MergeFunctions.cpp revision 7abe37e4aee38cc79d91dd069a37d7e91d5bef53
1//===- MergeFunctions.cpp - Merge identical functions ---------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass looks for equivalent functions that are mergable and folds them. 11// 12// A hash is computed from the function, based on its type and number of 13// basic blocks. 14// 15// Once all hashes are computed, we perform an expensive equality comparison 16// on each function pair. This takes n^2/2 comparisons per bucket, so it's 17// important that the hash function be high quality. The equality comparison 18// iterates through each instruction in each basic block. 19// 20// When a match is found, the functions are folded. We can only fold two 21// functions when we know that the definition of one of them is not 22// overridable. 23// 24//===----------------------------------------------------------------------===// 25// 26// Future work: 27// 28// * fold vector<T*>::push_back and vector<S*>::push_back. 29// 30// These two functions have different types, but in a way that doesn't matter 31// to us. As long as we never see an S or T itself, using S* and S** is the 32// same as using a T* and T**. 33// 34// * virtual functions. 35// 36// Many functions have their address taken by the virtual function table for 37// the object they belong to. However, as long as it's only used for a lookup 38// and call, this is irrelevant, and we'd like to fold such implementations. 39// 40//===----------------------------------------------------------------------===// 41 42#define DEBUG_TYPE "mergefunc" 43#include "llvm/Transforms/IPO.h" 44#include "llvm/ADT/DenseMap.h" 45#include "llvm/ADT/FoldingSet.h" 46#include "llvm/ADT/Statistic.h" 47#include "llvm/Constants.h" 48#include "llvm/InlineAsm.h" 49#include "llvm/Instructions.h" 50#include "llvm/LLVMContext.h" 51#include "llvm/Module.h" 52#include "llvm/Pass.h" 53#include "llvm/Support/CallSite.h" 54#include "llvm/Support/Debug.h" 55#include "llvm/Support/ErrorHandling.h" 56#include "llvm/Support/raw_ostream.h" 57#include <map> 58#include <vector> 59using namespace llvm; 60 61STATISTIC(NumFunctionsMerged, "Number of functions merged"); 62 63namespace { 64 struct MergeFunctions : public ModulePass { 65 static char ID; // Pass identification, replacement for typeid 66 MergeFunctions() : ModulePass(&ID) {} 67 68 bool runOnModule(Module &M); 69 }; 70} 71 72char MergeFunctions::ID = 0; 73static RegisterPass<MergeFunctions> 74X("mergefunc", "Merge Functions"); 75 76ModulePass *llvm::createMergeFunctionsPass() { 77 return new MergeFunctions(); 78} 79 80// ===----------------------------------------------------------------------=== 81// Comparison of functions 82// ===----------------------------------------------------------------------=== 83 84static unsigned long fhash(const Function *F) { 85 const FunctionType *FTy = F->getFunctionType(); 86 87 FoldingSetNodeID ID; 88 ID.AddInteger(F->size()); 89 ID.AddInteger(F->getCallingConv()); 90 ID.AddBoolean(F->hasGC()); 91 ID.AddBoolean(FTy->isVarArg()); 92 ID.AddInteger(FTy->getReturnType()->getTypeID()); 93 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i) 94 ID.AddInteger(FTy->getParamType(i)->getTypeID()); 95 return ID.ComputeHash(); 96} 97 98/// IgnoreBitcasts - given a bitcast, returns the first non-bitcast found by 99/// walking the chain of cast operands. Otherwise, returns the argument. 