDNode *n2 = MDNode::get(M.getContext(), e2); NewCall->setMetadata("scheduler_hint",n2); if(type==2) { //add meta data for reduction Function *join=get_join_func(M,CS); Value *Elts[] = {MDString::get(M.getContext(), join->getName())}; MDNode *Node = MDNode::get(M.getContext(), Elts); NewCall->setMetadata("join_cpu",Node); Function *Njoin=create_new_join(M,join); Value *Elts1[] = {MDString::get(M.getContext(), Njoin->getName())}; MDNode *Node1 = MDNode::get(M.getContext(), Elts1); NewCall->setMetadata("join_gpu",Node1); char buffer[32]; sprintf(buffer,"%d",object_sizes[NF]); Value *Elts2[] = {MDString::get(M.getContext(),buffer )}; MDNode *Node2 = MDNode::get(M.getContext(), Elts2); NewCall->setMetadata("object_size",Node2); } //NewCall->stripPointerCasts(); //DEBUG(dbgs() << "Newly created instruction:"); //NewCall->dump(); }
// RemoveDeadStuffFromFunction - Remove any arguments and return values from F // that are not in LiveValues. Transform the function and all of the callees of // the function to not have these arguments and return values. // bool DAE::RemoveDeadStuffFromFunction(Function *F) { // Don't modify fully live functions if (LiveFunctions.count(F)) return false; // Start by computing a new prototype for the function, which is the same as // the old function, but has fewer arguments and a different return type. FunctionType *FTy = F->getFunctionType(); std::vector<Type*> Params; // Set up to build a new list of parameter attributes. SmallVector<AttributeWithIndex, 8> AttributesVec; const AttrListPtr &PAL = F->getAttributes(); // The existing function return attributes. Attributes RAttrs = PAL.getRetAttributes(); Attributes FnAttrs = PAL.getFnAttributes(); // Find out the new return value. Type *RetTy = FTy->getReturnType(); Type *NRetTy = NULL; unsigned RetCount = NumRetVals(F); // -1 means unused, other numbers are the new index SmallVector<int, 5> NewRetIdxs(RetCount, -1); std::vector<Type*> RetTypes; if (RetTy->isVoidTy()) { NRetTy = RetTy; } else { StructType *STy = dyn_cast<StructType>(RetTy); if (STy) // Look at each of the original return values individually. for (unsigned i = 0; i != RetCount; ++i) { RetOrArg Ret = CreateRet(F, i); if (LiveValues.erase(Ret)) { RetTypes.push_back(STy->getElementType(i)); NewRetIdxs[i] = RetTypes.size() - 1; } else { ++NumRetValsEliminated; DEBUG(dbgs() << "DAE - Removing return value " << i << " from " << F->getName() << "\n"); } } else // We used to return a single value. if (LiveValues.erase(CreateRet(F, 0))) { RetTypes.push_back(RetTy); NewRetIdxs[0] = 0; } else { DEBUG(dbgs() << "DAE - Removing return value from " << F->getName() << "\n"); ++NumRetValsEliminated; } if (RetTypes.size() > 1) // More than one return type? Return a struct with them. Also, if we used // to return a struct and didn't change the number of return values, // return a struct again. This prevents changing {something} into // something and {} into void. // Make the new struct packed if we used to return a packed struct // already. NRetTy = StructType::get(STy->getContext(), RetTypes, STy->isPacked()); else if (RetTypes.size() == 1) // One return type? Just a simple value then, but only if we didn't use to // return a struct with that simple value before. NRetTy = RetTypes.front(); else if (RetTypes.size() == 0) // No return types? Make it void, but only if we didn't use to return {}. NRetTy = Type::getVoidTy(F->getContext()); } assert(NRetTy && "No new return type found?"); // Remove any incompatible attributes, but only if we removed all return // values. Otherwise, ensure that we don't have any conflicting attributes // here. Currently, this should not be possible, but special handling might be // required when new return value attributes are added. if (NRetTy->isVoidTy()) RAttrs = Attributes::get(NRetTy->getContext(), AttrBuilder(RAttrs). removeAttributes(Attributes::typeIncompatible(NRetTy))); else assert(!AttrBuilder(RAttrs). hasAttributes(Attributes::typeIncompatible(NRetTy)) && "Return attributes no longer compatible?"); if (RAttrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::ReturnIndex, RAttrs)); // Remember which arguments are still alive. SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false); // Construct