1 //===- ObjCARC.cpp - ObjC ARC Optimization --------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
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
20 // This file also defines a simple ARC-aware AliasAnalysis.
22 // WARNING: This file knows about certain library functions. It recognizes them
23 // by name, and hardwires knowedge of their semantics.
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.
29 //===----------------------------------------------------------------------===//
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"
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));
49 //===----------------------------------------------------------------------===//
51 //===----------------------------------------------------------------------===//
54 /// MapVector - An associative container with fast insertion-order
55 /// (deterministic) iteration over its elements. Plus the special
57 template<class KeyT, class ValueT>
59 /// Map - Map keys to indices in Vector.
60 typedef DenseMap<KeyT, size_t> MapTy;
63 /// Vector - Keys and values.
64 typedef std::vector<std::pair<KeyT, ValueT> > VectorTy;
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(); }
77 assert(Vector.size() >= Map.size()); // May differ due to blotting.
78 for (typename MapTy::const_iterator I = Map.begin(), E = Map.end();
80 assert(I->second < Vector.size());
81 assert(Vector[I->second].first == I->first);
83 for (typename VectorTy::const_iterator I = Vector.begin(),
84 E = Vector.end(); I != E; ++I)
86 (Map.count(I->first) &&
87 Map[I->first] == size_t(I - Vector.begin())));
91 ValueT &operator[](KeyT Arg) {
92 std::pair<typename MapTy::iterator, bool> Pair =
93 Map.insert(std::make_pair(Arg, size_t(0)));
95 Pair.first->second = Vector.size();
96 Vector.push_back(std::make_pair(Arg, ValueT()));
97 return Vector.back().second;
99 return Vector[Pair.first->second].second;
102 std::pair<iterator, bool>
103 insert(const std::pair<KeyT, ValueT> &InsertPair) {
104 std::pair<typename MapTy::iterator, bool> Pair =
105 Map.insert(std::make_pair(InsertPair.first, size_t(0)));
107 Pair.first->second = Vector.size();
108 Vector.push_back(InsertPair);
109 return std::make_pair(llvm::prior(Vector.end()), true);
111 return std::make_pair(Vector.begin() + Pair.first->second, false);
114 const_iterator find(KeyT Key) const {
115 typename MapTy::const_iterator It = Map.find(Key);
116 if (It == Map.end()) return Vector.end();
117 return Vector.begin() + It->second;
120 /// blot - This is similar to erase, but instead of removing the element
121 /// from the vector, it just zeros out the key in the vector. This leaves
122 /// iterators intact, but clients must be prepared for zeroed-out keys when
124 void blot(KeyT Key) {
125 typename MapTy::iterator It = Map.find(Key);
126 if (It == Map.end()) return;
127 Vector[It->second].first = KeyT();
138 //===----------------------------------------------------------------------===//
140 //===----------------------------------------------------------------------===//
143 /// InstructionClass - A simple classification for instructions.
144 enum InstructionClass {
145 IC_Retain, ///< objc_retain
146 IC_RetainRV, ///< objc_retainAutoreleasedReturnValue
147 IC_RetainBlock, ///< objc_retainBlock
148 IC_Release, ///< objc_release
149 IC_Autorelease, ///< objc_autorelease
150 IC_AutoreleaseRV, ///< objc_autoreleaseReturnValue
151 IC_AutoreleasepoolPush, ///< objc_autoreleasePoolPush
152 IC_AutoreleasepoolPop, ///< objc_autoreleasePoolPop
153 IC_NoopCast, ///< objc_retainedObject, etc.
154 IC_FusedRetainAutorelease, ///< objc_retainAutorelease
155 IC_FusedRetainAutoreleaseRV, ///< objc_retainAutoreleaseReturnValue
156 IC_LoadWeakRetained, ///< objc_loadWeakRetained (primitive)
157 IC_StoreWeak, ///< objc_storeWeak (primitive)
158 IC_InitWeak, ///< objc_initWeak (derived)
159 IC_LoadWeak, ///< objc_loadWeak (derived)
160 IC_MoveWeak, ///< objc_moveWeak (derived)
161 IC_CopyWeak, ///< objc_copyWeak (derived)
162 IC_DestroyWeak, ///< objc_destroyWeak (derived)
163 IC_CallOrUser, ///< could call objc_release and/or "use" pointers
164 IC_Call, ///< could call objc_release
165 IC_User, ///< could "use" a pointer
166 IC_None ///< anything else
170 /// IsPotentialUse - Test whether the given value is possible a
171 /// reference-counted pointer.
172 static bool IsPotentialUse(const Value *Op) {
173 // Pointers to static or stack storage are not reference-counted pointers.
174 if (isa<Constant>(Op) || isa<AllocaInst>(Op))
176 // Special arguments are not reference-counted.
177 if (const Argument *Arg = dyn_cast<Argument>(Op))
178 if (Arg->hasByValAttr() ||
179 Arg->hasNestAttr() ||
180 Arg->hasStructRetAttr())
182 // Only consider values with pointer types, and not function pointers.
183 PointerType *Ty = dyn_cast<PointerType>(Op->getType());
184 if (!Ty || isa<FunctionType>(Ty->getElementType()))
186 // Conservatively assume anything else is a potential use.
190 /// GetCallSiteClass - Helper for GetInstructionClass. Determines what kind
191 /// of construct CS is.
192 static InstructionClass GetCallSiteClass(ImmutableCallSite CS) {
193 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
195 if (IsPotentialUse(*I))
196 return CS.onlyReadsMemory() ? IC_User : IC_CallOrUser;
198 return CS.onlyReadsMemory() ? IC_None : IC_Call;
201 /// GetFunctionClass - Determine if F is one of the special known Functions.
202 /// If it isn't, return IC_CallOrUser.
203 static InstructionClass GetFunctionClass(const Function *F) {
204 Function::const_arg_iterator AI = F->arg_begin(), AE = F->arg_end();
208 return StringSwitch<InstructionClass>(F->getName())
209 .Case("objc_autoreleasePoolPush", IC_AutoreleasepoolPush)
210 .Default(IC_CallOrUser);
213 const Argument *A0 = AI++;
215 // Argument is a pointer.
216 if (PointerType *PTy = dyn_cast<PointerType>(A0->getType())) {
217 Type *ETy = PTy->getElementType();
219 if (ETy->isIntegerTy(8))
220 return StringSwitch<InstructionClass>(F->getName())
221 .Case("objc_retain", IC_Retain)
222 .Case("objc_retainAutoreleasedReturnValue", IC_RetainRV)
223 .Case("objc_retainBlock", IC_RetainBlock)
224 .Case("objc_release", IC_Release)
225 .Case("objc_autorelease", IC_Autorelease)
226 .Case("objc_autoreleaseReturnValue", IC_AutoreleaseRV)
227 .Case("objc_autoreleasePoolPop", IC_AutoreleasepoolPop)
228 .Case("objc_retainedObject", IC_NoopCast)
229 .Case("objc_unretainedObject", IC_NoopCast)
230 .Case("objc_unretainedPointer", IC_NoopCast)
231 .Case("objc_retain_autorelease", IC_FusedRetainAutorelease)
232 .Case("objc_retainAutorelease", IC_FusedRetainAutorelease)
233 .Case("objc_retainAutoreleaseReturnValue",IC_FusedRetainAutoreleaseRV)
234 .Default(IC_CallOrUser);
237 if (PointerType *Pte = dyn_cast<PointerType>(ETy))
238 if (Pte->getElementType()->isIntegerTy(8))
239 return StringSwitch<InstructionClass>(F->getName())
240 .Case("objc_loadWeakRetained", IC_LoadWeakRetained)
241 .Case("objc_loadWeak", IC_LoadWeak)
242 .Case("objc_destroyWeak", IC_DestroyWeak)
243 .Default(IC_CallOrUser);
246 // Two arguments, first is i8**.
247 const Argument *A1 = AI++;
249 if (PointerType *PTy = dyn_cast<PointerType>(A0->getType()))
250 if (PointerType *Pte = dyn_cast<PointerType>(PTy->getElementType()))
251 if (Pte->getElementType()->isIntegerTy(8))
252 if (PointerType *PTy1 = dyn_cast<PointerType>(A1->getType())) {
253 Type *ETy1 = PTy1->getElementType();
254 // Second argument is i8*
255 if (ETy1->isIntegerTy(8))
256 return StringSwitch<InstructionClass>(F->getName())
257 .Case("objc_storeWeak", IC_StoreWeak)
258 .Case("objc_initWeak", IC_InitWeak)
259 .Default(IC_CallOrUser);
260 // Second argument is i8**.
261 if (PointerType *Pte1 = dyn_cast<PointerType>(ETy1))
262 if (Pte1->getElementType()->isIntegerTy(8))
263 return StringSwitch<InstructionClass>(F->getName())
264 .Case("objc_moveWeak", IC_MoveWeak)
265 .Case("objc_copyWeak", IC_CopyWeak)
266 .Default(IC_CallOrUser);
270 return IC_CallOrUser;
273 /// GetInstructionClass - Determine what kind of construct V is.
274 static InstructionClass GetInstructionClass(const Value *V) {
275 if (const Instruction *I = dyn_cast<Instruction>(V)) {
276 // Any instruction other than bitcast and gep with a pointer operand have a
277 // use of an objc pointer. Bitcasts, GEPs, Selects, PHIs transfer a pointer
278 // to a subsequent use, rather than using it themselves, in this sense.
279 // As a short cut, several other opcodes are known to have no pointer
280 // operands of interest. And ret is never followed by a release, so it's
281 // not interesting to examine.
282 switch (I->getOpcode()) {
283 case Instruction::Call: {
284 const CallInst *CI = cast<CallInst>(I);
285 // Check for calls to special functions.
286 if (const Function *F = CI->getCalledFunction()) {
287 InstructionClass Class = GetFunctionClass(F);
288 if (Class != IC_CallOrUser)
291 // None of the intrinsic functions do objc_release. For intrinsics, the
292 // only question is whether or not they may be users.
293 switch (F->getIntrinsicID()) {
295 case Intrinsic::bswap: case Intrinsic::ctpop:
296 case Intrinsic::ctlz: case Intrinsic::cttz:
297 case Intrinsic::returnaddress: case Intrinsic::frameaddress:
298 case Intrinsic::stacksave: case Intrinsic::stackrestore:
299 case Intrinsic::vastart: case Intrinsic::vacopy: case Intrinsic::vaend:
300 // Don't let dbg info affect our results.
301 case Intrinsic::dbg_declare: case Intrinsic::dbg_value:
302 // Short cut: Some intrinsics obviously don't use ObjC pointers.
305 for (Function::const_arg_iterator AI = F->arg_begin(),
306 AE = F->arg_end(); AI != AE; ++AI)
307 if (IsPotentialUse(AI))
312 return GetCallSiteClass(CI);
314 case Instruction::Invoke:
315 return GetCallSiteClass(cast<InvokeInst>(I));
316 case Instruction::BitCast:
317 case Instruction::GetElementPtr:
318 case Instruction::Select: case Instruction::PHI:
319 case Instruction::Ret: case Instruction::Br:
320 case Instruction::Switch: case Instruction::IndirectBr:
321 case Instruction::Alloca: case Instruction::VAArg:
322 case Instruction::Add: case Instruction::FAdd:
323 case Instruction::Sub: case Instruction::FSub:
324 case Instruction::Mul: case Instruction::FMul:
325 case Instruction::SDiv: case Instruction::UDiv: case Instruction::FDiv:
326 case Instruction::SRem: case Instruction::URem: case Instruction::FRem:
327 case Instruction::Shl: case Instruction::LShr: case Instruction::AShr:
328 case Instruction::And: case Instruction::Or: case Instruction::Xor:
329 case Instruction::SExt: case Instruction::ZExt: case Instruction::Trunc:
330 case Instruction::IntToPtr: case Instruction::FCmp:
331 case Instruction::FPTrunc: case Instruction::FPExt:
332 case Instruction::FPToUI: case Instruction::FPToSI:
333 case Instruction::UIToFP: case Instruction::SIToFP:
334 case Instruction::InsertElement: case Instruction::ExtractElement:
335 case Instruction::ShuffleVector:
336 case Instruction::ExtractValue:
338 case Instruction::ICmp:
339 // Comparing a pointer with null, or any other constant, isn't an
340 // interesting use, because we don't care what the pointer points to, or
341 // about the values of any other dynamic reference-counted pointers.
342 if (IsPotentialUse(I->getOperand(1)))
346 // For anything else, check all the operands.
347 for (User::const_op_iterator OI = I->op_begin(), OE = I->op_end();
349 if (IsPotentialUse(*OI))
354 // Otherwise, it's totally inert for ARC purposes.
358 /// GetBasicInstructionClass - Determine what kind of construct V is. This is
359 /// similar to GetInstructionClass except that it only detects objc runtine
360 /// calls. This allows it to be faster.
361 static InstructionClass GetBasicInstructionClass(const Value *V) {
362 if (const CallInst *CI = dyn_cast<CallInst>(V)) {
363 if (const Function *F = CI->getCalledFunction())
364 return GetFunctionClass(F);
365 // Otherwise, be conservative.
366 return IC_CallOrUser;
369 // Otherwise, be conservative.
373 /// IsRetain - Test if the the given class is objc_retain or
375 static bool IsRetain(InstructionClass Class) {
376 return Class == IC_Retain ||
377 Class == IC_RetainRV;
380 /// IsAutorelease - Test if the the given class is objc_autorelease or
382 static bool IsAutorelease(InstructionClass Class) {
383 return Class == IC_Autorelease ||
384 Class == IC_AutoreleaseRV;
387 /// IsForwarding - Test if the given class represents instructions which return
388 /// their argument verbatim.
389 static bool IsForwarding(InstructionClass Class) {
390 // objc_retainBlock technically doesn't always return its argument
391 // verbatim, but it doesn't matter for our purposes here.
392 return Class == IC_Retain ||
393 Class == IC_RetainRV ||
394 Class == IC_Autorelease ||
395 Class == IC_AutoreleaseRV ||
396 Class == IC_RetainBlock ||
397 Class == IC_NoopCast;
400 /// IsNoopOnNull - Test if the given class represents instructions which do
401 /// nothing if passed a null pointer.
402 static bool IsNoopOnNull(InstructionClass Class) {
403 return Class == IC_Retain ||
404 Class == IC_RetainRV ||
405 Class == IC_Release ||
406 Class == IC_Autorelease ||
407 Class == IC_AutoreleaseRV ||
408 Class == IC_RetainBlock;
411 /// IsAlwaysTail - Test if the given class represents instructions which are
412 /// always safe to mark with the "tail" keyword.
