MergeFunctions.cpp revision 1f74590e9d1b9cf0f1f81a156efea73f76546e05
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. If both functions are 21// overridable, we move the functionality into a new internal function and 22// leave two overridable thunks to it. 23// 24//===----------------------------------------------------------------------===// 25// 26// Future work: 27// 28// * virtual functions. 29// 30// Many functions have their address taken by the virtual function table for 31// the object they belong to. However, as long as it's only used for a lookup 32// and call, this is irrelevant, and we'd like to fold such implementations. 33// 34// * switch from n^2 pair-wise comparisons to an n-way comparison for each 35// bucket. 36// 37// * be smarter about bitcast. 38// 39// In order to fold functions, we will sometimes add either bitcast instructions 40// or bitcast constant expressions. Unfortunately, this can confound further 41// analysis since the two functions differ where one has a bitcast and the 42// other doesn't. We should learn to peer through bitcasts without imposing bad 43// performance properties. 44// 45// * emit aliases for ELF 46// 47// ELF supports symbol aliases which are represented with GlobalAlias in the 48// Module, and we could emit them in the case that the addresses don't need to 49// be distinct. The problem is that not all object formats support equivalent 50// functionality. There's a few approaches to this problem; 51// a) teach codegen to lower global aliases to thunks on platforms which don't 52// support them. 53// b) always emit thunks, and create a separate thunk-to-alias pass which 54// runs on ELF systems. This has the added benefit of transforming other 55// thunks such as those produced by a C++ frontend into aliases when legal 56// to do so. 57// 58//===----------------------------------------------------------------------===// 59 60#define DEBUG_TYPE "mergefunc" 61#include "llvm/Transforms/IPO.h" 62#include "llvm/ADT/DenseMap.h" 63#include "llvm/ADT/FoldingSet.h" 64#include "llvm/ADT/SmallSet.h" 65#include "llvm/ADT/Statistic.h" 66#include "llvm/Constants.h" 67#include "llvm/InlineAsm.h" 68#include "llvm/Instructions.h" 69#include "llvm/LLVMContext.h" 70#include "llvm/Module.h" 71#include "llvm/Pass.h" 72#include "llvm/Support/CallSite.h" 73#include "llvm/Support/Debug.h" 74#include "llvm/Support/ErrorHandling.h" 75#include "llvm/Support/raw_ostream.h" 76#include "llvm/Target/TargetData.h" 77#include <map> 78#include <vector> 79using namespace llvm; 80 81STATISTIC(NumFunctionsMerged, "Number of functions merged"); 82 83namespace { 84 /// MergeFunctions finds functions which will generate identical machine code, 85 /// by considering all pointer types to be equivalent. Once identified, 86 /// MergeFunctions will fold them by replacing a call to one to a call to a 87 /// bitcast of the other. 88 /// 89 struct MergeFunctions : public ModulePass { 90 static char ID; // Pass identification, replacement for typeid 91 MergeFunctions() : ModulePass(&ID) {} 92 93 bool runOnModule(Module &M); 94 }; 95} 96 97char MergeFunctions::ID = 0; 98INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false); 99 100ModulePass *llvm::createMergeFunctionsPass() { 101 return new MergeFunctions(); 102} 103 104// ===----------------------------------------------------------------------=== 105// Comparison of functions 106// ===----------------------------------------------------------------------=== 107namespace { 108class FunctionComparator { 109public: 110 FunctionComparator(TargetData *TD, Function *F1, Function *F2) 111 : F1(F1), F2(F2), TD(TD) {} 112 113 // Compare - test whether the two functions have equivalent behaviour. 114 bool Compare(); 115 116private: 117 // Compare - test whether two basic blocks have equivalent behaviour. 