MergeFunctions.cpp revision b38824f866447ccf8dd0c76656755b05bcede1b1
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 functions. 33// 34// * switch from n^2 pair-wise comparisons to an n-way comparison for each 35// bucket. 36// 37// * be smarter about bitcasts. 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 look through bitcasts. 43// 44//===----------------------------------------------------------------------===// 45 46#define DEBUG_TYPE "mergefunc" 47#include "llvm/Transforms/IPO.h" 48#include "llvm/ADT/DenseSet.h" 49#include "llvm/ADT/FoldingSet.h" 50#include "llvm/ADT/SmallSet.h" 51#include "llvm/ADT/Statistic.h" 52#include "llvm/ADT/STLExtras.h" 53#include "llvm/Constants.h" 54#include "llvm/InlineAsm.h" 55#include "llvm/Instructions.h" 56#include "llvm/LLVMContext.h" 57#include "llvm/Module.h" 58#include "llvm/Pass.h" 59#include "llvm/Support/CallSite.h" 60#include "llvm/Support/Debug.h" 61#include "llvm/Support/ErrorHandling.h" 62#include "llvm/Support/IRBuilder.h" 63#include "llvm/Support/ValueHandle.h" 64#include "llvm/Support/raw_ostream.h" 65#include "llvm/Target/TargetData.h" 66#include <vector> 67using namespace llvm; 68 69STATISTIC(NumFunctionsMerged, "Number of functions merged"); 70STATISTIC(NumThunksWritten, "Number of thunks generated"); 71STATISTIC(NumAliasesWritten, "Number of aliases generated"); 72STATISTIC(NumDoubleWeak, "Number of new functions created"); 73 74/// ProfileFunction - Creates a hash-code for the function which is the same 75/// for any two functions that will compare equal, without looking at the 76/// instructions inside the function. 77static unsigned ProfileFunction(const Function *F) { 78 const FunctionType *FTy = F->getFunctionType(); 79 80 FoldingSetNodeID ID; 81 ID.AddInteger(F->size()); 82 ID.AddInteger(F->getCallingConv()); 83 ID.AddBoolean(F->hasGC()); 84 ID.AddBoolean(FTy->isVarArg()); 85 ID.AddInteger(FTy->getReturnType()->getTypeID()); 86 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i) 87 ID.AddInteger(FTy->getParamType(i)->getTypeID()); 88 return ID.ComputeHash(); 89} 90 91namespace { 92 93class ComparableFunction { 94public: 95 static const ComparableFunction EmptyKey; 96 static const ComparableFunction TombstoneKey; 97 98 ComparableFunction(Function *Func, TargetData *TD) 99 : Func(Func), Hash(ProfileFunction(Func)), TD(TD) {} 100 101 Function *getFunc() const { return Func; } 102 unsigned getHash() const { return Hash; } 103 TargetData *getTD() const { return TD; } 104 105 // Drops AssertingVH reference to the function. Outside of debug mode, this 106 // does nothing. 107 void release() { 108 assert(Func && 109 "Attempted to release function twice, or release empty/tombstone!"); 110 Func = NULL; 111 } 112 113 bool &getOrInsertCachedComparison(const ComparableFunction &Other, 114 bool &inserted) const { 115 typedef DenseMap<Function *, bool>::iterator iterator; 116 std::pair<iterator, bool> p = 117 CompareResultCache.insert(std::make_pair(Other.getFunc(), false)); 118 inserted = p.second; 119 return p.first->second; 120 } 121 122private: 123 explicit ComparableFunction(unsigned Hash) 124 : Func(NULL), Hash(Hash), TD(NULL) {} 125 126 // DenseMap::grow() triggers a recomparison of all keys in the map, which is 127 // wildly expensive. This cache tries to preserve known results. 128 mutable DenseMap<Function *, bool> CompareResultCache; 129 130 AssertingVH<Function> Func; 131 unsigned Hash; 132 TargetData *TD; 133}; 134 135const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0); 136const ComparableFunction ComparableFunction::TombstoneKey = 137 ComparableFunction(1); 138 139} 140 141namespace llvm { 142 template <> 143 struct DenseMapInfo<ComparableFunction> { 144 static ComparableFunction getEmptyKey() { 145 return ComparableFunction::EmptyKey; 146 } 147 static ComparableFunction getTombstoneKey() { 148 return ComparableFunction::TombstoneKey; 149 } 150 static unsigned getHashValue(const ComparableFunction &CF) { 151 return CF.getHash(); 152 } 153 static bool isEqual(const ComparableFunction &LHS, 154 const ComparableFunction &RHS); 155 }; 156} 157 158namespace { 159 160/// MergeFunctions finds functions which will generate identical machine code, 161/// by considering all pointer types to be equivalent. Once identified, 162/// MergeFunctions will fold them by replacing a call to one to a call to a 163/// bitcast of the other. 