413 static bool IsAlwaysTail(InstructionClass Class) {
414 // IC_RetainBlock may be given a stack argument.
415 return Class == IC_Retain ||
416 Class == IC_RetainRV ||
417 Class == IC_Autorelease ||
418 Class == IC_AutoreleaseRV;
421 /// IsNoThrow - Test if the given class represents instructions which are always
422 /// safe to mark with the nounwind attribute..
423 static bool IsNoThrow(InstructionClass Class) {
424 return Class == IC_Retain ||
425 Class == IC_RetainRV ||
426 Class == IC_RetainBlock ||
427 Class == IC_Release ||
428 Class == IC_Autorelease ||
429 Class == IC_AutoreleaseRV ||
430 Class == IC_AutoreleasepoolPush ||
431 Class == IC_AutoreleasepoolPop;
434 /// EraseInstruction - Erase the given instruction. ObjC calls return their
435 /// argument verbatim, so if it's such a call and the return value has users,
436 /// replace them with the argument value.
437 static void EraseInstruction(Instruction *CI) {
438 Value *OldArg = cast<CallInst>(CI)->getArgOperand(0);
440 bool Unused = CI->use_empty();
443 // Replace the return value with the argument.
444 assert(IsForwarding(GetBasicInstructionClass(CI)) &&
445 "Can't delete non-forwarding instruction with users!");
446 CI->replaceAllUsesWith(OldArg);
449 CI->eraseFromParent();
452 RecursivelyDeleteTriviallyDeadInstructions(OldArg);
455 /// GetUnderlyingObjCPtr - This is a wrapper around getUnderlyingObject which
456 /// also knows how to look through objc_retain and objc_autorelease calls, which
457 /// we know to return their argument verbatim.
458 static const Value *GetUnderlyingObjCPtr(const Value *V) {
460 V = GetUnderlyingObject(V);
461 if (!IsForwarding(GetBasicInstructionClass(V)))
463 V = cast<CallInst>(V)->getArgOperand(0);
469 /// StripPointerCastsAndObjCCalls - This is a wrapper around
470 /// Value::stripPointerCasts which also knows how to look through objc_retain
471 /// and objc_autorelease calls, which we know to return their argument verbatim.
472 static const Value *StripPointerCastsAndObjCCalls(const Value *V) {
474 V = V->stripPointerCasts();
475 if (!IsForwarding(GetBasicInstructionClass(V)))
477 V = cast<CallInst>(V)->getArgOperand(0);
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 Value *StripPointerCastsAndObjCCalls(Value *V) {
487 V = V->stripPointerCasts();
488 if (!IsForwarding(GetBasicInstructionClass(V)))
490 V = cast<CallInst>(V)->getArgOperand(0);
495 /// GetObjCArg - Assuming the given instruction is one of the special calls such
496 /// as objc_retain or objc_release, return the argument value, stripped of no-op
497 /// casts and forwarding calls.
498 static Value *GetObjCArg(Value *Inst) {
499 return StripPointerCastsAndObjCCalls(cast<CallInst>(Inst)->getArgOperand(0));
502 /// IsObjCIdentifiedObject - This is similar to AliasAnalysis'
503 /// isObjCIdentifiedObject, except that it uses special knowledge of
504 /// ObjC conventions...
505 static bool IsObjCIdentifiedObject(const Value *V) {
506 // Assume that call results and arguments have their own "provenance".
507 // Constants (including GlobalVariables) and Allocas are never
508 // reference-counted.
509 if (isa<CallInst>(V) || isa<InvokeInst>(V) ||
510 isa<Argument>(V) || isa<Constant>(V) ||
514 if (const LoadInst *LI = dyn_cast<LoadInst>(V)) {
515 const Value *Pointer =
516 StripPointerCastsAndObjCCalls(LI->getPointerOperand());
517 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Pointer)) {
518 // A constant pointer can't be pointing to an object on the heap. It may
519 // be reference-counted, but it won't be deleted.
520 if (GV->isConstant())
522 StringRef Name = GV->getName();
523 // These special variables are known to hold values which are not
524 // reference-counted pointers.
525 if (Name.startswith("\01L_OBJC_SELECTOR_REFERENCES_") ||
526 Name.startswith("\01L_OBJC_CLASSLIST_REFERENCES_") ||
527 Name.startswith("\01L_OBJC_CLASSLIST_SUP_REFS_$_") ||
528 Name.startswith("\01L_OBJC_METH_VAR_NAME_") ||
529 Name.startswith("\01l_objc_msgSend_fixup_"))
537 /// FindSingleUseIdentifiedObject - This is similar to
538 /// StripPointerCastsAndObjCCalls but it stops as soon as it finds a value
539 /// with multiple uses.
540 static const Value *FindSingleUseIdentifiedObject(const Value *Arg) {
541 if (Arg->hasOneUse()) {
542 if (const BitCastInst *BC = dyn_cast<BitCastInst>(Arg))
543 return FindSingleUseIdentifiedObject(BC->getOperand(0));
544 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Arg))
545 if (GEP->hasAllZeroIndices())
546 return FindSingleUseIdentifiedObject(GEP->getPointerOperand());
547 if (IsForwarding(GetBasicInstructionClass(Arg)))
548 return FindSingleUseIdentifiedObject(
549 cast<CallInst>(Arg)->getArgOperand(0));
550 if (!IsObjCIdentifiedObject(Arg))
555 // If we found an identifiable object but it has multiple uses, but they
556 // are trivial uses, we can still consider this to be a single-use
558 if (IsObjCIdentifiedObject(Arg)) {
559 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
562 if (!U->use_empty() || StripPointerCastsAndObjCCalls(U) != Arg)
572 /// ModuleHasARC - Test if the given module looks interesting to run ARC
574 static bool ModuleHasARC(const Module &M) {
576 M.getNamedValue("objc_retain") ||
577 M.getNamedValue("objc_release") ||
578 M.getNamedValue("objc_autorelease") ||
579 M.getNamedValue("objc_retainAutoreleasedReturnValue") ||
580 M.getNamedValue("objc_retainBlock") ||
581 M.getNamedValue("objc_autoreleaseReturnValue") ||
582 M.getNamedValue("objc_autoreleasePoolPush") ||
583 M.getNamedValue("objc_loadWeakRetained") ||
584 M.getNamedValue("objc_loadWeak") ||
585 M.getNamedValue("objc_destroyWeak") ||
586 M.getNamedValue("objc_storeWeak") ||
587 M.getNamedValue("objc_initWeak") ||
588 M.getNamedValue("objc_moveWeak") ||
589 M.getNamedValue("objc_copyWeak") ||
590 M.getNamedValue("objc_retainedObject") ||
591 M.getNamedValue("objc_unretainedObject") ||
592 M.getNamedValue("objc_unretainedPointer");
595 //===----------------------------------------------------------------------===//
596 // ARC AliasAnalysis.
597 //===----------------------------------------------------------------------===//
599 #include "llvm/Pass.h"
600 #include "llvm/Analysis/AliasAnalysis.h"
601 #include "llvm/Analysis/Passes.h"
604 /// ObjCARCAliasAnalysis - This is a simple alias analysis
605 /// implementation that uses knowledge of ARC constructs to answer queries.
607 /// TODO: This class could be generalized to know about other ObjC-specific
608 /// tricks. Such as knowing that ivars in the non-fragile ABI are non-aliasing
609 /// even though their offsets are dynamic.
610 class ObjCARCAliasAnalysis : public ImmutablePass,
611 public AliasAnalysis {
613 static char ID; // Class identification, replacement for typeinfo
614 ObjCARCAliasAnalysis() : ImmutablePass(ID) {
615 initializeObjCARCAliasAnalysisPass(*PassRegistry::getPassRegistry());
619 virtual void initializePass() {
620 InitializeAliasAnalysis(this);
623 /// getAdjustedAnalysisPointer - This method is used when a pass implements
624 /// an analysis interface through multiple inheritance. If needed, it
625 /// should override this to adjust the this pointer as needed for the
626 /// specified pass info.
627 virtual void *getAdjustedAnalysisPointer(const void *PI) {
628 if (PI == &AliasAnalysis::ID)
629 return (AliasAnalysis*)this;
633 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
634 virtual AliasResult alias(const Location &LocA, const Location &LocB);
635 virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
636 virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
637 virtual ModRefBehavior getModRefBehavior(const Function *F);
638 virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
639 const Location &Loc);
640 virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
641 ImmutableCallSite CS2);
643 } // End of anonymous namespace
645 // Register this pass...
646 char ObjCARCAliasAnalysis::ID = 0;
647 INITIALIZE_AG_PASS(ObjCARCAliasAnalysis, AliasAnalysis, "objc-arc-aa",
648 "ObjC-ARC-Based Alias Analysis", false, true, false)
650 ImmutablePass *llvm::createObjCARCAliasAnalysisPass() {
651 return new ObjCARCAliasAnalysis();
655 ObjCARCAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
656 AU.setPreservesAll();
657 AliasAnalysis::getAnalysisUsage(AU);
660 AliasAnalysis::AliasResult
661 ObjCARCAliasAnalysis::alias(const Location &LocA, const Location &LocB) {
663 return AliasAnalysis::alias(LocA, LocB);
665 // First, strip off no-ops, including ObjC-specific no-ops, and try making a
666 // precise alias query.
667 const Value *SA = StripPointerCastsAndObjCCalls(LocA.Ptr);
668 const Value *SB = StripPointerCastsAndObjCCalls(LocB.Ptr);
670 AliasAnalysis::alias(Location(SA, LocA.Size, LocA.TBAATag),
671 Location(SB, LocB.Size, LocB.TBAATag));
672 if (Result != MayAlias)
675 // If that failed, climb to the underlying object, including climbing through
676 // ObjC-specific no-ops, and try making an imprecise alias query.
677 const Value *UA = GetUnderlyingObjCPtr(SA);
678 const Value *UB = GetUnderlyingObjCPtr(SB);
679 if (UA != SA || UB != SB) {
680 Result = AliasAnalysis::alias(Location(UA), Location(UB));
681 // We can't use MustAlias or PartialAlias results here because
682 // GetUnderlyingObjCPtr may return an offsetted pointer value.
683 if (Result == NoAlias)
687 // If that failed, fail. We don't need to chain here, since that's covered
688 // by the earlier precise query.
693 ObjCARCAliasAnalysis::pointsToConstantMemory(const Location &Loc,
696 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
698 // First, strip off no-ops, including ObjC-specific no-ops, and try making
699 // a precise alias query.
700 const Value *S = StripPointerCastsAndObjCCalls(Loc.Ptr);
701 if (AliasAnalysis::pointsToConstantMemory(Location(S, Loc.Size, Loc.TBAATag),
705 // If that failed, climb to the underlying object, including climbing through
706 // ObjC-specific no-ops, and try making an imprecise alias query.
707 const Value *U = GetUnderlyingObjCPtr(S);
709 return AliasAnalysis::pointsToConstantMemory(Location(U), OrLocal);
711 // If that failed, fail. We don't need to chain here, since that's covered
712 // by the earlier precise query.
716 AliasAnalysis::ModRefBehavior
717 ObjCARCAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
718 // We have nothing to do. Just chain to the next AliasAnalysis.
719 return AliasAnalysis::getModRefBehavior(CS);
722 AliasAnalysis::ModRefBehavior
723 ObjCARCAliasAnalysis::getModRefBehavior(const Function *F) {
725 return AliasAnalysis::getModRefBehavior(F);
727 switch (GetFunctionClass(F)) {
729 return DoesNotAccessMemory;
734 return AliasAnalysis::getModRefBehavior(F);
737 AliasAnalysis::ModRefResult
738 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS, const Location &Loc) {
740 return AliasAnalysis::getModRefInfo(CS, Loc);
742 switch (GetBasicInstructionClass(CS.getInstruction())) {
747 case IC_AutoreleaseRV:
749 case IC_AutoreleasepoolPush:
750 case IC_FusedRetainAutorelease:
751 case IC_FusedRetainAutoreleaseRV:
752 // These functions don't access any memory visible to the compiler.
758 return AliasAnalysis::getModRefInfo(CS, Loc);
761 AliasAnalysis::ModRefResult
762 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS1,
763 ImmutableCallSite CS2) {
764 // TODO: Theoretically we could check for dependencies between objc_* calls
765 // and OnlyAccessesArgumentPointees calls or other well-behaved calls.
766 return AliasAnalysis::getModRefInfo(CS1, CS2);
769 //===----------------------------------------------------------------------===//
771 //===----------------------------------------------------------------------===//
773 #include "llvm/Support/InstIterator.h"
774 #include "llvm/Transforms/Scalar.h"
777 /// ObjCARCExpand - Early ARC transformations.
778 class ObjCARCExpand : public FunctionPass {
779 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
780 virtual bool doInitialization(Module &M);
781 virtual bool runOnFunction(Function &F);
783 /// Run - A flag indicating whether this optimization pass should run.
788 ObjCARCExpand() : FunctionPass(ID) {
789 initializeObjCARCExpandPass(*PassRegistry::getPassRegistry());
794 char ObjCARCExpand::ID = 0;
795 INITIALIZE_PASS(ObjCARCExpand,
796 "objc-arc-expand", "ObjC ARC expansion", false, false)
798 Pass *llvm::createObjCARCExpandPass() {
799 return new ObjCARCExpand();
802 void ObjCARCExpand::getAnalysisUsage(AnalysisUsage &AU) const {
803 AU.setPreservesCFG();
806 bool ObjCARCExpand::doInitialization(Module &M) {
807 Run = ModuleHasARC(M);
811 bool ObjCARCExpand::runOnFunction(Function &F) {
815 // If nothing in the Module uses ARC, don't do anything.
819 bool Changed = false;
821 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I) {
822 Instruction *Inst = &*I;
824 switch (GetBasicInstructionClass(Inst)) {
828 case IC_AutoreleaseRV:
829 case IC_FusedRetainAutorelease:
830 case IC_FusedRetainAutoreleaseRV:
831 // These calls return their argument verbatim, as a low-level
832 // optimization. However, this makes high-level optimizations
833 // harder. Undo any uses of this optimization that the front-end
834 // emitted here. We'll redo them in a later pass.
836 Inst->replaceAllUsesWith(cast<CallInst>(Inst)->getArgOperand(0));
846 //===----------------------------------------------------------------------===//
848 //===----------------------------------------------------------------------===//
850 // TODO: On code like this:
853 // stuff_that_cannot_release()
854 // objc_autorelease(%x)
855 // stuff_that_cannot_release()
857 // stuff_that_cannot_release()
858 // objc_autorelease(%x)
860 // The second retain and autorelease can be deleted.
