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 /// Returns the type id for a type to be hashed. We turn pointer types into 76 /// integers here because the actual compare logic below considers pointers and 77 /// integers of the same size as equal. 78 static Type::TypeID getTypeIDForHash(Type *Ty) { 79 if (Ty->isPointerTy()) 80 return Type::IntegerTyID; 81 return Ty->getTypeID(); 82 } 83 84 /// Creates a hash-code for the function which is the same for any two 85 /// functions that will compare equal, without looking at the instructions 86 /// inside the function. 87 static unsigned profileFunction(const Function *F) { 88 FunctionType *FTy = F->getFunctionType(); 89 90 FoldingSetNodeID ID; 91 ID.AddInteger(F->size()); 92 ID.AddInteger(F->getCallingConv()); 93 ID.AddBoolean(F->hasGC()); 94 ID.AddBoolean(FTy->isVarArg()); 95 ID.AddInteger(getTypeIDForHash(FTy->getReturnType())); 96 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i) 97 ID.AddInteger(getTypeIDForHash(FTy->getParamType(i))); 98 return ID.ComputeHash(); 99 } 100 101 namespace { 102 103 /// ComparableFunction - A struct that pairs together functions with a 104 /// DataLayout so that we can keep them together as elements in the DenseSet. 105 class ComparableFunction { 106 public: 107 static const ComparableFunction EmptyKey; 108 static const ComparableFunction TombstoneKey; 109 static DataLayout * const LookupOnly; 110 111 ComparableFunction(Function *Func, DataLayout *TD) 112 : Func(Func), Hash(profileFunction(Func)), TD(TD) {} 113 114 Function *getFunc() const { return Func; } 115 unsigned getHash() const { return Hash; } 116 DataLayout *getTD() const { return TD; } 117 118 // Drops AssertingVH reference to the function. Outside of debug mode, this 119 // does nothing. 120 void release() { 121 assert(Func && 122 "Attempted to release function twice, or release empty/tombstone!"); 123 Func = NULL; 124 } 125 126 private: 127 explicit ComparableFunction(unsigned Hash) 128 : Func(NULL), Hash(Hash), TD(NULL) {} 129 130 AssertingVH<Function> Func; 131 unsigned Hash; 132 DataLayout *TD; 133 }; 134 135 const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0); 136 const ComparableFunction ComparableFunction::TombstoneKey = 137 ComparableFunction(1); 138 DataLayout *const ComparableFunction::LookupOnly = (DataLayout*)(-1); 139 140 } 141 142 namespace llvm { 143 template <> 144 struct DenseMapInfo<ComparableFunction> { 145 static ComparableFunction getEmptyKey() { 146 return ComparableFunction::EmptyKey; 147 } 148 static ComparableFunction getTombstoneKey() { 149 return ComparableFunction::TombstoneKey; 150 } 151 static unsigned getHashValue(const ComparableFunction &CF) { 152 return CF.getHash(); 153 } 154 static bool isEqual(const ComparableFunction &LHS, 155 const ComparableFunction &RHS); 156 }; 157 } 158 159 namespace { 160 161 /// FunctionComparator - Compares two functions to determine whether or not 162 /// they will generate machine code with the same behaviour. DataLayout is 163 /// used if available. The comparator always fails conservatively (erring on the 164 /// side of claiming that two functions are different). 165 class FunctionComparator { 166 public: 167 FunctionComparator(const DataLayout *TD, const Function *F1, 168 const Function *F2) 169 : F1(F1), F2(F2), TD(TD) {} 170 171 /// Test whether the two functions have equivalent behaviour. 172 bool compare(); 173 174 private: 175 /// Test whether two basic blocks have equivalent behaviour. 176 bool compare(const BasicBlock *BB1, const BasicBlock *BB2); 177 178 /// Assign or look up previously assigned numbers for the two values, and 179 /// return whether the numbers are equal. Numbers are assigned in the order 180 /// visited. 181 bool enumerate(const Value *V1, const Value *V2); 182 183 /// Compare two Instructions for equivalence, similar to 184 /// Instruction::isSameOperationAs but with modifications to the type 185 /// comparison. 186 bool isEquivalentOperation(const Instruction *I1, 187 const Instruction *I2) const; 188 189 /// Compare two GEPs for equivalent pointer arithmetic. 