the new parameter list from non-dead arguments. Also construct // a new set of parameter attributes to correspond. Skip the first parameter // attribute, since that belongs to the return value. unsigned i = 0; for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I, ++i) { RetOrArg Arg = CreateArg(F, i); if (LiveValues.erase(Arg)) { Params.push_back(I->getType()); ArgAlive[i] = true; // Get the original parameter attributes (skipping the first one, that is // for the return value. Attributes Attrs = PAL.getParamAttributes(i + 1); if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Params.size(), Attrs)); } else { ++NumArgumentsEliminated; DEBUG(dbgs() << "DAE - Removing argument " << i << " (" << I->getName() << ") from " << F->getName() << "\n"); } } if (FnAttrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, FnAttrs)); // Reconstruct the AttributesList based on the vector we constructed. AttrListPtr NewPAL = AttrListPtr::get(AttributesVec); // Create the new function type based on the recomputed parameters. FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg()); // No change? if (NFTy == FTy) return false; // Create the new function body and insert it into the module... Function *NF = Function::Create(NFTy, F->getLinkage()); NF->copyAttributesFrom(F); NF->setAttributes(NewPAL); // Insert the new function before the old function, so we won't be processing // it again. F->getParent()->getFunctionList().insert(F, NF); NF->takeName(F); // Loop over all of the callers of the function, transforming the call sites // to pass in a smaller number of arguments into the new function. // std::vector<Value*> Args; while (!F->use_empty()) { CallSite CS(F->use_back()); Instruction *Call = CS.getInstruction(); AttributesVec.clear(); const AttrListPtr &CallPAL = CS.getAttributes(); // The call return attributes. Attributes RAttrs = CallPAL.getRetAttributes(); Attributes FnAttrs = CallPAL.getFnAttributes(); // Adjust in case the function was changed to return void. RAttrs = Attributes::get(NF->getContext(), AttrBuilder(RAttrs). removeAttributes(Attributes::typeIncompatible(NF->getReturnType()))); if (RAttrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::ReturnIndex, RAttrs)); // Declare these outside of the loops, so we can reuse them for the second // loop, which loops the varargs. CallSite::arg_iterator I = CS.arg_begin(); unsigned i = 0; // Loop over those operands, corresponding to the normal arguments to the // original function, and add those that are still alive. for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i) if (ArgAlive[i]) { Args.push_back(*I); // Get original parameter attributes, but skip return attributes. Attributes Attrs = CallPAL.getParamAttributes(i + 1); if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); } // Push any varargs arguments on the list. Don't forget their attributes. for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) { Args.push_back(*I); Attributes Attrs = CallPAL.getParamAttributes(i + 1); if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); } if (FnAttrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, FnAttrs)); // Reconstruct the AttributesList based on the vector we constructed. AttrListPtr NewCallPAL = AttrListPtr::get(AttributesVec); Instruction *New; if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), Args, "", Call); cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); cast<InvokeInst>(New)->setAttributes(NewCallPAL); } else { New = CallInst::Create(NF, Args, "", Call); cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); cast<CallInst>(New)->setAttributes(NewCallPAL); if (cast<CallInst>(Call)->isTailCall()) cast<CallInst>(New)->setTailCall(); } New->setDebugLoc(Call->getDebugLoc()); Args.clear(); if (!Call->use_empty()) { if (New->getType() == Call->getType()) { // Return type not changed? Just replace users then. Call->replaceAllUsesWith(New); New->takeName(Call); } else if (New->getType()->isVoidTy()) { // Our return value has uses, but they will get removed later on. // Replace by null for now. if (!Call->getType()->isX86_MMXTy()) Call->replaceAllUsesWith(Constant::getNullValue(Call->getType())); } else { assert(RetTy->isStructTy() && "Return type changed, but not into a void. The old return type" " must have been a struct!"); Instruction *InsertPt = Call; if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { BasicBlock::iterator IP = II->getNormalDest()->begin(); while (isa<PHINode>(IP)) ++IP; InsertPt = IP; } // We used to return a struct. Instead of doing smart stuff with all the // uses of this struct, we will just rebuild it using // extract/insertvalue chaining and let instcombine clean that up. // // Start out building up our return value from undef Value *RetVal = UndefValue::get(RetTy); for (unsigned i = 0; i != RetCount; ++i) if (NewRetIdxs[i] != -1) { Value *V; if (RetTypes.size() > 1) // We are still returning a struct, so extract the value from our // return value V = ExtractValueInst::Create(New, NewRetIdxs[i], "newret", InsertPt); else // We are now returning a single element, so just insert that V = New; // Insert the value at the old position RetVal = InsertValueInst::Create(RetVal, V, i, "oldret", InsertPt); } // Now, replace all uses of the old call instruction with the return // struct we built Call->replaceAllUsesWith(RetVal); New->takeName(Call); } } // Finally, remove the old call from the program, reducing the use-count of // F. Call->eraseFromParent(); } // Since we have now created the new function, splice the body of the old // function right into the new function, leaving the old rotting hulk of the // function empty. NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); // Loop over the argument list, transferring uses of the old arguments over to // the new arguments, also transferring over the names as well. i = 0; for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), I2 = NF->arg_begin(); I != E; ++I, ++i) if (ArgAlive[i]) { // If this is a live argument, move the name and users over to the new // version. I->replaceAllUsesWith(I2); I2->takeName(I); ++I2; } else { // If this argument is dead, replace any uses of it with null constants // (these are guaranteed to become unused later on). if (!I->getType()->isX86_MMXTy()) I->replaceAllUsesWith(Constant::getNullValue(I->getType())); } // If we change the return value of the function we must rewrite any return // instructions. Check this now. if (F->getReturnType() != NF->getReturnType()) for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB) if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { Value *RetVal; if (NFTy->getReturnType()->isVoidTy()) { RetVal = 0; } else { assert (RetTy->isStructTy()); // The original return value was a struct, insert // extractvalue/insertvalue chains to extract only the values we need // to return and insert them into our new result. // This does generate messy code, but we'll let it to instcombine to // clean that up. Value *OldRet = RI->getOperand(0); // Start out building up our return value from undef RetVal = UndefValue::get(NRetTy); for (unsigned i = 0; i != RetCount; ++i) if (NewRetIdxs[i] != -1) { ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i, "oldret", RI); if (RetTypes.size() > 1) { // We're still returning a struct, so reinsert the value into // our new return value at the new index RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i], "newret", RI); } else { // We are now only returning a simple value, so just return the // extracted value. RetVal = EV; } } } // Replace the return instruction with one returning the new return // value (possibly 0 if we became void). ReturnInst::Create(F->getContext(), RetVal, RI); BB->getInstList().erase(RI); } // Patch the pointer to LLVM function in debug info descriptor. FunctionDIMap::iterator DI = FunctionDIs.find(F); if (DI != FunctionDIs.end()) DI->second.replaceFunction(NF); // Now that the old function is dead, delete it. F->eraseFromParent(); return true; }
