ArgumentPromotion.cpp revision ed570a7dca45e001a6223e2a25d034b838934f88
1//===-- ArgumentPromotion.cpp - Promote 'by reference' arguments ----------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass promotes "by reference" arguments to be "by value" arguments. In 11// practice, this means looking for internal functions that have pointer 12// arguments. If we can prove, through the use of alias analysis, that that an 13// argument is *only* loaded, then we can pass the value into the function 14// instead of the address of the value. This can cause recursive simplification 15// of code, and lead to the elimination of allocas, especially in C++ template 16// code like the STL. 17// 18// Note that this transformation could also be done for arguments that are only 19// stored to (returning the value instead), but we do not currently handle that 20// case. 21// 22// Note that we should be able to promote pointers to structures that are only 23// loaded from as well. The danger is creating way to many arguments, so this 24// transformation should be limited to 3 element structs or something. 25// 26//===----------------------------------------------------------------------===// 27 28#include "llvm/Transforms/IPO.h" 29#include "llvm/Constants.h" 30#include "llvm/DerivedTypes.h" 31#include "llvm/Module.h" 32#include "llvm/Pass.h" 33#include "llvm/Instructions.h" 34#include "llvm/Analysis/AliasAnalysis.h" 35#include "llvm/Target/TargetData.h" 36#include "llvm/Support/CallSite.h" 37#include "llvm/Support/CFG.h" 38#include "Support/Debug.h" 39#include "Support/DepthFirstIterator.h" 40#include "Support/Statistic.h" 41#include <set> 42using namespace llvm; 43 44namespace { 45 Statistic<> NumArgumentsPromoted("argpromotion", 46 "Number of pointer arguments promoted"); 47 Statistic<> NumArgumentsDead("argpromotion", 48 "Number of dead pointer args eliminated"); 49 50 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass. 51 /// 52 class ArgPromotion : public Pass { 53 // WorkList - The set of internal functions that we have yet to process. As 54 // we eliminate arguments from a function, we push all callers into this set 55 // so that the by reference argument can be bubbled out as far as possible. 56 // This set contains only internal functions. 57 std::set<Function*> WorkList; 58 public: 59 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 60 AU.addRequired<AliasAnalysis>(); 61 AU.addRequired<TargetData>(); 62 } 63 64 virtual bool run(Module &M); 65 private: 66 bool PromoteArguments(Function *F); 67 bool isSafeToPromoteArgument(Argument *Arg) const; 68 void DoPromotion(Function *F, std::vector<Argument*> &ArgsToPromote); 69 }; 70 71 RegisterOpt<ArgPromotion> X("argpromotion", 72 "Promote 'by reference' arguments to scalars"); 73} 74 75Pass *llvm::createArgumentPromotionPass() { 76 return new ArgPromotion(); 77} 78 79bool ArgPromotion::run(Module &M) { 80 bool Changed = false; 81 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) 82 if (I->hasInternalLinkage()) { 83 WorkList.insert(I); 84 85 // If there are any constant pointer refs pointing to this function, 86 // eliminate them now if possible. 87 ConstantPointerRef *CPR = 0; 88 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 89 ++UI) 90 if ((CPR = dyn_cast<ConstantPointerRef>(*UI))) 91 break; // Found one! 92 if (CPR) { 93 // See if we can transform all users to use the function directly. 94 while (!CPR->use_empty()) { 95 User *TheUser = CPR->use_back(); 96 if (!isa<Constant>(TheUser)) { 97 Changed = true; 98 TheUser->replaceUsesOfWith(CPR, I); 99 } else { 100 // We won't be able to eliminate all users. :( 101 WorkList.erase(I); // Minor efficiency win. 102 break; 103 } 104 } 105 106 // If we nuked all users of the CPR, kill the CPR now! 107 if (CPR->use_empty()) { 108 CPR->destroyConstant(); 109 Changed = true; 110 } 111 } 112 } 113 114 while (!WorkList.empty()) { 115 Function *F = *WorkList.begin(); 116 WorkList.erase(WorkList.begin()); 117 118 if (PromoteArguments(F)) // Attempt to promote an argument. 