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 #include "llvm/GlobalAlias.h"
883 #include "llvm/Constants.h"
884 #include "llvm/LLVMContext.h"
885 #include "llvm/Support/ErrorHandling.h"
886 #include "llvm/Support/CFG.h"
887 #include "llvm/ADT/PostOrderIterator.h"
888 #include "llvm/ADT/Statistic.h"
890 STATISTIC(NumNoops, "Number of no-op objc calls eliminated");
891 STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated");
892 STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases");
893 STATISTIC(NumRets, "Number of return value forwarding "
894 "retain+autoreleaes eliminated");
895 STATISTIC(NumRRs, "Number of retain+release paths eliminated");
896 STATISTIC(NumPeeps, "Number of calls peephole-optimized");
899 /// ProvenanceAnalysis - This is similar to BasicAliasAnalysis, and it
900 /// uses many of the same techniques, except it uses special ObjC-specific
901 /// reasoning about pointer relationships.
902 class ProvenanceAnalysis {
905 typedef std::pair<const Value *, const Value *> ValuePairTy;
906 typedef DenseMap<ValuePairTy, bool> CachedResultsTy;
907 CachedResultsTy CachedResults;
909 bool relatedCheck(const Value *A, const Value *B);
910 bool relatedSelect(const SelectInst *A, const Value *B);
911 bool relatedPHI(const PHINode *A, const Value *B);
914 void operator=(const ProvenanceAnalysis &);
915 ProvenanceAnalysis(const ProvenanceAnalysis &);
918 ProvenanceAnalysis() {}
920 void setAA(AliasAnalysis *aa) { AA = aa; }
922 AliasAnalysis *getAA() const { return AA; }
924 bool related(const Value *A, const Value *B);
927 CachedResults.clear();
932 bool ProvenanceAnalysis::relatedSelect(const SelectInst *A, const Value *B) {
933 // If the values are Selects with the same condition, we can do a more precise
934 // check: just check for relations between the values on corresponding arms.
935 if (const SelectInst *SB = dyn_cast<SelectInst>(B))
936 if (A->getCondition() == SB->getCondition()) {
937 if (related(A->getTrueValue(), SB->getTrueValue()))
939 if (related(A->getFalseValue(), SB->getFalseValue()))
944 // Check both arms of the Select node individually.
945 if (related(A->getTrueValue(), B))
947 if (related(A->getFalseValue(), B))
950 // The arms both checked out.
954 bool ProvenanceAnalysis::relatedPHI(const PHINode *A, const Value *B) {
955 // If the values are PHIs in the same block, we can do a more precise as well
956 // as efficient check: just check for relations between the values on
957 // corresponding edges.
958 if (const PHINode *PNB = dyn_cast<PHINode>(B))
959 if (PNB->getParent() == A->getParent()) {
960 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i)
961 if (related(A->getIncomingValue(i),
962 PNB->getIncomingValueForBlock(A->getIncomingBlock(i))))
967 // Check each unique source of the PHI node against B.
968 SmallPtrSet<const Value *, 4> UniqueSrc;
969 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i) {
970 const Value *PV1 = A->getIncomingValue(i);
971 if (UniqueSrc.insert(PV1) && related(PV1, B))
975 // All of the arms checked out.
979 /// isStoredObjCPointer - Test if the value of P, or any value covered by its
980 /// provenance, is ever stored within the function (not counting callees).
981 static bool isStoredObjCPointer(const Value *P) {
982 SmallPtrSet<const Value *, 8> Visited;
983 SmallVector<const Value *, 8> Worklist;
984 Worklist.push_back(P);
987 P = Worklist.pop_back_val();
988 for (Value::const_use_iterator UI = P->use_begin(), UE = P->use_end();
990 const User *Ur = *UI;
991 if (isa<StoreInst>(Ur)) {
992 if (UI.getOperandNo() == 0)
993 // The pointer is stored.
995 // The pointed is stored through.
998 if (isa<CallInst>(Ur))
999 // The pointer is passed as an argument, ignore this.
1001 if (isa<PtrToIntInst>(P))
1002 // Assume the worst.
1004 if (Visited.insert(Ur))
1005 Worklist.push_back(Ur);
1007 } while (!Worklist.empty());
1009 // Everything checked out.
1013 bool ProvenanceAnalysis::relatedCheck(const Value *A, const Value *B) {
1014 // Skip past provenance pass-throughs.
1015 A = GetUnderlyingObjCPtr(A);
1016 B = GetUnderlyingObjCPtr(B);
1022 // Ask regular AliasAnalysis, for a first approximation.
1023 switch (AA->alias(A, B)) {
1024 case AliasAnalysis::NoAlias:
1026 case AliasAnalysis::MustAlias:
1027 case AliasAnalysis::PartialAlias:
1029 case AliasAnalysis::MayAlias:
1033 bool AIsIdentified = IsObjCIdentifiedObject(A);
1034 bool BIsIdentified = IsObjCIdentifiedObject(B);
1036 // An ObjC-Identified object can't alias a load if it is never locally stored.
1037 if (AIsIdentified) {
1038 if (BIsIdentified) {
1039 // If both pointers have provenance, they can be directly compared.
1043 if (isa<LoadInst>(B))
1044 return isStoredObjCPointer(A);
1047 if (BIsIdentified && isa<LoadInst>(A))
1048 return isStoredObjCPointer(B);
1051 // Special handling for PHI and Select.
1052 if (const PHINode *PN = dyn_cast<PHINode>(A))
1053 return relatedPHI(PN, B);
1054 if (const PHINode *PN = dyn_cast<PHINode>(B))
1055 return relatedPHI(PN, A);
1056 if (const SelectInst *S = dyn_cast<SelectInst>(A))
1057 return relatedSelect(S, B);
1058 if (const SelectInst *S = dyn_cast<SelectInst>(B))
1059 return relatedSelect(S, A);
1065 bool ProvenanceAnalysis::related(const Value *A, const Value *B) {
1066 // Begin by inserting a conservative value into the map. If the insertion
1067 // fails, we have the answer already. If it succeeds, leave it there until we
1068 // compute the real answer to guard against recursive queries.
1069 if (A > B) std::swap(A, B);
1070 std::pair<CachedResultsTy::iterator, bool> Pair =
1071 CachedResults.insert(std::make_pair(ValuePairTy(A, B), true));
1073 return Pair.first->second;
1075 bool Result = relatedCheck(A, B);
1076 CachedResults[ValuePairTy(A, B)] = Result;
1081 // Sequence - A sequence of states that a pointer may go through in which an
1082 // objc_retain and objc_release are actually needed.
1085 S_Retain, ///< objc_retain(x)
1086 S_CanRelease, ///< foo(x) -- x could possibly see a ref count decrement
1087 S_Use, ///< any use of x
1088 S_Stop, ///< like S_Release, but code motion is stopped
1089 S_Release, ///< objc_release(x)
1090 S_MovableRelease ///< objc_release(x), !clang.imprecise_release
1094 static Sequence MergeSeqs(Sequence A, Sequence B, bool TopDown) {
1098 if (A == S_None || B == S_None)
1101 // Note that we can't merge S_CanRelease and S_Use.
1102 if (A > B) std::swap(A, B);
1104 // Choose the side which is further along in the sequence.
1105 if (A == S_Retain && (B == S_CanRelease || B == S_Use))
1108 // Choose the side which is further along in the sequence.
1109 if ((A == S_Use || A == S_CanRelease) &&
1110 (B == S_Release || B == S_Stop || B == S_MovableRelease))
1112 // If both sides are releases, choose the more conservative one.
1113 if (A == S_Stop && (B == S_Release || B == S_MovableRelease))
1115 if (A == S_Release && B == S_MovableRelease)
1123 /// RRInfo - Unidirectional information about either a
1124 /// retain-decrement-use-release sequence or release-use-decrement-retain
1125 /// reverese sequence.
1127 /// KnownIncremented - After an objc_retain, the reference count of the
1128 /// referenced object is known to be positive. Similarly, before an
1129 /// objc_release, the reference count of the referenced object is known to
1130 /// be positive. If there are retain-release pairs in code regions where the
1131 /// retain count is known to be positive, they can be eliminated, regardless
1132 /// of any side effects between them.
1133 bool KnownIncremented;
1135 /// IsRetainBlock - True if the Calls are objc_retainBlock calls (as
1136 /// opposed to objc_retain calls).
1139 /// IsTailCallRelease - True of the objc_release calls are all marked
1140 /// with the "tail" keyword.
1141 bool IsTailCallRelease;
1143 /// ReleaseMetadata - If the Calls are objc_release calls and they all have
1144 /// a clang.imprecise_release tag, this is the metadata tag.
1145 MDNode *ReleaseMetadata;
1147 /// Calls - For a top-down sequence, the set of objc_retains or
1148 /// objc_retainBlocks. For bottom-up, the set of objc_releases.
1149 SmallPtrSet<Instruction *, 2> Calls;
1151 /// ReverseInsertPts - The set of optimal insert positions for
1152 /// moving calls in the opposite sequence.
1153 SmallPtrSet<Instruction *, 2> ReverseInsertPts;
1156 KnownIncremented(false), IsRetainBlock(false), IsTailCallRelease(false),
1157 ReleaseMetadata(0) {}
1163 void RRInfo::clear() {
1164 KnownIncremented = false;
1165 IsRetainBlock = false;
1166 IsTailCallRelease = false;
1167 ReleaseMetadata = 0;
1169 ReverseInsertPts.clear();
1173 /// PtrState - This class summarizes several per-pointer runtime properties
1174 /// which are propogated through the flow graph.
1176 /// RefCount - The known minimum number of reference count increments.
1179 /// Seq - The current position in the sequence.
1183 /// RRI - Unidirectional information about the current sequence.
1184 /// TODO: Encapsulate this better.
1187 PtrState() : RefCount(0), Seq(S_None) {}
1189 void IncrementRefCount() {
1190 if (RefCount != UINT_MAX) ++RefCount;
1193 void DecrementRefCount() {
1194 if (RefCount != 0) --RefCount;
1197 bool IsKnownIncremented() const {
1198 return RefCount > 0;
1201 void SetSeq(Sequence NewSeq) {
1205 void SetSeqToRelease(MDNode *M) {
1206 if (Seq == S_None || Seq == S_Use) {
1207 Seq = M ? S_MovableRelease : S_Release;
1208 RRI.ReleaseMetadata = M;
1209 } else if (Seq != S_MovableRelease || RRI.ReleaseMetadata != M) {
1211 RRI.ReleaseMetadata = 0;
1215 Sequence GetSeq() const {
1219 void ClearSequenceProgress() {
1224 void Merge(const PtrState &Other, bool TopDown);
1229 PtrState::Merge(const PtrState &Other, bool TopDown) {
1230 Seq = MergeSeqs(Seq, Other.Seq, TopDown);
1231 RefCount = std::min(RefCount, Other.RefCount);
1233 // We can't merge a plain objc_retain with an objc_retainBlock.