118 bool Compare(const BasicBlock *BB1, const BasicBlock *BB2); 119 120 // getDomain - a value's domain is its parent function if it is specific to a 121 // function, or NULL otherwise. 122 const Function *getDomain(const Value *V) const; 123 124 // Enumerate - Assign or look up previously assigned numbers for the two 125 // values, and return whether the numbers are equal. Numbers are assigned in 126 // the order visited. 127 bool Enumerate(const Value *V1, const Value *V2); 128 129 // isEquivalentOperation - Compare two Instructions for equivalence, similar 130 // to Instruction::isSameOperationAs but with modifications to the type 131 // comparison. 132 bool isEquivalentOperation(const Instruction *I1, 133 const Instruction *I2) const; 134 135 // isEquivalentGEP - Compare two GEPs for equivalent pointer arithmetic. 136 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2); 137 bool isEquivalentGEP(const GetElementPtrInst *GEP1, 138 const GetElementPtrInst *GEP2) { 139 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2)); 140 } 141 142 // isEquivalentType - Compare two Types, treating all pointer types as equal. 143 bool isEquivalentType(const Type *Ty1, const Type *Ty2) const; 144 145 // The two functions undergoing comparison. 146 Function *F1, *F2; 147 148 TargetData *TD; 149 150 typedef DenseMap<const Value *, unsigned long> IDMap; 151 IDMap Map; 152 DenseMap<const Function *, IDMap> Domains; 153 DenseMap<const Function *, unsigned long> DomainCount; 154}; 155} 156 157/// Compute a number which is guaranteed to be equal for two equivalent 158/// functions, but is very likely to be different for different functions. This 159/// needs to be computed as efficiently as possible. 160static unsigned long ProfileFunction(const Function *F) { 161 const FunctionType *FTy = F->getFunctionType(); 162 163 FoldingSetNodeID ID; 164 ID.AddInteger(F->size()); 165 ID.AddInteger(F->getCallingConv()); 166 ID.AddBoolean(F->hasGC()); 167 ID.AddBoolean(FTy->isVarArg()); 168 ID.AddInteger(FTy->getReturnType()->getTypeID()); 169 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i) 170 ID.AddInteger(FTy->getParamType(i)->getTypeID()); 171 return ID.ComputeHash(); 172} 173 174/// isEquivalentType - any two pointers are equivalent. Otherwise, standard 175/// type equivalence rules apply. 176bool FunctionComparator::isEquivalentType(const Type *Ty1, 177 const Type *Ty2) const { 178 if (Ty1 == Ty2) 179 return true; 180 if (Ty1->getTypeID() != Ty2->getTypeID()) 181 return false; 182 183 switch(Ty1->getTypeID()) { 184 default: 185 llvm_unreachable("Unknown type!"); 186 // Fall through in Release mode. 187 case Type::IntegerTyID: 188 case Type::OpaqueTyID: 189 // Ty1 == Ty2 would have returned true earlier. 190 return false; 191 192 case Type::VoidTyID: 193 case Type::FloatTyID: 194 case Type::DoubleTyID: 195 case Type::X86_FP80TyID: 196 case Type::FP128TyID: 197 case Type::PPC_FP128TyID: 198 case Type::LabelTyID: 199 case Type::MetadataTyID: 200 return true; 201 202 case Type::PointerTyID: { 203 const PointerType *PTy1 = cast<PointerType>(Ty1); 204 const PointerType *PTy2 = cast<PointerType>(Ty2); 205 return PTy1->getAddressSpace() == PTy2->getAddressSpace(); 206 } 207 208 case Type::StructTyID: { 209 const StructType *STy1 = cast<StructType>(Ty1); 210 const StructType *STy2 = cast<StructType>(Ty2); 211 if (STy1->getNumElements() != STy2->getNumElements()) 212 return false; 213 214 if (STy1->isPacked() != STy2->isPacked()) 215 return false; 216 217 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) { 218 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i))) 219 return false; 220 } 221 return true; 222 } 223 224 case Type::UnionTyID: { 225 const UnionType *UTy1 = cast<UnionType>(Ty1); 