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 StringRef Name = GV->getName();
519 // These special variables are known to hold values which are not
520 // reference-counted pointers.
521 if (Name.startswith("\01L_OBJC_SELECTOR_REFERENCES_") ||
522 Name.startswith("\01L_OBJC_CLASSLIST_REFERENCES_") ||
523 Name.startswith("\01L_OBJC_CLASSLIST_SUP_REFS_$_") ||
524 Name.startswith("\01L_OBJC_METH_VAR_NAME_") ||
525 Name.startswith("\01l_objc_msgSend_fixup_"))
533 /// FindSingleUseIdentifiedObject - This is similar to
534 /// StripPointerCastsAndObjCCalls but it stops as soon as it finds a value
535 /// with multiple uses.
536 static const Value *FindSingleUseIdentifiedObject(const Value *Arg) {
537 if (Arg->hasOneUse()) {
538 if (const BitCastInst *BC = dyn_cast<BitCastInst>(Arg))
539 return FindSingleUseIdentifiedObject(BC->getOperand(0));
540 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Arg))
541 if (GEP->hasAllZeroIndices())
542 return FindSingleUseIdentifiedObject(GEP->getPointerOperand());
543 if (IsForwarding(GetBasicInstructionClass(Arg)))
544 return FindSingleUseIdentifiedObject(
545 cast<CallInst>(Arg)->getArgOperand(0));
546 if (!IsObjCIdentifiedObject(Arg))
551 // If we found an identifiable object but it has multiple uses, but they
552 // are trivial uses, we can still consider this to be a single-use
554 if (IsObjCIdentifiedObject(Arg)) {
555 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
558 if (!U->use_empty() || StripPointerCastsAndObjCCalls(U) != Arg)
568 /// ModuleHasARC - Test if the given module looks interesting to run ARC
570 static bool ModuleHasARC(const Module &M) {
572 M.getNamedValue("objc_retain") ||
573 M.getNamedValue("objc_release") ||
574 M.getNamedValue("objc_autorelease") ||
575 M.getNamedValue("objc_retainAutoreleasedReturnValue") ||
576 M.getNamedValue("objc_retainBlock") ||
577 M.getNamedValue("objc_autoreleaseReturnValue") ||
578 M.getNamedValue("objc_autoreleasePoolPush") ||
579 M.getNamedValue("objc_loadWeakRetained") ||
580 M.getNamedValue("objc_loadWeak") ||
581 M.getNamedValue("objc_destroyWeak") ||
582 M.getNamedValue("objc_storeWeak") ||
583 M.getNamedValue("objc_initWeak") ||
584 M.getNamedValue("objc_moveWeak") ||
585 M.getNamedValue("objc_copyWeak") ||
586 M.getNamedValue("objc_retainedObject") ||
587 M.getNamedValue("objc_unretainedObject") ||
588 M.getNamedValue("objc_unretainedPointer");
591 //===----------------------------------------------------------------------===//
592 // ARC AliasAnalysis.
593 //===----------------------------------------------------------------------===//
595 #include "llvm/Pass.h"
596 #include "llvm/Analysis/AliasAnalysis.h"
597 #include "llvm/Analysis/Passes.h"
600 /// ObjCARCAliasAnalysis - This is a simple alias analysis
601 /// implementation that uses knowledge of ARC constructs to answer queries.
603 /// TODO: This class could be generalized to know about other ObjC-specific
604 /// tricks. Such as knowing that ivars in the non-fragile ABI are non-aliasing
605 /// even though their offsets are dynamic.
606 class ObjCARCAliasAnalysis : public ImmutablePass,
607 public AliasAnalysis {
609 static char ID; // Class identification, replacement for typeinfo
610 ObjCARCAliasAnalysis() : ImmutablePass(ID) {
611 initializeObjCARCAliasAnalysisPass(*PassRegistry::getPassRegistry());
615 virtual void initializePass() {
616 InitializeAliasAnalysis(this);
619 /// getAdjustedAnalysisPointer - This method is used when a pass implements
620 /// an analysis interface through multiple inheritance. If needed, it
621 /// should override this to adjust the this pointer as needed for the
622 /// specified pass info.
623 virtual void *getAdjustedAnalysisPointer(const void *PI) {
624 if (PI == &AliasAnalysis::ID)
625 return (AliasAnalysis*)this;
629 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
630 virtual AliasResult alias(const Location &LocA, const Location &LocB);
631 virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
632 virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
633 virtual ModRefBehavior getModRefBehavior(const Function *F);
634 virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
635 const Location &Loc);
636 virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
637 ImmutableCallSite CS2);
639 } // End of anonymous namespace
641 // Register this pass...
642 char ObjCARCAliasAnalysis::ID = 0;
643 INITIALIZE_AG_PASS(ObjCARCAliasAnalysis, AliasAnalysis, "objc-arc-aa",
644 "ObjC-ARC-Based Alias Analysis", false, true, false)
646 ImmutablePass *llvm::createObjCARCAliasAnalysisPass() {
647 return new ObjCARCAliasAnalysis();
651 ObjCARCAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
652 AU.setPreservesAll();
653 AliasAnalysis::getAnalysisUsage(AU);
656 AliasAnalysis::AliasResult
657 ObjCARCAliasAnalysis::alias(const Location &LocA, const Location &LocB) {
659 return AliasAnalysis::alias(LocA, LocB);
661 // First, strip off no-ops, including ObjC-specific no-ops, and try making a
662 // precise alias query.
663 const Value *SA = StripPointerCastsAndObjCCalls(LocA.Ptr);
664 const Value *SB = StripPointerCastsAndObjCCalls(LocB.Ptr);
666 AliasAnalysis::alias(Location(SA, LocA.Size, LocA.TBAATag),
667 Location(SB, LocB.Size, LocB.TBAATag));
668 if (Result != MayAlias)
671 // If that failed, climb to the underlying object, including climbing through
672 // ObjC-specific no-ops, and try making an imprecise alias query.
673 const Value *UA = GetUnderlyingObjCPtr(SA);
674 const Value *UB = GetUnderlyingObjCPtr(SB);
675 if (UA != SA || UB != SB) {
676 Result = AliasAnalysis::alias(Location(UA), Location(UB));
677 // We can't use MustAlias or PartialAlias results here because
678 // GetUnderlyingObjCPtr may return an offsetted pointer value.
679 if (Result == NoAlias)
683 // If that failed, fail. We don't need to chain here, since that's covered
684 // by the earlier precise query.
689 ObjCARCAliasAnalysis::pointsToConstantMemory(const Location &Loc,
692 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
694 // First, strip off no-ops, including ObjC-specific no-ops, and try making
695 // a precise alias query.
696 const Value *S = StripPointerCastsAndObjCCalls(Loc.Ptr);
697 if (AliasAnalysis::pointsToConstantMemory(Location(S, Loc.Size, Loc.TBAATag),
701 // If that failed, climb to the underlying object, including climbing through
702 // ObjC-specific no-ops, and try making an imprecise alias query.
703 const Value *U = GetUnderlyingObjCPtr(S);
705 return AliasAnalysis::pointsToConstantMemory(Location(U), OrLocal);
707 // If that failed, fail. We don't need to chain here, since that's covered
708 // by the earlier precise query.
712 AliasAnalysis::ModRefBehavior
713 ObjCARCAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
714 // We have nothing to do. Just chain to the next AliasAnalysis.
715 return AliasAnalysis::getModRefBehavior(CS);
718 AliasAnalysis::ModRefBehavior
719 ObjCARCAliasAnalysis::getModRefBehavior(const Function *F) {
721 return AliasAnalysis::getModRefBehavior(F);
723 switch (GetFunctionClass(F)) {
725 return DoesNotAccessMemory;
730 return AliasAnalysis::getModRefBehavior(F);
733 AliasAnalysis::ModRefResult
734 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS, const Location &Loc) {
736 return AliasAnalysis::getModRefInfo(CS, Loc);
738 switch (GetBasicInstructionClass(CS.getInstruction())) {
743 case IC_AutoreleaseRV:
745 case IC_AutoreleasepoolPush:
746 case IC_FusedRetainAutorelease:
747 case IC_FusedRetainAutoreleaseRV:
748 // These functions don't access any memory visible to the compiler.
754 return AliasAnalysis::getModRefInfo(CS, Loc);
757 AliasAnalysis::ModRefResult
758 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS1,
759 ImmutableCallSite CS2) {
760 // TODO: Theoretically we could check for dependencies between objc_* calls
761 // and OnlyAccessesArgumentPointees calls or other well-behaved calls.
762 return AliasAnalysis::getModRefInfo(CS1, CS2);
765 //===----------------------------------------------------------------------===//
767 //===----------------------------------------------------------------------===//
769 #include "llvm/Support/InstIterator.h"
770 #include "llvm/Transforms/Scalar.h"
773 /// ObjCARCExpand - Early ARC transformations.
774 class ObjCARCExpand : public FunctionPass {
775 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
776 virtual bool doInitialization(Module &M);
777 virtual bool runOnFunction(Function &F);
779 /// Run - A flag indicating whether this optimization pass should run.
784 ObjCARCExpand() : FunctionPass(ID) {
785 initializeObjCARCExpandPass(*PassRegistry::getPassRegistry());
790 char ObjCARCExpand::ID = 0;
791 INITIALIZE_PASS(ObjCARCExpand,
792 "objc-arc-expand", "ObjC ARC expansion", false, false)
794 Pass *llvm::createObjCARCExpandPass() {
795 return new ObjCARCExpand();
798 void ObjCARCExpand::getAnalysisUsage(AnalysisUsage &AU) const {
799 AU.setPreservesCFG();
802 bool ObjCARCExpand::doInitialization(Module &M) {
803 Run = ModuleHasARC(M);
807 bool ObjCARCExpand::runOnFunction(Function &F) {
811 // If nothing in the Module uses ARC, don't do anything.
815 bool Changed = false;
817 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I) {
818 Instruction *Inst = &*I;
820 switch (GetBasicInstructionClass(Inst)) {
824 case IC_AutoreleaseRV:
825 case IC_FusedRetainAutorelease:
826 case IC_FusedRetainAutoreleaseRV:
827 // These calls return their argument verbatim, as a low-level
828 // optimization. However, this makes high-level optimizations
829 // harder. Undo any uses of this optimization that the front-end
830 // emitted here. We'll redo them in a later pass.
832 Inst->replaceAllUsesWith(cast<CallInst>(Inst)->getArgOperand(0));
842 //===----------------------------------------------------------------------===//
844 //===----------------------------------------------------------------------===//
846 // TODO: On code like this:
849 // stuff_that_cannot_release()
850 // objc_autorelease(%x)
851 // stuff_that_cannot_release()
853 // stuff_that_cannot_release()
854 // objc_autorelease(%x)
856 // The second retain and autorelease can be deleted.
858 // TODO: It should be possible to delete
859 // objc_autoreleasePoolPush and objc_autoreleasePoolPop
860 // pairs if nothing is actually autoreleased between them. Also, autorelease
861 // calls followed by objc_autoreleasePoolPop calls (perhaps in ObjC++ code
862 // after inlining) can be turned into plain release calls.
864 // TODO: Critical-edge splitting. If the optimial insertion point is
865 // a critical edge, the current algorithm has to fail, because it doesn't
866 // know how to split edges. It should be possible to make the optimizer
867 // think in terms of edges, rather than blocks, and then split critical
870 // TODO: OptimizeSequences could generalized to be Interprocedural.
872 // TODO: Recognize that a bunch of other objc runtime calls have
873 // non-escaping arguments and non-releasing arguments, and may be
874 // non-autoreleasing.
876 // TODO: Sink autorelease calls as far as possible. Unfortunately we
877 // usually can't sink them past other calls, which would be the main
878 // case where it would be useful.
880 // TODO: The pointer returned from objc_loadWeakRetained is retained.
882 // TODO: Delete release+retain pairs (rare).