1234 if (RRI.IsRetainBlock != Other.RRI.IsRetainBlock)
1237 if (Seq == S_None) {
1240 // Conservatively merge the ReleaseMetadata information.
1241 if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
1242 RRI.ReleaseMetadata = 0;
1244 RRI.KnownIncremented = RRI.KnownIncremented && Other.RRI.KnownIncremented;
1245 RRI.IsTailCallRelease = RRI.IsTailCallRelease && Other.RRI.IsTailCallRelease;
1246 RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
1247 RRI.ReverseInsertPts.insert(Other.RRI.ReverseInsertPts.begin(),
1248 Other.RRI.ReverseInsertPts.end());
1253 /// BBState - Per-BasicBlock state.
1255 /// TopDownPathCount - The number of unique control paths from the entry
1256 /// which can reach this block.
1257 unsigned TopDownPathCount;
1259 /// BottomUpPathCount - The number of unique control paths to exits
1260 /// from this block.
1261 unsigned BottomUpPathCount;
1263 /// MapTy - A type for PerPtrTopDown and PerPtrBottomUp.
1264 typedef MapVector<const Value *, PtrState> MapTy;
1266 /// PerPtrTopDown - The top-down traversal uses this to record information
1267 /// known about a pointer at the bottom of each block.
1268 MapTy PerPtrTopDown;
1270 /// PerPtrBottomUp - The bottom-up traversal uses this to record information
1271 /// known about a pointer at the top of each block.
1272 MapTy PerPtrBottomUp;
1275 BBState() : TopDownPathCount(0), BottomUpPathCount(0) {}
1277 typedef MapTy::iterator ptr_iterator;
1278 typedef MapTy::const_iterator ptr_const_iterator;
1280 ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); }
1281 ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); }
1282 ptr_const_iterator top_down_ptr_begin() const {
1283 return PerPtrTopDown.begin();
1285 ptr_const_iterator top_down_ptr_end() const {
1286 return PerPtrTopDown.end();
1289 ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); }
1290 ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); }
1291 ptr_const_iterator bottom_up_ptr_begin() const {
1292 return PerPtrBottomUp.begin();
1294 ptr_const_iterator bottom_up_ptr_end() const {
1295 return PerPtrBottomUp.end();
1298 /// SetAsEntry - Mark this block as being an entry block, which has one
1299 /// path from the entry by definition.
1300 void SetAsEntry() { TopDownPathCount = 1; }
1302 /// SetAsExit - Mark this block as being an exit block, which has one
1303 /// path to an exit by definition.
1304 void SetAsExit() { BottomUpPathCount = 1; }
1306 PtrState &getPtrTopDownState(const Value *Arg) {
1307 return PerPtrTopDown[Arg];
1310 PtrState &getPtrBottomUpState(const Value *Arg) {
1311 return PerPtrBottomUp[Arg];
1314 void clearBottomUpPointers() {
1315 PerPtrBottomUp.clear();
1318 void clearTopDownPointers() {
1319 PerPtrTopDown.clear();
1322 void InitFromPred(const BBState &Other);
1323 void InitFromSucc(const BBState &Other);
1324 void MergePred(const BBState &Other);
1325 void MergeSucc(const BBState &Other);
1327 /// GetAllPathCount - Return the number of possible unique paths from an
1328 /// entry to an exit which pass through this block. This is only valid
1329 /// after both the top-down and bottom-up traversals are complete.
1330 unsigned GetAllPathCount() const {
1331 return TopDownPathCount * BottomUpPathCount;
1336 void BBState::InitFromPred(const BBState &Other) {
1337 PerPtrTopDown = Other.PerPtrTopDown;
1338 TopDownPathCount = Other.TopDownPathCount;
1341 void BBState::InitFromSucc(const BBState &Other) {
1342 PerPtrBottomUp = Other.PerPtrBottomUp;
1343 BottomUpPathCount = Other.BottomUpPathCount;
1346 /// MergePred - The top-down traversal uses this to merge information about
1347 /// predecessors to form the initial state for a new block.
1348 void BBState::MergePred(const BBState &Other) {
1349 // Other.TopDownPathCount can be 0, in which case it is either dead or a
1350 // loop backedge. Loop backedges are special.
1351 TopDownPathCount += Other.TopDownPathCount;
1353 // For each entry in the other set, if our set has an entry with the same key,
1354 // merge the entries. Otherwise, copy the entry and merge it with an empty
1356 for (ptr_const_iterator MI = Other.top_down_ptr_begin(),
1357 ME = Other.top_down_ptr_end(); MI != ME; ++MI) {
1358 std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI);
1359 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1363 // For each entry in our set, if the other set doesn't have an entry with the
1364 // same key, force it to merge with an empty entry.
1365 for (ptr_iterator MI = top_down_ptr_begin(),
1366 ME = top_down_ptr_end(); MI != ME; ++MI)
1367 if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end())
1368 MI->second.Merge(PtrState(), /*TopDown=*/true);
1371 /// MergeSucc - The bottom-up traversal uses this to merge information about
1372 /// successors to form the initial state for a new block.
1373 void BBState::MergeSucc(const BBState &Other) {
1374 // Other.BottomUpPathCount can be 0, in which case it is either dead or a
1375 // loop backedge. Loop backedges are special.
1376 BottomUpPathCount += Other.BottomUpPathCount;
1378 // For each entry in the other set, if our set has an entry with the
1379 // same key, merge the entries. Otherwise, copy the entry and merge
1380 // it with an empty entry.
1381 for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(),
1382 ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) {
1383 std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI);
1384 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1388 // For each entry in our set, if the other set doesn't have an entry
1389 // with the same key, force it to merge with an empty entry.
1390 for (ptr_iterator MI = bottom_up_ptr_begin(),
1391 ME = bottom_up_ptr_end(); MI != ME; ++MI)
1392 if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end())
1393 MI->second.Merge(PtrState(), /*TopDown=*/false);
1397 /// ObjCARCOpt - The main ARC optimization pass.
1398 class ObjCARCOpt : public FunctionPass {
1400 ProvenanceAnalysis PA;
1402 /// Run - A flag indicating whether this optimization pass should run.
1405 /// RetainRVCallee, etc. - Declarations for ObjC runtime
1406 /// functions, for use in creating calls to them. These are initialized
1407 /// lazily to avoid cluttering up the Module with unused declarations.
1408 Constant *RetainRVCallee, *AutoreleaseRVCallee, *ReleaseCallee,
1409 *RetainCallee, *RetainBlockCallee, *AutoreleaseCallee;
1411 /// UsedInThisFunciton - Flags which determine whether each of the
1412 /// interesting runtine functions is in fact used in the current function.
1413 unsigned UsedInThisFunction;
1415 /// ImpreciseReleaseMDKind - The Metadata Kind for clang.imprecise_release
1417 unsigned ImpreciseReleaseMDKind;
1419 Constant *getRetainRVCallee(Module *M);
1420 Constant *getAutoreleaseRVCallee(Module *M);
1421 Constant *getReleaseCallee(Module *M);
1422 Constant *getRetainCallee(Module *M);
1423 Constant *getRetainBlockCallee(Module *M);
1424 Constant *getAutoreleaseCallee(Module *M);
1426 void OptimizeRetainCall(Function &F, Instruction *Retain);
1427 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
1428 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV);
1429 void OptimizeIndividualCalls(Function &F);
1431 void CheckForCFGHazards(const BasicBlock *BB,
1432 DenseMap<const BasicBlock *, BBState> &BBStates,
1433 BBState &MyStates) const;
1434 bool VisitBottomUp(BasicBlock *BB,
1435 DenseMap<const BasicBlock *, BBState> &BBStates,
1436 MapVector<Value *, RRInfo> &Retains);
1437 bool VisitTopDown(BasicBlock *BB,
1438 DenseMap<const BasicBlock *, BBState> &BBStates,
1439 DenseMap<Value *, RRInfo> &Releases);
1440 bool Visit(Function &F,
1441 DenseMap<const BasicBlock *, BBState> &BBStates,
1442 MapVector<Value *, RRInfo> &Retains,
1443 DenseMap<Value *, RRInfo> &Releases);
1445 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1446 MapVector<Value *, RRInfo> &Retains,
1447 DenseMap<Value *, RRInfo> &Releases,
1448 SmallVectorImpl<Instruction *> &DeadInsts,
1451 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1452 MapVector<Value *, RRInfo> &Retains,
1453 DenseMap<Value *, RRInfo> &Releases,
1456 void OptimizeWeakCalls(Function &F);
1458 bool OptimizeSequences(Function &F);
1460 void OptimizeReturns(Function &F);
1462 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1463 virtual bool doInitialization(Module &M);
1464 virtual bool runOnFunction(Function &F);
1465 virtual void releaseMemory();
1469 ObjCARCOpt() : FunctionPass(ID) {
1470 initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1475 char ObjCARCOpt::ID = 0;
1476 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1477 "objc-arc", "ObjC ARC optimization", false, false)
1478 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1479 INITIALIZE_PASS_END(ObjCARCOpt,
1480 "objc-arc", "ObjC ARC optimization", false, false)
1482 Pass *llvm::createObjCARCOptPass() {
1483 return new ObjCARCOpt();
1486 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1487 AU.addRequired<ObjCARCAliasAnalysis>();
1488 AU.addRequired<AliasAnalysis>();
1489 // ARC optimization doesn't currently split critical edges.