164/// 165class MergeFunctions : public ModulePass { 166public: 167 static char ID; 168 MergeFunctions() 169 : ModulePass(ID), HasGlobalAliases(false) { 170 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry()); 171 } 172 173 bool runOnModule(Module &M); 174 175private: 176 typedef DenseSet<ComparableFunction> FnSetType; 177 178 /// A work queue of functions that may have been modified and should be 179 /// analyzed again. 180 std::vector<WeakVH> Deferred; 181 182 /// Insert a ComparableFunction into the FnSet, or merge it away if it's 183 /// equal to one that's already present. 184 bool Insert(ComparableFunction &NewF); 185 186 /// Remove a Function from the FnSet and queue it up for a second sweep of 187 /// analysis. 188 void Remove(Function *F); 189 190 /// Find the functions that use this Value and remove them from FnSet and 191 /// queue the functions. 192 void RemoveUsers(Value *V); 193 194 /// Replace all direct calls of Old with calls of New. Will bitcast New if 195 /// necessary to make types match. 196 void replaceDirectCallers(Function *Old, Function *New); 197 198 /// MergeTwoFunctions - Merge two equivalent functions. Upon completion, G 199 /// may be deleted, or may be converted into a thunk. In either case, it 200 /// should never be visited again. 201 void MergeTwoFunctions(Function *F, Function *G); 202 203 /// WriteThunkOrAlias - Replace G with a thunk or an alias to F. Deletes G. 204 void WriteThunkOrAlias(Function *F, Function *G); 205 206 /// WriteThunk - Replace G with a simple tail call to bitcast(F). Also 207 /// replace direct uses of G with bitcast(F). Deletes G. 208 void WriteThunk(Function *F, Function *G); 209 210 /// WriteAlias - Replace G with an alias to F. Deletes G. 211 void WriteAlias(Function *F, Function *G); 212 213 /// The set of all distinct functions. Use the Insert and Remove methods to 214 /// modify it. 215 FnSetType FnSet; 216 217 /// TargetData for more accurate GEP comparisons. May be NULL. 218 TargetData *TD; 219 220 /// Whether or not the target supports global aliases. 221 bool HasGlobalAliases; 222}; 223 224} // end anonymous namespace 225 226char MergeFunctions::ID = 0; 227INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false) 228 229ModulePass *llvm::createMergeFunctionsPass() { 230 return new MergeFunctions(); 231} 232 233namespace { 234/// FunctionComparator - Compares two functions to determine whether or not 235/// they will generate machine code with the same behaviour. TargetData is 236/// used if available. The comparator always fails conservatively (erring on the 237/// side of claiming that two functions are different). 238class FunctionComparator { 239public: 240 FunctionComparator(const TargetData *TD, const Function *F1, 241 const Function *F2) 242 : F1(F1), F2(F2), TD(TD), IDMap1Count(0), IDMap2Count(0) {} 243 244 /// Compare - test whether the two functions have equivalent behaviour. 245 bool Compare(); 246 247private: 248 /// Compare - test whether two basic blocks have equivalent behaviour. 249 bool Compare(const BasicBlock *BB1, const BasicBlock *BB2); 250 251 /// Enumerate - Assign or look up previously assigned numbers for the two 252 /// values, and return whether the numbers are equal. Numbers are assigned in 253 /// the order visited. 254 bool Enumerate(const Value *V1, const Value *V2); 255 256 /// isEquivalentOperation - Compare two Instructions for equivalence, similar 257 /// to Instruction::isSameOperationAs but with modifications to the type 258 /// comparison. 259 bool isEquivalentOperation(const Instruction *I1, 260 const Instruction *I2) const; 261 262 /// isEquivalentGEP - Compare two GEPs for equivalent pointer arithmetic. 