226 const UnionType *UTy2 = cast<UnionType>(Ty2); 227 228 // TODO: we could be fancy with union(A, union(A, B)) === union(A, B), etc. 229 if (UTy1->getNumElements() != UTy2->getNumElements()) 230 return false; 231 232 for (unsigned i = 0, e = UTy1->getNumElements(); i != e; ++i) { 233 if (!isEquivalentType(UTy1->getElementType(i), UTy2->getElementType(i))) 234 return false; 235 } 236 return true; 237 } 238 239 case Type::FunctionTyID: { 240 const FunctionType *FTy1 = cast<FunctionType>(Ty1); 241 const FunctionType *FTy2 = cast<FunctionType>(Ty2); 242 if (FTy1->getNumParams() != FTy2->getNumParams() || 243 FTy1->isVarArg() != FTy2->isVarArg()) 244 return false; 245 246 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType())) 247 return false; 248 249 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) { 250 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i))) 251 return false; 252 } 253 return true; 254 } 255 256 case Type::ArrayTyID: { 257 const ArrayType *ATy1 = cast<ArrayType>(Ty1); 258 const ArrayType *ATy2 = cast<ArrayType>(Ty2); 259 return ATy1->getNumElements() == ATy2->getNumElements() && 260 isEquivalentType(ATy1->getElementType(), ATy2->getElementType()); 261 } 262 case Type::VectorTyID: { 263 const VectorType *VTy1 = cast<VectorType>(Ty1); 264 const VectorType *VTy2 = cast<VectorType>(Ty2); 265 return VTy1->getNumElements() == VTy2->getNumElements() && 266 isEquivalentType(VTy1->getElementType(), VTy2->getElementType()); 267 } 268 } 269} 270 271/// isEquivalentOperation - determine whether the two operations are the same 272/// except that pointer-to-A and pointer-to-B are equivalent. This should be 273/// kept in sync with Instruction::isSameOperationAs. 274bool FunctionComparator::isEquivalentOperation(const Instruction *I1, 275 const Instruction *I2) const { 276 if (I1->getOpcode() != I2->getOpcode() || 277 I1->getNumOperands() != I2->getNumOperands() || 278 !isEquivalentType(I1->getType(), I2->getType()) || 279 !I1->hasSameSubclassOptionalData(I2)) 280 return false; 281 282 // We have two instructions of identical opcode and #operands. Check to see 283 // if all operands are the same type 284 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i) 285 if (!isEquivalentType(I1->getOperand(i)->getType(), 286 I2->getOperand(i)->getType())) 287 return false; 288 289 // Check special state that is a part of some instructions. 290 if (const LoadInst *LI = dyn_cast<LoadInst>(I1)) 291 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() && 292 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment(); 293 if (const StoreInst *SI = dyn_cast<StoreInst>(I1)) 294 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() && 295 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment(); 296 if (const CmpInst *CI = dyn_cast<CmpInst>(I1)) 297 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate(); 298 if (const CallInst *CI = dyn_cast<CallInst>(I1)) 299 return CI->isTailCall() == cast<CallInst>(I2)->isTailCall() && 300 CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() && 301 CI->getAttributes().getRawPointer() == 302 cast<CallInst>(I2)->getAttributes().getRawPointer(); 303 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1)) 304 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() && 305 CI->getAttributes().getRawPointer() == 306 cast<InvokeInst>(I2)->getAttributes().getRawPointer(); 307 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) { 308 if (IVI->getNumIndices() != cast<InsertValueInst>(I2)->getNumIndices()) 309 return false; 310 for (unsigned i = 0, e = IVI->getNumIndices(); i != e; ++i) 311 if (IVI->idx_begin()[i] != cast<InsertValueInst>(I2)->idx_begin()[i]) 312 return false; 313 return true; 314 } 315 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) { 316 if (EVI->getNumIndices() != cast<ExtractValueInst>(I2)->getNumIndices()) 317 return false; 318 for (unsigned i = 0, e = EVI->getNumIndices(); i != e; ++i) 319 if (EVI->idx_begin()[i] != cast<ExtractValueInst>(I2)->idx_begin()[i]) 320 return false; 321 return true; 322 } 323 324 return true; 325} 326 327/// isEquivalentGEP - determine whether two GEP operations perform the same 328/// underlying arithmetic. 329bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1, 330 const GEPOperator *GEP2) { 331 // When we have target data, we can reduce the GEP down to the value in bytes 332 // added to the address. 333 if (TD && GEP1->hasAllConstantIndices() && GEP2->hasAllConstantIndices()) { 334 SmallVector<Value *, 8> Indices1(GEP1->idx_begin(), GEP1->idx_end()); 335 SmallVector<Value *, 8> Indices2(GEP2->idx_begin(), GEP2->idx_end()); 336 uint64_t Offset1 = TD->getIndexedOffset(GEP1->getPointerOperandType(), 337 Indices1.data(), Indices1.size()); 338 uint64_t Offset2 = TD->getIndexedOffset(GEP2->getPointerOperandType(), 339 Indices2.data(), Indices2.size()); 340 return Offset1 == Offset2; 341 } 342 343 if (GEP1->getPointerOperand()->getType() != 344 GEP2->getPointerOperand()->getType()) 345 return false; 346 347 if (GEP1->getNumOperands() != GEP2->getNumOperands()) 348 return false; 349 350 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) { 351 if (!Enumerate(GEP1->getOperand(i), GEP2->getOperand(i))) 352 return false; 353 } 354 355 return true; 356} 357 358/// getDomain - a value's domain is its parent function if it is specific to a 359/// function, or NULL otherwise. 360const Function *FunctionComparator::getDomain(const Value *V) const { 361 if (const Argument *A = dyn_cast<Argument>(V)) { 362 return A->getParent(); 363 } else if (const BasicBlock *BB = dyn_cast<BasicBlock>(V)) { 364 return BB->getParent(); 365 } else if (const Instruction *I = dyn_cast<Instruction>(V)) { 366 return I->getParent()->getParent(); 367 } 368 return NULL; 369} 370 371/// Enumerate - Compare two values used by the two functions under pair-wise 372/// comparison. If this is the first time the values are seen, they're added to 373/// the mapping so that we will detect mismatches on next use. 374bool FunctionComparator::Enumerate(const Value *V1, const Value *V2) { 375 // Check for function @f1 referring to itself and function @f2 referring to 376 // itself, or referring to each other, or both referring to either of them. 377 // They're all equivalent if the two functions are otherwise equivalent. 378 if (V1 == F1 || V1 == F2) 379 if (V2 == F1 || V2 == F2) 380 return true; 381 382 // TODO: constant expressions with GEP or references to F1 or F2. 383 if (isa<Constant>(V1)) 384 return V1 == V2; 385 386 if (isa<InlineAsm>(V1) && isa<InlineAsm>(V2)) { 387 const InlineAsm *IA1 = cast<InlineAsm>(V1); 388 const InlineAsm *IA2 = cast<InlineAsm>(V2); 389 return IA1->getAsmString() == IA2->getAsmString() && 390 IA1->getConstraintString() == IA2->getConstraintString(); 391 } 392 393 // We enumerate constants globally and arguments, basic blocks or 394 // instructions within the function they belong to. 395 const Function *Domain1 = getDomain(V1); 396 const Function *Domain2 = getDomain(V2); 397 398 // The domains have to either be both NULL, or F1, F2. 399 if (Domain1 != Domain2) 400 if (Domain1 != F1 && Domain1 != F2) 401 return false; 402 403 IDMap &Map1 = Domains[Domain1]; 404 unsigned long &ID1 = Map1[V1]; 405 if (!ID1) 406 ID1 = ++DomainCount[Domain1]; 407 408 IDMap &Map2 = Domains[Domain2]; 409 unsigned long &ID2 = Map2[V2]; 410 if (!ID2) 411 ID2 = ++DomainCount[Domain2]; 412 413 return ID1 == ID2; 414} 415 416// Compare - test whether two basic blocks have equivalent behaviour. 