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/STLExtras.h" 51 #include "llvm/ADT/SmallSet.h" 52 #include "llvm/ADT/Statistic.h" 53 #include "llvm/IR/Constants.h" 54 #include "llvm/IR/DataLayout.h" 55 #include "llvm/IR/IRBuilder.h" 56 #include "llvm/IR/InlineAsm.h" 57 #include "llvm/IR/Instructions.h" 58 #include "llvm/IR/LLVMContext.h" 59 #include "llvm/IR/Module.h" 60 #include "llvm/IR/Operator.h" 61 #include "llvm/Pass.h" 62 #include "llvm/Support/CallSite.h" 63 #include "llvm/Support/Debug.h" 64 #include "llvm/Support/ErrorHandling.h" 65 #include "llvm/Support/ValueHandle.h" 66 #include "llvm/Support/raw_ostream.h" 67 #include <vector> 68 using namespace llvm; 69 70 STATISTIC(NumFunctionsMerged, "Number of functions merged"); 71 STATISTIC(NumThunksWritten, "Number of thunks generated"); 72 STATISTIC(NumAliasesWritten, "Number of aliases generated"); 73 STATISTIC(NumDoubleWeak, "Number of new functions created"); 74 75 /// Creates a hash-code for the function which is the same for any two 76 /// functions that will compare equal, without looking at the instructions 77 /// inside the function. 78 static unsigned profileFunction(const Function *F) { 79 FunctionType *FTy = F->getFunctionType(); 80 81 FoldingSetNodeID ID; 82 ID.AddInteger(F->size()); 83 ID.AddInteger(F->getCallingConv()); 84 ID.AddBoolean(F->hasGC()); 85 ID.AddBoolean(FTy->isVarArg()); 86 ID.AddInteger(FTy->getReturnType()->getTypeID()); 87 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i) 88 ID.AddInteger(FTy->getParamType(i)->getTypeID()); 89 return ID.ComputeHash(); 90 } 91 92 namespace { 93 94 /// ComparableFunction - A struct that pairs together functions with a 95 /// DataLayout so that we can keep them together as elements in the DenseSet. 96 class ComparableFunction { 97 public: 98 static const ComparableFunction EmptyKey; 99 static const ComparableFunction TombstoneKey; 100 static DataLayout * const LookupOnly; 101 102 ComparableFunction(Function *Func, DataLayout *TD) 103 : Func(Func), Hash(profileFunction(Func)), TD(TD) {} 104 105 Function *getFunc() const { return Func; } 106 unsigned getHash() const { return Hash; } 107 DataLayout *getTD() const { return TD; } 108 109 // Drops AssertingVH reference to the function. Outside of debug mode, this 110 // does nothing. 111 void release() { 112 assert(Func && 113 "Attempted to release function twice, or release empty/tombstone!"); 114 Func = NULL; 115 } 116 117 private: 118 explicit ComparableFunction(unsigned Hash) 119 : Func(NULL), Hash(Hash), TD(NULL) {} 120 121 AssertingVH<Function> Func; 122 unsigned Hash; 123 DataLayout *TD; 124 }; 125 126 const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0); 127 const ComparableFunction ComparableFunction::TombstoneKey = 128 ComparableFunction(1); 129 DataLayout *const ComparableFunction::LookupOnly = (DataLayout*)(-1); 130 131 } 132 133 namespace llvm { 134 template <> 135 struct DenseMapInfo<ComparableFunction> { 136 static ComparableFunction getEmptyKey() { 137 return ComparableFunction::EmptyKey; 138 } 139 static ComparableFunction getTombstoneKey() { 140 return ComparableFunction::TombstoneKey; 141 } 142 static unsigned getHashValue(const ComparableFunction &CF) { 143 return CF.getHash(); 144 } 145 static bool isEqual(const ComparableFunction &LHS, 146 const ComparableFunction &RHS); 147 }; 148 } 149 150 namespace { 151 152 /// FunctionComparator - Compares two functions to determine whether or not 153 /// they will generate machine code with the same behaviour. DataLayout is 154 /// used if available. The comparator always fails conservatively (erring on the 155 /// side of claiming that two functions are different). 156 class FunctionComparator { 157 public: 158 FunctionComparator(const DataLayout *TD, const Function *F1, 159 const Function *F2) 160 : F1(F1), F2(F2), TD(TD) {} 161 162 /// Test whether the two functions have equivalent behaviour. 