862 // TODO: It should be possible to delete
863 // objc_autoreleasePoolPush and objc_autoreleasePoolPop
864 // pairs if nothing is actually autoreleased between them. Also, autorelease
865 // calls followed by objc_autoreleasePoolPop calls (perhaps in ObjC++ code
866 // after inlining) can be turned into plain release calls.
868 // TODO: Critical-edge splitting. If the optimial insertion point is
869 // a critical edge, the current algorithm has to fail, because it doesn't
870 // know how to split edges. It should be possible to make the optimizer
871 // think in terms of edges, rather than blocks, and then split critical
874 // TODO: OptimizeSequences could generalized to be Interprocedural.
876 // TODO: Recognize that a bunch of other objc runtime calls have
877 // non-escaping arguments and non-releasing arguments, and may be
878 // non-autoreleasing.
880 // TODO: Sink autorelease calls as far as possible. Unfortunately we
881 // usually can't sink them past other calls, which would be the main
882 // case where it would be useful.
884 // TODO: The pointer returned from objc_loadWeakRetained is retained.
886 // TODO: Delete release+retain pairs (rare).
888 #include "llvm/GlobalAlias.h"
889 #include "llvm/Constants.h"
890 #include "llvm/LLVMContext.h"
891 #include "llvm/Support/ErrorHandling.h"
892 #include "llvm/Support/CFG.h"
893 #include "llvm/ADT/PostOrderIterator.h"
894 #include "llvm/ADT/Statistic.h"
896 STATISTIC(NumNoops, "Number of no-op objc calls eliminated");
897 STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated");
898 STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases");
899 STATISTIC(NumRets, "Number of return value forwarding "
900 "retain+autoreleaes eliminated");
901 STATISTIC(NumRRs, "Number of retain+release paths eliminated");
902 STATISTIC(NumPeeps, "Number of calls peephole-optimized");
905 /// ProvenanceAnalysis - This is similar to BasicAliasAnalysis, and it
906 /// uses many of the same techniques, except it uses special ObjC-specific
907 /// reasoning about pointer relationships.
908 class ProvenanceAnalysis {
911 typedef std::pair<const Value *, const Value *> ValuePairTy;
912 typedef DenseMap<ValuePairTy, bool> CachedResultsTy;
913 CachedResultsTy CachedResults;
915 bool relatedCheck(const Value *A, const Value *B);
916 bool relatedSelect(const SelectInst *A, const Value *B);
917 bool relatedPHI(const PHINode *A, const Value *B);
920 void operator=(const ProvenanceAnalysis &);
921 ProvenanceAnalysis(const ProvenanceAnalysis &);
924 ProvenanceAnalysis() {}
926 void setAA(AliasAnalysis *aa) { AA = aa; }
928 AliasAnalysis *getAA() const { return AA; }
930 bool related(const Value *A, const Value *B);
933 CachedResults.clear();
938 bool ProvenanceAnalysis::relatedSelect(const SelectInst *A, const Value *B) {
939 // If the values are Selects with the same condition, we can do a more precise
940 // check: just check for relations between the values on corresponding arms.
941 if (const SelectInst *SB = dyn_cast<SelectInst>(B))
942 if (A->getCondition() == SB->getCondition()) {
943 if (related(A->getTrueValue(), SB->getTrueValue()))
945 if (related(A->getFalseValue(), SB->getFalseValue()))
950 // Check both arms of the Select node individually.
951 if (related(A->getTrueValue(), B))
953 if (related(A->getFalseValue(), B))
956 // The arms both checked out.
960 bool ProvenanceAnalysis::relatedPHI(const PHINode *A, const Value *B) {
961 // If the values are PHIs in the same block, we can do a more precise as well
962 // as efficient check: just check for relations between the values on
963 // corresponding edges.
964 if (const PHINode *PNB = dyn_cast<PHINode>(B))
965 if (PNB->getParent() == A->getParent()) {
966 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i)
967 if (related(A->getIncomingValue(i),
968 PNB->getIncomingValueForBlock(A->getIncomingBlock(i))))
973 // Check each unique source of the PHI node against B.
974 SmallPtrSet<const Value *, 4> UniqueSrc;
975 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i) {
976 const Value *PV1 = A->getIncomingValue(i);
977 if (UniqueSrc.insert(PV1) && related(PV1, B))
981 // All of the arms checked out.
985 /// isStoredObjCPointer - Test if the value of P, or any value covered by its
986 /// provenance, is ever stored within the function (not counting callees).
987 static bool isStoredObjCPointer(const Value *P) {
988 SmallPtrSet<const Value *, 8> Visited;
989 SmallVector<const Value *, 8> Worklist;
990 Worklist.push_back(P);
993 P = Worklist.pop_back_val();
994 for (Value::const_use_iterator UI = P->use_begin(), UE = P->use_end();
996 const User *Ur = *UI;
997 if (isa<StoreInst>(Ur)) {
998 if (UI.getOperandNo() == 0)
999 // The pointer is stored.
1001 // The pointed is stored through.
1004 if (isa<CallInst>(Ur))
1005 // The pointer is passed as an argument, ignore this.
1007 if (isa<PtrToIntInst>(P))
1008 // Assume the worst.
1010 if (Visited.insert(Ur))
1011 Worklist.push_back(Ur);
1013 } while (!Worklist.empty());
1015 // Everything checked out.
1019 bool ProvenanceAnalysis::relatedCheck(const Value *A, const Value *B) {
1020 // Skip past provenance pass-throughs.
1021 A = GetUnderlyingObjCPtr(A);
1022 B = GetUnderlyingObjCPtr(B);
1028 // Ask regular AliasAnalysis, for a first approximation.
1029 switch (AA->alias(A, B)) {
1030 case AliasAnalysis::NoAlias:
1032 case AliasAnalysis::MustAlias:
1033 case AliasAnalysis::PartialAlias:
1035 case AliasAnalysis::MayAlias:
1039 bool AIsIdentified = IsObjCIdentifiedObject(A);
1040 bool BIsIdentified = IsObjCIdentifiedObject(B);
1042 // An ObjC-Identified object can't alias a load if it is never locally stored.
1043 if (AIsIdentified) {
1044 if (BIsIdentified) {
1045 // If both pointers have provenance, they can be directly compared.
1049 if (isa<LoadInst>(B))
1050 return isStoredObjCPointer(A);
1053 if (BIsIdentified && isa<LoadInst>(A))
1054 return isStoredObjCPointer(B);
1057 // Special handling for PHI and Select.
1058 if (const PHINode *PN = dyn_cast<PHINode>(A))
1059 return relatedPHI(PN, B);
1060 if (const PHINode *PN = dyn_cast<PHINode>(B))
1061 return relatedPHI(PN, A);
1062 if (const SelectInst *S = dyn_cast<SelectInst>(A))
1063 return relatedSelect(S, B);
1064 if (const SelectInst *S = dyn_cast<SelectInst>(B))
1065 return relatedSelect(S, A);
1071 bool ProvenanceAnalysis::related(const Value *A, const Value *B) {
1072 // Begin by inserting a conservative value into the map. If the insertion
1073 // fails, we have the answer already. If it succeeds, leave it there until we
1074 // compute the real answer to guard against recursive queries.
1075 if (A > B) std::swap(A, B);
1076 std::pair<CachedResultsTy::iterator, bool> Pair =
1077 CachedResults.insert(std::make_pair(ValuePairTy(A, B), true));
1079 return Pair.first->second;
1081 bool Result = relatedCheck(A, B);
1082 CachedResults[ValuePairTy(A, B)] = Result;
1087 // Sequence - A sequence of states that a pointer may go through in which an
1088 // objc_retain and objc_release are actually needed.
1091 S_Retain, ///< objc_retain(x)
1092 S_CanRelease, ///< foo(x) -- x could possibly see a ref count decrement
1093 S_Use, ///< any use of x
1094 S_Stop, ///< like S_Release, but code motion is stopped
1095 S_Release, ///< objc_release(x)
1096 S_MovableRelease ///< objc_release(x), !clang.imprecise_release
1100 static Sequence MergeSeqs(Sequence A, Sequence B, bool TopDown) {
1104 if (A == S_None || B == S_None)
1107 if (A > B) std::swap(A, B);
1109 // Choose the side which is further along in the sequence.
1110 if ((A == S_Retain || A == S_CanRelease) &&
1111 (B == S_CanRelease || B == S_Use))
1114 // Choose the side which is further along in the sequence.
1115 if ((A == S_Use || A == S_CanRelease) &&
1116 (B == S_Use || B == S_Release || B == S_Stop || B == S_MovableRelease))
1118 // If both sides are releases, choose the more conservative one.
1119 if (A == S_Stop && (B == S_Release || B == S_MovableRelease))
1121 if (A == S_Release && B == S_MovableRelease)
1129 /// RRInfo - Unidirectional information about either a
1130 /// retain-decrement-use-release sequence or release-use-decrement-retain
1131 /// reverese sequence.
1133 /// KnownSafe - After an objc_retain, the reference count of the referenced
1134 /// object is known to be positive. Similarly, before an objc_release, the
1135 /// reference count of the referenced object is known to be positive. If
1136 /// there are retain-release pairs in code regions where the retain count
1137 /// is known to be positive, they can be eliminated, regardless of any side
1138 /// effects between them.
1140 /// Also, a retain+release pair nested within another retain+release
1141 /// pair all on the known same pointer value can be eliminated, regardless
1142 /// of any intervening side effects.
1144 /// KnownSafe is true when either of these conditions is satisfied.
1147 /// IsRetainBlock - True if the Calls are objc_retainBlock calls (as
1148 /// opposed to objc_retain calls).
1151 /// IsTailCallRelease - True of the objc_release calls are all marked
1152 /// with the "tail" keyword.
1153 bool IsTailCallRelease;
1155 /// ReleaseMetadata - If the Calls are objc_release calls and they all have
1156 /// a clang.imprecise_release tag, this is the metadata tag.
1157 MDNode *ReleaseMetadata;
1159 /// Calls - For a top-down sequence, the set of objc_retains or
1160 /// objc_retainBlocks. For bottom-up, the set of objc_releases.
1161 SmallPtrSet<Instruction *, 2> Calls;
1163 /// ReverseInsertPts - The set of optimal insert positions for
1164 /// moving calls in the opposite sequence.
1165 SmallPtrSet<Instruction *, 2> ReverseInsertPts;
1168 KnownSafe(false), IsRetainBlock(false), IsTailCallRelease(false),
1169 ReleaseMetadata(0) {}
1175 void RRInfo::clear() {
1177 IsRetainBlock = false;
1178 IsTailCallRelease = false;
1179 ReleaseMetadata = 0;
1181 ReverseInsertPts.clear();
1185 /// PtrState - This class summarizes several per-pointer runtime properties
1186 /// which are propogated through the flow graph.
1188 /// RefCount - The known minimum number of reference count increments.
1191 /// NestCount - The known minimum level of retain+release nesting.
1194 /// Seq - The current position in the sequence.
1198 /// RRI - Unidirectional information about the current sequence.
1199 /// TODO: Encapsulate this better.
1202 PtrState() : RefCount(0), NestCount(0), Seq(S_None) {}
1204 void SetAtLeastOneRefCount() {
1205 if (RefCount == 0) RefCount = 1;
1208 void IncrementRefCount() {
1209 if (RefCount != UINT_MAX) ++RefCount;
1212 void DecrementRefCount() {
1213 if (RefCount != 0) --RefCount;
1216 bool IsKnownIncremented() const {
1217 return RefCount > 0;
1220 void IncrementNestCount() {
1221 if (NestCount != UINT_MAX) ++NestCount;
1224 void DecrementNestCount() {
1225 if (NestCount != 0) --NestCount;
1228 bool IsKnownNested() const {
1229 return NestCount > 0;
1232 void SetSeq(Sequence NewSeq) {
1236 void SetSeqToRelease(MDNode *M) {
1237 if (Seq == S_None || Seq == S_Use) {
1238 Seq = M ? S_MovableRelease : S_Release;
1239 RRI.ReleaseMetadata = M;
1240 } else if (Seq != S_MovableRelease || RRI.ReleaseMetadata != M) {
1242 RRI.ReleaseMetadata = 0;
1246 Sequence GetSeq() const {
1250 void ClearSequenceProgress() {
1255 void Merge(const PtrState &Other, bool TopDown);
1260 PtrState::Merge(const PtrState &Other, bool TopDown) {
1261 Seq = MergeSeqs(Seq, Other.Seq, TopDown);
1262 RefCount = std::min(RefCount, Other.RefCount);
1263 NestCount = std::min(NestCount, Other.NestCount);
1265 // We can't merge a plain objc_retain with an objc_retainBlock.
1266 if (RRI.IsRetainBlock != Other.RRI.IsRetainBlock)
1269 if (Seq == S_None) {
1272 // Conservatively merge the ReleaseMetadata information.
1273 if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
1274 RRI.ReleaseMetadata = 0;
1276 RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe;
1277 RRI.IsTailCallRelease = RRI.IsTailCallRelease && Other.RRI.IsTailCallRelease;
1278 RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
1279 RRI.ReverseInsertPts.insert(Other.RRI.ReverseInsertPts.begin(),
1280 Other.RRI.ReverseInsertPts.end());
1285 /// BBState - Per-BasicBlock state.
1287 /// TopDownPathCount - The number of unique control paths from the entry
1288 /// which can reach this block.
1289 unsigned TopDownPathCount;
1291 /// BottomUpPathCount - The number of unique control paths to exits
1292 /// from this block.
1293 unsigned BottomUpPathCount;
1295 /// MapTy - A type for PerPtrTopDown and PerPtrBottomUp.
1296 typedef MapVector<const Value *, PtrState> MapTy;
1298 /// PerPtrTopDown - The top-down traversal uses this to record information
1299 /// known about a pointer at the bottom of each block.
1300 MapTy PerPtrTopDown;
1302 /// PerPtrBottomUp - The bottom-up traversal uses this to record information
1303 /// known about a pointer at the top of each block.
1304 MapTy PerPtrBottomUp;
1307 BBState() : TopDownPathCount(0), BottomUpPathCount(0) {}
1309 typedef MapTy::iterator ptr_iterator;
1310 typedef MapTy::const_iterator ptr_const_iterator;
1312 ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); }
1313 ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); }
1314 ptr_const_iterator top_down_ptr_begin() const {
1315 return PerPtrTopDown.begin();
1317 ptr_const_iterator top_down_ptr_end() const {
1318 return PerPtrTopDown.end();
1321 ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); }
1322 ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); }
1323 ptr_const_iterator bottom_up_ptr_begin() const {
1324 return PerPtrBottomUp.begin();
1326 ptr_const_iterator bottom_up_ptr_end() const {
1327 return PerPtrBottomUp.end();
1330 /// SetAsEntry - Mark this block as being an entry block, which has one
1331 /// path from the entry by definition.