884 #include "llvm/GlobalAlias.h"
885 #include "llvm/Constants.h"
886 #include "llvm/LLVMContext.h"
887 #include "llvm/Support/ErrorHandling.h"
888 #include "llvm/Support/CFG.h"
889 #include "llvm/ADT/PostOrderIterator.h"
890 #include "llvm/ADT/Statistic.h"
892 STATISTIC(NumNoops, "Number of no-op objc calls eliminated");
893 STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated");
894 STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases");
895 STATISTIC(NumRets, "Number of return value forwarding "
896 "retain+autoreleaes eliminated");
897 STATISTIC(NumRRs, "Number of retain+release paths eliminated");
898 STATISTIC(NumPeeps, "Number of calls peephole-optimized");
901 /// ProvenanceAnalysis - This is similar to BasicAliasAnalysis, and it
902 /// uses many of the same techniques, except it uses special ObjC-specific
903 /// reasoning about pointer relationships.
904 class ProvenanceAnalysis {
907 typedef std::pair<const Value *, const Value *> ValuePairTy;
908 typedef DenseMap<ValuePairTy, bool> CachedResultsTy;
909 CachedResultsTy CachedResults;
911 bool relatedCheck(const Value *A, const Value *B);
912 bool relatedSelect(const SelectInst *A, const Value *B);
913 bool relatedPHI(const PHINode *A, const Value *B);
916 void operator=(const ProvenanceAnalysis &);
917 ProvenanceAnalysis(const ProvenanceAnalysis &);
920 ProvenanceAnalysis() {}
922 void setAA(AliasAnalysis *aa) { AA = aa; }
924 AliasAnalysis *getAA() const { return AA; }
926 bool related(const Value *A, const Value *B);
929 CachedResults.clear();
934 bool ProvenanceAnalysis::relatedSelect(const SelectInst *A, const Value *B) {
935 // If the values are Selects with the same condition, we can do a more precise
936 // check: just check for relations between the values on corresponding arms.
937 if (const SelectInst *SB = dyn_cast<SelectInst>(B))
938 if (A->getCondition() == SB->getCondition()) {
939 if (related(A->getTrueValue(), SB->getTrueValue()))
941 if (related(A->getFalseValue(), SB->getFalseValue()))
946 // Check both arms of the Select node individually.
947 if (related(A->getTrueValue(), B))
949 if (related(A->getFalseValue(), B))
952 // The arms both checked out.
956 bool ProvenanceAnalysis::relatedPHI(const PHINode *A, const Value *B) {
957 // If the values are PHIs in the same block, we can do a more precise as well
958 // as efficient check: just check for relations between the values on
959 // corresponding edges.
960 if (const PHINode *PNB = dyn_cast<PHINode>(B))
961 if (PNB->getParent() == A->getParent()) {
962 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i)
963 if (related(A->getIncomingValue(i),
964 PNB->getIncomingValueForBlock(A->getIncomingBlock(i))))
969 // Check each unique source of the PHI node against B.
970 SmallPtrSet<const Value *, 4> UniqueSrc;
971 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i) {
972 const Value *PV1 = A->getIncomingValue(i);
973 if (UniqueSrc.insert(PV1) && related(PV1, B))
977 // All of the arms checked out.
981 /// isStoredObjCPointer - Test if the value of P, or any value covered by its
982 /// provenance, is ever stored within the function (not counting callees).
983 static bool isStoredObjCPointer(const Value *P) {
984 SmallPtrSet<const Value *, 8> Visited;
985 SmallVector<const Value *, 8> Worklist;
986 Worklist.push_back(P);
989 P = Worklist.pop_back_val();
990 for (Value::const_use_iterator UI = P->use_begin(), UE = P->use_end();
992 const User *Ur = *UI;
993 if (isa<StoreInst>(Ur)) {
994 if (UI.getOperandNo() == 0)
995 // The pointer is stored.
997 // The pointed is stored through.
1000 if (isa<CallInst>(Ur))
1001 // The pointer is passed as an argument, ignore this.
1003 if (isa<PtrToIntInst>(P))
1004 // Assume the worst.
1006 if (Visited.insert(Ur))
1007 Worklist.push_back(Ur);
1009 } while (!Worklist.empty());
1011 // Everything checked out.
1015 bool ProvenanceAnalysis::relatedCheck(const Value *A, const Value *B) {
1016 // Skip past provenance pass-throughs.
1017 A = GetUnderlyingObjCPtr(A);
1018 B = GetUnderlyingObjCPtr(B);
1024 // Ask regular AliasAnalysis, for a first approximation.
1025 switch (AA->alias(A, B)) {
1026 case AliasAnalysis::NoAlias:
1028 case AliasAnalysis::MustAlias:
1029 case AliasAnalysis::PartialAlias:
1031 case AliasAnalysis::MayAlias:
1035 bool AIsIdentified = IsObjCIdentifiedObject(A);
1036 bool BIsIdentified = IsObjCIdentifiedObject(B);
1038 // An ObjC-Identified object can't alias a load if it is never locally stored.
1039 if (AIsIdentified) {
1040 if (BIsIdentified) {
1041 // If both pointers have provenance, they can be directly compared.
1045 if (isa<LoadInst>(B))
1046 return isStoredObjCPointer(A);
1049 if (BIsIdentified && isa<LoadInst>(A))
1050 return isStoredObjCPointer(B);
1053 // Special handling for PHI and Select.
1054 if (const PHINode *PN = dyn_cast<PHINode>(A))
1055 return relatedPHI(PN, B);
1056 if (const PHINode *PN = dyn_cast<PHINode>(B))
1057 return relatedPHI(PN, A);
1058 if (const SelectInst *S = dyn_cast<SelectInst>(A))
1059 return relatedSelect(S, B);
1060 if (const SelectInst *S = dyn_cast<SelectInst>(B))
1061 return relatedSelect(S, A);
1067 bool ProvenanceAnalysis::related(const Value *A, const Value *B) {
1068 // Begin by inserting a conservative value into the map. If the insertion
1069 // fails, we have the answer already. If it succeeds, leave it there until we
1070 // compute the real answer to guard against recursive queries.
1071 if (A > B) std::swap(A, B);
1072 std::pair<CachedResultsTy::iterator, bool> Pair =
1073 CachedResults.insert(std::make_pair(ValuePairTy(A, B), true));
1075 return Pair.first->second;
1077 bool Result = relatedCheck(A, B);
1078 CachedResults[ValuePairTy(A, B)] = Result;
1083 // Sequence - A sequence of states that a pointer may go through in which an
1084 // objc_retain and objc_release are actually needed.
1087 S_Retain, ///< objc_retain(x)
1088 S_CanRelease, ///< foo(x) -- x could possibly see a ref count decrement
1089 S_Use, ///< any use of x
1090 S_Stop, ///< like S_Release, but code motion is stopped
1091 S_Release, ///< objc_release(x)
1092 S_MovableRelease ///< objc_release(x), !clang.imprecise_release
1096 static Sequence MergeSeqs(Sequence A, Sequence B, bool TopDown) {
1100 if (A == S_None || B == S_None)
1103 if (A > B) std::swap(A, B);
1105 // Choose the side which is further along in the sequence.
1106 if ((A == S_Retain || A == S_CanRelease) &&
1107 (B == S_CanRelease || B == S_Use))
1110 // Choose the side which is further along in the sequence.
1111 if ((A == S_Use || A == S_CanRelease) &&
1112 (B == S_Use || B == S_Release || B == S_Stop || B == S_MovableRelease))
1114 // If both sides are releases, choose the more conservative one.
1115 if (A == S_Stop && (B == S_Release || B == S_MovableRelease))
1117 if (A == S_Release && B == S_MovableRelease)
1125 /// RRInfo - Unidirectional information about either a
1126 /// retain-decrement-use-release sequence or release-use-decrement-retain
1127 /// reverese sequence.
1129 /// KnownSafe - After an objc_retain, the reference count of the referenced
1130 /// object is known to be positive. Similarly, before an objc_release, the
1131 /// reference count of the referenced object is known to be positive. If
1132 /// there are retain-release pairs in code regions where the retain count
1133 /// is known to be positive, they can be eliminated, regardless of any side
1134 /// effects between them.
1136 /// Also, a retain+release pair nested within another retain+release
1137 /// pair all on the known same pointer value can be eliminated, regardless
1138 /// of any intervening side effects.
1140 /// KnownSafe is true when either of these conditions is satisfied.
1143 /// IsRetainBlock - True if the Calls are objc_retainBlock calls (as
1144 /// opposed to objc_retain calls).
1147 /// IsTailCallRelease - True of the objc_release calls are all marked
1148 /// with the "tail" keyword.
1149 bool IsTailCallRelease;
1151 /// ReleaseMetadata - If the Calls are objc_release calls and they all have
1152 /// a clang.imprecise_release tag, this is the metadata tag.
1153 MDNode *ReleaseMetadata;
1155 /// Calls - For a top-down sequence, the set of objc_retains or
1156 /// objc_retainBlocks. For bottom-up, the set of objc_releases.
1157 SmallPtrSet<Instruction *, 2> Calls;
1159 /// ReverseInsertPts - The set of optimal insert positions for
1160 /// moving calls in the opposite sequence.
1161 SmallPtrSet<Instruction *, 2> ReverseInsertPts;
1164 KnownSafe(false), IsRetainBlock(false), IsTailCallRelease(false),
1165 ReleaseMetadata(0) {}
1171 void RRInfo::clear() {
1173 IsRetainBlock = false;
1174 IsTailCallRelease = false;
1175 ReleaseMetadata = 0;
1177 ReverseInsertPts.clear();
1181 /// PtrState - This class summarizes several per-pointer runtime properties
1182 /// which are propogated through the flow graph.
1184 /// RefCount - The known minimum number of reference count increments.
1187 /// NestCount - The known minimum level of retain+release nesting.
1190 /// Seq - The current position in the sequence.
1194 /// RRI - Unidirectional information about the current sequence.
1195 /// TODO: Encapsulate this better.
1198 PtrState() : RefCount(0), NestCount(0), Seq(S_None) {}
1200 void SetAtLeastOneRefCount() {
1201 if (RefCount == 0) RefCount = 1;
1204 void IncrementRefCount() {
1205 if (RefCount != UINT_MAX) ++RefCount;
1208 void DecrementRefCount() {
1209 if (RefCount != 0) --RefCount;
1212 bool IsKnownIncremented() const {
1213 return RefCount > 0;
1216 void IncrementNestCount() {
1217 if (NestCount != UINT_MAX) ++NestCount;
1220 void DecrementNestCount() {
1221 if (NestCount != 0) --NestCount;
1224 bool IsKnownNested() const {
1225 return NestCount > 0;
1228 void SetSeq(Sequence NewSeq) {
1232 void SetSeqToRelease(MDNode *M) {
1233 if (Seq == S_None || Seq == S_Use) {
1234 Seq = M ? S_MovableRelease : S_Release;
1235 RRI.ReleaseMetadata = M;
1236 } else if (Seq != S_MovableRelease || RRI.ReleaseMetadata != M) {
1238 RRI.ReleaseMetadata = 0;
1242 Sequence GetSeq() const {
1246 void ClearSequenceProgress() {
1251 void Merge(const PtrState &Other, bool TopDown);
1256 PtrState::Merge(const PtrState &Other, bool TopDown) {
1257 Seq = MergeSeqs(Seq, Other.Seq, TopDown);
1258 RefCount = std::min(RefCount, Other.RefCount);
1259 NestCount = std::min(NestCount, Other.NestCount);
1261 // We can't merge a plain objc_retain with an objc_retainBlock.
1262 if (RRI.IsRetainBlock != Other.RRI.IsRetainBlock)
1265 if (Seq == S_None) {
1268 // Conservatively merge the ReleaseMetadata information.
1269 if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
1270 RRI.ReleaseMetadata = 0;
1272 RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe;
1273 RRI.IsTailCallRelease = RRI.IsTailCallRelease && Other.RRI.IsTailCallRelease;
1274 RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
1275 RRI.ReverseInsertPts.insert(Other.RRI.ReverseInsertPts.begin(),
1276 Other.RRI.ReverseInsertPts.end());
1281 /// BBState - Per-BasicBlock state.
1283 /// TopDownPathCount - The number of unique control paths from the entry
1284 /// which can reach this block.
1285 unsigned TopDownPathCount;
1287 /// BottomUpPathCount - The number of unique control paths to exits
1288 /// from this block.
1289 unsigned BottomUpPathCount;
1291 /// MapTy - A type for PerPtrTopDown and PerPtrBottomUp.