163 bool compare(); 164 165 private: 166 /// Test whether two basic blocks have equivalent behaviour. 167 bool compare(const BasicBlock *BB1, const BasicBlock *BB2); 168 169 /// Assign or look up previously assigned numbers for the two values, and 170 /// return whether the numbers are equal. Numbers are assigned in the order 171 /// visited. 172 bool enumerate(const Value *V1, const Value *V2); 173 174 /// Compare two Instructions for equivalence, similar to 175 /// Instruction::isSameOperationAs but with modifications to the type 176 /// comparison. 177 bool isEquivalentOperation(const Instruction *I1, 178 const Instruction *I2) const; 179 180 /// Compare two GEPs for equivalent pointer arithmetic. 181 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2); 182 bool isEquivalentGEP(const GetElementPtrInst *GEP1, 183 const GetElementPtrInst *GEP2) { 184 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2)); 185 } 186 187 /// Compare two Types, treating all pointer types as equal. 188 bool isEquivalentType(Type *Ty1, Type *Ty2) const; 189 190 // The two functions undergoing comparison. 191 const Function *F1, *F2; 192 193 const DataLayout *TD; 194 195 DenseMap<const Value *, const Value *> id_map; 196 DenseSet<const Value *> seen_values; 197 }; 198 199 } 200 201 // Any two pointers in the same address space are equivalent, intptr_t and 202 // pointers are equivalent. Otherwise, standard type equivalence rules apply. 203 bool FunctionComparator::isEquivalentType(Type *Ty1, 204 Type *Ty2) const { 205 if (Ty1 == Ty2) 206 return true; 207 if (Ty1->getTypeID() != Ty2->getTypeID()) { 208 if (TD) { 209 LLVMContext &Ctx = Ty1->getContext(); 210 if (isa<PointerType>(Ty1) && Ty2 == TD->getIntPtrType(Ctx)) return true; 211 if (isa<PointerType>(Ty2) && Ty1 == TD->getIntPtrType(Ctx)) return true; 212 } 213 return false; 214 } 215 216 switch (Ty1->getTypeID()) { 217 default: 218 llvm_unreachable("Unknown type!"); 219 // Fall through in Release mode. 220 case Type::IntegerTyID: 221 case Type::VectorTyID: 222 // Ty1 == Ty2 would have returned true earlier. 223 return false; 224 225 case Type::VoidTyID: 226 case Type::FloatTyID: 227 case Type::DoubleTyID: 228 case Type::X86_FP80TyID: 229 case Type::FP128TyID: 230 case Type::PPC_FP128TyID: 231 case Type::LabelTyID: 232 case Type::MetadataTyID: 233 return true; 234 235 case Type::PointerTyID: { 236 PointerType *PTy1 = cast<PointerType>(Ty1); 237 PointerType *PTy2 = cast<PointerType>(Ty2); 238 return PTy1->getAddressSpace() == PTy2->getAddressSpace(); 239 } 240 241 case Type::StructTyID: { 242 StructType *STy1 = cast<StructType>(Ty1); 243 StructType *STy2 = cast<StructType>(Ty2); 244 if (STy1->getNumElements() != STy2->getNumElements()) 245 return false; 246 247 if (STy1->isPacked() != STy2->isPacked()) 248 return false; 249 250 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) { 251 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i))) 252 return false; 253 } 254 return true; 255 } 256 257 case Type::FunctionTyID: { 258 FunctionType *FTy1 = cast<FunctionType>(Ty1); 259 FunctionType *FTy2 = cast<FunctionType>(Ty2); 260 if (FTy1->getNumParams() != FTy2->getNumParams() || 261 FTy1->isVarArg() != FTy2->isVarArg()) 262 return false; 263 264 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType())) 265 return false; 266 267 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) { 268 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i))) 269 return false; 270 } 271 return true; 272 } 273 274 case Type::ArrayTyID: { 275 ArrayType *ATy1 = cast<ArrayType>(Ty1); 276 ArrayType *ATy2 = cast<ArrayType>(Ty2); 277 return ATy1->getNumElements() == ATy2->getNumElements() && 278 isEquivalentType(ATy1->getElementType(), ATy2->getElementType()); 279 } 280 } 281 } 282 283 // Determine whether the two operations are the same except that pointer-to-A 284 // and pointer-to-B are equivalent. This should be kept in sync with 285 // Instruction::isSameOperationAs. 