190 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2); 191 bool isEquivalentGEP(const GetElementPtrInst *GEP1, 192 const GetElementPtrInst *GEP2) { 193 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2)); 194 } 195 196 /// Compare two Types, treating all pointer types as equal. 197 bool isEquivalentType(Type *Ty1, Type *Ty2) const; 198 199 // The two functions undergoing comparison. 200 const Function *F1, *F2; 201 202 const DataLayout *TD; 203 204 DenseMap<const Value *, const Value *> id_map; 205 DenseSet<const Value *> seen_values; 206 }; 207 208 } 209 210 // Any two pointers in the same address space are equivalent, intptr_t and 211 // pointers are equivalent. Otherwise, standard type equivalence rules apply. 212 bool FunctionComparator::isEquivalentType(Type *Ty1, Type *Ty2) const { 213 if (Ty1 == Ty2) 214 return true; 215 if (Ty1->getTypeID() != Ty2->getTypeID()) { 216 if (TD) { 217 218 if (isa<PointerType>(Ty1) && Ty2 == TD->getIntPtrType(Ty1)) 219 return true; 220 221 if (isa<PointerType>(Ty2) && Ty1 == TD->getIntPtrType(Ty2)) 222 return true; 223 } 224 return false; 225 } 226 227 switch (Ty1->getTypeID()) { 228 default: 229 llvm_unreachable("Unknown type!"); 230 // Fall through in Release mode. 231 case Type::IntegerTyID: 232 case Type::VectorTyID: 233 // Ty1 == Ty2 would have returned true earlier. 234 return false; 235 236 case Type::VoidTyID: 237 case Type::FloatTyID: 238 case Type::DoubleTyID: 239 case Type::X86_FP80TyID: 240 case Type::FP128TyID: 241 case Type::PPC_FP128TyID: 242 case Type::LabelTyID: 243 case Type::MetadataTyID: 244 return true; 245 246 case Type::PointerTyID: { 247 PointerType *PTy1 = cast<PointerType>(Ty1); 248 PointerType *PTy2 = cast<PointerType>(Ty2); 249 return PTy1->getAddressSpace() == PTy2->getAddressSpace(); 250 } 251 252 case Type::StructTyID: { 253 StructType *STy1 = cast<StructType>(Ty1); 254 StructType *STy2 = cast<StructType>(Ty2); 255 if (STy1->getNumElements() != STy2->getNumElements()) 256 return false; 257 258 if (STy1->isPacked() != STy2->isPacked()) 259 return false; 260 261 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) { 262 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i))) 263 return false; 264 } 265 return true; 266 } 267 268 case Type::FunctionTyID: { 269 FunctionType *FTy1 = cast<FunctionType>(Ty1); 270 FunctionType *FTy2 = cast<FunctionType>(Ty2); 271 if (FTy1->getNumParams() != FTy2->getNumParams() || 272 FTy1->isVarArg() != FTy2->isVarArg()) 273 return false; 274 275 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType())) 276 return false; 277 278 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) { 279 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i))) 280 return false; 281 } 282 return true; 283 } 284 285 case Type::ArrayTyID: { 286 ArrayType *ATy1 = cast<ArrayType>(Ty1); 287 ArrayType *ATy2 = cast<ArrayType>(Ty2); 288 return ATy1->getNumElements() == ATy2->getNumElements() && 289 isEquivalentType(ATy1->getElementType(), ATy2->getElementType()); 290 } 291 } 292 } 293 294 // Determine whether the two operations are the same except that pointer-to-A 295 // and pointer-to-B are equivalent. This should be kept in sync with 296 // Instruction::isSameOperationAs. 297 bool FunctionComparator::isEquivalentOperation(const Instruction *I1, 298 const Instruction *I2) const { 299 // Differences from Instruction::isSameOperationAs: 300 // * replace type comparison with calls to isEquivalentType. 301 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top 302 // * because of the above, we don't test for the tail bit on calls later on 303 if (I1->getOpcode() != I2->getOpcode() || 304 I1->getNumOperands() != I2->getNumOperands() || 305 !isEquivalentType(I1->getType(), I2->getType()) || 306 !I1->hasSameSubclassOptionalData(I2)) 307 return false; 308 309 // We have two instructions of identical opcode and #operands. Check to see 310 // if all operands are the same type 311 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i) 312 if (!isEquivalentType(I1->getOperand(i)->getType(), 313 I2->getOperand(i)->getType())) 314 return false; 315 316 // Check special state that is a part of some instructions. 