1292 typedef MapVector<const Value *, PtrState> MapTy;
1294 /// PerPtrTopDown - The top-down traversal uses this to record information
1295 /// known about a pointer at the bottom of each block.
1296 MapTy PerPtrTopDown;
1298 /// PerPtrBottomUp - The bottom-up traversal uses this to record information
1299 /// known about a pointer at the top of each block.
1300 MapTy PerPtrBottomUp;
1303 BBState() : TopDownPathCount(0), BottomUpPathCount(0) {}
1305 typedef MapTy::iterator ptr_iterator;
1306 typedef MapTy::const_iterator ptr_const_iterator;
1308 ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); }
1309 ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); }
1310 ptr_const_iterator top_down_ptr_begin() const {
1311 return PerPtrTopDown.begin();
1313 ptr_const_iterator top_down_ptr_end() const {
1314 return PerPtrTopDown.end();
1317 ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); }
1318 ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); }
1319 ptr_const_iterator bottom_up_ptr_begin() const {
1320 return PerPtrBottomUp.begin();
1322 ptr_const_iterator bottom_up_ptr_end() const {
1323 return PerPtrBottomUp.end();
1326 /// SetAsEntry - Mark this block as being an entry block, which has one
1327 /// path from the entry by definition.
1328 void SetAsEntry() { TopDownPathCount = 1; }
1330 /// SetAsExit - Mark this block as being an exit block, which has one
1331 /// path to an exit by definition.
1332 void SetAsExit() { BottomUpPathCount = 1; }
1334 PtrState &getPtrTopDownState(const Value *Arg) {
1335 return PerPtrTopDown[Arg];
1338 PtrState &getPtrBottomUpState(const Value *Arg) {
1339 return PerPtrBottomUp[Arg];
1342 void clearBottomUpPointers() {
1343 PerPtrBottomUp.clear();
1346 void clearTopDownPointers() {
1347 PerPtrTopDown.clear();
1350 void InitFromPred(const BBState &Other);
1351 void InitFromSucc(const BBState &Other);
1352 void MergePred(const BBState &Other);
1353 void MergeSucc(const BBState &Other);
1355 /// GetAllPathCount - Return the number of possible unique paths from an
1356 /// entry to an exit which pass through this block. This is only valid
1357 /// after both the top-down and bottom-up traversals are complete.
1358 unsigned GetAllPathCount() const {
1359 return TopDownPathCount * BottomUpPathCount;
1362 /// IsVisitedTopDown - Test whether the block for this BBState has been
1363 /// visited by the top-down portion of the algorithm.
1364 bool isVisitedTopDown() const {
1365 return TopDownPathCount != 0;
1370 void BBState::InitFromPred(const BBState &Other) {
1371 PerPtrTopDown = Other.PerPtrTopDown;
1372 TopDownPathCount = Other.TopDownPathCount;
1375 void BBState::InitFromSucc(const BBState &Other) {
1376 PerPtrBottomUp = Other.PerPtrBottomUp;
1377 BottomUpPathCount = Other.BottomUpPathCount;
1380 /// MergePred - The top-down traversal uses this to merge information about
1381 /// predecessors to form the initial state for a new block.
1382 void BBState::MergePred(const BBState &Other) {
1383 // Other.TopDownPathCount can be 0, in which case it is either dead or a
1384 // loop backedge. Loop backedges are special.
1385 TopDownPathCount += Other.TopDownPathCount;
1387 // For each entry in the other set, if our set has an entry with the same key,
1388 // merge the entries. Otherwise, copy the entry and merge it with an empty
1390 for (ptr_const_iterator MI = Other.top_down_ptr_begin(),
1391 ME = Other.top_down_ptr_end(); MI != ME; ++MI) {
1392 std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI);
1393 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1397 // For each entry in our set, if the other set doesn't have an entry with the
1398 // same key, force it to merge with an empty entry.
1399 for (ptr_iterator MI = top_down_ptr_begin(),
1400 ME = top_down_ptr_end(); MI != ME; ++MI)
1401 if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end())
1402 MI->second.Merge(PtrState(), /*TopDown=*/true);
1405 /// MergeSucc - The bottom-up traversal uses this to merge information about
1406 /// successors to form the initial state for a new block.
1407 void BBState::MergeSucc(const BBState &Other) {
1408 // Other.BottomUpPathCount can be 0, in which case it is either dead or a
1409 // loop backedge. Loop backedges are special.
1410 BottomUpPathCount += Other.BottomUpPathCount;
1412 // For each entry in the other set, if our set has an entry with the
1413 // same key, merge the entries. Otherwise, copy the entry and merge
1414 // it with an empty entry.
1415 for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(),
1416 ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) {
1417 std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI);
1418 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1422 // For each entry in our set, if the other set doesn't have an entry
1423 // with the same key, force it to merge with an empty entry.
1424 for (ptr_iterator MI = bottom_up_ptr_begin(),
1425 ME = bottom_up_ptr_end(); MI != ME; ++MI)
1426 if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end())
1427 MI->second.Merge(PtrState(), /*TopDown=*/false);
1431 /// ObjCARCOpt - The main ARC optimization pass.
1432 class ObjCARCOpt : public FunctionPass {
1434 ProvenanceAnalysis PA;
1436 /// Run - A flag indicating whether this optimization pass should run.
1439 /// RetainRVCallee, etc. - Declarations for ObjC runtime
1440 /// functions, for use in creating calls to them. These are initialized
1441 /// lazily to avoid cluttering up the Module with unused declarations.
1442 Constant *RetainRVCallee, *AutoreleaseRVCallee, *ReleaseCallee,
1443 *RetainCallee, *RetainBlockCallee, *AutoreleaseCallee;
1445 /// UsedInThisFunciton - Flags which determine whether each of the
1446 /// interesting runtine functions is in fact used in the current function.
1447 unsigned UsedInThisFunction;
1449 /// ImpreciseReleaseMDKind - The Metadata Kind for clang.imprecise_release
1451 unsigned ImpreciseReleaseMDKind;
1453 Constant *getRetainRVCallee(Module *M);
1454 Constant *getAutoreleaseRVCallee(Module *M);
1455 Constant *getReleaseCallee(Module *M);
1456 Constant *getRetainCallee(Module *M);
1457 Constant *getRetainBlockCallee(Module *M);
1458 Constant *getAutoreleaseCallee(Module *M);
1460 void OptimizeRetainCall(Function &F, Instruction *Retain);
1461 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
1462 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV);
1463 void OptimizeIndividualCalls(Function &F);
1465 void CheckForCFGHazards(const BasicBlock *BB,
1466 DenseMap<const BasicBlock *, BBState> &BBStates,
1467 BBState &MyStates) const;
1468 bool VisitBottomUp(BasicBlock *BB,
1469 DenseMap<const BasicBlock *, BBState> &BBStates,
1470 MapVector<Value *, RRInfo> &Retains);
1471 bool VisitTopDown(BasicBlock *BB,
1472 DenseMap<const BasicBlock *, BBState> &BBStates,
1473 DenseMap<Value *, RRInfo> &Releases);
1474 bool Visit(Function &F,
1475 DenseMap<const BasicBlock *, BBState> &BBStates,
1476 MapVector<Value *, RRInfo> &Retains,
1477 DenseMap<Value *, RRInfo> &Releases);
1479 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1480 MapVector<Value *, RRInfo> &Retains,
1481 DenseMap<Value *, RRInfo> &Releases,
1482 SmallVectorImpl<Instruction *> &DeadInsts,
1485 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1486 MapVector<Value *, RRInfo> &Retains,
1487 DenseMap<Value *, RRInfo> &Releases,
1490 void OptimizeWeakCalls(Function &F);
1492 bool OptimizeSequences(Function &F);
1494 void OptimizeReturns(Function &F);
1496 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1497 virtual bool doInitialization(Module &M);
1498 virtual bool runOnFunction(Function &F);
1499 virtual void releaseMemory();
1503 ObjCARCOpt() : FunctionPass(ID) {
1504 initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1509 char ObjCARCOpt::ID = 0;
1510 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1511 "objc-arc", "ObjC ARC optimization", false, false)
1512 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1513 INITIALIZE_PASS_END(ObjCARCOpt,
1514 "objc-arc", "ObjC ARC optimization", false, false)
1516 Pass *llvm::createObjCARCOptPass() {
1517 return new ObjCARCOpt();
1520 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1521 AU.addRequired<ObjCARCAliasAnalysis>();
1522 AU.addRequired<AliasAnalysis>();
1523 // ARC optimization doesn't currently split critical edges.
1524 AU.setPreservesCFG();
1527 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1528 if (!RetainRVCallee) {
1529 LLVMContext &C = M->getContext();
1530 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1531 std::vector<Type *> Params;
1532 Params.push_back(I8X);
1534 FunctionType::get(I8X, Params, /*isVarArg=*/false);
1535 AttrListPtr Attributes;
1536 Attributes.addAttr(~0u, Attribute::NoUnwind);
1538 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1541 return RetainRVCallee;
1544 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1545 if (!AutoreleaseRVCallee) {
1546 LLVMContext &C = M->getContext();
1547 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1548 std::vector<Type *> Params;
1549 Params.push_back(I8X);
1551 FunctionType::get(I8X, Params, /*isVarArg=*/false);
1552 AttrListPtr Attributes;
1553 Attributes.addAttr(~0u, Attribute::NoUnwind);
1554 AutoreleaseRVCallee =
1555 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1558 return AutoreleaseRVCallee;
1561 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1562 if (!ReleaseCallee) {
1563 LLVMContext &C = M->getContext();
1564 std::vector<Type *> Params;
1565 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1566 AttrListPtr Attributes;
1567 Attributes.addAttr(~0u, Attribute::NoUnwind);
1569 M->getOrInsertFunction(
1571 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1574 return ReleaseCallee;
1577 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1578 if (!RetainCallee) {
1579 LLVMContext &C = M->getContext();
1580 std::vector<Type *> Params;
1581 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1582 AttrListPtr Attributes;
1583 Attributes.addAttr(~0u, Attribute::NoUnwind);
1585 M->getOrInsertFunction(
1587 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1590 return RetainCallee;
1593 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1594 if (!RetainBlockCallee) {
1595 LLVMContext &C = M->getContext();
1596 std::vector<Type *> Params;
1597 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1598 AttrListPtr Attributes;
1599 Attributes.addAttr(~0u, Attribute::NoUnwind);
1601 M->getOrInsertFunction(
1603 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1606 return RetainBlockCallee;
1609 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1610 if (!AutoreleaseCallee) {
1611 LLVMContext &C = M->getContext();
1612 std::vector<Type *> Params;
1613 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1614 AttrListPtr Attributes;
1615 Attributes.addAttr(~0u, Attribute::NoUnwind);
1617 M->getOrInsertFunction(
1619 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1622 return AutoreleaseCallee;
1625 /// CanAlterRefCount - Test whether the given instruction can result in a
1626 /// reference count modification (positive or negative) for the pointer's
1629 CanAlterRefCount(const Instruction *Inst, const Value *Ptr,
1630 ProvenanceAnalysis &PA, InstructionClass Class) {
1632 case IC_Autorelease:
1633 case IC_AutoreleaseRV:
1635 // These operations never directly modify a reference count.
1640 ImmutableCallSite CS = static_cast<const Value *>(Inst);
1641 assert(CS && "Only calls can alter reference counts!");
1643 // See if AliasAnalysis can help us with the call.
1644 AliasAnalysis::ModRefBehavior MRB = PA.getAA()->getModRefBehavior(CS);
1645 if (AliasAnalysis::onlyReadsMemory(MRB))
1647 if (AliasAnalysis::onlyAccessesArgPointees(MRB)) {
1648 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1650 const Value *Op = *I;
1651 if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1657 // Assume the worst.
1661 /// CanUse - Test whether the given instruction can "use" the given pointer's
1662 /// object in a way that requires the reference count to be positive.
1664 CanUse(const Instruction *Inst, const Value *Ptr, ProvenanceAnalysis &PA,
1665 InstructionClass Class) {
1666 // IC_Call operations (as opposed to IC_CallOrUser) never "use" objc pointers.
1667 if (Class == IC_Call)
1670 // Consider various instructions which may have pointer arguments which are
1672 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(Inst)) {
1673 // Comparing a pointer with null, or any other constant, isn't really a use,
1674 // because we don't care what the pointer points to, or about the values
1675 // of any other dynamic reference-counted pointers.