1332 void SetAsEntry() { TopDownPathCount = 1; }
1334 /// SetAsExit - Mark this block as being an exit block, which has one
1335 /// path to an exit by definition.
1336 void SetAsExit() { BottomUpPathCount = 1; }
1338 PtrState &getPtrTopDownState(const Value *Arg) {
1339 return PerPtrTopDown[Arg];
1342 PtrState &getPtrBottomUpState(const Value *Arg) {
1343 return PerPtrBottomUp[Arg];
1346 void clearBottomUpPointers() {
1347 PerPtrBottomUp.clear();
1350 void clearTopDownPointers() {
1351 PerPtrTopDown.clear();
1354 void InitFromPred(const BBState &Other);
1355 void InitFromSucc(const BBState &Other);
1356 void MergePred(const BBState &Other);
1357 void MergeSucc(const BBState &Other);
1359 /// GetAllPathCount - Return the number of possible unique paths from an
1360 /// entry to an exit which pass through this block. This is only valid
1361 /// after both the top-down and bottom-up traversals are complete.
1362 unsigned GetAllPathCount() const {
1363 return TopDownPathCount * BottomUpPathCount;
1366 /// IsVisitedTopDown - Test whether the block for this BBState has been
1367 /// visited by the top-down portion of the algorithm.
1368 bool isVisitedTopDown() const {
1369 return TopDownPathCount != 0;
1374 void BBState::InitFromPred(const BBState &Other) {
1375 PerPtrTopDown = Other.PerPtrTopDown;
1376 TopDownPathCount = Other.TopDownPathCount;
1379 void BBState::InitFromSucc(const BBState &Other) {
1380 PerPtrBottomUp = Other.PerPtrBottomUp;
1381 BottomUpPathCount = Other.BottomUpPathCount;
1384 /// MergePred - The top-down traversal uses this to merge information about
1385 /// predecessors to form the initial state for a new block.
1386 void BBState::MergePred(const BBState &Other) {
1387 // Other.TopDownPathCount can be 0, in which case it is either dead or a
1388 // loop backedge. Loop backedges are special.
1389 TopDownPathCount += Other.TopDownPathCount;
1391 // For each entry in the other set, if our set has an entry with the same key,
1392 // merge the entries. Otherwise, copy the entry and merge it with an empty
1394 for (ptr_const_iterator MI = Other.top_down_ptr_begin(),
1395 ME = Other.top_down_ptr_end(); MI != ME; ++MI) {
1396 std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI);
1397 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1401 // For each entry in our set, if the other set doesn't have an entry with the
1402 // same key, force it to merge with an empty entry.
1403 for (ptr_iterator MI = top_down_ptr_begin(),
1404 ME = top_down_ptr_end(); MI != ME; ++MI)
1405 if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end())
1406 MI->second.Merge(PtrState(), /*TopDown=*/true);
1409 /// MergeSucc - The bottom-up traversal uses this to merge information about
1410 /// successors to form the initial state for a new block.
1411 void BBState::MergeSucc(const BBState &Other) {
1412 // Other.BottomUpPathCount can be 0, in which case it is either dead or a
1413 // loop backedge. Loop backedges are special.
1414 BottomUpPathCount += Other.BottomUpPathCount;
1416 // For each entry in the other set, if our set has an entry with the
1417 // same key, merge the entries. Otherwise, copy the entry and merge
1418 // it with an empty entry.
1419 for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(),
1420 ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) {
1421 std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI);
1422 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1426 // For each entry in our set, if the other set doesn't have an entry
1427 // with the same key, force it to merge with an empty entry.
1428 for (ptr_iterator MI = bottom_up_ptr_begin(),
1429 ME = bottom_up_ptr_end(); MI != ME; ++MI)
1430 if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end())
1431 MI->second.Merge(PtrState(), /*TopDown=*/false);
1435 /// ObjCARCOpt - The main ARC optimization pass.
1436 class ObjCARCOpt : public FunctionPass {
1438 ProvenanceAnalysis PA;
1440 /// Run - A flag indicating whether this optimization pass should run.
1443 /// RetainRVCallee, etc. - Declarations for ObjC runtime
1444 /// functions, for use in creating calls to them. These are initialized
1445 /// lazily to avoid cluttering up the Module with unused declarations.
1446 Constant *RetainRVCallee, *AutoreleaseRVCallee, *ReleaseCallee,
1447 *RetainCallee, *RetainBlockCallee, *AutoreleaseCallee;
1449 /// UsedInThisFunciton - Flags which determine whether each of the
1450 /// interesting runtine functions is in fact used in the current function.
1451 unsigned UsedInThisFunction;
1453 /// ImpreciseReleaseMDKind - The Metadata Kind for clang.imprecise_release
1455 unsigned ImpreciseReleaseMDKind;
1457 Constant *getRetainRVCallee(Module *M);
1458 Constant *getAutoreleaseRVCallee(Module *M);
1459 Constant *getReleaseCallee(Module *M);
1460 Constant *getRetainCallee(Module *M);
1461 Constant *getRetainBlockCallee(Module *M);
1462 Constant *getAutoreleaseCallee(Module *M);
1464 void OptimizeRetainCall(Function &F, Instruction *Retain);
1465 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
1466 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV);
1467 void OptimizeIndividualCalls(Function &F);
1469 void CheckForCFGHazards(const BasicBlock *BB,
1470 DenseMap<const BasicBlock *, BBState> &BBStates,
1471 BBState &MyStates) const;
1472 bool VisitBottomUp(BasicBlock *BB,
1473 DenseMap<const BasicBlock *, BBState> &BBStates,
1474 MapVector<Value *, RRInfo> &Retains);
1475 bool VisitTopDown(BasicBlock *BB,
1476 DenseMap<const BasicBlock *, BBState> &BBStates,
1477 DenseMap<Value *, RRInfo> &Releases);
1478 bool Visit(Function &F,
1479 DenseMap<const BasicBlock *, BBState> &BBStates,
1480 MapVector<Value *, RRInfo> &Retains,
1481 DenseMap<Value *, RRInfo> &Releases);
1483 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1484 MapVector<Value *, RRInfo> &Retains,
1485 DenseMap<Value *, RRInfo> &Releases,
1486 SmallVectorImpl<Instruction *> &DeadInsts,
1489 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1490 MapVector<Value *, RRInfo> &Retains,
1491 DenseMap<Value *, RRInfo> &Releases,
1494 void OptimizeWeakCalls(Function &F);
1496 bool OptimizeSequences(Function &F);
1498 void OptimizeReturns(Function &F);
1500 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1501 virtual bool doInitialization(Module &M);
1502 virtual bool runOnFunction(Function &F);
1503 virtual void releaseMemory();
1507 ObjCARCOpt() : FunctionPass(ID) {
1508 initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1513 char ObjCARCOpt::ID = 0;
1514 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1515 "objc-arc", "ObjC ARC optimization", false, false)
1516 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1517 INITIALIZE_PASS_END(ObjCARCOpt,
1518 "objc-arc", "ObjC ARC optimization", false, false)
1520 Pass *llvm::createObjCARCOptPass() {
1521 return new ObjCARCOpt();
1524 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1525 AU.addRequired<ObjCARCAliasAnalysis>();
1526 AU.addRequired<AliasAnalysis>();
1527 // ARC optimization doesn't currently split critical edges.
1528 AU.setPreservesCFG();
1531 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1532 if (!RetainRVCallee) {
1533 LLVMContext &C = M->getContext();
1534 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1535 std::vector<Type *> Params;
1536 Params.push_back(I8X);
1538 FunctionType::get(I8X, Params, /*isVarArg=*/false);
1539 AttrListPtr Attributes;
1540 Attributes.addAttr(~0u, Attribute::NoUnwind);
1542 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1545 return RetainRVCallee;
1548 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1549 if (!AutoreleaseRVCallee) {
1550 LLVMContext &C = M->getContext();
1551 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1552 std::vector<Type *> Params;
1553 Params.push_back(I8X);
1555 FunctionType::get(I8X, Params, /*isVarArg=*/false);
1556 AttrListPtr Attributes;
1557 Attributes.addAttr(~0u, Attribute::NoUnwind);
1558 AutoreleaseRVCallee =
1559 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1562 return AutoreleaseRVCallee;
1565 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1566 if (!ReleaseCallee) {
1567 LLVMContext &C = M->getContext();
1568 std::vector<Type *> Params;
1569 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1570 AttrListPtr Attributes;
1571 Attributes.addAttr(~0u, Attribute::NoUnwind);
1573 M->getOrInsertFunction(
1575 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1578 return ReleaseCallee;
1581 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1582 if (!RetainCallee) {
1583 LLVMContext &C = M->getContext();
1584 std::vector<Type *> Params;
1585 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1586 AttrListPtr Attributes;
1587 Attributes.addAttr(~0u, Attribute::NoUnwind);
1589 M->getOrInsertFunction(
1591 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1594 return RetainCallee;
1597 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1598 if (!RetainBlockCallee) {
1599 LLVMContext &C = M->getContext();
1600 std::vector<Type *> Params;
1601 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1602 AttrListPtr Attributes;
1603 Attributes.addAttr(~0u, Attribute::NoUnwind);
1605 M->getOrInsertFunction(
1607 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1610 return RetainBlockCallee;
1613 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1614 if (!AutoreleaseCallee) {
1615 LLVMContext &C = M->getContext();
1616 std::vector<Type *> Params;
1617 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1618 AttrListPtr Attributes;
1619 Attributes.addAttr(~0u, Attribute::NoUnwind);
1621 M->getOrInsertFunction(
1623 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1626 return AutoreleaseCallee;
1629 /// CanAlterRefCount - Test whether the given instruction can result in a
1630 /// reference count modification (positive or negative) for the pointer's
1633 CanAlterRefCount(const Instruction *Inst, const Value *Ptr,
1634 ProvenanceAnalysis &PA, InstructionClass Class) {
1636 case IC_Autorelease:
1637 case IC_AutoreleaseRV:
1639 // These operations never directly modify a reference count.
1644 ImmutableCallSite CS = static_cast<const Value *>(Inst);
1645 assert(CS && "Only calls can alter reference counts!");
1647 // See if AliasAnalysis can help us with the call.
1648 AliasAnalysis::ModRefBehavior MRB = PA.getAA()->getModRefBehavior(CS);
1649 if (AliasAnalysis::onlyReadsMemory(MRB))
1651 if (AliasAnalysis::onlyAccessesArgPointees(MRB)) {
1652 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1654 const Value *Op = *I;
1655 if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1661 // Assume the worst.
1665 /// CanUse - Test whether the given instruction can "use" the given pointer's
1666 /// object in a way that requires the reference count to be positive.
1668 CanUse(const Instruction *Inst, const Value *Ptr, ProvenanceAnalysis &PA,
1669 InstructionClass Class) {
1670 // IC_Call operations (as opposed to IC_CallOrUser) never "use" objc pointers.
1671 if (Class == IC_Call)
1674 // Consider various instructions which may have pointer arguments which are
1676 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(Inst)) {
1677 // Comparing a pointer with null, or any other constant, isn't really a use,
1678 // because we don't care what the pointer points to, or about the values
1679 // of any other dynamic reference-counted pointers.
1680 if (!IsPotentialUse(ICI->getOperand(1)))
1682 } else if (ImmutableCallSite CS = static_cast<const Value *>(Inst)) {
1683 // For calls, just check the arguments (and not the callee operand).
1684 for (ImmutableCallSite::arg_iterator OI = CS.arg_begin(),
1685 OE = CS.arg_end(); OI != OE; ++OI) {
1686 const Value *Op = *OI;
1687 if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1691 } else if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1692 // Special-case stores, because we don't care about the stored value, just
1693 // the store address.
1694 const Value *Op = GetUnderlyingObjCPtr(SI->getPointerOperand());
1695 // If we can't tell what the underlying object was, assume there is a
1697 return IsPotentialUse(Op) && PA.related(Op, Ptr);
1700 // Check each operand for a match.
1701 for (User::const_op_iterator OI = Inst->op_begin(), OE = Inst->op_end();
1703 const Value *Op = *OI;
1704 if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1710 /// CanInterruptRV - Test whether the given instruction can autorelease
1711 /// any pointer or cause an autoreleasepool pop.
1713 CanInterruptRV(InstructionClass Class) {
1715 case IC_AutoreleasepoolPop:
1718 case IC_Autorelease:
1719 case IC_AutoreleaseRV:
1720 case IC_FusedRetainAutorelease:
1721 case IC_FusedRetainAutoreleaseRV:
1729 /// DependenceKind - There are several kinds of dependence-like concepts in
1731 enum DependenceKind {
1732 NeedsPositiveRetainCount,
1733 CanChangeRetainCount,
1734 RetainAutoreleaseDep, ///< Blocks objc_retainAutorelease.
1735 RetainAutoreleaseRVDep, ///< Blocks objc_retainAutoreleaseReturnValue.
1736 RetainRVDep ///< Blocks objc_retainAutoreleasedReturnValue.
1740 /// Depends - Test if there can be dependencies on Inst through Arg. This
1741 /// function only tests dependencies relevant for removing pairs of calls.
1743 Depends(DependenceKind Flavor, Instruction *Inst, const Value *Arg,
1744 ProvenanceAnalysis &PA) {
1745 // If we've reached the definition of Arg, stop.
1750 case NeedsPositiveRetainCount: {
1751 InstructionClass Class = GetInstructionClass(Inst);
1753 case IC_AutoreleasepoolPop:
1754 case IC_AutoreleasepoolPush:
1758 return CanUse(Inst, Arg, PA, Class);
1762 case CanChangeRetainCount: {
1763 InstructionClass Class = GetInstructionClass(Inst);
1765 case IC_AutoreleasepoolPop:
1766 // Conservatively assume this can decrement any count.
1768 case IC_AutoreleasepoolPush:
1772 return CanAlterRefCount(Inst, Arg, PA, Class);
1776 case RetainAutoreleaseDep:
1777 switch (GetBasicInstructionClass(Inst)) {
1778 case IC_AutoreleasepoolPop:
1779 // Don't merge an objc_autorelease with an objc_retain inside a different
1780 // autoreleasepool scope.
1784 // Check for a retain of the same pointer for merging.
1785 return GetObjCArg(Inst) == Arg;
1787 // Nothing else matters for objc_retainAutorelease formation.