317 if (const LoadInst *LI = dyn_cast<LoadInst>(I1)) 318 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() && 319 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment() && 320 LI->getOrdering() == cast<LoadInst>(I2)->getOrdering() && 321 LI->getSynchScope() == cast<LoadInst>(I2)->getSynchScope(); 322 if (const StoreInst *SI = dyn_cast<StoreInst>(I1)) 323 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() && 324 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment() && 325 SI->getOrdering() == cast<StoreInst>(I2)->getOrdering() && 326 SI->getSynchScope() == cast<StoreInst>(I2)->getSynchScope(); 327 if (const CmpInst *CI = dyn_cast<CmpInst>(I1)) 328 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate(); 329 if (const CallInst *CI = dyn_cast<CallInst>(I1)) 330 return CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() && 331 CI->getAttributes() == cast<CallInst>(I2)->getAttributes(); 332 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1)) 333 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() && 334 CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes(); 335 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) 336 return IVI->getIndices() == cast<InsertValueInst>(I2)->getIndices(); 337 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) 338 return EVI->getIndices() == cast<ExtractValueInst>(I2)->getIndices(); 339 if (const FenceInst *FI = dyn_cast<FenceInst>(I1)) 340 return FI->getOrdering() == cast<FenceInst>(I2)->getOrdering() && 341 FI->getSynchScope() == cast<FenceInst>(I2)->getSynchScope(); 342 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I1)) 343 return CXI->isVolatile() == cast<AtomicCmpXchgInst>(I2)->isVolatile() && 344 CXI->getOrdering() == cast<AtomicCmpXchgInst>(I2)->getOrdering() && 345 CXI->getSynchScope() == cast<AtomicCmpXchgInst>(I2)->getSynchScope(); 346 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I1)) 347 return RMWI->getOperation() == cast<AtomicRMWInst>(I2)->getOperation() && 348 RMWI->isVolatile() == cast<AtomicRMWInst>(I2)->isVolatile() && 349 RMWI->getOrdering() == cast<AtomicRMWInst>(I2)->getOrdering() && 350 RMWI->getSynchScope() == cast<AtomicRMWInst>(I2)->getSynchScope(); 351 352 return true; 353 } 354 355 // Determine whether two GEP operations perform the same underlying arithmetic. 356 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1, 357 const GEPOperator *GEP2) { 358 unsigned AS = GEP1->getPointerAddressSpace(); 359 if (AS != GEP2->getPointerAddressSpace()) 360 return false; 361 362 if (TD) { 363 // When we have target data, we can reduce the GEP down to the value in bytes 364 // added to the address. 365 unsigned BitWidth = TD ? TD->getPointerSizeInBits(AS) : 1; 366 APInt Offset1(BitWidth, 0), Offset2(BitWidth, 0); 367 if (GEP1->accumulateConstantOffset(*TD, Offset1) && 368 GEP2->accumulateConstantOffset(*TD, Offset2)) { 369 return Offset1 == Offset2; 370 } 371 } 372 373 if (GEP1->getPointerOperand()->getType() != 374 GEP2->getPointerOperand()->getType()) 375 return false; 376 377 if (GEP1->getNumOperands() != GEP2->getNumOperands()) 378 return false; 379 380 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) { 381 if (!enumerate(GEP1->getOperand(i), GEP2->getOperand(i))) 382 return false; 383 } 384 385 return true; 386 } 387 388 // Compare two values used by the two functions under pair-wise comparison. If 389 // this is the first time the values are seen, they're added to the mapping so 390 // that we will detect mismatches on next use. 391 bool FunctionComparator::enumerate(const Value *V1, const Value *V2) { 392 // Check for function @f1 referring to itself and function @f2 referring to 393 // itself, or referring to each other, or both referring to either of them. 394 // They're all equivalent if the two functions are otherwise equivalent. 395 if (V1 == F1 && V2 == F2) 396 return true; 397 if (V1 == F2 && V2 == F1) 398 return true; 399 400 if (const Constant *C1 = dyn_cast<Constant>(V1)) { 401 if (V1 == V2) return true; 402 const Constant *C2 = dyn_cast<Constant>(V2); 403 if (!C2) return false; 404 // TODO: constant expressions with GEP or references to F1 or F2. 405 if (C1->isNullValue() && C2->isNullValue() && 406 isEquivalentType(C1->getType(), C2->getType())) 407 return true; 408 // Try bitcasting C2 to C1's type. If the bitcast is legal and returns C1 409 // then they must have equal bit patterns. 