1676 if (!IsPotentialUse(ICI->getOperand(1)))
1678 } else if (ImmutableCallSite CS = static_cast<const Value *>(Inst)) {
1679 // For calls, just check the arguments (and not the callee operand).
1680 for (ImmutableCallSite::arg_iterator OI = CS.arg_begin(),
1681 OE = CS.arg_end(); OI != OE; ++OI) {
1682 const Value *Op = *OI;
1683 if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1687 } else if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1688 // Special-case stores, because we don't care about the stored value, just
1689 // the store address.
1690 const Value *Op = GetUnderlyingObjCPtr(SI->getPointerOperand());
1691 // If we can't tell what the underlying object was, assume there is a
1693 return IsPotentialUse(Op) && PA.related(Op, Ptr);
1696 // Check each operand for a match.
1697 for (User::const_op_iterator OI = Inst->op_begin(), OE = Inst->op_end();
1699 const Value *Op = *OI;
1700 if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1706 /// CanInterruptRV - Test whether the given instruction can autorelease
1707 /// any pointer or cause an autoreleasepool pop.
1709 CanInterruptRV(InstructionClass Class) {
1711 case IC_AutoreleasepoolPop:
1714 case IC_Autorelease:
1715 case IC_AutoreleaseRV:
1716 case IC_FusedRetainAutorelease:
1717 case IC_FusedRetainAutoreleaseRV:
1725 /// DependenceKind - There are several kinds of dependence-like concepts in
1727 enum DependenceKind {
1728 NeedsPositiveRetainCount,
1729 CanChangeRetainCount,
1730 RetainAutoreleaseDep, ///< Blocks objc_retainAutorelease.
1731 RetainAutoreleaseRVDep, ///< Blocks objc_retainAutoreleaseReturnValue.
1732 RetainRVDep ///< Blocks objc_retainAutoreleasedReturnValue.
1736 /// Depends - Test if there can be dependencies on Inst through Arg. This
1737 /// function only tests dependencies relevant for removing pairs of calls.
1739 Depends(DependenceKind Flavor, Instruction *Inst, const Value *Arg,
1740 ProvenanceAnalysis &PA) {
1741 // If we've reached the definition of Arg, stop.
1746 case NeedsPositiveRetainCount: {
1747 InstructionClass Class = GetInstructionClass(Inst);
1749 case IC_AutoreleasepoolPop:
1750 case IC_AutoreleasepoolPush:
1754 return CanUse(Inst, Arg, PA, Class);
1758 case CanChangeRetainCount: {
1759 InstructionClass Class = GetInstructionClass(Inst);
1761 case IC_AutoreleasepoolPop:
1762 // Conservatively assume this can decrement any count.
1764 case IC_AutoreleasepoolPush:
1768 return CanAlterRefCount(Inst, Arg, PA, Class);
1772 case RetainAutoreleaseDep:
1773 switch (GetBasicInstructionClass(Inst)) {
1774 case IC_AutoreleasepoolPop:
1775 // Don't merge an objc_autorelease with an objc_retain inside a different
1776 // autoreleasepool scope.
1780 // Check for a retain of the same pointer for merging.
1781 return GetObjCArg(Inst) == Arg;
1783 // Nothing else matters for objc_retainAutorelease formation.
1788 case RetainAutoreleaseRVDep: {
1789 InstructionClass Class = GetBasicInstructionClass(Inst);
1793 // Check for a retain of the same pointer for merging.
1794 return GetObjCArg(Inst) == Arg;
1796 // Anything that can autorelease interrupts
1797 // retainAutoreleaseReturnValue formation.
1798 return CanInterruptRV(Class);
1804 return CanInterruptRV(GetBasicInstructionClass(Inst));
1807 llvm_unreachable("Invalid dependence flavor");
1811 /// FindDependencies - Walk up the CFG from StartPos (which is in StartBB) and
1812 /// find local and non-local dependencies on Arg.
1813 /// TODO: Cache results?
1815 FindDependencies(DependenceKind Flavor,
1817 BasicBlock *StartBB, Instruction *StartInst,
1818 SmallPtrSet<Instruction *, 4> &DependingInstructions,
1819 SmallPtrSet<const BasicBlock *, 4> &Visited,
1820 ProvenanceAnalysis &PA) {
1821 BasicBlock::iterator StartPos = StartInst;
1823 SmallVector<std::pair<BasicBlock *, BasicBlock::iterator>, 4> Worklist;
1824 Worklist.push_back(std::make_pair(StartBB, StartPos));
1826 std::pair<BasicBlock *, BasicBlock::iterator> Pair =
1827 Worklist.pop_back_val();
1828 BasicBlock *LocalStartBB = Pair.first;
1829 BasicBlock::iterator LocalStartPos = Pair.second;
1830 BasicBlock::iterator StartBBBegin = LocalStartBB->begin();
1832 if (LocalStartPos == StartBBBegin) {
1833 pred_iterator PI(LocalStartBB), PE(LocalStartBB, false);
1835 // If we've reached the function entry, produce a null dependence.
1836 DependingInstructions.insert(0);
1838 // Add the predecessors to the worklist.
1840 BasicBlock *PredBB = *PI;
1841 if (Visited.insert(PredBB))
1842 Worklist.push_back(std::make_pair(PredBB, PredBB->end()));
1843 } while (++PI != PE);
1847 Instruction *Inst = --LocalStartPos;
1848 if (Depends(Flavor, Inst, Arg, PA)) {
1849 DependingInstructions.insert(Inst);
1853 } while (!Worklist.empty());
1855 // Determine whether the original StartBB post-dominates all of the blocks we
1856 // visited. If not, insert a sentinal indicating that most optimizations are
1858 for (SmallPtrSet<const BasicBlock *, 4>::const_iterator I = Visited.begin(),
1859 E = Visited.end(); I != E; ++I) {
1860 const BasicBlock *BB = *I;
1863 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1864 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
1865 const BasicBlock *Succ = *SI;
1866 if (Succ != StartBB && !Visited.count(Succ)) {
1867 DependingInstructions.insert(reinterpret_cast<Instruction *>(-1));
1874 static bool isNullOrUndef(const Value *V) {
1875 return isa<ConstantPointerNull>(V) || isa<UndefValue>(V);
1878 static bool isNoopInstruction(const Instruction *I) {
1879 return isa<BitCastInst>(I) ||
1880 (isa<GetElementPtrInst>(I) &&
1881 cast<GetElementPtrInst>(I)->hasAllZeroIndices());
1884 /// OptimizeRetainCall - Turn objc_retain into
1885 /// objc_retainAutoreleasedReturnValue if the operand is a return value.
1887 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
1888 CallSite CS(GetObjCArg(Retain));
1889 Instruction *Call = CS.getInstruction();
1891 if (Call->getParent() != Retain->getParent()) return;
1893 // Check that the call is next to the retain.
1894 BasicBlock::iterator I = Call;
1896 while (isNoopInstruction(I)) ++I;
1900 // Turn it to an objc_retainAutoreleasedReturnValue..
1903 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
1906 /// OptimizeRetainRVCall - Turn objc_retainAutoreleasedReturnValue into
1907 /// objc_retain if the operand is not a return value. Or, if it can be
1908 /// paired with an objc_autoreleaseReturnValue, delete the pair and
1911 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
1912 // Check for the argument being from an immediately preceding call.
1913 Value *Arg = GetObjCArg(RetainRV);
1915 if (Instruction *Call = CS.getInstruction())
1916 if (Call->getParent() == RetainRV->getParent()) {
1917 BasicBlock::iterator I = Call;
1919 while (isNoopInstruction(I)) ++I;
1920 if (&*I == RetainRV)
1924 // Check for being preceded by an objc_autoreleaseReturnValue on the same
1925 // pointer. In this case, we can delete the pair.
1926 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
1928 do --I; while (I != Begin && isNoopInstruction(I));
1929 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
1930 GetObjCArg(I) == Arg) {
1933 EraseInstruction(I);
1934 EraseInstruction(RetainRV);
1939 // Turn it to a plain objc_retain.
1942 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
1946 /// OptimizeAutoreleaseRVCall - Turn objc_autoreleaseReturnValue into
1947 /// objc_autorelease if the result is not used as a return value.
1949 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV) {
1950 // Check for a return of the pointer value.
1951 const Value *Ptr = GetObjCArg(AutoreleaseRV);
1952 SmallVector<const Value *, 2> Users;
1953 Users.push_back(Ptr);
1955 Ptr = Users.pop_back_val();
1956 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
1958 const User *I = *UI;
1959 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
1961 if (isa<BitCastInst>(I))
1964 } while (!Users.empty());
1968 cast<CallInst>(AutoreleaseRV)->
1969 setCalledFunction(getAutoreleaseCallee(F.getParent()));
1972 /// OptimizeIndividualCalls - Visit each call, one at a time, and make
1973 /// simplifications without doing any additional analysis.
1974 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
1975 // Reset all the flags in preparation for recomputing them.
1976 UsedInThisFunction = 0;
1978 // Visit all objc_* calls in F.
1979 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
1980 Instruction *Inst = &*I++;
1981 InstructionClass Class = GetBasicInstructionClass(Inst);
1986 // Delete no-op casts. These function calls have special semantics, but
1987 // the semantics are entirely implemented via lowering in the front-end,
1988 // so by the time they reach the optimizer, they are just no-op calls
1989 // which return their argument.
1991 // There are gray areas here, as the ability to cast reference-counted
1992 // pointers to raw void* and back allows code to break ARC assumptions,
1993 // however these are currently considered to be unimportant.
1997 EraseInstruction(Inst);
2000 // If the pointer-to-weak-pointer is null, it's undefined behavior.
2003 case IC_LoadWeakRetained:
2005 case IC_DestroyWeak: {
2006 CallInst *CI = cast<CallInst>(Inst);
2007 if (isNullOrUndef(CI->getArgOperand(0))) {
2008 Type *Ty = CI->getArgOperand(0)->getType();
2009 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2010 Constant::getNullValue(Ty),
2012 CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
2013 CI->eraseFromParent();
2020 CallInst *CI = cast<CallInst>(Inst);
2021 if (isNullOrUndef(CI->getArgOperand(0)) ||
2022 isNullOrUndef(CI->getArgOperand(1))) {
2023 Type *Ty = CI->getArgOperand(0)->getType();
2024 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2025 Constant::getNullValue(Ty),
2027 CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
2028 CI->eraseFromParent();
2034 OptimizeRetainCall(F, Inst);
2037 if (OptimizeRetainRVCall(F, Inst))
2040 case IC_AutoreleaseRV:
2041 OptimizeAutoreleaseRVCall(F, Inst);
2045 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
2046 if (IsAutorelease(Class) && Inst->use_empty()) {
2047 CallInst *Call = cast<CallInst>(Inst);
2048 const Value *Arg = Call->getArgOperand(0);
2049 Arg = FindSingleUseIdentifiedObject(Arg);
2054 // Create the declaration lazily.
2055 LLVMContext &C = Inst->getContext();
2057 CallInst::Create(getReleaseCallee(F.getParent()),
2058 Call->getArgOperand(0), "", Call);
2059 NewCall->setMetadata(ImpreciseReleaseMDKind,
2060 MDNode::get(C, ArrayRef<Value *>()));
2061 EraseInstruction(Call);
2067 // For functions which can never be passed stack arguments, add
2069 if (IsAlwaysTail(Class)) {
2071 cast<CallInst>(Inst)->setTailCall();
2074 // Set nounwind as needed.
2075 if (IsNoThrow(Class)) {
2077 cast<CallInst>(Inst)->setDoesNotThrow();
2080 if (!IsNoopOnNull(Class)) {
2081 UsedInThisFunction |= 1 << Class;
2085 const Value *Arg = GetObjCArg(Inst);
2087 // ARC calls with null are no-ops. Delete them.
2088 if (isNullOrUndef(Arg)) {
2091 EraseInstruction(Inst);
2095 // Keep track of which of retain, release, autorelease, and retain_block
2096 // are actually present in this function.
2097 UsedInThisFunction |= 1 << Class;
2099 // If Arg is a PHI, and one or more incoming values to the
2100 // PHI are null, and the call is control-equivalent to the PHI, and there
2101 // are no relevant side effects between the PHI and the call, the call
2102 // could be pushed up to just those paths with non-null incoming values.
