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 Automatic
11 /// Reference Counting and is a system for managing reference counts for objects
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 knowledge 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/ADT/DenseMap.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_ostream.h"
39 /// \brief A handy option to enable/disable all optimizations in this file.
40 static cl::opt<bool> EnableARCOpts("enable-objc-arc-opts", cl::init(true));
42 /// \defgroup MiscUtils Miscellaneous utilities that are not ARC specific.
46 /// \brief An associative container with fast insertion-order (deterministic)
47 /// iteration over its elements. Plus the special blot operation.
48 template<class KeyT, class ValueT>
50 /// Map keys to indices in Vector.
51 typedef DenseMap<KeyT, size_t> MapTy;
54 typedef std::vector<std::pair<KeyT, ValueT> > VectorTy;
59 typedef typename VectorTy::iterator iterator;
60 typedef typename VectorTy::const_iterator const_iterator;
61 iterator begin() { return Vector.begin(); }
62 iterator end() { return Vector.end(); }
63 const_iterator begin() const { return Vector.begin(); }
64 const_iterator end() const { return Vector.end(); }
68 assert(Vector.size() >= Map.size()); // May differ due to blotting.
69 for (typename MapTy::const_iterator I = Map.begin(), E = Map.end();
71 assert(I->second < Vector.size());
72 assert(Vector[I->second].first == I->first);
74 for (typename VectorTy::const_iterator I = Vector.begin(),
75 E = Vector.end(); I != E; ++I)
77 (Map.count(I->first) &&
78 Map[I->first] == size_t(I - Vector.begin())));
82 ValueT &operator[](const KeyT &Arg) {
83 std::pair<typename MapTy::iterator, bool> Pair =
84 Map.insert(std::make_pair(Arg, size_t(0)));
86 size_t Num = Vector.size();
87 Pair.first->second = Num;
88 Vector.push_back(std::make_pair(Arg, ValueT()));
89 return Vector[Num].second;
91 return Vector[Pair.first->second].second;
94 std::pair<iterator, bool>
95 insert(const std::pair<KeyT, ValueT> &InsertPair) {
96 std::pair<typename MapTy::iterator, bool> Pair =
97 Map.insert(std::make_pair(InsertPair.first, size_t(0)));
99 size_t Num = Vector.size();
100 Pair.first->second = Num;
101 Vector.push_back(InsertPair);
102 return std::make_pair(Vector.begin() + Num, true);
104 return std::make_pair(Vector.begin() + Pair.first->second, false);
107 const_iterator find(const KeyT &Key) const {
108 typename MapTy::const_iterator It = Map.find(Key);
109 if (It == Map.end()) return Vector.end();
110 return Vector.begin() + It->second;
113 /// This is similar to erase, but instead of removing the element from the
114 /// vector, it just zeros out the key in the vector. This leaves iterators
115 /// intact, but clients must be prepared for zeroed-out keys when iterating.
116 void blot(const KeyT &Key) {
117 typename MapTy::iterator It = Map.find(Key);
118 if (It == Map.end()) return;
119 Vector[It->second].first = KeyT();
132 /// \defgroup ARCUtilities Utility declarations/definitions specific to ARC.
135 #include "llvm/ADT/StringSwitch.h"
136 #include "llvm/Analysis/ValueTracking.h"
137 #include "llvm/IR/Intrinsics.h"
138 #include "llvm/IR/Module.h"
139 #include "llvm/Support/CallSite.h"
140 #include "llvm/Transforms/Utils/Local.h"
143 /// \enum InstructionClass
144 /// \brief A simple classification for instructions.
145 enum InstructionClass {
146 IC_Retain, ///< objc_retain
147 IC_RetainRV, ///< objc_retainAutoreleasedReturnValue
148 IC_RetainBlock, ///< objc_retainBlock
149 IC_Release, ///< objc_release
150 IC_Autorelease, ///< objc_autorelease
151 IC_AutoreleaseRV, ///< objc_autoreleaseReturnValue
152 IC_AutoreleasepoolPush, ///< objc_autoreleasePoolPush
153 IC_AutoreleasepoolPop, ///< objc_autoreleasePoolPop
154 IC_NoopCast, ///< objc_retainedObject, etc.
155 IC_FusedRetainAutorelease, ///< objc_retainAutorelease
156 IC_FusedRetainAutoreleaseRV, ///< objc_retainAutoreleaseReturnValue
157 IC_LoadWeakRetained, ///< objc_loadWeakRetained (primitive)
158 IC_StoreWeak, ///< objc_storeWeak (primitive)
159 IC_InitWeak, ///< objc_initWeak (derived)
160 IC_LoadWeak, ///< objc_loadWeak (derived)
161 IC_MoveWeak, ///< objc_moveWeak (derived)
162 IC_CopyWeak, ///< objc_copyWeak (derived)
163 IC_DestroyWeak, ///< objc_destroyWeak (derived)
164 IC_StoreStrong, ///< objc_storeStrong (derived)
165 IC_CallOrUser, ///< could call objc_release and/or "use" pointers
166 IC_Call, ///< could call objc_release
167 IC_User, ///< could "use" a pointer
168 IC_None ///< anything else
171 raw_ostream &operator<<(raw_ostream &OS, const InstructionClass Class)
173 raw_ostream &operator<<(raw_ostream &OS, const InstructionClass Class) {
176 return OS << "IC_Retain";
178 return OS << "IC_RetainRV";
180 return OS << "IC_RetainBlock";
182 return OS << "IC_Release";
184 return OS << "IC_Autorelease";
185 case IC_AutoreleaseRV:
186 return OS << "IC_AutoreleaseRV";
187 case IC_AutoreleasepoolPush:
188 return OS << "IC_AutoreleasepoolPush";
189 case IC_AutoreleasepoolPop:
190 return OS << "IC_AutoreleasepoolPop";
192 return OS << "IC_NoopCast";
193 case IC_FusedRetainAutorelease:
194 return OS << "IC_FusedRetainAutorelease";
195 case IC_FusedRetainAutoreleaseRV:
196 return OS << "IC_FusedRetainAutoreleaseRV";
197 case IC_LoadWeakRetained:
198 return OS << "IC_LoadWeakRetained";
200 return OS << "IC_StoreWeak";
202 return OS << "IC_InitWeak";
204 return OS << "IC_LoadWeak";
206 return OS << "IC_MoveWeak";
208 return OS << "IC_CopyWeak";
210 return OS << "IC_DestroyWeak";
212 return OS << "IC_StoreStrong";
214 return OS << "IC_CallOrUser";
216 return OS << "IC_Call";
218 return OS << "IC_User";
220 return OS << "IC_None";
222 llvm_unreachable("Unknown instruction class!");
226 /// \brief Test whether the given value is possible a reference-counted pointer.
227 static bool IsPotentialUse(const Value *Op) {
228 // Pointers to static or stack storage are not reference-counted pointers.
229 if (isa<Constant>(Op) || isa<AllocaInst>(Op))
231 // Special arguments are not reference-counted.
232 if (const Argument *Arg = dyn_cast<Argument>(Op))
233 if (Arg->hasByValAttr() ||
234 Arg->hasNestAttr() ||
235 Arg->hasStructRetAttr())
237 // Only consider values with pointer types.
238 // It seemes intuitive to exclude function pointer types as well, since
239 // functions are never reference-counted, however clang occasionally
240 // bitcasts reference-counted pointers to function-pointer type
242 PointerType *Ty = dyn_cast<PointerType>(Op->getType());
245 // Conservatively assume anything else is a potential use.
249 /// \brief Helper for GetInstructionClass. Determines what kind of construct CS
251 static InstructionClass GetCallSiteClass(ImmutableCallSite CS) {
252 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
254 if (IsPotentialUse(*I))
255 return CS.onlyReadsMemory() ? IC_User : IC_CallOrUser;
257 return CS.onlyReadsMemory() ? IC_None : IC_Call;
260 /// \brief Determine if F is one of the special known Functions. If it isn't,
261 /// return IC_CallOrUser.
262 static InstructionClass GetFunctionClass(const Function *F) {
263 Function::const_arg_iterator AI = F->arg_begin(), AE = F->arg_end();
267 return StringSwitch<InstructionClass>(F->getName())
268 .Case("objc_autoreleasePoolPush", IC_AutoreleasepoolPush)
269 .Default(IC_CallOrUser);
272 const Argument *A0 = AI++;
274 // Argument is a pointer.
275 if (PointerType *PTy = dyn_cast<PointerType>(A0->getType())) {
276 Type *ETy = PTy->getElementType();
278 if (ETy->isIntegerTy(8))
279 return StringSwitch<InstructionClass>(F->getName())
280 .Case("objc_retain", IC_Retain)
281 .Case("objc_retainAutoreleasedReturnValue", IC_RetainRV)
282 .Case("objc_retainBlock", IC_RetainBlock)
283 .Case("objc_release", IC_Release)
284 .Case("objc_autorelease", IC_Autorelease)
285 .Case("objc_autoreleaseReturnValue", IC_AutoreleaseRV)
286 .Case("objc_autoreleasePoolPop", IC_AutoreleasepoolPop)
287 .Case("objc_retainedObject", IC_NoopCast)
288 .Case("objc_unretainedObject", IC_NoopCast)
289 .Case("objc_unretainedPointer", IC_NoopCast)
290 .Case("objc_retain_autorelease", IC_FusedRetainAutorelease)
291 .Case("objc_retainAutorelease", IC_FusedRetainAutorelease)
292 .Case("objc_retainAutoreleaseReturnValue",IC_FusedRetainAutoreleaseRV)
293 .Default(IC_CallOrUser);
296 if (PointerType *Pte = dyn_cast<PointerType>(ETy))
297 if (Pte->getElementType()->isIntegerTy(8))
298 return StringSwitch<InstructionClass>(F->getName())
299 .Case("objc_loadWeakRetained", IC_LoadWeakRetained)
300 .Case("objc_loadWeak", IC_LoadWeak)
301 .Case("objc_destroyWeak", IC_DestroyWeak)
302 .Default(IC_CallOrUser);
305 // Two arguments, first is i8**.
306 const Argument *A1 = AI++;
308 if (PointerType *PTy = dyn_cast<PointerType>(A0->getType()))
309 if (PointerType *Pte = dyn_cast<PointerType>(PTy->getElementType()))
310 if (Pte->getElementType()->isIntegerTy(8))
311 if (PointerType *PTy1 = dyn_cast<PointerType>(A1->getType())) {
312 Type *ETy1 = PTy1->getElementType();
313 // Second argument is i8*
314 if (ETy1->isIntegerTy(8))
315 return StringSwitch<InstructionClass>(F->getName())
316 .Case("objc_storeWeak", IC_StoreWeak)
317 .Case("objc_initWeak", IC_InitWeak)
318 .Case("objc_storeStrong", IC_StoreStrong)
319 .Default(IC_CallOrUser);
320 // Second argument is i8**.
321 if (PointerType *Pte1 = dyn_cast<PointerType>(ETy1))
322 if (Pte1->getElementType()->isIntegerTy(8))
323 return StringSwitch<InstructionClass>(F->getName())
324 .Case("objc_moveWeak", IC_MoveWeak)
325 .Case("objc_copyWeak", IC_CopyWeak)
326 .Default(IC_CallOrUser);
330 return IC_CallOrUser;
333 /// \brief Determine what kind of construct V is.
334 static InstructionClass GetInstructionClass(const Value *V) {
335 if (const Instruction *I = dyn_cast<Instruction>(V)) {
336 // Any instruction other than bitcast and gep with a pointer operand have a
337 // use of an objc pointer. Bitcasts, GEPs, Selects, PHIs transfer a pointer
338 // to a subsequent use, rather than using it themselves, in this sense.
339 // As a short cut, several other opcodes are known to have no pointer
340 // operands of interest. And ret is never followed by a release, so it's
341 // not interesting to examine.
342 switch (I->getOpcode()) {
343 case Instruction::Call: {
344 const CallInst *CI = cast<CallInst>(I);
345 // Check for calls to special functions.