/// DoPromotion - This method actually performs the promotion of the specified /// arguments, and returns the new function. At this point, we know that it's /// safe to do so. CallGraphNode *ArgPromotion::DoPromotion(Function *F, SmallPtrSet<Argument*, 8> &ArgsToPromote, SmallPtrSet<Argument*, 8> &ByValArgsToTransform) { // Start by computing a new prototype for the function, which is the same as // the old function, but has modified arguments. FunctionType *FTy = F->getFunctionType(); std::vector<Type*> Params; typedef std::set<IndicesVector> ScalarizeTable; // ScalarizedElements - If we are promoting a pointer that has elements // accessed out of it, keep track of which elements are accessed so that we // can add one argument for each. // // Arguments that are directly loaded will have a zero element value here, to // handle cases where there are both a direct load and GEP accesses. // std::map<Argument*, ScalarizeTable> ScalarizedElements; // OriginalLoads - Keep track of a representative load instruction from the // original function so that we can tell the alias analysis implementation // what the new GEP/Load instructions we are inserting look like. std::map<IndicesVector, LoadInst*> OriginalLoads; // Attributes - Keep track of the parameter attributes for the arguments // that we are *not* promoting. For the ones that we do promote, the parameter // attributes are lost SmallVector<AttributeWithIndex, 8> AttributesVec; const AttributeSet &PAL = F->getAttributes(); // Add any return attributes. Attributes attrs = PAL.getRetAttributes(); if (attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(AttributeSet::ReturnIndex, attrs)); // First, determine the new argument list unsigned ArgIndex = 1; for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I, ++ArgIndex) { if (ByValArgsToTransform.count(I)) { // Simple byval argument? Just add all the struct element types. Type *AgTy = cast<PointerType>(I->getType())->getElementType(); StructType *STy = cast<StructType>(AgTy); for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) Params.push_back(STy->getElementType(i)); ++NumByValArgsPromoted; } else if (!ArgsToPromote.count(I)) { // Unchanged argument Params.push_back(I->getType()); Attributes attrs = PAL.getParamAttributes(ArgIndex); if (attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs)); } else if (I->use_empty()) { // Dead argument (which are always marked as promotable) ++NumArgumentsDead; } else { // Okay, this is being promoted. This means that the only uses are loads // or GEPs which are only used by loads // In this table, we will track which indices are loaded from the argument // (where direct loads are tracked as no indices). ScalarizeTable &ArgIndices = ScalarizedElements[I]; for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI) { Instruction *User = cast<Instruction>(*UI); assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User)); IndicesVector Indices; Indices.reserve(User->getNumOperands() - 1); // Since loads will only have a single operand, and GEPs only a single // non-index operand, this will record direct loads without any indices, // and gep+loads with the GEP indices. for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end(); II != IE; ++II) Indices.push_back(cast<ConstantInt>(*II)->getSExtValue()); // GEPs with a single 0 index can be merged with direct loads if (Indices.size() == 1 && Indices.front() == 0) Indices.clear(); ArgIndices.insert(Indices); LoadInst *OrigLoad; if (LoadInst *L = dyn_cast<LoadInst>(User)) OrigLoad = L; else // Take any load, we will use it only to update Alias Analysis OrigLoad = cast<LoadInst>(User->use_back()); OriginalLoads[Indices] = OrigLoad; } // Add a parameter to the function for each element passed in. for (ScalarizeTable::iterator SI = ArgIndices.begin(), E = ArgIndices.end(); SI != E; ++SI) { // not allowed to dereference ->begin() if size() is 0 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI)); assert(Params.back()); } if (ArgIndices.size() == 1 && ArgIndices.begin()->empty()) ++NumArgumentsPromoted; else ++NumAggregatesPromoted; } } // Add any function attributes. attrs = PAL.getFnAttributes(); if (attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(AttributeSet::FunctionIndex, attrs)); Type *RetTy = FTy->getReturnType(); // Construct the new function type using the new arguments. FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); // Create the new function body and insert it into the module. Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); NF->copyAttributesFrom(F); DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n" << "From: " << *F); // Recompute the parameter attributes list based on the new arguments for // the function. NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec)); AttributesVec.clear(); F->getParent()->getFunctionList().insert(F, NF); NF->takeName(F); // Get the alias analysis information that we need to update to reflect our // changes. AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); // Get the callgraph information that we need to update to reflect our // changes. CallGraph &CG = getAnalysis<CallGraph>(); // Get a new callgraph node for NF. CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF); // Loop over all of the callers of the function, transforming the call sites // to pass in the loaded pointers. // SmallVector<Value*, 16> Args; while (!F->use_empty()) { CallSite CS(F->use_back()); assert(CS.getCalledFunction() == F); Instruction *Call = CS.getInstruction(); const AttributeSet &CallPAL = CS.getAttributes(); // Add any return attributes. Attributes attrs = CallPAL.getRetAttributes(); if (attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(AttributeSet::ReturnIndex, attrs)); // Loop over the operands, inserting GEP and loads in the caller as // appropriate. CallSite::arg_iterator AI = CS.arg_begin(); ArgIndex = 1; for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I, ++AI, ++ArgIndex) if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { Args.push_back(*AI); // Unmodified argument Attributes Attrs = CallPAL.getParamAttributes(ArgIndex); if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); } else if (ByValArgsToTransform.count(I)) { // Emit a GEP and load for each element of the struct. Type *AgTy = cast<PointerType>(I->getType())->getElementType(); StructType *STy = cast<StructType>(AgTy); Value *Idxs[2] = { ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); Value *Idx = GetElementPtrInst::Create(*AI, Idxs, (*AI)->getName()+"."+utostr(i), Call); // TODO: Tell AA about the new values? Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call)); } } else if (!I->use_empty()) { // Non-dead argument: insert GEPs and loads as appropriate. ScalarizeTable &ArgIndices = ScalarizedElements[I]; // Store the Value* version of the indices in here, but declare it now // for reuse. std::vector<Value*> Ops; for (ScalarizeTable::iterator SI = ArgIndices.begin(), E = ArgIndices.end(); SI != E; ++SI) { Value *V = *AI; LoadInst *OrigLoad = OriginalLoads[*SI]; if (!SI->empty()) { Ops.reserve(SI->size()); Type *ElTy = V->getType(); for (IndicesVector::const_iterator II = SI->begin(), IE = SI->end(); II != IE; ++II) { // Use i32 to index structs, and i64 for others (pointers/arrays). // This satisfies GEP constraints. Type *IdxTy = (ElTy->isStructTy() ? Type::getInt32Ty(F->getContext()) : Type::getInt64Ty(F->getContext())); Ops.push_back(ConstantInt::get(IdxTy, *II)); // Keep track of the type we're currently indexing. ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II); } // And create a GEP to extract those indices. V = GetElementPtrInst::Create(V, Ops, V->getName()+".idx", Call); Ops.clear(); AA.copyValue(OrigLoad->getOperand(0), V); } // Since we're replacing a load make sure we take the alignment // of the previous load. LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call); newLoad->setAlignment(OrigLoad->getAlignment()); // Transfer the TBAA info too. newLoad->setMetadata(LLVMContext::MD_tbaa, OrigLoad->getMetadata(LLVMContext::MD_tbaa)); Args.push_back(newLoad); AA.copyValue(OrigLoad, Args.back()); } } // Push any varargs arguments on the list. for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { Args.push_back(*AI); Attributes Attrs = CallPAL.getParamAttributes(ArgIndex); if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); } // Add any function attributes. attrs = CallPAL.getFnAttributes(); if (attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(AttributeSet::FunctionIndex, attrs)); Instruction *New; if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), Args, "", Call); cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(), AttributesVec)); } else { New = CallInst::Create(NF, Args, "", Call); cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(), AttributesVec)); if (cast<CallInst>(Call)->isTailCall()) cast<CallInst>(New)->setTailCall(); } Args.clear(); AttributesVec.clear(); // Update the alias analysis implementation to know that we are replacing // the old call with a new one. AA.replaceWithNewValue(Call, New); // Update the callgraph to know that the callsite has been transformed. CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()]; CalleeNode->replaceCallEdge(Call, New, NF_CGN); if (!Call->use_empty()) { Call->replaceAllUsesWith(New); New->takeName(Call); } // Finally, remove the old call from the program, reducing the use-count of // F. Call->eraseFromParent(); } // Since we have now created the new function, splice the body of the old // function right into the new function, leaving the old rotting hulk of the // function empty. NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); // Loop over the argument list, transferring uses of the old arguments over to // the new arguments, also transferring over the names as well. // for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), I2 = NF->arg_begin(); I != E; ++I) { if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { // If this is an unmodified argument, move the name and users over to the // new version. I->replaceAllUsesWith(I2); I2->takeName(I); AA.replaceWithNewValue(I, I2); ++I2; continue; } if (ByValArgsToTransform.count(I)) { // In the callee, we create an alloca, and store each of the new incoming // arguments into the alloca. Instruction *InsertPt = NF->begin()->begin(); // Just add all the struct element types. Type *AgTy = cast<PointerType>(I->getType())->getElementType(); Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt); StructType *STy = cast<StructType>(AgTy); Value *Idxs[2] = { ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); Value *Idx = GetElementPtrInst::Create(TheAlloca, Idxs, TheAlloca->getName()+"."+Twine(i), InsertPt); I2->setName(I->getName()+"."+Twine(i)); new StoreInst(I2++, Idx, InsertPt); } // Anything that used the arg should now use the alloca. I->replaceAllUsesWith(TheAlloca); TheAlloca->takeName(I); AA.replaceWithNewValue(I, TheAlloca); continue; } if (I->use_empty()) { AA.deleteValue(I); continue; } // Otherwise, if we promoted this argument, then all users are load // instructions (or GEPs with only load users), and all loads should be // using the new argument that we added. ScalarizeTable &ArgIndices = ScalarizedElements[I]; while (!I->use_empty()) { if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) { assert(ArgIndices.begin()->empty() && "Load element should sort to front!"); I2->setName(I->getName()+".val"); LI->replaceAllUsesWith(I2); AA.replaceWithNewValue(LI, I2); LI->eraseFromParent(); DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName() << "' in function '" << F->getName() << "'\n"); } else { GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back()); IndicesVector Operands; Operands.reserve(GEP->getNumIndices()); for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); II != IE; ++II) Operands.push_back(cast<ConstantInt>(*II)->getSExtValue()); // GEPs with a single 0 index can be merged with direct loads if (Operands.size() == 1 && Operands.front() == 0) Operands.clear(); Function::arg_iterator TheArg = I2; for (ScalarizeTable::iterator It = ArgIndices.begin(); *It != Operands; ++It, ++TheArg) { assert(It != ArgIndices.end() && "GEP not handled??"); } std::string NewName = I->getName(); for (unsigned i = 0, e = Operands.size(); i != e; ++i) { NewName += "." + utostr(Operands[i]); } NewName += ".val"; TheArg->setName(NewName); DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName() << "' of function '" << NF->getName() << "'\n"); // All of the uses must be load instructions. Replace them all with // the argument specified by ArgNo. while (!GEP->use_empty()) { LoadInst *L = cast<LoadInst>(GEP->use_back()); L->replaceAllUsesWith(TheArg); AA.replaceWithNewValue(L, TheArg); L->eraseFromParent(); } AA.deleteValue(GEP); GEP->eraseFromParent(); } } // Increment I2 past all of the arguments added for this promoted pointer. std::advance(I2, ArgIndices.size()); } // Tell the alias analysis that the old function is about to disappear. AA.replaceWithNewValue(F, NF); NF_CGN->stealCalledFunctionsFrom(CG[F]); // Now that the old function is dead, delete it. If there is a dangling // reference to the CallgraphNode, just leave the dead function around for // someone else to nuke. CallGraphNode *CGN = CG[F]; if (CGN->getNumReferences() == 0) delete CG.removeFunctionFromModule(CGN); else F->setLinkage(Function::ExternalLinkage); return NF_CGN; }