2103 // For now, don't bother splitting critical edges for this.
2104 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
2105 Worklist.push_back(std::make_pair(Inst, Arg));
2107 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
2111 const PHINode *PN = dyn_cast<PHINode>(Arg);
2114 // Determine if the PHI has any null operands, or any incoming
2116 bool HasNull = false;
2117 bool HasCriticalEdges = false;
2118 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2120 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2121 if (isNullOrUndef(Incoming))
2123 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
2124 .getNumSuccessors() != 1) {
2125 HasCriticalEdges = true;
2129 // If we have null operands and no critical edges, optimize.
2130 if (!HasCriticalEdges && HasNull) {
2131 SmallPtrSet<Instruction *, 4> DependingInstructions;
2132 SmallPtrSet<const BasicBlock *, 4> Visited;
2134 // Check that there is nothing that cares about the reference
2135 // count between the call and the phi.
2136 FindDependencies(NeedsPositiveRetainCount, Arg,
2137 Inst->getParent(), Inst,
2138 DependingInstructions, Visited, PA);
2139 if (DependingInstructions.size() == 1 &&
2140 *DependingInstructions.begin() == PN) {
2143 // Clone the call into each predecessor that has a non-null value.
2144 CallInst *CInst = cast<CallInst>(Inst);
2145 Type *ParamTy = CInst->getArgOperand(0)->getType();
2146 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2148 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2149 if (!isNullOrUndef(Incoming)) {
2150 CallInst *Clone = cast<CallInst>(CInst->clone());
2151 Value *Op = PN->getIncomingValue(i);
2152 Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
2153 if (Op->getType() != ParamTy)
2154 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
2155 Clone->setArgOperand(0, Op);
2156 Clone->insertBefore(InsertPos);
2157 Worklist.push_back(std::make_pair(Clone, Incoming));
2160 // Erase the original call.
2161 EraseInstruction(CInst);
2165 } while (!Worklist.empty());
2169 /// CheckForCFGHazards - Check for critical edges, loop boundaries, irreducible
2170 /// control flow, or other CFG structures where moving code across the edge
2171 /// would result in it being executed more.
2173 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
2174 DenseMap<const BasicBlock *, BBState> &BBStates,
2175 BBState &MyStates) const {
2176 // If any top-down local-use or possible-dec has a succ which is earlier in
2177 // the sequence, forget it.
2178 for (BBState::ptr_const_iterator I = MyStates.top_down_ptr_begin(),
2179 E = MyStates.top_down_ptr_end(); I != E; ++I)
2180 switch (I->second.GetSeq()) {
2183 const Value *Arg = I->first;
2184 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2185 bool SomeSuccHasSame = false;
2186 bool AllSuccsHaveSame = true;
2187 PtrState &S = MyStates.getPtrTopDownState(Arg);
2188 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
2189 PtrState &SuccS = BBStates[*SI].getPtrBottomUpState(Arg);
2190 switch (SuccS.GetSeq()) {
2192 case S_CanRelease: {
2193 if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2194 S.ClearSequenceProgress();
2198 SomeSuccHasSame = true;
2202 case S_MovableRelease:
2203 if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2204 AllSuccsHaveSame = false;
2207 llvm_unreachable("bottom-up pointer in retain state!");
2210 // If the state at the other end of any of the successor edges
2211 // matches the current state, require all edges to match. This
2212 // guards against loops in the middle of a sequence.
2213 if (SomeSuccHasSame && !AllSuccsHaveSame)
2214 S.ClearSequenceProgress();
2216 case S_CanRelease: {
2217 const Value *Arg = I->first;
2218 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2219 bool SomeSuccHasSame = false;
2220 bool AllSuccsHaveSame = true;
2221 PtrState &S = MyStates.getPtrTopDownState(Arg);
2222 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
2223 PtrState &SuccS = BBStates[*SI].getPtrBottomUpState(Arg);
2224 switch (SuccS.GetSeq()) {
2226 if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2227 S.ClearSequenceProgress();
2231 SomeSuccHasSame = true;
2235 case S_MovableRelease:
2237 if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2238 AllSuccsHaveSame = false;
2241 llvm_unreachable("bottom-up pointer in retain state!");
2244 // If the state at the other end of any of the successor edges
2245 // matches the current state, require all edges to match. This
2246 // guards against loops in the middle of a sequence.
2247 if (SomeSuccHasSame && !AllSuccsHaveSame)
2248 S.ClearSequenceProgress();
2254 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
2255 DenseMap<const BasicBlock *, BBState> &BBStates,
2256 MapVector<Value *, RRInfo> &Retains) {
2257 bool NestingDetected = false;
2258 BBState &MyStates = BBStates[BB];
2260 // Merge the states from each successor to compute the initial state
2261 // for the current block.
2262 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2263 succ_const_iterator SI(TI), SE(TI, false);
2265 MyStates.SetAsExit();
2268 const BasicBlock *Succ = *SI++;
2271 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
2272 // If we haven't seen this node yet, then we've found a CFG cycle.
2273 // Be optimistic here; it's CheckForCFGHazards' job detect trouble.
2274 if (I == BBStates.end())
2276 MyStates.InitFromSucc(I->second);
2280 I = BBStates.find(Succ);
2281 if (I != BBStates.end())
2282 MyStates.MergeSucc(I->second);
2288 // Visit all the instructions, bottom-up.
2289 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
2290 Instruction *Inst = llvm::prior(I);
2291 InstructionClass Class = GetInstructionClass(Inst);
2292 const Value *Arg = 0;
2296 Arg = GetObjCArg(Inst);
2298 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2300 // If we see two releases in a row on the same pointer. If so, make
2301 // a note, and we'll cicle back to revisit it after we've
2302 // hopefully eliminated the second release, which may allow us to
2303 // eliminate the first release too.
2304 // Theoretically we could implement removal of nested retain+release
2305 // pairs by making PtrState hold a stack of states, but this is
2306 // simple and avoids adding overhead for the non-nested case.
2307 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease)
2308 NestingDetected = true;
2310 S.SetSeqToRelease(Inst->getMetadata(ImpreciseReleaseMDKind));
2312 S.RRI.KnownSafe = S.IsKnownNested() || S.IsKnownIncremented();
2313 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2314 S.RRI.Calls.insert(Inst);
2316 S.IncrementRefCount();
2317 S.IncrementNestCount();
2320 case IC_RetainBlock:
2323 Arg = GetObjCArg(Inst);
2325 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2326 S.DecrementRefCount();
2327 S.SetAtLeastOneRefCount();
2328 S.DecrementNestCount();
2330 switch (S.GetSeq()) {
2333 case S_MovableRelease:
2335 S.RRI.ReverseInsertPts.clear();
2338 // Don't do retain+release tracking for IC_RetainRV, because it's
2339 // better to let it remain as the first instruction after a call.
2340 if (Class != IC_RetainRV) {
2341 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2342 Retains[Inst] = S.RRI;
2344 S.ClearSequenceProgress();
2349 llvm_unreachable("bottom-up pointer in retain state!");
2353 case IC_AutoreleasepoolPop:
2354 // Conservatively, clear MyStates for all known pointers.
2355 MyStates.clearBottomUpPointers();
2357 case IC_AutoreleasepoolPush:
2359 // These are irrelevant.
2365 // Consider any other possible effects of this instruction on each
2366 // pointer being tracked.
2367 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
2368 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
2369 const Value *Ptr = MI->first;
2371 continue; // Handled above.
2372 PtrState &S = MI->second;
2373 Sequence Seq = S.GetSeq();
2375 // Check for possible releases.
2376 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2377 S.DecrementRefCount();
2380 S.SetSeq(S_CanRelease);
2384 case S_MovableRelease:
2389 llvm_unreachable("bottom-up pointer in retain state!");
2393 // Check for possible direct uses.
2396 case S_MovableRelease:
2397 if (CanUse(Inst, Ptr, PA, Class)) {
2398 S.RRI.ReverseInsertPts.clear();
2399 S.RRI.ReverseInsertPts.insert(Inst);
2401 } else if (Seq == S_Release &&
2402 (Class == IC_User || Class == IC_CallOrUser)) {
2403 // Non-movable releases depend on any possible objc pointer use.
2405 S.RRI.ReverseInsertPts.clear();
2406 S.RRI.ReverseInsertPts.insert(Inst);
2410 if (CanUse(Inst, Ptr, PA, Class))
2418 llvm_unreachable("bottom-up pointer in retain state!");
2423 return NestingDetected;
2427 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
2428 DenseMap<const BasicBlock *, BBState> &BBStates,
2429 DenseMap<Value *, RRInfo> &Releases) {
2430 bool NestingDetected = false;
2431 BBState &MyStates = BBStates[BB];
2433 // Merge the states from each predecessor to compute the initial state
2434 // for the current block.
2435 const_pred_iterator PI(BB), PE(BB, false);
2437 MyStates.SetAsEntry();
2440 const BasicBlock *Pred = *PI++;
2443 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
2444 assert(I != BBStates.end());
2445 // If we haven't seen this node yet, then we've found a CFG cycle.
2446 // Be optimistic here; it's CheckForCFGHazards' job detect trouble.
2447 if (!I->second.isVisitedTopDown())
2449 MyStates.InitFromPred(I->second);
2453 I = BBStates.find(Pred);
2454 assert(I != BBStates.end());
2455 if (I->second.isVisitedTopDown())
2456 MyStates.MergePred(I->second);
2462 // Visit all the instructions, top-down.
2463 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2464 Instruction *Inst = I;
2465 InstructionClass Class = GetInstructionClass(Inst);
2466 const Value *Arg = 0;
2469 case IC_RetainBlock:
2472 Arg = GetObjCArg(Inst);
2474 PtrState &S = MyStates.getPtrTopDownState(Arg);
2476 // Don't do retain+release tracking for IC_RetainRV, because it's
2477 // better to let it remain as the first instruction after a call.
2478 if (Class != IC_RetainRV) {
2479 // If we see two retains in a row on the same pointer. If so, make
2480 // a note, and we'll cicle back to revisit it after we've
2481 // hopefully eliminated the second retain, which may allow us to
2482 // eliminate the first retain too.
2483 // Theoretically we could implement removal of nested retain+release
2484 // pairs by making PtrState hold a stack of states, but this is
2485 // simple and avoids adding overhead for the non-nested case.
2486 if (S.GetSeq() == S_Retain)
2487 NestingDetected = true;
2491 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2492 // Don't check S.IsKnownIncremented() here because it's not
2494 S.RRI.KnownSafe = S.IsKnownNested();
2495 S.RRI.Calls.insert(Inst);
2498 S.SetAtLeastOneRefCount();
2499 S.IncrementRefCount();
2500 S.IncrementNestCount();
2504 Arg = GetObjCArg(Inst);
2506 PtrState &S = MyStates.getPtrTopDownState(Arg);
2507 S.DecrementRefCount();
2508 S.DecrementNestCount();
2510 switch (S.GetSeq()) {
2513 S.RRI.ReverseInsertPts.clear();
2516 S.RRI.ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2517 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2518 Releases[Inst] = S.RRI;
2519 S.ClearSequenceProgress();
2525 case S_MovableRelease:
2526 llvm_unreachable("top-down pointer in release state!");
2530 case IC_AutoreleasepoolPop:
2531 // Conservatively, clear MyStates for all known pointers.
2532 MyStates.clearTopDownPointers();
2534 case IC_AutoreleasepoolPush:
2536 // These are irrelevant.
2542 // Consider any other possible effects of this instruction on each
2543 // pointer being tracked.
2544 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
2545 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
2546 const Value *Ptr = MI->first;
2548 continue; // Handled above.
2549 PtrState &S = MI->second;
2550 Sequence Seq = S.GetSeq();
2552 // Check for possible releases.
2553 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2554 S.DecrementRefCount();
2557 S.SetSeq(S_CanRelease);
2558 S.RRI.ReverseInsertPts.clear();
2559 S.RRI.ReverseInsertPts.insert(Inst);
2561 // One call can't cause a transition from S_Retain to S_CanRelease
2562 // and S_CanRelease to S_Use. If we've made the first transition,
2571 case S_MovableRelease:
2572 llvm_unreachable("top-down pointer in release state!");
2576 // Check for possible direct uses.