286 bool FunctionComparator::isEquivalentOperation(const Instruction *I1, 287 const Instruction *I2) const { 288 // Differences from Instruction::isSameOperationAs: 289 // * replace type comparison with calls to isEquivalentType. 290 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top 291 // * because of the above, we don't test for the tail bit on calls later on 292 if (I1->getOpcode() != I2->getOpcode() || 293 I1->getNumOperands() != I2->getNumOperands() || 294 !isEquivalentType(I1->getType(), I2->getType()) || 295 !I1->hasSameSubclassOptionalData(I2)) 296 return false; 297 298 // We have two instructions of identical opcode and #operands. Check to see 299 // if all operands are the same type 300 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i) 301 if (!isEquivalentType(I1->getOperand(i)->getType(), 302 I2->getOperand(i)->getType())) 303 return false; 304 305 // Check special state that is a part of some instructions. 306 if (const LoadInst *LI = dyn_cast<LoadInst>(I1)) 307 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() && 308 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment() && 309 LI->getOrdering() == cast<LoadInst>(I2)->getOrdering() && 310 LI->getSynchScope() == cast<LoadInst>(I2)->getSynchScope(); 311 if (const StoreInst *SI = dyn_cast<StoreInst>(I1)) 312 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() && 313 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment() && 314 SI->getOrdering() == cast<StoreInst>(I2)->getOrdering() && 315 SI->getSynchScope() == cast<StoreInst>(I2)->getSynchScope(); 316 if (const CmpInst *CI = dyn_cast<CmpInst>(I1)) 317 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate(); 318 if (const CallInst *CI = dyn_cast<CallInst>(I1)) 319 return CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() && 320 CI->getAttributes() == cast<CallInst>(I2)->getAttributes(); 321 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1)) 322 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() && 323 CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes(); 324 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) 325 return IVI->getIndices() == cast<InsertValueInst>(I2)->getIndices(); 326 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) 327 return EVI->getIndices() == cast<ExtractValueInst>(I2)->getIndices(); 328 if (const FenceInst *FI = dyn_cast<FenceInst>(I1)) 329 return FI->getOrdering() == cast<FenceInst>(I2)->getOrdering() && 330 FI->getSynchScope() == cast<FenceInst>(I2)->getSynchScope(); 331 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I1)) 332 return CXI->isVolatile() == cast<AtomicCmpXchgInst>(I2)->isVolatile() && 333 CXI->getOrdering() == cast<AtomicCmpXchgInst>(I2)->getOrdering() && 334 CXI->getSynchScope() == cast<AtomicCmpXchgInst>(I2)->getSynchScope(); 335 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I1)) 336 return RMWI->getOperation() == cast<AtomicRMWInst>(I2)->getOperation() && 337 RMWI->isVolatile() == cast<AtomicRMWInst>(I2)->isVolatile() && 338 RMWI->getOrdering() == cast<AtomicRMWInst>(I2)->getOrdering() && 339 RMWI->getSynchScope() == cast<AtomicRMWInst>(I2)->getSynchScope(); 340 341 return true; 342 } 343 344 // Determine whether two GEP operations perform the same underlying arithmetic. 345 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1, 346 const GEPOperator *GEP2) { 347 // When we have target data, we can reduce the GEP down to the value in bytes 348 // added to the address. 