1 //===- ObjCARC.cpp - ObjC ARC Optimization --------------------------------===// 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 file defines ObjC ARC optimizations. ARC stands for 11 // Automatic Reference Counting and is a system for managing reference counts 12 // for objects in Objective C. 13 // 14 // The optimizations performed include elimination of redundant, partially 15 // redundant, and inconsequential reference count operations, elimination of 16 // redundant weak pointer operations, pattern-matching and replacement of 17 // low-level operations into higher-level operations, and numerous minor 18 // simplifications. 19 // 20 // This file also defines a simple ARC-aware AliasAnalysis. 21 // 22 // WARNING: This file knows about certain library functions. It recognizes them 23 // by name, and hardwires knowedge of their semantics. 24 // 25 // WARNING: This file knows about how certain Objective-C library functions are 26 // used. Naive LLVM IR transformations which would otherwise be 27 // behavior-preserving may break these assumptions. 28 // 29 //===----------------------------------------------------------------------===// 30 31 #define DEBUG_TYPE "objc-arc" 32 #include "llvm/Function.h" 33 #include "llvm/Intrinsics.h" 34 #include "llvm/GlobalVariable.h" 35 #include "llvm/DerivedTypes.h" 36 #include "llvm/Module.h" 37 #include "llvm/Analysis/ValueTracking.h" 38 #include "llvm/Transforms/Utils/Local.h" 39 #include "llvm/Support/CallSite.h" 40 #include "llvm/Support/CommandLine.h" 41 #include "llvm/ADT/StringSwitch.h" 42 #include "llvm/ADT/DenseMap.h" 43 #include "llvm/ADT/STLExtras.h" 44 using namespace llvm; 45 46 // A handy option to enable/disable all optimizations in this file. 47 static cl::opt<bool> EnableARCOpts("enable-objc-arc-opts", cl::init(true)); 48 49 //===----------------------------------------------------------------------===// 50 // Misc. Utilities 51 //===----------------------------------------------------------------------===// 52 53 namespace { 54 /// MapVector - An associative container with fast insertion-order 55 /// (deterministic) iteration over its elements. Plus the special 56 /// blot operation. 57 template<class KeyT, class ValueT> 58 class MapVector { 59 /// Map - Map keys to indices in Vector. 60 typedef DenseMap<KeyT, size_t> MapTy; 61 MapTy Map; 62 63 /// Vector - Keys and values. 64 typedef std::vector<std::pair<KeyT, ValueT> > VectorTy; 65 VectorTy Vector; 66 67 public: 68 typedef typename VectorTy::iterator iterator; 69 typedef typename VectorTy::const_iterator const_iterator; 70 iterator begin() { return Vector.begin(); } 71 iterator end() { return Vector.end(); } 72 const_iterator begin() const { return Vector.begin(); } 73 const_iterator end() const { return Vector.end(); } 74 75 #ifdef XDEBUG 76 ~MapVector() { 77 assert(Vector.size() >= Map.size()); // May differ due to blotting. 78 for (typename MapTy::const_iterator I = Map.begin(), E = Map.end(); 79 I != E; ++I) { 80 assert(I->second < Vector.size()); 81 assert(Vector[I->second].first == I->first); 82 } 83 for (typename VectorTy::const_iterator I = Vector.begin(), 84 E = Vector.end(); I != E; ++I) 85 assert(!I->first || 86 (Map.count(I->first) && 87 Map[I->first] == size_t(I - Vector.begin()))); 88 } 89 #endif 90 91 ValueT &operator[](const KeyT &Arg) { 92 std::pair<typename MapTy::iterator, bool> Pair = 93 Map.insert(std::make_pair(Arg, size_t(0))); 94 if (Pair.second) { 95 size_t Num = Vector.size(); 96 Pair.first->second = Num; 97 Vector.push_back(std::make_pair(Arg, ValueT())); 98 return Vector[Num].second; 99 } 100 return Vector[Pair.first->second].second; 101 } 102 103 std::pair<iterator, bool> 104 insert(const std::pair<KeyT, ValueT> &InsertPair) { 105 std::pair<typename MapTy::iterator, bool> Pair = 106 Map.insert(std::make_pair(InsertPair.first, size_t(0))); 107 if (Pair.second) { 108 size_t Num = Vector.size(); 109 Pair.first->second = Num; 110 Vector.push_back(InsertPair); 111 return std::make_pair(Vector.begin() + Num, true); 112 } 113 return std::make_pair(Vector.begin() + Pair.first->second, false); 114 } 115 116 const_iterator find(const KeyT &Key) const { 117 typename MapTy::const_iterator It = Map.find(Key); 118 if (It == Map.end()) return Vector.end(); 119 return Vector.begin() + It->second; 120 } 121 122 /// blot - This is similar to erase, but instead of removing the element 123 /// from the vector, it just zeros out the key in the vector. This leaves 124 /// iterators intact, but clients must be prepared for zeroed-out keys when 125 /// iterating. 126 void blot(const KeyT &Key) { 127 typename MapTy::iterator It = Map.find(Key); 128 if (It == Map.end()) return; 129 Vector[It->second].first = KeyT(); 130 Map.erase(It); 131 } 132 133 void clear() { 134 Map.clear(); 135 Vector.clear(); 136 } 137 }; 138 } 139 140 //===----------------------------------------------------------------------===// 141 // ARC Utilities. 142 //===----------------------------------------------------------------------===// 143 144 namespace { 145 /// InstructionClass - A simple classification for instructions. 146 enum InstructionClass { 147 IC_Retain, ///< objc_retain 148 IC_RetainRV, ///< objc_retainAutoreleasedReturnValue 149 IC_RetainBlock, ///< objc_retainBlock 150 IC_Release, ///< objc_release 151 IC_Autorelease, ///< objc_autorelease 152 IC_AutoreleaseRV, ///< objc_autoreleaseReturnValue 153 IC_AutoreleasepoolPush, ///< objc_autoreleasePoolPush 154 IC_AutoreleasepoolPop, ///< objc_autoreleasePoolPop 155 IC_NoopCast, ///< objc_retainedObject, etc. 156 IC_FusedRetainAutorelease, ///< objc_retainAutorelease 157 IC_FusedRetainAutoreleaseRV, ///< objc_retainAutoreleaseReturnValue 158 IC_LoadWeakRetained, ///< objc_loadWeakRetained (primitive) 159 IC_StoreWeak, ///< objc_storeWeak (primitive) 160 IC_InitWeak, ///< objc_initWeak (derived) 161 IC_LoadWeak, ///< objc_loadWeak (derived) 162 IC_MoveWeak, ///< objc_moveWeak (derived) 163 IC_CopyWeak, ///< objc_copyWeak (derived) 164 IC_DestroyWeak, ///< objc_destroyWeak (derived) 165 IC_StoreStrong, ///< objc_storeStrong (derived) 166 IC_CallOrUser, ///< could call objc_release and/or "use" pointers 167 IC_Call, ///< could call objc_release 168 IC_User, ///< could "use" a pointer 169 IC_None ///< anything else 170 }; 171 } 172 173 /// IsPotentialUse - Test whether the given value is possible a 174 /// reference-counted pointer. 175 static bool IsPotentialUse(const Value *Op) { 176 // Pointers to static or stack storage are not reference-counted pointers. 177 if (isa<Constant>(Op) || isa<AllocaInst>(Op)) 178 return false; 179 // Special arguments are not reference-counted. 180 if (const Argument *Arg = dyn_cast<Argument>(Op)) 181 if (Arg->hasByValAttr() || 182 Arg->hasNestAttr() || 183 Arg->hasStructRetAttr()) 184 return false; 185 // Only consider values with pointer types. 186 // It seemes intuitive to exclude function pointer types as well, since 187 // functions are never reference-counted, however clang occasionally 188 // bitcasts reference-counted pointers to function-pointer type 189 // temporarily. 190 PointerType *Ty = dyn_cast<PointerType>(Op->getType()); 191 if (!Ty) 192 return false; 193 // Conservatively assume anything else is a potential use. 194 return true; 195 } 196 197 /// GetCallSiteClass - Helper for GetInstructionClass. Determines what kind 198 /// of construct CS is. 199 static InstructionClass GetCallSiteClass(ImmutableCallSite CS) { 200 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 201 I != E; ++I) 202 if (IsPotentialUse(*I)) 203 return CS.onlyReadsMemory() ? IC_User : IC_CallOrUser; 204 205 return CS.onlyReadsMemory() ? IC_None : IC_Call; 206 } 207 208 /// GetFunctionClass - Determine if F is one of the special known Functions. 209 /// If it isn't, return IC_CallOrUser. 210 static InstructionClass GetFunctionClass(const Function *F) { 211 Function::const_arg_iterator AI = F->arg_begin(), AE = F->arg_end(); 212 213 // No arguments. 214 if (AI == AE) 215 return StringSwitch<InstructionClass>(F->getName()) 216 .Case("objc_autoreleasePoolPush", IC_AutoreleasepoolPush) 217 .Default(IC_CallOrUser); 218 219 // One argument. 220 const Argument *A0 = AI++; 221 if (AI == AE) 222 // Argument is a pointer. 223 if (PointerType *PTy = dyn_cast<PointerType>(A0->getType())) { 224 Type *ETy = PTy->getElementType(); 225 // Argument is i8*. 226 if (ETy->isIntegerTy(8)) 227 return StringSwitch<InstructionClass>(F->getName()) 228 .Case("objc_retain", IC_Retain) 229 .Case("objc_retainAutoreleasedReturnValue", IC_RetainRV) 230 .Case("objc_retainBlock", IC_RetainBlock) 231 .Case("objc_release", IC_Release) 232 .Case("objc_autorelease", IC_Autorelease) 233 .Case("objc_autoreleaseReturnValue", IC_AutoreleaseRV) 234 .Case("objc_autoreleasePoolPop", IC_AutoreleasepoolPop) 235 .Case("objc_retainedObject", IC_NoopCast) 236 .Case("objc_unretainedObject", IC_NoopCast) 237 .Case("objc_unretainedPointer", IC_NoopCast) 238 .Case("objc_retain_autorelease", IC_FusedRetainAutorelease) 239 .Case("objc_retainAutorelease", IC_FusedRetainAutorelease) 240 .Case("objc_retainAutoreleaseReturnValue",IC_FusedRetainAutoreleaseRV) 241 .Default(IC_CallOrUser); 242 243 // Argument is i8** 244 if (PointerType *Pte = dyn_cast<PointerType>(ETy)) 245 if (Pte->getElementType()->isIntegerTy(8)) 246 return StringSwitch<InstructionClass>(F->getName()) 247 .Case("objc_loadWeakRetained", IC_LoadWeakRetained) 248 .Case("objc_loadWeak", IC_LoadWeak) 249 .Case("objc_destroyWeak", IC_DestroyWeak) 250 .Default(IC_CallOrUser); 251 } 252 253 // Two arguments, first is i8**. 254 const Argument *A1 = AI++; 255 if (AI == AE) 256 if (PointerType *PTy = dyn_cast<PointerType>(A0->getType())) 257 if (PointerType *Pte = dyn_cast<PointerType>(PTy->getElementType())) 258 if (Pte->getElementType()->isIntegerTy(8)) 259 if (PointerType *PTy1 = dyn_cast<PointerType>(A1->getType())) { 260 Type *ETy1 = PTy1->getElementType(); 261 // Second argument is i8* 262 if (ETy1->isIntegerTy(8)) 263 return StringSwitch<InstructionClass>(F->getName()) 264 .Case("objc_storeWeak", IC_StoreWeak) 265 .Case("objc_initWeak", IC_InitWeak) 266 .Case("objc_storeStrong", IC_StoreStrong) 267 .Default(IC_CallOrUser); 268 // Second argument is i8**. 269 if (PointerType *Pte1 = dyn_cast<PointerType>(ETy1)) 270 if (Pte1->getElementType()->isIntegerTy(8)) 271 return StringSwitch<InstructionClass>(F->getName()) 272 .Case("objc_moveWeak", IC_MoveWeak) 273 .Case("objc_copyWeak", IC_CopyWeak) 274 .Default(IC_CallOrUser); 275 } 276 277 // Anything else. 278 return IC_CallOrUser; 279 } 280 281 /// GetInstructionClass - Determine what kind of construct V is. 282 static InstructionClass GetInstructionClass(const Value *V) { 283 if (const Instruction *I = dyn_cast<Instruction>(V)) { 284 // Any instruction other than bitcast and gep with a pointer operand have a 285 // use of an objc pointer. Bitcasts, GEPs, Selects, PHIs transfer a pointer 286 // to a subsequent use, rather than using it themselves, in this sense. 287 // As a short cut, several other opcodes are known to have no pointer 288 // operands of interest. And ret is never followed by a release, so it's 289 // not interesting to examine. 290 switch (I->getOpcode()) { 291 case Instruction::Call: { 292 const CallInst *CI = cast<CallInst>(I); 293 // Check for calls to special functions. 294 if (const Function *F = CI->getCalledFunction()) { 295 InstructionClass Class = GetFunctionClass(F); 296 if (Class != IC_CallOrUser) 297 return Class; 298 299 // None of the intrinsic functions do objc_release. For intrinsics, the 300 // only question is whether or not they may be users. 301 switch (F->getIntrinsicID()) { 302 case 0: break; 303 case Intrinsic::bswap: case Intrinsic::ctpop: 304 case Intrinsic::ctlz: case Intrinsic::cttz: 305 case Intrinsic::returnaddress: case Intrinsic::frameaddress: 306 case Intrinsic::stacksave: case Intrinsic::stackrestore: 307 case Intrinsic::vastart: case Intrinsic::vacopy: case Intrinsic::vaend: 308 // Don't let dbg info affect our results. 309 case Intrinsic::dbg_declare: case Intrinsic::dbg_value: 310 // Short cut: Some intrinsics obviously don't use ObjC pointers. 311 return IC_None; 312 default: 313 for (Function::const_arg_iterator AI = F->arg_begin(), 314 AE = F->arg_end(); AI != AE; ++AI) 315 if (IsPotentialUse(AI)) 316 return IC_User; 317 return IC_None; 318 } 319 } 320 return GetCallSiteClass(CI); 321 } 322 case Instruction::Invoke: 323 return GetCallSiteClass(cast<InvokeInst>(I)); 324 case Instruction::BitCast: 325 case Instruction::GetElementPtr: 326 case Instruction::Select: case Instruction::PHI: 327 case Instruction::Ret: case Instruction::Br: 328 case Instruction::Switch: case Instruction::IndirectBr: 329 case Instruction::Alloca: case Instruction::VAArg: 330 case Instruction::Add: case Instruction::FAdd: 331 case Instruction::Sub: case Instruction::FSub: 332 case Instruction::Mul: case Instruction::FMul: 333 case Instruction::SDiv: case Instruction::UDiv: case Instruction::FDiv: 334 case Instruction::SRem: case Instruction::URem: case Instruction::FRem: 335 case Instruction::Shl: case Instruction::LShr: case Instruction::AShr: 336 case Instruction::And: case Instruction::Or: case Instruction::Xor: 337 case Instruction::SExt: case Instruction::ZExt: case Instruction::Trunc: 338 case Instruction::IntToPtr: case Instruction::FCmp: 339 case Instruction::FPTrunc: case Instruction::FPExt: 340 case Instruction::FPToUI: case Instruction::FPToSI: 341 case Instruction::UIToFP: case Instruction::SIToFP: 342 case Instruction::InsertElement: case Instruction::ExtractElement: 343 case Instruction::ShuffleVector: 344 case Instruction::ExtractValue: 345 break; 346 case Instruction::ICmp: 347 // Comparing a pointer with null, or any other constant, isn't an 348 // interesting use, because we don't care what the pointer points to, or 349 // about the values of any other dynamic reference-counted pointers. 350 if (IsPotentialUse(I->getOperand(1))) 351 return IC_User; 352 break; 353 default: 354 // For anything else, check all the operands. 355 // Note that this includes both operands of a Store: while the first 356 // operand isn't actually being dereferenced, it is being stored to 357 // memory where we can no longer track who might read it and dereference 358 // it, so we have to consider it potentially used. 359 for (User::const_op_iterator OI = I->op_begin(), OE = I->op_end(); 360 OI != OE; ++OI) 361 if (IsPotentialUse(*OI)) 362 return IC_User; 363 } 364 } 365 366 // Otherwise, it's totally inert for ARC purposes. 367 return IC_None; 368 } 369 370 /// GetBasicInstructionClass - Determine what kind of construct V is. This is 371 /// similar to GetInstructionClass except that it only detects objc runtine 372 /// calls. This allows it to be faster. 373 static InstructionClass GetBasicInstructionClass(const Value *V) { 374 if (const CallInst *CI = dyn_cast<CallInst>(V)) { 375 if (const Function *F = CI->getCalledFunction()) 376 return GetFunctionClass(F); 377 // Otherwise, be conservative. 378 return IC_CallOrUser; 379 } 380 381 // Otherwise, be conservative. 382 return isa<InvokeInst>(V) ? IC_CallOrUser : IC_User; 383 } 384 385 /// IsRetain - Test if the the given class is objc_retain or 386 /// equivalent. 387 static bool IsRetain(InstructionClass Class) { 388 return Class == IC_Retain || 389 Class == IC_RetainRV; 390 } 391 392 /// IsAutorelease - Test if the the given class is objc_autorelease or 393 /// equivalent. 394 static bool IsAutorelease(InstructionClass Class) { 395 return Class == IC_Autorelease || 396 Class == IC_AutoreleaseRV; 397 } 398 399 /// IsForwarding - Test if the given class represents instructions which return 400 /// their argument verbatim. 401 static bool IsForwarding(InstructionClass Class) { 402 // objc_retainBlock technically doesn't always return its argument 403 // verbatim, but it doesn't matter for our purposes here. 404 return Class == IC_Retain || 405 Class == IC_RetainRV || 406 Class == IC_Autorelease || 407 Class == IC_AutoreleaseRV || 408 Class == IC_RetainBlock || 409 Class == IC_NoopCast; 410 } 411 412 /// IsNoopOnNull - Test if the given class represents instructions which do 413 /// nothing if passed a null pointer. 414 static bool IsNoopOnNull(InstructionClass Class) { 415 return Class == IC_Retain || 416 Class == IC_RetainRV || 417 Class == IC_Release || 418 Class == IC_Autorelease || 419 Class == IC_AutoreleaseRV || 420 Class == IC_RetainBlock; 421 } 422 423 /// IsAlwaysTail - Test if the given class represents instructions which are 424 /// always safe to mark with the "tail" keyword. 425 static bool IsAlwaysTail(InstructionClass Class) { 426 // IC_RetainBlock may be given a stack argument. 427 return Class == IC_Retain || 428 Class == IC_RetainRV || 429 Class == IC_Autorelease || 430 Class == IC_AutoreleaseRV; 431 } 432 433 /// IsNoThrow - Test if the given class represents instructions which are always 434 /// safe to mark with the nounwind attribute.. 435 static bool IsNoThrow(InstructionClass Class) { 436 // objc_retainBlock is not nounwind because it calls user copy constructors 437 // which could theoretically throw. 438 return Class == IC_Retain || 439 Class == IC_RetainRV || 440 Class == IC_Release || 441 Class == IC_Autorelease || 442 Class == IC_AutoreleaseRV || 443 Class == IC_AutoreleasepoolPush || 444 Class == IC_AutoreleasepoolPop; 445 } 446 447 /// EraseInstruction - Erase the given instruction. ObjC calls return their 448 /// argument verbatim, so if it's such a call and the return value has users, 449 /// replace them with the argument value. 450 static void EraseInstruction(Instruction *CI) { 451 Value *OldArg = cast<CallInst>(CI)->getArgOperand(0); 452 453 bool Unused = CI->use_empty(); 454 455 if (!Unused) { 456 // Replace the return value with the argument. 457 assert(IsForwarding(GetBasicInstructionClass(CI)) && 458 "Can't delete non-forwarding instruction with users!"); 459 CI->replaceAllUsesWith(OldArg); 460 } 461 462 CI->eraseFromParent(); 463 464 if (Unused) 465 RecursivelyDeleteTriviallyDeadInstructions(OldArg); 466 } 467 468 /// GetUnderlyingObjCPtr - This is a wrapper around getUnderlyingObject which 469 /// also knows how to look through objc_retain and objc_autorelease calls, which 470 /// we know to return their argument verbatim. 471 static const Value *GetUnderlyingObjCPtr(const Value *V) { 472 for (;;) { 473 V = GetUnderlyingObject(V); 474 if (!IsForwarding(GetBasicInstructionClass(V))) 475 break; 476 V = cast<CallInst>(V)->getArgOperand(0); 477 } 478 479 return V; 480 } 481 482 /// StripPointerCastsAndObjCCalls - This is a wrapper around 483 /// Value::stripPointerCasts which also knows how to look through objc_retain 484 /// and objc_autorelease calls, which we know to return their argument verbatim. 