2579 if (CanUse(Inst, Ptr, PA, Class))
2588 case S_MovableRelease:
2589 llvm_unreachable("top-down pointer in release state!");
2594 CheckForCFGHazards(BB, BBStates, MyStates);
2595 return NestingDetected;
2598 // Visit - Visit the function both top-down and bottom-up.
2600 ObjCARCOpt::Visit(Function &F,
2601 DenseMap<const BasicBlock *, BBState> &BBStates,
2602 MapVector<Value *, RRInfo> &Retains,
2603 DenseMap<Value *, RRInfo> &Releases) {
2604 // Use reverse-postorder on the reverse CFG for bottom-up, because we
2605 // magically know that loops will be well behaved, i.e. they won't repeatedly
2606 // call retain on a single pointer without doing a release. We can't use
2607 // ReversePostOrderTraversal here because we want to walk up from each
2608 // function exit point.
2609 SmallPtrSet<BasicBlock *, 16> Visited;
2610 SmallVector<std::pair<BasicBlock *, pred_iterator>, 16> Stack;
2611 SmallVector<BasicBlock *, 16> Order;
2612 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2614 if (BB->getTerminator()->getNumSuccessors() == 0)
2615 Stack.push_back(std::make_pair(BB, pred_begin(BB)));
2617 while (!Stack.empty()) {
2618 pred_iterator End = pred_end(Stack.back().first);
2619 while (Stack.back().second != End) {
2620 BasicBlock *BB = *Stack.back().second++;
2621 if (Visited.insert(BB))
2622 Stack.push_back(std::make_pair(BB, pred_begin(BB)));
2624 Order.push_back(Stack.pop_back_val().first);
2626 bool BottomUpNestingDetected = false;
2627 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2628 Order.rbegin(), E = Order.rend(); I != E; ++I) {
2629 BasicBlock *BB = *I;
2630 BottomUpNestingDetected |= VisitBottomUp(BB, BBStates, Retains);
2633 // Use regular reverse-postorder for top-down.
2634 bool TopDownNestingDetected = false;
2635 typedef ReversePostOrderTraversal<Function *> RPOTType;
2637 for (RPOTType::rpo_iterator I = RPOT.begin(), E = RPOT.end(); I != E; ++I) {
2638 BasicBlock *BB = *I;
2639 TopDownNestingDetected |= VisitTopDown(BB, BBStates, Releases);
2642 return TopDownNestingDetected && BottomUpNestingDetected;
2645 /// MoveCalls - Move the calls in RetainsToMove and ReleasesToMove.
2646 void ObjCARCOpt::MoveCalls(Value *Arg,
2647 RRInfo &RetainsToMove,
2648 RRInfo &ReleasesToMove,
2649 MapVector<Value *, RRInfo> &Retains,
2650 DenseMap<Value *, RRInfo> &Releases,
2651 SmallVectorImpl<Instruction *> &DeadInsts,
2653 Type *ArgTy = Arg->getType();
2654 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
2656 // Insert the new retain and release calls.
2657 for (SmallPtrSet<Instruction *, 2>::const_iterator
2658 PI = ReleasesToMove.ReverseInsertPts.begin(),
2659 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2660 Instruction *InsertPt = *PI;
2661 Value *MyArg = ArgTy == ParamTy ? Arg :
2662 new BitCastInst(Arg, ParamTy, "", InsertPt);
2664 CallInst::Create(RetainsToMove.IsRetainBlock ?
2665 getRetainBlockCallee(M) : getRetainCallee(M),
2666 MyArg, "", InsertPt);
2667 Call->setDoesNotThrow();
2668 if (!RetainsToMove.IsRetainBlock)
2669 Call->setTailCall();
2671 for (SmallPtrSet<Instruction *, 2>::const_iterator
2672 PI = RetainsToMove.ReverseInsertPts.begin(),
2673 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2674 Instruction *LastUse = *PI;
2675 Instruction *InsertPts[] = { 0, 0, 0 };
2676 if (InvokeInst *II = dyn_cast<InvokeInst>(LastUse)) {
2677 // We can't insert code immediately after an invoke instruction, so
2678 // insert code at the beginning of both successor blocks instead.
2679 // The invoke's return value isn't available in the unwind block,
2680 // but our releases will never depend on it, because they must be
2681 // paired with retains from before the invoke.
2682 InsertPts[0] = II->getNormalDest()->getFirstNonPHI();
2683 InsertPts[1] = II->getUnwindDest()->getFirstNonPHI();
2685 // Insert code immediately after the last use.
2686 InsertPts[0] = llvm::next(BasicBlock::iterator(LastUse));
2689 for (Instruction **I = InsertPts; *I; ++I) {
2690 Instruction *InsertPt = *I;
2691 Value *MyArg = ArgTy == ParamTy ? Arg :
2692 new BitCastInst(Arg, ParamTy, "", InsertPt);
2693 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
2695 // Attach a clang.imprecise_release metadata tag, if appropriate.
2696 if (MDNode *M = ReleasesToMove.ReleaseMetadata)
2697 Call->setMetadata(ImpreciseReleaseMDKind, M);
2698 Call->setDoesNotThrow();
2699 if (ReleasesToMove.IsTailCallRelease)
2700 Call->setTailCall();
2704 // Delete the original retain and release calls.
2705 for (SmallPtrSet<Instruction *, 2>::const_iterator
2706 AI = RetainsToMove.Calls.begin(),
2707 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
2708 Instruction *OrigRetain = *AI;
2709 Retains.blot(OrigRetain);
2710 DeadInsts.push_back(OrigRetain);
2712 for (SmallPtrSet<Instruction *, 2>::const_iterator
2713 AI = ReleasesToMove.Calls.begin(),
2714 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
2715 Instruction *OrigRelease = *AI;
2716 Releases.erase(OrigRelease);
2717 DeadInsts.push_back(OrigRelease);
2722 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
2724 MapVector<Value *, RRInfo> &Retains,
2725 DenseMap<Value *, RRInfo> &Releases,
2727 bool AnyPairsCompletelyEliminated = false;
2728 RRInfo RetainsToMove;
2729 RRInfo ReleasesToMove;
2730 SmallVector<Instruction *, 4> NewRetains;
2731 SmallVector<Instruction *, 4> NewReleases;
2732 SmallVector<Instruction *, 8> DeadInsts;
2734 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
2735 E = Retains.end(); I != E; ) {
2736 Value *V = (I++)->first;
2737 if (!V) continue; // blotted
2739 Instruction *Retain = cast<Instruction>(V);
2740 Value *Arg = GetObjCArg(Retain);
2742 // If the object being released is in static or stack storage, we know it's
2743 // not being managed by ObjC reference counting, so we can delete pairs
2744 // regardless of what possible decrements or uses lie between them.
2745 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg);
2747 // If a pair happens in a region where it is known that the reference count
2748 // is already incremented, we can similarly ignore possible decrements.
2749 bool KnownSafeTD = true, KnownSafeBU = true;
2751 // Connect the dots between the top-down-collected RetainsToMove and
2752 // bottom-up-collected ReleasesToMove to form sets of related calls.
2753 // This is an iterative process so that we connect multiple releases
2754 // to multiple retains if needed.
2755 unsigned OldDelta = 0;
2756 unsigned NewDelta = 0;
2757 unsigned OldCount = 0;
2758 unsigned NewCount = 0;
2759 bool FirstRelease = true;
2760 bool FirstRetain = true;
2761 NewRetains.push_back(Retain);
2763 for (SmallVectorImpl<Instruction *>::const_iterator
2764 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
2765 Instruction *NewRetain = *NI;
2766 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
2767 assert(It != Retains.end());
2768 const RRInfo &NewRetainRRI = It->second;
2769 KnownSafeTD &= NewRetainRRI.KnownSafe;
2770 for (SmallPtrSet<Instruction *, 2>::const_iterator
2771 LI = NewRetainRRI.Calls.begin(),
2772 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
2773 Instruction *NewRetainRelease = *LI;
2774 DenseMap<Value *, RRInfo>::const_iterator Jt =
2775 Releases.find(NewRetainRelease);
2776 if (Jt == Releases.end())
2778 const RRInfo &NewRetainReleaseRRI = Jt->second;
2779 assert(NewRetainReleaseRRI.Calls.count(NewRetain));
2780 if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
2782 BBStates[NewRetainRelease->getParent()].GetAllPathCount();
2784 // Merge the ReleaseMetadata and IsTailCallRelease values.
2786 ReleasesToMove.ReleaseMetadata =
2787 NewRetainReleaseRRI.ReleaseMetadata;
2788 ReleasesToMove.IsTailCallRelease =
2789 NewRetainReleaseRRI.IsTailCallRelease;
2790 FirstRelease = false;
2792 if (ReleasesToMove.ReleaseMetadata !=
2793 NewRetainReleaseRRI.ReleaseMetadata)
2794 ReleasesToMove.ReleaseMetadata = 0;
2795 if (ReleasesToMove.IsTailCallRelease !=
2796 NewRetainReleaseRRI.IsTailCallRelease)
2797 ReleasesToMove.IsTailCallRelease = false;
2800 // Collect the optimal insertion points.
2802 for (SmallPtrSet<Instruction *, 2>::const_iterator
2803 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
2804 RE = NewRetainReleaseRRI.ReverseInsertPts.end();
2806 Instruction *RIP = *RI;
2807 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
2808 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
2810 NewReleases.push_back(NewRetainRelease);
2815 if (NewReleases.empty()) break;
2817 // Back the other way.
2818 for (SmallVectorImpl<Instruction *>::const_iterator
2819 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
2820 Instruction *NewRelease = *NI;
2821 DenseMap<Value *, RRInfo>::const_iterator It =
2822 Releases.find(NewRelease);
2823 assert(It != Releases.end());
2824 const RRInfo &NewReleaseRRI = It->second;
2825 KnownSafeBU &= NewReleaseRRI.KnownSafe;
2826 for (SmallPtrSet<Instruction *, 2>::const_iterator
2827 LI = NewReleaseRRI.Calls.begin(),
2828 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
2829 Instruction *NewReleaseRetain = *LI;
2830 MapVector<Value *, RRInfo>::const_iterator Jt =
2831 Retains.find(NewReleaseRetain);
2832 if (Jt == Retains.end())
2834 const RRInfo &NewReleaseRetainRRI = Jt->second;
2835 assert(NewReleaseRetainRRI.Calls.count(NewRelease));
2836 if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
2837 unsigned PathCount =
2838 BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
2839 OldDelta += PathCount;
2840 OldCount += PathCount;
2842 // Merge the IsRetainBlock values.
2844 RetainsToMove.IsRetainBlock = NewReleaseRetainRRI.IsRetainBlock;
2845 FirstRetain = false;
2846 } else if (ReleasesToMove.IsRetainBlock !=
2847 NewReleaseRetainRRI.IsRetainBlock)
2848 // It's not possible to merge the sequences if one uses
2849 // objc_retain and the other uses objc_retainBlock.
2852 // Collect the optimal insertion points.
2854 for (SmallPtrSet<Instruction *, 2>::const_iterator
2855 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
2856 RE = NewReleaseRetainRRI.ReverseInsertPts.end();
2858 Instruction *RIP = *RI;
2859 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
2860 PathCount = BBStates[RIP->getParent()].GetAllPathCount();
2861 NewDelta += PathCount;
2862 NewCount += PathCount;
2865 NewRetains.push_back(NewReleaseRetain);
2869 NewReleases.clear();
2870 if (NewRetains.empty()) break;
2873 // If the pointer is known incremented or nested, we can safely delete the
2874 // pair regardless of what's between them.
2875 if (KnownSafeTD || KnownSafeBU) {
2876 RetainsToMove.ReverseInsertPts.clear();
2877 ReleasesToMove.ReverseInsertPts.clear();
2880 // Determine whether the new insertion points we computed preserve the
2881 // balance of retain and release calls through the program.
2882 // TODO: If the fully aggressive solution isn't valid, try to find a
2883 // less aggressive solution which is.
2888 // Determine whether the original call points are balanced in the retain and
2889 // release calls through the program. If not, conservatively don't touch
2891 // TODO: It's theoretically possible to do code motion in this case, as
2892 // long as the existing imbalances are maintained.
2896 // Ok, everything checks out and we're all set. Let's move some code!