119 Changed = true; // Remember that we changed something. 120 } 121 122 return Changed; 123} 124 125 126bool ArgPromotion::PromoteArguments(Function *F) { 127 assert(F->hasInternalLinkage() && "We can only process internal functions!"); 128 129 // First check: see if there are any pointer arguments! If not, quick exit. 130 std::vector<Argument*> PointerArgs; 131 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I) 132 if (isa<PointerType>(I->getType())) 133 PointerArgs.push_back(I); 134 if (PointerArgs.empty()) return false; 135 136 // Second check: make sure that all callers are direct callers. We can't 137 // transform functions that have indirect callers. 138 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); 139 UI != E; ++UI) 140 // What about CPRs? 141 if (!CallSite::get(*UI).getInstruction()) 142 return false; // Cannot promote an indirect call! 143 144 // Check to see which arguments are promotable. If an argument is not 145 // promotable, remove it from the PointerArgs vector. 146 for (unsigned i = 0; i != PointerArgs.size(); ++i) 147 if (!isSafeToPromoteArgument(PointerArgs[i])) { 148 std::swap(PointerArgs[i--], PointerArgs.back()); 149 PointerArgs.pop_back(); 150 } 151 152 // No promotable pointer arguments. 153 if (PointerArgs.empty()) return false; 154 155 // Okay, promote all of the arguments are rewrite the callees! 156 DoPromotion(F, PointerArgs); 157 return true; 158} 159 160bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg) const { 161 // We can only promote this argument if all of the uses are loads... 162 std::vector<LoadInst*> Loads; 163 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end(); 164 UI != E; ++UI) 165 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 166 if (LI->isVolatile()) return false; // Don't hack volatile loads 167 Loads.push_back(LI); 168 } else 169 return false; 170 171 if (Loads.empty()) return true; // No users, dead argument. 172 173 const Type *LoadTy = cast<PointerType>(Arg->getType())->getElementType(); 174 unsigned LoadSize = getAnalysis<TargetData>().getTypeSize(LoadTy); 175 176 // Okay, now we know that the argument is only used by load instructions. 177 // Check to see if the pointer is guaranteed to not be modified from entry of 178 // the function to each of the load instructions. 179 Function &F = *Arg->getParent(); 180 181 // Because there could be several/many load instructions, remember which 182 // blocks we know to be transparent to the load. 183 std::set<BasicBlock*> TranspBlocks; 184 185 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 186 187 for (unsigned i = 0, e = Loads.size(); i != e; ++i) { 188 // Check to see if the load is invalidated from the start of the block to 189 // the load itself. 190 LoadInst *Load = Loads[i]; 191 BasicBlock *BB = Load->getParent(); 192 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize)) 193 return false; // Pointer is invalidated! 194 195 // Now check every path from the entry block to the load for transparency. 196 // To do this, we perform a depth first search on the inverse CFG from the 197 // loading block. 198 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 199 for (idf_ext_iterator<BasicBlock*> I = idf_ext_begin(*PI, TranspBlocks), 200 E = idf_ext_end(*PI, TranspBlocks); I != E; ++I) 201 if (AA.canBasicBlockModify(**I, Arg, LoadSize)) 202 return false; 203 } 204 205 // If the path from the entry of the function to each load is free of 206 // instructions that potentially invalidate the load, we can make the 207 // transformation! 208 return true; 209} 210 211 212void ArgPromotion::DoPromotion(Function *F, std::vector<Argument*> &Args2Prom) { 213 std::set<Argument*> ArgsToPromote(Args2Prom.begin(), Args2Prom.end()); 214 215 // Start by computing a new prototype for the function, which is the same as 216 // the old function, but has modified arguments. 