485 static const Value *StripPointerCastsAndObjCCalls(const Value *V) { 486 for (;;) { 487 V = V->stripPointerCasts(); 488 if (!IsForwarding(GetBasicInstructionClass(V))) 489 break; 490 V = cast<CallInst>(V)->getArgOperand(0); 491 } 492 return V; 493 } 494 495 /// StripPointerCastsAndObjCCalls - This is a wrapper around 496 /// Value::stripPointerCasts which also knows how to look through objc_retain 497 /// and objc_autorelease calls, which we know to return their argument verbatim. 498 static Value *StripPointerCastsAndObjCCalls(Value *V) { 499 for (;;) { 500 V = V->stripPointerCasts(); 501 if (!IsForwarding(GetBasicInstructionClass(V))) 502 break; 503 V = cast<CallInst>(V)->getArgOperand(0); 504 } 505 return V; 506 } 507 508 /// GetObjCArg - Assuming the given instruction is one of the special calls such 509 /// as objc_retain or objc_release, return the argument value, stripped of no-op 510 /// casts and forwarding calls. 511 static Value *GetObjCArg(Value *Inst) { 512 return StripPointerCastsAndObjCCalls(cast<CallInst>(Inst)->getArgOperand(0)); 513 } 514 515 /// IsObjCIdentifiedObject - This is similar to AliasAnalysis' 516 /// isObjCIdentifiedObject, except that it uses special knowledge of 517 /// ObjC conventions... 518 static bool IsObjCIdentifiedObject(const Value *V) { 519 // Assume that call results and arguments have their own "provenance". 520 // Constants (including GlobalVariables) and Allocas are never 521 // reference-counted. 522 if (isa<CallInst>(V) || isa<InvokeInst>(V) || 523 isa<Argument>(V) || isa<Constant>(V) || 524 isa<AllocaInst>(V)) 525 return true; 526 527 if (const LoadInst *LI = dyn_cast<LoadInst>(V)) { 528 const Value *Pointer = 529 StripPointerCastsAndObjCCalls(LI->getPointerOperand()); 530 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Pointer)) { 531 // A constant pointer can't be pointing to an object on the heap. It may 532 // be reference-counted, but it won't be deleted. 533 if (GV->isConstant()) 534 return true; 535 StringRef Name = GV->getName(); 536 // These special variables are known to hold values which are not 537 // reference-counted pointers. 538 if (Name.startswith("\01L_OBJC_SELECTOR_REFERENCES_") || 539 Name.startswith("\01L_OBJC_CLASSLIST_REFERENCES_") || 540 Name.startswith("\01L_OBJC_CLASSLIST_SUP_REFS_$_") || 541 Name.startswith("\01L_OBJC_METH_VAR_NAME_") || 542 Name.startswith("\01l_objc_msgSend_fixup_")) 543 return true; 544 } 545 } 546 547 return false; 548 } 549 550 /// FindSingleUseIdentifiedObject - This is similar to 551 /// StripPointerCastsAndObjCCalls but it stops as soon as it finds a value 552 /// with multiple uses. 553 static const Value *FindSingleUseIdentifiedObject(const Value *Arg) { 554 if (Arg->hasOneUse()) { 555 if (const BitCastInst *BC = dyn_cast<BitCastInst>(Arg)) 556 return FindSingleUseIdentifiedObject(BC->getOperand(0)); 557 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Arg)) 558 if (GEP->hasAllZeroIndices()) 559 return FindSingleUseIdentifiedObject(GEP->getPointerOperand()); 560 if (IsForwarding(GetBasicInstructionClass(Arg))) 561 return FindSingleUseIdentifiedObject( 562 cast<CallInst>(Arg)->getArgOperand(0)); 563 if (!IsObjCIdentifiedObject(Arg)) 564 return 0; 565 return Arg; 566 } 567 568 // If we found an identifiable object but it has multiple uses, but they 569 // are trivial uses, we can still consider this to be a single-use 570 // value. 571 if (IsObjCIdentifiedObject(Arg)) { 572 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end(); 573 UI != UE; ++UI) { 574 const User *U = *UI; 575 if (!U->use_empty() || StripPointerCastsAndObjCCalls(U) != Arg) 576 return 0; 577 } 578 579 return Arg; 580 } 581 582 return 0; 583 } 584 585 /// ModuleHasARC - Test if the given module looks interesting to run ARC 586 /// optimization on. 587 static bool ModuleHasARC(const Module &M) { 588 return 589 M.getNamedValue("objc_retain") || 590 M.getNamedValue("objc_release") || 591 M.getNamedValue("objc_autorelease") || 592 M.getNamedValue("objc_retainAutoreleasedReturnValue") || 593 M.getNamedValue("objc_retainBlock") || 594 M.getNamedValue("objc_autoreleaseReturnValue") || 595 M.getNamedValue("objc_autoreleasePoolPush") || 596 M.getNamedValue("objc_loadWeakRetained") || 597 M.getNamedValue("objc_loadWeak") || 598 M.getNamedValue("objc_destroyWeak") || 599 M.getNamedValue("objc_storeWeak") || 600 M.getNamedValue("objc_initWeak") || 601 M.getNamedValue("objc_moveWeak") || 602 M.getNamedValue("objc_copyWeak") || 603 M.getNamedValue("objc_retainedObject") || 604 M.getNamedValue("objc_unretainedObject") || 605 M.getNamedValue("objc_unretainedPointer"); 606 } 607 608 /// DoesObjCBlockEscape - Test whether the given pointer, which is an 609 /// Objective C block pointer, does not "escape". This differs from regular 610 /// escape analysis in that a use as an argument to a call is not considered 611 /// an escape. 612 static bool DoesObjCBlockEscape(const Value *BlockPtr) { 613 // Walk the def-use chains. 614 SmallVector<const Value *, 4> Worklist; 615 Worklist.push_back(BlockPtr); 616 do { 617 const Value *V = Worklist.pop_back_val(); 618 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end(); 619 UI != UE; ++UI) { 620 const User *UUser = *UI; 621 // Special - Use by a call (callee or argument) is not considered 622 // to be an escape. 623 switch (GetBasicInstructionClass(UUser)) { 624 case IC_StoreWeak: 625 case IC_InitWeak: 626 case IC_StoreStrong: 627 case IC_Autorelease: 628 case IC_AutoreleaseRV: 629 // These special functions make copies of their pointer arguments. 630 return true; 631 case IC_User: 632 case IC_None: 633 // Use by an instruction which copies the value is an escape if the 634 // result is an escape. 635 if (isa<BitCastInst>(UUser) || isa<GetElementPtrInst>(UUser) || 636 isa<PHINode>(UUser) || isa<SelectInst>(UUser)) { 637 Worklist.push_back(UUser); 638 continue; 639 } 640 // Use by a load is not an escape. 641 if (isa<LoadInst>(UUser)) 642 continue; 643 // Use by a store is not an escape if the use is the address. 644 if (const StoreInst *SI = dyn_cast<StoreInst>(UUser)) 645 if (V != SI->getValueOperand()) 646 continue; 647 break; 648 default: 649 // Regular calls and other stuff are not considered escapes. 650 continue; 651 } 652 // Otherwise, conservatively assume an escape. 653 return true; 654 } 655 } while (!Worklist.empty()); 656 657 // No escapes found. 658 return false; 659 } 660 661 //===----------------------------------------------------------------------===// 662 // ARC AliasAnalysis. 663 //===----------------------------------------------------------------------===// 664 665 #include "llvm/Pass.h" 666 #include "llvm/Analysis/AliasAnalysis.h" 667 #include "llvm/Analysis/Passes.h" 668 669 namespace { 670 /// ObjCARCAliasAnalysis - This is a simple alias analysis 671 /// implementation that uses knowledge of ARC constructs to answer queries. 672 /// 673 /// TODO: This class could be generalized to know about other ObjC-specific 674 /// tricks. Such as knowing that ivars in the non-fragile ABI are non-aliasing 675 /// even though their offsets are dynamic. 676 class ObjCARCAliasAnalysis : public ImmutablePass, 677 public AliasAnalysis { 678 public: 679 static char ID; // Class identification, replacement for typeinfo 680 ObjCARCAliasAnalysis() : ImmutablePass(ID) { 681 initializeObjCARCAliasAnalysisPass(*PassRegistry::getPassRegistry()); 682 } 683 684 private: 685 virtual void initializePass() { 686 InitializeAliasAnalysis(this); 687 } 688 689 /// getAdjustedAnalysisPointer - This method is used when a pass implements 690 /// an analysis interface through multiple inheritance. If needed, it 691 /// should override this to adjust the this pointer as needed for the 692 /// specified pass info. 693 virtual void *getAdjustedAnalysisPointer(const void *PI) { 694 if (PI == &AliasAnalysis::ID) 695 return (AliasAnalysis*)this; 696 return this; 697 } 698 699 virtual void getAnalysisUsage(AnalysisUsage &AU) const; 700 virtual AliasResult alias(const Location &LocA, const Location &LocB); 701 virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal); 702 virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS); 703 virtual ModRefBehavior getModRefBehavior(const Function *F); 704 virtual ModRefResult getModRefInfo(ImmutableCallSite CS, 705 const Location &Loc); 706 virtual ModRefResult getModRefInfo(ImmutableCallSite CS1, 707 ImmutableCallSite CS2); 708 }; 709 } // End of anonymous namespace 710 711 // Register this pass... 712 char ObjCARCAliasAnalysis::ID = 0; 713 INITIALIZE_AG_PASS(ObjCARCAliasAnalysis, AliasAnalysis, "objc-arc-aa", 714 "ObjC-ARC-Based Alias Analysis", false, true, false) 715 716 ImmutablePass *llvm::createObjCARCAliasAnalysisPass() { 717 return new ObjCARCAliasAnalysis(); 718 } 719 720 void 721 ObjCARCAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 722 AU.setPreservesAll(); 723 AliasAnalysis::getAnalysisUsage(AU); 724 } 725 726 AliasAnalysis::AliasResult 727 ObjCARCAliasAnalysis::alias(const Location &LocA, const Location &LocB) { 728 if (!EnableARCOpts) 729 return AliasAnalysis::alias(LocA, LocB); 730 731 // First, strip off no-ops, including ObjC-specific no-ops, and try making a 732 // precise alias query. 733 const Value *SA = StripPointerCastsAndObjCCalls(LocA.Ptr); 734 const Value *SB = StripPointerCastsAndObjCCalls(LocB.Ptr); 735 AliasResult Result = 736 AliasAnalysis::alias(Location(SA, LocA.Size, LocA.TBAATag), 737 Location(SB, LocB.Size, LocB.TBAATag)); 738 if (Result != MayAlias) 739 return Result; 740 741 // If that failed, climb to the underlying object, including climbing through 742 // ObjC-specific no-ops, and try making an imprecise alias query. 743 const Value *UA = GetUnderlyingObjCPtr(SA); 744 const Value *UB = GetUnderlyingObjCPtr(SB); 745 if (UA != SA || UB != SB) { 746 Result = AliasAnalysis::alias(Location(UA), Location(UB)); 747 // We can't use MustAlias or PartialAlias results here because 748 // GetUnderlyingObjCPtr may return an offsetted pointer value. 749 if (Result == NoAlias) 750 return NoAlias; 751 } 752 753 // If that failed, fail. We don't need to chain here, since that's covered 754 // by the earlier precise query. 755 return MayAlias; 756 } 757 758 bool 759 ObjCARCAliasAnalysis::pointsToConstantMemory(const Location &Loc, 760 bool OrLocal) { 761 if (!EnableARCOpts) 762 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal); 763 764 // First, strip off no-ops, including ObjC-specific no-ops, and try making 765 // a precise alias query. 766 const Value *S = StripPointerCastsAndObjCCalls(Loc.Ptr); 767 if (AliasAnalysis::pointsToConstantMemory(Location(S, Loc.Size, Loc.TBAATag), 768 OrLocal)) 769 return true; 770 771 // If that failed, climb to the underlying object, including climbing through 772 // ObjC-specific no-ops, and try making an imprecise alias query. 773 const Value *U = GetUnderlyingObjCPtr(S); 774 if (U != S) 775 return AliasAnalysis::pointsToConstantMemory(Location(U), OrLocal); 776 777 // If that failed, fail. We don't need to chain here, since that's covered 778 // by the earlier precise query. 779 return false; 780 } 781 782 AliasAnalysis::ModRefBehavior 783 ObjCARCAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) { 784 // We have nothing to do. Just chain to the next AliasAnalysis. 785 return AliasAnalysis::getModRefBehavior(CS); 786 } 787 788 AliasAnalysis::ModRefBehavior 789 ObjCARCAliasAnalysis::getModRefBehavior(const Function *F) { 790 if (!EnableARCOpts) 791 return AliasAnalysis::getModRefBehavior(F); 792 793 switch (GetFunctionClass(F)) { 794 case IC_NoopCast: 795 return DoesNotAccessMemory; 796 default: 797 break; 798 } 799 800 return AliasAnalysis::getModRefBehavior(F); 801 } 802 803 AliasAnalysis::ModRefResult 804 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS, const Location &Loc) { 805 if (!EnableARCOpts) 806 return AliasAnalysis::getModRefInfo(CS, Loc); 807 808 switch (GetBasicInstructionClass(CS.getInstruction())) { 809 case IC_Retain: 810 case IC_RetainRV: 811 case IC_Autorelease: 812 case IC_AutoreleaseRV: 813 case IC_NoopCast: 814 case IC_AutoreleasepoolPush: 815 case IC_FusedRetainAutorelease: 816 case IC_FusedRetainAutoreleaseRV: 817 // These functions don't access any memory visible to the compiler. 818 // Note that this doesn't include objc_retainBlock, becuase it updates 819 // pointers when it copies block data. 820 return NoModRef; 821 default: 822 break; 823 } 824 825 return AliasAnalysis::getModRefInfo(CS, Loc); 826 } 827 828 AliasAnalysis::ModRefResult 829 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS1, 830 ImmutableCallSite CS2) { 831 // TODO: Theoretically we could check for dependencies between objc_* calls 832 // and OnlyAccessesArgumentPointees calls or other well-behaved calls. 833 return AliasAnalysis::getModRefInfo(CS1, CS2); 834 } 835 836 //===----------------------------------------------------------------------===// 837 // ARC expansion. 838 //===----------------------------------------------------------------------===// 839 840 #include "llvm/Support/InstIterator.h" 841 #include "llvm/Transforms/Scalar.h" 842 843 namespace { 844 /// ObjCARCExpand - Early ARC transformations. 845 class ObjCARCExpand : public FunctionPass { 846 virtual void getAnalysisUsage(AnalysisUsage &AU) const; 847 virtual bool doInitialization(Module &M); 848 virtual bool runOnFunction(Function &F); 849 850 /// Run - A flag indicating whether this optimization pass should run. 851 bool Run; 852 853 public: 854 static char ID; 855 ObjCARCExpand() : FunctionPass(ID) { 856 initializeObjCARCExpandPass(*PassRegistry::getPassRegistry()); 857 } 858 }; 859 } 860 861 char ObjCARCExpand::ID = 0; 862 INITIALIZE_PASS(ObjCARCExpand, 863 "objc-arc-expand", "ObjC ARC expansion", false, false) 864 865 Pass *llvm::createObjCARCExpandPass() { 866 return new ObjCARCExpand(); 867 } 868 869 void ObjCARCExpand::getAnalysisUsage(AnalysisUsage &AU) const { 870 AU.setPreservesCFG(); 871 } 872 873 bool ObjCARCExpand::doInitialization(Module &M) { 874 Run = ModuleHasARC(M); 875 return false; 876 } 877 878 bool ObjCARCExpand::runOnFunction(Function &F) { 879 if (!EnableARCOpts) 880 return false; 881 882 // If nothing in the Module uses ARC, don't do anything. 883 if (!Run) 884 return false; 885 886 bool Changed = false; 887 888 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I) { 889 Instruction *Inst = &*I; 890 891 switch (GetBasicInstructionClass(Inst)) { 892 case IC_Retain: 893 case IC_RetainRV: 894 case IC_Autorelease: 895 case IC_AutoreleaseRV: 896 case IC_FusedRetainAutorelease: 897 case IC_FusedRetainAutoreleaseRV: 898 // These calls return their argument verbatim, as a low-level 899 // optimization. However, this makes high-level optimizations 900 // harder. Undo any uses of this optimization that the front-end 901 // emitted here. We'll redo them in the contract pass. 902 Changed = true; 903 Inst->replaceAllUsesWith(cast<CallInst>(Inst)->getArgOperand(0)); 904 break; 905 default: 906 break; 907 } 908 } 909 910 return Changed; 911 } 912 913 //===----------------------------------------------------------------------===// 914 // ARC autorelease pool elimination. 915 //===----------------------------------------------------------------------===// 916 917 #include "llvm/Constants.h" 918 919 namespace { 920 /// ObjCARCAPElim - Autorelease pool elimination. 921 class ObjCARCAPElim : public ModulePass { 922 virtual void getAnalysisUsage(AnalysisUsage &AU) const; 923 virtual bool runOnModule(Module &M); 924 925 bool MayAutorelease(CallSite CS, unsigned Depth = 0); 926 bool OptimizeBB(BasicBlock *BB); 927 928 public: 929 static char ID; 930 ObjCARCAPElim() : ModulePass(ID) { 931 initializeObjCARCAPElimPass(*PassRegistry::getPassRegistry()); 932 } 933 }; 934 } 935 936 char ObjCARCAPElim::ID = 0; 937 INITIALIZE_PASS(ObjCARCAPElim, 938 "objc-arc-apelim", 939 "ObjC ARC autorelease pool elimination", 940 false, false) 941 942 Pass *llvm::createObjCARCAPElimPass() { 943 return new ObjCARCAPElim(); 944 } 945 946 void ObjCARCAPElim::getAnalysisUsage(AnalysisUsage &AU) const { 947 AU.setPreservesCFG(); 948 } 949 950 /// MayAutorelease - Interprocedurally determine if calls made by the 951 /// given call site can possibly produce autoreleases. 952 bool ObjCARCAPElim::MayAutorelease(CallSite CS, unsigned Depth) { 953 if (Function *Callee = CS.getCalledFunction()) { 954 if (Callee->isDeclaration() || Callee->mayBeOverridden()) 955 return true; 956 for (Function::iterator I = Callee->begin(), E = Callee->end(); 957 I != E; ++I) { 958 BasicBlock *BB = I; 959 for (BasicBlock::iterator J = BB->begin(), F = BB->end(); J != F; ++J) 960 if (CallSite JCS = CallSite(J)) 961 // This recursion depth limit is arbitrary. It's just great 962 // enough to cover known interesting testcases. 963 if (Depth < 3 && 964 !JCS.onlyReadsMemory() && 965 MayAutorelease(JCS, Depth + 1)) 966 return true; 967 } 968 return false; 969 } 970 971 return true; 972 } 973 974 bool ObjCARCAPElim::OptimizeBB(BasicBlock *BB) { 975 bool Changed = false; 976 977 Instruction *Push = 0; 978 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 979 Instruction *Inst = I++; 980 switch (GetBasicInstructionClass(Inst)) { 981 case IC_AutoreleasepoolPush: 982 Push = Inst; 983 break; 984 case IC_AutoreleasepoolPop: 985 // If this pop matches a push and nothing in between can autorelease, 986 // zap the pair. 987 if (Push && cast<CallInst>(Inst)->getArgOperand(0) == Push) { 988 Changed = true; 989 Inst->eraseFromParent(); 990 Push->eraseFromParent(); 991 } 992 Push = 0; 993 break; 994 case IC_CallOrUser: 995 if (MayAutorelease(CallSite(Inst))) 996 Push = 0; 997 break; 998 default: 999 break; 1000 } 1001 } 1002 1003 return Changed; 1004 } 1005 1006 bool ObjCARCAPElim::runOnModule(Module &M) { 1007 if (!EnableARCOpts) 1008 return false; 1009 1010 // If nothing in the Module uses ARC, don't do anything. 1011 if (!ModuleHasARC(M)) 1012 return false; 1013 1014 // Find the llvm.global_ctors variable, as the first step in 1015 // identifying the global constructors. In theory, unnecessary autorelease 1016 // pools could occur anywhere, but in practice it's pretty rare. Global 1017 // ctors are a place where autorelease pools get inserted automatically, 1018 // so it's pretty common for them to be unnecessary, and it's pretty 1019 // profitable to eliminate them. 1020 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors"); 1021 if (!GV) 1022 return false; 1023 1024 assert(GV->hasDefinitiveInitializer() && 1025 "llvm.global_ctors is uncooperative!"); 1026 1027 bool Changed = false; 1028 1029 // Dig the constructor functions out of GV's initializer. 