1490 AU.setPreservesCFG();
1493 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1494 if (!RetainRVCallee) {
1495 LLVMContext &C = M->getContext();
1496 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1497 std::vector<Type *> Params;
1498 Params.push_back(I8X);
1500 FunctionType::get(I8X, Params, /*isVarArg=*/false);
1501 AttrListPtr Attributes;
1502 Attributes.addAttr(~0u, Attribute::NoUnwind);
1504 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1507 return RetainRVCallee;
1510 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1511 if (!AutoreleaseRVCallee) {
1512 LLVMContext &C = M->getContext();
1513 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1514 std::vector<Type *> Params;
1515 Params.push_back(I8X);
1517 FunctionType::get(I8X, Params, /*isVarArg=*/false);
1518 AttrListPtr Attributes;
1519 Attributes.addAttr(~0u, Attribute::NoUnwind);
1520 AutoreleaseRVCallee =
1521 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1524 return AutoreleaseRVCallee;
1527 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1528 if (!ReleaseCallee) {
1529 LLVMContext &C = M->getContext();
1530 std::vector<Type *> Params;
1531 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1532 AttrListPtr Attributes;
1533 Attributes.addAttr(~0u, Attribute::NoUnwind);
1535 M->getOrInsertFunction(
1537 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1540 return ReleaseCallee;
1543 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1544 if (!RetainCallee) {
1545 LLVMContext &C = M->getContext();
1546 std::vector<Type *> Params;
1547 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1548 AttrListPtr Attributes;
1549 Attributes.addAttr(~0u, Attribute::NoUnwind);
1551 M->getOrInsertFunction(
1553 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1556 return RetainCallee;
1559 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1560 if (!RetainBlockCallee) {
1561 LLVMContext &C = M->getContext();
1562 std::vector<Type *> Params;
1563 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1564 AttrListPtr Attributes;
1565 Attributes.addAttr(~0u, Attribute::NoUnwind);
1567 M->getOrInsertFunction(
1569 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1572 return RetainBlockCallee;
1575 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1576 if (!AutoreleaseCallee) {
1577 LLVMContext &C = M->getContext();
1578 std::vector<Type *> Params;
1579 Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1580 AttrListPtr Attributes;
1581 Attributes.addAttr(~0u, Attribute::NoUnwind);
1583 M->getOrInsertFunction(
1585 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1588 return AutoreleaseCallee;
1591 /// CanAlterRefCount - Test whether the given instruction can result in a
1592 /// reference count modification (positive or negative) for the pointer's
1595 CanAlterRefCount(const Instruction *Inst, const Value *Ptr,
1596 ProvenanceAnalysis &PA, InstructionClass Class) {
1598 case IC_Autorelease:
1599 case IC_AutoreleaseRV:
1601 // These operations never directly modify a reference count.
1606 ImmutableCallSite CS = static_cast<const Value *>(Inst);
1607 assert(CS && "Only calls can alter reference counts!");
1609 // See if AliasAnalysis can help us with the call.
1610 AliasAnalysis::ModRefBehavior MRB = PA.getAA()->getModRefBehavior(CS);
1611 if (AliasAnalysis::onlyReadsMemory(MRB))
1613 if (AliasAnalysis::onlyAccessesArgPointees(MRB)) {
1614 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1616 const Value *Op = *I;
1617 if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1623 // Assume the worst.
1627 /// CanUse - Test whether the given instruction can "use" the given pointer's
1628 /// object in a way that requires the reference count to be positive.
1630 CanUse(const Instruction *Inst, const Value *Ptr, ProvenanceAnalysis &PA,
1631 InstructionClass Class) {
1632 // IC_Call operations (as opposed to IC_CallOrUser) never "use" objc pointers.
1633 if (Class == IC_Call)
1636 // Consider various instructions which may have pointer arguments which are
1638 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(Inst)) {
1639 // Comparing a pointer with null, or any other constant, isn't really a use,
1640 // because we don't care what the pointer points to, or about the values
1641 // of any other dynamic reference-counted pointers.
1642 if (!IsPotentialUse(ICI->getOperand(1)))
1644 } else if (ImmutableCallSite CS = static_cast<const Value *>(Inst)) {
1645 // For calls, just check the arguments (and not the callee operand).
1646 for (ImmutableCallSite::arg_iterator OI = CS.arg_begin(),
1647 OE = CS.arg_end(); OI != OE; ++OI) {
1648 const Value *Op = *OI;
1649 if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1653 } else if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1654 // Special-case stores, because we don't care about the stored value, just
1655 // the store address.
1656 const Value *Op = GetUnderlyingObjCPtr(SI->getPointerOperand());
1657 // If we can't tell what the underlying object was, assume there is a
1659 return IsPotentialUse(Op) && PA.related(Op, Ptr);
1662 // Check each operand for a match.
1663 for (User::const_op_iterator OI = Inst->op_begin(), OE = Inst->op_end();
1665 const Value *Op = *OI;
1666 if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1672 /// CanInterruptRV - Test whether the given instruction can autorelease
1673 /// any pointer or cause an autoreleasepool pop.
1675 CanInterruptRV(InstructionClass Class) {
1677 case IC_AutoreleasepoolPop:
1680 case IC_Autorelease:
1681 case IC_AutoreleaseRV:
1682 case IC_FusedRetainAutorelease:
1683 case IC_FusedRetainAutoreleaseRV:
1691 /// DependenceKind - There are several kinds of dependence-like concepts in
1693 enum DependenceKind {
1694 NeedsPositiveRetainCount,
1695 CanChangeRetainCount,
1696 RetainAutoreleaseDep, ///< Blocks objc_retainAutorelease.
1697 RetainAutoreleaseRVDep, ///< Blocks objc_retainAutoreleaseReturnValue.
1698 RetainRVDep ///< Blocks objc_retainAutoreleasedReturnValue.
1702 /// Depends - Test if there can be dependencies on Inst through Arg. This
1703 /// function only tests dependencies relevant for removing pairs of calls.
1705 Depends(DependenceKind Flavor, Instruction *Inst, const Value *Arg,
1706 ProvenanceAnalysis &PA) {
1707 // If we've reached the definition of Arg, stop.
1712 case NeedsPositiveRetainCount: {
1713 InstructionClass Class = GetInstructionClass(Inst);
1715 case IC_AutoreleasepoolPop:
1716 case IC_AutoreleasepoolPush:
1720 return CanUse(Inst, Arg, PA, Class);
1724 case CanChangeRetainCount: {
1725 InstructionClass Class = GetInstructionClass(Inst);
1727 case IC_AutoreleasepoolPop:
1728 // Conservatively assume this can decrement any count.
1730 case IC_AutoreleasepoolPush:
1734 return CanAlterRefCount(Inst, Arg, PA, Class);
1738 case RetainAutoreleaseDep:
1739 switch (GetBasicInstructionClass(Inst)) {
1740 case IC_AutoreleasepoolPop:
1741 // Don't merge an objc_autorelease with an objc_retain inside a different
1742 // autoreleasepool scope.
1746 // Check for a retain of the same pointer for merging.
1747 return GetObjCArg(Inst) == Arg;
1749 // Nothing else matters for objc_retainAutorelease formation.
1754 case RetainAutoreleaseRVDep: {
1755 InstructionClass Class = GetBasicInstructionClass(Inst);
1759 // Check for a retain of the same pointer for merging.
1760 return GetObjCArg(Inst) == Arg;
1762 // Anything that can autorelease interrupts
1763 // retainAutoreleaseReturnValue formation.
1764 return CanInterruptRV(Class);
1770 return CanInterruptRV(GetBasicInstructionClass(Inst));
1773 llvm_unreachable("Invalid dependence flavor");
1777 /// FindDependencies - Walk up the CFG from StartPos (which is in StartBB) and
1778 /// find local and non-local dependencies on Arg.
1779 /// TODO: Cache results?
1781 FindDependencies(DependenceKind Flavor,
1783 BasicBlock *StartBB, Instruction *StartInst,
1784 SmallPtrSet<Instruction *, 4> &DependingInstructions,
1785 SmallPtrSet<const BasicBlock *, 4> &Visited,
1786 ProvenanceAnalysis &PA) {
1787 BasicBlock::iterator StartPos = StartInst;
1789 SmallVector<std::pair<BasicBlock *, BasicBlock::iterator>, 4> Worklist;
1790 Worklist.push_back(std::make_pair(StartBB, StartPos));
1792 std::pair<BasicBlock *, BasicBlock::iterator> Pair =
1793 Worklist.pop_back_val();
1794 BasicBlock *LocalStartBB = Pair.first;
1795 BasicBlock::iterator LocalStartPos = Pair.second;
1796 BasicBlock::iterator StartBBBegin = LocalStartBB->begin();
1798 if (LocalStartPos == StartBBBegin) {
1799 pred_iterator PI(LocalStartBB), PE(LocalStartBB, false);
1801 // If we've reached the function entry, produce a null dependence.
1802 DependingInstructions.insert(0);
1804 // Add the predecessors to the worklist.
1806 BasicBlock *PredBB = *PI;
1807 if (Visited.insert(PredBB))
1808 Worklist.push_back(std::make_pair(PredBB, PredBB->end()));
1809 } while (++PI != PE);
1813 Instruction *Inst = --LocalStartPos;
1814 if (Depends(Flavor, Inst, Arg, PA)) {
1815 DependingInstructions.insert(Inst);
1819 } while (!Worklist.empty());
1821 // Determine whether the original StartBB post-dominates all of the blocks we
1822 // visited. If not, insert a sentinal indicating that most optimizations are
1824 for (SmallPtrSet<const BasicBlock *, 4>::const_iterator I = Visited.begin(),
1825 E = Visited.end(); I != E; ++I) {
1826 const BasicBlock *BB = *I;
1829 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1830 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
1831 const BasicBlock *Succ = *SI;
1832 if (Succ != StartBB && !Visited.count(Succ)) {
1833 DependingInstructions.insert(reinterpret_cast<Instruction *>(-1));
1840 static bool isNullOrUndef(const Value *V) {
1841 return isa<ConstantPointerNull>(V) || isa<UndefValue>(V);
1844 static bool isNoopInstruction(const Instruction *I) {
1845 return isa<BitCastInst>(I) ||
1846 (isa<GetElementPtrInst>(I) &&
1847 cast<GetElementPtrInst>(I)->hasAllZeroIndices());
1850 /// OptimizeRetainCall - Turn objc_retain into
1851 /// objc_retainAutoreleasedReturnValue if the operand is a return value.
1853 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
1854 CallSite CS(GetObjCArg(Retain));
1855 Instruction *Call = CS.getInstruction();
1857 if (Call->getParent() != Retain->getParent()) return;
1859 // Check that the call is next to the retain.
1860 BasicBlock::iterator I = Call;
1862 while (isNoopInstruction(I)) ++I;
1866 // Turn it to an objc_retainAutoreleasedReturnValue..
1869 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
1872 /// OptimizeRetainRVCall - Turn objc_retainAutoreleasedReturnValue into
1873 /// objc_retain if the operand is not a return value. Or, if it can be
1874 /// paired with an objc_autoreleaseReturnValue, delete the pair and
1877 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
1878 // Check for the argument being from an immediately preceding call.