/// DeleteDeadVarargs - If this is an function that takes a ... list, and if /// llvm.vastart is never called, the varargs list is dead for the function. bool DAE::DeleteDeadVarargs(Function &Fn) { assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!"); if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false; // Ensure that the function is only directly called. if (Fn.hasAddressTaken()) return false; // Okay, we know we can transform this function if safe. Scan its body // looking for calls to llvm.vastart. for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { if (II->getIntrinsicID() == Intrinsic::vastart) return false; } } } // If we get here, there are no calls to llvm.vastart in the function body, // remove the "..." and adjust all the calls. // Start by computing a new prototype for the function, which is the same as // the old function, but doesn't have isVarArg set. FunctionType *FTy = Fn.getFunctionType(); std::vector<Type*> Params(FTy->param_begin(), FTy->param_end()); FunctionType *NFTy = FunctionType::get(FTy->getReturnType(), Params, false); unsigned NumArgs = Params.size(); // Create the new function body and insert it into the module... Function *NF = Function::Create(NFTy, Fn.getLinkage()); NF->copyAttributesFrom(&Fn); Fn.getParent()->getFunctionList().insert(&Fn, NF); NF->takeName(&Fn); // Loop over all of the callers of the function, transforming the call sites // to pass in a smaller number of arguments into the new function. // std::vector<Value*> Args; while (!Fn.use_empty()) { CallSite CS(Fn.use_back()); Instruction *Call = CS.getInstruction(); // Pass all the same arguments. Args.assign(CS.arg_begin(), CS.arg_begin() + NumArgs); // Drop any attributes that were on the vararg arguments. AttrListPtr PAL = CS.getAttributes(); if (!PAL.isEmpty() && PAL.getSlot(PAL.getNumSlots() - 1).Index > NumArgs) { SmallVector<AttributeWithIndex, 8> AttributesVec; for (unsigned i = 0; PAL.getSlot(i).Index <= NumArgs; ++i) AttributesVec.push_back(PAL.getSlot(i)); Attributes FnAttrs = PAL.getFnAttributes(); if (FnAttrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, FnAttrs)); PAL = AttrListPtr::get(AttributesVec); } Instruction *New; if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), Args, "", Call); cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); cast<InvokeInst>(New)->setAttributes(PAL); } else { New = CallInst::Create(NF, Args, "", Call); cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); cast<CallInst>(New)->setAttributes(PAL); if (cast<CallInst>(Call)->isTailCall()) cast<CallInst>(New)->setTailCall(); } New->setDebugLoc(Call->getDebugLoc()); Args.clear(); if (!Call->use_empty()) Call->replaceAllUsesWith(New); New->takeName(Call); // Finally, remove the old call from the program, reducing the use-count of // F. Call->eraseFromParent(); } // Since we have now created the new function, splice the body of the old // function right into the new function, leaving the old rotting hulk of the // function empty. NF->getBasicBlockList().splice(NF->begin(), Fn.getBasicBlockList()); // Loop over the argument list, transferring uses of the old arguments over to // the new arguments, also transferring over the names as well. While we're at // it, remove the dead arguments from the DeadArguments list. // for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(), I2 = NF->arg_begin(); I != E; ++I, ++I2) { // Move the name and users over to the new version. I->replaceAllUsesWith(I2); I2->takeName(I); } // Patch the pointer to LLVM function in debug info descriptor. FunctionDIMap::iterator DI = FunctionDIs.find(&Fn); if (DI != FunctionDIs.end()) DI->second.replaceFunction(NF); // Finally, nuke the old function. Fn.eraseFromParent(); return true; }