1792 case RetainAutoreleaseRVDep: {
1793 InstructionClass Class = GetBasicInstructionClass(Inst);
1797 // Check for a retain of the same pointer for merging.
1798 return GetObjCArg(Inst) == Arg;
1800 // Anything that can autorelease interrupts
1801 // retainAutoreleaseReturnValue formation.
1802 return CanInterruptRV(Class);
1808 return CanInterruptRV(GetBasicInstructionClass(Inst));
1811 llvm_unreachable("Invalid dependence flavor");
1815 /// FindDependencies - Walk up the CFG from StartPos (which is in StartBB) and
1816 /// find local and non-local dependencies on Arg.
1817 /// TODO: Cache results?
1819 FindDependencies(DependenceKind Flavor,
1821 BasicBlock *StartBB, Instruction *StartInst,
1822 SmallPtrSet<Instruction *, 4> &DependingInstructions,
1823 SmallPtrSet<const BasicBlock *, 4> &Visited,
1824 ProvenanceAnalysis &PA) {
1825 BasicBlock::iterator StartPos = StartInst;
1827 SmallVector<std::pair<BasicBlock *, BasicBlock::iterator>, 4> Worklist;
1828 Worklist.push_back(std::make_pair(StartBB, StartPos));
1830 std::pair<BasicBlock *, BasicBlock::iterator> Pair =
1831 Worklist.pop_back_val();
1832 BasicBlock *LocalStartBB = Pair.first;
1833 BasicBlock::iterator LocalStartPos = Pair.second;
1834 BasicBlock::iterator StartBBBegin = LocalStartBB->begin();
1836 if (LocalStartPos == StartBBBegin) {
1837 pred_iterator PI(LocalStartBB), PE(LocalStartBB, false);
1839 // If we've reached the function entry, produce a null dependence.
1840 DependingInstructions.insert(0);
1842 // Add the predecessors to the worklist.
1844 BasicBlock *PredBB = *PI;
1845 if (Visited.insert(PredBB))
1846 Worklist.push_back(std::make_pair(PredBB, PredBB->end()));
1847 } while (++PI != PE);
1851 Instruction *Inst = --LocalStartPos;
1852 if (Depends(Flavor, Inst, Arg, PA)) {
1853 DependingInstructions.insert(Inst);
1857 } while (!Worklist.empty());
1859 // Determine whether the original StartBB post-dominates all of the blocks we
1860 // visited. If not, insert a sentinal indicating that most optimizations are
1862 for (SmallPtrSet<const BasicBlock *, 4>::const_iterator I = Visited.begin(),
1863 E = Visited.end(); I != E; ++I) {
1864 const BasicBlock *BB = *I;
1867 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1868 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
1869 const BasicBlock *Succ = *SI;
1870 if (Succ != StartBB && !Visited.count(Succ)) {
1871 DependingInstructions.insert(reinterpret_cast<Instruction *>(-1));
1878 static bool isNullOrUndef(const Value *V) {
1879 return isa<ConstantPointerNull>(V) || isa<UndefValue>(V);
1882 static bool isNoopInstruction(const Instruction *I) {
1883 return isa<BitCastInst>(I) ||
1884 (isa<GetElementPtrInst>(I) &&
1885 cast<GetElementPtrInst>(I)->hasAllZeroIndices());
1888 /// OptimizeRetainCall - Turn objc_retain into
1889 /// objc_retainAutoreleasedReturnValue if the operand is a return value.
1891 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
1892 CallSite CS(GetObjCArg(Retain));
1893 Instruction *Call = CS.getInstruction();
1895 if (Call->getParent() != Retain->getParent()) return;
1897 // Check that the call is next to the retain.
1898 BasicBlock::iterator I = Call;
1900 while (isNoopInstruction(I)) ++I;
1904 // Turn it to an objc_retainAutoreleasedReturnValue..
1907 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
1910 /// OptimizeRetainRVCall - Turn objc_retainAutoreleasedReturnValue into
1911 /// objc_retain if the operand is not a return value. Or, if it can be
1912 /// paired with an objc_autoreleaseReturnValue, delete the pair and
1915 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
1916 // Check for the argument being from an immediately preceding call.
1917 Value *Arg = GetObjCArg(RetainRV);
1919 if (Instruction *Call = CS.getInstruction())
1920 if (Call->getParent() == RetainRV->getParent()) {
1921 BasicBlock::iterator I = Call;
1923 while (isNoopInstruction(I)) ++I;
1924 if (&*I == RetainRV)
1928 // Check for being preceded by an objc_autoreleaseReturnValue on the same
1929 // pointer. In this case, we can delete the pair.
1930 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
1932 do --I; while (I != Begin && isNoopInstruction(I));
1933 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
1934 GetObjCArg(I) == Arg) {
1937 EraseInstruction(I);
1938 EraseInstruction(RetainRV);
1943 // Turn it to a plain objc_retain.
1946 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
1950 /// OptimizeAutoreleaseRVCall - Turn objc_autoreleaseReturnValue into
1951 /// objc_autorelease if the result is not used as a return value.
1953 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV) {
1954 // Check for a return of the pointer value.
1955 const Value *Ptr = GetObjCArg(AutoreleaseRV);
1956 SmallVector<const Value *, 2> Users;
1957 Users.push_back(Ptr);
1959 Ptr = Users.pop_back_val();
1960 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
1962 const User *I = *UI;
1963 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
1965 if (isa<BitCastInst>(I))
1968 } while (!Users.empty());
1972 cast<CallInst>(AutoreleaseRV)->
1973 setCalledFunction(getAutoreleaseCallee(F.getParent()));
1976 /// OptimizeIndividualCalls - Visit each call, one at a time, and make
1977 /// simplifications without doing any additional analysis.
1978 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
1979 // Reset all the flags in preparation for recomputing them.
1980 UsedInThisFunction = 0;
1982 // Visit all objc_* calls in F.
1983 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
1984 Instruction *Inst = &*I++;
1985 InstructionClass Class = GetBasicInstructionClass(Inst);
1990 // Delete no-op casts. These function calls have special semantics, but
1991 // the semantics are entirely implemented via lowering in the front-end,
1992 // so by the time they reach the optimizer, they are just no-op calls
1993 // which return their argument.
1995 // There are gray areas here, as the ability to cast reference-counted
1996 // pointers to raw void* and back allows code to break ARC assumptions,
1997 // however these are currently considered to be unimportant.
2001 EraseInstruction(Inst);
2004 // If the pointer-to-weak-pointer is null, it's undefined behavior.
2007 case IC_LoadWeakRetained:
2009 case IC_DestroyWeak: {
2010 CallInst *CI = cast<CallInst>(Inst);
2011 if (isNullOrUndef(CI->getArgOperand(0))) {
2012 Type *Ty = CI->getArgOperand(0)->getType();
2013 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2014 Constant::getNullValue(Ty),
2016 CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
2017 CI->eraseFromParent();
2024 CallInst *CI = cast<CallInst>(Inst);
2025 if (isNullOrUndef(CI->getArgOperand(0)) ||
2026 isNullOrUndef(CI->getArgOperand(1))) {
2027 Type *Ty = CI->getArgOperand(0)->getType();
2028 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2029 Constant::getNullValue(Ty),
2031 CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
2032 CI->eraseFromParent();
2038 OptimizeRetainCall(F, Inst);
2041 if (OptimizeRetainRVCall(F, Inst))
2044 case IC_AutoreleaseRV:
2045 OptimizeAutoreleaseRVCall(F, Inst);
2049 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
2050 if (IsAutorelease(Class) && Inst->use_empty()) {
2051 CallInst *Call = cast<CallInst>(Inst);
2052 const Value *Arg = Call->getArgOperand(0);
2053 Arg = FindSingleUseIdentifiedObject(Arg);
2058 // Create the declaration lazily.
2059 LLVMContext &C = Inst->getContext();
2061 CallInst::Create(getReleaseCallee(F.getParent()),
2062 Call->getArgOperand(0), "", Call);
2063 NewCall->setMetadata(ImpreciseReleaseMDKind,
2064 MDNode::get(C, ArrayRef<Value *>()));
2065 EraseInstruction(Call);
2071 // For functions which can never be passed stack arguments, add
2073 if (IsAlwaysTail(Class)) {
2075 cast<CallInst>(Inst)->setTailCall();
2078 // Set nounwind as needed.
2079 if (IsNoThrow(Class)) {
2081 cast<CallInst>(Inst)->setDoesNotThrow();
2084 if (!IsNoopOnNull(Class)) {
2085 UsedInThisFunction |= 1 << Class;
2089 const Value *Arg = GetObjCArg(Inst);
2091 // ARC calls with null are no-ops. Delete them.
2092 if (isNullOrUndef(Arg)) {
2095 EraseInstruction(Inst);
2099 // Keep track of which of retain, release, autorelease, and retain_block
2100 // are actually present in this function.
2101 UsedInThisFunction |= 1 << Class;
2103 // If Arg is a PHI, and one or more incoming values to the
2104 // PHI are null, and the call is control-equivalent to the PHI, and there
2105 // are no relevant side effects between the PHI and the call, the call
2106 // could be pushed up to just those paths with non-null incoming values.
2107 // For now, don't bother splitting critical edges for this.
2108 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
2109 Worklist.push_back(std::make_pair(Inst, Arg));
2111 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
2115 const PHINode *PN = dyn_cast<PHINode>(Arg);
2118 // Determine if the PHI has any null operands, or any incoming
2120 bool HasNull = false;
2121 bool HasCriticalEdges = false;
2122 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2124 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2125 if (isNullOrUndef(Incoming))
2127 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
2128 .getNumSuccessors() != 1) {
2129 HasCriticalEdges = true;
2133 // If we have null operands and no critical edges, optimize.
2134 if (!HasCriticalEdges && HasNull) {
2135 SmallPtrSet<Instruction *, 4> DependingInstructions;
2136 SmallPtrSet<const BasicBlock *, 4> Visited;
2138 // Check that there is nothing that cares about the reference
2139 // count between the call and the phi.
2140 FindDependencies(NeedsPositiveRetainCount, Arg,
2141 Inst->getParent(), Inst,
2142 DependingInstructions, Visited, PA);
2143 if (DependingInstructions.size() == 1 &&
2144 *DependingInstructions.begin() == PN) {
2147 // Clone the call into each predecessor that has a non-null value.
2148 CallInst *CInst = cast<CallInst>(Inst);
2149 Type *ParamTy = CInst->getArgOperand(0)->getType();
2150 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2152 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2153 if (!isNullOrUndef(Incoming)) {
2154 CallInst *Clone = cast<CallInst>(CInst->clone());
2155 Value *Op = PN->getIncomingValue(i);
2156 Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
2157 if (Op->getType() != ParamTy)
2158 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
2159 Clone->setArgOperand(0, Op);
2160 Clone->insertBefore(InsertPos);
2161 Worklist.push_back(std::make_pair(Clone, Incoming));
2164 // Erase the original call.
2165 EraseInstruction(CInst);
2169 } while (!Worklist.empty());
2173 /// CheckForCFGHazards - Check for critical edges, loop boundaries, irreducible
2174 /// control flow, or other CFG structures where moving code across the edge
2175 /// would result in it being executed more.
2177 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
2178 DenseMap<const BasicBlock *, BBState> &BBStates,
2179 BBState &MyStates) const {
2180 // If any top-down local-use or possible-dec has a succ which is earlier in
2181 // the sequence, forget it.
2182 for (BBState::ptr_const_iterator I = MyStates.top_down_ptr_begin(),
2183 E = MyStates.top_down_ptr_end(); I != E; ++I)
2184 switch (I->second.GetSeq()) {
2187 const Value *Arg = I->first;
2188 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2189 bool SomeSuccHasSame = false;
2190 bool AllSuccsHaveSame = true;
2191 PtrState &S = MyStates.getPtrTopDownState(Arg);
2192 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
2193 PtrState &SuccS = BBStates[*SI].getPtrBottomUpState(Arg);
2194 switch (SuccS.GetSeq()) {
2196 case S_CanRelease: {
2197 if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2198 S.ClearSequenceProgress();
2202 SomeSuccHasSame = true;
2206 case S_MovableRelease:
2207 if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2208 AllSuccsHaveSame = false;
2211 llvm_unreachable("bottom-up pointer in retain state!");
2214 // If the state at the other end of any of the successor edges
2215 // matches the current state, require all edges to match. This
2216 // guards against loops in the middle of a sequence.
2217 if (SomeSuccHasSame && !AllSuccsHaveSame)
2218 S.ClearSequenceProgress();
2220 case S_CanRelease: {
2221 const Value *Arg = I->first;
2222 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2223 bool SomeSuccHasSame = false;
2224 bool AllSuccsHaveSame = true;
2225 PtrState &S = MyStates.getPtrTopDownState(Arg);
2226 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
2227 PtrState &SuccS = BBStates[*SI].getPtrBottomUpState(Arg);
2228 switch (SuccS.GetSeq()) {
2230 if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2231 S.ClearSequenceProgress();
2235 SomeSuccHasSame = true;
2239 case S_MovableRelease:
2241 if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2242 AllSuccsHaveSame = false;
2245 llvm_unreachable("bottom-up pointer in retain state!");
2248 // If the state at the other end of any of the successor edges
2249 // matches the current state, require all edges to match. This
2250 // guards against loops in the middle of a sequence.
2251 if (SomeSuccHasSame && !AllSuccsHaveSame)
2252 S.ClearSequenceProgress();
2258 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
2259 DenseMap<const BasicBlock *, BBState> &BBStates,
2260 MapVector<Value *, RRInfo> &Retains) {
2261 bool NestingDetected = false;
2262 BBState &MyStates = BBStates[BB];
2264 // Merge the states from each successor to compute the initial state
2265 // for the current block.
2266 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2267 succ_const_iterator SI(TI), SE(TI, false);
2269 MyStates.SetAsExit();
2272 const BasicBlock *Succ = *SI++;
2275 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
2276 // If we haven't seen this node yet, then we've found a CFG cycle.
2277 // Be optimistic here; it's CheckForCFGHazards' job detect trouble.
2278 if (I == BBStates.end())
2280 MyStates.InitFromSucc(I->second);
2284 I = BBStates.find(Succ);
2285 if (I != BBStates.end())
2286 MyStates.MergeSucc(I->second);
2292 // Visit all the instructions, bottom-up.
2293 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
2294 Instruction *Inst = llvm::prior(I);
2295 InstructionClass Class = GetInstructionClass(Inst);
2296 const Value *Arg = 0;
2300 Arg = GetObjCArg(Inst);
2302 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2304 // If we see two releases in a row on the same pointer. If so, make
2305 // a note, and we'll cicle back to revisit it after we've
2306 // hopefully eliminated the second release, which may allow us to
2307 // eliminate the first release too.