417bool FunctionComparator::Compare(const BasicBlock *BB1, const BasicBlock *BB2) { 418 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end(); 419 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end(); 420 421 do { 422 if (!Enumerate(F1I, F2I)) 423 return false; 424 425 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) { 426 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I); 427 if (!GEP2) 428 return false; 429 430 if (!Enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand())) 431 return false; 432 433 if (!isEquivalentGEP(GEP1, GEP2)) 434 return false; 435 } else { 436 if (!isEquivalentOperation(F1I, F2I)) 437 return false; 438 439 assert(F1I->getNumOperands() == F2I->getNumOperands()); 440 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) { 441 Value *OpF1 = F1I->getOperand(i); 442 Value *OpF2 = F2I->getOperand(i); 443 444 if (!Enumerate(OpF1, OpF2)) 445 return false; 446 447 if (OpF1->getValueID() != OpF2->getValueID() || 448 !isEquivalentType(OpF1->getType(), OpF2->getType())) 449 return false; 450 } 451 } 452 453 ++F1I, ++F2I; 454 } while (F1I != F1E && F2I != F2E); 455 456 return F1I == F1E && F2I == F2E; 457} 458 459bool FunctionComparator::Compare() { 460 // We need to recheck everything, but check the things that weren't included 461 // in the hash first. 462 463 if (F1->getAttributes() != F2->getAttributes()) 464 return false; 465 466 if (F1->hasGC() != F2->hasGC()) 467 return false; 468 469 if (F1->hasGC() && F1->getGC() != F2->getGC()) 470 return false; 471 472 if (F1->hasSection() != F2->hasSection()) 473 return false; 474 475 if (F1->hasSection() && F1->getSection() != F2->getSection()) 476 return false; 477 478 if (F1->isVarArg() != F2->isVarArg()) 479 return false; 480 481 // TODO: if it's internal and only used in direct calls, we could handle this 482 // case too. 483 if (F1->getCallingConv() != F2->getCallingConv()) 484 return false; 485 486 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType())) 487 return false; 488 489 assert(F1->arg_size() == F2->arg_size() && 490 "Identical functions have a different number of args."); 491 492 // Visit the arguments so that they get enumerated in the order they're 493 // passed in. 494 for (Function::const_arg_iterator f1i = F1->arg_begin(), 495 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) { 496 if (!Enumerate(f1i, f2i)) 497 llvm_unreachable("Arguments repeat"); 498 } 499 500 // We need to do an ordered walk since the actual ordering of the blocks in 501 // the linked list is immaterial. Our walk starts at the entry block for both 502 // functions, then takes each block from each terminator in order. As an 503 // artifact, this also means that unreachable blocks are ignored. 504 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs; 505 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1. 506 F1BBs.push_back(&F1->getEntryBlock()); 507 F2BBs.push_back(&F2->getEntryBlock()); 508 VisitedBBs.insert(F1BBs[0]); 509 while (!F1BBs.empty()) { 510 const BasicBlock *F1BB = F1BBs.pop_back_val(); 511 const BasicBlock *F2BB = F2BBs.pop_back_val(); 512 if (!Enumerate(F1BB, F2BB) || !Compare(F1BB, F2BB)) 513 return false; 514 const TerminatorInst *F1TI = F1BB->getTerminator(); 515 const TerminatorInst *F2TI = F2BB->getTerminator(); 516 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors()); 517 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) { 518 if (!VisitedBBs.insert(F1TI->getSuccessor(i))) 519 continue; 520 F1BBs.push_back(F1TI->getSuccessor(i)); 521 F2BBs.push_back(F2TI->getSuccessor(i)); 522 } 523 } 524 return true; 525} 526 527// ===----------------------------------------------------------------------=== 528// Folding of functions 529// ===----------------------------------------------------------------------=== 530 531// Cases: 532// * F is external strong, G