1030 ConstantArray *Init = cast<ConstantArray>(GV->getInitializer()); 1031 for (User::op_iterator OI = Init->op_begin(), OE = Init->op_end(); 1032 OI != OE; ++OI) { 1033 Value *Op = *OI; 1034 // llvm.global_ctors is an array of pairs where the second members 1035 // are constructor functions. 1036 Function *F = dyn_cast<Function>(cast<ConstantStruct>(Op)->getOperand(1)); 1037 // If the user used a constructor function with the wrong signature and 1038 // it got bitcasted or whatever, look the other way. 1039 if (!F) 1040 continue; 1041 // Only look at function definitions. 1042 if (F->isDeclaration()) 1043 continue; 1044 // Only look at functions with one basic block. 1045 if (llvm::next(F->begin()) != F->end()) 1046 continue; 1047 // Ok, a single-block constructor function definition. Try to optimize it. 1048 Changed |= OptimizeBB(F->begin()); 1049 } 1050 1051 return Changed; 1052 } 1053 1054 //===----------------------------------------------------------------------===// 1055 // ARC optimization. 1056 //===----------------------------------------------------------------------===// 1057 1058 // TODO: On code like this: 1059 // 1060 // objc_retain(%x) 1061 // stuff_that_cannot_release() 1062 // objc_autorelease(%x) 1063 // stuff_that_cannot_release() 1064 // objc_retain(%x) 1065 // stuff_that_cannot_release() 1066 // objc_autorelease(%x) 1067 // 1068 // The second retain and autorelease can be deleted. 1069 1070 // TODO: It should be possible to delete 1071 // objc_autoreleasePoolPush and objc_autoreleasePoolPop 1072 // pairs if nothing is actually autoreleased between them. Also, autorelease 1073 // calls followed by objc_autoreleasePoolPop calls (perhaps in ObjC++ code 1074 // after inlining) can be turned into plain release calls. 1075 1076 // TODO: Critical-edge splitting. If the optimial insertion point is 1077 // a critical edge, the current algorithm has to fail, because it doesn't 1078 // know how to split edges. It should be possible to make the optimizer 1079 // think in terms of edges, rather than blocks, and then split critical 1080 // edges on demand. 1081 1082 // TODO: OptimizeSequences could generalized to be Interprocedural. 1083 1084 // TODO: Recognize that a bunch of other objc runtime calls have 1085 // non-escaping arguments and non-releasing arguments, and may be 1086 // non-autoreleasing. 1087 1088 // TODO: Sink autorelease calls as far as possible. Unfortunately we 1089 // usually can't sink them past other calls, which would be the main 1090 // case where it would be useful. 1091 1092 // TODO: The pointer returned from objc_loadWeakRetained is retained. 1093 1094 // TODO: Delete release+retain pairs (rare). 1095 1096 #include "llvm/GlobalAlias.h" 1097 #include "llvm/Constants.h" 1098 #include "llvm/LLVMContext.h" 1099 #include "llvm/Support/ErrorHandling.h" 1100 #include "llvm/Support/CFG.h" 1101 #include "llvm/ADT/Statistic.h" 1102 #include "llvm/ADT/SmallPtrSet.h" 1103 #include "llvm/ADT/DenseSet.h" 1104 1105 STATISTIC(NumNoops, "Number of no-op objc calls eliminated"); 1106 STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated"); 1107 STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases"); 1108 STATISTIC(NumRets, "Number of return value forwarding " 1109 "retain+autoreleaes eliminated"); 1110 STATISTIC(NumRRs, "Number of retain+release paths eliminated"); 1111 STATISTIC(NumPeeps, "Number of calls peephole-optimized"); 1112 1113 namespace { 1114 /// ProvenanceAnalysis - This is similar to BasicAliasAnalysis, and it 1115 /// uses many of the same techniques, except it uses special ObjC-specific 1116 /// reasoning about pointer relationships. 1117 class ProvenanceAnalysis { 1118 AliasAnalysis *AA; 1119 1120 typedef std::pair<const Value *, const Value *> ValuePairTy; 1121 typedef DenseMap<ValuePairTy, bool> CachedResultsTy; 1122 CachedResultsTy CachedResults; 1123 1124 bool relatedCheck(const Value *A, const Value *B); 1125 bool relatedSelect(const SelectInst *A, const Value *B); 1126 bool relatedPHI(const PHINode *A, const Value *B); 1127 1128 // Do not implement. 1129 void operator=(const ProvenanceAnalysis &); 1130 ProvenanceAnalysis(const ProvenanceAnalysis &); 1131 1132 public: 1133 ProvenanceAnalysis() {} 1134 1135 void setAA(AliasAnalysis *aa) { AA = aa; } 1136 1137 AliasAnalysis *getAA() const { return AA; } 1138 1139 bool related(const Value *A, const Value *B); 1140 1141 void clear() { 1142 CachedResults.clear(); 1143 } 1144 }; 1145 } 1146 1147 bool ProvenanceAnalysis::relatedSelect(const SelectInst *A, const Value *B) { 1148 // If the values are Selects with the same condition, we can do a more precise 1149 // check: just check for relations between the values on corresponding arms. 1150 if (const SelectInst *SB = dyn_cast<SelectInst>(B)) 1151 if (A->getCondition() == SB->getCondition()) { 1152 if (related(A->getTrueValue(), SB->getTrueValue())) 1153 return true; 1154 if (related(A->getFalseValue(), SB->getFalseValue())) 1155 return true; 1156 return false; 1157 } 1158 1159 // Check both arms of the Select node individually. 1160 if (related(A->getTrueValue(), B)) 1161 return true; 1162 if (related(A->getFalseValue(), B)) 1163 return true; 1164 1165 // The arms both checked out. 1166 return false; 1167 } 1168 1169 bool ProvenanceAnalysis::relatedPHI(const PHINode *A, const Value *B) { 1170 // If the values are PHIs in the same block, we can do a more precise as well 1171 // as efficient check: just check for relations between the values on 1172 // corresponding edges. 1173 if (const PHINode *PNB = dyn_cast<PHINode>(B)) 1174 if (PNB->getParent() == A->getParent()) { 1175 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i) 1176 if (related(A->getIncomingValue(i), 1177 PNB->getIncomingValueForBlock(A->getIncomingBlock(i)))) 1178 return true; 1179 return false; 1180 } 1181 1182 // Check each unique source of the PHI node against B. 1183 SmallPtrSet<const Value *, 4> UniqueSrc; 1184 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i) { 1185 const Value *PV1 = A->getIncomingValue(i); 1186 if (UniqueSrc.insert(PV1) && related(PV1, B)) 1187 return true; 1188 } 1189 1190 // All of the arms checked out. 1191 return false; 1192 } 1193 1194 /// isStoredObjCPointer - Test if the value of P, or any value covered by its 1195 /// provenance, is ever stored within the function (not counting callees). 1196 static bool isStoredObjCPointer(const Value *P) { 1197 SmallPtrSet<const Value *, 8> Visited; 1198 SmallVector<const Value *, 8> Worklist; 1199 Worklist.push_back(P); 1200 Visited.insert(P); 1201 do { 1202 P = Worklist.pop_back_val(); 1203 for (Value::const_use_iterator UI = P->use_begin(), UE = P->use_end(); 1204 UI != UE; ++UI) { 1205 const User *Ur = *UI; 1206 if (isa<StoreInst>(Ur)) { 1207 if (UI.getOperandNo() == 0) 1208 // The pointer is stored. 1209 return true; 1210 // The pointed is stored through. 1211 continue; 1212 } 1213 if (isa<CallInst>(Ur)) 1214 // The pointer is passed as an argument, ignore this. 1215 continue; 1216 if (isa<PtrToIntInst>(P)) 1217 // Assume the worst. 1218 return true; 1219 if (Visited.insert(Ur)) 1220 Worklist.push_back(Ur); 1221 } 1222 } while (!Worklist.empty()); 1223 1224 // Everything checked out. 1225 return false; 1226 } 1227 1228 bool ProvenanceAnalysis::relatedCheck(const Value *A, const Value *B) { 1229 // Skip past provenance pass-throughs. 1230 A = GetUnderlyingObjCPtr(A); 1231 B = GetUnderlyingObjCPtr(B); 1232 1233 // Quick check. 1234 if (A == B) 1235 return true; 1236 1237 // Ask regular AliasAnalysis, for a first approximation. 1238 switch (AA->alias(A, B)) { 1239 case AliasAnalysis::NoAlias: 1240 return false; 1241 case AliasAnalysis::MustAlias: 1242 case AliasAnalysis::PartialAlias: 1243 return true; 1244 case AliasAnalysis::MayAlias: 1245 break; 1246 } 1247 1248 bool AIsIdentified = IsObjCIdentifiedObject(A); 1249 bool BIsIdentified = IsObjCIdentifiedObject(B); 1250 1251 // An ObjC-Identified object can't alias a load if it is never locally stored. 1252 if (AIsIdentified) { 1253 if (BIsIdentified) { 1254 // If both pointers have provenance, they can be directly compared. 1255 if (A != B) 1256 return false; 1257 } else { 1258 if (isa<LoadInst>(B)) 1259 return isStoredObjCPointer(A); 1260 } 1261 } else { 1262 if (BIsIdentified && isa<LoadInst>(A)) 1263 return isStoredObjCPointer(B); 1264 } 1265 1266 // Special handling for PHI and Select. 1267 if (const PHINode *PN = dyn_cast<PHINode>(A)) 1268 return relatedPHI(PN, B); 1269 if (const PHINode *PN = dyn_cast<PHINode>(B)) 1270 return relatedPHI(PN, A); 1271 if (const SelectInst *S = dyn_cast<SelectInst>(A)) 1272 return relatedSelect(S, B); 1273 if (const SelectInst *S = dyn_cast<SelectInst>(B)) 1274 return relatedSelect(S, A); 1275 1276 // Conservative. 1277 return true; 1278 } 1279 1280 bool ProvenanceAnalysis::related(const Value *A, const Value *B) { 1281 // Begin by inserting a conservative value into the map. If the insertion 1282 // fails, we have the answer already. If it succeeds, leave it there until we 1283 // compute the real answer to guard against recursive queries. 1284 if (A > B) std::swap(A, B); 1285 std::pair<CachedResultsTy::iterator, bool> Pair = 1286 CachedResults.insert(std::make_pair(ValuePairTy(A, B), true)); 1287 if (!Pair.second) 1288 return Pair.first->second; 1289 1290 bool Result = relatedCheck(A, B); 1291 CachedResults[ValuePairTy(A, B)] = Result; 1292 return Result; 1293 } 1294 1295 namespace { 1296 // Sequence - A sequence of states that a pointer may go through in which an 1297 // objc_retain and objc_release are actually needed. 1298 enum Sequence { 1299 S_None, 1300 S_Retain, ///< objc_retain(x) 1301 S_CanRelease, ///< foo(x) -- x could possibly see a ref count decrement 1302 S_Use, ///< any use of x 1303 S_Stop, ///< like S_Release, but code motion is stopped 1304 S_Release, ///< objc_release(x) 1305 S_MovableRelease ///< objc_release(x), !clang.imprecise_release 1306 }; 1307 } 1308 1309 static Sequence MergeSeqs(Sequence A, Sequence B, bool TopDown) { 1310 // The easy cases. 1311 if (A == B) 1312 return A; 1313 if (A == S_None || B == S_None) 1314 return S_None; 1315 1316 if (A > B) std::swap(A, B); 1317 if (TopDown) { 1318 // Choose the side which is further along in the sequence. 1319 if ((A == S_Retain || A == S_CanRelease) && 1320 (B == S_CanRelease || B == S_Use)) 1321 return B; 1322 } else { 1323 // Choose the side which is further along in the sequence. 1324 if ((A == S_Use || A == S_CanRelease) && 1325 (B == S_Use || B == S_Release || B == S_Stop || B == S_MovableRelease)) 1326 return A; 1327 // If both sides are releases, choose the more conservative one. 1328 if (A == S_Stop && (B == S_Release || B == S_MovableRelease)) 1329 return A; 1330 if (A == S_Release && B == S_MovableRelease) 1331 return A; 1332 } 1333 1334 return S_None; 1335 } 1336 1337 namespace { 1338 /// RRInfo - Unidirectional information about either a 1339 /// retain-decrement-use-release sequence or release-use-decrement-retain 1340 /// reverese sequence. 1341 struct RRInfo { 1342 /// KnownSafe - After an objc_retain, the reference count of the referenced 1343 /// object is known to be positive. Similarly, before an objc_release, the 1344 /// reference count of the referenced object is known to be positive. If 1345 /// there are retain-release pairs in code regions where the retain count 1346 /// is known to be positive, they can be eliminated, regardless of any side 1347 /// effects between them. 1348 /// 1349 /// Also, a retain+release pair nested within another retain+release 1350 /// pair all on the known same pointer value can be eliminated, regardless 1351 /// of any intervening side effects. 1352 /// 1353 /// KnownSafe is true when either of these conditions is satisfied. 1354 bool KnownSafe; 1355 1356 /// IsRetainBlock - True if the Calls are objc_retainBlock calls (as 1357 /// opposed to objc_retain calls). 1358 bool IsRetainBlock; 1359 1360 /// IsTailCallRelease - True of the objc_release calls are all marked 1361 /// with the "tail" keyword. 1362 bool IsTailCallRelease; 1363 1364 /// Partial - True of we've seen an opportunity for partial RR elimination, 1365 /// such as pushing calls into a CFG triangle or into one side of a 1366 /// CFG diamond. 1367 /// TODO: Consider moving this to PtrState. 1368 bool Partial; 1369 1370 /// ReleaseMetadata - If the Calls are objc_release calls and they all have 1371 /// a clang.imprecise_release tag, this is the metadata tag. 1372 MDNode *ReleaseMetadata; 1373 1374 /// Calls - For a top-down sequence, the set of objc_retains or 1375 /// objc_retainBlocks. For bottom-up, the set of objc_releases. 1376 SmallPtrSet<Instruction *, 2> Calls; 1377 1378 /// ReverseInsertPts - The set of optimal insert positions for 1379 /// moving calls in the opposite sequence. 1380 SmallPtrSet<Instruction *, 2> ReverseInsertPts; 1381 1382 RRInfo() : 1383 KnownSafe(false), IsRetainBlock(false), 1384 IsTailCallRelease(false), Partial(false), 1385 ReleaseMetadata(0) {} 1386 1387 void clear(); 1388 }; 1389 } 1390 1391 void RRInfo::clear() { 1392 KnownSafe = false; 1393 IsRetainBlock = false; 1394 IsTailCallRelease = false; 1395 Partial = false; 1396 ReleaseMetadata = 0; 1397 Calls.clear(); 1398 ReverseInsertPts.clear(); 1399 } 1400 1401 namespace { 1402 /// PtrState - This class summarizes several per-pointer runtime properties 1403 /// which are propogated through the flow graph. 1404 class PtrState { 1405 /// RefCount - The known minimum number of reference count increments. 1406 unsigned RefCount; 1407 1408 /// NestCount - The known minimum level of retain+release nesting. 1409 unsigned NestCount; 1410 1411 /// Seq - The current position in the sequence. 1412 Sequence Seq; 1413 1414 public: 1415 /// RRI - Unidirectional information about the current sequence. 1416 /// TODO: Encapsulate this better. 1417 RRInfo RRI; 1418 1419 PtrState() : RefCount(0), NestCount(0), Seq(S_None) {} 1420 1421 void SetAtLeastOneRefCount() { 1422 if (RefCount == 0) RefCount = 1; 1423 } 1424 1425 void IncrementRefCount() { 1426 if (RefCount != UINT_MAX) ++RefCount; 1427 } 1428 1429 void DecrementRefCount() { 1430 if (RefCount != 0) --RefCount; 1431 } 1432 1433 bool IsKnownIncremented() const { 1434 return RefCount > 0; 1435 } 1436 1437 void IncrementNestCount() { 1438 if (NestCount != UINT_MAX) ++NestCount; 1439 } 1440 1441 void DecrementNestCount() { 1442 if (NestCount != 0) --NestCount; 1443 } 1444 1445 bool IsKnownNested() const { 1446 return NestCount > 0; 1447 } 1448 1449 void SetSeq(Sequence NewSeq) { 1450 Seq = NewSeq; 1451 } 1452 1453 Sequence GetSeq() const { 1454 return Seq; 1455 } 1456 1457 void ClearSequenceProgress() { 1458 Seq = S_None; 1459 RRI.clear(); 1460 } 1461 1462 void Merge(const PtrState &Other, bool TopDown); 1463 }; 1464 } 1465 1466 void 1467 PtrState::Merge(const PtrState &Other, bool TopDown) { 1468 Seq = MergeSeqs(Seq, Other.Seq, TopDown); 1469 RefCount = std::min(RefCount, Other.RefCount); 1470 NestCount = std::min(NestCount, Other.NestCount); 1471 1472 // We can't merge a plain objc_retain with an objc_retainBlock. 1473 if (RRI.IsRetainBlock != Other.RRI.IsRetainBlock) 1474 Seq = S_None; 1475 1476 // If we're not in a sequence (anymore), drop all associated state. 1477 if (Seq == S_None) { 1478 RRI.clear(); 1479 } else if (RRI.Partial || Other.RRI.Partial) { 1480 // If we're doing a merge on a path that's previously seen a partial 1481 // merge, conservatively drop the sequence, to avoid doing partial 1482 // RR elimination. If the branch predicates for the two merge differ, 1483 // mixing them is unsafe. 1484 Seq = S_None; 1485 RRI.clear(); 1486 } else { 1487 // Conservatively merge the ReleaseMetadata information. 1488 if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata) 1489 RRI.ReleaseMetadata = 0; 1490 1491 RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe; 1492 RRI.IsTailCallRelease = RRI.IsTailCallRelease && Other.RRI.IsTailCallRelease; 1493 RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end()); 1494 1495 // Merge the insert point sets. If there are any differences, 1496 // that makes this a partial merge. 1497 RRI.Partial = RRI.ReverseInsertPts.size() != 1498 Other.RRI.ReverseInsertPts.size(); 1499 for (SmallPtrSet<Instruction *, 2>::const_iterator 1500 I = Other.RRI.ReverseInsertPts.begin(), 1501 E = Other.RRI.ReverseInsertPts.end(); I != E; ++I) 1502 RRI.Partial |= RRI.ReverseInsertPts.insert(*I); 1503 } 1504 } 1505 1506 namespace { 1507 /// BBState - Per-BasicBlock state. 1508 class BBState { 1509 /// TopDownPathCount - The number of unique control paths from the entry 1510 /// which can reach this block. 1511 unsigned TopDownPathCount; 1512 1513 /// BottomUpPathCount - The number of unique control paths to exits 1514 /// from this block. 1515 unsigned BottomUpPathCount; 1516 1517 /// MapTy - A type for PerPtrTopDown and PerPtrBottomUp. 1518 typedef MapVector<const Value *, PtrState> MapTy; 1519 1520 /// PerPtrTopDown - The top-down traversal uses this to record information 1521 /// known about a pointer at the bottom of each block. 1522 MapTy PerPtrTopDown; 1523 1524 /// PerPtrBottomUp - The bottom-up traversal uses this to record information 1525 /// known about a pointer at the top of each block. 1526 MapTy PerPtrBottomUp; 1527 1528 public: 1529 BBState() : TopDownPathCount(0), BottomUpPathCount(0) {} 1530 1531 typedef MapTy::iterator ptr_iterator; 1532 typedef MapTy::const_iterator ptr_const_iterator; 1533 1534 ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); } 1535 ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); } 1536 ptr_const_iterator top_down_ptr_begin() const { 1537 return PerPtrTopDown.begin(); 1538 } 1539 ptr_const_iterator top_down_ptr_end() const { 1540 return PerPtrTopDown.end(); 1541 } 1542 1543 ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); } 1544 ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); } 1545 ptr_const_iterator bottom_up_ptr_begin() const { 1546 return PerPtrBottomUp.begin(); 1547 } 1548 ptr_const_iterator bottom_up_ptr_end() const { 1549 return PerPtrBottomUp.end(); 1550 } 1551 1552 /// SetAsEntry - Mark this block as being an entry block, which has one 1553 /// path from the entry by definition. 1554 void SetAsEntry() { TopDownPathCount = 1; } 1555 1556 /// SetAsExit - Mark this block as being an exit block, which has one 1557 /// path to an exit by definition. 