217 const FunctionType *FTy = F->getFunctionType(); 218 std::vector<const Type*> Params; 219 220 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I) 221 if (!ArgsToPromote.count(I)) { 222 Params.push_back(I->getType()); 223 } else if (!I->use_empty()) { 224 Params.push_back(cast<PointerType>(I->getType())->getElementType()); 225 ++NumArgumentsPromoted; 226 } else { 227 ++NumArgumentsDead; 228 } 229 230 const Type *RetTy = FTy->getReturnType(); 231 232 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which 233 // have zero fixed arguments. 234 bool ExtraArgHack = false; 235 if (Params.empty() && FTy->isVarArg()) { 236 ExtraArgHack = true; 237 Params.push_back(Type::IntTy); 238 } 239 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); 240 241 // Create the new function body and insert it into the module... 242 Function *NF = new Function(NFTy, F->getLinkage(), F->getName()); 243 F->getParent()->getFunctionList().insert(F, NF); 244 245 // Loop over all of the callers of the function, transforming the call sites 246 // to pass in the loaded pointers. 247 // 248 std::vector<Value*> Args; 249 while (!F->use_empty()) { 250 CallSite CS = CallSite::get(F->use_back()); 251 Instruction *Call = CS.getInstruction(); 252 253 // Make sure the caller of this function is revisited. 254 if (Call->getParent()->getParent()->hasInternalLinkage()) 255 WorkList.insert(Call->getParent()->getParent()); 256 257 // Loop over the operands, deleting dead ones... 258 CallSite::arg_iterator AI = CS.arg_begin(); 259 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I, ++AI) 260 if (!ArgsToPromote.count(I)) 261 Args.push_back(*AI); // Unmodified argument 262 else if (!I->use_empty()) { 263 // Non-dead instruction 264 Args.push_back(new LoadInst(*AI, (*AI)->getName()+".val", Call)); 265 } 266 267 if (ExtraArgHack) 268 Args.push_back(Constant::getNullValue(Type::IntTy)); 269 270 // Push any varargs arguments on the list 271 for (; AI != CS.arg_end(); ++AI) 272 Args.push_back(*AI); 273 274 Instruction *New; 275 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 276 New = new InvokeInst(NF, II->getNormalDest(), II->getUnwindDest(), 277 Args, "", Call); 278 } else { 279 New = new CallInst(NF, Args, "", Call); 280 } 281 Args.clear(); 282 283 if (!Call->use_empty()) { 284 Call->replaceAllUsesWith(New); 285 std::string Name = Call->getName(); 286 Call->setName(""); 287 New->setName(Name); 288 } 289 290 // Finally, remove the old call from the program, reducing the use-count of 291 // F. 292 Call->getParent()->getInstList().erase(Call); 293 } 294 295 // Since we have now created the new function, splice the body of the old 296 // function right into the new function, leaving the old rotting hulk of the 297 // function empty. 298 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 299 300 // Loop over the argument list, transfering uses of the old arguments over to 301 // the new arguments, also transfering over the names as well. 302 // 303 for (Function::aiterator I = F->abegin(), E = F->aend(), I2 = NF->abegin(); 304 I != E; ++I) 305 if (!ArgsToPromote.count(I)) { 306 // If this is an unmodified argument, move the name and users over to the 307 // new version. 308 I->replaceAllUsesWith(I2); 309 I2->setName(I->getName()); 310 ++I2; 311 } else if (!I->use_empty()) { 312 // Otherwise, if we promoted this argument, then all users are load 313 // instructions, and all loads should be using the new argument that we 314 // added. 315 /*DEBUG*/(std::cerr << "*** Promoted argument '" << I->getName() 316 << "' of function '" << F->getName() << "'\n"); 317 I2->setName(I->getName()+".val"); 318 while (!I->use_empty()) { 319 LoadInst *LI = cast<LoadInst>(I->use_back()); 320 LI->replaceAllUsesWith(I2); 321 LI->getParent()->getInstList().erase(LI); 322 } 323 ++I2; 324 } 325 326 // Now that the old function is dead, delete it. 327 F->getParent()->getFunctionList().erase(F); 328} 329