346 if (const Function *F = CI->getCalledFunction()) {
347 InstructionClass Class = GetFunctionClass(F);
348 if (Class != IC_CallOrUser)
351 // None of the intrinsic functions do objc_release. For intrinsics, the
352 // only question is whether or not they may be users.
353 switch (F->getIntrinsicID()) {
354 case Intrinsic::returnaddress: case Intrinsic::frameaddress:
355 case Intrinsic::stacksave: case Intrinsic::stackrestore:
356 case Intrinsic::vastart: case Intrinsic::vacopy: case Intrinsic::vaend:
357 case Intrinsic::objectsize: case Intrinsic::prefetch:
358 case Intrinsic::stackprotector:
359 case Intrinsic::eh_return_i32: case Intrinsic::eh_return_i64:
360 case Intrinsic::eh_typeid_for: case Intrinsic::eh_dwarf_cfa:
361 case Intrinsic::eh_sjlj_lsda: case Intrinsic::eh_sjlj_functioncontext:
362 case Intrinsic::init_trampoline: case Intrinsic::adjust_trampoline:
363 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
364 case Intrinsic::invariant_start: case Intrinsic::invariant_end:
365 // Don't let dbg info affect our results.
366 case Intrinsic::dbg_declare: case Intrinsic::dbg_value:
367 // Short cut: Some intrinsics obviously don't use ObjC pointers.
373 return GetCallSiteClass(CI);
375 case Instruction::Invoke:
376 return GetCallSiteClass(cast<InvokeInst>(I));
377 case Instruction::BitCast:
378 case Instruction::GetElementPtr:
379 case Instruction::Select: case Instruction::PHI:
380 case Instruction::Ret: case Instruction::Br:
381 case Instruction::Switch: case Instruction::IndirectBr:
382 case Instruction::Alloca: case Instruction::VAArg:
383 case Instruction::Add: case Instruction::FAdd:
384 case Instruction::Sub: case Instruction::FSub:
385 case Instruction::Mul: case Instruction::FMul:
386 case Instruction::SDiv: case Instruction::UDiv: case Instruction::FDiv:
387 case Instruction::SRem: case Instruction::URem: case Instruction::FRem:
388 case Instruction::Shl: case Instruction::LShr: case Instruction::AShr:
389 case Instruction::And: case Instruction::Or: case Instruction::Xor:
390 case Instruction::SExt: case Instruction::ZExt: case Instruction::Trunc:
391 case Instruction::IntToPtr: case Instruction::FCmp:
392 case Instruction::FPTrunc: case Instruction::FPExt:
393 case Instruction::FPToUI: case Instruction::FPToSI:
394 case Instruction::UIToFP: case Instruction::SIToFP:
395 case Instruction::InsertElement: case Instruction::ExtractElement:
396 case Instruction::ShuffleVector:
397 case Instruction::ExtractValue:
399 case Instruction::ICmp:
400 // Comparing a pointer with null, or any other constant, isn't an
401 // interesting use, because we don't care what the pointer points to, or
402 // about the values of any other dynamic reference-counted pointers.
403 if (IsPotentialUse(I->getOperand(1)))
407 // For anything else, check all the operands.
408 // Note that this includes both operands of a Store: while the first
409 // operand isn't actually being dereferenced, it is being stored to
410 // memory where we can no longer track who might read it and dereference
411 // it, so we have to consider it potentially used.
412 for (User::const_op_iterator OI = I->op_begin(), OE = I->op_end();
414 if (IsPotentialUse(*OI))
419 // Otherwise, it's totally inert for ARC purposes.
423 /// \brief Determine which objc runtime call instruction class V belongs to.
425 /// This is similar to GetInstructionClass except that it only detects objc
426 /// runtime calls. This allows it to be faster.
428 static InstructionClass GetBasicInstructionClass(const Value *V) {
429 if (const CallInst *CI = dyn_cast<CallInst>(V)) {
430 if (const Function *F = CI->getCalledFunction())
431 return GetFunctionClass(F);
432 // Otherwise, be conservative.
433 return IC_CallOrUser;
436 // Otherwise, be conservative.
437 return isa<InvokeInst>(V) ? IC_CallOrUser : IC_User;
440 /// \brief Test if the given class is objc_retain or equivalent.
441 static bool IsRetain(InstructionClass Class) {
442 return Class == IC_Retain ||
443 Class == IC_RetainRV;
446 /// \brief Test if the given class is objc_autorelease or equivalent.
447 static bool IsAutorelease(InstructionClass Class) {
448 return Class == IC_Autorelease ||
449 Class == IC_AutoreleaseRV;
452 /// \brief Test if the given class represents instructions which return their
453 /// argument verbatim.
454 static bool IsForwarding(InstructionClass Class) {
455 // objc_retainBlock technically doesn't always return its argument
456 // verbatim, but it doesn't matter for our purposes here.
457 return Class == IC_Retain ||
458 Class == IC_RetainRV ||
459 Class == IC_Autorelease ||
460 Class == IC_AutoreleaseRV ||
461 Class == IC_RetainBlock ||
462 Class == IC_NoopCast;
465 /// \brief Test if the given class represents instructions which do nothing if
466 /// passed a null pointer.
467 static bool IsNoopOnNull(InstructionClass Class) {
468 return Class == IC_Retain ||
469 Class == IC_RetainRV ||
470 Class == IC_Release ||
471 Class == IC_Autorelease ||
472 Class == IC_AutoreleaseRV ||
473 Class == IC_RetainBlock;
476 /// \brief Test if the given class represents instructions which are always safe
477 /// to mark with the "tail" keyword.
478 static bool IsAlwaysTail(InstructionClass Class) {
479 // IC_RetainBlock may be given a stack argument.
480 return Class == IC_Retain ||
481 Class == IC_RetainRV ||
482 Class == IC_AutoreleaseRV;
485 /// \brief Test if the given class represents instructions which are never safe
486 /// to mark with the "tail" keyword.
487 static bool IsNeverTail(InstructionClass Class) {
488 /// It is never safe to tail call objc_autorelease since by tail calling
489 /// objc_autorelease, we also tail call -[NSObject autorelease] which supports
490 /// fast autoreleasing causing our object to be potentially reclaimed from the
491 /// autorelease pool which violates the semantics of __autoreleasing types in
493 return Class == IC_Autorelease;
496 /// \brief Test if the given class represents instructions which are always safe
497 /// to mark with the nounwind attribute.
498 static bool IsNoThrow(InstructionClass Class) {
499 // objc_retainBlock is not nounwind because it calls user copy constructors
500 // which could theoretically throw.
501 return Class == IC_Retain ||
502 Class == IC_RetainRV ||
503 Class == IC_Release ||
504 Class == IC_Autorelease ||
505 Class == IC_AutoreleaseRV ||
506 Class == IC_AutoreleasepoolPush ||
507 Class == IC_AutoreleasepoolPop;
510 /// \brief Erase the given instruction.
512 /// Many ObjC calls return their argument verbatim,
513 /// so if it's such a call and the return value has users, replace them with the
516 static void EraseInstruction(Instruction *CI) {
517 Value *OldArg = cast<CallInst>(CI)->getArgOperand(0);
519 bool Unused = CI->use_empty();
522 // Replace the return value with the argument.
523 assert(IsForwarding(GetBasicInstructionClass(CI)) &&
524 "Can't delete non-forwarding instruction with users!");
525 CI->replaceAllUsesWith(OldArg);
528 CI->eraseFromParent();
531 RecursivelyDeleteTriviallyDeadInstructions(OldArg);
534 /// \brief This is a wrapper around getUnderlyingObject which also knows how to
535 /// look through objc_retain and objc_autorelease calls, which we know to return
536 /// their argument verbatim.
537 static const Value *GetUnderlyingObjCPtr(const Value *V) {
539 V = GetUnderlyingObject(V);
540 if (!IsForwarding(GetBasicInstructionClass(V)))
542 V = cast<CallInst>(V)->getArgOperand(0);
548 /// \brief This is a wrapper around Value::stripPointerCasts which also knows
549 /// how to look through objc_retain and objc_autorelease calls, which we know to
550 /// return their argument verbatim.
551 static const Value *StripPointerCastsAndObjCCalls(const Value *V) {
553 V = V->stripPointerCasts();
554 if (!IsForwarding(GetBasicInstructionClass(V)))
556 V = cast<CallInst>(V)->getArgOperand(0);
561 /// \brief This is a wrapper around Value::stripPointerCasts which also knows
562 /// how to look through objc_retain and objc_autorelease calls, which we know to
563 /// return their argument verbatim.
564 static Value *StripPointerCastsAndObjCCalls(Value *V) {
566 V = V->stripPointerCasts();
567 if (!IsForwarding(GetBasicInstructionClass(V)))
569 V = cast<CallInst>(V)->getArgOperand(0);
574 /// \brief Assuming the given instruction is one of the special calls such as
575 /// objc_retain or objc_release, return the argument value, stripped of no-op
576 /// casts and forwarding calls.
577 static Value *GetObjCArg(Value *Inst) {
578 return StripPointerCastsAndObjCCalls(cast<CallInst>(Inst)->getArgOperand(0));
581 /// \brief Return true if this value refers to a distinct and identifiable
584 /// This is similar to AliasAnalysis's isIdentifiedObject, except that it uses
585 /// special knowledge of ObjC conventions.
586 static bool IsObjCIdentifiedObject(const Value *V) {
587 // Assume that call results and arguments have their own "provenance".
588 // Constants (including GlobalVariables) and Allocas are never
589 // reference-counted.
590 if (isa<CallInst>(V) || isa<InvokeInst>(V) ||
591 isa<Argument>(V) || isa<Constant>(V) ||
595 if (const LoadInst *LI = dyn_cast<LoadInst>(V)) {
596 const Value *Pointer =
597 StripPointerCastsAndObjCCalls(LI->getPointerOperand());
598 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Pointer)) {
599 // A constant pointer can't be pointing to an object on the heap. It may
600 // be reference-counted, but it won't be deleted.
601 if (GV->isConstant())
603 StringRef Name = GV->getName();
604 // These special variables are known to hold values which are not
605 // reference-counted pointers.
606 if (Name.startswith("\01L_OBJC_SELECTOR_REFERENCES_") ||
607 Name.startswith("\01L_OBJC_CLASSLIST_REFERENCES_") ||
608 Name.startswith("\01L_OBJC_CLASSLIST_SUP_REFS_$_") ||
609 Name.startswith("\01L_OBJC_METH_VAR_NAME_") ||
610 Name.startswith("\01l_objc_msgSend_fixup_"))
618 /// \brief This is similar to StripPointerCastsAndObjCCalls but it stops as soon
619 /// as it finds a value with multiple uses.
620 static const Value *FindSingleUseIdentifiedObject(const Value *Arg) {
621 if (Arg->hasOneUse()) {
622 if (const BitCastInst *BC = dyn_cast<BitCastInst>(Arg))
623 return FindSingleUseIdentifiedObject(BC->getOperand(0));
624 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Arg))
625 if (GEP->hasAllZeroIndices())
626 return FindSingleUseIdentifiedObject(GEP->getPointerOperand());
627 if (IsForwarding(GetBasicInstructionClass(Arg)))
628 return FindSingleUseIdentifiedObject(
629 cast<CallInst>(Arg)->getArgOperand(0));
630 if (!IsObjCIdentifiedObject(Arg))
635 // If we found an identifiable object but it has multiple uses, but they are
636 // trivial uses, we can still consider this to be a single-use value.
637 if (IsObjCIdentifiedObject(Arg)) {
638 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
641 if (!U->use_empty() || StripPointerCastsAndObjCCalls(U) != Arg)
651 /// \brief Test if the given module looks interesting to run ARC optimization
653 static bool ModuleHasARC(const Module &M) {
655 M.getNamedValue("objc_retain") ||
656 M.getNamedValue("objc_release") ||
657 M.getNamedValue("objc_autorelease") ||
658 M.getNamedValue("objc_retainAutoreleasedReturnValue") ||
659 M.getNamedValue("objc_retainBlock") ||
660 M.getNamedValue("objc_autoreleaseReturnValue") ||
661 M.getNamedValue("objc_autoreleasePoolPush") ||
662 M.getNamedValue("objc_loadWeakRetained") ||
663 M.getNamedValue("objc_loadWeak") ||
664 M.getNamedValue("objc_destroyWeak") ||
665 M.getNamedValue("objc_storeWeak") ||
666 M.getNamedValue("objc_initWeak") ||
667 M.getNamedValue("objc_moveWeak") ||
668 M.getNamedValue("objc_copyWeak") ||
669 M.getNamedValue("objc_retainedObject") ||
670 M.getNamedValue("objc_unretainedObject") ||
671 M.getNamedValue("objc_unretainedPointer");
674 /// \brief Test whether the given pointer, which is an Objective C block
675 /// pointer, does not "escape".