349 unsigned BitWidth = TD ? TD->getPointerSizeInBits() : 1; 350 APInt Offset1(BitWidth, 0), Offset2(BitWidth, 0); 351 if (TD && 352 GEP1->accumulateConstantOffset(*TD, Offset1) && 353 GEP2->accumulateConstantOffset(*TD, Offset2)) { 354 return Offset1 == Offset2; 355 } 356 357 if (GEP1->getPointerOperand()->getType() != 358 GEP2->getPointerOperand()->getType()) 359 return false; 360 361 if (GEP1->getNumOperands() != GEP2->getNumOperands()) 362 return false; 363 364 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) { 365 if (!enumerate(GEP1->getOperand(i), GEP2->getOperand(i))) 366 return false; 367 } 368 369 return true; 370 } 371 372 // Compare two values used by the two functions under pair-wise comparison. If 373 // this is the first time the values are seen, they're added to the mapping so 374 // that we will detect mismatches on next use. 375 bool FunctionComparator::enumerate(const Value *V1, const Value *V2) { 376 // Check for function @f1 referring to itself and function @f2 referring to 377 // itself, or referring to each other, or both referring to either of them. 378 // They're all equivalent if the two functions are otherwise equivalent. 379 if (V1 == F1 && V2 == F2) 380 return true; 381 if (V1 == F2 && V2 == F1) 382 return true; 383 384 if (const Constant *C1 = dyn_cast<Constant>(V1)) { 385 if (V1 == V2) return true; 386 const Constant *C2 = dyn_cast<Constant>(V2); 387 if (!C2) return false; 388 // TODO: constant expressions with GEP or references to F1 or F2. 389 if (C1->isNullValue() && C2->isNullValue() && 390 isEquivalentType(C1->getType(), C2->getType())) 391 return true; 392 // Try bitcasting C2 to C1's type. If the bitcast is legal and returns C1 393 // then they must have equal bit patterns. 394 return C1->getType()->canLosslesslyBitCastTo(C2->getType()) && 395 C1 == ConstantExpr::getBitCast(const_cast<Constant*>(C2), C1->getType()); 396 } 397 398 if (isa<InlineAsm>(V1) || isa<InlineAsm>(V2)) 399 return V1 == V2; 400 401 // Check that V1 maps to V2. If we find a value that V1 maps to then we simply 402 // check whether it's equal to V2. When there is no mapping then we need to 403 // ensure that V2 isn't already equivalent to something else. For this 404 // purpose, we track the V2 values in a set. 405 406 const Value *&map_elem = id_map[V1]; 407 if (map_elem) 408 return map_elem == V2; 409 if (!seen_values.insert(V2).second) 410 return false; 411 map_elem = V2; 412 return true; 413 } 414 415 // Test whether two basic blocks have equivalent behaviour. 416 bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) { 417 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end(); 418 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end(); 419 420 do { 421 if (!enumerate(F1I, F2I)) 422 return false; 423 424 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) { 425 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I); 426 if (!GEP2) 427 return false; 428 429 if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand())) 430 return false; 431 432 if (!isEquivalentGEP(GEP1, GEP2)) 433 return false; 434 } else { 435 if (!isEquivalentOperation(F1I, F2I)) 436 return false; 437 438 assert(F1I->getNumOperands() == F2I->getNumOperands()); 439 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) { 440 Value *OpF1 = F1I->getOperand(i); 441 Value *OpF2 = F2I->getOperand(i); 442 443 if (!enumerate(OpF1, OpF2)) 444 return false; 445 446 if (OpF1->getValueID() != OpF2->getValueID() || 447 !isEquivalentType(OpF1->getType(), OpF2->getType())) 448 return false; 449 } 450 } 451 452 ++F1I, ++F2I; 453 } while (F1I != F1E && F2I != F2E); 454 455 return F1I == F1E && F2I == F2E; 456 } 457 458 // Test whether the two functions have equivalent behaviour. 459 bool 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 "Identically typed functions have different numbers 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 do a CFG-ordered walk since the actual ordering of the blocks in the 501 // 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. 