is external strong: 533// turn G into a thunk to F (1) 534// * F is external strong, G is external weak: 535// turn G into a thunk to F (1) 536// * F is external weak, G is external weak: 537// unfoldable 538// * F is external strong, G is internal: 539// address of G taken: 540// turn G into a thunk to F (1) 541// address of G not taken: 542// make G an alias to F (2) 543// * F is internal, G is external weak 544// address of F is taken: 545// turn G into a thunk to F (1) 546// address of F is not taken: 547// make G an alias of F (2) 548// * F is internal, G is internal: 549// address of F and G are taken: 550// turn G into a thunk to F (1) 551// address of G is not taken: 552// make G an alias to F (2) 553// 554// alias requires linkage == (external,local,weak) fallback to creating a thunk 555// external means 'externally visible' linkage != (internal,private) 556// internal means linkage == (internal,private) 557// weak means linkage mayBeOverridable 558// being external implies that the address is taken 559// 560// 1. turn G into a thunk to F 561// 2. make G an alias to F 562 563enum LinkageCategory { 564 ExternalStrong, 565 ExternalWeak, 566 Internal 567}; 568 569static LinkageCategory categorize(const Function *F) { 570 switch (F->getLinkage()) { 571 case GlobalValue::InternalLinkage: 572 case GlobalValue::PrivateLinkage: 573 case GlobalValue::LinkerPrivateLinkage: 574 return Internal; 575 576 case GlobalValue::WeakAnyLinkage: 577 case GlobalValue::WeakODRLinkage: 578 case GlobalValue::ExternalWeakLinkage: 579 case GlobalValue::LinkerPrivateWeakLinkage: 580 return ExternalWeak; 581 582 case GlobalValue::ExternalLinkage: 583 case GlobalValue::AvailableExternallyLinkage: 584 case GlobalValue::LinkOnceAnyLinkage: 585 case GlobalValue::LinkOnceODRLinkage: 586 case GlobalValue::AppendingLinkage: 587 case GlobalValue::DLLImportLinkage: 588 case GlobalValue::DLLExportLinkage: 589 case GlobalValue::CommonLinkage: 590 return ExternalStrong; 591 } 592 593 llvm_unreachable("Unknown LinkageType."); 594 return ExternalWeak; 595} 596 597static void ThunkGToF(Function *F, Function *G) { 598 if (!G->mayBeOverridden()) { 599 // Redirect direct callers of G to F. 600 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType()); 601 for (Value::use_iterator UI = G->use_begin(), UE = G->use_end(); 602 UI != UE;) { 603 Value::use_iterator TheIter = UI; 604 ++UI; 605 CallSite CS(*TheIter); 606 if (CS && CS.isCallee(TheIter)) 607 TheIter.getUse().set(BitcastF); 608 } 609 } 610 611 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "", 612 G->getParent()); 613 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG); 614 615 SmallVector<Value *, 16> Args; 616 unsigned i = 0; 617 const FunctionType *FFTy = F->getFunctionType(); 618 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end(); 619 AI != AE; ++AI) { 620 if (FFTy->getParamType(i) == AI->getType()) { 621 Args.push_back(AI); 622 } else { 623 Args.push_back(new BitCastInst(AI, FFTy->getParamType(i), "", BB)); 624 } 625 ++i; 626 } 627 628 CallInst *CI = CallInst::Create(F, Args.begin(), Args.end(), "", BB); 629 CI->setTailCall(); 630 CI->setCallingConv(F->getCallingConv()); 631 if (NewG->getReturnType()->isVoidTy()) { 632 ReturnInst::Create(F->getContext(), BB); 633 } else if (CI->getType() != NewG->getReturnType()) { 634 Value *BCI = new BitCastInst(CI, NewG->getReturnType(), "", BB); 635 ReturnInst::Create(F->getContext(), BCI, BB); 636 } else { 637 ReturnInst::Create(F->getContext(), CI, BB); 638 } 639 640 NewG->copyAttributesFrom(G); 641 NewG->takeName(G); 642 G->replaceAllUsesWith(NewG); 643 G->eraseFromParent(); 644} 645 646static void AliasGToF(Function *F, Function *G) { 647 // Darwin will trigger llvm_unreachable if asked to codegen an alias. 