2308 // Theoretically we could implement removal of nested retain+release
2309 // pairs by making PtrState hold a stack of states, but this is
2310 // simple and avoids adding overhead for the non-nested case.
2311 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease)
2312 NestingDetected = true;
2314 S.SetSeqToRelease(Inst->getMetadata(ImpreciseReleaseMDKind));
2316 S.RRI.KnownSafe = S.IsKnownNested() || S.IsKnownIncremented();
2317 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2318 S.RRI.Calls.insert(Inst);
2320 S.IncrementRefCount();
2321 S.IncrementNestCount();
2324 case IC_RetainBlock:
2327 Arg = GetObjCArg(Inst);
2329 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2330 S.DecrementRefCount();
2331 S.SetAtLeastOneRefCount();
2332 S.DecrementNestCount();
2334 switch (S.GetSeq()) {
2337 case S_MovableRelease:
2339 S.RRI.ReverseInsertPts.clear();
2342 // Don't do retain+release tracking for IC_RetainRV, because it's
2343 // better to let it remain as the first instruction after a call.
2344 if (Class != IC_RetainRV) {
2345 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2346 Retains[Inst] = S.RRI;
2348 S.ClearSequenceProgress();
2353 llvm_unreachable("bottom-up pointer in retain state!");
2357 case IC_AutoreleasepoolPop:
2358 // Conservatively, clear MyStates for all known pointers.
2359 MyStates.clearBottomUpPointers();
2361 case IC_AutoreleasepoolPush:
2363 // These are irrelevant.
2369 // Consider any other possible effects of this instruction on each
2370 // pointer being tracked.
2371 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
2372 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
2373 const Value *Ptr = MI->first;
2375 continue; // Handled above.
2376 PtrState &S = MI->second;
2377 Sequence Seq = S.GetSeq();
2379 // Check for possible releases.
2380 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2381 S.DecrementRefCount();
2384 S.SetSeq(S_CanRelease);
2388 case S_MovableRelease:
2393 llvm_unreachable("bottom-up pointer in retain state!");
2397 // Check for possible direct uses.
2400 case S_MovableRelease:
2401 if (CanUse(Inst, Ptr, PA, Class)) {
2402 S.RRI.ReverseInsertPts.clear();
2403 S.RRI.ReverseInsertPts.insert(Inst);
2405 } else if (Seq == S_Release &&
2406 (Class == IC_User || Class == IC_CallOrUser)) {
2407 // Non-movable releases depend on any possible objc pointer use.
2409 S.RRI.ReverseInsertPts.clear();
2410 S.RRI.ReverseInsertPts.insert(Inst);
2414 if (CanUse(Inst, Ptr, PA, Class))
2422 llvm_unreachable("bottom-up pointer in retain state!");
2427 return NestingDetected;
2431 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
2432 DenseMap<const BasicBlock *, BBState> &BBStates,
2433 DenseMap<Value *, RRInfo> &Releases) {
2434 bool NestingDetected = false;
2435 BBState &MyStates = BBStates[BB];
2437 // Merge the states from each predecessor to compute the initial state
2438 // for the current block.
2439 const_pred_iterator PI(BB), PE(BB, false);
2441 MyStates.SetAsEntry();
2444 const BasicBlock *Pred = *PI++;
2447 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
2448 assert(I != BBStates.end());
2449 // If we haven't seen this node yet, then we've found a CFG cycle.
2450 // Be optimistic here; it's CheckForCFGHazards' job detect trouble.
2451 if (!I->second.isVisitedTopDown())
2453 MyStates.InitFromPred(I->second);
2457 I = BBStates.find(Pred);
2458 assert(I != BBStates.end());
2459 if (I->second.isVisitedTopDown())
2460 MyStates.MergePred(I->second);
2466 // Visit all the instructions, top-down.
2467 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2468 Instruction *Inst = I;
2469 InstructionClass Class = GetInstructionClass(Inst);
2470 const Value *Arg = 0;
2473 case IC_RetainBlock:
2476 Arg = GetObjCArg(Inst);
2478 PtrState &S = MyStates.getPtrTopDownState(Arg);
2480 // Don't do retain+release tracking for IC_RetainRV, because it's
2481 // better to let it remain as the first instruction after a call.
2482 if (Class != IC_RetainRV) {
2483 // If we see two retains in a row on the same pointer. If so, make
2484 // a note, and we'll cicle back to revisit it after we've
2485 // hopefully eliminated the second retain, which may allow us to
2486 // eliminate the first retain too.
2487 // Theoretically we could implement removal of nested retain+release
2488 // pairs by making PtrState hold a stack of states, but this is
2489 // simple and avoids adding overhead for the non-nested case.
2490 if (S.GetSeq() == S_Retain)
2491 NestingDetected = true;
2495 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2496 // Don't check S.IsKnownIncremented() here because it's not
2498 S.RRI.KnownSafe = S.IsKnownNested();
2499 S.RRI.Calls.insert(Inst);
2502 S.SetAtLeastOneRefCount();
2503 S.IncrementRefCount();
2504 S.IncrementNestCount();
2508 Arg = GetObjCArg(Inst);
2510 PtrState &S = MyStates.getPtrTopDownState(Arg);
2511 S.DecrementRefCount();
2512 S.DecrementNestCount();
2514 switch (S.GetSeq()) {
2517 S.RRI.ReverseInsertPts.clear();
2520 S.RRI.ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2521 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2522 Releases[Inst] = S.RRI;
2523 S.ClearSequenceProgress();
2529 case S_MovableRelease:
2530 llvm_unreachable("top-down pointer in release state!");
2534 case IC_AutoreleasepoolPop:
2535 // Conservatively, clear MyStates for all known pointers.
2536 MyStates.clearTopDownPointers();
2538 case IC_AutoreleasepoolPush:
2540 // These are irrelevant.
2546 // Consider any other possible effects of this instruction on each
2547 // pointer being tracked.
2548 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
2549 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
2550 const Value *Ptr = MI->first;
2552 continue; // Handled above.
2553 PtrState &S = MI->second;
2554 Sequence Seq = S.GetSeq();
2556 // Check for possible releases.
2557 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2558 S.DecrementRefCount();
2561 S.SetSeq(S_CanRelease);
2562 S.RRI.ReverseInsertPts.clear();
2563 S.RRI.ReverseInsertPts.insert(Inst);
2565 // One call can't cause a transition from S_Retain to S_CanRelease
2566 // and S_CanRelease to S_Use. If we've made the first transition,
2575 case S_MovableRelease:
2576 llvm_unreachable("top-down pointer in release state!");
2580 // Check for possible direct uses.
2583 if (CanUse(Inst, Ptr, PA, Class))
2592 case S_MovableRelease:
2593 llvm_unreachable("top-down pointer in release state!");
2598 CheckForCFGHazards(BB, BBStates, MyStates);
2599 return NestingDetected;
2602 // Visit - Visit the function both top-down and bottom-up.
2604 ObjCARCOpt::Visit(Function &F,
2605 DenseMap<const BasicBlock *, BBState> &BBStates,
2606 MapVector<Value *, RRInfo> &Retains,
2607 DenseMap<Value *, RRInfo> &Releases) {
2608 // Use reverse-postorder on the reverse CFG for bottom-up, because we
2609 // magically know that loops will be well behaved, i.e. they won't repeatedly
2610 // call retain on a single pointer without doing a release. We can't use
2611 // ReversePostOrderTraversal here because we want to walk up from each
2612 // function exit point.
2613 SmallPtrSet<BasicBlock *, 16> Visited;
2614 SmallVector<std::pair<BasicBlock *, pred_iterator>, 16> Stack;
2615 SmallVector<BasicBlock *, 16> Order;
2616 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2618 if (BB->getTerminator()->getNumSuccessors() == 0)
2619 Stack.push_back(std::make_pair(BB, pred_begin(BB)));
2621 while (!Stack.empty()) {
2622 pred_iterator End = pred_end(Stack.back().first);
2623 while (Stack.back().second != End) {
2624 BasicBlock *BB = *Stack.back().second++;
2625 if (Visited.insert(BB))
2626 Stack.push_back(std::make_pair(BB, pred_begin(BB)));
2628 Order.push_back(Stack.pop_back_val().first);
2630 bool BottomUpNestingDetected = false;
2631 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2632 Order.rbegin(), E = Order.rend(); I != E; ++I) {
2633 BasicBlock *BB = *I;
2634 BottomUpNestingDetected |= VisitBottomUp(BB, BBStates, Retains);
2637 // Use regular reverse-postorder for top-down.
2638 bool TopDownNestingDetected = false;
2639 typedef ReversePostOrderTraversal<Function *> RPOTType;
2641 for (RPOTType::rpo_iterator I = RPOT.begin(), E = RPOT.end(); I != E; ++I) {
2642 BasicBlock *BB = *I;
2643 TopDownNestingDetected |= VisitTopDown(BB, BBStates, Releases);
2646 return TopDownNestingDetected && BottomUpNestingDetected;
2649 /// MoveCalls - Move the calls in RetainsToMove and ReleasesToMove.
2650 void ObjCARCOpt::MoveCalls(Value *Arg,
2651 RRInfo &RetainsToMove,
2652 RRInfo &ReleasesToMove,
2653 MapVector<Value *, RRInfo> &Retains,
2654 DenseMap<Value *, RRInfo> &Releases,
2655 SmallVectorImpl<Instruction *> &DeadInsts,
2657 Type *ArgTy = Arg->getType();
2658 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
2660 // Insert the new retain and release calls.
2661 for (SmallPtrSet<Instruction *, 2>::const_iterator
2662 PI = ReleasesToMove.ReverseInsertPts.begin(),
2663 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2664 Instruction *InsertPt = *PI;
2665 Value *MyArg = ArgTy == ParamTy ? Arg :
2666 new BitCastInst(Arg, ParamTy, "", InsertPt);
2668 CallInst::Create(RetainsToMove.IsRetainBlock ?
2669 getRetainBlockCallee(M) : getRetainCallee(M),
2670 MyArg, "", InsertPt);
2671 Call->setDoesNotThrow();
2672 if (!RetainsToMove.IsRetainBlock)
2673 Call->setTailCall();
2675 for (SmallPtrSet<Instruction *, 2>::const_iterator
2676 PI = RetainsToMove.ReverseInsertPts.begin(),
2677 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2678 Instruction *LastUse = *PI;
2679 Instruction *InsertPts[] = { 0, 0, 0 };
2680 if (InvokeInst *II = dyn_cast<InvokeInst>(LastUse)) {
2681 // We can't insert code immediately after an invoke instruction, so
2682 // insert code at the beginning of both successor blocks instead.
2683 // The invoke's return value isn't available in the unwind block,
2684 // but our releases will never depend on it, because they must be
2685 // paired with retains from before the invoke.
2686 InsertPts[0] = II->getNormalDest()->getFirstNonPHI();
2687 InsertPts[1] = II->getUnwindDest()->getFirstNonPHI();
2689 // Insert code immediately after the last use.
2690 InsertPts[0] = llvm::next(BasicBlock::iterator(LastUse));
2693 for (Instruction **I = InsertPts; *I; ++I) {
2694 Instruction *InsertPt = *I;
2695 Value *MyArg = ArgTy == ParamTy ? Arg :
2696 new BitCastInst(Arg, ParamTy, "", InsertPt);
2697 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
2699 // Attach a clang.imprecise_release metadata tag, if appropriate.
2700 if (MDNode *M = ReleasesToMove.ReleaseMetadata)
2701 Call->setMetadata(ImpreciseReleaseMDKind, M);
2702 Call->setDoesNotThrow();
2703 if (ReleasesToMove.IsTailCallRelease)
2704 Call->setTailCall();
2708 // Delete the original retain and release calls.
2709 for (SmallPtrSet<Instruction *, 2>::const_iterator
2710 AI = RetainsToMove.Calls.begin(),
2711 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
2712 Instruction *OrigRetain = *AI;
2713 Retains.blot(OrigRetain);
2714 DeadInsts.push_back(OrigRetain);
2716 for (SmallPtrSet<Instruction *, 2>::const_iterator
2717 AI = ReleasesToMove.Calls.begin(),
2718 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
2719 Instruction *OrigRelease = *AI;
2720 Releases.erase(OrigRelease);
2721 DeadInsts.push_back(OrigRelease);
2726 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
2728 MapVector<Value *, RRInfo> &Retains,
2729 DenseMap<Value *, RRInfo> &Releases,
2731 bool AnyPairsCompletelyEliminated = false;
2732 RRInfo RetainsToMove;
2733 RRInfo ReleasesToMove;
2734 SmallVector<Instruction *, 4> NewRetains;
2735 SmallVector<Instruction *, 4> NewReleases;
2736 SmallVector<Instruction *, 8> DeadInsts;
2738 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
2739 E = Retains.end(); I != E; ) {
2740 Value *V = (I++)->first;
2741 if (!V) continue; // blotted
2743 Instruction *Retain = cast<Instruction>(V);
2744 Value *Arg = GetObjCArg(Retain);
2746 // If the object being released is in static or stack storage, we know it's
2747 // not being managed by ObjC reference counting, so we can delete pairs
2748 // regardless of what possible decrements or uses lie between them.
2749 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg);
2751 // A constant pointer can't be pointing to an object on the heap. It may
2752 // be reference-counted, but it won't be deleted.
2753 if (const LoadInst *LI = dyn_cast<LoadInst>(Arg))
2754 if (const GlobalVariable *GV =
2755 dyn_cast<GlobalVariable>(
2756 StripPointerCastsAndObjCCalls(LI->getPointerOperand())))
2757 if (GV->isConstant())
2760 // If a pair happens in a region where it is known that the reference count
2761 // is already incremented, we can similarly ignore possible decrements.
2762 bool KnownSafeTD = true, KnownSafeBU = true;
2764 // Connect the dots between the top-down-collected RetainsToMove and
2765 // bottom-up-collected ReleasesToMove to form sets of related calls.
2766 // This is an iterative process so that we connect multiple releases
2767 // to multiple retains if needed.