677 /// This differs from regular escape analysis in that a use as an
678 /// argument to a call is not considered an escape.
680 static bool DoesObjCBlockEscape(const Value *BlockPtr) {
682 DEBUG(dbgs() << "DoesObjCBlockEscape: Target: " << *BlockPtr << "\n");
684 // Walk the def-use chains.
685 SmallVector<const Value *, 4> Worklist;
686 Worklist.push_back(BlockPtr);
688 // Ensure we do not visit any value twice.
689 SmallPtrSet<const Value *, 4> VisitedSet;
692 const Value *V = Worklist.pop_back_val();
694 DEBUG(dbgs() << "DoesObjCBlockEscape: Visiting: " << *V << "\n");
696 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
698 const User *UUser = *UI;
700 DEBUG(dbgs() << "DoesObjCBlockEscape: User: " << *UUser << "\n");
702 // Special - Use by a call (callee or argument) is not considered
704 switch (GetBasicInstructionClass(UUser)) {
709 case IC_AutoreleaseRV: {
710 DEBUG(dbgs() << "DoesObjCBlockEscape: User copies pointer arguments. "
712 // These special functions make copies of their pointer arguments.
717 // Use by an instruction which copies the value is an escape if the
718 // result is an escape.
719 if (isa<BitCastInst>(UUser) || isa<GetElementPtrInst>(UUser) ||
720 isa<PHINode>(UUser) || isa<SelectInst>(UUser)) {
722 if (!VisitedSet.insert(UUser)) {
723 DEBUG(dbgs() << "DoesObjCBlockEscape: User copies value. Escapes "
724 "if result escapes. Adding to list.\n");
725 Worklist.push_back(UUser);
727 DEBUG(dbgs() << "DoesObjCBlockEscape: Already visited node.\n");
731 // Use by a load is not an escape.
732 if (isa<LoadInst>(UUser))
734 // Use by a store is not an escape if the use is the address.
735 if (const StoreInst *SI = dyn_cast<StoreInst>(UUser))
736 if (V != SI->getValueOperand())
740 // Regular calls and other stuff are not considered escapes.
743 // Otherwise, conservatively assume an escape.
744 DEBUG(dbgs() << "DoesObjCBlockEscape: Assuming block escapes.\n");
747 } while (!Worklist.empty());
750 DEBUG(dbgs() << "DoesObjCBlockEscape: Block does not escape.\n");
756 /// \defgroup ARCAA Extends alias analysis using ObjC specific knowledge.
759 #include "llvm/Analysis/AliasAnalysis.h"
760 #include "llvm/Analysis/Passes.h"
761 #include "llvm/Pass.h"
764 /// \brief This is a simple alias analysis implementation that uses knowledge
765 /// of ARC constructs to answer queries.
767 /// TODO: This class could be generalized to know about other ObjC-specific
768 /// tricks. Such as knowing that ivars in the non-fragile ABI are non-aliasing
769 /// even though their offsets are dynamic.
770 class ObjCARCAliasAnalysis : public ImmutablePass,
771 public AliasAnalysis {
773 static char ID; // Class identification, replacement for typeinfo
774 ObjCARCAliasAnalysis() : ImmutablePass(ID) {
775 initializeObjCARCAliasAnalysisPass(*PassRegistry::getPassRegistry());
779 virtual void initializePass() {
780 InitializeAliasAnalysis(this);
783 /// This method is used when a pass implements an analysis interface through
784 /// multiple inheritance. If needed, it should override this to adjust the
785 /// this pointer as needed for the specified pass info.
786 virtual void *getAdjustedAnalysisPointer(const void *PI) {
787 if (PI == &AliasAnalysis::ID)
788 return static_cast<AliasAnalysis *>(this);
792 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
793 virtual AliasResult alias(const Location &LocA, const Location &LocB);
794 virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
795 virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
796 virtual ModRefBehavior getModRefBehavior(const Function *F);
797 virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
798 const Location &Loc);
799 virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
800 ImmutableCallSite CS2);
802 } // End of anonymous namespace
804 // Register this pass...
805 char ObjCARCAliasAnalysis::ID = 0;
806 INITIALIZE_AG_PASS(ObjCARCAliasAnalysis, AliasAnalysis, "objc-arc-aa",
807 "ObjC-ARC-Based Alias Analysis", false, true, false)
809 ImmutablePass *llvm::createObjCARCAliasAnalysisPass() {
810 return new ObjCARCAliasAnalysis();
814 ObjCARCAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
815 AU.setPreservesAll();
816 AliasAnalysis::getAnalysisUsage(AU);
819 AliasAnalysis::AliasResult
820 ObjCARCAliasAnalysis::alias(const Location &LocA, const Location &LocB) {
822 return AliasAnalysis::alias(LocA, LocB);
824 // First, strip off no-ops, including ObjC-specific no-ops, and try making a
825 // precise alias query.
826 const Value *SA = StripPointerCastsAndObjCCalls(LocA.Ptr);
827 const Value *SB = StripPointerCastsAndObjCCalls(LocB.Ptr);
829 AliasAnalysis::alias(Location(SA, LocA.Size, LocA.TBAATag),
830 Location(SB, LocB.Size, LocB.TBAATag));
831 if (Result != MayAlias)
834 // If that failed, climb to the underlying object, including climbing through
835 // ObjC-specific no-ops, and try making an imprecise alias query.
836 const Value *UA = GetUnderlyingObjCPtr(SA);
837 const Value *UB = GetUnderlyingObjCPtr(SB);
838 if (UA != SA || UB != SB) {
839 Result = AliasAnalysis::alias(Location(UA), Location(UB));
840 // We can't use MustAlias or PartialAlias results here because
841 // GetUnderlyingObjCPtr may return an offsetted pointer value.
842 if (Result == NoAlias)
846 // If that failed, fail. We don't need to chain here, since that's covered
847 // by the earlier precise query.
852 ObjCARCAliasAnalysis::pointsToConstantMemory(const Location &Loc,
855 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
857 // First, strip off no-ops, including ObjC-specific no-ops, and try making
858 // a precise alias query.
859 const Value *S = StripPointerCastsAndObjCCalls(Loc.Ptr);
860 if (AliasAnalysis::pointsToConstantMemory(Location(S, Loc.Size, Loc.TBAATag),
864 // If that failed, climb to the underlying object, including climbing through
865 // ObjC-specific no-ops, and try making an imprecise alias query.
866 const Value *U = GetUnderlyingObjCPtr(S);
868 return AliasAnalysis::pointsToConstantMemory(Location(U), OrLocal);
870 // If that failed, fail. We don't need to chain here, since that's covered
871 // by the earlier precise query.
875 AliasAnalysis::ModRefBehavior
876 ObjCARCAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
877 // We have nothing to do. Just chain to the next AliasAnalysis.
878 return AliasAnalysis::getModRefBehavior(CS);
881 AliasAnalysis::ModRefBehavior
882 ObjCARCAliasAnalysis::getModRefBehavior(const Function *F) {
884 return AliasAnalysis::getModRefBehavior(F);
886 switch (GetFunctionClass(F)) {
888 return DoesNotAccessMemory;
893 return AliasAnalysis::getModRefBehavior(F);
896 AliasAnalysis::ModRefResult
897 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS, const Location &Loc) {
899 return AliasAnalysis::getModRefInfo(CS, Loc);
901 switch (GetBasicInstructionClass(CS.getInstruction())) {
905 case IC_AutoreleaseRV:
907 case IC_AutoreleasepoolPush:
908 case IC_FusedRetainAutorelease:
909 case IC_FusedRetainAutoreleaseRV:
910 // These functions don't access any memory visible to the compiler.
911 // Note that this doesn't include objc_retainBlock, because it updates
912 // pointers when it copies block data.
918 return AliasAnalysis::getModRefInfo(CS, Loc);
921 AliasAnalysis::ModRefResult
922 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS1,
923 ImmutableCallSite CS2) {
924 // TODO: Theoretically we could check for dependencies between objc_* calls
925 // and OnlyAccessesArgumentPointees calls or other well-behaved calls.
926 return AliasAnalysis::getModRefInfo(CS1, CS2);
931 /// \defgroup ARCExpansion Early ARC Optimizations.
934 #include "llvm/Support/InstIterator.h"
935 #include "llvm/Transforms/Scalar.h"
938 /// \brief Early ARC transformations.
939 class ObjCARCExpand : public FunctionPass {
940 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
941 virtual bool doInitialization(Module &M);
942 virtual bool runOnFunction(Function &F);
944 /// A flag indicating whether this optimization pass should run.
949 ObjCARCExpand() : FunctionPass(ID) {
950 initializeObjCARCExpandPass(*PassRegistry::getPassRegistry());
955 char ObjCARCExpand::ID = 0;
956 INITIALIZE_PASS(ObjCARCExpand,
957 "objc-arc-expand", "ObjC ARC expansion", false, false)
959 Pass *llvm::createObjCARCExpandPass() {
960 return new ObjCARCExpand();
963 void ObjCARCExpand::getAnalysisUsage(AnalysisUsage &AU) const {
964 AU.setPreservesCFG();
967 bool ObjCARCExpand::doInitialization(Module &M) {
968 Run = ModuleHasARC(M);
972 bool ObjCARCExpand::runOnFunction(Function &F) {
976 // If nothing in the Module uses ARC, don't do anything.
980 bool Changed = false;
982 DEBUG(dbgs() << "ObjCARCExpand: Visiting Function: " << F.getName() << "\n");
984 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I) {
985 Instruction *Inst = &*I;
987 DEBUG(dbgs() << "ObjCARCExpand: Visiting: " << *Inst << "\n");
989 switch (GetBasicInstructionClass(Inst)) {
993 case IC_AutoreleaseRV:
994 case IC_FusedRetainAutorelease:
995 case IC_FusedRetainAutoreleaseRV: {
996 // These calls return their argument verbatim, as a low-level
997 // optimization. However, this makes high-level optimizations
998 // harder. Undo any uses of this optimization that the front-end
999 // emitted here. We'll redo them in the contract pass.
1001 Value *Value = cast<CallInst>(Inst)->getArgOperand(0);
1002 DEBUG(dbgs() << "ObjCARCExpand: Old = " << *Inst << "\n"
1003 " New = " << *Value << "\n");
1004 Inst->replaceAllUsesWith(Value);
1012 DEBUG(dbgs() << "ObjCARCExpand: Finished List.\n\n");
1019 /// \defgroup ARCAPElim ARC Autorelease Pool Elimination.
1022 #include "llvm/ADT/STLExtras.h"
1023 #include "llvm/IR/Constants.h"
1026 /// \brief Autorelease pool elimination.
1027 class ObjCARCAPElim : public ModulePass {
1028 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1029 virtual bool runOnModule(Module &M);
1031 static bool MayAutorelease(ImmutableCallSite CS, unsigned Depth = 0);
1032 static bool OptimizeBB(BasicBlock *BB);
1036 ObjCARCAPElim() : ModulePass(ID) {
1037 initializeObjCARCAPElimPass(*PassRegistry::getPassRegistry());
1042 char ObjCARCAPElim::ID = 0;
1043 INITIALIZE_PASS(ObjCARCAPElim,
1045 "ObjC ARC autorelease pool elimination",
1048 Pass *llvm::createObjCARCAPElimPass() {
1049 return new ObjCARCAPElim();
1052 void ObjCARCAPElim::getAnalysisUsage(AnalysisUsage &AU) const {
1053 AU.setPreservesCFG();
1056 /// Interprocedurally determine if calls made by the given call site can
1057 /// possibly produce autoreleases.