100static Value* IgnoreBitcasts(Value *V) { 101 while (BitCastInst *BC = dyn_cast<BitCastInst>(V)) 102 V = BC->getOperand(0); 103 104 return V; 105} 106 107/// isEquivalentType - any two pointers are equivalent. Otherwise, standard 108/// type equivalence rules apply. 109static bool isEquivalentType(const Type *Ty1, const Type *Ty2) { 110 if (Ty1 == Ty2) 111 return true; 112 if (Ty1->getTypeID() != Ty2->getTypeID()) 113 return false; 114 115 switch(Ty1->getTypeID()) { 116 case Type::VoidTyID: 117 case Type::FloatTyID: 118 case Type::DoubleTyID: 119 case Type::X86_FP80TyID: 120 case Type::FP128TyID: 121 case Type::PPC_FP128TyID: 122 case Type::LabelTyID: 123 case Type::MetadataTyID: 124 return true; 125 126 case Type::IntegerTyID: 127 case Type::OpaqueTyID: 128 // Ty1 == Ty2 would have returned true earlier. 129 return false; 130 131 default: 132 llvm_unreachable("Unknown type!"); 133 return false; 134 135 case Type::PointerTyID: { 136 const PointerType *PTy1 = cast<PointerType>(Ty1); 137 const PointerType *PTy2 = cast<PointerType>(Ty2); 138 return PTy1->getAddressSpace() == PTy2->getAddressSpace(); 139 } 140 141 case Type::StructTyID: { 142 const StructType *STy1 = cast<StructType>(Ty1); 143 const StructType *STy2 = cast<StructType>(Ty2); 144 if (STy1->getNumElements() != STy2->getNumElements()) 145 return false; 146 147 if (STy1->isPacked() != STy2->isPacked()) 148 return false; 149 150 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) { 151 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i))) 152 return false; 153 } 154 return true; 155 } 156 157 case Type::FunctionTyID: { 158 const FunctionType *FTy1 = cast<FunctionType>(Ty1); 159 const FunctionType *FTy2 = cast<FunctionType>(Ty2); 160 if (FTy1->getNumParams() != FTy2->getNumParams() || 161 FTy1->isVarArg() != FTy2->isVarArg()) 162 return false; 163 164 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType())) 165 return false; 166 167 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) { 168 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i))) 169 return false; 170 } 171 return true; 172 } 173 174 case Type::ArrayTyID: 175 case Type::VectorTyID: { 176 const SequentialType *STy1 = cast<SequentialType>(Ty1); 177 const SequentialType *STy2 = cast<SequentialType>(Ty2); 178 return isEquivalentType(STy1->getElementType(), STy2->getElementType()); 179 } 180 } 181} 182 183/// isEquivalentOperation - determine whether the two operations are the same 184/// except that pointer-to-A and pointer-to-B are equivalent. This should be 185/// kept in sync with Instruction::isSameOperationAs. 186static bool 187isEquivalentOperation(const Instruction *I1, const Instruction *I2) { 188 if (I1->getOpcode() != I2->getOpcode() || 189 I1->getNumOperands() != I2->getNumOperands() || 190 !isEquivalentType(I1->getType(), I2->getType()) || 191 !I1->hasSameSubclassOptionalData(I2)) 192 return false; 193 194 // We have two instructions of identical opcode and #operands. Check to see 195 // if all operands are the same type 196 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i) 197 if (!isEquivalentType(I1->getOperand(i)->getType(), 198 I2->getOperand(i)->getType())) 199 return false; 200 201 // Check special state that is a part of some instructions. 