506 507 F1BBs.push_back(&F1->getEntryBlock()); 508 F2BBs.push_back(&F2->getEntryBlock()); 509 510 VisitedBBs.insert(F1BBs[0]); 511 while (!F1BBs.empty()) { 512 const BasicBlock *F1BB = F1BBs.pop_back_val(); 513 const BasicBlock *F2BB = F2BBs.pop_back_val(); 514 515 if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB)) 516 return false; 517 518 const TerminatorInst *F1TI = F1BB->getTerminator(); 519 const TerminatorInst *F2TI = F2BB->getTerminator(); 520 521 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors()); 522 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) { 523 if (!VisitedBBs.insert(F1TI->getSuccessor(i))) 524 continue; 525 526 F1BBs.push_back(F1TI->getSuccessor(i)); 527 F2BBs.push_back(F2TI->getSuccessor(i)); 528 } 529 } 530 return true; 531 } 532 533 namespace { 534 535 /// MergeFunctions finds functions which will generate identical machine code, 536 /// by considering all pointer types to be equivalent. Once identified, 537 /// MergeFunctions will fold them by replacing a call to one to a call to a 538 /// bitcast of the other. 539 /// 540 class MergeFunctions : public ModulePass { 541 public: 542 static char ID; 543 MergeFunctions() 544 : ModulePass(ID), HasGlobalAliases(false) { 545 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry()); 546 } 547 548 bool runOnModule(Module &M); 549 550 private: 551 typedef DenseSet<ComparableFunction> FnSetType; 552 553 /// A work queue of functions that may have been modified and should be 554 /// analyzed again. 555 std::vector<WeakVH> Deferred; 556 557 /// Insert a ComparableFunction into the FnSet, or merge it away if it's 558 /// equal to one that's already present. 559 bool insert(ComparableFunction &NewF); 560 561 /// Remove a Function from the FnSet and queue it up for a second sweep of 562 /// analysis. 563 void remove(Function *F); 564 565 /// Find the functions that use this Value and remove them from FnSet and 566 /// queue the functions. 567 void removeUsers(Value *V); 568 569 /// Replace all direct calls of Old with calls of New. Will bitcast New if 570 /// necessary to make types match. 571 void replaceDirectCallers(Function *Old, Function *New); 572 573 /// Merge two equivalent functions. Upon completion, G may be deleted, or may 574 /// be converted into a thunk. In either case, it should never be visited 575 /// again. 576 void mergeTwoFunctions(Function *F, Function *G); 577 578 /// Replace G with a thunk or an alias to F. Deletes G. 579 void writeThunkOrAlias(Function *F, Function *G); 580 581 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses 582 /// of G with bitcast(F). Deletes G. 583 void writeThunk(Function *F, Function *G); 584 585 /// Replace G with an alias to F. Deletes G. 586 void writeAlias(Function *F, Function *G); 587 588 /// The set of all distinct functions. Use the insert() and remove() methods 589 /// to modify it. 590 FnSetType FnSet; 591 592 /// DataLayout for more accurate GEP comparisons. May be NULL. 593 DataLayout *TD; 594 595 /// Whether or not the target supports global aliases. 596 bool HasGlobalAliases; 597 }; 598 599 } // end anonymous namespace 600 601 char MergeFunctions::ID = 0; 602 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false) 603 604 ModulePass *llvm::createMergeFunctionsPass() { 605 return new MergeFunctions(); 606 } 607 608 bool MergeFunctions::runOnModule(Module &M) { 609 bool Changed = false; 610 TD = getAnalysisIfAvailable<DataLayout>(); 611 612 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { 613 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage()) 614 Deferred.push_back(WeakVH(I)); 615 } 616 FnSet.resize(Deferred.size()); 617 618 do { 619 std::vector<WeakVH> Worklist; 620 Deferred.swap(Worklist); 621 622 DEBUG(dbgs() << "size of module: " << M.size() << '\n'); 623 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n'); 624 625 // Insert only strong functions and merge them. Strong function merging 626 // always deletes one of them. 627 for (std::vector<WeakVH>::iterator I = Worklist.begin(), 628 E = Worklist.end(); I != E; ++I) { 629 if (!