263 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2); 264 bool isEquivalentGEP(const GetElementPtrInst *GEP1, 265 const GetElementPtrInst *GEP2) { 266 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2)); 267 } 268 269 /// isEquivalentType - Compare two Types, treating all pointer types as equal. 270 bool isEquivalentType(const Type *Ty1, const Type *Ty2) const; 271 272 // The two functions undergoing comparison. 273 const Function *F1, *F2; 274 275 const TargetData *TD; 276 277 typedef DenseMap<const Value *, unsigned long> IDMap; 278 IDMap Map1, Map2; 279 unsigned long IDMap1Count, IDMap2Count; 280}; 281} 282 283/// isEquivalentType - any two pointers in the same address space are 284/// equivalent. Otherwise, standard type equivalence rules apply. 285bool FunctionComparator::isEquivalentType(const Type *Ty1, 286 const Type *Ty2) const { 287 if (Ty1 == Ty2) 288 return true; 289 if (Ty1->getTypeID() != Ty2->getTypeID()) 290 return false; 291 292 switch(Ty1->getTypeID()) { 293 default: 294 llvm_unreachable("Unknown type!"); 295 // Fall through in Release mode. 296 case Type::IntegerTyID: 297 case Type::OpaqueTyID: 298 // Ty1 == Ty2 would have returned true earlier. 299 return false; 300 301 case Type::VoidTyID: 302 case Type::FloatTyID: 303 case Type::DoubleTyID: 304 case Type::X86_FP80TyID: 305 case Type::FP128TyID: 306 case Type::PPC_FP128TyID: 307 case Type::LabelTyID: 308 case Type::MetadataTyID: 309 return true; 310 311 case Type::PointerTyID: { 312 const PointerType *PTy1 = cast<PointerType>(Ty1); 313 const PointerType *PTy2 = cast<PointerType>(Ty2); 314 return PTy1->getAddressSpace() == PTy2->getAddressSpace(); 315 } 316 317 case Type::StructTyID: { 318 const StructType *STy1 = cast<StructType>(Ty1); 319 const StructType *STy2 = cast<StructType>(Ty2); 320 if (STy1->getNumElements() != STy2->getNumElements()) 321 return false; 322 323 if (STy1->isPacked() != STy2->isPacked()) 324 return false; 325 326 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) { 327 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i))) 328 return false; 329 } 330 return true; 331 } 332 333 case Type::FunctionTyID: { 334 const FunctionType *FTy1 = cast<FunctionType>(Ty1); 335 const FunctionType *FTy2 = cast<FunctionType>(Ty2); 336 if (FTy1->getNumParams() != FTy2->getNumParams() || 337 FTy1->isVarArg() != FTy2->isVarArg()) 338 return false; 339 340 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType())) 341 return false; 342 343 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) { 344 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i))) 345 return false; 346 } 347 return true; 348 } 349 350 case Type::ArrayTyID: { 351 const ArrayType *ATy1 = cast<ArrayType>(Ty1); 352 const ArrayType *ATy2 = cast<ArrayType>(Ty2); 353 return ATy1->getNumElements() == ATy2->getNumElements() && 354 isEquivalentType(ATy1->getElementType(), ATy2->getElementType()); 355 } 356 357 case Type::VectorTyID: { 358 const VectorType *VTy1 = cast<VectorType>(Ty1); 359 const VectorType *VTy2 = cast<VectorType>(Ty2); 360 return VTy1->getNumElements() == VTy2->getNumElements() && 361 isEquivalentType(VTy1->getElementType(), VTy2->getElementType()); 362 } 363 } 364} 365 366/// isEquivalentOperation - determine whether the two operations are the same 367/// except that pointer-to-A and pointer-to-B are equivalent. This should be 368/// kept in sync with Instruction::isSameOperationAs. 369bool FunctionComparator::isEquivalentOperation(const Instruction *I1, 370 const Instruction *I2) const { 371 if (I1->getOpcode() != I2->getOpcode() || 372 I1->getNumOperands() != I2->getNumOperands() || 373 !isEquivalentType(I1->getType(), I2->getType()) || 374 !I1->hasSameSubclassOptionalData(I2)) 375 return false; 376 377 // We have two instructions of identical opcode and #operands. Check to see 378 // if all operands are the same type 379 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i) 380 if (!isEquivalentType(I1->getOperand(i)->getType(), 381 I2->getOperand(i)->getType())) 382 return false; 383 384 // Check special state that is a part of some instructions. 