202 if (const LoadInst *LI = dyn_cast<LoadInst>(I1)) 203 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() && 204 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment(); 205 if (const StoreInst *SI = dyn_cast<StoreInst>(I1)) 206 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() && 207 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment(); 208 if (const CmpInst *CI = dyn_cast<CmpInst>(I1)) 209 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate(); 210 if (const CallInst *CI = dyn_cast<CallInst>(I1)) 211 return CI->isTailCall() == cast<CallInst>(I2)->isTailCall() && 212 CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() && 213 CI->getAttributes().getRawPointer() == 214 cast<CallInst>(I2)->getAttributes().getRawPointer(); 215 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1)) 216 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() && 217 CI->getAttributes().getRawPointer() == 218 cast<InvokeInst>(I2)->getAttributes().getRawPointer(); 219 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) { 220 if (IVI->getNumIndices() != cast<InsertValueInst>(I2)->getNumIndices()) 221 return false; 222 for (unsigned i = 0, e = IVI->getNumIndices(); i != e; ++i) 223 if (IVI->idx_begin()[i] != cast<InsertValueInst>(I2)->idx_begin()[i]) 224 return false; 225 return true; 226 } 227 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) { 228 if (EVI->getNumIndices() != cast<ExtractValueInst>(I2)->getNumIndices()) 229 return false; 230 for (unsigned i = 0, e = EVI->getNumIndices(); i != e; ++i) 231 if (EVI->idx_begin()[i] != cast<ExtractValueInst>(I2)->idx_begin()[i]) 232 return false; 233 return true; 234 } 235 236 return true; 237} 238 239static bool compare(const Value *V, const Value *U) { 240 assert(!isa<BasicBlock>(V) && !isa<BasicBlock>(U) && 241 "Must not compare basic blocks."); 242 243 assert(isEquivalentType(V->getType(), U->getType()) && 244 "Two of the same operation have operands of different type."); 245 246 // TODO: If the constant is an expression of F, we should accept that it's 247 // equal to the same expression in terms of G. 248 if (isa<Constant>(V)) 249 return V == U; 250 251 // The caller has ensured that ValueMap[V] != U. Since Arguments are 252 // pre-loaded into the ValueMap, and Instructions are added as we go, we know 253 // that this can only be a mis-match. 254 if (isa<Instruction>(V) || isa<Argument>(V)) 255 return false; 256 257 if (isa<InlineAsm>(V) && isa<InlineAsm>(U)) { 258 const InlineAsm *IAF = cast<InlineAsm>(V); 259 const InlineAsm *IAG = cast<InlineAsm>(U); 260 return IAF->getAsmString() == IAG->getAsmString() && 261 IAF->getConstraintString() == IAG->getConstraintString(); 262 } 263 264 return false; 265} 266 267static bool equals(const BasicBlock *BB1, const BasicBlock *BB2, 268 DenseMap<const Value *, const Value *> &ValueMap, 269 DenseMap<const Value *, const Value *> &SpeculationMap) { 270 // Speculatively add it anyways. If it's false, we'll notice a difference 271 // later, and this won't matter. 272 ValueMap[BB1] = BB2; 273 274 BasicBlock::const_iterator FI = BB1->begin(), FE = BB1->end(); 275 BasicBlock::const_iterator GI = BB2->begin(), GE = BB2->end(); 276 277 do { 278 if (isa<BitCastInst>(FI)) { 279 ++FI; 280 continue; 281 } 282 if (isa<BitCastInst>(GI)) { 283 ++GI; 284 continue; 285 } 286 287 if (!isEquivalentOperation(FI, GI)) 288 return false; 289 290 if (isa<GetElementPtrInst>(FI)) { 291 const GetElementPtrInst *GEPF = cast<GetElementPtrInst>(FI); 292 const GetElementPtrInst *GEPG = cast<GetElementPtrInst>(GI); 293 if (GEPF->hasAllZeroIndices() && GEPG->hasAllZeroIndices()) { 294 // It's effectively a bitcast. 