1058 bool ObjCARCAPElim::MayAutorelease(ImmutableCallSite CS, unsigned Depth) {
1059 if (const Function *Callee = CS.getCalledFunction()) {
1060 if (Callee->isDeclaration() || Callee->mayBeOverridden())
1062 for (Function::const_iterator I = Callee->begin(), E = Callee->end();
1064 const BasicBlock *BB = I;
1065 for (BasicBlock::const_iterator J = BB->begin(), F = BB->end();
1067 if (ImmutableCallSite JCS = ImmutableCallSite(J))
1068 // This recursion depth limit is arbitrary. It's just great
1069 // enough to cover known interesting testcases.
1071 !JCS.onlyReadsMemory() &&
1072 MayAutorelease(JCS, Depth + 1))
1081 bool ObjCARCAPElim::OptimizeBB(BasicBlock *BB) {
1082 bool Changed = false;
1084 Instruction *Push = 0;
1085 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
1086 Instruction *Inst = I++;
1087 switch (GetBasicInstructionClass(Inst)) {
1088 case IC_AutoreleasepoolPush:
1091 case IC_AutoreleasepoolPop:
1092 // If this pop matches a push and nothing in between can autorelease,
1094 if (Push && cast<CallInst>(Inst)->getArgOperand(0) == Push) {
1096 DEBUG(dbgs() << "ObjCARCAPElim::OptimizeBB: Zapping push pop "
1097 "autorelease pair:\n"
1098 " Pop: " << *Inst << "\n"
1099 << " Push: " << *Push << "\n");
1100 Inst->eraseFromParent();
1101 Push->eraseFromParent();
1106 if (MayAutorelease(ImmutableCallSite(Inst)))
1117 bool ObjCARCAPElim::runOnModule(Module &M) {
1121 // If nothing in the Module uses ARC, don't do anything.
1122 if (!ModuleHasARC(M))
1125 // Find the llvm.global_ctors variable, as the first step in
1126 // identifying the global constructors. In theory, unnecessary autorelease
1127 // pools could occur anywhere, but in practice it's pretty rare. Global
1128 // ctors are a place where autorelease pools get inserted automatically,
1129 // so it's pretty common for them to be unnecessary, and it's pretty
1130 // profitable to eliminate them.
1131 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1135 assert(GV->hasDefinitiveInitializer() &&
1136 "llvm.global_ctors is uncooperative!");
1138 bool Changed = false;
1140 // Dig the constructor functions out of GV's initializer.
1141 ConstantArray *Init = cast<ConstantArray>(GV->getInitializer());
1142 for (User::op_iterator OI = Init->op_begin(), OE = Init->op_end();
1145 // llvm.global_ctors is an array of pairs where the second members
1146 // are constructor functions.
1147 Function *F = dyn_cast<Function>(cast<ConstantStruct>(Op)->getOperand(1));
1148 // If the user used a constructor function with the wrong signature and
1149 // it got bitcasted or whatever, look the other way.
1152 // Only look at function definitions.
1153 if (F->isDeclaration())
1155 // Only look at functions with one basic block.
1156 if (llvm::next(F->begin()) != F->end())
1158 // Ok, a single-block constructor function definition. Try to optimize it.
1159 Changed |= OptimizeBB(F->begin());
1167 /// \defgroup ARCOpt ARC Optimization.
1170 // TODO: On code like this:
1173 // stuff_that_cannot_release()
1174 // objc_autorelease(%x)
1175 // stuff_that_cannot_release()
1177 // stuff_that_cannot_release()
1178 // objc_autorelease(%x)
1180 // The second retain and autorelease can be deleted.
1182 // TODO: It should be possible to delete
1183 // objc_autoreleasePoolPush and objc_autoreleasePoolPop
1184 // pairs if nothing is actually autoreleased between them. Also, autorelease
1185 // calls followed by objc_autoreleasePoolPop calls (perhaps in ObjC++ code
1186 // after inlining) can be turned into plain release calls.
1188 // TODO: Critical-edge splitting. If the optimial insertion point is
1189 // a critical edge, the current algorithm has to fail, because it doesn't
1190 // know how to split edges. It should be possible to make the optimizer
1191 // think in terms of edges, rather than blocks, and then split critical
1194 // TODO: OptimizeSequences could generalized to be Interprocedural.
1196 // TODO: Recognize that a bunch of other objc runtime calls have
1197 // non-escaping arguments and non-releasing arguments, and may be
1198 // non-autoreleasing.
1200 // TODO: Sink autorelease calls as far as possible. Unfortunately we
1201 // usually can't sink them past other calls, which would be the main
1202 // case where it would be useful.
1204 // TODO: The pointer returned from objc_loadWeakRetained is retained.
1206 // TODO: Delete release+retain pairs (rare).
1208 #include "llvm/ADT/SmallPtrSet.h"
1209 #include "llvm/ADT/Statistic.h"
1210 #include "llvm/IR/LLVMContext.h"
1211 #include "llvm/Support/CFG.h"
1213 STATISTIC(NumNoops, "Number of no-op objc calls eliminated");
1214 STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated");
1215 STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases");
1216 STATISTIC(NumRets, "Number of return value forwarding "
1217 "retain+autoreleaes eliminated");
1218 STATISTIC(NumRRs, "Number of retain+release paths eliminated");
1219 STATISTIC(NumPeeps, "Number of calls peephole-optimized");
1222 /// \brief This is similar to BasicAliasAnalysis, and it uses many of the same
1223 /// techniques, except it uses special ObjC-specific reasoning about pointer
1225 class ProvenanceAnalysis {
1228 typedef std::pair<const Value *, const Value *> ValuePairTy;
1229 typedef DenseMap<ValuePairTy, bool> CachedResultsTy;
1230 CachedResultsTy CachedResults;
1232 bool relatedCheck(const Value *A, const Value *B);
1233 bool relatedSelect(const SelectInst *A, const Value *B);
1234 bool relatedPHI(const PHINode *A, const Value *B);
1236 void operator=(const ProvenanceAnalysis &) LLVM_DELETED_FUNCTION;
1237 ProvenanceAnalysis(const ProvenanceAnalysis &) LLVM_DELETED_FUNCTION;
1240 ProvenanceAnalysis() {}
1242 void setAA(AliasAnalysis *aa) { AA = aa; }
1244 AliasAnalysis *getAA() const { return AA; }
1246 bool related(const Value *A, const Value *B);
1249 CachedResults.clear();
1254 bool ProvenanceAnalysis::relatedSelect(const SelectInst *A, const Value *B) {
1255 // If the values are Selects with the same condition, we can do a more precise
1256 // check: just check for relations between the values on corresponding arms.
1257 if (const SelectInst *SB = dyn_cast<SelectInst>(B))
1258 if (A->getCondition() == SB->getCondition())
1259 return related(A->getTrueValue(), SB->getTrueValue()) ||
1260 related(A->getFalseValue(), SB->getFalseValue());
1262 // Check both arms of the Select node individually.
1263 return related(A->getTrueValue(), B) ||
1264 related(A->getFalseValue(), B);
1267 bool ProvenanceAnalysis::relatedPHI(const PHINode *A, const Value *B) {
1268 // If the values are PHIs in the same block, we can do a more precise as well
1269 // as efficient check: just check for relations between the values on
1270 // corresponding edges.
1271 if (const PHINode *PNB = dyn_cast<PHINode>(B))
1272 if (PNB->getParent() == A->getParent()) {
1273 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i)
1274 if (related(A->getIncomingValue(i),
1275 PNB->getIncomingValueForBlock(A->getIncomingBlock(i))))
1280 // Check each unique source of the PHI node against B.
1281 SmallPtrSet<const Value *, 4> UniqueSrc;
1282 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i) {
1283 const Value *PV1 = A->getIncomingValue(i);
1284 if (UniqueSrc.insert(PV1) && related(PV1, B))
1288 // All of the arms checked out.
1292 /// Test if the value of P, or any value covered by its provenance, is ever
1293 /// stored within the function (not counting callees).
1294 static bool isStoredObjCPointer(const Value *P) {
1295 SmallPtrSet<const Value *, 8> Visited;
1296 SmallVector<const Value *, 8> Worklist;
1297 Worklist.push_back(P);
1300 P = Worklist.pop_back_val();
1301 for (Value::const_use_iterator UI = P->use_begin(), UE = P->use_end();
1303 const User *Ur = *UI;
1304 if (isa<StoreInst>(Ur)) {
1305 if (UI.getOperandNo() == 0)
1306 // The pointer is stored.
1308 // The pointed is stored through.
1311 if (isa<CallInst>(Ur))
1312 // The pointer is passed as an argument, ignore this.
1314 if (isa<PtrToIntInst>(P))
1315 // Assume the worst.
1317 if (Visited.insert(Ur))
1318 Worklist.push_back(Ur);
1320 } while (!Worklist.empty());
1322 // Everything checked out.
1326 bool ProvenanceAnalysis::relatedCheck(const Value *A, const Value *B) {
1327 // Skip past provenance pass-throughs.
1328 A = GetUnderlyingObjCPtr(A);
1329 B = GetUnderlyingObjCPtr(B);
1335 // Ask regular AliasAnalysis, for a first approximation.
1336 switch (AA->alias(A, B)) {
1337 case AliasAnalysis::NoAlias:
1339 case AliasAnalysis::MustAlias:
1340 case AliasAnalysis::PartialAlias:
1342 case AliasAnalysis::MayAlias:
1346 bool AIsIdentified = IsObjCIdentifiedObject(A);
1347 bool BIsIdentified = IsObjCIdentifiedObject(B);
1349 // An ObjC-Identified object can't alias a load if it is never locally stored.
1350 if (AIsIdentified) {
1351 // Check for an obvious escape.
1352 if (isa<LoadInst>(B))
1353 return isStoredObjCPointer(A);
1354 if (BIsIdentified) {
1355 // Check for an obvious escape.
1356 if (isa<LoadInst>(A))
1357 return isStoredObjCPointer(B);
1358 // Both pointers are identified and escapes aren't an evident problem.
1361 } else if (BIsIdentified) {
1362 // Check for an obvious escape.
1363 if (isa<LoadInst>(A))
1364 return isStoredObjCPointer(B);
1367 // Special handling for PHI and Select.
1368 if (const PHINode *PN = dyn_cast<PHINode>(A))
1369 return relatedPHI(PN, B);
1370 if (const PHINode *PN = dyn_cast<PHINode>(B))
1371 return relatedPHI(PN, A);
1372 if (const SelectInst *S = dyn_cast<SelectInst>(A))
1373 return relatedSelect(S, B);
1374 if (const SelectInst *S = dyn_cast<SelectInst>(B))
1375 return relatedSelect(S, A);
1381 bool ProvenanceAnalysis::related(const Value *A, const Value *B) {
1382 // Begin by inserting a conservative value into the map. If the insertion
1383 // fails, we have the answer already. If it succeeds, leave it there until we
1384 // compute the real answer to guard against recursive queries.
1385 if (A > B) std::swap(A, B);
1386 std::pair<CachedResultsTy::iterator, bool> Pair =
1387 CachedResults.insert(std::make_pair(ValuePairTy(A, B), true));
1389 return Pair.first->second;
1391 bool Result = relatedCheck(A, B);
1392 CachedResults[ValuePairTy(A, B)] = Result;
1399 /// \brief A sequence of states that a pointer may go through in which an
1400 /// objc_retain and objc_release are actually needed.
1403 S_Retain, ///< objc_retain(x)
1404 S_CanRelease, ///< foo(x) -- x could possibly see a ref count decrement
1405 S_Use, ///< any use of x
1406 S_Stop, ///< like S_Release, but code motion is stopped
1407 S_Release, ///< objc_release(x)
1408 S_MovableRelease ///< objc_release(x), !clang.imprecise_release
1412 static Sequence MergeSeqs(Sequence A, Sequence B, bool TopDown) {
1416 if (A == S_None || B == S_None)
1419 if (A > B) std::swap(A, B);
1421 // Choose the side which is further along in the sequence.