385 if (const LoadInst *LI = dyn_cast<LoadInst>(I1)) 386 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() && 387 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment(); 388 if (const StoreInst *SI = dyn_cast<StoreInst>(I1)) 389 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() && 390 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment(); 391 if (const CmpInst *CI = dyn_cast<CmpInst>(I1)) 392 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate(); 393 if (const CallInst *CI = dyn_cast<CallInst>(I1)) 394 return CI->isTailCall() == cast<CallInst>(I2)->isTailCall() && 395 CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() && 396 CI->getAttributes().getRawPointer() == 397 cast<CallInst>(I2)->getAttributes().getRawPointer(); 398 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1)) 399 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() && 400 CI->getAttributes().getRawPointer() == 401 cast<InvokeInst>(I2)->getAttributes().getRawPointer(); 402 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) { 403 if (IVI->getNumIndices() != cast<InsertValueInst>(I2)->getNumIndices()) 404 return false; 405 for (unsigned i = 0, e = IVI->getNumIndices(); i != e; ++i) 406 if (IVI->idx_begin()[i] != cast<InsertValueInst>(I2)->idx_begin()[i]) 407 return false; 408 return true; 409 } 410 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) { 411 if (EVI->getNumIndices() != cast<ExtractValueInst>(I2)->getNumIndices()) 412 return false; 413 for (unsigned i = 0, e = EVI->getNumIndices(); i != e; ++i) 414 if (EVI->idx_begin()[i] != cast<ExtractValueInst>(I2)->idx_begin()[i]) 415 return false; 416 return true; 417 } 418 419 return true; 420} 421 422/// isEquivalentGEP - determine whether two GEP operations perform the same 423/// underlying arithmetic. 424bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1, 425 const GEPOperator *GEP2) { 426 // When we have target data, we can reduce the GEP down to the value in bytes 427 // added to the address. 428 if (TD && GEP1->hasAllConstantIndices() && GEP2->hasAllConstantIndices()) { 429 SmallVector<Value *, 8> Indices1(GEP1->idx_begin(), GEP1->idx_end()); 430 SmallVector<Value *, 8> Indices2(GEP2->idx_begin(), GEP2->idx_end()); 431 uint64_t Offset1 = TD->getIndexedOffset(GEP1->getPointerOperandType(), 432 Indices1.data(), Indices1.size()); 433 uint64_t Offset2 = TD->getIndexedOffset(GEP2->getPointerOperandType(), 434 Indices2.data(), Indices2.size()); 435 return Offset1 == Offset2; 436 } 437 438 if (GEP1->getPointerOperand()->getType() != 439 GEP2->getPointerOperand()->getType()) 440 return false; 441 442 if (GEP1->getNumOperands() != GEP2->getNumOperands()) 443 return false; 444 445 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) { 446 if (!Enumerate(GEP1->getOperand(i), GEP2->getOperand(i))) 447 return false; 448 } 449 450 return true; 451} 452 453/// Enumerate - Compare two values used by the two functions under pair-wise 454/// comparison. If this is the first time the values are seen, they're added to 455/// the mapping so that we will detect mismatches on next use. 456bool FunctionComparator::Enumerate(const Value *V1, const Value *V2) { 457 // Check for function @f1 referring to itself and function @f2 referring to 458 // itself, or referring to each other, or both referring to either of them. 459 // They're all equivalent if the two functions are otherwise equivalent. 460 if (V1 == F1 && V2 == F2) 461 return true; 462 if (V1 == F2 && V2 == F1) 463 return true; 464 465 // TODO: constant expressions with GEP or references to F1 or F2. 466 if (isa<Constant>(V1)) 467 return V1 == V2; 468 469 if (isa<InlineAsm>(V1) && isa<InlineAsm>(V2)) { 470 const InlineAsm *IA1 = cast<InlineAsm>(V1); 471 const InlineAsm *IA2 = cast<InlineAsm>(V2); 472 return IA1->getAsmString() == IA2->getAsmString() && 473 IA1->getConstraintString() == IA2->getConstraintString(); 474 } 475 476 unsigned long &ID1 = Map1[V1]; 477 if (!ID1) 478 ID1 = ++IDMap1Count; 479 480 unsigned long &ID2 = Map2[V2]; 481 if (!ID2) 482 ID2 = ++IDMap2Count; 483 484 return ID1 == ID2; 485} 486 487/// Compare - test whether two basic blocks have equivalent behaviour. 