2768 unsigned OldDelta = 0;
2769 unsigned NewDelta = 0;
2770 unsigned OldCount = 0;
2771 unsigned NewCount = 0;
2772 bool FirstRelease = true;
2773 bool FirstRetain = true;
2774 NewRetains.push_back(Retain);
2776 for (SmallVectorImpl<Instruction *>::const_iterator
2777 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
2778 Instruction *NewRetain = *NI;
2779 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
2780 assert(It != Retains.end());
2781 const RRInfo &NewRetainRRI = It->second;
2782 KnownSafeTD &= NewRetainRRI.KnownSafe;
2783 for (SmallPtrSet<Instruction *, 2>::const_iterator
2784 LI = NewRetainRRI.Calls.begin(),
2785 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
2786 Instruction *NewRetainRelease = *LI;
2787 DenseMap<Value *, RRInfo>::const_iterator Jt =
2788 Releases.find(NewRetainRelease);
2789 if (Jt == Releases.end())
2791 const RRInfo &NewRetainReleaseRRI = Jt->second;
2792 assert(NewRetainReleaseRRI.Calls.count(NewRetain));
2793 if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
2795 BBStates[NewRetainRelease->getParent()].GetAllPathCount();
2797 // Merge the ReleaseMetadata and IsTailCallRelease values.
2799 ReleasesToMove.ReleaseMetadata =
2800 NewRetainReleaseRRI.ReleaseMetadata;
2801 ReleasesToMove.IsTailCallRelease =
2802 NewRetainReleaseRRI.IsTailCallRelease;
2803 FirstRelease = false;
2805 if (ReleasesToMove.ReleaseMetadata !=
2806 NewRetainReleaseRRI.ReleaseMetadata)
2807 ReleasesToMove.ReleaseMetadata = 0;
2808 if (ReleasesToMove.IsTailCallRelease !=
2809 NewRetainReleaseRRI.IsTailCallRelease)
2810 ReleasesToMove.IsTailCallRelease = false;
2813 // Collect the optimal insertion points.
2815 for (SmallPtrSet<Instruction *, 2>::const_iterator
2816 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
2817 RE = NewRetainReleaseRRI.ReverseInsertPts.end();
2819 Instruction *RIP = *RI;
2820 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
2821 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
2823 NewReleases.push_back(NewRetainRelease);
2828 if (NewReleases.empty()) break;
2830 // Back the other way.
2831 for (SmallVectorImpl<Instruction *>::const_iterator
2832 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
2833 Instruction *NewRelease = *NI;
2834 DenseMap<Value *, RRInfo>::const_iterator It =
2835 Releases.find(NewRelease);
2836 assert(It != Releases.end());
2837 const RRInfo &NewReleaseRRI = It->second;
2838 KnownSafeBU &= NewReleaseRRI.KnownSafe;
2839 for (SmallPtrSet<Instruction *, 2>::const_iterator
2840 LI = NewReleaseRRI.Calls.begin(),
2841 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
2842 Instruction *NewReleaseRetain = *LI;
2843 MapVector<Value *, RRInfo>::const_iterator Jt =
2844 Retains.find(NewReleaseRetain);
2845 if (Jt == Retains.end())
2847 const RRInfo &NewReleaseRetainRRI = Jt->second;
2848 assert(NewReleaseRetainRRI.Calls.count(NewRelease));
2849 if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
2850 unsigned PathCount =
2851 BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
2852 OldDelta += PathCount;
2853 OldCount += PathCount;
2855 // Merge the IsRetainBlock values.
2857 RetainsToMove.IsRetainBlock = NewReleaseRetainRRI.IsRetainBlock;
2858 FirstRetain = false;
2859 } else if (ReleasesToMove.IsRetainBlock !=
2860 NewReleaseRetainRRI.IsRetainBlock)
2861 // It's not possible to merge the sequences if one uses
2862 // objc_retain and the other uses objc_retainBlock.
2865 // Collect the optimal insertion points.
2867 for (SmallPtrSet<Instruction *, 2>::const_iterator
2868 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
2869 RE = NewReleaseRetainRRI.ReverseInsertPts.end();
2871 Instruction *RIP = *RI;
2872 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
2873 PathCount = BBStates[RIP->getParent()].GetAllPathCount();
2874 NewDelta += PathCount;
2875 NewCount += PathCount;
2878 NewRetains.push_back(NewReleaseRetain);
2882 NewReleases.clear();
2883 if (NewRetains.empty()) break;
2886 // If the pointer is known incremented or nested, we can safely delete the
2887 // pair regardless of what's between them.
2888 if (KnownSafeTD || KnownSafeBU) {
2889 RetainsToMove.ReverseInsertPts.clear();
2890 ReleasesToMove.ReverseInsertPts.clear();
2893 // Determine whether the new insertion points we computed preserve the
2894 // balance of retain and release calls through the program.
2895 // TODO: If the fully aggressive solution isn't valid, try to find a
2896 // less aggressive solution which is.
2901 // Determine whether the original call points are balanced in the retain and
2902 // release calls through the program. If not, conservatively don't touch
2904 // TODO: It's theoretically possible to do code motion in this case, as
2905 // long as the existing imbalances are maintained.
2909 // Ok, everything checks out and we're all set. Let's move some code!
2911 AnyPairsCompletelyEliminated = NewCount == 0;
2912 NumRRs += OldCount - NewCount;
2913 MoveCalls(Arg, RetainsToMove, ReleasesToMove,
2914 Retains, Releases, DeadInsts, M);
2917 NewReleases.clear();
2919 RetainsToMove.clear();
2920 ReleasesToMove.clear();
2923 // Now that we're done moving everything, we can delete the newly dead
2924 // instructions, as we no longer need them as insert points.
2925 while (!DeadInsts.empty())
2926 EraseInstruction(DeadInsts.pop_back_val());
2928 return AnyPairsCompletelyEliminated;
2931 /// OptimizeWeakCalls - Weak pointer optimizations.
2932 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
2933 // First, do memdep-style RLE and S2L optimizations. We can't use memdep
2934 // itself because it uses AliasAnalysis and we need to do provenance
2936 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2937 Instruction *Inst = &*I++;
2938 InstructionClass Class = GetBasicInstructionClass(Inst);
2939 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
2942 // Delete objc_loadWeak calls with no users.
2943 if (Class == IC_LoadWeak && Inst->use_empty()) {
2944 Inst->eraseFromParent();
2948 // TODO: For now, just look for an earlier available version of this value
2949 // within the same block. Theoretically, we could do memdep-style non-local
2950 // analysis too, but that would want caching. A better approach would be to
2951 // use the technique that EarlyCSE uses.
2952 inst_iterator Current = llvm::prior(I);
2953 BasicBlock *CurrentBB = Current.getBasicBlockIterator();
2954 for (BasicBlock::iterator B = CurrentBB->begin(),
2955 J = Current.getInstructionIterator();
2957 Instruction *EarlierInst = &*llvm::prior(J);
2958 InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
2959 switch (EarlierClass) {
2961 case IC_LoadWeakRetained: {
2962 // If this is loading from the same pointer, replace this load's value
2964 CallInst *Call = cast<CallInst>(Inst);
2965 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2966 Value *Arg = Call->getArgOperand(0);
2967 Value *EarlierArg = EarlierCall->getArgOperand(0);
2968 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2969 case AliasAnalysis::MustAlias:
2971 // If the load has a builtin retain, insert a plain retain for it.
2972 if (Class == IC_LoadWeakRetained) {
2974 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2978 // Zap the fully redundant load.
2979 Call->replaceAllUsesWith(EarlierCall);
2980 Call->eraseFromParent();
2982 case AliasAnalysis::MayAlias:
2983 case AliasAnalysis::PartialAlias:
2985 case AliasAnalysis::NoAlias:
2992 // If this is storing to the same pointer and has the same size etc.
2993 // replace this load's value with the stored value.
2994 CallInst *Call = cast<CallInst>(Inst);
2995 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2996 Value *Arg = Call->getArgOperand(0);
2997 Value *EarlierArg = EarlierCall->getArgOperand(0);
2998 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2999 case AliasAnalysis::MustAlias:
3001 // If the load has a builtin retain, insert a plain retain for it.
3002 if (Class == IC_LoadWeakRetained) {
3004 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
3008 // Zap the fully redundant load.
3009 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
3010 Call->eraseFromParent();
3012 case AliasAnalysis::MayAlias:
3013 case AliasAnalysis::PartialAlias:
3015 case AliasAnalysis::NoAlias:
3022 // TOOD: Grab the copied value.
3024 case IC_AutoreleasepoolPush:
3027 // Weak pointers are only modified through the weak entry points
3028 // (and arbitrary calls, which could call the weak entry points).
3031 // Anything else could modify the weak pointer.
3038 // Then, for each destroyWeak with an alloca operand, check to see if
3039 // the alloca and all its users can be zapped.
3040 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3041 Instruction *Inst = &*I++;
3042 InstructionClass Class = GetBasicInstructionClass(Inst);
3043 if (Class != IC_DestroyWeak)
3046 CallInst *Call = cast<CallInst>(Inst);
3047 Value *Arg = Call->getArgOperand(0);
3048 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
3049 for (Value::use_iterator UI = Alloca->use_begin(),
3050 UE = Alloca->use_end(); UI != UE; ++UI) {
3051 Instruction *UserInst = cast<Instruction>(*UI);
3052 switch (GetBasicInstructionClass(UserInst)) {
3055 case IC_DestroyWeak:
3062 for (Value::use_iterator UI = Alloca->use_begin(),
3063 UE = Alloca->use_end(); UI != UE; ) {
3064 CallInst *UserInst = cast<CallInst>(*UI++);
3065 if (!UserInst->use_empty())
3066 UserInst->replaceAllUsesWith(UserInst->getOperand(1));
3067 UserInst->eraseFromParent();
3069 Alloca->eraseFromParent();
3075 /// OptimizeSequences - Identify program paths which execute sequences of
3076 /// retains and releases which can be eliminated.
3077 bool ObjCARCOpt::OptimizeSequences(Function &F) {
3078 /// Releases, Retains - These are used to store the results of the main flow
3079 /// analysis. These use Value* as the key instead of Instruction* so that the
3080 /// map stays valid when we get around to rewriting code and calls get
3081 /// replaced by arguments.
3082 DenseMap<Value *, RRInfo> Releases;
3083 MapVector<Value *, RRInfo> Retains;
3085 /// BBStates, This is used during the traversal of the function to track the
3086 /// states for each identified object at each block.
3087 DenseMap<const BasicBlock *, BBState> BBStates;
3089 // Analyze the CFG of the function, and all instructions.
3090 bool NestingDetected = Visit(F, BBStates, Retains, Releases);
3093 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
3097 /// OptimizeReturns - Look for this pattern:
3099 /// %call = call i8* @something(...)
3100 /// %2 = call i8* @objc_retain(i8* %call)
3101 /// %3 = call i8* @objc_autorelease(i8* %2)
3104 /// And delete the retain and autorelease.
3106 /// Otherwise if it's just this:
3108 /// %3 = call i8* @objc_autorelease(i8* %2)
3111 /// convert the autorelease to autoreleaseRV.
3112 void ObjCARCOpt::OptimizeReturns(Function &F) {
3113 if (!F.getReturnType()->isPointerTy())
3116 SmallPtrSet<Instruction *, 4> DependingInstructions;
3117 SmallPtrSet<const BasicBlock *, 4> Visited;
3118 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
3119 BasicBlock *BB = FI;
3120 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
3123 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
3124 FindDependencies(NeedsPositiveRetainCount, Arg,
3125 BB, Ret, DependingInstructions, Visited, PA);
3126 if (DependingInstructions.size() != 1)
3130 CallInst *Autorelease =
3131 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3134 InstructionClass AutoreleaseClass =
3135 GetBasicInstructionClass(Autorelease);
3136 if (!IsAutorelease(AutoreleaseClass))
3138 if (GetObjCArg(Autorelease) != Arg)
3141 DependingInstructions.clear();
3144 // Check that there is nothing that can affect the reference
3145 // count between the autorelease and the retain.
3146 FindDependencies(CanChangeRetainCount, Arg,
3147 BB, Autorelease, DependingInstructions, Visited, PA);
3148 if (DependingInstructions.size() != 1)
3153 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3155 // Check that we found a retain with the same argument.
3157 !IsRetain(GetBasicInstructionClass(Retain)) ||
3158 GetObjCArg(Retain) != Arg)
3161 DependingInstructions.clear();
3164 // Convert the autorelease to an autoreleaseRV, since it's
3165 // returning the value.
3166 if (AutoreleaseClass == IC_Autorelease) {
3167 Autorelease->setCalledFunction(getAutoreleaseRVCallee(F.getParent()));
3168 AutoreleaseClass = IC_AutoreleaseRV;
3171 // Check that there is nothing that can affect the reference
3172 // count between the retain and the call.
3173 FindDependencies(CanChangeRetainCount, Arg, BB, Retain,
3174 DependingInstructions, Visited, PA);
3175 if (DependingInstructions.size() != 1)
3180 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3182 // Check that the pointer is the return value of the call.
3183 if (!Call || Arg != Call)
3186 // Check that the call is a regular call.
3187 InstructionClass Class = GetBasicInstructionClass(Call);
3188 if (Class != IC_CallOrUser && Class != IC_Call)
3191 // If so, we can zap the retain and autorelease.
3194 EraseInstruction(Retain);
3195 EraseInstruction(Autorelease);
3201 DependingInstructions.clear();
3206 bool ObjCARCOpt::doInitialization(Module &M) {
3210 Run = ModuleHasARC(M);
3214 // Identify the imprecise release metadata kind.
3215 ImpreciseReleaseMDKind =
3216 M.getContext().getMDKindID("clang.imprecise_release");
3218 // Intuitively, objc_retain and others are nocapture, however in practice
3219 // they are not, because they return their argument value. And objc_release
3220 // calls finalizers.
3222 // These are initialized lazily.
3224 AutoreleaseRVCallee = 0;
3227 RetainBlockCallee = 0;
3228 AutoreleaseCallee = 0;
3233 bool ObjCARCOpt::runOnFunction(Function &F) {
3237 // If nothing in the Module uses ARC, don't do anything.
3243 PA.setAA(&getAnalysis<AliasAnalysis>());
3245 // This pass performs several distinct transformations. As a compile-time aid
3246 // when compiling code that isn't ObjC, skip these if the relevant ObjC
3247 // library functions aren't declared.
3249 // Preliminary optimizations. This also computs UsedInThisFunction.
3250 OptimizeIndividualCalls(F);
3252 // Optimizations for weak pointers.
3253 if (UsedInThisFunction & ((1 << IC_LoadWeak) |
3254 (1 << IC_LoadWeakRetained) |
3255 (1 << IC_StoreWeak) |
3256 (1 << IC_InitWeak) |
3257 (1 << IC_CopyWeak) |
3258 (1 << IC_MoveWeak) |
3259 (1 << IC_DestroyWeak)))
3260 OptimizeWeakCalls(F);
3262 // Optimizations for retain+release pairs.