410 return C1->getType()->canLosslesslyBitCastTo(C2->getType()) && 411 C1 == ConstantExpr::getBitCast(const_cast<Constant*>(C2), C1->getType()); 412 } 413 414 if (isa<InlineAsm>(V1) || isa<InlineAsm>(V2)) 415 return V1 == V2; 416 417 // Check that V1 maps to V2. If we find a value that V1 maps to then we simply 418 // check whether it's equal to V2. When there is no mapping then we need to 419 // ensure that V2 isn't already equivalent to something else. For this 420 // purpose, we track the V2 values in a set. 421 422 const Value *&map_elem = id_map[V1]; 423 if (map_elem) 424 return map_elem == V2; 425 if (!seen_values.insert(V2).second) 426 return false; 427 map_elem = V2; 428 return true; 429 } 430 431 // Test whether two basic blocks have equivalent behaviour. 432 bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) { 433 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end(); 434 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end(); 435 436 do { 437 if (!enumerate(F1I, F2I)) 438 return false; 439 440 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) { 441 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I); 442 if (!GEP2) 443 return false; 444 445 if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand())) 446 return false; 447 448 if (!isEquivalentGEP(GEP1, GEP2)) 449 return false; 450 } else { 451 if (!isEquivalentOperation(F1I, F2I)) 452 return false; 453 454 assert(F1I->getNumOperands() == F2I->getNumOperands()); 455 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) { 456 Value *OpF1 = F1I->getOperand(i); 457 Value *OpF2 = F2I->getOperand(i); 458 459 if (!enumerate(OpF1, OpF2)) 460 return false; 461 462 if (OpF1->getValueID() != OpF2->getValueID() || 463 !isEquivalentType(OpF1->getType(), OpF2->getType())) 464 return false; 465 } 466 } 467 468 ++F1I, ++F2I; 469 } while (F1I != F1E && F2I != F2E); 470 471 return F1I == F1E && F2I == F2E; 472 } 473 474 // Test whether the two functions have equivalent behaviour. 475 bool FunctionComparator::compare() { 476 // We need to recheck everything, but check the things that weren't included 477 // in the hash first. 478 479 if (F1->getAttributes() != F2->getAttributes()) 480 return false; 481 482 if (F1->hasGC() != F2->hasGC()) 483 return false; 484 485 if (F1->hasGC() && F1->getGC() != F2->getGC()) 486 return false; 487 488 if (F1->hasSection() != F2->hasSection()) 489 return false; 490 491 if (F1->hasSection() && F1->getSection() != F2->getSection()) 492 return false; 493 494 if (F1->isVarArg() != F2->isVarArg()) 495 return false; 496 497 // TODO: if it's internal and only used in direct calls, we could handle this 498 // case too. 499 if (F1->getCallingConv() != F2->getCallingConv()) 500 return false; 501 502 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType())) 503 return false; 504 505 assert(F1->arg_size() == F2->arg_size() && 506 "Identically typed functions have different numbers of args!"); 507 508 // Visit the arguments so that they get enumerated in the order they're 509 // passed in. 510 for (Function::const_arg_iterator f1i = F1->arg_begin(), 511 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) { 512 if (!enumerate(f1i, f2i)) 513 llvm_unreachable("Arguments repeat!"); 514 } 515 516 // We do a CFG-ordered walk since the actual ordering of the blocks in the 517 // linked list is immaterial. Our walk starts at the entry block for both 518 // functions, then takes each block from each terminator in order. As an 519 // artifact, this also means that unreachable blocks are ignored. 520 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs; 521 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1. 