1422 if ((A == S_Retain || A == S_CanRelease) &&
1423 (B == S_CanRelease || B == S_Use))
1426 // Choose the side which is further along in the sequence.
1427 if ((A == S_Use || A == S_CanRelease) &&
1428 (B == S_Use || B == S_Release || B == S_Stop || B == S_MovableRelease))
1430 // If both sides are releases, choose the more conservative one.
1431 if (A == S_Stop && (B == S_Release || B == S_MovableRelease))
1433 if (A == S_Release && B == S_MovableRelease)
1441 /// \brief Unidirectional information about either a
1442 /// retain-decrement-use-release sequence or release-use-decrement-retain
1443 /// reverese sequence.
1445 /// After an objc_retain, the reference count of the referenced
1446 /// object is known to be positive. Similarly, before an objc_release, the
1447 /// reference count of the referenced object is known to be positive. If
1448 /// there are retain-release pairs in code regions where the retain count
1449 /// is known to be positive, they can be eliminated, regardless of any side
1450 /// effects between them.
1452 /// Also, a retain+release pair nested within another retain+release
1453 /// pair all on the known same pointer value can be eliminated, regardless
1454 /// of any intervening side effects.
1456 /// KnownSafe is true when either of these conditions is satisfied.
1459 /// True if the Calls are objc_retainBlock calls (as opposed to objc_retain
1463 /// True of the objc_release calls are all marked with the "tail" keyword.
1464 bool IsTailCallRelease;
1466 /// If the Calls are objc_release calls and they all have a
1467 /// clang.imprecise_release tag, this is the metadata tag.
1468 MDNode *ReleaseMetadata;
1470 /// For a top-down sequence, the set of objc_retains or
1471 /// objc_retainBlocks. For bottom-up, the set of objc_releases.
1472 SmallPtrSet<Instruction *, 2> Calls;
1474 /// The set of optimal insert positions for moving calls in the opposite
1476 SmallPtrSet<Instruction *, 2> ReverseInsertPts;
1479 KnownSafe(false), IsRetainBlock(false),
1480 IsTailCallRelease(false),
1481 ReleaseMetadata(0) {}
1487 void RRInfo::clear() {
1489 IsRetainBlock = false;
1490 IsTailCallRelease = false;
1491 ReleaseMetadata = 0;
1493 ReverseInsertPts.clear();
1497 /// \brief This class summarizes several per-pointer runtime properties which
1498 /// are propogated through the flow graph.
1500 /// True if the reference count is known to be incremented.
1501 bool KnownPositiveRefCount;
1503 /// True of we've seen an opportunity for partial RR elimination, such as
1504 /// pushing calls into a CFG triangle or into one side of a CFG diamond.
1507 /// The current position in the sequence.
1511 /// Unidirectional information about the current sequence.
1513 /// TODO: Encapsulate this better.
1516 PtrState() : KnownPositiveRefCount(false), Partial(false),
1519 void SetKnownPositiveRefCount() {
1520 KnownPositiveRefCount = true;
1523 void ClearRefCount() {
1524 KnownPositiveRefCount = false;
1527 bool IsKnownIncremented() const {
1528 return KnownPositiveRefCount;
1531 void SetSeq(Sequence NewSeq) {
1535 Sequence GetSeq() const {
1539 void ClearSequenceProgress() {
1540 ResetSequenceProgress(S_None);
1543 void ResetSequenceProgress(Sequence NewSeq) {
1549 void Merge(const PtrState &Other, bool TopDown);
1554 PtrState::Merge(const PtrState &Other, bool TopDown) {
1555 Seq = MergeSeqs(Seq, Other.Seq, TopDown);
1556 KnownPositiveRefCount = KnownPositiveRefCount && Other.KnownPositiveRefCount;
1558 // We can't merge a plain objc_retain with an objc_retainBlock.
1559 if (RRI.IsRetainBlock != Other.RRI.IsRetainBlock)
1562 // If we're not in a sequence (anymore), drop all associated state.
1563 if (Seq == S_None) {
1566 } else if (Partial || Other.Partial) {
1567 // If we're doing a merge on a path that's previously seen a partial
1568 // merge, conservatively drop the sequence, to avoid doing partial
1569 // RR elimination. If the branch predicates for the two merge differ,
1570 // mixing them is unsafe.
1571 ClearSequenceProgress();
1573 // Conservatively merge the ReleaseMetadata information.
1574 if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
1575 RRI.ReleaseMetadata = 0;
1577 RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe;
1578 RRI.IsTailCallRelease = RRI.IsTailCallRelease &&
1579 Other.RRI.IsTailCallRelease;
1580 RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
1582 // Merge the insert point sets. If there are any differences,
1583 // that makes this a partial merge.
1584 Partial = RRI.ReverseInsertPts.size() != Other.RRI.ReverseInsertPts.size();
1585 for (SmallPtrSet<Instruction *, 2>::const_iterator
1586 I = Other.RRI.ReverseInsertPts.begin(),
1587 E = Other.RRI.ReverseInsertPts.end(); I != E; ++I)
1588 Partial |= RRI.ReverseInsertPts.insert(*I);
1593 /// \brief Per-BasicBlock state.
1595 /// The number of unique control paths from the entry which can reach this
1597 unsigned TopDownPathCount;
1599 /// The number of unique control paths to exits from this block.
1600 unsigned BottomUpPathCount;
1602 /// A type for PerPtrTopDown and PerPtrBottomUp.
1603 typedef MapVector<const Value *, PtrState> MapTy;
1605 /// The top-down traversal uses this to record information known about a
1606 /// pointer at the bottom of each block.
1607 MapTy PerPtrTopDown;
1609 /// The bottom-up traversal uses this to record information known about a
1610 /// pointer at the top of each block.
1611 MapTy PerPtrBottomUp;
1613 /// Effective predecessors of the current block ignoring ignorable edges and
1614 /// ignored backedges.
1615 SmallVector<BasicBlock *, 2> Preds;
1616 /// Effective successors of the current block ignoring ignorable edges and
1617 /// ignored backedges.
1618 SmallVector<BasicBlock *, 2> Succs;
1621 BBState() : TopDownPathCount(0), BottomUpPathCount(0) {}
1623 typedef MapTy::iterator ptr_iterator;
1624 typedef MapTy::const_iterator ptr_const_iterator;
1626 ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); }
1627 ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); }
1628 ptr_const_iterator top_down_ptr_begin() const {
1629 return PerPtrTopDown.begin();
1631 ptr_const_iterator top_down_ptr_end() const {
1632 return PerPtrTopDown.end();
1635 ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); }
1636 ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); }
1637 ptr_const_iterator bottom_up_ptr_begin() const {
1638 return PerPtrBottomUp.begin();
1640 ptr_const_iterator bottom_up_ptr_end() const {
1641 return PerPtrBottomUp.end();
1644 /// Mark this block as being an entry block, which has one path from the
1645 /// entry by definition.
1646 void SetAsEntry() { TopDownPathCount = 1; }
1648 /// Mark this block as being an exit block, which has one path to an exit by
1650 void SetAsExit() { BottomUpPathCount = 1; }
1652 PtrState &getPtrTopDownState(const Value *Arg) {
1653 return PerPtrTopDown[Arg];
1656 PtrState &getPtrBottomUpState(const Value *Arg) {
1657 return PerPtrBottomUp[Arg];
1660 void clearBottomUpPointers() {
1661 PerPtrBottomUp.clear();
1664 void clearTopDownPointers() {
1665 PerPtrTopDown.clear();
1668 void InitFromPred(const BBState &Other);
1669 void InitFromSucc(const BBState &Other);
1670 void MergePred(const BBState &Other);
1671 void MergeSucc(const BBState &Other);
1673 /// Return the number of possible unique paths from an entry to an exit
1674 /// which pass through this block. This is only valid after both the
1675 /// top-down and bottom-up traversals are complete.
1676 unsigned GetAllPathCount() const {
1677 assert(TopDownPathCount != 0);
1678 assert(BottomUpPathCount != 0);
1679 return TopDownPathCount * BottomUpPathCount;
1682 // Specialized CFG utilities.
1683 typedef SmallVectorImpl<BasicBlock *>::const_iterator edge_iterator;
1684 edge_iterator pred_begin() { return Preds.begin(); }
1685 edge_iterator pred_end() { return Preds.end(); }
1686 edge_iterator succ_begin() { return Succs.begin(); }
1687 edge_iterator succ_end() { return Succs.end(); }
1689 void addSucc(BasicBlock *Succ) { Succs.push_back(Succ); }
1690 void addPred(BasicBlock *Pred) { Preds.push_back(Pred); }
1692 bool isExit() const { return Succs.empty(); }
1696 void BBState::InitFromPred(const BBState &Other) {
1697 PerPtrTopDown = Other.PerPtrTopDown;
1698 TopDownPathCount = Other.TopDownPathCount;
1701 void BBState::InitFromSucc(const BBState &Other) {
1702 PerPtrBottomUp = Other.PerPtrBottomUp;
1703 BottomUpPathCount = Other.BottomUpPathCount;
1706 /// The top-down traversal uses this to merge information about predecessors to
1707 /// form the initial state for a new block.
1708 void BBState::MergePred(const BBState &Other) {
1709 // Other.TopDownPathCount can be 0, in which case it is either dead or a
1710 // loop backedge. Loop backedges are special.
1711 TopDownPathCount += Other.TopDownPathCount;
1713 // Check for overflow. If we have overflow, fall back to conservative
1715 if (TopDownPathCount < Other.TopDownPathCount) {
1716 clearTopDownPointers();
1720 // For each entry in the other set, if our set has an entry with the same key,
1721 // merge the entries. Otherwise, copy the entry and merge it with an empty
1723 for (ptr_const_iterator MI = Other.top_down_ptr_begin(),
1724 ME = Other.top_down_ptr_end(); MI != ME; ++MI) {
1725 std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI);
1726 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1730 // For each entry in our set, if the other set doesn't have an entry with the
1731 // same key, force it to merge with an empty entry.
1732 for (ptr_iterator MI = top_down_ptr_begin(),
1733 ME = top_down_ptr_end(); MI != ME; ++MI)
1734 if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end())
1735 MI->second.Merge(PtrState(), /*TopDown=*/true);
1738 /// The bottom-up traversal uses this to merge information about successors to
1739 /// form the initial state for a new block.
1740 void BBState::MergeSucc(const BBState &Other) {
1741 // Other.BottomUpPathCount can be 0, in which case it is either dead or a
1742 // loop backedge. Loop backedges are special.
1743 BottomUpPathCount += Other.BottomUpPathCount;
1745 // Check for overflow. If we have overflow, fall back to conservative
1747 if (BottomUpPathCount < Other.BottomUpPathCount) {
1748 clearBottomUpPointers();
1752 // For each entry in the other set, if our set has an entry with the
1753 // same key, merge the entries. Otherwise, copy the entry and merge
1754 // it with an empty entry.
1755 for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(),
1756 ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) {
1757 std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI);
1758 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1762 // For each entry in our set, if the other set doesn't have an entry
1763 // with the same key, force it to merge with an empty entry.
1764 for (ptr_iterator MI = bottom_up_ptr_begin(),
1765 ME = bottom_up_ptr_end(); MI != ME; ++MI)
1766 if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end())
1767 MI->second.Merge(PtrState(), /*TopDown=*/false);
1771 /// \brief The main ARC optimization pass.
1772 class ObjCARCOpt : public FunctionPass {
1774 ProvenanceAnalysis PA;
1776 /// A flag indicating whether this optimization pass should run.
1779 /// Declarations for ObjC runtime functions, for use in creating calls to
1780 /// them. These are initialized lazily to avoid cluttering up the Module
1781 /// with unused declarations.
1783 /// Declaration for ObjC runtime function
1784 /// objc_retainAutoreleasedReturnValue.
1785 Constant *RetainRVCallee;
1786 /// Declaration for ObjC runtime function objc_autoreleaseReturnValue.
1787 Constant *AutoreleaseRVCallee;
1788 /// Declaration for ObjC runtime function objc_release.