1879 Value *Arg = GetObjCArg(RetainRV);
1881 if (Instruction *Call = CS.getInstruction())
1882 if (Call->getParent() == RetainRV->getParent()) {
1883 BasicBlock::iterator I = Call;
1885 while (isNoopInstruction(I)) ++I;
1886 if (&*I == RetainRV)
1890 // Check for being preceded by an objc_autoreleaseReturnValue on the same
1891 // pointer. In this case, we can delete the pair.
1892 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
1894 do --I; while (I != Begin && isNoopInstruction(I));
1895 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
1896 GetObjCArg(I) == Arg) {
1899 EraseInstruction(I);
1900 EraseInstruction(RetainRV);
1905 // Turn it to a plain objc_retain.
1908 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
1912 /// OptimizeAutoreleaseRVCall - Turn objc_autoreleaseReturnValue into
1913 /// objc_autorelease if the result is not used as a return value.
1915 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV) {
1916 // Check for a return of the pointer value.
1917 const Value *Ptr = GetObjCArg(AutoreleaseRV);
1918 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
1920 const User *I = *UI;
1921 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
1927 cast<CallInst>(AutoreleaseRV)->
1928 setCalledFunction(getAutoreleaseCallee(F.getParent()));
1931 /// OptimizeIndividualCalls - Visit each call, one at a time, and make
1932 /// simplifications without doing any additional analysis.
1933 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
1934 // Reset all the flags in preparation for recomputing them.
1935 UsedInThisFunction = 0;
1937 // Visit all objc_* calls in F.
1938 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
1939 Instruction *Inst = &*I++;
1940 InstructionClass Class = GetBasicInstructionClass(Inst);
1945 // Delete no-op casts. These function calls have special semantics, but
1946 // the semantics are entirely implemented via lowering in the front-end,
1947 // so by the time they reach the optimizer, they are just no-op calls
1948 // which return their argument.
1950 // There are gray areas here, as the ability to cast reference-counted
1951 // pointers to raw void* and back allows code to break ARC assumptions,
1952 // however these are currently considered to be unimportant.
1956 EraseInstruction(Inst);
1959 // If the pointer-to-weak-pointer is null, it's undefined behavior.
1962 case IC_LoadWeakRetained:
1964 case IC_DestroyWeak: {
1965 CallInst *CI = cast<CallInst>(Inst);
1966 if (isNullOrUndef(CI->getArgOperand(0))) {
1967 Type *Ty = CI->getArgOperand(0)->getType();
1968 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1969 Constant::getNullValue(Ty),
1971 CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
1972 CI->eraseFromParent();
1979 CallInst *CI = cast<CallInst>(Inst);
1980 if (isNullOrUndef(CI->getArgOperand(0)) ||
1981 isNullOrUndef(CI->getArgOperand(1))) {
1982 Type *Ty = CI->getArgOperand(0)->getType();
1983 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1984 Constant::getNullValue(Ty),
1986 CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
1987 CI->eraseFromParent();
1993 OptimizeRetainCall(F, Inst);
1996 if (OptimizeRetainRVCall(F, Inst))
1999 case IC_AutoreleaseRV:
2000 OptimizeAutoreleaseRVCall(F, Inst);
2004 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
2005 if (IsAutorelease(Class) && Inst->use_empty()) {
2006 CallInst *Call = cast<CallInst>(Inst);
2007 const Value *Arg = Call->getArgOperand(0);
2008 Arg = FindSingleUseIdentifiedObject(Arg);
2013 // Create the declaration lazily.
2014 LLVMContext &C = Inst->getContext();
2016 CallInst::Create(getReleaseCallee(F.getParent()),
2017 Call->getArgOperand(0), "", Call);
2018 NewCall->setMetadata(ImpreciseReleaseMDKind,
2019 MDNode::get(C, ArrayRef<Value *>()));
2020 EraseInstruction(Call);
2026 // For functions which can never be passed stack arguments, add
2028 if (IsAlwaysTail(Class)) {
2030 cast<CallInst>(Inst)->setTailCall();
2033 // Set nounwind as needed.
2034 if (IsNoThrow(Class)) {
2036 cast<CallInst>(Inst)->setDoesNotThrow();
2039 if (!IsNoopOnNull(Class)) {
2040 UsedInThisFunction |= 1 << Class;
2044 const Value *Arg = GetObjCArg(Inst);
2046 // ARC calls with null are no-ops. Delete them.
2047 if (isNullOrUndef(Arg)) {
2050 EraseInstruction(Inst);
2054 // Keep track of which of retain, release, autorelease, and retain_block
2055 // are actually present in this function.
2056 UsedInThisFunction |= 1 << Class;
2058 // If Arg is a PHI, and one or more incoming values to the
2059 // PHI are null, and the call is control-equivalent to the PHI, and there
2060 // are no relevant side effects between the PHI and the call, the call
2061 // could be pushed up to just those paths with non-null incoming values.
2062 // For now, don't bother splitting critical edges for this.
2063 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
2064 Worklist.push_back(std::make_pair(Inst, Arg));
2066 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
2070 const PHINode *PN = dyn_cast<PHINode>(Arg);
2073 // Determine if the PHI has any null operands, or any incoming
2075 bool HasNull = false;
2076 bool HasCriticalEdges = false;
2077 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2079 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2080 if (isNullOrUndef(Incoming))
2082 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
2083 .getNumSuccessors() != 1) {
2084 HasCriticalEdges = true;
2088 // If we have null operands and no critical edges, optimize.
2089 if (!HasCriticalEdges && HasNull) {
2090 SmallPtrSet<Instruction *, 4> DependingInstructions;
2091 SmallPtrSet<const BasicBlock *, 4> Visited;
2093 // Check that there is nothing that cares about the reference
2094 // count between the call and the phi.
2095 FindDependencies(NeedsPositiveRetainCount, Arg,
2096 Inst->getParent(), Inst,
2097 DependingInstructions, Visited, PA);
2098 if (DependingInstructions.size() == 1 &&
2099 *DependingInstructions.begin() == PN) {
2102 // Clone the call into each predecessor that has a non-null value.
2103 CallInst *CInst = cast<CallInst>(Inst);
2104 Type *ParamTy = CInst->getArgOperand(0)->getType();
2105 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2107 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2108 if (!isNullOrUndef(Incoming)) {
2109 CallInst *Clone = cast<CallInst>(CInst->clone());
2110 Value *Op = PN->getIncomingValue(i);
2111 Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
2112 if (Op->getType() != ParamTy)
2113 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
2114 Clone->setArgOperand(0, Op);
2115 Clone->insertBefore(InsertPos);
2116 Worklist.push_back(std::make_pair(Clone, Incoming));
2119 // Erase the original call.
2120 EraseInstruction(CInst);
2124 } while (!Worklist.empty());
2128 /// CheckForCFGHazards - Check for critical edges, loop boundaries, irreducible
2129 /// control flow, or other CFG structures where moving code across the edge
2130 /// would result in it being executed more.
2132 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
2133 DenseMap<const BasicBlock *, BBState> &BBStates,
2134 BBState &MyStates) const {
2135 // If any top-down local-use or possible-dec has a succ which is earlier in
2136 // the sequence, forget it.
2137 for (BBState::ptr_const_iterator I = MyStates.top_down_ptr_begin(),
2138 E = MyStates.top_down_ptr_end(); I != E; ++I)
2139 switch (I->second.GetSeq()) {
2142 const Value *Arg = I->first;
2143 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2144 bool SomeSuccHasSame = false;
2145 bool AllSuccsHaveSame = true;
2146 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI)
2147 switch (BBStates[*SI].getPtrBottomUpState(Arg).GetSeq()) {
2150 MyStates.getPtrTopDownState(Arg).ClearSequenceProgress();
2151 SomeSuccHasSame = false;
2154 SomeSuccHasSame = true;
2158 case S_MovableRelease:
2159 AllSuccsHaveSame = false;
2162 llvm_unreachable("bottom-up pointer in retain state!");
2164 // If the state at the other end of any of the successor edges
2165 // matches the current state, require all edges to match. This
2166 // guards against loops in the middle of a sequence.
2167 if (SomeSuccHasSame && !AllSuccsHaveSame)
2168 MyStates.getPtrTopDownState(Arg).ClearSequenceProgress();
2170 case S_CanRelease: {
2171 const Value *Arg = I->first;
2172 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2173 bool SomeSuccHasSame = false;
2174 bool AllSuccsHaveSame = true;
2175 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI)
2176 switch (BBStates[*SI].getPtrBottomUpState(Arg).GetSeq()) {
2178 MyStates.getPtrTopDownState(Arg).ClearSequenceProgress();
2179 SomeSuccHasSame = false;
2182 SomeSuccHasSame = true;
2186 case S_MovableRelease:
2188 AllSuccsHaveSame = false;
2191 llvm_unreachable("bottom-up pointer in retain state!");
2193 // If the state at the other end of any of the successor edges
2194 // matches the current state, require all edges to match. This
2195 // guards against loops in the middle of a sequence.
2196 if (SomeSuccHasSame && !AllSuccsHaveSame)
2197 MyStates.getPtrTopDownState(Arg).ClearSequenceProgress();
2203 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
2204 DenseMap<const BasicBlock *, BBState> &BBStates,
2205 MapVector<Value *, RRInfo> &Retains) {
2206 bool NestingDetected = false;
2207 BBState &MyStates = BBStates[BB];
2209 // Merge the states from each successor to compute the initial state
2210 // for the current block.
2211 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2212 succ_const_iterator SI(TI), SE(TI, false);
2214 MyStates.SetAsExit();
2217 const BasicBlock *Succ = *SI++;
2220 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
2221 if (I == BBStates.end())
2223 MyStates.InitFromSucc(I->second);
2227 I = BBStates.find(Succ);
2228 if (I != BBStates.end())
2229 MyStates.MergeSucc(I->second);
2235 // Visit all the instructions, bottom-up.
2236 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
2237 Instruction *Inst = llvm::prior(I);
2238 InstructionClass Class = GetInstructionClass(Inst);
2239 const Value *Arg = 0;
2243 Arg = GetObjCArg(Inst);
2245 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2247 // If we see two releases in a row on the same pointer. If so, make
2248 // a note, and we'll cicle back to revisit it after we've
2249 // hopefully eliminated the second release, which may allow us to
2250 // eliminate the first release too.
2251 // Theoretically we could implement removal of nested retain+release
2252 // pairs by making PtrState hold a stack of states, but this is
2253 // simple and avoids adding overhead for the non-nested case.