1558 void SetAsExit() { BottomUpPathCount = 1; } 1559 1560 PtrState &getPtrTopDownState(const Value *Arg) { 1561 return PerPtrTopDown[Arg]; 1562 } 1563 1564 PtrState &getPtrBottomUpState(const Value *Arg) { 1565 return PerPtrBottomUp[Arg]; 1566 } 1567 1568 void clearBottomUpPointers() { 1569 PerPtrBottomUp.clear(); 1570 } 1571 1572 void clearTopDownPointers() { 1573 PerPtrTopDown.clear(); 1574 } 1575 1576 void InitFromPred(const BBState &Other); 1577 void InitFromSucc(const BBState &Other); 1578 void MergePred(const BBState &Other); 1579 void MergeSucc(const BBState &Other); 1580 1581 /// GetAllPathCount - Return the number of possible unique paths from an 1582 /// entry to an exit which pass through this block. This is only valid 1583 /// after both the top-down and bottom-up traversals are complete. 1584 unsigned GetAllPathCount() const { 1585 return TopDownPathCount * BottomUpPathCount; 1586 } 1587 1588 /// IsVisitedTopDown - Test whether the block for this BBState has been 1589 /// visited by the top-down portion of the algorithm. 1590 bool isVisitedTopDown() const { 1591 return TopDownPathCount != 0; 1592 } 1593 }; 1594 } 1595 1596 void BBState::InitFromPred(const BBState &Other) { 1597 PerPtrTopDown = Other.PerPtrTopDown; 1598 TopDownPathCount = Other.TopDownPathCount; 1599 } 1600 1601 void BBState::InitFromSucc(const BBState &Other) { 1602 PerPtrBottomUp = Other.PerPtrBottomUp; 1603 BottomUpPathCount = Other.BottomUpPathCount; 1604 } 1605 1606 /// MergePred - The top-down traversal uses this to merge information about 1607 /// predecessors to form the initial state for a new block. 1608 void BBState::MergePred(const BBState &Other) { 1609 // Other.TopDownPathCount can be 0, in which case it is either dead or a 1610 // loop backedge. Loop backedges are special. 1611 TopDownPathCount += Other.TopDownPathCount; 1612 1613 // For each entry in the other set, if our set has an entry with the same key, 1614 // merge the entries. Otherwise, copy the entry and merge it with an empty 1615 // entry. 1616 for (ptr_const_iterator MI = Other.top_down_ptr_begin(), 1617 ME = Other.top_down_ptr_end(); MI != ME; ++MI) { 1618 std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI); 1619 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second, 1620 /*TopDown=*/true); 1621 } 1622 1623 // For each entry in our set, if the other set doesn't have an entry with the 1624 // same key, force it to merge with an empty entry. 1625 for (ptr_iterator MI = top_down_ptr_begin(), 1626 ME = top_down_ptr_end(); MI != ME; ++MI) 1627 if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end()) 1628 MI->second.Merge(PtrState(), /*TopDown=*/true); 1629 } 1630 1631 /// MergeSucc - The bottom-up traversal uses this to merge information about 1632 /// successors to form the initial state for a new block. 1633 void BBState::MergeSucc(const BBState &Other) { 1634 // Other.BottomUpPathCount can be 0, in which case it is either dead or a 1635 // loop backedge. Loop backedges are special. 1636 BottomUpPathCount += Other.BottomUpPathCount; 1637 1638 // For each entry in the other set, if our set has an entry with the 1639 // same key, merge the entries. Otherwise, copy the entry and merge 1640 // it with an empty entry. 1641 for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(), 1642 ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) { 1643 std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI); 1644 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second, 1645 /*TopDown=*/false); 1646 } 1647 1648 // For each entry in our set, if the other set doesn't have an entry 1649 // with the same key, force it to merge with an empty entry. 1650 for (ptr_iterator MI = bottom_up_ptr_begin(), 1651 ME = bottom_up_ptr_end(); MI != ME; ++MI) 1652 if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end()) 1653 MI->second.Merge(PtrState(), /*TopDown=*/false); 1654 } 1655 1656 namespace { 1657 /// ObjCARCOpt - The main ARC optimization pass. 1658 class ObjCARCOpt : public FunctionPass { 1659 bool Changed; 1660 ProvenanceAnalysis PA; 1661 1662 /// Run - A flag indicating whether this optimization pass should run. 1663 bool Run; 1664 1665 /// RetainRVCallee, etc. - Declarations for ObjC runtime 1666 /// functions, for use in creating calls to them. These are initialized 1667 /// lazily to avoid cluttering up the Module with unused declarations. 1668 Constant *RetainRVCallee, *AutoreleaseRVCallee, *ReleaseCallee, 1669 *RetainCallee, *RetainBlockCallee, *AutoreleaseCallee; 1670 1671 /// UsedInThisFunciton - Flags which determine whether each of the 1672 /// interesting runtine functions is in fact used in the current function. 1673 unsigned UsedInThisFunction; 1674 1675 /// ImpreciseReleaseMDKind - The Metadata Kind for clang.imprecise_release 1676 /// metadata. 1677 unsigned ImpreciseReleaseMDKind; 1678 1679 /// CopyOnEscapeMDKind - The Metadata Kind for clang.arc.copy_on_escape 1680 /// metadata. 1681 unsigned CopyOnEscapeMDKind; 1682 1683 /// NoObjCARCExceptionsMDKind - The Metadata Kind for 1684 /// clang.arc.no_objc_arc_exceptions metadata. 1685 unsigned NoObjCARCExceptionsMDKind; 1686 1687 Constant *getRetainRVCallee(Module *M); 1688 Constant *getAutoreleaseRVCallee(Module *M); 1689 Constant *getReleaseCallee(Module *M); 1690 Constant *getRetainCallee(Module *M); 1691 Constant *getRetainBlockCallee(Module *M); 1692 Constant *getAutoreleaseCallee(Module *M); 1693 1694 bool IsRetainBlockOptimizable(const Instruction *Inst); 1695 1696 void OptimizeRetainCall(Function &F, Instruction *Retain); 1697 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV); 1698 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV); 1699 void OptimizeIndividualCalls(Function &F); 1700 1701 void CheckForCFGHazards(const BasicBlock *BB, 1702 DenseMap<const BasicBlock *, BBState> &BBStates, 1703 BBState &MyStates) const; 1704 bool VisitInstructionBottomUp(Instruction *Inst, 1705 BasicBlock *BB, 1706 MapVector<Value *, RRInfo> &Retains, 1707 BBState &MyStates); 1708 bool VisitBottomUp(BasicBlock *BB, 1709 DenseMap<const BasicBlock *, BBState> &BBStates, 1710 MapVector<Value *, RRInfo> &Retains); 1711 bool VisitInstructionTopDown(Instruction *Inst, 1712 DenseMap<Value *, RRInfo> &Releases, 1713 BBState &MyStates); 1714 bool VisitTopDown(BasicBlock *BB, 1715 DenseMap<const BasicBlock *, BBState> &BBStates, 1716 DenseMap<Value *, RRInfo> &Releases); 1717 bool Visit(Function &F, 1718 DenseMap<const BasicBlock *, BBState> &BBStates, 1719 MapVector<Value *, RRInfo> &Retains, 1720 DenseMap<Value *, RRInfo> &Releases); 1721 1722 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove, 1723 MapVector<Value *, RRInfo> &Retains, 1724 DenseMap<Value *, RRInfo> &Releases, 1725 SmallVectorImpl<Instruction *> &DeadInsts, 1726 Module *M); 1727 1728 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates, 1729 MapVector<Value *, RRInfo> &Retains, 1730 DenseMap<Value *, RRInfo> &Releases, 1731 Module *M); 1732 1733 void OptimizeWeakCalls(Function &F); 1734 1735 bool OptimizeSequences(Function &F); 1736 1737 void OptimizeReturns(Function &F); 1738 1739 virtual void getAnalysisUsage(AnalysisUsage &AU) const; 1740 virtual bool doInitialization(Module &M); 1741 virtual bool runOnFunction(Function &F); 1742 virtual void releaseMemory(); 1743 1744 public: 1745 static char ID; 1746 ObjCARCOpt() : FunctionPass(ID) { 1747 initializeObjCARCOptPass(*PassRegistry::getPassRegistry()); 1748 } 1749 }; 1750 } 1751 1752 char ObjCARCOpt::ID = 0; 1753 INITIALIZE_PASS_BEGIN(ObjCARCOpt, 1754 "objc-arc", "ObjC ARC optimization", false, false) 1755 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis) 1756 INITIALIZE_PASS_END(ObjCARCOpt, 1757 "objc-arc", "ObjC ARC optimization", false, false) 1758 1759 Pass *llvm::createObjCARCOptPass() { 1760 return new ObjCARCOpt(); 1761 } 1762 1763 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const { 1764 AU.addRequired<ObjCARCAliasAnalysis>(); 1765 AU.addRequired<AliasAnalysis>(); 1766 // ARC optimization doesn't currently split critical edges. 1767 AU.setPreservesCFG(); 1768 } 1769 1770 bool ObjCARCOpt::IsRetainBlockOptimizable(const Instruction *Inst) { 1771 // Without the magic metadata tag, we have to assume this might be an 1772 // objc_retainBlock call inserted to convert a block pointer to an id, 1773 // in which case it really is needed. 1774 if (!Inst->getMetadata(CopyOnEscapeMDKind)) 1775 return false; 1776 1777 // If the pointer "escapes" (not including being used in a call), 1778 // the copy may be needed. 1779 if (DoesObjCBlockEscape(Inst)) 1780 return false; 1781 1782 // Otherwise, it's not needed. 1783 return true; 1784 } 1785 1786 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) { 1787 if (!RetainRVCallee) { 1788 LLVMContext &C = M->getContext(); 1789 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); 1790 std::vector<Type *> Params; 1791 Params.push_back(I8X); 1792 FunctionType *FTy = 1793 FunctionType::get(I8X, Params, /*isVarArg=*/false); 1794 AttrListPtr Attributes; 1795 Attributes.addAttr(~0u, Attribute::NoUnwind); 1796 RetainRVCallee = 1797 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy, 1798 Attributes); 1799 } 1800 return RetainRVCallee; 1801 } 1802 1803 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) { 1804 if (!AutoreleaseRVCallee) { 1805 LLVMContext &C = M->getContext(); 1806 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); 1807 std::vector<Type *> Params; 1808 Params.push_back(I8X); 1809 FunctionType *FTy = 1810 FunctionType::get(I8X, Params, /*isVarArg=*/false); 1811 AttrListPtr Attributes; 1812 Attributes.addAttr(~0u, Attribute::NoUnwind); 1813 AutoreleaseRVCallee = 1814 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy, 1815 Attributes); 1816 } 1817 return AutoreleaseRVCallee; 1818 } 1819 1820 Constant *ObjCARCOpt::getReleaseCallee(Module *M) { 1821 if (!ReleaseCallee) { 1822 LLVMContext &C = M->getContext(); 1823 std::vector<Type *> Params; 1824 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C))); 1825 AttrListPtr Attributes; 1826 Attributes.addAttr(~0u, Attribute::NoUnwind); 1827 ReleaseCallee = 1828 M->getOrInsertFunction( 1829 "objc_release", 1830 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false), 1831 Attributes); 1832 } 1833 return ReleaseCallee; 1834 } 1835 1836 Constant *ObjCARCOpt::getRetainCallee(Module *M) { 1837 if (!RetainCallee) { 1838 LLVMContext &C = M->getContext(); 1839 std::vector<Type *> Params; 1840 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C))); 1841 AttrListPtr Attributes; 1842 Attributes.addAttr(~0u, Attribute::NoUnwind); 1843 RetainCallee = 1844 M->getOrInsertFunction( 1845 "objc_retain", 1846 FunctionType::get(Params[0], Params, /*isVarArg=*/false), 1847 Attributes); 1848 } 1849 return RetainCallee; 1850 } 1851 1852 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) { 1853 if (!RetainBlockCallee) { 1854 LLVMContext &C = M->getContext(); 1855 std::vector<Type *> Params; 1856 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C))); 1857 AttrListPtr Attributes; 1858 // objc_retainBlock is not nounwind because it calls user copy constructors 1859 // which could theoretically throw. 1860 RetainBlockCallee = 1861 M->getOrInsertFunction( 1862 "objc_retainBlock", 1863 FunctionType::get(Params[0], Params, /*isVarArg=*/false), 1864 Attributes); 1865 } 1866 return RetainBlockCallee; 1867 } 1868 1869 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) { 1870 if (!AutoreleaseCallee) { 1871 LLVMContext &C = M->getContext(); 1872 std::vector<Type *> Params; 1873 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C))); 1874 AttrListPtr Attributes; 1875 Attributes.addAttr(~0u, Attribute::NoUnwind); 1876 AutoreleaseCallee = 1877 M->getOrInsertFunction( 1878 "objc_autorelease", 1879 FunctionType::get(Params[0], Params, /*isVarArg=*/false), 1880 Attributes); 1881 } 1882 return AutoreleaseCallee; 1883 } 1884 1885 /// CanAlterRefCount - Test whether the given instruction can result in a 1886 /// reference count modification (positive or negative) for the pointer's 1887 /// object. 1888 static bool 1889 CanAlterRefCount(const Instruction *Inst, const Value *Ptr, 1890 ProvenanceAnalysis &PA, InstructionClass Class) { 1891 switch (Class) { 1892 case IC_Autorelease: 1893 case IC_AutoreleaseRV: 1894 case IC_User: 1895 // These operations never directly modify a reference count. 1896 return false; 1897 default: break; 1898 } 1899 1900 ImmutableCallSite CS = static_cast<const Value *>(Inst); 1901 assert(CS && "Only calls can alter reference counts!"); 1902 1903 // See if AliasAnalysis can help us with the call. 1904 AliasAnalysis::ModRefBehavior MRB = PA.getAA()->getModRefBehavior(CS); 1905 if (AliasAnalysis::onlyReadsMemory(MRB)) 1906 return false; 1907 if (AliasAnalysis::onlyAccessesArgPointees(MRB)) { 1908 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1909 I != E; ++I) { 1910 const Value *Op = *I; 1911 if (IsPotentialUse(Op) && PA.related(Ptr, Op)) 1912 return true; 1913 } 1914 return false; 1915 } 1916 1917 // Assume the worst. 1918 return true; 1919 } 1920 1921 /// CanUse - Test whether the given instruction can "use" the given pointer's 1922 /// object in a way that requires the reference count to be positive. 1923 static bool 1924 CanUse(const Instruction *Inst, const Value *Ptr, ProvenanceAnalysis &PA, 1925 InstructionClass Class) { 1926 // IC_Call operations (as opposed to IC_CallOrUser) never "use" objc pointers. 1927 if (Class == IC_Call) 1928 return false; 1929 1930 // Consider various instructions which may have pointer arguments which are 1931 // not "uses". 1932 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(Inst)) { 1933 // Comparing a pointer with null, or any other constant, isn't really a use, 1934 // because we don't care what the pointer points to, or about the values 1935 // of any other dynamic reference-counted pointers. 1936 if (!IsPotentialUse(ICI->getOperand(1))) 1937 return false; 1938 } else if (ImmutableCallSite CS = static_cast<const Value *>(Inst)) { 1939 // For calls, just check the arguments (and not the callee operand). 1940 for (ImmutableCallSite::arg_iterator OI = CS.arg_begin(), 1941 OE = CS.arg_end(); OI != OE; ++OI) { 1942 const Value *Op = *OI; 1943 if (IsPotentialUse(Op) && PA.related(Ptr, Op)) 1944 return true; 1945 } 1946 return false; 1947 } else if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 1948 // Special-case stores, because we don't care about the stored value, just 1949 // the store address. 1950 const Value *Op = GetUnderlyingObjCPtr(SI->getPointerOperand()); 1951 // If we can't tell what the underlying object was, assume there is a 1952 // dependence. 1953 return IsPotentialUse(Op) && PA.related(Op, Ptr); 1954 } 1955 1956 // Check each operand for a match. 1957 for (User::const_op_iterator OI = Inst->op_begin(), OE = Inst->op_end(); 1958 OI != OE; ++OI) { 1959 const Value *Op = *OI; 1960 if (IsPotentialUse(Op) && PA.related(Ptr, Op)) 1961 return true; 1962 } 1963 return false; 1964 } 1965 1966 /// CanInterruptRV - Test whether the given instruction can autorelease 1967 /// any pointer or cause an autoreleasepool pop. 1968 static bool 1969 CanInterruptRV(InstructionClass Class) { 1970 switch (Class) { 1971 case IC_AutoreleasepoolPop: 1972 case IC_CallOrUser: 1973 case IC_Call: 1974 case IC_Autorelease: 1975 case IC_AutoreleaseRV: 1976 case IC_FusedRetainAutorelease: 1977 case IC_FusedRetainAutoreleaseRV: 1978 return true; 1979 default: 1980 return false; 1981 } 1982 } 1983 1984 namespace { 1985 /// DependenceKind - There are several kinds of dependence-like concepts in 1986 /// use here. 1987 enum DependenceKind { 1988 NeedsPositiveRetainCount, 1989 AutoreleasePoolBoundary, 1990 CanChangeRetainCount, 1991 RetainAutoreleaseDep, ///< Blocks objc_retainAutorelease. 1992 RetainAutoreleaseRVDep, ///< Blocks objc_retainAutoreleaseReturnValue. 1993 RetainRVDep ///< Blocks objc_retainAutoreleasedReturnValue. 1994 }; 1995 } 1996 1997 /// Depends - Test if there can be dependencies on Inst through Arg. This 1998 /// function only tests dependencies relevant for removing pairs of calls. 1999 static bool 2000 Depends(DependenceKind Flavor, Instruction *Inst, const Value *Arg, 2001 ProvenanceAnalysis &PA) { 2002 // If we've reached the definition of Arg, stop. 2003 if (Inst == Arg) 2004 return true; 2005 2006 switch (Flavor) { 2007 case NeedsPositiveRetainCount: { 2008 InstructionClass Class = GetInstructionClass(Inst); 2009 switch (Class) { 2010 case IC_AutoreleasepoolPop: 2011 case IC_AutoreleasepoolPush: 2012 case IC_None: 2013 return false; 2014 default: 2015 return CanUse(Inst, Arg, PA, Class); 2016 } 2017 } 2018 2019 case AutoreleasePoolBoundary: { 2020 InstructionClass Class = GetInstructionClass(Inst); 2021 switch (Class) { 2022 case IC_AutoreleasepoolPop: 2023 case IC_AutoreleasepoolPush: 2024 // These mark the end and begin of an autorelease pool scope. 2025 return true; 2026 default: 2027 // Nothing else does this. 2028 return false; 2029 } 2030 } 2031 2032 case CanChangeRetainCount: { 2033 InstructionClass Class = GetInstructionClass(Inst); 2034 switch (Class) { 2035 case IC_AutoreleasepoolPop: 2036 // Conservatively assume this can decrement any count. 2037 return true; 2038 case IC_AutoreleasepoolPush: 2039 case IC_None: 2040 return false; 2041 default: 2042 return CanAlterRefCount(Inst, Arg, PA, Class); 2043 } 2044 } 2045 2046 case RetainAutoreleaseDep: 2047 switch (GetBasicInstructionClass(Inst)) { 2048 case IC_AutoreleasepoolPop: 2049 case IC_AutoreleasepoolPush: 2050 // Don't merge an objc_autorelease with an objc_retain inside a different 2051 // autoreleasepool scope. 2052 return true; 2053 case IC_Retain: 2054 case IC_RetainRV: 2055 // Check for a retain of the same pointer for merging. 2056 return GetObjCArg(Inst) == Arg; 2057 default: 2058 // Nothing else matters for objc_retainAutorelease formation. 2059 return false; 2060 } 2061 2062 case RetainAutoreleaseRVDep: { 2063 InstructionClass Class = GetBasicInstructionClass(Inst); 2064 switch (Class) { 2065 case IC_Retain: 2066 case IC_RetainRV: 2067 // Check for a retain of the same pointer for merging. 2068 return GetObjCArg(Inst) == Arg; 2069 default: 2070 // Anything that can autorelease interrupts 2071 // retainAutoreleaseReturnValue formation. 2072 return CanInterruptRV(Class); 2073 } 2074 } 2075 2076 case RetainRVDep: 2077 return CanInterruptRV(GetBasicInstructionClass(Inst)); 2078 } 2079 2080 llvm_unreachable("Invalid dependence flavor"); 2081 } 2082 2083 /// FindDependencies - Walk up the CFG from StartPos (which is in StartBB) and 2084 /// find local and non-local dependencies on Arg. 2085 /// TODO: Cache results? 