648 return ThunkGToF(F, G); 649 650#if 0 651 if (!G->hasExternalLinkage() && !G->hasLocalLinkage() && !G->hasWeakLinkage()) 652 return ThunkGToF(F, G); 653 654 GlobalAlias *GA = new GlobalAlias( 655 G->getType(), G->getLinkage(), "", 656 ConstantExpr::getBitCast(F, G->getType()), G->getParent()); 657 F->setAlignment(std::max(F->getAlignment(), G->getAlignment())); 658 GA->takeName(G); 659 GA->setVisibility(G->getVisibility()); 660 G->replaceAllUsesWith(GA); 661 G->eraseFromParent(); 662#endif 663} 664 665static bool fold(std::vector<Function *> &FnVec, unsigned i, unsigned j) { 666 Function *F = FnVec[i]; 667 Function *G = FnVec[j]; 668 669 LinkageCategory catF = categorize(F); 670 LinkageCategory catG = categorize(G); 671 672 if (catF == ExternalWeak || (catF == Internal && catG == ExternalStrong)) { 673 std::swap(FnVec[i], FnVec[j]); 674 std::swap(F, G); 675 std::swap(catF, catG); 676 } 677 678 switch (catF) { 679 case ExternalStrong: 680 switch (catG) { 681 case ExternalStrong: 682 case ExternalWeak: 683 ThunkGToF(F, G); 684 break; 685 case Internal: 686 if (G->hasAddressTaken()) 687 ThunkGToF(F, G); 688 else 689 AliasGToF(F, G); 690 break; 691 } 692 break; 693 694 case ExternalWeak: { 695 assert(catG == ExternalWeak); 696 697 // Make them both thunks to the same internal function. 698 F->setAlignment(std::max(F->getAlignment(), G->getAlignment())); 699 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "", 700 F->getParent()); 701 H->copyAttributesFrom(F); 702 H->takeName(F); 703 F->replaceAllUsesWith(H); 704 705 ThunkGToF(F, G); 706 ThunkGToF(F, H); 707 708 F->setLinkage(GlobalValue::InternalLinkage); 709 } break; 710 711 case Internal: 712 switch (catG) { 713 case ExternalStrong: 714 llvm_unreachable(0); 715 // fall-through 716 case ExternalWeak: 717 if (F->hasAddressTaken()) 718 ThunkGToF(F, G); 719 else 720 AliasGToF(F, G); 721 break; 722 case Internal: { 723 bool addrTakenF = F->hasAddressTaken(); 724 bool addrTakenG = G->hasAddressTaken(); 725 if (!addrTakenF && addrTakenG) { 726 std::swap(FnVec[i], FnVec[j]); 727 std::swap(F, G); 728 std::swap(addrTakenF, addrTakenG); 729 } 730 731 if (addrTakenF && addrTakenG) { 732 ThunkGToF(F, G); 733 } else { 734 assert(!addrTakenG); 735 AliasGToF(F, G); 736 } 737 } break; 738 } break; 739 } 740 741 ++NumFunctionsMerged; 742 return true; 743} 744 745// ===----------------------------------------------------------------------=== 746// Pass definition 747// ===----------------------------------------------------------------------=== 748 749bool MergeFunctions::runOnModule(Module &M) { 750 bool Changed = false; 751 752 std::map<unsigned long, std::vector<Function *> > FnMap; 753 754 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) { 755 if (F->isDeclaration()) 756 continue; 757 758 FnMap[ProfileFunction(F)].push_back(F); 759 } 760 761 TargetData *TD = getAnalysisIfAvailable<TargetData>(); 762 763 bool LocalChanged; 764 do { 765 LocalChanged = false; 766 DEBUG(dbgs() << "size: " << FnMap.size() << "\n"); 767 for (std::map<unsigned long, std::vector<Function *> >::iterator 768 I = FnMap.begin(), E = FnMap.end(); I != E; ++I) { 769 std::vector<Function *> &FnVec = I->second; 770 DEBUG(dbgs() << "hash (" << I->first << "): " << FnVec.size() << "\n"); 771 772 for (int i = 0, e = FnVec.size(); i != e; ++i) { 773 for (int j = i + 1; j != e; ++j) { 774 bool isEqual = FunctionComparator(TD, FnVec[i], FnVec[j]).Compare(); 775 776 DEBUG(dbgs() << " " << FnVec[i]->getName() 777 << (isEqual ? " == " : " != ") 778 << FnVec[j]->getName() << "\n"); 779 780 if (isEqual) { 781 if (fold(FnVec, i, j)) { 782 LocalChanged = true; 783 FnVec.erase(FnVec.begin() + j); 784 --j, --e; 785 } 786 } 787 } 788 } 789 790 } 791 Changed |= LocalChanged; 792 } while (LocalChanged); 793 794 return Changed; 795} 796