295 ++FI, ++GI; 296 continue; 297 } 298 299 // TODO: we only really care about the elements before the index 300 if (FI->getOperand(0)->getType() != GI->getOperand(0)->getType()) 301 return false; 302 } 303 304 if (ValueMap[FI] == GI) { 305 ++FI, ++GI; 306 continue; 307 } 308 309 if (ValueMap[FI] != NULL) 310 return false; 311 312 for (unsigned i = 0, e = FI->getNumOperands(); i != e; ++i) { 313 Value *OpF = IgnoreBitcasts(FI->getOperand(i)); 314 Value *OpG = IgnoreBitcasts(GI->getOperand(i)); 315 316 if (ValueMap[OpF] == OpG) 317 continue; 318 319 if (ValueMap[OpF] != NULL) 320 return false; 321 322 if (OpF->getValueID() != OpG->getValueID() || 323 !isEquivalentType(OpF->getType(), OpG->getType())) 324 return false; 325 326 if (isa<PHINode>(FI)) { 327 if (SpeculationMap[OpF] == NULL) 328 SpeculationMap[OpF] = OpG; 329 else if (SpeculationMap[OpF] != OpG) 330 return false; 331 continue; 332 } else if (isa<BasicBlock>(OpF)) { 333 assert(isa<TerminatorInst>(FI) && 334 "BasicBlock referenced by non-Terminator non-PHI"); 335 // This call changes the ValueMap, hence we can't use 336 // Value *& = ValueMap[...] 337 if (!equals(cast<BasicBlock>(OpF), cast<BasicBlock>(OpG), ValueMap, 338 SpeculationMap)) 339 return false; 340 } else { 341 if (!compare(OpF, OpG)) 342 return false; 343 } 344 345 ValueMap[OpF] = OpG; 346 } 347 348 ValueMap[FI] = GI; 349 ++FI, ++GI; 350 } while (FI != FE && GI != GE); 351 352 return FI == FE && GI == GE; 353} 354 355static bool equals(const Function *F, const Function *G) { 356 // We need to recheck everything, but check the things that weren't included 357 // in the hash first. 358 359 if (F->getAttributes() != G->getAttributes()) 360 return false; 361 362 if (F->hasGC() != G->hasGC()) 363 return false; 364 365 if (F->hasGC() && F->getGC() != G->getGC()) 366 return false; 367 368 if (F->hasSection() != G->hasSection()) 369 return false; 370 371 if (F->hasSection() && F->getSection() != G->getSection()) 372 return false; 373 374 if (F->isVarArg() != G->isVarArg()) 375 return false; 376 377 // TODO: if it's internal and only used in direct calls, we could handle this 378 // case too. 379 if (F->getCallingConv() != G->getCallingConv()) 380 return false; 381 382 if (!isEquivalentType(F->getFunctionType(), G->getFunctionType())) 383 return false; 384 385 DenseMap<const Value *, const Value *> ValueMap; 386 DenseMap<const Value *, const Value *> SpeculationMap; 387 ValueMap[F] = G; 388 389 assert(F->arg_size() == G->arg_size() && 390 "Identical functions have a different number of args."); 391 392 for (Function::const_arg_iterator fi = F->arg_begin(), gi = G->arg_begin(), 393 fe = F->arg_end(); fi != fe; ++fi, ++gi) 394 ValueMap[fi] = gi; 395 396 if (!equals(&F->getEntryBlock(), &G->getEntryBlock(), ValueMap, 397 SpeculationMap)) 398 return false; 399 400 for (DenseMap<const Value *, const Value *>::iterator 401 I = SpeculationMap.begin(), E = SpeculationMap.end(); I != E; ++I) { 402 if (ValueMap[I->first] != I->second) 403 return false; 404 } 405 406 return true; 407} 408 409// ===----------------------------------------------------------------------=== 410// Folding of functions 411// ===----------------------------------------------------------------------=== 412 413// Cases: 414// * F is external strong, G is external strong: 415// turn G into a thunk to F (1) 416// * F is external strong, G is external weak: 417// turn G into a thunk to F (1) 418// * F is external weak, G is external weak: 419// unfoldable 420// * F is external strong, G is internal: 421// address of G taken: 422// turn G into