2086 static void 2087 FindDependencies(DependenceKind Flavor, 2088 const Value *Arg, 2089 BasicBlock *StartBB, Instruction *StartInst, 2090 SmallPtrSet<Instruction *, 4> &DependingInstructions, 2091 SmallPtrSet<const BasicBlock *, 4> &Visited, 2092 ProvenanceAnalysis &PA) { 2093 BasicBlock::iterator StartPos = StartInst; 2094 2095 SmallVector<std::pair<BasicBlock *, BasicBlock::iterator>, 4> Worklist; 2096 Worklist.push_back(std::make_pair(StartBB, StartPos)); 2097 do { 2098 std::pair<BasicBlock *, BasicBlock::iterator> Pair = 2099 Worklist.pop_back_val(); 2100 BasicBlock *LocalStartBB = Pair.first; 2101 BasicBlock::iterator LocalStartPos = Pair.second; 2102 BasicBlock::iterator StartBBBegin = LocalStartBB->begin(); 2103 for (;;) { 2104 if (LocalStartPos == StartBBBegin) { 2105 pred_iterator PI(LocalStartBB), PE(LocalStartBB, false); 2106 if (PI == PE) 2107 // If we've reached the function entry, produce a null dependence. 2108 DependingInstructions.insert(0); 2109 else 2110 // Add the predecessors to the worklist. 2111 do { 2112 BasicBlock *PredBB = *PI; 2113 if (Visited.insert(PredBB)) 2114 Worklist.push_back(std::make_pair(PredBB, PredBB->end())); 2115 } while (++PI != PE); 2116 break; 2117 } 2118 2119 Instruction *Inst = --LocalStartPos; 2120 if (Depends(Flavor, Inst, Arg, PA)) { 2121 DependingInstructions.insert(Inst); 2122 break; 2123 } 2124 } 2125 } while (!Worklist.empty()); 2126 2127 // Determine whether the original StartBB post-dominates all of the blocks we 2128 // visited. If not, insert a sentinal indicating that most optimizations are 2129 // not safe. 2130 for (SmallPtrSet<const BasicBlock *, 4>::const_iterator I = Visited.begin(), 2131 E = Visited.end(); I != E; ++I) { 2132 const BasicBlock *BB = *I; 2133 if (BB == StartBB) 2134 continue; 2135 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back()); 2136 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) { 2137 const BasicBlock *Succ = *SI; 2138 if (Succ != StartBB && !Visited.count(Succ)) { 2139 DependingInstructions.insert(reinterpret_cast<Instruction *>(-1)); 2140 return; 2141 } 2142 } 2143 } 2144 } 2145 2146 static bool isNullOrUndef(const Value *V) { 2147 return isa<ConstantPointerNull>(V) || isa<UndefValue>(V); 2148 } 2149 2150 static bool isNoopInstruction(const Instruction *I) { 2151 return isa<BitCastInst>(I) || 2152 (isa<GetElementPtrInst>(I) && 2153 cast<GetElementPtrInst>(I)->hasAllZeroIndices()); 2154 } 2155 2156 /// OptimizeRetainCall - Turn objc_retain into 2157 /// objc_retainAutoreleasedReturnValue if the operand is a return value. 2158 void 2159 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) { 2160 CallSite CS(GetObjCArg(Retain)); 2161 Instruction *Call = CS.getInstruction(); 2162 if (!Call) return; 2163 if (Call->getParent() != Retain->getParent()) return; 2164 2165 // Check that the call is next to the retain. 2166 BasicBlock::iterator I = Call; 2167 ++I; 2168 while (isNoopInstruction(I)) ++I; 2169 if (&*I != Retain) 2170 return; 2171 2172 // Turn it to an objc_retainAutoreleasedReturnValue.. 2173 Changed = true; 2174 ++NumPeeps; 2175 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent())); 2176 } 2177 2178 /// OptimizeRetainRVCall - Turn objc_retainAutoreleasedReturnValue into 2179 /// objc_retain if the operand is not a return value. Or, if it can be 2180 /// paired with an objc_autoreleaseReturnValue, delete the pair and 2181 /// return true. 2182 bool 2183 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) { 2184 // Check for the argument being from an immediately preceding call or invoke. 2185 Value *Arg = GetObjCArg(RetainRV); 2186 CallSite CS(Arg); 2187 if (Instruction *Call = CS.getInstruction()) { 2188 if (Call->getParent() == RetainRV->getParent()) { 2189 BasicBlock::iterator I = Call; 2190 ++I; 2191 while (isNoopInstruction(I)) ++I; 2192 if (&*I == RetainRV) 2193 return false; 2194 } else if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 2195 BasicBlock *RetainRVParent = RetainRV->getParent(); 2196 if (II->getNormalDest() == RetainRVParent) { 2197 BasicBlock::iterator I = RetainRVParent->begin(); 2198 while (isNoopInstruction(I)) ++I; 2199 if (&*I == RetainRV) 2200 return false; 2201 } 2202 } 2203 } 2204 2205 // Check for being preceded by an objc_autoreleaseReturnValue on the same 2206 // pointer. In this case, we can delete the pair. 2207 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin(); 2208 if (I != Begin) { 2209 do --I; while (I != Begin && isNoopInstruction(I)); 2210 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV && 2211 GetObjCArg(I) == Arg) { 2212 Changed = true; 2213 ++NumPeeps; 2214 EraseInstruction(I); 2215 EraseInstruction(RetainRV); 2216 return true; 2217 } 2218 } 2219 2220 // Turn it to a plain objc_retain. 2221 Changed = true; 2222 ++NumPeeps; 2223 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent())); 2224 return false; 2225 } 2226 2227 /// OptimizeAutoreleaseRVCall - Turn objc_autoreleaseReturnValue into 2228 /// objc_autorelease if the result is not used as a return value. 2229 void 2230 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV) { 2231 // Check for a return of the pointer value. 2232 const Value *Ptr = GetObjCArg(AutoreleaseRV); 2233 SmallVector<const Value *, 2> Users; 2234 Users.push_back(Ptr); 2235 do { 2236 Ptr = Users.pop_back_val(); 2237 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end(); 2238 UI != UE; ++UI) { 2239 const User *I = *UI; 2240 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV) 2241 return; 2242 if (isa<BitCastInst>(I)) 2243 Users.push_back(I); 2244 } 2245 } while (!Users.empty()); 2246 2247 Changed = true; 2248 ++NumPeeps; 2249 cast<CallInst>(AutoreleaseRV)-> 2250 setCalledFunction(getAutoreleaseCallee(F.getParent())); 2251 } 2252 2253 /// OptimizeIndividualCalls - Visit each call, one at a time, and make 2254 /// simplifications without doing any additional analysis. 2255 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) { 2256 // Reset all the flags in preparation for recomputing them. 2257 UsedInThisFunction = 0; 2258 2259 // Visit all objc_* calls in F. 2260 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) { 2261 Instruction *Inst = &*I++; 2262 InstructionClass Class = GetBasicInstructionClass(Inst); 2263 2264 switch (Class) { 2265 default: break; 2266 2267 // Delete no-op casts. These function calls have special semantics, but 2268 // the semantics are entirely implemented via lowering in the front-end, 2269 // so by the time they reach the optimizer, they are just no-op calls 2270 // which return their argument. 2271 // 2272 // There are gray areas here, as the ability to cast reference-counted 2273 // pointers to raw void* and back allows code to break ARC assumptions, 2274 // however these are currently considered to be unimportant. 2275 case IC_NoopCast: 2276 Changed = true; 2277 ++NumNoops; 2278 EraseInstruction(Inst); 2279 continue; 2280 2281 // If the pointer-to-weak-pointer is null, it's undefined behavior. 2282 case IC_StoreWeak: 2283 case IC_LoadWeak: 2284 case IC_LoadWeakRetained: 2285 case IC_InitWeak: 2286 case IC_DestroyWeak: { 2287 CallInst *CI = cast<CallInst>(Inst); 2288 if (isNullOrUndef(CI->getArgOperand(0))) { 2289 Changed = true; 2290 Type *Ty = CI->getArgOperand(0)->getType(); 2291 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()), 2292 Constant::getNullValue(Ty), 2293 CI); 2294 CI->replaceAllUsesWith(UndefValue::get(CI->getType())); 2295 CI->eraseFromParent(); 2296 continue; 2297 } 2298 break; 2299 } 2300 case IC_CopyWeak: 2301 case IC_MoveWeak: { 2302 CallInst *CI = cast<CallInst>(Inst); 2303 if (isNullOrUndef(CI->getArgOperand(0)) || 2304 isNullOrUndef(CI->getArgOperand(1))) { 2305 Changed = true; 2306 Type *Ty = CI->getArgOperand(0)->getType(); 2307 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()), 2308 Constant::getNullValue(Ty), 2309 CI); 2310 CI->replaceAllUsesWith(UndefValue::get(CI->getType())); 2311 CI->eraseFromParent(); 2312 continue; 2313 } 2314 break; 2315 } 2316 case IC_Retain: 2317 OptimizeRetainCall(F, Inst); 2318 break; 2319 case IC_RetainRV: 2320 if (OptimizeRetainRVCall(F, Inst)) 2321 continue; 2322 break; 2323 case IC_AutoreleaseRV: 2324 OptimizeAutoreleaseRVCall(F, Inst); 2325 break; 2326 } 2327 2328 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused. 2329 if (IsAutorelease(Class) && Inst->use_empty()) { 2330 CallInst *Call = cast<CallInst>(Inst); 2331 const Value *Arg = Call->getArgOperand(0); 2332 Arg = FindSingleUseIdentifiedObject(Arg); 2333 if (Arg) { 2334 Changed = true; 2335 ++NumAutoreleases; 2336 2337 // Create the declaration lazily. 2338 LLVMContext &C = Inst->getContext(); 2339 CallInst *NewCall = 2340 CallInst::Create(getReleaseCallee(F.getParent()), 2341 Call->getArgOperand(0), "", Call); 2342 NewCall->setMetadata(ImpreciseReleaseMDKind, 2343 MDNode::get(C, ArrayRef<Value *>())); 2344 EraseInstruction(Call); 2345 Inst = NewCall; 2346 Class = IC_Release; 2347 } 2348 } 2349 2350 // For functions which can never be passed stack arguments, add 2351 // a tail keyword. 2352 if (IsAlwaysTail(Class)) { 2353 Changed = true; 2354 cast<CallInst>(Inst)->setTailCall(); 2355 } 2356 2357 // Set nounwind as needed. 2358 if (IsNoThrow(Class)) { 2359 Changed = true; 2360 cast<CallInst>(Inst)->setDoesNotThrow(); 2361 } 2362 2363 if (!IsNoopOnNull(Class)) { 2364 UsedInThisFunction |= 1 << Class; 2365 continue; 2366 } 2367 2368 const Value *Arg = GetObjCArg(Inst); 2369 2370 // ARC calls with null are no-ops. Delete them. 2371 if (isNullOrUndef(Arg)) { 2372 Changed = true; 2373 ++NumNoops; 2374 EraseInstruction(Inst); 2375 continue; 2376 } 2377 2378 // Keep track of which of retain, release, autorelease, and retain_block 2379 // are actually present in this function. 2380 UsedInThisFunction |= 1 << Class; 2381 2382 // If Arg is a PHI, and one or more incoming values to the 2383 // PHI are null, and the call is control-equivalent to the PHI, and there 2384 // are no relevant side effects between the PHI and the call, the call 2385 // could be pushed up to just those paths with non-null incoming values. 2386 // For now, don't bother splitting critical edges for this. 2387 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist; 2388 Worklist.push_back(std::make_pair(Inst, Arg)); 2389 do { 2390 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val(); 2391 Inst = Pair.first; 2392 Arg = Pair.second; 2393 2394 const PHINode *PN = dyn_cast<PHINode>(Arg); 2395 if (!PN) continue; 2396 2397 // Determine if the PHI has any null operands, or any incoming 2398 // critical edges. 2399 bool HasNull = false; 2400 bool HasCriticalEdges = false; 2401 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 2402 Value *Incoming = 2403 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i)); 2404 if (isNullOrUndef(Incoming)) 2405 HasNull = true; 2406 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back()) 2407 .getNumSuccessors() != 1) { 2408 HasCriticalEdges = true; 2409 break; 2410 } 2411 } 2412 // If we have null operands and no critical edges, optimize. 2413 if (!HasCriticalEdges && HasNull) { 2414 SmallPtrSet<Instruction *, 4> DependingInstructions; 2415 SmallPtrSet<const BasicBlock *, 4> Visited; 2416 2417 // Check that there is nothing that cares about the reference 2418 // count between the call and the phi. 2419 switch (Class) { 2420 case IC_Retain: 2421 case IC_RetainBlock: 2422 // These can always be moved up. 2423 break; 2424 case IC_Release: 2425 // These can't be moved across things that care about the retain count. 2426 FindDependencies(NeedsPositiveRetainCount, Arg, 2427 Inst->getParent(), Inst, 2428 DependingInstructions, Visited, PA); 2429 break; 2430 case IC_Autorelease: 2431 // These can't be moved across autorelease pool scope boundaries. 2432 FindDependencies(AutoreleasePoolBoundary, Arg, 2433 Inst->getParent(), Inst, 2434 DependingInstructions, Visited, PA); 2435 break; 2436 case IC_RetainRV: 2437 case IC_AutoreleaseRV: 2438 // Don't move these; the RV optimization depends on the autoreleaseRV 2439 // being tail called, and the retainRV being immediately after a call 2440 // (which might still happen if we get lucky with codegen layout, but 2441 // it's not worth taking the chance). 2442 continue; 2443 default: 2444 llvm_unreachable("Invalid dependence flavor"); 2445 } 2446 2447 if (DependingInstructions.size() == 1 && 2448 *DependingInstructions.begin() == PN) { 2449 Changed = true; 2450 ++NumPartialNoops; 2451 // Clone the call into each predecessor that has a non-null value. 2452 CallInst *CInst = cast<CallInst>(Inst); 2453 Type *ParamTy = CInst->getArgOperand(0)->getType(); 2454 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 2455 Value *Incoming = 2456 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i)); 2457 if (!isNullOrUndef(Incoming)) { 2458 CallInst *Clone = cast<CallInst>(CInst->clone()); 2459 Value *Op = PN->getIncomingValue(i); 2460 Instruction *InsertPos = &PN->getIncomingBlock(i)->back(); 2461 if (Op->getType() != ParamTy) 2462 Op = new BitCastInst(Op, ParamTy, "", InsertPos); 2463 Clone->setArgOperand(0, Op); 2464 Clone->insertBefore(InsertPos); 2465 Worklist.push_back(std::make_pair(Clone, Incoming)); 2466 } 2467 } 2468 // Erase the original call. 2469 EraseInstruction(CInst); 2470 continue; 2471 } 2472 } 2473 } while (!Worklist.empty()); 2474 } 2475 } 2476 2477 /// CheckForCFGHazards - Check for critical edges, loop boundaries, irreducible 2478 /// control flow, or other CFG structures where moving code across the edge 2479 /// would result in it being executed more. 2480 void 2481 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB, 2482 DenseMap<const BasicBlock *, BBState> &BBStates, 2483 BBState &MyStates) const { 2484 // If any top-down local-use or possible-dec has a succ which is earlier in 2485 // the sequence, forget it. 2486 for (BBState::ptr_iterator I = MyStates.top_down_ptr_begin(), 2487 E = MyStates.top_down_ptr_end(); I != E; ++I) 2488 switch (I->second.GetSeq()) { 2489 default: break; 2490 case S_Use: { 2491 const Value *Arg = I->first; 2492 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back()); 2493 bool SomeSuccHasSame = false; 2494 bool AllSuccsHaveSame = true; 2495 PtrState &S = I->second; 2496 succ_const_iterator SI(TI), SE(TI, false); 2497 2498 // If the terminator is an invoke marked with the 2499 // clang.arc.no_objc_arc_exceptions metadata, the unwind edge can be 2500 // ignored, for ARC purposes. 2501 if (isa<InvokeInst>(TI) && TI->getMetadata(NoObjCARCExceptionsMDKind)) 2502 --SE; 2503 2504 for (; SI != SE; ++SI) { 2505 Sequence SuccSSeq = S_None; 2506 bool SuccSRRIKnownSafe = false; 2507 // If VisitBottomUp has visited this successor, take what we know about it. 2508 DenseMap<const BasicBlock *, BBState>::iterator BBI = BBStates.find(*SI); 2509 if (BBI != BBStates.end()) { 2510 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg); 2511 SuccSSeq = SuccS.GetSeq(); 2512 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe; 2513 } 2514 switch (SuccSSeq) { 2515 case S_None: 2516 case S_CanRelease: { 2517 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) { 2518 S.ClearSequenceProgress(); 2519 break; 2520 } 2521 continue; 2522 } 2523 case S_Use: 2524 SomeSuccHasSame = true; 2525 break; 2526 case S_Stop: 2527 case S_Release: 2528 case S_MovableRelease: 2529 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) 2530 AllSuccsHaveSame = false; 2531 break; 2532 case S_Retain: 2533 llvm_unreachable("bottom-up pointer in retain state!"); 2534 } 2535 } 2536 // If the state at the other end of any of the successor edges 2537 // matches the current state, require all edges to match. This 2538 // guards against loops in the middle of a sequence. 2539 if (SomeSuccHasSame && !AllSuccsHaveSame) 2540 S.ClearSequenceProgress(); 2541 break; 2542 } 2543 case S_CanRelease: { 2544 const Value *Arg = I->first; 2545 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back()); 2546 bool SomeSuccHasSame = false; 2547 bool AllSuccsHaveSame = true; 2548 PtrState &S = I->second; 2549 succ_const_iterator SI(TI), SE(TI, false); 2550 2551 // If the terminator is an invoke marked with the 2552 // clang.arc.no_objc_arc_exceptions metadata, the unwind edge can be 2553 // ignored, for ARC purposes. 2554 if (isa<InvokeInst>(TI) && TI->getMetadata(NoObjCARCExceptionsMDKind)) 2555 --SE; 2556 2557 for (; SI != SE; ++SI) { 2558 Sequence SuccSSeq = S_None; 2559 bool SuccSRRIKnownSafe = false; 2560 // If VisitBottomUp has visited this successor, take what we know about it. 2561 DenseMap<const BasicBlock *, BBState>::iterator BBI = BBStates.find(*SI); 2562 if (BBI != BBStates.end()) { 2563 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg); 2564 SuccSSeq = SuccS.GetSeq(); 2565 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe; 2566 } 2567 switch (SuccSSeq) { 2568 case S_None: { 2569 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) { 2570 S.ClearSequenceProgress(); 2571 break; 2572 } 2573 continue; 2574 } 2575 case S_CanRelease: 2576 SomeSuccHasSame = true; 2577 break; 2578 case S_Stop: 2579 case S_Release: 2580 case S_MovableRelease: 2581 case S_Use: 2582 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) 2583 AllSuccsHaveSame = false; 2584 break; 2585 case S_Retain: 2586 llvm_unreachable("bottom-up pointer in retain state!"); 2587 } 2588 } 2589 // If the state at the other end of any of the successor edges 2590 // matches the current state, require all edges to match. This 2591 // guards against loops in the middle of a sequence. 