522 523 F1BBs.push_back(&F1->getEntryBlock()); 524 F2BBs.push_back(&F2->getEntryBlock()); 525 526 VisitedBBs.insert(F1BBs[0]); 527 while (!F1BBs.empty()) { 528 const BasicBlock *F1BB = F1BBs.pop_back_val(); 529 const BasicBlock *F2BB = F2BBs.pop_back_val(); 530 531 if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB)) 532 return false; 533 534 const TerminatorInst *F1TI = F1BB->getTerminator(); 535 const TerminatorInst *F2TI = F2BB->getTerminator(); 536 537 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors()); 538 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) { 539 if (!VisitedBBs.insert(F1TI->getSuccessor(i))) 540 continue; 541 542 F1BBs.push_back(F1TI->getSuccessor(i)); 543 F2BBs.push_back(F2TI->getSuccessor(i)); 544 } 545 } 546 return true; 547 } 548 549 namespace { 550 551 /// MergeFunctions finds functions which will generate identical machine code, 552 /// by considering all pointer types to be equivalent. Once identified, 553 /// MergeFunctions will fold them by replacing a call to one to a call to a 554 /// bitcast of the other. 555 /// 556 class MergeFunctions : public ModulePass { 557 public: 558 static char ID; 559 MergeFunctions() 560 : ModulePass(ID), HasGlobalAliases(false) { 561 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry()); 562 } 563 564 bool runOnModule(Module &M); 565 566 private: 567 typedef DenseSet<ComparableFunction> FnSetType; 568 569 /// A work queue of functions that may have been modified and should be 570 /// analyzed again. 571 std::vector<WeakVH> Deferred; 572 573 /// Insert a ComparableFunction into the FnSet, or merge it away if it's 574 /// equal to one that's already present. 575 bool insert(ComparableFunction &NewF); 576 577 /// Remove a Function from the FnSet and queue it up for a second sweep of 578 /// analysis. 579 void remove(Function *F); 580 581 /// Find the functions that use this Value and remove them from FnSet and 582 /// queue the functions. 583 void removeUsers(Value *V); 584 585 /// Replace all direct calls of Old with calls of New. Will bitcast New if 586 /// necessary to make types match. 587 void replaceDirectCallers(Function *Old, Function *New); 588 589 /// Merge two equivalent functions. Upon completion, G may be deleted, or may 590 /// be converted into a thunk. In either case, it should never be visited 591 /// again. 592 void mergeTwoFunctions(Function *F, Function *G); 593 594 /// Replace G with a thunk or an alias to F. Deletes G. 595 void writeThunkOrAlias(Function *F, Function *G); 596 597 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses 598 /// of G with bitcast(F). Deletes G. 599 void writeThunk(Function *F, Function *G); 600 601 /// Replace G with an alias to F. Deletes G. 602 void writeAlias(Function *F, Function *G); 603 604 /// The set of all distinct functions. Use the insert() and remove() methods 605 /// to modify it. 606 FnSetType FnSet; 607 608 /// DataLayout for more accurate GEP comparisons. May be NULL. 609 DataLayout *TD; 610 611 /// Whether or not the target supports global aliases. 612 bool HasGlobalAliases; 613 }; 614 615 } // end anonymous namespace 616 617 char MergeFunctions::ID = 0; 618 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false) 619 620 ModulePass *llvm::createMergeFunctionsPass() { 621 return new MergeFunctions(); 622 } 623 624 bool MergeFunctions::runOnModule(Module &M) { 625 bool Changed = false; 626 TD = getAnalysisIfAvailable<DataLayout>(); 627 628 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { 629 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage()) 630 Deferred.push_back(WeakVH(I)); 631 } 632 FnSet.resize(Deferred.size()); 633 634 do { 635 std::vector<WeakVH> Worklist; 636 Deferred.swap(Worklist); 637 638 DEBUG(dbgs() << "size of module: " << M.size() << '\n'); 639 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n'); 640 641 // Insert only strong functions and merge them. Strong function merging 642 // always deletes one of them. 643 for (std::vector<WeakVH>::iterator I = Worklist.begin(), 644 E = Worklist.end(); I != E; ++I) { 645 if (!*I) continue; 646 Function *F = cast<Function>(*I); 647 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() && 648 !F->mayBeOverridden()) { 649 ComparableFunction CF = ComparableFunction(F, TD); 650 Changed |= insert(CF); 651 } 652 } 653 654 // Insert only weak functions and merge them. By doing these second we 655 // create thunks to the strong function when possible. When two weak 656 // functions are identical, we create a new strong function with two weak 657 // weak thunks to it which are identical but not mergable. 