1789 Constant *ReleaseCallee;
1790 /// Declaration for ObjC runtime function objc_retain.
1791 Constant *RetainCallee;
1792 /// Declaration for ObjC runtime function objc_retainBlock.
1793 Constant *RetainBlockCallee;
1794 /// Declaration for ObjC runtime function objc_autorelease.
1795 Constant *AutoreleaseCallee;
1797 /// Flags which determine whether each of the interesting runtine functions
1798 /// is in fact used in the current function.
1799 unsigned UsedInThisFunction;
1801 /// The Metadata Kind for clang.imprecise_release metadata.
1802 unsigned ImpreciseReleaseMDKind;
1804 /// The Metadata Kind for clang.arc.copy_on_escape metadata.
1805 unsigned CopyOnEscapeMDKind;
1807 /// The Metadata Kind for clang.arc.no_objc_arc_exceptions metadata.
1808 unsigned NoObjCARCExceptionsMDKind;
1810 Constant *getRetainRVCallee(Module *M);
1811 Constant *getAutoreleaseRVCallee(Module *M);
1812 Constant *getReleaseCallee(Module *M);
1813 Constant *getRetainCallee(Module *M);
1814 Constant *getRetainBlockCallee(Module *M);
1815 Constant *getAutoreleaseCallee(Module *M);
1817 bool IsRetainBlockOptimizable(const Instruction *Inst);
1819 void OptimizeRetainCall(Function &F, Instruction *Retain);
1820 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
1821 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
1822 InstructionClass &Class);
1823 void OptimizeIndividualCalls(Function &F);
1825 void CheckForCFGHazards(const BasicBlock *BB,
1826 DenseMap<const BasicBlock *, BBState> &BBStates,
1827 BBState &MyStates) const;
1828 bool VisitInstructionBottomUp(Instruction *Inst,
1830 MapVector<Value *, RRInfo> &Retains,
1832 bool VisitBottomUp(BasicBlock *BB,
1833 DenseMap<const BasicBlock *, BBState> &BBStates,
1834 MapVector<Value *, RRInfo> &Retains);
1835 bool VisitInstructionTopDown(Instruction *Inst,
1836 DenseMap<Value *, RRInfo> &Releases,
1838 bool VisitTopDown(BasicBlock *BB,
1839 DenseMap<const BasicBlock *, BBState> &BBStates,
1840 DenseMap<Value *, RRInfo> &Releases);
1841 bool Visit(Function &F,
1842 DenseMap<const BasicBlock *, BBState> &BBStates,
1843 MapVector<Value *, RRInfo> &Retains,
1844 DenseMap<Value *, RRInfo> &Releases);
1846 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1847 MapVector<Value *, RRInfo> &Retains,
1848 DenseMap<Value *, RRInfo> &Releases,
1849 SmallVectorImpl<Instruction *> &DeadInsts,
1852 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1853 MapVector<Value *, RRInfo> &Retains,
1854 DenseMap<Value *, RRInfo> &Releases,
1857 void OptimizeWeakCalls(Function &F);
1859 bool OptimizeSequences(Function &F);
1861 void OptimizeReturns(Function &F);
1863 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1864 virtual bool doInitialization(Module &M);
1865 virtual bool runOnFunction(Function &F);
1866 virtual void releaseMemory();
1870 ObjCARCOpt() : FunctionPass(ID) {
1871 initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1876 char ObjCARCOpt::ID = 0;
1877 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1878 "objc-arc", "ObjC ARC optimization", false, false)
1879 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1880 INITIALIZE_PASS_END(ObjCARCOpt,
1881 "objc-arc", "ObjC ARC optimization", false, false)
1883 Pass *llvm::createObjCARCOptPass() {
1884 return new ObjCARCOpt();
1887 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1888 AU.addRequired<ObjCARCAliasAnalysis>();
1889 AU.addRequired<AliasAnalysis>();
1890 // ARC optimization doesn't currently split critical edges.
1891 AU.setPreservesCFG();
1894 bool ObjCARCOpt::IsRetainBlockOptimizable(const Instruction *Inst) {
1895 // Without the magic metadata tag, we have to assume this might be an
1896 // objc_retainBlock call inserted to convert a block pointer to an id,
1897 // in which case it really is needed.
1898 if (!Inst->getMetadata(CopyOnEscapeMDKind))
1901 // If the pointer "escapes" (not including being used in a call),
1902 // the copy may be needed.
1903 if (DoesObjCBlockEscape(Inst))
1906 // Otherwise, it's not needed.
1910 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1911 if (!RetainRVCallee) {
1912 LLVMContext &C = M->getContext();
1913 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1914 Type *Params[] = { I8X };
1915 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1916 AttributeSet Attribute =
1917 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1918 Attribute::NoUnwind);
1920 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1923 return RetainRVCallee;
1926 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1927 if (!AutoreleaseRVCallee) {
1928 LLVMContext &C = M->getContext();
1929 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1930 Type *Params[] = { I8X };
1931 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1932 AttributeSet Attribute =
1933 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1934 Attribute::NoUnwind);
1935 AutoreleaseRVCallee =
1936 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1939 return AutoreleaseRVCallee;
1942 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1943 if (!ReleaseCallee) {
1944 LLVMContext &C = M->getContext();
1945 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1946 AttributeSet Attribute =
1947 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1948 Attribute::NoUnwind);
1950 M->getOrInsertFunction(
1952 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1955 return ReleaseCallee;
1958 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1959 if (!RetainCallee) {
1960 LLVMContext &C = M->getContext();
1961 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1962 AttributeSet Attribute =
1963 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1964 Attribute::NoUnwind);
1966 M->getOrInsertFunction(
1968 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1971 return RetainCallee;
1974 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1975 if (!RetainBlockCallee) {
1976 LLVMContext &C = M->getContext();
1977 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1978 // objc_retainBlock is not nounwind because it calls user copy constructors
1979 // which could theoretically throw.
1981 M->getOrInsertFunction(
1983 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1986 return RetainBlockCallee;
1989 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1990 if (!AutoreleaseCallee) {
1991 LLVMContext &C = M->getContext();
1992 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1993 AttributeSet Attribute =
1994 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1995 Attribute::NoUnwind);
1997 M->getOrInsertFunction(
1999 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
2002 return AutoreleaseCallee;
2005 /// Test whether the given value is possible a reference-counted pointer,
2006 /// including tests which utilize AliasAnalysis.
2007 static bool IsPotentialUse(const Value *Op, AliasAnalysis &AA) {
2008 // First make the rudimentary check.
2009 if (!IsPotentialUse(Op))
2012 // Objects in constant memory are not reference-counted.
2013 if (AA.pointsToConstantMemory(Op))
2016 // Pointers in constant memory are not pointing to reference-counted objects.
2017 if (const LoadInst *LI = dyn_cast<LoadInst>(Op))
2018 if (AA.pointsToConstantMemory(LI->getPointerOperand()))
2021 // Otherwise assume the worst.
2025 /// Test whether the given instruction can result in a reference count
2026 /// modification (positive or negative) for the pointer's object.
2028 CanAlterRefCount(const Instruction *Inst, const Value *Ptr,
2029 ProvenanceAnalysis &PA, InstructionClass Class) {
2031 case IC_Autorelease:
2032 case IC_AutoreleaseRV:
2034 // These operations never directly modify a reference count.
2039 ImmutableCallSite CS = static_cast<const Value *>(Inst);
2040 assert(CS && "Only calls can alter reference counts!");
2042 // See if AliasAnalysis can help us with the call.
2043 AliasAnalysis::ModRefBehavior MRB = PA.getAA()->getModRefBehavior(CS);
2044 if (AliasAnalysis::onlyReadsMemory(MRB))
2046 if (AliasAnalysis::onlyAccessesArgPointees(MRB)) {
2047 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
2049 const Value *Op = *I;
2050 if (IsPotentialUse(Op, *PA.getAA()) && PA.related(Ptr, Op))
2056 // Assume the worst.
2060 /// Test whether the given instruction can "use" the given pointer's object in a
2061 /// way that requires the reference count to be positive.
2063 CanUse(const Instruction *Inst, const Value *Ptr, ProvenanceAnalysis &PA,
2064 InstructionClass Class) {
2065 // IC_Call operations (as opposed to IC_CallOrUser) never "use" objc pointers.
2066 if (Class == IC_Call)
2069 // Consider various instructions which may have pointer arguments which are
2071 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(Inst)) {
2072 // Comparing a pointer with null, or any other constant, isn't really a use,
2073 // because we don't care what the pointer points to, or about the values
2074 // of any other dynamic reference-counted pointers.
2075 if (!IsPotentialUse(ICI->getOperand(1), *PA.getAA()))
2077 } else if (ImmutableCallSite CS = static_cast<const Value *>(Inst)) {
2078 // For calls, just check the arguments (and not the callee operand).
2079 for (ImmutableCallSite::arg_iterator OI = CS.arg_begin(),
2080 OE = CS.arg_end(); OI != OE; ++OI) {
2081 const Value *Op = *OI;
2082 if (IsPotentialUse(Op, *PA.getAA()) && PA.related(Ptr, Op))
2086 } else if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
2087 // Special-case stores, because we don't care about the stored value, just
2088 // the store address.
2089 const Value *Op = GetUnderlyingObjCPtr(SI->getPointerOperand());
2090 // If we can't tell what the underlying object was, assume there is a
2092 return IsPotentialUse(Op, *PA.getAA()) && PA.related(Op, Ptr);
2095 // Check each operand for a match.
2096 for (User::const_op_iterator OI = Inst->op_begin(), OE = Inst->op_end();
2098 const Value *Op = *OI;
2099 if (IsPotentialUse(Op, *PA.getAA()) && PA.related(Ptr, Op))
2105 /// Test whether the given instruction can autorelease any pointer or cause an
2106 /// autoreleasepool pop.
2108 CanInterruptRV(InstructionClass Class) {
2110 case IC_AutoreleasepoolPop:
2113 case IC_Autorelease:
2114 case IC_AutoreleaseRV:
2115 case IC_FusedRetainAutorelease:
2116 case IC_FusedRetainAutoreleaseRV:
2124 /// \enum DependenceKind
2125 /// \brief Defines different dependence kinds among various ARC constructs.
2127 /// There are several kinds of dependence-like concepts in use here.
2129 enum DependenceKind {
2130 NeedsPositiveRetainCount,
2131 AutoreleasePoolBoundary,
2132 CanChangeRetainCount,
2133 RetainAutoreleaseDep, ///< Blocks objc_retainAutorelease.
2134 RetainAutoreleaseRVDep, ///< Blocks objc_retainAutoreleaseReturnValue.
2135 RetainRVDep ///< Blocks objc_retainAutoreleasedReturnValue.
2139 /// Test if there can be dependencies on Inst through Arg. This function only
2140 /// tests dependencies relevant for removing pairs of calls.
2142 Depends(DependenceKind Flavor, Instruction *Inst, const Value *Arg,
2143 ProvenanceAnalysis &PA) {
2144 // If we've reached the definition of Arg, stop.
2149 case NeedsPositiveRetainCount: {
2150 InstructionClass Class = GetInstructionClass(Inst);
2152 case IC_AutoreleasepoolPop:
2153 case IC_AutoreleasepoolPush:
2157 return CanUse(Inst, Arg, PA, Class);
2161 case AutoreleasePoolBoundary: {
2162 InstructionClass Class = GetInstructionClass(Inst);
2164 case IC_AutoreleasepoolPop:
2165 case IC_AutoreleasepoolPush:
2166 // These mark the end and begin of an autorelease pool scope.
2169 // Nothing else does this.
2174 case CanChangeRetainCount: {
2175 InstructionClass Class = GetInstructionClass(Inst);
2177 case IC_AutoreleasepoolPop:
2178 // Conservatively assume this can decrement any count.
2180 case IC_AutoreleasepoolPush:
2184 return CanAlterRefCount(Inst, Arg, PA, Class);
2188 case RetainAutoreleaseDep:
2189 switch (GetBasicInstructionClass(Inst)) {
2190 case IC_AutoreleasepoolPop:
2191 case IC_AutoreleasepoolPush:
2192 // Don't merge an objc_autorelease with an objc_retain inside a different
2193 // autoreleasepool scope.