2898 AnyPairsCompletelyEliminated = NewCount == 0;
2899 NumRRs += OldCount - NewCount;
2900 MoveCalls(Arg, RetainsToMove, ReleasesToMove,
2901 Retains, Releases, DeadInsts, M);
2904 NewReleases.clear();
2906 RetainsToMove.clear();
2907 ReleasesToMove.clear();
2910 // Now that we're done moving everything, we can delete the newly dead
2911 // instructions, as we no longer need them as insert points.
2912 while (!DeadInsts.empty())
2913 EraseInstruction(DeadInsts.pop_back_val());
2915 return AnyPairsCompletelyEliminated;
2918 /// OptimizeWeakCalls - Weak pointer optimizations.
2919 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
2920 // First, do memdep-style RLE and S2L optimizations. We can't use memdep
2921 // itself because it uses AliasAnalysis and we need to do provenance
2923 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2924 Instruction *Inst = &*I++;
2925 InstructionClass Class = GetBasicInstructionClass(Inst);
2926 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
2929 // Delete objc_loadWeak calls with no users.
2930 if (Class == IC_LoadWeak && Inst->use_empty()) {
2931 Inst->eraseFromParent();
2935 // TODO: For now, just look for an earlier available version of this value
2936 // within the same block. Theoretically, we could do memdep-style non-local
2937 // analysis too, but that would want caching. A better approach would be to
2938 // use the technique that EarlyCSE uses.
2939 inst_iterator Current = llvm::prior(I);
2940 BasicBlock *CurrentBB = Current.getBasicBlockIterator();
2941 for (BasicBlock::iterator B = CurrentBB->begin(),
2942 J = Current.getInstructionIterator();
2944 Instruction *EarlierInst = &*llvm::prior(J);
2945 InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
2946 switch (EarlierClass) {
2948 case IC_LoadWeakRetained: {
2949 // If this is loading from the same pointer, replace this load's value
2951 CallInst *Call = cast<CallInst>(Inst);
2952 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2953 Value *Arg = Call->getArgOperand(0);
2954 Value *EarlierArg = EarlierCall->getArgOperand(0);
2955 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2956 case AliasAnalysis::MustAlias:
2958 // If the load has a builtin retain, insert a plain retain for it.
2959 if (Class == IC_LoadWeakRetained) {
2961 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2965 // Zap the fully redundant load.
2966 Call->replaceAllUsesWith(EarlierCall);
2967 Call->eraseFromParent();
2969 case AliasAnalysis::MayAlias:
2970 case AliasAnalysis::PartialAlias:
2972 case AliasAnalysis::NoAlias:
2979 // If this is storing to the same pointer and has the same size etc.
2980 // replace this load's value with the stored value.
2981 CallInst *Call = cast<CallInst>(Inst);
2982 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2983 Value *Arg = Call->getArgOperand(0);
2984 Value *EarlierArg = EarlierCall->getArgOperand(0);
2985 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2986 case AliasAnalysis::MustAlias:
2988 // If the load has a builtin retain, insert a plain retain for it.
2989 if (Class == IC_LoadWeakRetained) {
2991 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2995 // Zap the fully redundant load.
2996 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
2997 Call->eraseFromParent();
2999 case AliasAnalysis::MayAlias:
3000 case AliasAnalysis::PartialAlias:
3002 case AliasAnalysis::NoAlias:
3009 // TOOD: Grab the copied value.
3011 case IC_AutoreleasepoolPush:
3014 // Weak pointers are only modified through the weak entry points
3015 // (and arbitrary calls, which could call the weak entry points).
3018 // Anything else could modify the weak pointer.
3025 // Then, for each destroyWeak with an alloca operand, check to see if
3026 // the alloca and all its users can be zapped.
3027 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3028 Instruction *Inst = &*I++;
3029 InstructionClass Class = GetBasicInstructionClass(Inst);
3030 if (Class != IC_DestroyWeak)
3033 CallInst *Call = cast<CallInst>(Inst);
3034 Value *Arg = Call->getArgOperand(0);
3035 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
3036 for (Value::use_iterator UI = Alloca->use_begin(),
3037 UE = Alloca->use_end(); UI != UE; ++UI) {
3038 Instruction *UserInst = cast<Instruction>(*UI);
3039 switch (GetBasicInstructionClass(UserInst)) {
3042 case IC_DestroyWeak:
3049 for (Value::use_iterator UI = Alloca->use_begin(),
3050 UE = Alloca->use_end(); UI != UE; ) {
3051 CallInst *UserInst = cast<CallInst>(*UI++);
3052 if (!UserInst->use_empty())
3053 UserInst->replaceAllUsesWith(UserInst->getOperand(1));
3054 UserInst->eraseFromParent();
3056 Alloca->eraseFromParent();
3062 /// OptimizeSequences - Identify program paths which execute sequences of
3063 /// retains and releases which can be eliminated.
3064 bool ObjCARCOpt::OptimizeSequences(Function &F) {
3065 /// Releases, Retains - These are used to store the results of the main flow
3066 /// analysis. These use Value* as the key instead of Instruction* so that the
3067 /// map stays valid when we get around to rewriting code and calls get
3068 /// replaced by arguments.
3069 DenseMap<Value *, RRInfo> Releases;
3070 MapVector<Value *, RRInfo> Retains;
3072 /// BBStates, This is used during the traversal of the function to track the
3073 /// states for each identified object at each block.
3074 DenseMap<const BasicBlock *, BBState> BBStates;
3076 // Analyze the CFG of the function, and all instructions.
3077 bool NestingDetected = Visit(F, BBStates, Retains, Releases);
3080 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
3084 /// OptimizeReturns - Look for this pattern:
3086 /// %call = call i8* @something(...)
3087 /// %2 = call i8* @objc_retain(i8* %call)
3088 /// %3 = call i8* @objc_autorelease(i8* %2)
3091 /// And delete the retain and autorelease.
3093 /// Otherwise if it's just this:
3095 /// %3 = call i8* @objc_autorelease(i8* %2)
3098 /// convert the autorelease to autoreleaseRV.
3099 void ObjCARCOpt::OptimizeReturns(Function &F) {
3100 if (!F.getReturnType()->isPointerTy())
3103 SmallPtrSet<Instruction *, 4> DependingInstructions;
3104 SmallPtrSet<const BasicBlock *, 4> Visited;
3105 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
3106 BasicBlock *BB = FI;
3107 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
3110 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
3111 FindDependencies(NeedsPositiveRetainCount, Arg,
3112 BB, Ret, DependingInstructions, Visited, PA);
3113 if (DependingInstructions.size() != 1)
3117 CallInst *Autorelease =
3118 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3121 InstructionClass AutoreleaseClass =
3122 GetBasicInstructionClass(Autorelease);
3123 if (!IsAutorelease(AutoreleaseClass))
3125 if (GetObjCArg(Autorelease) != Arg)
3128 DependingInstructions.clear();
3131 // Check that there is nothing that can affect the reference
3132 // count between the autorelease and the retain.
3133 FindDependencies(CanChangeRetainCount, Arg,
3134 BB, Autorelease, DependingInstructions, Visited, PA);
3135 if (DependingInstructions.size() != 1)
3140 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3142 // Check that we found a retain with the same argument.
3144 !IsRetain(GetBasicInstructionClass(Retain)) ||
3145 GetObjCArg(Retain) != Arg)
3148 DependingInstructions.clear();
3151 // Convert the autorelease to an autoreleaseRV, since it's
3152 // returning the value.
3153 if (AutoreleaseClass == IC_Autorelease) {
3154 Autorelease->setCalledFunction(getAutoreleaseRVCallee(F.getParent()));
3155 AutoreleaseClass = IC_AutoreleaseRV;
3158 // Check that there is nothing that can affect the reference
3159 // count between the retain and the call.
3160 FindDependencies(CanChangeRetainCount, Arg, BB, Retain,
3161 DependingInstructions, Visited, PA);
3162 if (DependingInstructions.size() != 1)
3167 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3169 // Check that the pointer is the return value of the call.
3170 if (!Call || Arg != Call)
3173 // Check that the call is a regular call.
3174 InstructionClass Class = GetBasicInstructionClass(Call);
3175 if (Class != IC_CallOrUser && Class != IC_Call)
3178 // If so, we can zap the retain and autorelease.
3181 EraseInstruction(Retain);
3182 EraseInstruction(Autorelease);
3188 DependingInstructions.clear();
3193 bool ObjCARCOpt::doInitialization(Module &M) {
3197 Run = ModuleHasARC(M);
3201 // Identify the imprecise release metadata kind.
3202 ImpreciseReleaseMDKind =
3203 M.getContext().getMDKindID("clang.imprecise_release");
3205 // Intuitively, objc_retain and others are nocapture, however in practice
3206 // they are not, because they return their argument value. And objc_release
3207 // calls finalizers.
3209 // These are initialized lazily.
3211 AutoreleaseRVCallee = 0;
3214 RetainBlockCallee = 0;
3215 AutoreleaseCallee = 0;
3220 bool ObjCARCOpt::runOnFunction(Function &F) {
3224 // If nothing in the Module uses ARC, don't do anything.
3230 PA.setAA(&getAnalysis<AliasAnalysis>());
3232 // This pass performs several distinct transformations. As a compile-time aid
3233 // when compiling code that isn't ObjC, skip these if the relevant ObjC
3234 // library functions aren't declared.
3236 // Preliminary optimizations. This also computs UsedInThisFunction.
3237 OptimizeIndividualCalls(F);
3239 // Optimizations for weak pointers.
3240 if (UsedInThisFunction & ((1 << IC_LoadWeak) |
3241 (1 << IC_LoadWeakRetained) |
3242 (1 << IC_StoreWeak) |
3243 (1 << IC_InitWeak) |
3244 (1 << IC_CopyWeak) |
3245 (1 << IC_MoveWeak) |
3246 (1 << IC_DestroyWeak)))
3247 OptimizeWeakCalls(F);
3249 // Optimizations for retain+release pairs.
3250 if (UsedInThisFunction & ((1 << IC_Retain) |
3251 (1 << IC_RetainRV) |
3252 (1 << IC_RetainBlock)))
3253 if (UsedInThisFunction & (1 << IC_Release))
3254 // Run OptimizeSequences until it either stops making changes or
3255 // no retain+release pair nesting is detected.
3256 while (OptimizeSequences(F)) {}
3258 // Optimizations if objc_autorelease is used.
3259 if (UsedInThisFunction &
3260 ((1 << IC_Autorelease) | (1 << IC_AutoreleaseRV)))
3266 void ObjCARCOpt::releaseMemory() {
3270 //===----------------------------------------------------------------------===//
3272 //===----------------------------------------------------------------------===//
3274 // TODO: ObjCARCContract could insert PHI nodes when uses aren't
3275 // dominated by single calls.
3277 #include "llvm/Operator.h"
3278 #include "llvm/InlineAsm.h"
3279 #include "llvm/Analysis/Dominators.h"
3281 STATISTIC(NumStoreStrongs, "Number objc_storeStrong calls formed");
3284 /// ObjCARCContract - Late ARC optimizations. These change the IR in a way
3285 /// that makes it difficult to be analyzed by ObjCARCOpt, so it's run late.
3286 class ObjCARCContract : public FunctionPass {
3290 ProvenanceAnalysis PA;
3292 /// Run - A flag indicating whether this optimization pass should run.
3295 /// StoreStrongCallee, etc. - Declarations for ObjC runtime
3296 /// functions, for use in creating calls to them. These are initialized
3297 /// lazily to avoid cluttering up the Module with unused declarations.
3298 Constant *StoreStrongCallee,
3299 *RetainAutoreleaseCallee, *RetainAutoreleaseRVCallee;
3301 /// RetainRVMarker - The inline asm string to insert between calls and
3302 /// RetainRV calls to make the optimization work on targets which need it.