2254 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease)
2255 NestingDetected = true;
2257 S.SetSeqToRelease(Inst->getMetadata(ImpreciseReleaseMDKind));
2259 S.RRI.KnownIncremented = S.IsKnownIncremented();
2260 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2261 S.RRI.Calls.insert(Inst);
2263 S.IncrementRefCount();
2266 case IC_RetainBlock:
2269 Arg = GetObjCArg(Inst);
2271 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2272 S.DecrementRefCount();
2274 switch (S.GetSeq()) {
2277 case S_MovableRelease:
2279 S.RRI.ReverseInsertPts.clear();
2282 // Don't do retain+release tracking for IC_RetainRV, because it's
2283 // better to let it remain as the first instruction after a call.
2284 if (Class != IC_RetainRV) {
2285 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2286 Retains[Inst] = S.RRI;
2288 S.ClearSequenceProgress();
2293 llvm_unreachable("bottom-up pointer in retain state!");
2297 case IC_AutoreleasepoolPop:
2298 // Conservatively, clear MyStates for all known pointers.
2299 MyStates.clearBottomUpPointers();
2301 case IC_AutoreleasepoolPush:
2303 // These are irrelevant.
2309 // Consider any other possible effects of this instruction on each
2310 // pointer being tracked.
2311 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
2312 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
2313 const Value *Ptr = MI->first;
2315 continue; // Handled above.
2316 PtrState &S = MI->second;
2317 Sequence Seq = S.GetSeq();
2319 // Check for possible retains and releases.
2320 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2321 // Check for a retain (we're going bottom-up here).
2322 S.DecrementRefCount();
2324 // Check for a release.
2325 if (!IsRetain(Class) && Class != IC_RetainBlock)
2328 S.SetSeq(S_CanRelease);
2332 case S_MovableRelease:
2337 llvm_unreachable("bottom-up pointer in retain state!");
2341 // Check for possible direct uses.
2344 case S_MovableRelease:
2345 if (CanUse(Inst, Ptr, PA, Class)) {
2346 S.RRI.ReverseInsertPts.clear();
2347 S.RRI.ReverseInsertPts.insert(Inst);
2349 } else if (Seq == S_Release &&
2350 (Class == IC_User || Class == IC_CallOrUser)) {
2351 // Non-movable releases depend on any possible objc pointer use.
2353 S.RRI.ReverseInsertPts.clear();
2354 S.RRI.ReverseInsertPts.insert(Inst);
2358 if (CanUse(Inst, Ptr, PA, Class))
2366 llvm_unreachable("bottom-up pointer in retain state!");
2371 return NestingDetected;
2375 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
2376 DenseMap<const BasicBlock *, BBState> &BBStates,
2377 DenseMap<Value *, RRInfo> &Releases) {
2378 bool NestingDetected = false;
2379 BBState &MyStates = BBStates[BB];
2381 // Merge the states from each predecessor to compute the initial state
2382 // for the current block.
2383 const_pred_iterator PI(BB), PE(BB, false);
2385 MyStates.SetAsEntry();
2388 const BasicBlock *Pred = *PI++;
2391 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
2392 if (I == BBStates.end())
2394 MyStates.InitFromPred(I->second);
2398 I = BBStates.find(Pred);
2399 if (I != BBStates.end())
2400 MyStates.MergePred(I->second);
2406 // Visit all the instructions, top-down.
2407 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2408 Instruction *Inst = I;
2409 InstructionClass Class = GetInstructionClass(Inst);
2410 const Value *Arg = 0;
2413 case IC_RetainBlock:
2416 Arg = GetObjCArg(Inst);
2418 PtrState &S = MyStates.getPtrTopDownState(Arg);
2420 // Don't do retain+release tracking for IC_RetainRV, because it's
2421 // better to let it remain as the first instruction after a call.
2422 if (Class != IC_RetainRV) {
2423 // If we see two retains in a row on the same pointer. If so, make
2424 // a note, and we'll cicle back to revisit it after we've
2425 // hopefully eliminated the second retain, which may allow us to
2426 // eliminate the first retain too.
2427 // Theoretically we could implement removal of nested retain+release
2428 // pairs by making PtrState hold a stack of states, but this is
2429 // simple and avoids adding overhead for the non-nested case.
2430 if (S.GetSeq() == S_Retain)
2431 NestingDetected = true;
2435 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2436 S.RRI.KnownIncremented = S.IsKnownIncremented();
2437 S.RRI.Calls.insert(Inst);
2440 S.IncrementRefCount();
2444 Arg = GetObjCArg(Inst);
2446 PtrState &S = MyStates.getPtrTopDownState(Arg);
2447 S.DecrementRefCount();
2449 switch (S.GetSeq()) {
2452 S.RRI.ReverseInsertPts.clear();
2455 S.RRI.ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2456 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2457 Releases[Inst] = S.RRI;
2458 S.ClearSequenceProgress();
2464 case S_MovableRelease:
2465 llvm_unreachable("top-down pointer in release state!");
2469 case IC_AutoreleasepoolPop:
2470 // Conservatively, clear MyStates for all known pointers.
2471 MyStates.clearTopDownPointers();
2473 case IC_AutoreleasepoolPush:
2475 // These are irrelevant.
2481 // Consider any other possible effects of this instruction on each
2482 // pointer being tracked.
2483 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
2484 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
2485 const Value *Ptr = MI->first;
2487 continue; // Handled above.
2488 PtrState &S = MI->second;
2489 Sequence Seq = S.GetSeq();
2491 // Check for possible releases.
2492 if (!IsRetain(Class) && Class != IC_RetainBlock &&
2493 CanAlterRefCount(Inst, Ptr, PA, Class)) {
2494 // Check for a release.
2495 S.DecrementRefCount();
2497 // Check for a release.
2500 S.SetSeq(S_CanRelease);
2501 S.RRI.ReverseInsertPts.clear();
2502 S.RRI.ReverseInsertPts.insert(Inst);
2504 // One call can't cause a transition from S_Retain to S_CanRelease
2505 // and S_CanRelease to S_Use. If we've made the first transition,
2514 case S_MovableRelease:
2515 llvm_unreachable("top-down pointer in release state!");
2519 // Check for possible direct uses.
2522 if (CanUse(Inst, Ptr, PA, Class))
2531 case S_MovableRelease:
2532 llvm_unreachable("top-down pointer in release state!");
2537 CheckForCFGHazards(BB, BBStates, MyStates);
2538 return NestingDetected;
2541 // Visit - Visit the function both top-down and bottom-up.
2543 ObjCARCOpt::Visit(Function &F,
2544 DenseMap<const BasicBlock *, BBState> &BBStates,
2545 MapVector<Value *, RRInfo> &Retains,
2546 DenseMap<Value *, RRInfo> &Releases) {
2547 // Use postorder for bottom-up, and reverse-postorder for top-down, because we
2548 // magically know that loops will be well behaved, i.e. they won't repeatedly
2549 // call retain on a single pointer without doing a release.
2550 bool BottomUpNestingDetected = false;
2551 SmallVector<BasicBlock *, 8> PostOrder;
2552 for (po_iterator<Function *> I = po_begin(&F), E = po_end(&F); I != E; ++I) {
2553 BasicBlock *BB = *I;
2554 PostOrder.push_back(BB);
2556 BottomUpNestingDetected |= VisitBottomUp(BB, BBStates, Retains);
2559 // Iterate through the post-order in reverse order, achieving a
2560 // reverse-postorder traversal. We don't use the ReversePostOrderTraversal
2561 // class here because it works by computing its own full postorder iteration,
2562 // recording the sequence, and playing it back in reverse. Since we're already
2563 // doing a full iteration above, we can just record the sequence manually and
2564 // avoid the cost of having ReversePostOrderTraversal compute it.
2565 bool TopDownNestingDetected = false;
2566 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator
2567 RI = PostOrder.rbegin(), RE = PostOrder.rend(); RI != RE; ++RI)
2568 TopDownNestingDetected |= VisitTopDown(*RI, BBStates, Releases);
2570 return TopDownNestingDetected && BottomUpNestingDetected;
2573 /// MoveCalls - Move the calls in RetainsToMove and ReleasesToMove.
2574 void ObjCARCOpt::MoveCalls(Value *Arg,
2575 RRInfo &RetainsToMove,
2576 RRInfo &ReleasesToMove,
2577 MapVector<Value *, RRInfo> &Retains,
2578 DenseMap<Value *, RRInfo> &Releases,
2579 SmallVectorImpl<Instruction *> &DeadInsts,
2581 Type *ArgTy = Arg->getType();
2582 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
2584 // Insert the new retain and release calls.
2585 for (SmallPtrSet<Instruction *, 2>::const_iterator
2586 PI = ReleasesToMove.ReverseInsertPts.begin(),
2587 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2588 Instruction *InsertPt = *PI;
2589 Value *MyArg = ArgTy == ParamTy ? Arg :
2590 new BitCastInst(Arg, ParamTy, "", InsertPt);
2592 CallInst::Create(RetainsToMove.IsRetainBlock ?
2593 getRetainBlockCallee(M) : getRetainCallee(M),
2594 MyArg, "", InsertPt);
2595 Call->setDoesNotThrow();
2596 if (!RetainsToMove.IsRetainBlock)
2597 Call->setTailCall();
2599 for (SmallPtrSet<Instruction *, 2>::const_iterator
2600 PI = RetainsToMove.ReverseInsertPts.begin(),
2601 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2602 Instruction *LastUse = *PI;
2603 Instruction *InsertPts[] = { 0, 0, 0 };
2604 if (InvokeInst *II = dyn_cast<InvokeInst>(LastUse)) {
2605 // We can't insert code immediately after an invoke instruction, so
2606 // insert code at the beginning of both successor blocks instead.
2607 // The invoke's return value isn't available in the unwind block,
2608 // but our releases will never depend on it, because they must be
2609 // paired with retains from before the invoke.
2610 InsertPts[0] = II->getNormalDest()->getFirstNonPHI();
2611 InsertPts[1] = II->getUnwindDest()->getFirstNonPHI();
2613 // Insert code immediately after the last use.
2614 InsertPts[0] = llvm::next(BasicBlock::iterator(LastUse));
2617 for (Instruction **I = InsertPts; *I; ++I) {
2618 Instruction *InsertPt = *I;
2619 Value *MyArg = ArgTy == ParamTy ? Arg :
2620 new BitCastInst(Arg, ParamTy, "", InsertPt);
2621 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
2623 // Attach a clang.imprecise_release metadata tag, if appropriate.
2624 if (MDNode *M = ReleasesToMove.ReleaseMetadata)
2625 Call->setMetadata(ImpreciseReleaseMDKind, M);
2626 Call->setDoesNotThrow();
2627 if (ReleasesToMove.IsTailCallRelease)
2628 Call->setTailCall();
2632 // Delete the original retain and release calls.