488bool FunctionComparator::Compare(const BasicBlock *BB1, const BasicBlock *BB2) { 489 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end(); 490 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end(); 491 492 do { 493 if (!Enumerate(F1I, F2I)) 494 return false; 495 496 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) { 497 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I); 498 if (!GEP2) 499 return false; 500 501 if (!Enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand())) 502 return false; 503 504 if (!isEquivalentGEP(GEP1, GEP2)) 505 return false; 506 } else { 507 if (!isEquivalentOperation(F1I, F2I)) 508 return false; 509 510 assert(F1I->getNumOperands() == F2I->getNumOperands()); 511 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) { 512 Value *OpF1 = F1I->getOperand(i); 513 Value *OpF2 = F2I->getOperand(i); 514 515 if (!Enumerate(OpF1, OpF2)) 516 return false; 517 518 if (OpF1->getValueID() != OpF2->getValueID() || 519 !isEquivalentType(OpF1->getType(), OpF2->getType())) 520 return false; 521 } 522 } 523 524 ++F1I, ++F2I; 525 } while (F1I != F1E && F2I != F2E); 526 527 return F1I == F1E && F2I == F2E; 528} 529 530/// Compare - test whether the two functions have equivalent behaviour. 531bool FunctionComparator::Compare() { 532 // We need to recheck everything, but check the things that weren't included 533 // in the hash first. 534 535 if (F1->getAttributes() != F2->getAttributes()) 536 return false; 537 538 if (F1->hasGC() != F2->hasGC()) 539 return false; 540 541 if (F1->hasGC() && F1->getGC() != F2->getGC()) 542 return false; 543 544 if (F1->hasSection() != F2->hasSection()) 545 return false; 546 547 if (F1->hasSection() && F1->getSection() != F2->getSection()) 548 return false; 549 550 if (F1->isVarArg() != F2->isVarArg()) 551 return false; 552 553 // TODO: if it's internal and only used in direct calls, we could handle this 554 // case too. 555 if (F1->getCallingConv() != F2->getCallingConv()) 556 return false; 557 558 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType())) 559 return false; 560 561 assert(F1->arg_size() == F2->arg_size() && 562 "Identically typed functions have different numbers of args!"); 563 564 // Visit the arguments so that they get enumerated in the order they're 565 // passed in. 566 for (Function::const_arg_iterator f1i = F1->arg_begin(), 567 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) { 568 if (!Enumerate(f1i, f2i)) 569 llvm_unreachable("Arguments repeat!"); 570 } 571 572 // We do a CFG-ordered walk since the actual ordering of the blocks in the 573 // linked list is immaterial. Our walk starts at the entry block for both 574 // functions, then takes each block from each terminator in order. As an 575 // artifact, this also means that unreachable blocks are ignored. 576 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs; 577 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1. 578 579 F1BBs.push_back(&F1->getEntryBlock()); 580 F2BBs.push_back(&F2->getEntryBlock()); 581 582 VisitedBBs.insert(F1BBs[0]); 583 while (!F1BBs.empty()) { 584 const BasicBlock *F1BB = F1BBs.pop_back_val(); 585 const BasicBlock *F2BB = F2BBs.pop_back_val(); 586 587 if (!Enumerate(F1BB, F2BB) || !Compare(F1BB, F2BB)) 588 return false; 589 590 const TerminatorInst *F1TI = F1BB->getTerminator(); 591 const TerminatorInst *F2TI = F2BB->getTerminator(); 592 593 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors()); 594 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) { 595 if (!VisitedBBs.insert(F1TI->getSuccessor(i))) 596 continue; 597 598 F1BBs.push_back(F1TI->getSuccessor(i)); 599 F2BBs.push_back(F2TI->getSuccessor(i)); 600 } 601 } 602 return true; 603} 604 605/// Replace direct callers of Old with New. 606void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) { 607 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType()); 