658 for (std::vector<WeakVH>::iterator I = Worklist.begin(), 659 E = Worklist.end(); I != E; ++I) { 660 if (!*I) continue; 661 Function *F = cast<Function>(*I); 662 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() && 663 F->mayBeOverridden()) { 664 ComparableFunction CF = ComparableFunction(F, TD); 665 Changed |= insert(CF); 666 } 667 } 668 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n'); 669 } while (!Deferred.empty()); 670 671 FnSet.clear(); 672 673 return Changed; 674 } 675 676 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS, 677 const ComparableFunction &RHS) { 678 if (LHS.getFunc() == RHS.getFunc() && 679 LHS.getHash() == RHS.getHash()) 680 return true; 681 if (!LHS.getFunc() || !RHS.getFunc()) 682 return false; 683 684 // One of these is a special "underlying pointer comparison only" object. 685 if (LHS.getTD() == ComparableFunction::LookupOnly || 686 RHS.getTD() == ComparableFunction::LookupOnly) 687 return false; 688 689 assert(LHS.getTD() == RHS.getTD() && 690 "Comparing functions for different targets"); 691 692 return FunctionComparator(LHS.getTD(), LHS.getFunc(), 693 RHS.getFunc()).compare(); 694 } 695 696 // Replace direct callers of Old with New. 697 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) { 698 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType()); 699 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end(); 700 UI != UE;) { 701 Value::use_iterator TheIter = UI; 702 ++UI; 703 CallSite CS(*TheIter); 704 if (CS && CS.isCallee(TheIter)) { 705 remove(CS.getInstruction()->getParent()->getParent()); 706 TheIter.getUse().set(BitcastNew); 707 } 708 } 709 } 710 711 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G. 712 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) { 713 if (HasGlobalAliases && G->hasUnnamedAddr()) { 714 if (G->hasExternalLinkage() || G->hasLocalLinkage() || 715 G->hasWeakLinkage()) { 716 writeAlias(F, G); 717 return; 718 } 719 } 720 721 writeThunk(F, G); 722 } 723 724 // Helper for writeThunk, 725 // Selects proper bitcast operation, 726 // but a bit simplier then CastInst::getCastOpcode. 727 static Value* createCast(IRBuilder<false> &Builder, Value *V, Type *DestTy) { 728 Type *SrcTy = V->getType(); 729 if (SrcTy->isIntegerTy() && DestTy->isPointerTy()) 730 return Builder.CreateIntToPtr(V, DestTy); 731 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy()) 732 return Builder.CreatePtrToInt(V, DestTy); 733 else 734 return Builder.CreateBitCast(V, DestTy); 735 } 736 737 // Replace G with a simple tail call to bitcast(F). Also replace direct uses 738 // of G with bitcast(F). Deletes G. 739 void MergeFunctions::writeThunk(Function *F, Function *G) { 740 if (!G->mayBeOverridden()) { 741 // Redirect direct callers of G to F. 742 replaceDirectCallers(G, F); 743 } 744 745 // If G was internal then we may have replaced all uses of G with F. If so, 746 // stop here and delete G. There's no need for a thunk. 747 if (G->hasLocalLinkage() && G->use_empty()) { 748 G->eraseFromParent(); 749 return; 750 } 751 752 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "", 753 G->getParent()); 754 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG); 755 IRBuilder<false> Builder(BB); 756 757 SmallVector<Value *, 16> Args; 758 unsigned i = 0; 759 FunctionType *FFTy = F->getFunctionType(); 760 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end(); 761 AI != AE; ++AI) { 762 Args.push_back(createCast(Builder, (Value*)AI, FFTy->getParamType(i))); 763 ++i; 764 } 765 766 CallInst *CI = Builder.CreateCall(F, Args); 767 CI->setTailCall(); 768 CI->setCallingConv(F->getCallingConv()); 769 if (NewG->getReturnType()->isVoidTy()) { 770 Builder.CreateRetVoid(); 771 } else { 772 Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType())); 773 } 774 775 NewG->copyAttributesFrom(G); 776 NewG->takeName(G); 777 removeUsers(G); 778 G->replaceAllUsesWith(NewG); 779 G->eraseFromParent(); 780 781 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n'); 782 ++NumThunksWritten; 783 } 784 785 // Replace G with an alias to F and delete G. 786 void MergeFunctions::writeAlias(Function *F, Function *G) { 787 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType()); 788 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "", 789 BitcastF, G->getParent()); 790 F->setAlignment(std::max(F->getAlignment(), G->getAlignment())); 791 GA->takeName(G); 792 GA->setVisibility(G->getVisibility()); 793 removeUsers(G); 794 G->replaceAllUsesWith(GA); 795 G->eraseFromParent(); 796 797 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n'); 798 ++NumAliasesWritten; 799 } 800 801 // Merge two equivalent functions. Upon completion, Function G is deleted. 