a thunk to F (1) 423// address of G not taken: 424// make G an alias to F (2) 425// * F is internal, G is external weak 426// address of F is taken: 427// turn G into a thunk to F (1) 428// address of F is not taken: 429// make G an alias of F (2) 430// * F is internal, G is internal: 431// address of F and G are taken: 432// turn G into a thunk to F (1) 433// address of G is not taken: 434// make G an alias to F (2) 435// 436// alias requires linkage == (external,local,weak) fallback to creating a thunk 437// external means 'externally visible' linkage != (internal,private) 438// internal means linkage == (internal,private) 439// weak means linkage mayBeOverridable 440// being external implies that the address is taken 441// 442// 1. turn G into a thunk to F 443// 2. make G an alias to F 444 445enum LinkageCategory { 446 ExternalStrong, 447 ExternalWeak, 448 Internal 449}; 450 451static LinkageCategory categorize(const Function *F) { 452 switch (F->getLinkage()) { 453 case GlobalValue::InternalLinkage: 454 case GlobalValue::PrivateLinkage: 455 case GlobalValue::LinkerPrivateLinkage: 456 return Internal; 457 458 case GlobalValue::WeakAnyLinkage: 459 case GlobalValue::WeakODRLinkage: 460 case GlobalValue::ExternalWeakLinkage: 461 return ExternalWeak; 462 463 case GlobalValue::ExternalLinkage: 464 case GlobalValue::AvailableExternallyLinkage: 465 case GlobalValue::LinkOnceAnyLinkage: 466 case GlobalValue::LinkOnceODRLinkage: 467 case GlobalValue::AppendingLinkage: 468 case GlobalValue::DLLImportLinkage: 469 case GlobalValue::DLLExportLinkage: 470 case GlobalValue::CommonLinkage: 471 return ExternalStrong; 472 } 473 474 llvm_unreachable("Unknown LinkageType."); 475 return ExternalWeak; 476} 477 478static void ThunkGToF(Function *F, Function *G) { 479 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "", 480 G->getParent()); 481 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG); 482 483 std::vector<Value *> Args; 484 unsigned i = 0; 485 const FunctionType *FFTy = F->getFunctionType(); 486 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end(); 487 AI != AE; ++AI) { 488 if (FFTy->getParamType(i) == AI->getType()) 489 Args.push_back(AI); 490 else { 491 Value *BCI = new BitCastInst(AI, FFTy->getParamType(i), "", BB); 492 Args.push_back(BCI); 493 } 494 ++i; 495 } 496 497 CallInst *CI = CallInst::Create(F, Args.begin(), Args.end(), "", BB); 498 CI->setTailCall(); 499 CI->setCallingConv(F->getCallingConv()); 500 if (NewG->getReturnType()->isVoidTy()) { 501 ReturnInst::Create(F->getContext(), BB); 502 } else if (CI->getType() != NewG->getReturnType()) { 503 Value *BCI = new BitCastInst(CI, NewG->getReturnType(), "", BB); 504 ReturnInst::Create(F->getContext(), BCI, BB); 505 } else { 506 ReturnInst::Create(F->getContext(), CI, BB); 507 } 508 509 NewG->copyAttributesFrom(G); 510 NewG->takeName(G); 511 G->replaceAllUsesWith(NewG); 512 G->eraseFromParent(); 513 514 // TODO: look at direct callers to G and make them all direct callers to F. 515} 516 517static void AliasGToF(Function *F, Function *G) { 518 if (!G->hasExternalLinkage() && !G->hasLocalLinkage() && !G->hasWeakLinkage()) 519 return ThunkGToF(F, G); 520 521 GlobalAlias *GA = new GlobalAlias( 522 G->getType(), G->getLinkage(), "", 523 ConstantExpr::getBitCast(F, G->getType()), G->getParent()); 524 F->setAlignment(std::max(F->getAlignment(), G->getAlignment())); 525 GA->takeName(G); 526 GA->setVisibility(G->getVisibility()); 527 G->replaceAllUsesWith(GA); 528 