608 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end(); 609 UI != UE;) { 610 Value::use_iterator TheIter = UI; 611 ++UI; 612 CallSite CS(*TheIter); 613 if (CS && CS.isCallee(TheIter)) { 614 Remove(CS.getInstruction()->getParent()->getParent()); 615 TheIter.getUse().set(BitcastNew); 616 } 617 } 618} 619 620void MergeFunctions::WriteThunkOrAlias(Function *F, Function *G) { 621 if (HasGlobalAliases && G->hasUnnamedAddr()) { 622 if (G->hasExternalLinkage() || G->hasLocalLinkage() || 623 G->hasWeakLinkage()) { 624 WriteAlias(F, G); 625 return; 626 } 627 } 628 629 WriteThunk(F, G); 630} 631 632/// WriteThunk - Replace G with a simple tail call to bitcast(F). Also replace 633/// direct uses of G with bitcast(F). Deletes G. 634void MergeFunctions::WriteThunk(Function *F, Function *G) { 635 if (!G->mayBeOverridden()) { 636 // Redirect direct callers of G to F. 637 replaceDirectCallers(G, F); 638 } 639 640 // If G was internal then we may have replaced all uses of G with F. If so, 641 // stop here and delete G. There's no need for a thunk. 642 if (G->hasLocalLinkage() && G->use_empty()) { 643 G->eraseFromParent(); 644 return; 645 } 646 647 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "", 648 G->getParent()); 649 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG); 650 IRBuilder<false> Builder(BB); 651 652 SmallVector<Value *, 16> Args; 653 unsigned i = 0; 654 const FunctionType *FFTy = F->getFunctionType(); 655 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end(); 656 AI != AE; ++AI) { 657 Args.push_back(Builder.CreateBitCast(AI, FFTy->getParamType(i))); 658 ++i; 659 } 660 661 CallInst *CI = Builder.CreateCall(F, Args.begin(), Args.end()); 662 CI->setTailCall(); 663 CI->setCallingConv(F->getCallingConv()); 664 if (NewG->getReturnType()->isVoidTy()) { 665 Builder.CreateRetVoid(); 666 } else { 667 Builder.CreateRet(Builder.CreateBitCast(CI, NewG->getReturnType())); 668 } 669 670 NewG->copyAttributesFrom(G); 671 NewG->takeName(G); 672 RemoveUsers(G); 673 G->replaceAllUsesWith(NewG); 674 G->eraseFromParent(); 675 676 DEBUG(dbgs() << "WriteThunk: " << NewG->getName() << '\n'); 677 ++NumThunksWritten; 678} 679 680/// WriteAlias - Replace G with an alias to F and delete G. 681void MergeFunctions::WriteAlias(Function *F, Function *G) { 682 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType()); 683 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "", 684 BitcastF, G->getParent()); 685 F->setAlignment(std::max(F->getAlignment(), G->getAlignment())); 686 GA->takeName(G); 687 GA->setVisibility(G->getVisibility()); 688 RemoveUsers(G); 689 G->replaceAllUsesWith(GA); 690 G->eraseFromParent(); 691 692 DEBUG(dbgs() << "WriteAlias: " << GA->getName() << '\n'); 693 ++NumAliasesWritten; 694} 695 696/// MergeTwoFunctions - Merge two equivalent functions. Upon completion, 697/// Function G is deleted. 698void MergeFunctions::MergeTwoFunctions(Function *F, Function *G) { 699 if (F->mayBeOverridden()) { 700 assert(G->mayBeOverridden()); 701 702 if (HasGlobalAliases) { 703 // Make them both thunks to the same internal function. 704 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "", 705 F->getParent()); 706 H->copyAttributesFrom(F); 707 H->takeName(F); 708 RemoveUsers(F); 709 F->replaceAllUsesWith(H); 710 711 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment()); 712 713 WriteAlias(F, G); 714 WriteAlias(F, H); 715 716 F->setAlignment(MaxAlignment); 717 F->setLinkage(GlobalValue::PrivateLinkage); 718 } else { 719 // We can't merge them. Instead, pick one and update all direct callers 720 // to call it and hope that we improve the instruction cache hit rate. 721 replaceDirectCallers(G, F); 722 } 723 724 ++NumDoubleWeak; 725 } else { 726 WriteThunkOrAlias(F, G); 727 } 728 729 ++NumFunctionsMerged; 730} 731 732// Insert - Insert a ComparableFunction into the FnSet, or merge it away if 733// equal to one that's already inserted. 