2197 // Check for a retain of the same pointer for merging.
2198 return GetObjCArg(Inst) == Arg;
2200 // Nothing else matters for objc_retainAutorelease formation.
2204 case RetainAutoreleaseRVDep: {
2205 InstructionClass Class = GetBasicInstructionClass(Inst);
2209 // Check for a retain of the same pointer for merging.
2210 return GetObjCArg(Inst) == Arg;
2212 // Anything that can autorelease interrupts
2213 // retainAutoreleaseReturnValue formation.
2214 return CanInterruptRV(Class);
2219 return CanInterruptRV(GetBasicInstructionClass(Inst));
2222 llvm_unreachable("Invalid dependence flavor");
2225 /// Walk up the CFG from StartPos (which is in StartBB) and find local and
2226 /// non-local dependencies on Arg.
2228 /// TODO: Cache results?
2230 FindDependencies(DependenceKind Flavor,
2232 BasicBlock *StartBB, Instruction *StartInst,
2233 SmallPtrSet<Instruction *, 4> &DependingInstructions,
2234 SmallPtrSet<const BasicBlock *, 4> &Visited,
2235 ProvenanceAnalysis &PA) {
2236 BasicBlock::iterator StartPos = StartInst;
2238 SmallVector<std::pair<BasicBlock *, BasicBlock::iterator>, 4> Worklist;
2239 Worklist.push_back(std::make_pair(StartBB, StartPos));
2241 std::pair<BasicBlock *, BasicBlock::iterator> Pair =
2242 Worklist.pop_back_val();
2243 BasicBlock *LocalStartBB = Pair.first;
2244 BasicBlock::iterator LocalStartPos = Pair.second;
2245 BasicBlock::iterator StartBBBegin = LocalStartBB->begin();
2247 if (LocalStartPos == StartBBBegin) {
2248 pred_iterator PI(LocalStartBB), PE(LocalStartBB, false);
2250 // If we've reached the function entry, produce a null dependence.
2251 DependingInstructions.insert(0);
2253 // Add the predecessors to the worklist.
2255 BasicBlock *PredBB = *PI;
2256 if (Visited.insert(PredBB))
2257 Worklist.push_back(std::make_pair(PredBB, PredBB->end()));
2258 } while (++PI != PE);
2262 Instruction *Inst = --LocalStartPos;
2263 if (Depends(Flavor, Inst, Arg, PA)) {
2264 DependingInstructions.insert(Inst);
2268 } while (!Worklist.empty());
2270 // Determine whether the original StartBB post-dominates all of the blocks we
2271 // visited. If not, insert a sentinal indicating that most optimizations are
2273 for (SmallPtrSet<const BasicBlock *, 4>::const_iterator I = Visited.begin(),
2274 E = Visited.end(); I != E; ++I) {
2275 const BasicBlock *BB = *I;
2278 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2279 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
2280 const BasicBlock *Succ = *SI;
2281 if (Succ != StartBB && !Visited.count(Succ)) {
2282 DependingInstructions.insert(reinterpret_cast<Instruction *>(-1));
2289 static bool isNullOrUndef(const Value *V) {
2290 return isa<ConstantPointerNull>(V) || isa<UndefValue>(V);
2293 static bool isNoopInstruction(const Instruction *I) {
2294 return isa<BitCastInst>(I) ||
2295 (isa<GetElementPtrInst>(I) &&
2296 cast<GetElementPtrInst>(I)->hasAllZeroIndices());
2299 /// Turn objc_retain into objc_retainAutoreleasedReturnValue if the operand is a
2302 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
2303 ImmutableCallSite CS(GetObjCArg(Retain));
2304 const Instruction *Call = CS.getInstruction();
2306 if (Call->getParent() != Retain->getParent()) return;
2308 // Check that the call is next to the retain.
2309 BasicBlock::const_iterator I = Call;
2311 while (isNoopInstruction(I)) ++I;
2315 // Turn it to an objc_retainAutoreleasedReturnValue..
2319 DEBUG(dbgs() << "ObjCARCOpt::OptimizeRetainCall: Transforming "
2320 "objc_retain => objc_retainAutoreleasedReturnValue"
2321 " since the operand is a return value.\n"
2323 << *Retain << "\n");
2325 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
2327 DEBUG(dbgs() << " New: "
2328 << *Retain << "\n");
2331 /// Turn objc_retainAutoreleasedReturnValue into objc_retain if the operand is
2332 /// not a return value. Or, if it can be paired with an
2333 /// objc_autoreleaseReturnValue, delete the pair and return true.
2335 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
2336 // Check for the argument being from an immediately preceding call or invoke.
2337 const Value *Arg = GetObjCArg(RetainRV);
2338 ImmutableCallSite CS(Arg);
2339 if (const Instruction *Call = CS.getInstruction()) {
2340 if (Call->getParent() == RetainRV->getParent()) {
2341 BasicBlock::const_iterator I = Call;
2343 while (isNoopInstruction(I)) ++I;
2344 if (&*I == RetainRV)
2346 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
2347 BasicBlock *RetainRVParent = RetainRV->getParent();
2348 if (II->getNormalDest() == RetainRVParent) {
2349 BasicBlock::const_iterator I = RetainRVParent->begin();
2350 while (isNoopInstruction(I)) ++I;
2351 if (&*I == RetainRV)
2357 // Check for being preceded by an objc_autoreleaseReturnValue on the same
2358 // pointer. In this case, we can delete the pair.
2359 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
2361 do --I; while (I != Begin && isNoopInstruction(I));
2362 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
2363 GetObjCArg(I) == Arg) {
2367 DEBUG(dbgs() << "ObjCARCOpt::OptimizeRetainRVCall: Erasing " << *I << "\n"
2368 << " Erasing " << *RetainRV
2371 EraseInstruction(I);
2372 EraseInstruction(RetainRV);
2377 // Turn it to a plain objc_retain.
2381 DEBUG(dbgs() << "ObjCARCOpt::OptimizeRetainRVCall: Transforming "
2382 "objc_retainAutoreleasedReturnValue => "
2383 "objc_retain since the operand is not a return value.\n"
2385 << *RetainRV << "\n");
2387 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
2389 DEBUG(dbgs() << " New: "
2390 << *RetainRV << "\n");
2395 /// Turn objc_autoreleaseReturnValue into objc_autorelease if the result is not
2396 /// used as a return value.
2398 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
2399 InstructionClass &Class) {
2400 // Check for a return of the pointer value.
2401 const Value *Ptr = GetObjCArg(AutoreleaseRV);
2402 SmallVector<const Value *, 2> Users;
2403 Users.push_back(Ptr);
2405 Ptr = Users.pop_back_val();
2406 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
2408 const User *I = *UI;
2409 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
2411 if (isa<BitCastInst>(I))
2414 } while (!Users.empty());
2419 DEBUG(dbgs() << "ObjCARCOpt::OptimizeAutoreleaseRVCall: Transforming "
2420 "objc_autoreleaseReturnValue => "
2421 "objc_autorelease since its operand is not used as a return "
2424 << *AutoreleaseRV << "\n");
2426 CallInst *AutoreleaseRVCI = cast<CallInst>(AutoreleaseRV);
2428 setCalledFunction(getAutoreleaseCallee(F.getParent()));
2429 AutoreleaseRVCI->setTailCall(false); // Never tail call objc_autorelease.
2430 Class = IC_Autorelease;
2432 DEBUG(dbgs() << " New: "
2433 << *AutoreleaseRV << "\n");
2437 /// Visit each call, one at a time, and make simplifications without doing any
2438 /// additional analysis.
2439 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
2440 // Reset all the flags in preparation for recomputing them.
2441 UsedInThisFunction = 0;
2443 // Visit all objc_* calls in F.
2444 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2445 Instruction *Inst = &*I++;
2447 InstructionClass Class = GetBasicInstructionClass(Inst);
2449 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Visiting: Class: "
2450 << Class << "; " << *Inst << "\n");
2455 // Delete no-op casts. These function calls have special semantics, but
2456 // the semantics are entirely implemented via lowering in the front-end,
2457 // so by the time they reach the optimizer, they are just no-op calls
2458 // which return their argument.
2460 // There are gray areas here, as the ability to cast reference-counted
2461 // pointers to raw void* and back allows code to break ARC assumptions,
2462 // however these are currently considered to be unimportant.
2466 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Erasing no-op cast:"
2467 " " << *Inst << "\n");
2468 EraseInstruction(Inst);
2471 // If the pointer-to-weak-pointer is null, it's undefined behavior.
2474 case IC_LoadWeakRetained:
2476 case IC_DestroyWeak: {
2477 CallInst *CI = cast<CallInst>(Inst);
2478 if (isNullOrUndef(CI->getArgOperand(0))) {
2480 Type *Ty = CI->getArgOperand(0)->getType();
2481 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2482 Constant::getNullValue(Ty),
2484 llvm::Value *NewValue = UndefValue::get(CI->getType());
2485 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: A null "
2486 "pointer-to-weak-pointer is undefined behavior.\n"
2490 CI->replaceAllUsesWith(NewValue);
2491 CI->eraseFromParent();
2498 CallInst *CI = cast<CallInst>(Inst);
2499 if (isNullOrUndef(CI->getArgOperand(0)) ||
2500 isNullOrUndef(CI->getArgOperand(1))) {
2502 Type *Ty = CI->getArgOperand(0)->getType();
2503 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2504 Constant::getNullValue(Ty),
2507 llvm::Value *NewValue = UndefValue::get(CI->getType());
2508 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: A null "
2509 "pointer-to-weak-pointer is undefined behavior.\n"
2514 CI->replaceAllUsesWith(NewValue);
2515 CI->eraseFromParent();
2521 OptimizeRetainCall(F, Inst);
2524 if (OptimizeRetainRVCall(F, Inst))
2527 case IC_AutoreleaseRV:
2528 OptimizeAutoreleaseRVCall(F, Inst, Class);
2532 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
2533 if (IsAutorelease(Class) && Inst->use_empty()) {
2534 CallInst *Call = cast<CallInst>(Inst);
2535 const Value *Arg = Call->getArgOperand(0);
2536 Arg = FindSingleUseIdentifiedObject(Arg);
2541 // Create the declaration lazily.
2542 LLVMContext &C = Inst->getContext();
2544 CallInst::Create(getReleaseCallee(F.getParent()),
2545 Call->getArgOperand(0), "", Call);
2546 NewCall->setMetadata(ImpreciseReleaseMDKind,
2547 MDNode::get(C, ArrayRef<Value *>()));
2549 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Replacing "
2550 "objc_autorelease(x) with objc_release(x) since x is "
2551 "otherwise unused.\n"
2552 " Old: " << *Call <<
2556 EraseInstruction(Call);
2562 // For functions which can never be passed stack arguments, add
2564 if (IsAlwaysTail(Class)) {
2566 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Adding tail keyword"
2567 " to function since it can never be passed stack args: " << *Inst <<
2569 cast<CallInst>(Inst)->setTailCall();
2572 // Ensure that functions that can never have a "tail" keyword due to the
2573 // semantics of ARC truly do not do so.
2574 if (IsNeverTail(Class)) {
2576 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Removing tail "
2577 "keyword from function: " << *Inst <<
2579 cast<CallInst>(Inst)->setTailCall(false);
2582 // Set nounwind as needed.
2583 if (IsNoThrow(Class)) {
2585 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Found no throw"
2586 " class. Setting nounwind on: " << *Inst << "\n");
2587 cast<CallInst>(Inst)->setDoesNotThrow();
2590 if (!IsNoopOnNull(Class)) {
2591 UsedInThisFunction |= 1 << Class;
2595 const Value *Arg = GetObjCArg(Inst);
2597 // ARC calls with null are no-ops. Delete them.
2598 if (isNullOrUndef(Arg)) {
2601 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: ARC calls with "
2602 " null are no-ops. Erasing: " << *Inst << "\n");
2603 EraseInstruction(Inst);
2607 // Keep track of which of retain, release, autorelease, and retain_block
2608 // are actually present in this function.