*I) continue; 630 Function *F = cast<Function>(*I); 631 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() && 632 !F->mayBeOverridden()) { 633 ComparableFunction CF = ComparableFunction(F, TD); 634 Changed |= insert(CF); 635 } 636 } 637 638 // Insert only weak functions and merge them. By doing these second we 639 // create thunks to the strong function when possible. When two weak 640 // functions are identical, we create a new strong function with two weak 641 // weak thunks to it which are identical but not mergable. 642 for (std::vector<WeakVH>::iterator I = Worklist.begin(), 643 E = Worklist.end(); I != E; ++I) { 644 if (!*I) continue; 645 Function *F = cast<Function>(*I); 646 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() && 647 F->mayBeOverridden()) { 648 ComparableFunction CF = ComparableFunction(F, TD); 649 Changed |= insert(CF); 650 } 651 } 652 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n'); 653 } while (!Deferred.empty()); 654 655 FnSet.clear(); 656 657 return Changed; 658 } 659 660 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS, 661 const ComparableFunction &RHS) { 662 if (LHS.getFunc() == RHS.getFunc() && 663 LHS.getHash() == RHS.getHash()) 664 return true; 665 if (!LHS.getFunc() || !RHS.getFunc()) 666 return false; 667 668 // One of these is a special "underlying pointer comparison only" object. 669 if (LHS.getTD() == ComparableFunction::LookupOnly || 670 RHS.getTD() == ComparableFunction::LookupOnly) 671 return false; 672 673 assert(LHS.getTD() == RHS.getTD() && 674 "Comparing functions for different targets"); 675 676 return FunctionComparator(LHS.getTD(), LHS.getFunc(), 677 RHS.getFunc()).compare(); 678 } 679 680 // Replace direct callers of Old with New. 681 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) { 682 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType()); 683 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end(); 684 UI != UE;) { 685 Value::use_iterator TheIter = UI; 686 ++UI; 687 CallSite CS(*TheIter); 688 if (CS && CS.isCallee(TheIter)) { 689 remove(CS.getInstruction()->getParent()->getParent()); 690 TheIter.getUse().set(BitcastNew); 691 } 692 } 693 } 694 695 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G. 696 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) { 697 if (HasGlobalAliases && G->hasUnnamedAddr()) { 698 if (G->hasExternalLinkage() || G->hasLocalLinkage() || 699 G->hasWeakLinkage()) { 700 writeAlias(F, G); 701 return; 702 } 703 } 704 705 writeThunk(F, G); 706 } 707 708 // Replace G with a simple tail call to bitcast(F). Also replace direct uses 709 // of G with bitcast(F). Deletes G. 710 void MergeFunctions::writeThunk(Function *F, Function *G) { 711 if (!G->mayBeOverridden()) { 712 // Redirect direct callers of G to F. 713 replaceDirectCallers(G, F); 714 } 715 716 // If G was internal then we may have replaced all uses of G with F. If so, 717 // stop here and delete G. There's no need for a thunk. 718 if (G->hasLocalLinkage() && G->use_empty()) { 719 G->eraseFromParent(); 720 return; 721 } 722 723 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "", 724 G->getParent()); 725 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG); 726 IRBuilder<false> Builder(BB); 727 728 SmallVector<Value *, 16> Args; 729 unsigned i = 0; 730 FunctionType *FFTy = F->getFunctionType(); 731 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end(); 732 AI != AE; ++AI) { 733 Args.push_back(Builder.CreateBitCast(AI, FFTy->getParamType(i))); 734 ++i; 735 } 736 737 CallInst *CI = Builder.CreateCall(F, Args); 738 CI->setTailCall(); 739 CI->setCallingConv(F->getCallingConv()); 740 if (NewG->getReturnType()->isVoidTy()) { 741 Builder.CreateRetVoid(); 742 } else { 743 Builder.CreateRet(Builder.CreateBitCast(CI, NewG->getReturnType())); 744 } 745 746 NewG->copyAttributesFrom(G); 747 NewG->takeName(G); 748 removeUsers(G); 749 G->replaceAllUsesWith(NewG); 750 G->eraseFromParent(); 751 752 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n'); 753 ++NumThunksWritten; 754 } 755 756 // Replace G with an alias to F and delete G. 