734bool MergeFunctions::Insert(ComparableFunction &NewF) { 735 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF); 736 if (Result.second) 737 return false; 738 739 const ComparableFunction &OldF = *Result.first; 740 741 // Never thunk a strong function to a weak function. 742 assert(!OldF.getFunc()->mayBeOverridden() || 743 NewF.getFunc()->mayBeOverridden()); 744 745 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == " 746 << NewF.getFunc()->getName() << '\n'); 747 748 Function *DeleteF = NewF.getFunc(); 749 NewF.release(); 750 MergeTwoFunctions(OldF.getFunc(), DeleteF); 751 return true; 752} 753 754// Remove - Remove a function from FnSet. If it was already in FnSet, add it to 755// Deferred so that we'll look at it in the next round. 756void MergeFunctions::Remove(Function *F) { 757 ComparableFunction CF = ComparableFunction(F, TD); 758 if (FnSet.erase(CF)) { 759 Deferred.push_back(F); 760 } 761} 762 763// RemoveUsers - For each instruction used by the value, Remove() the function 764// that contains the instruction. This should happen right before a call to RAUW. 765void MergeFunctions::RemoveUsers(Value *V) { 766 std::vector<Value *> Worklist; 767 Worklist.push_back(V); 768 while (!Worklist.empty()) { 769 Value *V = Worklist.back(); 770 Worklist.pop_back(); 771 772 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 773 UI != UE; ++UI) { 774 Use &U = UI.getUse(); 775 if (Instruction *I = dyn_cast<Instruction>(U.getUser())) { 776 Remove(I->getParent()->getParent()); 777 } else if (isa<GlobalValue>(U.getUser())) { 778 // do nothing 779 } else if (Constant *C = dyn_cast<Constant>(U.getUser())) { 780 for (Value::use_iterator CUI = C->use_begin(), CUE = C->use_end(); 781 CUI != CUE; ++CUI) 782 Worklist.push_back(*CUI); 783 } 784 } 785 } 786} 787 788bool MergeFunctions::runOnModule(Module &M) { 789 bool Changed = false; 790 TD = getAnalysisIfAvailable<TargetData>(); 791 792 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { 793 Deferred.push_back(WeakVH(I)); 794 } 795 796 do { 797 std::vector<WeakVH> Worklist; 798 Deferred.swap(Worklist); 799 800 DEBUG(dbgs() << "size of module: " << M.size() << '\n'); 801 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n'); 802 803 // Insert only strong functions and merge them. Strong function merging 804 // always deletes one of them. 805 for (std::vector<WeakVH>::iterator I = Worklist.begin(), 806 E = Worklist.end(); I != E; ++I) { 807 if (!*I) continue; 808 Function *F = cast<Function>(*I); 809 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() && 810 !F->mayBeOverridden()) { 811 ComparableFunction CF = ComparableFunction(F, TD); 812 Changed |= Insert(CF); 813 } 814 } 815 816 // Insert only weak functions and merge them. By doing these second we 817 // create thunks to the strong function when possible. When two weak 818 // functions are identical, we create a new strong function with two weak 819 // weak thunks to it which are identical but not mergable. 820 for (std::vector<WeakVH>::iterator I = Worklist.begin(), 821 E = Worklist.end(); I != E; ++I) { 822 if (!*I) continue; 823 Function *F = cast<Function>(*I); 824 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() && 825 F->mayBeOverridden()) { 826 ComparableFunction CF = ComparableFunction(F, TD); 827 Changed |= Insert(CF); 828 } 829 } 830 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n'); 831 } while (!Deferred.empty()); 832 833 FnSet.clear(); 834 835 return Changed; 836} 837 838bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS, 839 const ComparableFunction &RHS) { 840 if (LHS.getFunc() == RHS.getFunc() && 841 LHS.getHash() == RHS.getHash()) 842 return true; 843 if (!LHS.getFunc() || !RHS.getFunc()) 844 return false; 845 assert(LHS.getTD() == RHS.getTD() && 846 "Comparing functions for different targets"); 847 848 bool inserted; 849 bool &result1 = LHS.getOrInsertCachedComparison(RHS, inserted); 850 if (!inserted) 851 return result1; 852 bool &result2 = RHS.getOrInsertCachedComparison(LHS, inserted); 853 if (!inserted) 854 return result1 = result2; 855 856 return result1 = result2 = FunctionComparator(LHS.getTD(), LHS.getFunc(), 857 RHS.getFunc()).Compare(); 858} 859