2633 for (SmallPtrSet<Instruction *, 2>::const_iterator
2634 AI = RetainsToMove.Calls.begin(),
2635 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
2636 Instruction *OrigRetain = *AI;
2637 Retains.blot(OrigRetain);
2638 DeadInsts.push_back(OrigRetain);
2640 for (SmallPtrSet<Instruction *, 2>::const_iterator
2641 AI = ReleasesToMove.Calls.begin(),
2642 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
2643 Instruction *OrigRelease = *AI;
2644 Releases.erase(OrigRelease);
2645 DeadInsts.push_back(OrigRelease);
2650 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
2652 MapVector<Value *, RRInfo> &Retains,
2653 DenseMap<Value *, RRInfo> &Releases,
2655 bool AnyPairsCompletelyEliminated = false;
2656 RRInfo RetainsToMove;
2657 RRInfo ReleasesToMove;
2658 SmallVector<Instruction *, 4> NewRetains;
2659 SmallVector<Instruction *, 4> NewReleases;
2660 SmallVector<Instruction *, 8> DeadInsts;
2662 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
2663 E = Retains.end(); I != E; ) {
2664 Value *V = (I++)->first;
2665 if (!V) continue; // blotted
2667 Instruction *Retain = cast<Instruction>(V);
2668 Value *Arg = GetObjCArg(Retain);
2670 // If the object being released is in static or stack storage, we know it's
2671 // not being managed by ObjC reference counting, so we can delete pairs
2672 // regardless of what possible decrements or uses lie between them.
2673 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg);
2675 // If a pair happens in a region where it is known that the reference count
2676 // is already incremented, we can similarly ignore possible decrements.
2677 bool KnownIncrementedTD = true, KnownIncrementedBU = true;
2679 // Connect the dots between the top-down-collected RetainsToMove and
2680 // bottom-up-collected ReleasesToMove to form sets of related calls.
2681 // This is an iterative process so that we connect multiple releases
2682 // to multiple retains if needed.
2683 unsigned OldDelta = 0;
2684 unsigned NewDelta = 0;
2685 unsigned OldCount = 0;
2686 unsigned NewCount = 0;
2687 bool FirstRelease = true;
2688 bool FirstRetain = true;
2689 NewRetains.push_back(Retain);
2691 for (SmallVectorImpl<Instruction *>::const_iterator
2692 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
2693 Instruction *NewRetain = *NI;
2694 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
2695 assert(It != Retains.end());
2696 const RRInfo &NewRetainRRI = It->second;
2697 KnownIncrementedTD &= NewRetainRRI.KnownIncremented;
2698 for (SmallPtrSet<Instruction *, 2>::const_iterator
2699 LI = NewRetainRRI.Calls.begin(),
2700 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
2701 Instruction *NewRetainRelease = *LI;
2702 DenseMap<Value *, RRInfo>::const_iterator Jt =
2703 Releases.find(NewRetainRelease);
2704 if (Jt == Releases.end())
2706 const RRInfo &NewRetainReleaseRRI = Jt->second;
2707 assert(NewRetainReleaseRRI.Calls.count(NewRetain));
2708 if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
2710 BBStates[NewRetainRelease->getParent()].GetAllPathCount();
2712 // Merge the ReleaseMetadata and IsTailCallRelease values.
2714 ReleasesToMove.ReleaseMetadata =
2715 NewRetainReleaseRRI.ReleaseMetadata;
2716 ReleasesToMove.IsTailCallRelease =
2717 NewRetainReleaseRRI.IsTailCallRelease;
2718 FirstRelease = false;
2720 if (ReleasesToMove.ReleaseMetadata !=
2721 NewRetainReleaseRRI.ReleaseMetadata)
2722 ReleasesToMove.ReleaseMetadata = 0;
2723 if (ReleasesToMove.IsTailCallRelease !=
2724 NewRetainReleaseRRI.IsTailCallRelease)
2725 ReleasesToMove.IsTailCallRelease = false;
2728 // Collect the optimal insertion points.
2730 for (SmallPtrSet<Instruction *, 2>::const_iterator
2731 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
2732 RE = NewRetainReleaseRRI.ReverseInsertPts.end();
2734 Instruction *RIP = *RI;
2735 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
2736 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
2738 NewReleases.push_back(NewRetainRelease);
2743 if (NewReleases.empty()) break;
2745 // Back the other way.
2746 for (SmallVectorImpl<Instruction *>::const_iterator
2747 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
2748 Instruction *NewRelease = *NI;
2749 DenseMap<Value *, RRInfo>::const_iterator It =
2750 Releases.find(NewRelease);
2751 assert(It != Releases.end());
2752 const RRInfo &NewReleaseRRI = It->second;
2753 KnownIncrementedBU &= NewReleaseRRI.KnownIncremented;
2754 for (SmallPtrSet<Instruction *, 2>::const_iterator
2755 LI = NewReleaseRRI.Calls.begin(),
2756 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
2757 Instruction *NewReleaseRetain = *LI;
2758 MapVector<Value *, RRInfo>::const_iterator Jt =
2759 Retains.find(NewReleaseRetain);
2760 if (Jt == Retains.end())
2762 const RRInfo &NewReleaseRetainRRI = Jt->second;
2763 assert(NewReleaseRetainRRI.Calls.count(NewRelease));
2764 if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
2765 unsigned PathCount =
2766 BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
2767 OldDelta += PathCount;
2768 OldCount += PathCount;
2770 // Merge the IsRetainBlock values.
2772 RetainsToMove.IsRetainBlock = NewReleaseRetainRRI.IsRetainBlock;
2773 FirstRetain = false;
2774 } else if (ReleasesToMove.IsRetainBlock !=
2775 NewReleaseRetainRRI.IsRetainBlock)
2776 // It's not possible to merge the sequences if one uses
2777 // objc_retain and the other uses objc_retainBlock.
2780 // Collect the optimal insertion points.
2782 for (SmallPtrSet<Instruction *, 2>::const_iterator
2783 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
2784 RE = NewReleaseRetainRRI.ReverseInsertPts.end();
2786 Instruction *RIP = *RI;
2787 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
2788 PathCount = BBStates[RIP->getParent()].GetAllPathCount();
2789 NewDelta += PathCount;
2790 NewCount += PathCount;
2793 NewRetains.push_back(NewReleaseRetain);
2797 NewReleases.clear();
2798 if (NewRetains.empty()) break;
2801 // If the pointer is known incremented, we can safely delete the pair
2802 // regardless of what's between them.
2803 if (KnownIncrementedTD || KnownIncrementedBU) {
2804 RetainsToMove.ReverseInsertPts.clear();
2805 ReleasesToMove.ReverseInsertPts.clear();
2809 // Determine whether the original call points are balanced in the retain and
2810 // release calls through the program. If not, conservatively don't touch
2812 // TODO: It's theoretically possible to do code motion in this case, as
2813 // long as the existing imbalances are maintained.
2817 // Determine whether the new insertion points we computed preserve the
2818 // balance of retain and release calls through the program.
2819 // TODO: If the fully aggressive solution isn't valid, try to find a
2820 // less aggressive solution which is.
2824 // Ok, everything checks out and we're all set. Let's move some code!
2826 AnyPairsCompletelyEliminated = NewCount == 0;
2827 NumRRs += OldCount - NewCount;
2828 MoveCalls(Arg, RetainsToMove, ReleasesToMove,
2829 Retains, Releases, DeadInsts, M);
2832 NewReleases.clear();
2834 RetainsToMove.clear();
2835 ReleasesToMove.clear();
2838 // Now that we're done moving everything, we can delete the newly dead
2839 // instructions, as we no longer need them as insert points.
2840 while (!DeadInsts.empty())
2841 EraseInstruction(DeadInsts.pop_back_val());
2843 return AnyPairsCompletelyEliminated;
2846 /// OptimizeWeakCalls - Weak pointer optimizations.
2847 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
2848 // First, do memdep-style RLE and S2L optimizations. We can't use memdep
2849 // itself because it uses AliasAnalysis and we need to do provenance
2851 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2852 Instruction *Inst = &*I++;
2853 InstructionClass Class = GetBasicInstructionClass(Inst);
2854 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
2857 // Delete objc_loadWeak calls with no users.
2858 if (Class == IC_LoadWeak && Inst->use_empty()) {
2859 Inst->eraseFromParent();
2863 // TODO: For now, just look for an earlier available version of this value
2864 // within the same block. Theoretically, we could do memdep-style non-local
2865 // analysis too, but that would want caching. A better approach would be to
2866 // use the technique that EarlyCSE uses.
2867 inst_iterator Current = llvm::prior(I);
2868 BasicBlock *CurrentBB = Current.getBasicBlockIterator();
2869 for (BasicBlock::iterator B = CurrentBB->begin(),
2870 J = Current.getInstructionIterator();
2872 Instruction *EarlierInst = &*llvm::prior(J);
2873 InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
2874 switch (EarlierClass) {
2876 case IC_LoadWeakRetained: {
2877 // If this is loading from the same pointer, replace this load's value
2879 CallInst *Call = cast<CallInst>(Inst);
2880 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2881 Value *Arg = Call->getArgOperand(0);
2882 Value *EarlierArg = EarlierCall->getArgOperand(0);
2883 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2884 case AliasAnalysis::MustAlias:
2886 // If the load has a builtin retain, insert a plain retain for it.
2887 if (Class == IC_LoadWeakRetained) {
2889 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2893 // Zap the fully redundant load.
2894 Call->replaceAllUsesWith(EarlierCall);
2895 Call->eraseFromParent();
2897 case AliasAnalysis::MayAlias:
2898 case AliasAnalysis::PartialAlias:
2900 case AliasAnalysis::NoAlias:
2907 // If this is storing to the same pointer and has the same size etc.
2908 // replace this load's value with the stored value.
2909 CallInst *Call = cast<CallInst>(Inst);
2910 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2911 Value *Arg = Call->getArgOperand(0);
2912 Value *EarlierArg = EarlierCall->getArgOperand(0);
2913 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2914 case AliasAnalysis::MustAlias:
2916 // If the load has a builtin retain, insert a plain retain for it.
2917 if (Class == IC_LoadWeakRetained) {
2919 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2923 // Zap the fully redundant load.
2924 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
2925 Call->eraseFromParent();
2927 case AliasAnalysis::MayAlias:
2928 case AliasAnalysis::PartialAlias:
2930 case AliasAnalysis::NoAlias:
2937 // TOOD: Grab the copied value.
2939 case IC_AutoreleasepoolPush:
2942 // Weak pointers are only modified through the weak entry points
2943 // (and arbitrary calls, which could call the weak entry points).