G->eraseFromParent(); 529} 530 531static bool fold(std::vector<Function *> &FnVec, unsigned i, unsigned j) { 532 Function *F = FnVec[i]; 533 Function *G = FnVec[j]; 534 535 LinkageCategory catF = categorize(F); 536 LinkageCategory catG = categorize(G); 537 538 if (catF == ExternalWeak || (catF == Internal && catG == ExternalStrong)) { 539 std::swap(FnVec[i], FnVec[j]); 540 std::swap(F, G); 541 std::swap(catF, catG); 542 } 543 544 switch (catF) { 545 case ExternalStrong: 546 switch (catG) { 547 case ExternalStrong: 548 case ExternalWeak: 549 ThunkGToF(F, G); 550 break; 551 case Internal: 552 if (G->hasAddressTaken()) 553 ThunkGToF(F, G); 554 else 555 AliasGToF(F, G); 556 break; 557 } 558 break; 559 560 case ExternalWeak: { 561 assert(catG == ExternalWeak); 562 563 // Make them both thunks to the same internal function. 564 F->setAlignment(std::max(F->getAlignment(), G->getAlignment())); 565 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "", 566 F->getParent()); 567 H->copyAttributesFrom(F); 568 H->takeName(F); 569 F->replaceAllUsesWith(H); 570 571 ThunkGToF(F, G); 572 ThunkGToF(F, H); 573 574 F->setLinkage(GlobalValue::InternalLinkage); 575 } break; 576 577 case Internal: 578 switch (catG) { 579 case ExternalStrong: 580 llvm_unreachable(0); 581 // fall-through 582 case ExternalWeak: 583 if (F->hasAddressTaken()) 584 ThunkGToF(F, G); 585 else 586 AliasGToF(F, G); 587 break; 588 case Internal: { 589 bool addrTakenF = F->hasAddressTaken(); 590 bool addrTakenG = G->hasAddressTaken(); 591 if (!addrTakenF && addrTakenG) { 592 std::swap(FnVec[i], FnVec[j]); 593 std::swap(F, G); 594 std::swap(addrTakenF, addrTakenG); 595 } 596 597 if (addrTakenF && addrTakenG) { 598 ThunkGToF(F, G); 599 } else { 600 assert(!addrTakenG); 601 AliasGToF(F, G); 602 } 603 } break; 604 } 605 break; 606 } 607 608 ++NumFunctionsMerged; 609 return true; 610} 611 612// ===----------------------------------------------------------------------=== 613// Pass definition 614// ===----------------------------------------------------------------------=== 615 616bool MergeFunctions::runOnModule(Module &M) { 617 bool Changed = false; 618 619 std::map<unsigned long, std::vector<Function *> > FnMap; 620 621 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) { 622 if (F->isDeclaration() || F->isIntrinsic()) 623 continue; 624 625 FnMap[fhash(F)].push_back(F); 626 } 627 628 // TODO: instead of running in a loop, we could also fold functions in 629 // callgraph order. Constructing the CFG probably isn't cheaper than just 630 // running in a loop, unless it happened to already be available. 631 632 bool LocalChanged; 633 do { 634 LocalChanged = false; 635 DEBUG(dbgs() << "size: " << FnMap.size() << "\n"); 636 for (std::map<unsigned long, std::vector<Function *> >::iterator 637 I = FnMap.begin(), E = FnMap.end(); I != E; ++I) { 638 std::vector<Function *> &FnVec = I->second; 639 DEBUG(dbgs() << "fhash (" << I->first << "): " << FnVec.size() << "\n"); 640 641 for (int i = 0, e = FnVec.size(); i != e; ++i) { 642 for (int j = i + 1; j != e; ++j) { 643 bool isEqual = equals(FnVec[i], FnVec[j]); 644 645 DEBUG(dbgs() << " " << FnVec[i]->getName() 646 << (isEqual ? " == " : " != ") 647 << FnVec[j]->getName() << "\n"); 648 649 if (isEqual) { 650 if (fold(FnVec, i, j)) { 651 LocalChanged = true; 652 FnVec.erase(FnVec.begin() + j); 653 --j, --e; 654 } 655 } 656 } 657 } 658 659 } 660 Changed |= LocalChanged; 661 } while (LocalChanged); 662 663 return Changed; 664} 665