3303 const MDString *RetainRVMarker;
3305 Constant *getStoreStrongCallee(Module *M);
3306 Constant *getRetainAutoreleaseCallee(Module *M);
3307 Constant *getRetainAutoreleaseRVCallee(Module *M);
3309 bool ContractAutorelease(Function &F, Instruction *Autorelease,
3310 InstructionClass Class,
3311 SmallPtrSet<Instruction *, 4>
3312 &DependingInstructions,
3313 SmallPtrSet<const BasicBlock *, 4>
3316 void ContractRelease(Instruction *Release,
3317 inst_iterator &Iter);
3319 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
3320 virtual bool doInitialization(Module &M);
3321 virtual bool runOnFunction(Function &F);
3325 ObjCARCContract() : FunctionPass(ID) {
3326 initializeObjCARCContractPass(*PassRegistry::getPassRegistry());
3331 char ObjCARCContract::ID = 0;
3332 INITIALIZE_PASS_BEGIN(ObjCARCContract,
3333 "objc-arc-contract", "ObjC ARC contraction", false, false)
3334 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3335 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3336 INITIALIZE_PASS_END(ObjCARCContract,
3337 "objc-arc-contract", "ObjC ARC contraction", false, false)
3339 Pass *llvm::createObjCARCContractPass() {
3340 return new ObjCARCContract();
3343 void ObjCARCContract::getAnalysisUsage(AnalysisUsage &AU) const {
3344 AU.addRequired<AliasAnalysis>();
3345 AU.addRequired<DominatorTree>();
3346 AU.setPreservesCFG();
3349 Constant *ObjCARCContract::getStoreStrongCallee(Module *M) {
3350 if (!StoreStrongCallee) {
3351 LLVMContext &C = M->getContext();
3352 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3353 Type *I8XX = PointerType::getUnqual(I8X);
3354 std::vector<Type *> Params;
3355 Params.push_back(I8XX);
3356 Params.push_back(I8X);
3358 AttrListPtr Attributes;
3359 Attributes.addAttr(~0u, Attribute::NoUnwind);
3360 Attributes.addAttr(1, Attribute::NoCapture);
3363 M->getOrInsertFunction(
3365 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
3368 return StoreStrongCallee;
3371 Constant *ObjCARCContract::getRetainAutoreleaseCallee(Module *M) {
3372 if (!RetainAutoreleaseCallee) {
3373 LLVMContext &C = M->getContext();
3374 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3375 std::vector<Type *> Params;
3376 Params.push_back(I8X);
3378 FunctionType::get(I8X, Params, /*isVarArg=*/false);
3379 AttrListPtr Attributes;
3380 Attributes.addAttr(~0u, Attribute::NoUnwind);
3381 RetainAutoreleaseCallee =
3382 M->getOrInsertFunction("objc_retainAutorelease", FTy, Attributes);
3384 return RetainAutoreleaseCallee;
3387 Constant *ObjCARCContract::getRetainAutoreleaseRVCallee(Module *M) {
3388 if (!RetainAutoreleaseRVCallee) {
3389 LLVMContext &C = M->getContext();
3390 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3391 std::vector<Type *> Params;
3392 Params.push_back(I8X);
3394 FunctionType::get(I8X, Params, /*isVarArg=*/false);
3395 AttrListPtr Attributes;
3396 Attributes.addAttr(~0u, Attribute::NoUnwind);
3397 RetainAutoreleaseRVCallee =
3398 M->getOrInsertFunction("objc_retainAutoreleaseReturnValue", FTy,
3401 return RetainAutoreleaseRVCallee;
3404 /// ContractAutorelease - Merge an autorelease with a retain into a fused
3407 ObjCARCContract::ContractAutorelease(Function &F, Instruction *Autorelease,
3408 InstructionClass Class,
3409 SmallPtrSet<Instruction *, 4>
3410 &DependingInstructions,
3411 SmallPtrSet<const BasicBlock *, 4>
3413 const Value *Arg = GetObjCArg(Autorelease);
3415 // Check that there are no instructions between the retain and the autorelease
3416 // (such as an autorelease_pop) which may change the count.
3417 CallInst *Retain = 0;
3418 if (Class == IC_AutoreleaseRV)
3419 FindDependencies(RetainAutoreleaseRVDep, Arg,
3420 Autorelease->getParent(), Autorelease,
3421 DependingInstructions, Visited, PA);
3423 FindDependencies(RetainAutoreleaseDep, Arg,
3424 Autorelease->getParent(), Autorelease,
3425 DependingInstructions, Visited, PA);
3428 if (DependingInstructions.size() != 1) {
3429 DependingInstructions.clear();
3433 Retain = dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3434 DependingInstructions.clear();
3437 GetBasicInstructionClass(Retain) != IC_Retain ||
3438 GetObjCArg(Retain) != Arg)
3444 if (Class == IC_AutoreleaseRV)
3445 Retain->setCalledFunction(getRetainAutoreleaseRVCallee(F.getParent()));
3447 Retain->setCalledFunction(getRetainAutoreleaseCallee(F.getParent()));
3449 EraseInstruction(Autorelease);
3453 /// ContractRelease - Attempt to merge an objc_release with a store, load, and
3454 /// objc_retain to form an objc_storeStrong. This can be a little tricky because
3455 /// the instructions don't always appear in order, and there may be unrelated
3456 /// intervening instructions.
3457 void ObjCARCContract::ContractRelease(Instruction *Release,
3458 inst_iterator &Iter) {
3459 LoadInst *Load = dyn_cast<LoadInst>(GetObjCArg(Release));
3460 if (!Load || Load->isVolatile()) return;
3462 // For now, require everything to be in one basic block.
3463 BasicBlock *BB = Release->getParent();
3464 if (Load->getParent() != BB) return;
3466 // Walk down to find the store.
3467 BasicBlock::iterator I = Load, End = BB->end();
3469 AliasAnalysis::Location Loc = AA->getLocation(Load);
3472 IsRetain(GetBasicInstructionClass(I)) ||
3473 !(AA->getModRefInfo(I, Loc) & AliasAnalysis::Mod)))
3475 StoreInst *Store = dyn_cast<StoreInst>(I);
3476 if (!Store || Store->isVolatile()) return;
3477 if (Store->getPointerOperand() != Loc.Ptr) return;
3479 Value *New = StripPointerCastsAndObjCCalls(Store->getValueOperand());
3481 // Walk up to find the retain.
3483 BasicBlock::iterator Begin = BB->begin();
3484 while (I != Begin && GetBasicInstructionClass(I) != IC_Retain)
3486 Instruction *Retain = I;
3487 if (GetBasicInstructionClass(Retain) != IC_Retain) return;
3488 if (GetObjCArg(Retain) != New) return;
3493 LLVMContext &C = Release->getContext();
3494 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3495 Type *I8XX = PointerType::getUnqual(I8X);
3497 Value *Args[] = { Load->getPointerOperand(), New };
3498 if (Args[0]->getType() != I8XX)
3499 Args[0] = new BitCastInst(Args[0], I8XX, "", Store);
3500 if (Args[1]->getType() != I8X)
3501 Args[1] = new BitCastInst(Args[1], I8X, "", Store);
3502 CallInst *StoreStrong =
3503 CallInst::Create(getStoreStrongCallee(BB->getParent()->getParent()),
3505 StoreStrong->setDoesNotThrow();
3506 StoreStrong->setDebugLoc(Store->getDebugLoc());
3508 if (&*Iter == Store) ++Iter;
3509 Store->eraseFromParent();
3510 Release->eraseFromParent();
3511 EraseInstruction(Retain);
3512 if (Load->use_empty())
3513 Load->eraseFromParent();
3516 bool ObjCARCContract::doInitialization(Module &M) {
3517 Run = ModuleHasARC(M);
3521 // These are initialized lazily.
3522 StoreStrongCallee = 0;
3523 RetainAutoreleaseCallee = 0;
3524 RetainAutoreleaseRVCallee = 0;
3526 // Initialize RetainRVMarker.
3528 if (NamedMDNode *NMD =
3529 M.getNamedMetadata("clang.arc.retainAutoreleasedReturnValueMarker"))
3530 if (NMD->getNumOperands() == 1) {
3531 const MDNode *N = NMD->getOperand(0);
3532 if (N->getNumOperands() == 1)
3533 if (const MDString *S = dyn_cast<MDString>(N->getOperand(0)))
3540 bool ObjCARCContract::runOnFunction(Function &F) {
3544 // If nothing in the Module uses ARC, don't do anything.
3549 AA = &getAnalysis<AliasAnalysis>();
3550 DT = &getAnalysis<DominatorTree>();
3552 PA.setAA(&getAnalysis<AliasAnalysis>());
3554 // For ObjC library calls which return their argument, replace uses of the
3555 // argument with uses of the call return value, if it dominates the use. This
3556 // reduces register pressure.
3557 SmallPtrSet<Instruction *, 4> DependingInstructions;
3558 SmallPtrSet<const BasicBlock *, 4> Visited;
3559 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3560 Instruction *Inst = &*I++;
3562 // Only these library routines return their argument. In particular,
3563 // objc_retainBlock does not necessarily return its argument.
3564 InstructionClass Class = GetBasicInstructionClass(Inst);
3567 case IC_FusedRetainAutorelease:
3568 case IC_FusedRetainAutoreleaseRV:
3570 case IC_Autorelease:
3571 case IC_AutoreleaseRV:
3572 if (ContractAutorelease(F, Inst, Class, DependingInstructions, Visited))
3576 // If we're compiling for a target which needs a special inline-asm
3577 // marker to do the retainAutoreleasedReturnValue optimization,
3579 if (!RetainRVMarker)
3581 BasicBlock::iterator BBI = Inst;
3583 while (isNoopInstruction(BBI)) --BBI;
3584 if (&*BBI == GetObjCArg(Inst)) {
3586 InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
3587 /*isVarArg=*/false),
3588 RetainRVMarker->getString(),
3589 /*Constraints=*/"", /*hasSideEffects=*/true);
3590 CallInst::Create(IA, "", Inst);
3595 // objc_initWeak(p, null) => *p = null
3596 CallInst *CI = cast<CallInst>(Inst);
3597 if (isNullOrUndef(CI->getArgOperand(1))) {
3599 ConstantPointerNull::get(cast<PointerType>(CI->getType()));
3601 new StoreInst(Null, CI->getArgOperand(0), CI);
3602 CI->replaceAllUsesWith(Null);
3603 CI->eraseFromParent();
3608 ContractRelease(Inst, I);
3614 // Don't use GetObjCArg because we don't want to look through bitcasts
3615 // and such; to do the replacement, the argument must have type i8*.
3616 const Value *Arg = cast<CallInst>(Inst)->getArgOperand(0);
3618 // If we're compiling bugpointed code, don't get in trouble.
3619 if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))
3621 // Look through the uses of the pointer.
3622 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
3624 Use &U = UI.getUse();
3625 unsigned OperandNo = UI.getOperandNo();
3626 ++UI; // Increment UI now, because we may unlink its element.
3627 if (Instruction *UserInst = dyn_cast<Instruction>(U.getUser()))
3628 if (Inst != UserInst && DT->dominates(Inst, UserInst)) {
3630 Instruction *Replacement = Inst;
3631 Type *UseTy = U.get()->getType();
3632 if (PHINode *PHI = dyn_cast<PHINode>(UserInst)) {
3633 // For PHI nodes, insert the bitcast in the predecessor block.
3635 PHINode::getIncomingValueNumForOperand(OperandNo);
3637 PHI->getIncomingBlock(ValNo);
3638 if (Replacement->getType() != UseTy)
3639 Replacement = new BitCastInst(Replacement, UseTy, "",
3641 for (unsigned i = 0, e = PHI->getNumIncomingValues();
3643 if (PHI->getIncomingBlock(i) == BB) {
3644 // Keep the UI iterator valid.
3645 if (&PHI->getOperandUse(
3646 PHINode::getOperandNumForIncomingValue(i)) ==
3649 PHI->setIncomingValue(i, Replacement);
3652 if (Replacement->getType() != UseTy)
3653 Replacement = new BitCastInst(Replacement, UseTy, "", UserInst);
3659 // If Arg is a no-op casted pointer, strip one level of casts and
3661 if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg))
3662 Arg = BI->getOperand(0);
3663 else if (isa<GEPOperator>(Arg) &&
3664 cast<GEPOperator>(Arg)->hasAllZeroIndices())
3665 Arg = cast<GEPOperator>(Arg)->getPointerOperand();
3666 else if (isa<GlobalAlias>(Arg) &&
3667 !cast<GlobalAlias>(Arg)->mayBeOverridden())
3668 Arg = cast<GlobalAlias>(Arg)->getAliasee();