757 void MergeFunctions::writeAlias(Function *F, Function *G) { 758 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType()); 759 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "", 760 BitcastF, G->getParent()); 761 F->setAlignment(std::max(F->getAlignment(), G->getAlignment())); 762 GA->takeName(G); 763 GA->setVisibility(G->getVisibility()); 764 removeUsers(G); 765 G->replaceAllUsesWith(GA); 766 G->eraseFromParent(); 767 768 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n'); 769 ++NumAliasesWritten; 770 } 771 772 // Merge two equivalent functions. Upon completion, Function G is deleted. 773 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) { 774 if (F->mayBeOverridden()) { 775 assert(G->mayBeOverridden()); 776 777 if (HasGlobalAliases) { 778 // Make them both thunks to the same internal function. 779 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "", 780 F->getParent()); 781 H->copyAttributesFrom(F); 782 H->takeName(F); 783 removeUsers(F); 784 F->replaceAllUsesWith(H); 785 786 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment()); 787 788 writeAlias(F, G); 789 writeAlias(F, H); 790 791 F->setAlignment(MaxAlignment); 792 F->setLinkage(GlobalValue::PrivateLinkage); 793 } else { 794 // We can't merge them. Instead, pick one and update all direct callers 795 // to call it and hope that we improve the instruction cache hit rate. 796 replaceDirectCallers(G, F); 797 } 798 799 ++NumDoubleWeak; 800 } else { 801 writeThunkOrAlias(F, G); 802 } 803 804 ++NumFunctionsMerged; 805 } 806 807 // Insert a ComparableFunction into the FnSet, or merge it away if equal to one 808 // that was already inserted. 809 bool MergeFunctions::insert(ComparableFunction &NewF) { 810 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF); 811 if (Result.second) { 812 DEBUG(dbgs() << "Inserting as unique: " << NewF.getFunc()->getName() << '\n'); 813 return false; 814 } 815 816 const ComparableFunction &OldF = *Result.first; 817 818 // Never thunk a strong function to a weak function. 819 assert(!OldF.getFunc()->mayBeOverridden() || 820 NewF.getFunc()->mayBeOverridden()); 821 822 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == " 823 << NewF.getFunc()->getName() << '\n'); 824 825 Function *DeleteF = NewF.getFunc(); 826 NewF.release(); 827 mergeTwoFunctions(OldF.getFunc(), DeleteF); 828 return true; 829 } 830 831 // Remove a function from FnSet. If it was already in FnSet, add it to Deferred 832 // so that we'll look at it in the next round. 833 void MergeFunctions::remove(Function *F) { 834 // We need to make sure we remove F, not a function "equal" to F per the 835 // function equality comparator. 836 // 837 // The special "lookup only" ComparableFunction bypasses the expensive 838 // function comparison in favour of a pointer comparison on the underlying 839 // Function*'s. 840 ComparableFunction CF = ComparableFunction(F, ComparableFunction::LookupOnly); 841 if (FnSet.erase(CF)) { 842 DEBUG(dbgs() << "Removed " << F->getName() << " from set and deferred it.\n"); 843 Deferred.push_back(F); 844 } 845 } 846 847 // For each instruction used by the value, remove() the function that contains 848 // the instruction. This should happen right before a call to RAUW. 849 void MergeFunctions::removeUsers(Value *V) { 850 std::vector<Value *> Worklist; 851 Worklist.push_back(V); 852 while (!Worklist.empty()) { 853 Value *V = Worklist.back(); 854 Worklist.pop_back(); 855 856 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 857 UI != UE; ++UI) { 858 Use &U = UI.getUse(); 859 if (Instruction *I = dyn_cast<Instruction>(U.getUser())) { 860 remove(I->getParent()->getParent()); 861 } else if (isa<GlobalValue>(U.getUser())) { 862 // do nothing 863 } else if (Constant *C = dyn_cast<Constant>(U.getUser())) { 864 for (Value::use_iterator CUI = C->use_begin(), CUE = C->use_end(); 865 CUI != CUE; ++CUI) 866 Worklist.push_back(*CUI); 867 } 868 } 869 } 870 } 871