2592 if (SomeSuccHasSame && !AllSuccsHaveSame) 2593 S.ClearSequenceProgress(); 2594 break; 2595 } 2596 } 2597 } 2598 2599 bool 2600 ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst, 2601 BasicBlock *BB, 2602 MapVector<Value *, RRInfo> &Retains, 2603 BBState &MyStates) { 2604 bool NestingDetected = false; 2605 InstructionClass Class = GetInstructionClass(Inst); 2606 const Value *Arg = 0; 2607 2608 switch (Class) { 2609 case IC_Release: { 2610 Arg = GetObjCArg(Inst); 2611 2612 PtrState &S = MyStates.getPtrBottomUpState(Arg); 2613 2614 // If we see two releases in a row on the same pointer. If so, make 2615 // a note, and we'll cicle back to revisit it after we've 2616 // hopefully eliminated the second release, which may allow us to 2617 // eliminate the first release too. 2618 // Theoretically we could implement removal of nested retain+release 2619 // pairs by making PtrState hold a stack of states, but this is 2620 // simple and avoids adding overhead for the non-nested case. 2621 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease) 2622 NestingDetected = true; 2623 2624 S.RRI.clear(); 2625 2626 MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind); 2627 S.SetSeq(ReleaseMetadata ? S_MovableRelease : S_Release); 2628 S.RRI.ReleaseMetadata = ReleaseMetadata; 2629 S.RRI.KnownSafe = S.IsKnownNested() || S.IsKnownIncremented(); 2630 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall(); 2631 S.RRI.Calls.insert(Inst); 2632 2633 S.IncrementRefCount(); 2634 S.IncrementNestCount(); 2635 break; 2636 } 2637 case IC_RetainBlock: 2638 // An objc_retainBlock call with just a use may need to be kept, 2639 // because it may be copying a block from the stack to the heap. 2640 if (!IsRetainBlockOptimizable(Inst)) 2641 break; 2642 // FALLTHROUGH 2643 case IC_Retain: 2644 case IC_RetainRV: { 2645 Arg = GetObjCArg(Inst); 2646 2647 PtrState &S = MyStates.getPtrBottomUpState(Arg); 2648 S.DecrementRefCount(); 2649 S.SetAtLeastOneRefCount(); 2650 S.DecrementNestCount(); 2651 2652 switch (S.GetSeq()) { 2653 case S_Stop: 2654 case S_Release: 2655 case S_MovableRelease: 2656 case S_Use: 2657 S.RRI.ReverseInsertPts.clear(); 2658 // FALL THROUGH 2659 case S_CanRelease: 2660 // Don't do retain+release tracking for IC_RetainRV, because it's 2661 // better to let it remain as the first instruction after a call. 2662 if (Class != IC_RetainRV) { 2663 S.RRI.IsRetainBlock = Class == IC_RetainBlock; 2664 Retains[Inst] = S.RRI; 2665 } 2666 S.ClearSequenceProgress(); 2667 break; 2668 case S_None: 2669 break; 2670 case S_Retain: 2671 llvm_unreachable("bottom-up pointer in retain state!"); 2672 } 2673 return NestingDetected; 2674 } 2675 case IC_AutoreleasepoolPop: 2676 // Conservatively, clear MyStates for all known pointers. 2677 MyStates.clearBottomUpPointers(); 2678 return NestingDetected; 2679 case IC_AutoreleasepoolPush: 2680 case IC_None: 2681 // These are irrelevant. 2682 return NestingDetected; 2683 default: 2684 break; 2685 } 2686 2687 // Consider any other possible effects of this instruction on each 2688 // pointer being tracked. 2689 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(), 2690 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) { 2691 const Value *Ptr = MI->first; 2692 if (Ptr == Arg) 2693 continue; // Handled above. 2694 PtrState &S = MI->second; 2695 Sequence Seq = S.GetSeq(); 2696 2697 // Check for possible releases. 2698 if (CanAlterRefCount(Inst, Ptr, PA, Class)) { 2699 S.DecrementRefCount(); 2700 switch (Seq) { 2701 case S_Use: 2702 S.SetSeq(S_CanRelease); 2703 continue; 2704 case S_CanRelease: 2705 case S_Release: 2706 case S_MovableRelease: 2707 case S_Stop: 2708 case S_None: 2709 break; 2710 case S_Retain: 2711 llvm_unreachable("bottom-up pointer in retain state!"); 2712 } 2713 } 2714 2715 // Check for possible direct uses. 2716 switch (Seq) { 2717 case S_Release: 2718 case S_MovableRelease: 2719 if (CanUse(Inst, Ptr, PA, Class)) { 2720 assert(S.RRI.ReverseInsertPts.empty()); 2721 // If this is an invoke instruction, we're scanning it as part of 2722 // one of its successor blocks, since we can't insert code after it 2723 // in its own block, and we don't want to split critical edges. 2724 if (isa<InvokeInst>(Inst)) 2725 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt()); 2726 else 2727 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst))); 2728 S.SetSeq(S_Use); 2729 } else if (Seq == S_Release && 2730 (Class == IC_User || Class == IC_CallOrUser)) { 2731 // Non-movable releases depend on any possible objc pointer use. 2732 S.SetSeq(S_Stop); 2733 assert(S.RRI.ReverseInsertPts.empty()); 2734 // As above; handle invoke specially. 2735 if (isa<InvokeInst>(Inst)) 2736 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt()); 2737 else 2738 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst))); 2739 } 2740 break; 2741 case S_Stop: 2742 if (CanUse(Inst, Ptr, PA, Class)) 2743 S.SetSeq(S_Use); 2744 break; 2745 case S_CanRelease: 2746 case S_Use: 2747 case S_None: 2748 break; 2749 case S_Retain: 2750 llvm_unreachable("bottom-up pointer in retain state!"); 2751 } 2752 } 2753 2754 return NestingDetected; 2755 } 2756 2757 bool 2758 ObjCARCOpt::VisitBottomUp(BasicBlock *BB, 2759 DenseMap<const BasicBlock *, BBState> &BBStates, 2760 MapVector<Value *, RRInfo> &Retains) { 2761 bool NestingDetected = false; 2762 BBState &MyStates = BBStates[BB]; 2763 2764 // Merge the states from each successor to compute the initial state 2765 // for the current block. 2766 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back()); 2767 succ_const_iterator SI(TI), SE(TI, false); 2768 if (SI == SE) 2769 MyStates.SetAsExit(); 2770 else { 2771 // If the terminator is an invoke marked with the 2772 // clang.arc.no_objc_arc_exceptions metadata, the unwind edge can be 2773 // ignored, for ARC purposes. 2774 if (isa<InvokeInst>(TI) && TI->getMetadata(NoObjCARCExceptionsMDKind)) 2775 --SE; 2776 2777 do { 2778 const BasicBlock *Succ = *SI++; 2779 if (Succ == BB) 2780 continue; 2781 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ); 2782 // If we haven't seen this node yet, then we've found a CFG cycle. 2783 // Be optimistic here; it's CheckForCFGHazards' job detect trouble. 2784 if (I == BBStates.end()) 2785 continue; 2786 MyStates.InitFromSucc(I->second); 2787 while (SI != SE) { 2788 Succ = *SI++; 2789 if (Succ != BB) { 2790 I = BBStates.find(Succ); 2791 if (I != BBStates.end()) 2792 MyStates.MergeSucc(I->second); 2793 } 2794 } 2795 break; 2796 } while (SI != SE); 2797 } 2798 2799 // Visit all the instructions, bottom-up. 2800 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) { 2801 Instruction *Inst = llvm::prior(I); 2802 2803 // Invoke instructions are visited as part of their successors (below). 2804 if (isa<InvokeInst>(Inst)) 2805 continue; 2806 2807 NestingDetected |= VisitInstructionBottomUp(Inst, BB, Retains, MyStates); 2808 } 2809 2810 // If there's a predecessor with an invoke, visit the invoke as 2811 // if it were part of this block, since we can't insert code after 2812 // an invoke in its own block, and we don't want to split critical 2813 // edges. 2814 for (pred_iterator PI(BB), PE(BB, false); PI != PE; ++PI) { 2815 BasicBlock *Pred = *PI; 2816 TerminatorInst *PredTI = cast<TerminatorInst>(&Pred->back()); 2817 if (isa<InvokeInst>(PredTI)) 2818 NestingDetected |= VisitInstructionBottomUp(PredTI, BB, Retains, MyStates); 2819 } 2820 2821 return NestingDetected; 2822 } 2823 2824 bool 2825 ObjCARCOpt::VisitInstructionTopDown(Instruction *Inst, 2826 DenseMap<Value *, RRInfo> &Releases, 2827 BBState &MyStates) { 2828 bool NestingDetected = false; 2829 InstructionClass Class = GetInstructionClass(Inst); 2830 const Value *Arg = 0; 2831 2832 switch (Class) { 2833 case IC_RetainBlock: 2834 // An objc_retainBlock call with just a use may need to be kept, 2835 // because it may be copying a block from the stack to the heap. 2836 if (!IsRetainBlockOptimizable(Inst)) 2837 break; 2838 // FALLTHROUGH 2839 case IC_Retain: 2840 case IC_RetainRV: { 2841 Arg = GetObjCArg(Inst); 2842 2843 PtrState &S = MyStates.getPtrTopDownState(Arg); 2844 2845 // Don't do retain+release tracking for IC_RetainRV, because it's 2846 // better to let it remain as the first instruction after a call. 2847 if (Class != IC_RetainRV) { 2848 // If we see two retains in a row on the same pointer. If so, make 2849 // a note, and we'll cicle back to revisit it after we've 2850 // hopefully eliminated the second retain, which may allow us to 2851 // eliminate the first retain too. 2852 // Theoretically we could implement removal of nested retain+release 2853 // pairs by making PtrState hold a stack of states, but this is 2854 // simple and avoids adding overhead for the non-nested case. 2855 if (S.GetSeq() == S_Retain) 2856 NestingDetected = true; 2857 2858 S.SetSeq(S_Retain); 2859 S.RRI.clear(); 2860 S.RRI.IsRetainBlock = Class == IC_RetainBlock; 2861 // Don't check S.IsKnownIncremented() here because it's not 2862 // sufficient. 2863 S.RRI.KnownSafe = S.IsKnownNested(); 2864 S.RRI.Calls.insert(Inst); 2865 } 2866 2867 S.SetAtLeastOneRefCount(); 2868 S.IncrementRefCount(); 2869 S.IncrementNestCount(); 2870 return NestingDetected; 2871 } 2872 case IC_Release: { 2873 Arg = GetObjCArg(Inst); 2874 2875 PtrState &S = MyStates.getPtrTopDownState(Arg); 2876 S.DecrementRefCount(); 2877 S.DecrementNestCount(); 2878 2879 switch (S.GetSeq()) { 2880 case S_Retain: 2881 case S_CanRelease: 2882 S.RRI.ReverseInsertPts.clear(); 2883 // FALL THROUGH 2884 case S_Use: 2885 S.RRI.ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind); 2886 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall(); 2887 Releases[Inst] = S.RRI; 2888 S.ClearSequenceProgress(); 2889 break; 2890 case S_None: 2891 break; 2892 case S_Stop: 2893 case S_Release: 2894 case S_MovableRelease: 2895 llvm_unreachable("top-down pointer in release state!"); 2896 } 2897 break; 2898 } 2899 case IC_AutoreleasepoolPop: 2900 // Conservatively, clear MyStates for all known pointers. 2901 MyStates.clearTopDownPointers(); 2902 return NestingDetected; 2903 case IC_AutoreleasepoolPush: 2904 case IC_None: 2905 // These are irrelevant. 2906 return NestingDetected; 2907 default: 2908 break; 2909 } 2910 2911 // Consider any other possible effects of this instruction on each 2912 // pointer being tracked. 2913 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(), 2914 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) { 2915 const Value *Ptr = MI->first; 2916 if (Ptr == Arg) 2917 continue; // Handled above. 2918 PtrState &S = MI->second; 2919 Sequence Seq = S.GetSeq(); 2920 2921 // Check for possible releases. 2922 if (CanAlterRefCount(Inst, Ptr, PA, Class)) { 2923 S.DecrementRefCount(); 2924 switch (Seq) { 2925 case S_Retain: 2926 S.SetSeq(S_CanRelease); 2927 assert(S.RRI.ReverseInsertPts.empty()); 2928 S.RRI.ReverseInsertPts.insert(Inst); 2929 2930 // One call can't cause a transition from S_Retain to S_CanRelease 2931 // and S_CanRelease to S_Use. If we've made the first transition, 2932 // we're done. 2933 continue; 2934 case S_Use: 2935 case S_CanRelease: 2936 case S_None: 2937 break; 2938 case S_Stop: 2939 case S_Release: 2940 case S_MovableRelease: 2941 llvm_unreachable("top-down pointer in release state!"); 2942 } 2943 } 2944 2945 // Check for possible direct uses. 2946 switch (Seq) { 2947 case S_CanRelease: 2948 if (CanUse(Inst, Ptr, PA, Class)) 2949 S.SetSeq(S_Use); 2950 break; 2951 case S_Retain: 2952 case S_Use: 2953 case S_None: 2954 break; 2955 case S_Stop: 2956 case S_Release: 2957 case S_MovableRelease: 2958 llvm_unreachable("top-down pointer in release state!"); 2959 } 2960 } 2961 2962 return NestingDetected; 2963 } 2964 2965 bool 2966 ObjCARCOpt::VisitTopDown(BasicBlock *BB, 2967 DenseMap<const BasicBlock *, BBState> &BBStates, 2968 DenseMap<Value *, RRInfo> &Releases) { 2969 bool NestingDetected = false; 2970 BBState &MyStates = BBStates[BB]; 2971 2972 // Merge the states from each predecessor to compute the initial state 2973 // for the current block. 2974 const_pred_iterator PI(BB), PE(BB, false); 2975 if (PI == PE) 2976 MyStates.SetAsEntry(); 2977 else 2978 do { 2979 unsigned OperandNo = PI.getOperandNo(); 2980 const Use &Us = PI.getUse(); 2981 ++PI; 2982 2983 // Skip invoke unwind edges on invoke instructions marked with 2984 // clang.arc.no_objc_arc_exceptions. 2985 if (const InvokeInst *II = dyn_cast<InvokeInst>(Us.getUser())) 2986 if (OperandNo == II->getNumArgOperands() + 2 && 2987 II->getMetadata(NoObjCARCExceptionsMDKind)) 2988 continue; 2989 2990 const BasicBlock *Pred = cast<TerminatorInst>(Us.getUser())->getParent(); 2991 if (Pred == BB) 2992 continue; 2993 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred); 2994 // If we haven't seen this node yet, then we've found a CFG cycle. 2995 // Be optimistic here; it's CheckForCFGHazards' job detect trouble. 2996 if (I == BBStates.end() || !I->second.isVisitedTopDown()) 2997 continue; 2998 MyStates.InitFromPred(I->second); 2999 while (PI != PE) { 3000 Pred = *PI++; 3001 if (Pred != BB) { 3002 I = BBStates.find(Pred); 3003 if (I != BBStates.end() && I->second.isVisitedTopDown()) 3004 MyStates.MergePred(I->second); 3005 } 3006 } 3007 break; 3008 } while (PI != PE); 3009 3010 // Visit all the instructions, top-down. 3011 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 3012 Instruction *Inst = I; 3013 NestingDetected |= VisitInstructionTopDown(Inst, Releases, MyStates); 3014 } 3015 3016 CheckForCFGHazards(BB, BBStates, MyStates); 3017 return NestingDetected; 3018 } 3019 3020 static void 3021 ComputePostOrders(Function &F, 3022 SmallVectorImpl<BasicBlock *> &PostOrder, 3023 SmallVectorImpl<BasicBlock *> &ReverseCFGPostOrder) { 3024 /// Backedges - Backedges detected in the DFS. These edges will be 3025 /// ignored in the reverse-CFG DFS, so that loops with multiple exits will be 3026 /// traversed in the desired order. 3027 DenseSet<std::pair<BasicBlock *, BasicBlock *> > Backedges; 3028 3029 /// Visited - The visited set, for doing DFS walks. 3030 SmallPtrSet<BasicBlock *, 16> Visited; 3031 3032 // Do DFS, computing the PostOrder. 3033 SmallPtrSet<BasicBlock *, 16> OnStack; 3034 SmallVector<std::pair<BasicBlock *, succ_iterator>, 16> SuccStack; 3035 BasicBlock *EntryBB = &F.getEntryBlock(); 3036 SuccStack.push_back(std::make_pair(EntryBB, succ_begin(EntryBB))); 3037 Visited.insert(EntryBB); 3038 OnStack.insert(EntryBB); 3039 do { 3040 dfs_next_succ: 3041 TerminatorInst *TI = cast<TerminatorInst>(&SuccStack.back().first->back()); 3042 succ_iterator End = succ_iterator(TI, true); 3043 while (SuccStack.back().second != End) { 3044 BasicBlock *BB = *SuccStack.back().second++; 3045 if (Visited.insert(BB)) { 3046 SuccStack.push_back(std::make_pair(BB, succ_begin(BB))); 3047 OnStack.insert(BB); 3048 goto dfs_next_succ; 3049 } 3050 if (OnStack.count(BB)) 3051 Backedges.insert(std::make_pair(SuccStack.back().first, BB)); 3052 } 3053 OnStack.erase(SuccStack.back().first); 3054 PostOrder.push_back(SuccStack.pop_back_val().first); 3055 } while (!SuccStack.empty()); 3056 3057 Visited.clear(); 3058 3059 // Compute the exits, which are the starting points for reverse-CFG DFS. 3060 // This includes blocks where all the successors are backedges that 3061 // we're skipping. 3062 SmallVector<BasicBlock *, 4> Exits; 3063 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) { 3064 BasicBlock *BB = I; 3065 TerminatorInst *TI = cast<TerminatorInst>(&BB->back()); 3066 for (succ_iterator SI(TI), SE(TI, true); SI != SE; ++SI) 3067 if (!Backedges.count(std::make_pair(BB, *SI))) 3068 goto HasNonBackedgeSucc; 3069 Exits.push_back(BB); 3070 HasNonBackedgeSucc:; 3071 } 3072 3073 // Do reverse-CFG DFS, computing the reverse-CFG PostOrder. 3074 SmallVector<std::pair<BasicBlock *, pred_iterator>, 16> PredStack; 3075 for (SmallVectorImpl<BasicBlock *>::iterator I = Exits.begin(), E = Exits.end(); 3076 I != E; ++I) { 3077 BasicBlock *ExitBB = *I; 3078 PredStack.push_back(std::make_pair(ExitBB, pred_begin(ExitBB))); 3079 Visited.insert(ExitBB); 3080 while (!PredStack.empty()) { 3081 reverse_dfs_next_succ: 3082 pred_iterator End = pred_end(PredStack.back().first); 3083 while (PredStack.back().second != End) { 3084 BasicBlock *BB = *PredStack.back().second++; 3085 // Skip backedges detected in the forward-CFG DFS. 3086 if (Backedges.count(std::make_pair(BB, PredStack.back().first))) 3087 continue; 3088 if (Visited.insert(BB)) { 3089 PredStack.push_back(std::make_pair(BB, pred_begin(BB))); 3090 goto reverse_dfs_next_succ; 3091 } 3092 } 3093 ReverseCFGPostOrder.push_back(PredStack.pop_back_val().first); 3094 } 3095 } 3096 } 3097 3098 // Visit - Visit the function both top-down and bottom-up. 3099 bool 3100 ObjCARCOpt::Visit(Function &F, 3101 DenseMap<const BasicBlock *, BBState> &BBStates, 3102 MapVector<Value *, RRInfo> &Retains, 3103 DenseMap<Value *, RRInfo> &Releases) { 3104 3105 // Use reverse-postorder traversals, because we magically know that loops 3106 // will be well behaved, i.e. they won't repeatedly call retain on a single 3107 // pointer without doing a release. We can't use the ReversePostOrderTraversal 3108 // class here because we want the reverse-CFG postorder to consider each 3109 // function exit point, and we want to ignore selected cycle edges. 3110 SmallVector<BasicBlock *, 16> PostOrder; 3111 SmallVector<BasicBlock *, 16> ReverseCFGPostOrder; 3112 ComputePostOrders(F, PostOrder, ReverseCFGPostOrder); 3113 3114 // Use reverse-postorder on the reverse CFG for bottom-up. 3115 bool BottomUpNestingDetected = false; 3116 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I = 3117 ReverseCFGPostOrder.rbegin(), E = ReverseCFGPostOrder.rend(); 3118 I != E; ++I) 3119 BottomUpNestingDetected |= VisitBottomUp(*I, BBStates, Retains); 3120 3121 // Use reverse-postorder for top-down. 3122 bool TopDownNestingDetected = false; 3123 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I = 3124 PostOrder.rbegin(), E = PostOrder.rend(); 3125 I != E; ++I) 3126 TopDownNestingDetected |= VisitTopDown(*I, BBStates, Releases); 3127 3128 return TopDownNestingDetected && BottomUpNestingDetected; 3129 } 3130 3131 /// MoveCalls - Move the calls in RetainsToMove and ReleasesToMove. 3132 void ObjCARCOpt::MoveCalls(Value *Arg, 3133 RRInfo &RetainsToMove, 3134 RRInfo &ReleasesToMove, 3135 MapVector<Value *, RRInfo> &Retains, 3136 DenseMap<Value *, RRInfo> &Releases, 3137 SmallVectorImpl<Instruction *> &DeadInsts, 3138 Module *M) { 3139 Type *ArgTy = Arg->getType(); 3140 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext())); 3141 3142 // Insert the new retain and release calls. 3143 for (SmallPtrSet<Instruction *, 2>::const_iterator 3144 PI = ReleasesToMove.ReverseInsertPts.begin(), 3145 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) { 3146 Instruction *InsertPt = *PI; 3147 Value *MyArg = ArgTy == ParamTy ? Arg : 3148 new BitCastInst(Arg, ParamTy, "", InsertPt); 3149 CallInst *Call = 3150 CallInst::Create(RetainsToMove.IsRetainBlock ? 3151 getRetainBlockCallee(M) : getRetainCallee(M), 3152 MyArg, "", InsertPt); 3153 Call->setDoesNotThrow(); 3154 if (RetainsToMove.IsRetainBlock) 3155 Call->setMetadata(CopyOnEscapeMDKind, 3156 MDNode::get(M->getContext(), ArrayRef<Value *>())); 3157 else 3158 Call->setTailCall(); 3159 } 3160 for (SmallPtrSet<Instruction *, 2>::const_iterator 3161 PI = RetainsToMove.ReverseInsertPts.begin(), 3162 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) { 3163 Instruction *InsertPt = *PI; 3164 Value *MyArg = ArgTy == ParamTy ? Arg : 3165 new BitCastInst(Arg, ParamTy, "", InsertPt); 3166 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg, 3167 "", InsertPt); 3168 // Attach a clang.imprecise_release metadata tag, if appropriate. 