2946 // Anything else could modify the weak pointer.
2953 // Then, for each destroyWeak with an alloca operand, check to see if
2954 // the alloca and all its users can be zapped.
2955 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2956 Instruction *Inst = &*I++;
2957 InstructionClass Class = GetBasicInstructionClass(Inst);
2958 if (Class != IC_DestroyWeak)
2961 CallInst *Call = cast<CallInst>(Inst);
2962 Value *Arg = Call->getArgOperand(0);
2963 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
2964 for (Value::use_iterator UI = Alloca->use_begin(),
2965 UE = Alloca->use_end(); UI != UE; ++UI) {
2966 Instruction *UserInst = cast<Instruction>(*UI);
2967 switch (GetBasicInstructionClass(UserInst)) {
2970 case IC_DestroyWeak:
2977 for (Value::use_iterator UI = Alloca->use_begin(),
2978 UE = Alloca->use_end(); UI != UE; ) {
2979 CallInst *UserInst = cast<CallInst>(*UI++);
2980 if (!UserInst->use_empty())
2981 UserInst->replaceAllUsesWith(UserInst->getOperand(1));
2982 UserInst->eraseFromParent();
2984 Alloca->eraseFromParent();
2990 /// OptimizeSequences - Identify program paths which execute sequences of
2991 /// retains and releases which can be eliminated.
2992 bool ObjCARCOpt::OptimizeSequences(Function &F) {
2993 /// Releases, Retains - These are used to store the results of the main flow
2994 /// analysis. These use Value* as the key instead of Instruction* so that the
2995 /// map stays valid when we get around to rewriting code and calls get
2996 /// replaced by arguments.
2997 DenseMap<Value *, RRInfo> Releases;
2998 MapVector<Value *, RRInfo> Retains;
3000 /// BBStates, This is used during the traversal of the function to track the
3001 /// states for each identified object at each block.
3002 DenseMap<const BasicBlock *, BBState> BBStates;
3004 // Analyze the CFG of the function, and all instructions.
3005 bool NestingDetected = Visit(F, BBStates, Retains, Releases);
3008 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
3012 /// OptimizeReturns - Look for this pattern:
3014 /// %call = call i8* @something(...)
3015 /// %2 = call i8* @objc_retain(i8* %call)
3016 /// %3 = call i8* @objc_autorelease(i8* %2)
3019 /// And delete the retain and autorelease.
3021 /// Otherwise if it's just this:
3023 /// %3 = call i8* @objc_autorelease(i8* %2)
3026 /// convert the autorelease to autoreleaseRV.
3027 void ObjCARCOpt::OptimizeReturns(Function &F) {
3028 if (!F.getReturnType()->isPointerTy())
3031 SmallPtrSet<Instruction *, 4> DependingInstructions;
3032 SmallPtrSet<const BasicBlock *, 4> Visited;
3033 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
3034 BasicBlock *BB = FI;
3035 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
3038 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
3039 FindDependencies(NeedsPositiveRetainCount, Arg,
3040 BB, Ret, DependingInstructions, Visited, PA);
3041 if (DependingInstructions.size() != 1)
3045 CallInst *Autorelease =
3046 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3049 InstructionClass AutoreleaseClass =
3050 GetBasicInstructionClass(Autorelease);
3051 if (!IsAutorelease(AutoreleaseClass))
3053 if (GetObjCArg(Autorelease) != Arg)
3056 DependingInstructions.clear();
3059 // Check that there is nothing that can affect the reference
3060 // count between the autorelease and the retain.
3061 FindDependencies(CanChangeRetainCount, Arg,
3062 BB, Autorelease, DependingInstructions, Visited, PA);
3063 if (DependingInstructions.size() != 1)
3068 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3070 // Check that we found a retain with the same argument.
3072 !IsRetain(GetBasicInstructionClass(Retain)) ||
3073 GetObjCArg(Retain) != Arg)
3076 DependingInstructions.clear();
3079 // Convert the autorelease to an autoreleaseRV, since it's
3080 // returning the value.
3081 if (AutoreleaseClass == IC_Autorelease) {
3082 Autorelease->setCalledFunction(getAutoreleaseRVCallee(F.getParent()));
3083 AutoreleaseClass = IC_AutoreleaseRV;
3086 // Check that there is nothing that can affect the reference
3087 // count between the retain and the call.
3088 FindDependencies(CanChangeRetainCount, Arg, BB, Retain,
3089 DependingInstructions, Visited, PA);
3090 if (DependingInstructions.size() != 1)
3095 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3097 // Check that the pointer is the return value of the call.
3098 if (!Call || Arg != Call)
3101 // Check that the call is a regular call.
3102 InstructionClass Class = GetBasicInstructionClass(Call);
3103 if (Class != IC_CallOrUser && Class != IC_Call)
3106 // If so, we can zap the retain and autorelease.
3109 EraseInstruction(Retain);
3110 EraseInstruction(Autorelease);
3116 DependingInstructions.clear();
3121 bool ObjCARCOpt::doInitialization(Module &M) {
3125 Run = ModuleHasARC(M);
3129 // Identify the imprecise release metadata kind.
3130 ImpreciseReleaseMDKind =
3131 M.getContext().getMDKindID("clang.imprecise_release");
3133 // Intuitively, objc_retain and others are nocapture, however in practice
3134 // they are not, because they return their argument value. And objc_release
3135 // calls finalizers.
3137 // These are initialized lazily.
3139 AutoreleaseRVCallee = 0;
3142 RetainBlockCallee = 0;
3143 AutoreleaseCallee = 0;
3148 bool ObjCARCOpt::runOnFunction(Function &F) {
3152 // If nothing in the Module uses ARC, don't do anything.
3158 PA.setAA(&getAnalysis<AliasAnalysis>());
3160 // This pass performs several distinct transformations. As a compile-time aid
3161 // when compiling code that isn't ObjC, skip these if the relevant ObjC
3162 // library functions aren't declared.
3164 // Preliminary optimizations. This also computs UsedInThisFunction.
3165 OptimizeIndividualCalls(F);
3167 // Optimizations for weak pointers.
3168 if (UsedInThisFunction & ((1 << IC_LoadWeak) |
3169 (1 << IC_LoadWeakRetained) |
3170 (1 << IC_StoreWeak) |
3171 (1 << IC_InitWeak) |
3172 (1 << IC_CopyWeak) |
3173 (1 << IC_MoveWeak) |
3174 (1 << IC_DestroyWeak)))
3175 OptimizeWeakCalls(F);
3177 // Optimizations for retain+release pairs.
3178 if (UsedInThisFunction & ((1 << IC_Retain) |
3179 (1 << IC_RetainRV) |
3180 (1 << IC_RetainBlock)))
3181 if (UsedInThisFunction & (1 << IC_Release))
3182 // Run OptimizeSequences until it either stops making changes or
3183 // no retain+release pair nesting is detected.
3184 while (OptimizeSequences(F)) {}
3186 // Optimizations if objc_autorelease is used.
3187 if (UsedInThisFunction &
3188 ((1 << IC_Autorelease) | (1 << IC_AutoreleaseRV)))
3194 void ObjCARCOpt::releaseMemory() {
3198 //===----------------------------------------------------------------------===//
3200 //===----------------------------------------------------------------------===//
3202 // TODO: ObjCARCContract could insert PHI nodes when uses aren't
3203 // dominated by single calls.
3205 #include "llvm/Operator.h"
3206 #include "llvm/InlineAsm.h"
3207 #include "llvm/Analysis/Dominators.h"
3209 STATISTIC(NumStoreStrongs, "Number objc_storeStrong calls formed");
3212 /// ObjCARCContract - Late ARC optimizations. These change the IR in a way
3213 /// that makes it difficult to be analyzed by ObjCARCOpt, so it's run late.
3214 class ObjCARCContract : public FunctionPass {
3218 ProvenanceAnalysis PA;
3220 /// Run - A flag indicating whether this optimization pass should run.
3223 /// StoreStrongCallee, etc. - Declarations for ObjC runtime
3224 /// functions, for use in creating calls to them. These are initialized
3225 /// lazily to avoid cluttering up the Module with unused declarations.
3226 Constant *StoreStrongCallee,
3227 *RetainAutoreleaseCallee, *RetainAutoreleaseRVCallee;
3229 /// RetainRVMarker - The inline asm string to insert between calls and
3230 /// RetainRV calls to make the optimization work on targets which need it.
3231 const MDString *RetainRVMarker;
3233 Constant *getStoreStrongCallee(Module *M);
3234 Constant *getRetainAutoreleaseCallee(Module *M);
3235 Constant *getRetainAutoreleaseRVCallee(Module *M);
3237 bool ContractAutorelease(Function &F, Instruction *Autorelease,
3238 InstructionClass Class,
3239 SmallPtrSet<Instruction *, 4>
3240 &DependingInstructions,
3241 SmallPtrSet<const BasicBlock *, 4>
3244 void ContractRelease(Instruction *Release,
3245 inst_iterator &Iter);
3247 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
3248 virtual bool doInitialization(Module &M);
3249 virtual bool runOnFunction(Function &F);
3253 ObjCARCContract() : FunctionPass(ID) {
3254 initializeObjCARCContractPass(*PassRegistry::getPassRegistry());
3259 char ObjCARCContract::ID = 0;
3260 INITIALIZE_PASS_BEGIN(ObjCARCContract,
3261 "objc-arc-contract", "ObjC ARC contraction", false, false)
3262 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3263 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3264 INITIALIZE_PASS_END(ObjCARCContract,
3265 "objc-arc-contract", "ObjC ARC contraction", false, false)
3267 Pass *llvm::createObjCARCContractPass() {
3268 return new ObjCARCContract();
3271 void ObjCARCContract::getAnalysisUsage(AnalysisUsage &AU) const {
3272 AU.addRequired<AliasAnalysis>();
3273 AU.addRequired<DominatorTree>();
3274 AU.setPreservesCFG();
3277 Constant *ObjCARCContract::getStoreStrongCallee(Module *M) {
3278 if (!StoreStrongCallee) {
3279 LLVMContext &C = M->getContext();
3280 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3281 Type *I8XX = PointerType::getUnqual(I8X);
3282 std::vector<Type *> Params;
3283 Params.push_back(I8XX);
3284 Params.push_back(I8X);
3286 AttrListPtr Attributes;
3287 Attributes.addAttr(~0u, Attribute::NoUnwind);
3288 Attributes.addAttr(1, Attribute::NoCapture);
3291 M->getOrInsertFunction(
3293 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
3296 return StoreStrongCallee;
3299 Constant *ObjCARCContract::getRetainAutoreleaseCallee(Module *M) {
3300 if (!RetainAutoreleaseCallee) {
3301 LLVMContext &C = M->getContext();
3302 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3303 std::vector<Type *> Params;
3304 Params.push_back(I8X);
3306 FunctionType::get(I8X, Params, /*isVarArg=*/false);
3307 AttrListPtr Attributes;
3308 Attributes.addAttr(~0u, Attribute::NoUnwind);
3309 RetainAutoreleaseCallee =
3310 M->getOrInsertFunction("objc_retainAutorelease", FTy, Attributes);
3312 return RetainAutoreleaseCallee;
3315 Constant *ObjCARCContract::getRetainAutoreleaseRVCallee(Module *M) {
3316 if (!RetainAutoreleaseRVCallee) {
3317 LLVMContext &C = M->getContext();
3318 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3319 std::vector<Type *> Params;
3320 Params.push_back(I8X);
3322 FunctionType::get(I8X, Params, /*isVarArg=*/false);
3323 AttrListPtr Attributes;
3324 Attributes.addAttr(~0u, Attribute::NoUnwind);
3325 RetainAutoreleaseRVCallee =
3326 M->getOrInsertFunction("objc_retainAutoreleaseReturnValue", FTy,
3329 return RetainAutoreleaseRVCallee;
3332 /// ContractAutorelease - Merge an autorelease with a retain into a fused
3335 ObjCARCContract::ContractAutorelease(Function &F, Instruction *Autorelease,
3336 InstructionClass Class,
3337 SmallPtrSet<Instruction *, 4>
3338 &DependingInstructions,
3339 SmallPtrSet<const BasicBlock *, 4>
3341 const Value *Arg = GetObjCArg(Autorelease);
3343 // Check that there are no instructions between the retain and the autorelease
3344 // (such as an autorelease_pop) which may change the count.