3263 if (UsedInThisFunction & ((1 << IC_Retain) |
3264 (1 << IC_RetainRV) |
3265 (1 << IC_RetainBlock)))
3266 if (UsedInThisFunction & (1 << IC_Release))
3267 // Run OptimizeSequences until it either stops making changes or
3268 // no retain+release pair nesting is detected.
3269 while (OptimizeSequences(F)) {}
3271 // Optimizations if objc_autorelease is used.
3272 if (UsedInThisFunction &
3273 ((1 << IC_Autorelease) | (1 << IC_AutoreleaseRV)))
3279 void ObjCARCOpt::releaseMemory() {
3283 //===----------------------------------------------------------------------===//
3285 //===----------------------------------------------------------------------===//
3287 // TODO: ObjCARCContract could insert PHI nodes when uses aren't
3288 // dominated by single calls.
3290 #include "llvm/Operator.h"
3291 #include "llvm/InlineAsm.h"
3292 #include "llvm/Analysis/Dominators.h"
3294 STATISTIC(NumStoreStrongs, "Number objc_storeStrong calls formed");
3297 /// ObjCARCContract - Late ARC optimizations. These change the IR in a way
3298 /// that makes it difficult to be analyzed by ObjCARCOpt, so it's run late.
3299 class ObjCARCContract : public FunctionPass {
3303 ProvenanceAnalysis PA;
3305 /// Run - A flag indicating whether this optimization pass should run.
3308 /// StoreStrongCallee, etc. - Declarations for ObjC runtime
3309 /// functions, for use in creating calls to them. These are initialized
3310 /// lazily to avoid cluttering up the Module with unused declarations.
3311 Constant *StoreStrongCallee,
3312 *RetainAutoreleaseCallee, *RetainAutoreleaseRVCallee;
3314 /// RetainRVMarker - The inline asm string to insert between calls and
3315 /// RetainRV calls to make the optimization work on targets which need it.
3316 const MDString *RetainRVMarker;
3318 Constant *getStoreStrongCallee(Module *M);
3319 Constant *getRetainAutoreleaseCallee(Module *M);
3320 Constant *getRetainAutoreleaseRVCallee(Module *M);
3322 bool ContractAutorelease(Function &F, Instruction *Autorelease,
3323 InstructionClass Class,
3324 SmallPtrSet<Instruction *, 4>
3325 &DependingInstructions,
3326 SmallPtrSet<const BasicBlock *, 4>
3329 void ContractRelease(Instruction *Release,
3330 inst_iterator &Iter);
3332 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
3333 virtual bool doInitialization(Module &M);
3334 virtual bool runOnFunction(Function &F);
3338 ObjCARCContract() : FunctionPass(ID) {
3339 initializeObjCARCContractPass(*PassRegistry::getPassRegistry());
3344 char ObjCARCContract::ID = 0;
3345 INITIALIZE_PASS_BEGIN(ObjCARCContract,
3346 "objc-arc-contract", "ObjC ARC contraction", false, false)
3347 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3348 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3349 INITIALIZE_PASS_END(ObjCARCContract,
3350 "objc-arc-contract", "ObjC ARC contraction", false, false)
3352 Pass *llvm::createObjCARCContractPass() {
3353 return new ObjCARCContract();
3356 void ObjCARCContract::getAnalysisUsage(AnalysisUsage &AU) const {
3357 AU.addRequired<AliasAnalysis>();
3358 AU.addRequired<DominatorTree>();
3359 AU.setPreservesCFG();
3362 Constant *ObjCARCContract::getStoreStrongCallee(Module *M) {
3363 if (!StoreStrongCallee) {
3364 LLVMContext &C = M->getContext();
3365 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3366 Type *I8XX = PointerType::getUnqual(I8X);
3367 std::vector<Type *> Params;
3368 Params.push_back(I8XX);
3369 Params.push_back(I8X);
3371 AttrListPtr Attributes;
3372 Attributes.addAttr(~0u, Attribute::NoUnwind);
3373 Attributes.addAttr(1, Attribute::NoCapture);
3376 M->getOrInsertFunction(
3378 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
3381 return StoreStrongCallee;
3384 Constant *ObjCARCContract::getRetainAutoreleaseCallee(Module *M) {
3385 if (!RetainAutoreleaseCallee) {
3386 LLVMContext &C = M->getContext();
3387 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3388 std::vector<Type *> Params;
3389 Params.push_back(I8X);
3391 FunctionType::get(I8X, Params, /*isVarArg=*/false);
3392 AttrListPtr Attributes;
3393 Attributes.addAttr(~0u, Attribute::NoUnwind);
3394 RetainAutoreleaseCallee =
3395 M->getOrInsertFunction("objc_retainAutorelease", FTy, Attributes);
3397 return RetainAutoreleaseCallee;
3400 Constant *ObjCARCContract::getRetainAutoreleaseRVCallee(Module *M) {
3401 if (!RetainAutoreleaseRVCallee) {
3402 LLVMContext &C = M->getContext();
3403 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3404 std::vector<Type *> Params;
3405 Params.push_back(I8X);
3407 FunctionType::get(I8X, Params, /*isVarArg=*/false);
3408 AttrListPtr Attributes;
3409 Attributes.addAttr(~0u, Attribute::NoUnwind);
3410 RetainAutoreleaseRVCallee =
3411 M->getOrInsertFunction("objc_retainAutoreleaseReturnValue", FTy,
3414 return RetainAutoreleaseRVCallee;
3417 /// ContractAutorelease - Merge an autorelease with a retain into a fused
3420 ObjCARCContract::ContractAutorelease(Function &F, Instruction *Autorelease,
3421 InstructionClass Class,
3422 SmallPtrSet<Instruction *, 4>
3423 &DependingInstructions,
3424 SmallPtrSet<const BasicBlock *, 4>
3426 const Value *Arg = GetObjCArg(Autorelease);
3428 // Check that there are no instructions between the retain and the autorelease
3429 // (such as an autorelease_pop) which may change the count.
3430 CallInst *Retain = 0;
3431 if (Class == IC_AutoreleaseRV)
3432 FindDependencies(RetainAutoreleaseRVDep, Arg,
3433 Autorelease->getParent(), Autorelease,
3434 DependingInstructions, Visited, PA);
3436 FindDependencies(RetainAutoreleaseDep, Arg,
3437 Autorelease->getParent(), Autorelease,
3438 DependingInstructions, Visited, PA);
3441 if (DependingInstructions.size() != 1) {
3442 DependingInstructions.clear();
3446 Retain = dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3447 DependingInstructions.clear();
3450 GetBasicInstructionClass(Retain) != IC_Retain ||
3451 GetObjCArg(Retain) != Arg)
3457 if (Class == IC_AutoreleaseRV)
3458 Retain->setCalledFunction(getRetainAutoreleaseRVCallee(F.getParent()));
3460 Retain->setCalledFunction(getRetainAutoreleaseCallee(F.getParent()));
3462 EraseInstruction(Autorelease);
3466 /// ContractRelease - Attempt to merge an objc_release with a store, load, and
3467 /// objc_retain to form an objc_storeStrong. This can be a little tricky because
3468 /// the instructions don't always appear in order, and there may be unrelated
3469 /// intervening instructions.
3470 void ObjCARCContract::ContractRelease(Instruction *Release,
3471 inst_iterator &Iter) {
3472 LoadInst *Load = dyn_cast<LoadInst>(GetObjCArg(Release));
3473 if (!Load || Load->isVolatile()) return;
3475 // For now, require everything to be in one basic block.
3476 BasicBlock *BB = Release->getParent();
3477 if (Load->getParent() != BB) return;
3479 // Walk down to find the store.
3480 BasicBlock::iterator I = Load, End = BB->end();
3482 AliasAnalysis::Location Loc = AA->getLocation(Load);
3485 IsRetain(GetBasicInstructionClass(I)) ||
3486 !(AA->getModRefInfo(I, Loc) & AliasAnalysis::Mod)))
3488 StoreInst *Store = dyn_cast<StoreInst>(I);
3489 if (!Store || Store->isVolatile()) return;
3490 if (Store->getPointerOperand() != Loc.Ptr) return;
3492 Value *New = StripPointerCastsAndObjCCalls(Store->getValueOperand());
3494 // Walk up to find the retain.
3496 BasicBlock::iterator Begin = BB->begin();
3497 while (I != Begin && GetBasicInstructionClass(I) != IC_Retain)
3499 Instruction *Retain = I;
3500 if (GetBasicInstructionClass(Retain) != IC_Retain) return;
3501 if (GetObjCArg(Retain) != New) return;
3506 LLVMContext &C = Release->getContext();
3507 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3508 Type *I8XX = PointerType::getUnqual(I8X);
3510 Value *Args[] = { Load->getPointerOperand(), New };
3511 if (Args[0]->getType() != I8XX)
3512 Args[0] = new BitCastInst(Args[0], I8XX, "", Store);
3513 if (Args[1]->getType() != I8X)
3514 Args[1] = new BitCastInst(Args[1], I8X, "", Store);
3515 CallInst *StoreStrong =
3516 CallInst::Create(getStoreStrongCallee(BB->getParent()->getParent()),
3518 StoreStrong->setDoesNotThrow();
3519 StoreStrong->setDebugLoc(Store->getDebugLoc());
3521 if (&*Iter == Store) ++Iter;
3522 Store->eraseFromParent();
3523 Release->eraseFromParent();
3524 EraseInstruction(Retain);
3525 if (Load->use_empty())
3526 Load->eraseFromParent();
3529 bool ObjCARCContract::doInitialization(Module &M) {
3530 Run = ModuleHasARC(M);
3534 // These are initialized lazily.
3535 StoreStrongCallee = 0;
3536 RetainAutoreleaseCallee = 0;
3537 RetainAutoreleaseRVCallee = 0;
3539 // Initialize RetainRVMarker.
3541 if (NamedMDNode *NMD =
3542 M.getNamedMetadata("clang.arc.retainAutoreleasedReturnValueMarker"))
3543 if (NMD->getNumOperands() == 1) {
3544 const MDNode *N = NMD->getOperand(0);
3545 if (N->getNumOperands() == 1)
3546 if (const MDString *S = dyn_cast<MDString>(N->getOperand(0)))
3553 bool ObjCARCContract::runOnFunction(Function &F) {
3557 // If nothing in the Module uses ARC, don't do anything.
3562 AA = &getAnalysis<AliasAnalysis>();
3563 DT = &getAnalysis<DominatorTree>();
3565 PA.setAA(&getAnalysis<AliasAnalysis>());
3567 // For ObjC library calls which return their argument, replace uses of the
3568 // argument with uses of the call return value, if it dominates the use. This
3569 // reduces register pressure.
3570 SmallPtrSet<Instruction *, 4> DependingInstructions;
3571 SmallPtrSet<const BasicBlock *, 4> Visited;
3572 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3573 Instruction *Inst = &*I++;
3575 // Only these library routines return their argument. In particular,
3576 // objc_retainBlock does not necessarily return its argument.
3577 InstructionClass Class = GetBasicInstructionClass(Inst);
3580 case IC_FusedRetainAutorelease:
3581 case IC_FusedRetainAutoreleaseRV:
3583 case IC_Autorelease:
3584 case IC_AutoreleaseRV:
3585 if (ContractAutorelease(F, Inst, Class, DependingInstructions, Visited))
3589 // If we're compiling for a target which needs a special inline-asm
3590 // marker to do the retainAutoreleasedReturnValue optimization,
3592 if (!RetainRVMarker)
3594 BasicBlock::iterator BBI = Inst;
3596 while (isNoopInstruction(BBI)) --BBI;
3597 if (&*BBI == GetObjCArg(Inst)) {
3599 InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
3600 /*isVarArg=*/false),
3601 RetainRVMarker->getString(),
3602 /*Constraints=*/"", /*hasSideEffects=*/true);
3603 CallInst::Create(IA, "", Inst);
3608 // objc_initWeak(p, null) => *p = null
3609 CallInst *CI = cast<CallInst>(Inst);
3610 if (isNullOrUndef(CI->getArgOperand(1))) {
3612 ConstantPointerNull::get(cast<PointerType>(CI->getType()));
3614 new StoreInst(Null, CI->getArgOperand(0), CI);
3615 CI->replaceAllUsesWith(Null);
3616 CI->eraseFromParent();
3621 ContractRelease(Inst, I);
3627 // Don't use GetObjCArg because we don't want to look through bitcasts
3628 // and such; to do the replacement, the argument must have type i8*.
3629 const Value *Arg = cast<CallInst>(Inst)->getArgOperand(0);
3631 // If we're compiling bugpointed code, don't get in trouble.
3632 if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))
3634 // Look through the uses of the pointer.
3635 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
3637 Use &U = UI.getUse();
3638 unsigned OperandNo = UI.getOperandNo();
3639 ++UI; // Increment UI now, because we may unlink its element.
3640 if (Instruction *UserInst = dyn_cast<Instruction>(U.getUser()))
3641 if (Inst != UserInst && DT->dominates(Inst, UserInst)) {
3643 Instruction *Replacement = Inst;
3644 Type *UseTy = U.get()->getType();
3645 if (PHINode *PHI = dyn_cast<PHINode>(UserInst)) {
3646 // For PHI nodes, insert the bitcast in the predecessor block.
3648 PHINode::getIncomingValueNumForOperand(OperandNo);
3650 PHI->getIncomingBlock(ValNo);
3651 if (Replacement->getType() != UseTy)
3652 Replacement = new BitCastInst(Replacement, UseTy, "",
3654 for (unsigned i = 0, e = PHI->getNumIncomingValues();
3656 if (PHI->getIncomingBlock(i) == BB) {
3657 // Keep the UI iterator valid.
3658 if (&PHI->getOperandUse(
3659 PHINode::getOperandNumForIncomingValue(i)) ==
3662 PHI->setIncomingValue(i, Replacement);
3665 if (Replacement->getType() != UseTy)
3666 Replacement = new BitCastInst(Replacement, UseTy, "", UserInst);
3672 // If Arg is a no-op casted pointer, strip one level of casts and
3674 if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg))
3675 Arg = BI->getOperand(0);
3676 else if (isa<GEPOperator>(Arg) &&
3677 cast<GEPOperator>(Arg)->hasAllZeroIndices())
3678 Arg = cast<GEPOperator>(Arg)->getPointerOperand();
3679 else if (isa<GlobalAlias>(Arg) &&
3680 !cast<GlobalAlias>(Arg)->mayBeOverridden())
3681 Arg = cast<GlobalAlias>(Arg)->getAliasee();