3169 if (MDNode *M = ReleasesToMove.ReleaseMetadata) 3170 Call->setMetadata(ImpreciseReleaseMDKind, M); 3171 Call->setDoesNotThrow(); 3172 if (ReleasesToMove.IsTailCallRelease) 3173 Call->setTailCall(); 3174 } 3175 3176 // Delete the original retain and release calls. 3177 for (SmallPtrSet<Instruction *, 2>::const_iterator 3178 AI = RetainsToMove.Calls.begin(), 3179 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) { 3180 Instruction *OrigRetain = *AI; 3181 Retains.blot(OrigRetain); 3182 DeadInsts.push_back(OrigRetain); 3183 } 3184 for (SmallPtrSet<Instruction *, 2>::const_iterator 3185 AI = ReleasesToMove.Calls.begin(), 3186 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) { 3187 Instruction *OrigRelease = *AI; 3188 Releases.erase(OrigRelease); 3189 DeadInsts.push_back(OrigRelease); 3190 } 3191 } 3192 3193 /// PerformCodePlacement - Identify pairings between the retains and releases, 3194 /// and delete and/or move them. 3195 bool 3196 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState> 3197 &BBStates, 3198 MapVector<Value *, RRInfo> &Retains, 3199 DenseMap<Value *, RRInfo> &Releases, 3200 Module *M) { 3201 bool AnyPairsCompletelyEliminated = false; 3202 RRInfo RetainsToMove; 3203 RRInfo ReleasesToMove; 3204 SmallVector<Instruction *, 4> NewRetains; 3205 SmallVector<Instruction *, 4> NewReleases; 3206 SmallVector<Instruction *, 8> DeadInsts; 3207 3208 // Visit each retain. 3209 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(), 3210 E = Retains.end(); I != E; ++I) { 3211 Value *V = I->first; 3212 if (!V) continue; // blotted 3213 3214 Instruction *Retain = cast<Instruction>(V); 3215 Value *Arg = GetObjCArg(Retain); 3216 3217 // If the object being released is in static or stack storage, we know it's 3218 // not being managed by ObjC reference counting, so we can delete pairs 3219 // regardless of what possible decrements or uses lie between them. 3220 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg); 3221 3222 // A constant pointer can't be pointing to an object on the heap. It may 3223 // be reference-counted, but it won't be deleted. 3224 if (const LoadInst *LI = dyn_cast<LoadInst>(Arg)) 3225 if (const GlobalVariable *GV = 3226 dyn_cast<GlobalVariable>( 3227 StripPointerCastsAndObjCCalls(LI->getPointerOperand()))) 3228 if (GV->isConstant()) 3229 KnownSafe = true; 3230 3231 // If a pair happens in a region where it is known that the reference count 3232 // is already incremented, we can similarly ignore possible decrements. 3233 bool KnownSafeTD = true, KnownSafeBU = true; 3234 3235 // Connect the dots between the top-down-collected RetainsToMove and 3236 // bottom-up-collected ReleasesToMove to form sets of related calls. 3237 // This is an iterative process so that we connect multiple releases 3238 // to multiple retains if needed. 3239 unsigned OldDelta = 0; 3240 unsigned NewDelta = 0; 3241 unsigned OldCount = 0; 3242 unsigned NewCount = 0; 3243 bool FirstRelease = true; 3244 bool FirstRetain = true; 3245 NewRetains.push_back(Retain); 3246 for (;;) { 3247 for (SmallVectorImpl<Instruction *>::const_iterator 3248 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) { 3249 Instruction *NewRetain = *NI; 3250 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain); 3251 assert(It != Retains.end()); 3252 const RRInfo &NewRetainRRI = It->second; 3253 KnownSafeTD &= NewRetainRRI.KnownSafe; 3254 for (SmallPtrSet<Instruction *, 2>::const_iterator 3255 LI = NewRetainRRI.Calls.begin(), 3256 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) { 3257 Instruction *NewRetainRelease = *LI; 3258 DenseMap<Value *, RRInfo>::const_iterator Jt = 3259 Releases.find(NewRetainRelease); 3260 if (Jt == Releases.end()) 3261 goto next_retain; 3262 const RRInfo &NewRetainReleaseRRI = Jt->second; 3263 assert(NewRetainReleaseRRI.Calls.count(NewRetain)); 3264 if (ReleasesToMove.Calls.insert(NewRetainRelease)) { 3265 OldDelta -= 3266 BBStates[NewRetainRelease->getParent()].GetAllPathCount(); 3267 3268 // Merge the ReleaseMetadata and IsTailCallRelease values. 3269 if (FirstRelease) { 3270 ReleasesToMove.ReleaseMetadata = 3271 NewRetainReleaseRRI.ReleaseMetadata; 3272 ReleasesToMove.IsTailCallRelease = 3273 NewRetainReleaseRRI.IsTailCallRelease; 3274 FirstRelease = false; 3275 } else { 3276 if (ReleasesToMove.ReleaseMetadata != 3277 NewRetainReleaseRRI.ReleaseMetadata) 3278 ReleasesToMove.ReleaseMetadata = 0; 3279 if (ReleasesToMove.IsTailCallRelease != 3280 NewRetainReleaseRRI.IsTailCallRelease) 3281 ReleasesToMove.IsTailCallRelease = false; 3282 } 3283 3284 // Collect the optimal insertion points. 3285 if (!KnownSafe) 3286 for (SmallPtrSet<Instruction *, 2>::const_iterator 3287 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(), 3288 RE = NewRetainReleaseRRI.ReverseInsertPts.end(); 3289 RI != RE; ++RI) { 3290 Instruction *RIP = *RI; 3291 if (ReleasesToMove.ReverseInsertPts.insert(RIP)) 3292 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount(); 3293 } 3294 NewReleases.push_back(NewRetainRelease); 3295 } 3296 } 3297 } 3298 NewRetains.clear(); 3299 if (NewReleases.empty()) break; 3300 3301 // Back the other way. 3302 for (SmallVectorImpl<Instruction *>::const_iterator 3303 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) { 3304 Instruction *NewRelease = *NI; 3305 DenseMap<Value *, RRInfo>::const_iterator It = 3306 Releases.find(NewRelease); 3307 assert(It != Releases.end()); 3308 const RRInfo &NewReleaseRRI = It->second; 3309 KnownSafeBU &= NewReleaseRRI.KnownSafe; 3310 for (SmallPtrSet<Instruction *, 2>::const_iterator 3311 LI = NewReleaseRRI.Calls.begin(), 3312 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) { 3313 Instruction *NewReleaseRetain = *LI; 3314 MapVector<Value *, RRInfo>::const_iterator Jt = 3315 Retains.find(NewReleaseRetain); 3316 if (Jt == Retains.end()) 3317 goto next_retain; 3318 const RRInfo &NewReleaseRetainRRI = Jt->second; 3319 assert(NewReleaseRetainRRI.Calls.count(NewRelease)); 3320 if (RetainsToMove.Calls.insert(NewReleaseRetain)) { 3321 unsigned PathCount = 3322 BBStates[NewReleaseRetain->getParent()].GetAllPathCount(); 3323 OldDelta += PathCount; 3324 OldCount += PathCount; 3325 3326 // Merge the IsRetainBlock values. 3327 if (FirstRetain) { 3328 RetainsToMove.IsRetainBlock = NewReleaseRetainRRI.IsRetainBlock; 3329 FirstRetain = false; 3330 } else if (ReleasesToMove.IsRetainBlock != 3331 NewReleaseRetainRRI.IsRetainBlock) 3332 // It's not possible to merge the sequences if one uses 3333 // objc_retain and the other uses objc_retainBlock. 3334 goto next_retain; 3335 3336 // Collect the optimal insertion points. 3337 if (!KnownSafe) 3338 for (SmallPtrSet<Instruction *, 2>::const_iterator 3339 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(), 3340 RE = NewReleaseRetainRRI.ReverseInsertPts.end(); 3341 RI != RE; ++RI) { 3342 Instruction *RIP = *RI; 3343 if (RetainsToMove.ReverseInsertPts.insert(RIP)) { 3344 PathCount = BBStates[RIP->getParent()].GetAllPathCount(); 3345 NewDelta += PathCount; 3346 NewCount += PathCount; 3347 } 3348 } 3349 NewRetains.push_back(NewReleaseRetain); 3350 } 3351 } 3352 } 3353 NewReleases.clear(); 3354 if (NewRetains.empty()) break; 3355 } 3356 3357 // If the pointer is known incremented or nested, we can safely delete the 3358 // pair regardless of what's between them. 3359 if (KnownSafeTD || KnownSafeBU) { 3360 RetainsToMove.ReverseInsertPts.clear(); 3361 ReleasesToMove.ReverseInsertPts.clear(); 3362 NewCount = 0; 3363 } else { 3364 // Determine whether the new insertion points we computed preserve the 3365 // balance of retain and release calls through the program. 3366 // TODO: If the fully aggressive solution isn't valid, try to find a 3367 // less aggressive solution which is. 3368 if (NewDelta != 0) 3369 goto next_retain; 3370 } 3371 3372 // Determine whether the original call points are balanced in the retain and 3373 // release calls through the program. If not, conservatively don't touch 3374 // them. 3375 // TODO: It's theoretically possible to do code motion in this case, as 3376 // long as the existing imbalances are maintained. 3377 if (OldDelta != 0) 3378 goto next_retain; 3379 3380 // Ok, everything checks out and we're all set. Let's move some code! 3381 Changed = true; 3382 AnyPairsCompletelyEliminated = NewCount == 0; 3383 NumRRs += OldCount - NewCount; 3384 MoveCalls(Arg, RetainsToMove, ReleasesToMove, 3385 Retains, Releases, DeadInsts, M); 3386 3387 next_retain: 3388 NewReleases.clear(); 3389 NewRetains.clear(); 3390 RetainsToMove.clear(); 3391 ReleasesToMove.clear(); 3392 } 3393 3394 // Now that we're done moving everything, we can delete the newly dead 3395 // instructions, as we no longer need them as insert points. 3396 while (!DeadInsts.empty()) 3397 EraseInstruction(DeadInsts.pop_back_val()); 3398 3399 return AnyPairsCompletelyEliminated; 3400 } 3401 3402 /// OptimizeWeakCalls - Weak pointer optimizations. 3403 void ObjCARCOpt::OptimizeWeakCalls(Function &F) { 3404 // First, do memdep-style RLE and S2L optimizations. We can't use memdep 3405 // itself because it uses AliasAnalysis and we need to do provenance 3406 // queries instead. 3407 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) { 3408 Instruction *Inst = &*I++; 3409 InstructionClass Class = GetBasicInstructionClass(Inst); 3410 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained) 3411 continue; 3412 3413 // Delete objc_loadWeak calls with no users. 3414 if (Class == IC_LoadWeak && Inst->use_empty()) { 3415 Inst->eraseFromParent(); 3416 continue; 3417 } 3418 3419 // TODO: For now, just look for an earlier available version of this value 3420 // within the same block. Theoretically, we could do memdep-style non-local 3421 // analysis too, but that would want caching. A better approach would be to 3422 // use the technique that EarlyCSE uses. 3423 inst_iterator Current = llvm::prior(I); 3424 BasicBlock *CurrentBB = Current.getBasicBlockIterator(); 3425 for (BasicBlock::iterator B = CurrentBB->begin(), 3426 J = Current.getInstructionIterator(); 3427 J != B; --J) { 3428 Instruction *EarlierInst = &*llvm::prior(J); 3429 InstructionClass EarlierClass = GetInstructionClass(EarlierInst); 3430 switch (EarlierClass) { 3431 case IC_LoadWeak: 3432 case IC_LoadWeakRetained: { 3433 // If this is loading from the same pointer, replace this load's value 3434 // with that one. 3435 CallInst *Call = cast<CallInst>(Inst); 3436 CallInst *EarlierCall = cast<CallInst>(EarlierInst); 3437 Value *Arg = Call->getArgOperand(0); 3438 Value *EarlierArg = EarlierCall->getArgOperand(0); 3439 switch (PA.getAA()->alias(Arg, EarlierArg)) { 3440 case AliasAnalysis::MustAlias: 3441 Changed = true; 3442 // If the load has a builtin retain, insert a plain retain for it. 3443 if (Class == IC_LoadWeakRetained) { 3444 CallInst *CI = 3445 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall, 3446 "", Call); 3447 CI->setTailCall(); 3448 } 3449 // Zap the fully redundant load. 3450 Call->replaceAllUsesWith(EarlierCall); 3451 Call->eraseFromParent(); 3452 goto clobbered; 3453 case AliasAnalysis::MayAlias: 3454 case AliasAnalysis::PartialAlias: 3455 goto clobbered; 3456 case AliasAnalysis::NoAlias: 3457 break; 3458 } 3459 break; 3460 } 3461 case IC_StoreWeak: 3462 case IC_InitWeak: { 3463 // If this is storing to the same pointer and has the same size etc. 3464 // replace this load's value with the stored value. 3465 CallInst *Call = cast<CallInst>(Inst); 3466 CallInst *EarlierCall = cast<CallInst>(EarlierInst); 3467 Value *Arg = Call->getArgOperand(0); 3468 Value *EarlierArg = EarlierCall->getArgOperand(0); 3469 switch (PA.getAA()->alias(Arg, EarlierArg)) { 3470 case AliasAnalysis::MustAlias: 3471 Changed = true; 3472 // If the load has a builtin retain, insert a plain retain for it. 3473 if (Class == IC_LoadWeakRetained) { 3474 CallInst *CI = 3475 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall, 3476 "", Call); 3477 CI->setTailCall(); 3478 } 3479 // Zap the fully redundant load. 3480 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1)); 3481 Call->eraseFromParent(); 3482 goto clobbered; 3483 case AliasAnalysis::MayAlias: 3484 case AliasAnalysis::PartialAlias: 3485 goto clobbered; 3486 case AliasAnalysis::NoAlias: 3487 break; 3488 } 3489 break; 3490 } 3491 case IC_MoveWeak: 3492 case IC_CopyWeak: 3493 // TOOD: Grab the copied value. 3494 goto clobbered; 3495 case IC_AutoreleasepoolPush: 3496 case IC_None: 3497 case IC_User: 3498 // Weak pointers are only modified through the weak entry points 3499 // (and arbitrary calls, which could call the weak entry points). 3500 break; 3501 default: 3502 // Anything else could modify the weak pointer. 3503 goto clobbered; 3504 } 3505 } 3506 clobbered:; 3507 } 3508 3509 // Then, for each destroyWeak with an alloca operand, check to see if 3510 // the alloca and all its users can be zapped. 3511 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) { 3512 Instruction *Inst = &*I++; 3513 InstructionClass Class = GetBasicInstructionClass(Inst); 3514 if (Class != IC_DestroyWeak) 3515 continue; 3516 3517 CallInst *Call = cast<CallInst>(Inst); 3518 Value *Arg = Call->getArgOperand(0); 3519 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) { 3520 for (Value::use_iterator UI = Alloca->use_begin(), 3521 UE = Alloca->use_end(); UI != UE; ++UI) { 3522 Instruction *UserInst = cast<Instruction>(*UI); 3523 switch (GetBasicInstructionClass(UserInst)) { 3524 case IC_InitWeak: 3525 case IC_StoreWeak: 3526 case IC_DestroyWeak: 3527 continue; 3528 default: 3529 goto done; 3530 } 3531 } 3532 Changed = true; 3533 for (Value::use_iterator UI = Alloca->use_begin(), 3534 UE = Alloca->use_end(); UI != UE; ) { 3535 CallInst *UserInst = cast<CallInst>(*UI++); 3536 if (!UserInst->use_empty()) 3537 UserInst->replaceAllUsesWith(UserInst->getArgOperand(0)); 3538 UserInst->eraseFromParent(); 3539 } 3540 Alloca->eraseFromParent(); 3541 done:; 3542 } 3543 } 3544 } 3545 3546 /// OptimizeSequences - Identify program paths which execute sequences of 3547 /// retains and releases which can be eliminated. 3548 bool ObjCARCOpt::OptimizeSequences(Function &F) { 3549 /// Releases, Retains - These are used to store the results of the main flow 3550 /// analysis. These use Value* as the key instead of Instruction* so that the 3551 /// map stays valid when we get around to rewriting code and calls get 3552 /// replaced by arguments. 3553 DenseMap<Value *, RRInfo> Releases; 3554 MapVector<Value *, RRInfo> Retains; 3555 3556 /// BBStates, This is used during the traversal of the function to track the 3557 /// states for each identified object at each block. 3558 DenseMap<const BasicBlock *, BBState> BBStates; 3559 3560 // Analyze the CFG of the function, and all instructions. 3561 bool NestingDetected = Visit(F, BBStates, Retains, Releases); 3562 3563 // Transform. 3564 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) && 3565 NestingDetected; 3566 } 3567 3568 /// OptimizeReturns - Look for this pattern: 3569 /// 3570 /// %call = call i8* @something(...) 3571 /// %2 = call i8* @objc_retain(i8* %call) 3572 /// %3 = call i8* @objc_autorelease(i8* %2) 3573 /// ret i8* %3 3574 /// 3575 /// And delete the retain and autorelease. 3576 /// 3577 /// Otherwise if it's just this: 3578 /// 3579 /// %3 = call i8* @objc_autorelease(i8* %2) 3580 /// ret i8* %3 3581 /// 3582 /// convert the autorelease to autoreleaseRV. 3583 void ObjCARCOpt::OptimizeReturns(Function &F) { 3584 if (!F.getReturnType()->isPointerTy()) 3585 return; 3586 3587 SmallPtrSet<Instruction *, 4> DependingInstructions; 3588 SmallPtrSet<const BasicBlock *, 4> Visited; 3589 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) { 3590 BasicBlock *BB = FI; 3591 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back()); 3592 if (!Ret) continue; 3593 3594 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0)); 3595 FindDependencies(NeedsPositiveRetainCount, Arg, 3596 BB, Ret, DependingInstructions, Visited, PA); 3597 if (DependingInstructions.size() != 1) 3598 goto next_block; 3599 3600 { 3601 CallInst *Autorelease = 3602 dyn_cast_or_null<CallInst>(*DependingInstructions.begin()); 3603 if (!Autorelease) 3604 goto next_block; 3605 InstructionClass AutoreleaseClass = 3606 GetBasicInstructionClass(Autorelease); 3607 if (!IsAutorelease(AutoreleaseClass)) 3608 goto next_block; 3609 if (GetObjCArg(Autorelease) != Arg) 3610 goto next_block; 3611 3612 DependingInstructions.clear(); 3613 Visited.clear(); 3614 3615 // Check that there is nothing that can affect the reference 3616 // count between the autorelease and the retain. 3617 FindDependencies(CanChangeRetainCount, Arg, 3618 BB, Autorelease, DependingInstructions, Visited, PA); 3619 if (DependingInstructions.size() != 1) 3620 goto next_block; 3621 3622 { 3623 CallInst *Retain = 3624 dyn_cast_or_null<CallInst>(*DependingInstructions.begin()); 3625 3626 // Check that we found a retain with the same argument. 3627 if (!Retain || 3628 !IsRetain(GetBasicInstructionClass(Retain)) || 3629 GetObjCArg(Retain) != Arg) 3630 goto next_block; 3631 3632 DependingInstructions.clear(); 3633 Visited.clear(); 3634 3635 // Convert the autorelease to an autoreleaseRV, since it's 3636 // returning the value. 3637 if (AutoreleaseClass == IC_Autorelease) { 3638 Autorelease->setCalledFunction(getAutoreleaseRVCallee(F.getParent())); 3639 AutoreleaseClass = IC_AutoreleaseRV; 3640 } 3641 3642 // Check that there is nothing that can affect the reference 3643 // count between the retain and the call. 3644 // Note that Retain need not be in BB. 3645 FindDependencies(CanChangeRetainCount, Arg, Retain->getParent(), Retain, 3646 DependingInstructions, Visited, PA); 3647 if (DependingInstructions.size() != 1) 3648 goto next_block; 3649 3650 { 3651 CallInst *Call = 3652 dyn_cast_or_null<CallInst>(*DependingInstructions.begin()); 3653 3654 // Check that the pointer is the return value of the call. 3655 if (!Call || Arg != Call) 3656 goto next_block; 3657 3658 // Check that the call is a regular call. 3659 InstructionClass Class = GetBasicInstructionClass(Call); 3660 if (Class != IC_CallOrUser && Class != IC_Call) 3661 goto next_block; 3662 3663 // If so, we can zap the retain and autorelease. 3664 Changed = true; 3665 ++NumRets; 3666 EraseInstruction(Retain); 3667 EraseInstruction(Autorelease); 3668 } 3669 } 3670 } 3671 3672 next_block: 3673 DependingInstructions.clear(); 3674 Visited.clear(); 3675 } 3676 } 3677 3678 bool ObjCARCOpt::doInitialization(Module &M) { 3679 if (!EnableARCOpts) 3680 return false; 3681 3682 // If nothing in the Module uses ARC, don't do anything. 