3345 CallInst *Retain = 0;
3346 if (Class == IC_AutoreleaseRV)
3347 FindDependencies(RetainAutoreleaseRVDep, Arg,
3348 Autorelease->getParent(), Autorelease,
3349 DependingInstructions, Visited, PA);
3351 FindDependencies(RetainAutoreleaseDep, Arg,
3352 Autorelease->getParent(), Autorelease,
3353 DependingInstructions, Visited, PA);
3356 if (DependingInstructions.size() != 1) {
3357 DependingInstructions.clear();
3361 Retain = dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3362 DependingInstructions.clear();
3365 GetBasicInstructionClass(Retain) != IC_Retain ||
3366 GetObjCArg(Retain) != Arg)
3372 if (Class == IC_AutoreleaseRV)
3373 Retain->setCalledFunction(getRetainAutoreleaseRVCallee(F.getParent()));
3375 Retain->setCalledFunction(getRetainAutoreleaseCallee(F.getParent()));
3377 EraseInstruction(Autorelease);
3381 /// ContractRelease - Attempt to merge an objc_release with a store, load, and
3382 /// objc_retain to form an objc_storeStrong. This can be a little tricky because
3383 /// the instructions don't always appear in order, and there may be unrelated
3384 /// intervening instructions.
3385 void ObjCARCContract::ContractRelease(Instruction *Release,
3386 inst_iterator &Iter) {
3387 LoadInst *Load = dyn_cast<LoadInst>(GetObjCArg(Release));
3388 if (!Load || Load->isVolatile()) return;
3390 // For now, require everything to be in one basic block.
3391 BasicBlock *BB = Release->getParent();
3392 if (Load->getParent() != BB) return;
3394 // Walk down to find the store.
3395 BasicBlock::iterator I = Load, End = BB->end();
3397 AliasAnalysis::Location Loc = AA->getLocation(Load);
3400 IsRetain(GetBasicInstructionClass(I)) ||
3401 !(AA->getModRefInfo(I, Loc) & AliasAnalysis::Mod)))
3403 StoreInst *Store = dyn_cast<StoreInst>(I);
3404 if (!Store || Store->isVolatile()) return;
3405 if (Store->getPointerOperand() != Loc.Ptr) return;
3407 Value *New = StripPointerCastsAndObjCCalls(Store->getValueOperand());
3409 // Walk up to find the retain.
3411 BasicBlock::iterator Begin = BB->begin();
3412 while (I != Begin && GetBasicInstructionClass(I) != IC_Retain)
3414 Instruction *Retain = I;
3415 if (GetBasicInstructionClass(Retain) != IC_Retain) return;
3416 if (GetObjCArg(Retain) != New) return;
3421 LLVMContext &C = Release->getContext();
3422 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3423 Type *I8XX = PointerType::getUnqual(I8X);
3425 Value *Args[] = { Load->getPointerOperand(), New };
3426 if (Args[0]->getType() != I8XX)
3427 Args[0] = new BitCastInst(Args[0], I8XX, "", Store);
3428 if (Args[1]->getType() != I8X)
3429 Args[1] = new BitCastInst(Args[1], I8X, "", Store);
3430 CallInst *StoreStrong =
3431 CallInst::Create(getStoreStrongCallee(BB->getParent()->getParent()),
3433 StoreStrong->setDoesNotThrow();
3434 StoreStrong->setDebugLoc(Store->getDebugLoc());
3436 if (&*Iter == Store) ++Iter;
3437 Store->eraseFromParent();
3438 Release->eraseFromParent();
3439 EraseInstruction(Retain);
3440 if (Load->use_empty())
3441 Load->eraseFromParent();
3444 bool ObjCARCContract::doInitialization(Module &M) {
3445 Run = ModuleHasARC(M);
3449 // These are initialized lazily.
3450 StoreStrongCallee = 0;
3451 RetainAutoreleaseCallee = 0;
3452 RetainAutoreleaseRVCallee = 0;
3454 // Initialize RetainRVMarker.
3456 if (NamedMDNode *NMD =
3457 M.getNamedMetadata("clang.arc.retainAutoreleasedReturnValueMarker"))
3458 if (NMD->getNumOperands() == 1) {
3459 const MDNode *N = NMD->getOperand(0);
3460 if (N->getNumOperands() == 1)
3461 if (const MDString *S = dyn_cast<MDString>(N->getOperand(0)))
3468 bool ObjCARCContract::runOnFunction(Function &F) {
3472 // If nothing in the Module uses ARC, don't do anything.
3477 AA = &getAnalysis<AliasAnalysis>();
3478 DT = &getAnalysis<DominatorTree>();
3480 PA.setAA(&getAnalysis<AliasAnalysis>());
3482 // For ObjC library calls which return their argument, replace uses of the
3483 // argument with uses of the call return value, if it dominates the use. This
3484 // reduces register pressure.
3485 SmallPtrSet<Instruction *, 4> DependingInstructions;
3486 SmallPtrSet<const BasicBlock *, 4> Visited;
3487 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3488 Instruction *Inst = &*I++;
3490 // Only these library routines return their argument. In particular,
3491 // objc_retainBlock does not necessarily return its argument.
3492 InstructionClass Class = GetBasicInstructionClass(Inst);
3495 case IC_FusedRetainAutorelease:
3496 case IC_FusedRetainAutoreleaseRV:
3498 case IC_Autorelease:
3499 case IC_AutoreleaseRV:
3500 if (ContractAutorelease(F, Inst, Class, DependingInstructions, Visited))
3504 // If we're compiling for a target which needs a special inline-asm
3505 // marker to do the retainAutoreleasedReturnValue optimization,
3507 if (!RetainRVMarker)
3509 BasicBlock::iterator BBI = Inst;
3511 while (isNoopInstruction(BBI)) --BBI;
3512 if (&*BBI == GetObjCArg(Inst)) {
3514 InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
3515 /*isVarArg=*/false),
3516 RetainRVMarker->getString(),
3517 /*Constraints=*/"", /*hasSideEffects=*/true);
3518 CallInst::Create(IA, "", Inst);
3523 // objc_initWeak(p, null) => *p = null
3524 CallInst *CI = cast<CallInst>(Inst);
3525 if (isNullOrUndef(CI->getArgOperand(1))) {
3527 ConstantPointerNull::get(cast<PointerType>(CI->getType()));
3529 new StoreInst(Null, CI->getArgOperand(0), CI);
3530 CI->replaceAllUsesWith(Null);
3531 CI->eraseFromParent();
3536 ContractRelease(Inst, I);
3542 // Don't use GetObjCArg because we don't want to look through bitcasts
3543 // and such; to do the replacement, the argument must have type i8*.
3544 const Value *Arg = cast<CallInst>(Inst)->getArgOperand(0);
3546 // If we're compiling bugpointed code, don't get in trouble.
3547 if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))
3549 // Look through the uses of the pointer.
3550 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
3552 Use &U = UI.getUse();
3553 unsigned OperandNo = UI.getOperandNo();
3554 ++UI; // Increment UI now, because we may unlink its element.
3555 if (Instruction *UserInst = dyn_cast<Instruction>(U.getUser()))
3556 if (Inst != UserInst && DT->dominates(Inst, UserInst)) {
3558 Instruction *Replacement = Inst;
3559 Type *UseTy = U.get()->getType();
3560 if (PHINode *PHI = dyn_cast<PHINode>(UserInst)) {
3561 // For PHI nodes, insert the bitcast in the predecessor block.
3563 PHINode::getIncomingValueNumForOperand(OperandNo);
3565 PHI->getIncomingBlock(ValNo);
3566 if (Replacement->getType() != UseTy)
3567 Replacement = new BitCastInst(Replacement, UseTy, "",
3569 for (unsigned i = 0, e = PHI->getNumIncomingValues();
3571 if (PHI->getIncomingBlock(i) == BB) {
3572 // Keep the UI iterator valid.
3573 if (&PHI->getOperandUse(
3574 PHINode::getOperandNumForIncomingValue(i)) ==
3577 PHI->setIncomingValue(i, Replacement);
3580 if (Replacement->getType() != UseTy)
3581 Replacement = new BitCastInst(Replacement, UseTy, "", UserInst);
3587 // If Arg is a no-op casted pointer, strip one level of casts and
3589 if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg))
3590 Arg = BI->getOperand(0);
3591 else if (isa<GEPOperator>(Arg) &&
3592 cast<GEPOperator>(Arg)->hasAllZeroIndices())
3593 Arg = cast<GEPOperator>(Arg)->getPointerOperand();
3594 else if (isa<GlobalAlias>(Arg) &&
3595 !cast<GlobalAlias>(Arg)->mayBeOverridden())
3596 Arg = cast<GlobalAlias>(Arg)->getAliasee();