802 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) { 803 if (F->mayBeOverridden()) { 804 assert(G->mayBeOverridden()); 805 806 if (HasGlobalAliases) { 807 // Make them both thunks to the same internal function. 808 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "", 809 F->getParent()); 810 H->copyAttributesFrom(F); 811 H->takeName(F); 812 removeUsers(F); 813 F->replaceAllUsesWith(H); 814 815 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment()); 816 817 writeAlias(F, G); 818 writeAlias(F, H); 819 820 F->setAlignment(MaxAlignment); 821 F->setLinkage(GlobalValue::PrivateLinkage); 822 } else { 823 // We can't merge them. Instead, pick one and update all direct callers 824 // to call it and hope that we improve the instruction cache hit rate. 825 replaceDirectCallers(G, F); 826 } 827 828 ++NumDoubleWeak; 829 } else { 830 writeThunkOrAlias(F, G); 831 } 832 833 ++NumFunctionsMerged; 834 } 835 836 // Insert a ComparableFunction into the FnSet, or merge it away if equal to one 837 // that was already inserted. 838 bool MergeFunctions::insert(ComparableFunction &NewF) { 839 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF); 840 if (Result.second) { 841 DEBUG(dbgs() << "Inserting as unique: " << NewF.getFunc()->getName() << '\n'); 842 return false; 843 } 844 845 const ComparableFunction &OldF = *Result.first; 846 847 // Don't merge tiny functions, since it can just end up making the function 848 // larger. 849 // FIXME: Should still merge them if they are unnamed_addr and produce an 850 // alias. 851 if (NewF.getFunc()->size() == 1) { 852 if (NewF.getFunc()->front().size() <= 2) { 853 DEBUG(dbgs() << NewF.getFunc()->getName() 854 << " is to small to bother merging\n"); 855 return false; 856 } 857 } 858 859 // Never thunk a strong function to a weak function. 860 assert(!OldF.getFunc()->mayBeOverridden() || 861 NewF.getFunc()->mayBeOverridden()); 862 863 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == " 864 << NewF.getFunc()->getName() << '\n'); 865 866 Function *DeleteF = NewF.getFunc(); 867 NewF.release(); 868 mergeTwoFunctions(OldF.getFunc(), DeleteF); 869 return true; 870 } 871 872 // Remove a function from FnSet. If it was already in FnSet, add it to Deferred 873 // so that we'll look at it in the next round. 874 void MergeFunctions::remove(Function *F) { 875 // We need to make sure we remove F, not a function "equal" to F per the 876 // function equality comparator. 877 // 878 // The special "lookup only" ComparableFunction bypasses the expensive 879 // function comparison in favour of a pointer comparison on the underlying 880 // Function*'s. 881 ComparableFunction CF = ComparableFunction(F, ComparableFunction::LookupOnly); 882 if (FnSet.erase(CF)) { 883 DEBUG(dbgs() << "Removed " << F->getName() << " from set and deferred it.\n"); 884 Deferred.push_back(F); 885 } 886 } 887 888 // For each instruction used by the value, remove() the function that contains 889 // the instruction. This should happen right before a call to RAUW. 890 void MergeFunctions::removeUsers(Value *V) { 891 std::vector<Value *> Worklist; 892 Worklist.push_back(V); 893 while (!Worklist.empty()) { 894 Value *V = Worklist.back(); 895 Worklist.pop_back(); 896 897 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 898 UI != UE; ++UI) { 899 Use &U = UI.getUse(); 900 if (Instruction *I = dyn_cast<Instruction>(U.getUser())) { 901 remove(I->getParent()->getParent()); 902 } else if (isa<GlobalValue>(U.getUser())) { 903 // do nothing 904 } else if (Constant *C = dyn_cast<Constant>(U.getUser())) { 905 for (Value::use_iterator CUI = C->use_begin(), CUE = C->use_end(); 906 CUI != CUE; ++CUI) 907 Worklist.push_back(*CUI); 908 } 909 } 910 } 911 } 912