3683 Run = ModuleHasARC(M); 3684 if (!Run) 3685 return false; 3686 3687 // Identify the imprecise release metadata kind. 3688 ImpreciseReleaseMDKind = 3689 M.getContext().getMDKindID("clang.imprecise_release"); 3690 CopyOnEscapeMDKind = 3691 M.getContext().getMDKindID("clang.arc.copy_on_escape"); 3692 NoObjCARCExceptionsMDKind = 3693 M.getContext().getMDKindID("clang.arc.no_objc_arc_exceptions"); 3694 3695 // Intuitively, objc_retain and others are nocapture, however in practice 3696 // they are not, because they return their argument value. And objc_release 3697 // calls finalizers. 3698 3699 // These are initialized lazily. 3700 RetainRVCallee = 0; 3701 AutoreleaseRVCallee = 0; 3702 ReleaseCallee = 0; 3703 RetainCallee = 0; 3704 RetainBlockCallee = 0; 3705 AutoreleaseCallee = 0; 3706 3707 return false; 3708 } 3709 3710 bool ObjCARCOpt::runOnFunction(Function &F) { 3711 if (!EnableARCOpts) 3712 return false; 3713 3714 // If nothing in the Module uses ARC, don't do anything. 3715 if (!Run) 3716 return false; 3717 3718 Changed = false; 3719 3720 PA.setAA(&getAnalysis<AliasAnalysis>()); 3721 3722 // This pass performs several distinct transformations. As a compile-time aid 3723 // when compiling code that isn't ObjC, skip these if the relevant ObjC 3724 // library functions aren't declared. 3725 3726 // Preliminary optimizations. This also computs UsedInThisFunction. 3727 OptimizeIndividualCalls(F); 3728 3729 // Optimizations for weak pointers. 3730 if (UsedInThisFunction & ((1 << IC_LoadWeak) | 3731 (1 << IC_LoadWeakRetained) | 3732 (1 << IC_StoreWeak) | 3733 (1 << IC_InitWeak) | 3734 (1 << IC_CopyWeak) | 3735 (1 << IC_MoveWeak) | 3736 (1 << IC_DestroyWeak))) 3737 OptimizeWeakCalls(F); 3738 3739 // Optimizations for retain+release pairs. 3740 if (UsedInThisFunction & ((1 << IC_Retain) | 3741 (1 << IC_RetainRV) | 3742 (1 << IC_RetainBlock))) 3743 if (UsedInThisFunction & (1 << IC_Release)) 3744 // Run OptimizeSequences until it either stops making changes or 3745 // no retain+release pair nesting is detected. 3746 while (OptimizeSequences(F)) {} 3747 3748 // Optimizations if objc_autorelease is used. 3749 if (UsedInThisFunction & 3750 ((1 << IC_Autorelease) | (1 << IC_AutoreleaseRV))) 3751 OptimizeReturns(F); 3752 3753 return Changed; 3754 } 3755 3756 void ObjCARCOpt::releaseMemory() { 3757 PA.clear(); 3758 } 3759 3760 //===----------------------------------------------------------------------===// 3761 // ARC contraction. 3762 //===----------------------------------------------------------------------===// 3763 3764 // TODO: ObjCARCContract could insert PHI nodes when uses aren't 3765 // dominated by single calls. 3766 3767 #include "llvm/Operator.h" 3768 #include "llvm/InlineAsm.h" 3769 #include "llvm/Analysis/Dominators.h" 3770 3771 STATISTIC(NumStoreStrongs, "Number objc_storeStrong calls formed"); 3772 3773 namespace { 3774 /// ObjCARCContract - Late ARC optimizations. These change the IR in a way 3775 /// that makes it difficult to be analyzed by ObjCARCOpt, so it's run late. 3776 class ObjCARCContract : public FunctionPass { 3777 bool Changed; 3778 AliasAnalysis *AA; 3779 DominatorTree *DT; 3780 ProvenanceAnalysis PA; 3781 3782 /// Run - A flag indicating whether this optimization pass should run. 3783 bool Run; 3784 3785 /// StoreStrongCallee, etc. - Declarations for ObjC runtime 3786 /// functions, for use in creating calls to them. These are initialized 3787 /// lazily to avoid cluttering up the Module with unused declarations. 3788 Constant *StoreStrongCallee, 3789 *RetainAutoreleaseCallee, *RetainAutoreleaseRVCallee; 3790 3791 /// RetainRVMarker - The inline asm string to insert between calls and 3792 /// RetainRV calls to make the optimization work on targets which need it. 3793 const MDString *RetainRVMarker; 3794 3795 /// StoreStrongCalls - The set of inserted objc_storeStrong calls. If 3796 /// at the end of walking the function we have found no alloca 3797 /// instructions, these calls can be marked "tail". 3798 DenseSet<CallInst *> StoreStrongCalls; 3799 3800 Constant *getStoreStrongCallee(Module *M); 3801 Constant *getRetainAutoreleaseCallee(Module *M); 3802 Constant *getRetainAutoreleaseRVCallee(Module *M); 3803 3804 bool ContractAutorelease(Function &F, Instruction *Autorelease, 3805 InstructionClass Class, 3806 SmallPtrSet<Instruction *, 4> 3807 &DependingInstructions, 3808 SmallPtrSet<const BasicBlock *, 4> 3809 &Visited); 3810 3811 void ContractRelease(Instruction *Release, 3812 inst_iterator &Iter); 3813 3814 virtual void getAnalysisUsage(AnalysisUsage &AU) const; 3815 virtual bool doInitialization(Module &M); 3816 virtual bool runOnFunction(Function &F); 3817 3818 public: 3819 static char ID; 3820 ObjCARCContract() : FunctionPass(ID) { 3821 initializeObjCARCContractPass(*PassRegistry::getPassRegistry()); 3822 } 3823 }; 3824 } 3825 3826 char ObjCARCContract::ID = 0; 3827 INITIALIZE_PASS_BEGIN(ObjCARCContract, 3828 "objc-arc-contract", "ObjC ARC contraction", false, false) 3829 INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3830 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 3831 INITIALIZE_PASS_END(ObjCARCContract, 3832 "objc-arc-contract", "ObjC ARC contraction", false, false) 3833 3834 Pass *llvm::createObjCARCContractPass() { 3835 return new ObjCARCContract(); 3836 } 3837 3838 void ObjCARCContract::getAnalysisUsage(AnalysisUsage &AU) const { 3839 AU.addRequired<AliasAnalysis>(); 3840 AU.addRequired<DominatorTree>(); 3841 AU.setPreservesCFG(); 3842 } 3843 3844 Constant *ObjCARCContract::getStoreStrongCallee(Module *M) { 3845 if (!StoreStrongCallee) { 3846 LLVMContext &C = M->getContext(); 3847 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); 3848 Type *I8XX = PointerType::getUnqual(I8X); 3849 std::vector<Type *> Params; 3850 Params.push_back(I8XX); 3851 Params.push_back(I8X); 3852 3853 AttrListPtr Attributes; 3854 Attributes.addAttr(~0u, Attribute::NoUnwind); 3855 Attributes.addAttr(1, Attribute::NoCapture); 3856 3857 StoreStrongCallee = 3858 M->getOrInsertFunction( 3859 "objc_storeStrong", 3860 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false), 3861 Attributes); 3862 } 3863 return StoreStrongCallee; 3864 } 3865 3866 Constant *ObjCARCContract::getRetainAutoreleaseCallee(Module *M) { 3867 if (!RetainAutoreleaseCallee) { 3868 LLVMContext &C = M->getContext(); 3869 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); 3870 std::vector<Type *> Params; 3871 Params.push_back(I8X); 3872 FunctionType *FTy = 3873 FunctionType::get(I8X, Params, /*isVarArg=*/false); 3874 AttrListPtr Attributes; 3875 Attributes.addAttr(~0u, Attribute::NoUnwind); 3876 RetainAutoreleaseCallee = 3877 M->getOrInsertFunction("objc_retainAutorelease", FTy, Attributes); 3878 } 3879 return RetainAutoreleaseCallee; 3880 } 3881 3882 Constant *ObjCARCContract::getRetainAutoreleaseRVCallee(Module *M) { 3883 if (!RetainAutoreleaseRVCallee) { 3884 LLVMContext &C = M->getContext(); 3885 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); 3886 std::vector<Type *> Params; 3887 Params.push_back(I8X); 3888 FunctionType *FTy = 3889 FunctionType::get(I8X, Params, /*isVarArg=*/false); 3890 AttrListPtr Attributes; 3891 Attributes.addAttr(~0u, Attribute::NoUnwind); 3892 RetainAutoreleaseRVCallee = 3893 M->getOrInsertFunction("objc_retainAutoreleaseReturnValue", FTy, 3894 Attributes); 3895 } 3896 return RetainAutoreleaseRVCallee; 3897 } 3898 3899 /// ContractAutorelease - Merge an autorelease with a retain into a fused 3900 /// call. 3901 bool 3902 ObjCARCContract::ContractAutorelease(Function &F, Instruction *Autorelease, 3903 InstructionClass Class, 3904 SmallPtrSet<Instruction *, 4> 3905 &DependingInstructions, 3906 SmallPtrSet<const BasicBlock *, 4> 3907 &Visited) { 3908 const Value *Arg = GetObjCArg(Autorelease); 3909 3910 // Check that there are no instructions between the retain and the autorelease 3911 // (such as an autorelease_pop) which may change the count. 3912 CallInst *Retain = 0; 3913 if (Class == IC_AutoreleaseRV) 3914 FindDependencies(RetainAutoreleaseRVDep, Arg, 3915 Autorelease->getParent(), Autorelease, 3916 DependingInstructions, Visited, PA); 3917 else 3918 FindDependencies(RetainAutoreleaseDep, Arg, 3919 Autorelease->getParent(), Autorelease, 3920 DependingInstructions, Visited, PA); 3921 3922 Visited.clear(); 3923 if (DependingInstructions.size() != 1) { 3924 DependingInstructions.clear(); 3925 return false; 3926 } 3927 3928 Retain = dyn_cast_or_null<CallInst>(*DependingInstructions.begin()); 3929 DependingInstructions.clear(); 3930 3931 if (!Retain || 3932 GetBasicInstructionClass(Retain) != IC_Retain || 3933 GetObjCArg(Retain) != Arg) 3934 return false; 3935 3936 Changed = true; 3937 ++NumPeeps; 3938 3939 if (Class == IC_AutoreleaseRV) 3940 Retain->setCalledFunction(getRetainAutoreleaseRVCallee(F.getParent())); 3941 else 3942 Retain->setCalledFunction(getRetainAutoreleaseCallee(F.getParent())); 3943 3944 EraseInstruction(Autorelease); 3945 return true; 3946 } 3947 3948 /// ContractRelease - Attempt to merge an objc_release with a store, load, and 3949 /// objc_retain to form an objc_storeStrong. This can be a little tricky because 3950 /// the instructions don't always appear in order, and there may be unrelated 3951 /// intervening instructions. 3952 void ObjCARCContract::ContractRelease(Instruction *Release, 3953 inst_iterator &Iter) { 3954 LoadInst *Load = dyn_cast<LoadInst>(GetObjCArg(Release)); 3955 if (!Load || !Load->isSimple()) return; 3956 3957 // For now, require everything to be in one basic block. 3958 BasicBlock *BB = Release->getParent(); 3959 if (Load->getParent() != BB) return; 3960 3961 // Walk down to find the store. 3962 BasicBlock::iterator I = Load, End = BB->end(); 3963 ++I; 3964 AliasAnalysis::Location Loc = AA->getLocation(Load); 3965 while (I != End && 3966 (&*I == Release || 3967 IsRetain(GetBasicInstructionClass(I)) || 3968 !(AA->getModRefInfo(I, Loc) & AliasAnalysis::Mod))) 3969 ++I; 3970 StoreInst *Store = dyn_cast<StoreInst>(I); 3971 if (!Store || !Store->isSimple()) return; 3972 if (Store->getPointerOperand() != Loc.Ptr) return; 3973 3974 Value *New = StripPointerCastsAndObjCCalls(Store->getValueOperand()); 3975 3976 // Walk up to find the retain. 3977 I = Store; 3978 BasicBlock::iterator Begin = BB->begin(); 3979 while (I != Begin && GetBasicInstructionClass(I) != IC_Retain) 3980 --I; 3981 Instruction *Retain = I; 3982 if (GetBasicInstructionClass(Retain) != IC_Retain) return; 3983 if (GetObjCArg(Retain) != New) return; 3984 3985 Changed = true; 3986 ++NumStoreStrongs; 3987 3988 LLVMContext &C = Release->getContext(); 3989 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); 3990 Type *I8XX = PointerType::getUnqual(I8X); 3991 3992 Value *Args[] = { Load->getPointerOperand(), New }; 3993 if (Args[0]->getType() != I8XX) 3994 Args[0] = new BitCastInst(Args[0], I8XX, "", Store); 3995 if (Args[1]->getType() != I8X) 3996 Args[1] = new BitCastInst(Args[1], I8X, "", Store); 3997 CallInst *StoreStrong = 3998 CallInst::Create(getStoreStrongCallee(BB->getParent()->getParent()), 3999 Args, "", Store); 4000 StoreStrong->setDoesNotThrow(); 4001 StoreStrong->setDebugLoc(Store->getDebugLoc()); 4002 4003 // We can't set the tail flag yet, because we haven't yet determined 4004 // whether there are any escaping allocas. Remember this call, so that 4005 // we can set the tail flag once we know it's safe. 4006 StoreStrongCalls.insert(StoreStrong); 4007 4008 if (&*Iter == Store) ++Iter; 4009 Store->eraseFromParent(); 4010 Release->eraseFromParent(); 4011 EraseInstruction(Retain); 4012 if (Load->use_empty()) 4013 Load->eraseFromParent(); 4014 } 4015 4016 bool ObjCARCContract::doInitialization(Module &M) { 4017 // If nothing in the Module uses ARC, don't do anything. 4018 Run = ModuleHasARC(M); 4019 if (!Run) 4020 return false; 4021 4022 // These are initialized lazily. 4023 StoreStrongCallee = 0; 4024 RetainAutoreleaseCallee = 0; 4025 RetainAutoreleaseRVCallee = 0; 4026 4027 // Initialize RetainRVMarker. 4028 RetainRVMarker = 0; 4029 if (NamedMDNode *NMD = 4030 M.getNamedMetadata("clang.arc.retainAutoreleasedReturnValueMarker")) 4031 if (NMD->getNumOperands() == 1) { 4032 const MDNode *N = NMD->getOperand(0); 4033 if (N->getNumOperands() == 1) 4034 if (const MDString *S = dyn_cast<MDString>(N->getOperand(0))) 4035 RetainRVMarker = S; 4036 } 4037 4038 return false; 4039 } 4040 4041 bool ObjCARCContract::runOnFunction(Function &F) { 4042 if (!EnableARCOpts) 4043 return false; 4044 4045 // If nothing in the Module uses ARC, don't do anything. 4046 if (!Run) 4047 return false; 4048 4049 Changed = false; 4050 AA = &getAnalysis<AliasAnalysis>(); 4051 DT = &getAnalysis<DominatorTree>(); 4052 4053 PA.setAA(&getAnalysis<AliasAnalysis>()); 4054 4055 // Track whether it's ok to mark objc_storeStrong calls with the "tail" 4056 // keyword. Be conservative if the function has variadic arguments. 4057 // It seems that functions which "return twice" are also unsafe for the 4058 // "tail" argument, because they are setjmp, which could need to 4059 // return to an earlier stack state. 4060 bool TailOkForStoreStrongs = !F.isVarArg() && !F.callsFunctionThatReturnsTwice(); 4061 4062 // For ObjC library calls which return their argument, replace uses of the 4063 // argument with uses of the call return value, if it dominates the use. This 4064 // reduces register pressure. 4065 SmallPtrSet<Instruction *, 4> DependingInstructions; 4066 SmallPtrSet<const BasicBlock *, 4> Visited; 4067 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) { 4068 Instruction *Inst = &*I++; 4069 4070 // Only these library routines return their argument. In particular, 4071 // objc_retainBlock does not necessarily return its argument. 4072 InstructionClass Class = GetBasicInstructionClass(Inst); 4073 switch (Class) { 4074 case IC_Retain: 4075 case IC_FusedRetainAutorelease: 4076 case IC_FusedRetainAutoreleaseRV: 4077 break; 4078 case IC_Autorelease: 4079 case IC_AutoreleaseRV: 4080 if (ContractAutorelease(F, Inst, Class, DependingInstructions, Visited)) 4081 continue; 4082 break; 4083 case IC_RetainRV: { 4084 // If we're compiling for a target which needs a special inline-asm 4085 // marker to do the retainAutoreleasedReturnValue optimization, 4086 // insert it now. 4087 if (!RetainRVMarker) 4088 break; 4089 BasicBlock::iterator BBI = Inst; 4090 --BBI; 4091 while (isNoopInstruction(BBI)) --BBI; 4092 if (&*BBI == GetObjCArg(Inst)) { 4093 Changed = true; 4094 InlineAsm *IA = 4095 InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()), 4096 /*isVarArg=*/false), 4097 RetainRVMarker->getString(), 4098 /*Constraints=*/"", /*hasSideEffects=*/true); 4099 CallInst::Create(IA, "", Inst); 4100 } 4101 break; 4102 } 4103 case IC_InitWeak: { 4104 // objc_initWeak(p, null) => *p = null 4105 CallInst *CI = cast<CallInst>(Inst); 4106 if (isNullOrUndef(CI->getArgOperand(1))) { 4107 Value *Null = 4108 ConstantPointerNull::get(cast<PointerType>(CI->getType())); 4109 Changed = true; 4110 new StoreInst(Null, CI->getArgOperand(0), CI); 4111 CI->replaceAllUsesWith(Null); 4112 CI->eraseFromParent(); 4113 } 4114 continue; 4115 } 4116 case IC_Release: 4117 ContractRelease(Inst, I); 4118 continue; 4119 case IC_User: 4120 // Be conservative if the function has any alloca instructions. 4121 // Technically we only care about escaping alloca instructions, 4122 // but this is sufficient to handle some interesting cases. 4123 if (isa<AllocaInst>(Inst)) 4124 TailOkForStoreStrongs = false; 4125 continue; 4126 default: 4127 continue; 4128 } 4129 4130 // Don't use GetObjCArg because we don't want to look through bitcasts 4131 // and such; to do the replacement, the argument must have type i8*. 4132 const Value *Arg = cast<CallInst>(Inst)->getArgOperand(0); 4133 for (;;) { 4134 // If we're compiling bugpointed code, don't get in trouble. 4135 if (!isa<Instruction>(Arg) && !isa<Argument>(Arg)) 4136 break; 4137 // Look through the uses of the pointer. 4138 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end(); 4139 UI != UE; ) { 4140 Use &U = UI.getUse(); 4141 unsigned OperandNo = UI.getOperandNo(); 4142 ++UI; // Increment UI now, because we may unlink its element. 4143 4144 // If the call's return value dominates a use of the call's argument 4145 // value, rewrite the use to use the return value. We check for 4146 // reachability here because an unreachable call is considered to 4147 // trivially dominate itself, which would lead us to rewriting its 4148 // argument in terms of its return value, which would lead to 4149 // infinite loops in GetObjCArg. 4150 if (DT->isReachableFromEntry(U) && 4151 DT->dominates(Inst, U)) { 4152 Changed = true; 4153 Instruction *Replacement = Inst; 4154 Type *UseTy = U.get()->getType(); 4155 if (PHINode *PHI = dyn_cast<PHINode>(U.getUser())) { 4156 // For PHI nodes, insert the bitcast in the predecessor block. 4157 unsigned ValNo = 4158 PHINode::getIncomingValueNumForOperand(OperandNo); 4159 BasicBlock *BB = 4160 PHI->getIncomingBlock(ValNo); 4161 if (Replacement->getType() != UseTy) 4162 Replacement = new BitCastInst(Replacement, UseTy, "", 4163 &BB->back()); 4164 // While we're here, rewrite all edges for this PHI, rather 4165 // than just one use at a time, to minimize the number of 4166 // bitcasts we emit. 4167 for (unsigned i = 0, e = PHI->getNumIncomingValues(); 4168 i != e; ++i) 4169 if (PHI->getIncomingBlock(i) == BB) { 4170 // Keep the UI iterator valid. 4171 if (&PHI->getOperandUse( 4172 PHINode::getOperandNumForIncomingValue(i)) == 4173 &UI.getUse()) 4174 ++UI; 4175 PHI->setIncomingValue(i, Replacement); 4176 } 4177 } else { 4178 if (Replacement->getType() != UseTy) 4179 Replacement = new BitCastInst(Replacement, UseTy, "", 4180 cast<Instruction>(U.getUser())); 4181 U.set(Replacement); 4182 } 4183 } 4184 } 4185 4186 // If Arg is a no-op casted pointer, strip one level of casts and 4187 // iterate. 4188 if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg)) 4189 Arg = BI->getOperand(0); 4190 else if (isa<GEPOperator>(Arg) && 4191 cast<GEPOperator>(Arg)->hasAllZeroIndices()) 4192 Arg = cast<GEPOperator>(Arg)->getPointerOperand(); 4193 else if (isa<GlobalAlias>(Arg) && 4194 !cast<GlobalAlias>(Arg)->mayBeOverridden()) 4195 Arg = cast<GlobalAlias>(Arg)->getAliasee(); 4196 else 4197 break; 4198 } 4199 } 4200 4201 // If this function has no escaping allocas or suspicious vararg usage, 4202 // objc_storeStrong calls can be marked with the "tail" keyword. 4203 if (TailOkForStoreStrongs) 4204 for (DenseSet<CallInst *>::iterator I = StoreStrongCalls.begin(), 4205 E = StoreStrongCalls.end(); I != E; ++I) 4206 (*I)->setTailCall(); 4207 StoreStrongCalls.clear(); 4208 4209 return Changed; 4210 } 4211