2609 UsedInThisFunction |= 1 << Class;
2611 // If Arg is a PHI, and one or more incoming values to the
2612 // PHI are null, and the call is control-equivalent to the PHI, and there
2613 // are no relevant side effects between the PHI and the call, the call
2614 // could be pushed up to just those paths with non-null incoming values.
2615 // For now, don't bother splitting critical edges for this.
2616 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
2617 Worklist.push_back(std::make_pair(Inst, Arg));
2619 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
2623 const PHINode *PN = dyn_cast<PHINode>(Arg);
2626 // Determine if the PHI has any null operands, or any incoming
2628 bool HasNull = false;
2629 bool HasCriticalEdges = false;
2630 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2632 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2633 if (isNullOrUndef(Incoming))
2635 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
2636 .getNumSuccessors() != 1) {
2637 HasCriticalEdges = true;
2641 // If we have null operands and no critical edges, optimize.
2642 if (!HasCriticalEdges && HasNull) {
2643 SmallPtrSet<Instruction *, 4> DependingInstructions;
2644 SmallPtrSet<const BasicBlock *, 4> Visited;
2646 // Check that there is nothing that cares about the reference
2647 // count between the call and the phi.
2650 case IC_RetainBlock:
2651 // These can always be moved up.
2654 // These can't be moved across things that care about the retain
2656 FindDependencies(NeedsPositiveRetainCount, Arg,
2657 Inst->getParent(), Inst,
2658 DependingInstructions, Visited, PA);
2660 case IC_Autorelease:
2661 // These can't be moved across autorelease pool scope boundaries.
2662 FindDependencies(AutoreleasePoolBoundary, Arg,
2663 Inst->getParent(), Inst,
2664 DependingInstructions, Visited, PA);
2667 case IC_AutoreleaseRV:
2668 // Don't move these; the RV optimization depends on the autoreleaseRV
2669 // being tail called, and the retainRV being immediately after a call
2670 // (which might still happen if we get lucky with codegen layout, but
2671 // it's not worth taking the chance).
2674 llvm_unreachable("Invalid dependence flavor");
2677 if (DependingInstructions.size() == 1 &&
2678 *DependingInstructions.begin() == PN) {
2681 // Clone the call into each predecessor that has a non-null value.
2682 CallInst *CInst = cast<CallInst>(Inst);
2683 Type *ParamTy = CInst->getArgOperand(0)->getType();
2684 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2686 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2687 if (!isNullOrUndef(Incoming)) {
2688 CallInst *Clone = cast<CallInst>(CInst->clone());
2689 Value *Op = PN->getIncomingValue(i);
2690 Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
2691 if (Op->getType() != ParamTy)
2692 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
2693 Clone->setArgOperand(0, Op);
2694 Clone->insertBefore(InsertPos);
2696 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Cloning "
2699 "clone at " << *InsertPos << "\n");
2700 Worklist.push_back(std::make_pair(Clone, Incoming));
2703 // Erase the original call.
2704 DEBUG(dbgs() << "Erasing: " << *CInst << "\n");
2705 EraseInstruction(CInst);
2709 } while (!Worklist.empty());
2711 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Finished List.\n");
2714 /// Check for critical edges, loop boundaries, irreducible control flow, or
2715 /// other CFG structures where moving code across the edge would result in it
2716 /// being executed more.
2718 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
2719 DenseMap<const BasicBlock *, BBState> &BBStates,
2720 BBState &MyStates) const {
2721 // If any top-down local-use or possible-dec has a succ which is earlier in
2722 // the sequence, forget it.
2723 for (BBState::ptr_iterator I = MyStates.top_down_ptr_begin(),
2724 E = MyStates.top_down_ptr_end(); I != E; ++I)
2725 switch (I->second.GetSeq()) {
2728 const Value *Arg = I->first;
2729 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2730 bool SomeSuccHasSame = false;
2731 bool AllSuccsHaveSame = true;
2732 PtrState &S = I->second;
2733 succ_const_iterator SI(TI), SE(TI, false);
2735 for (; SI != SE; ++SI) {
2736 Sequence SuccSSeq = S_None;
2737 bool SuccSRRIKnownSafe = false;
2738 // If VisitBottomUp has pointer information for this successor, take
2739 // what we know about it.
2740 DenseMap<const BasicBlock *, BBState>::iterator BBI =
2742 assert(BBI != BBStates.end());
2743 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
2744 SuccSSeq = SuccS.GetSeq();
2745 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
2748 case S_CanRelease: {
2749 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
2750 S.ClearSequenceProgress();
2756 SomeSuccHasSame = true;
2760 case S_MovableRelease:
2761 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
2762 AllSuccsHaveSame = false;
2765 llvm_unreachable("bottom-up pointer in retain state!");
2768 // If the state at the other end of any of the successor edges
2769 // matches the current state, require all edges to match. This
2770 // guards against loops in the middle of a sequence.
2771 if (SomeSuccHasSame && !AllSuccsHaveSame)
2772 S.ClearSequenceProgress();
2775 case S_CanRelease: {
2776 const Value *Arg = I->first;
2777 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2778 bool SomeSuccHasSame = false;
2779 bool AllSuccsHaveSame = true;
2780 PtrState &S = I->second;
2781 succ_const_iterator SI(TI), SE(TI, false);
2783 for (; SI != SE; ++SI) {
2784 Sequence SuccSSeq = S_None;
2785 bool SuccSRRIKnownSafe = false;
2786 // If VisitBottomUp has pointer information for this successor, take
2787 // what we know about it.
2788 DenseMap<const BasicBlock *, BBState>::iterator BBI =
2790 assert(BBI != BBStates.end());
2791 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
2792 SuccSSeq = SuccS.GetSeq();
2793 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
2796 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
2797 S.ClearSequenceProgress();
2803 SomeSuccHasSame = true;
2807 case S_MovableRelease:
2809 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
2810 AllSuccsHaveSame = false;
2813 llvm_unreachable("bottom-up pointer in retain state!");
2816 // If the state at the other end of any of the successor edges
2817 // matches the current state, require all edges to match. This
2818 // guards against loops in the middle of a sequence.
2819 if (SomeSuccHasSame && !AllSuccsHaveSame)
2820 S.ClearSequenceProgress();
2827 ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst,
2829 MapVector<Value *, RRInfo> &Retains,
2830 BBState &MyStates) {
2831 bool NestingDetected = false;
2832 InstructionClass Class = GetInstructionClass(Inst);
2833 const Value *Arg = 0;
2837 Arg = GetObjCArg(Inst);
2839 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2841 // If we see two releases in a row on the same pointer. If so, make
2842 // a note, and we'll cicle back to revisit it after we've
2843 // hopefully eliminated the second release, which may allow us to
2844 // eliminate the first release too.
2845 // Theoretically we could implement removal of nested retain+release
2846 // pairs by making PtrState hold a stack of states, but this is
2847 // simple and avoids adding overhead for the non-nested case.
2848 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease) {
2849 DEBUG(dbgs() << "ObjCARCOpt::VisitInstructionBottomUp: Found nested "
2850 "releases (i.e. a release pair)\n");
2851 NestingDetected = true;
2854 MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2855 S.ResetSequenceProgress(ReleaseMetadata ? S_MovableRelease : S_Release);
2856 S.RRI.ReleaseMetadata = ReleaseMetadata;
2857 S.RRI.KnownSafe = S.IsKnownIncremented();
2858 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2859 S.RRI.Calls.insert(Inst);
2861 S.SetKnownPositiveRefCount();
2864 case IC_RetainBlock:
2865 // An objc_retainBlock call with just a use may need to be kept,
2866 // because it may be copying a block from the stack to the heap.
2867 if (!IsRetainBlockOptimizable(Inst))
2872 Arg = GetObjCArg(Inst);
2874 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2875 S.SetKnownPositiveRefCount();
2877 switch (S.GetSeq()) {
2880 case S_MovableRelease:
2882 S.RRI.ReverseInsertPts.clear();
2885 // Don't do retain+release tracking for IC_RetainRV, because it's
2886 // better to let it remain as the first instruction after a call.
2887 if (Class != IC_RetainRV) {
2888 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2889 Retains[Inst] = S.RRI;
2891 S.ClearSequenceProgress();
2896 llvm_unreachable("bottom-up pointer in retain state!");
2898 return NestingDetected;
2900 case IC_AutoreleasepoolPop:
2901 // Conservatively, clear MyStates for all known pointers.
2902 MyStates.clearBottomUpPointers();
2903 return NestingDetected;
2904 case IC_AutoreleasepoolPush:
2906 // These are irrelevant.
2907 return NestingDetected;
2912 // Consider any other possible effects of this instruction on each
2913 // pointer being tracked.
2914 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
2915 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
2916 const Value *Ptr = MI->first;
2918 continue; // Handled above.
2919 PtrState &S = MI->second;
2920 Sequence Seq = S.GetSeq();
2922 // Check for possible releases.
2923 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2927 S.SetSeq(S_CanRelease);
2931 case S_MovableRelease:
2936 llvm_unreachable("bottom-up pointer in retain state!");
2940 // Check for possible direct uses.
2943 case S_MovableRelease:
2944 if (CanUse(Inst, Ptr, PA, Class)) {
2945 assert(S.RRI.ReverseInsertPts.empty());
2946 // If this is an invoke instruction, we're scanning it as part of
2947 // one of its successor blocks, since we can't insert code after it
2948 // in its own block, and we don't want to split critical edges.
2949 if (isa<InvokeInst>(Inst))
2950 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
2952 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
2954 } else if (Seq == S_Release &&
2955 (Class == IC_User || Class == IC_CallOrUser)) {
2956 // Non-movable releases depend on any possible objc pointer use.
2958 assert(S.RRI.ReverseInsertPts.empty());
2959 // As above; handle invoke specially.
2960 if (isa<InvokeInst>(Inst))
2961 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
2963 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
2967 if (CanUse(Inst, Ptr, PA, Class))
2975 llvm_unreachable("bottom-up pointer in retain state!");
2979 return NestingDetected;
2983 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
2984 DenseMap<const BasicBlock *, BBState> &BBStates,
2985 MapVector<Value *, RRInfo> &Retains) {
2986 bool NestingDetected = false;
2987 BBState &MyStates = BBStates[BB];
2989 // Merge the states from each successor to compute the initial state
2990 // for the current block.
2991 BBState::edge_iterator SI(MyStates.succ_begin()),
2992 SE(MyStates.succ_end());
2994 const BasicBlock *Succ = *SI;
2995 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
2996 assert(I != BBStates.end());
2997 MyStates.InitFromSucc(I->second);
2999 for (; SI != SE; ++SI) {
3001 I = BBStates.find(Succ);
3002 assert(I != BBStates.end());
3003 MyStates.MergeSucc(I->second);
3007 // Visit all the instructions, bottom-up.
3008 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
3009 Instruction *Inst = llvm::prior(I);
3011 // Invoke instructions are visited as part of their successors (below).
3012 if (isa<InvokeInst>(Inst))
3015 DEBUG(dbgs() << "ObjCARCOpt::VisitButtonUp: Visiting " << *Inst << "\n");
3017 NestingDetected |= VisitInstructionBottomUp(Inst, BB, Retains, MyStates);
3020 // If there's a predecessor with an invoke, visit the invoke as if it were
3021 // part of this block, since we can't insert code after an invoke in its own
3022 // block, and we don't want to split critical edges.
3023 for (BBState::edge_iterator PI(MyStates.pred_begin()),
3024 PE(MyStates.pred_end()); PI != PE; ++PI) {
3025 BasicBlock *Pred = *PI;
3026 if (InvokeInst *II = dyn_cast<InvokeInst>(&Pred->back()))
3027 NestingDetected |= VisitInstructionBottomUp(II, BB, Retains, MyStates);
3030 return NestingDetected;
3034 ObjCARCOpt::VisitInstructionTopDown(Instruction *Inst,
3035 DenseMap<Value *, RRInfo> &Releases,
3036 BBState &MyStates) {
3037 bool NestingDetected = false;
3038 InstructionClass Class = GetInstructionClass(Inst);
3039 const Value *Arg = 0;
3042 case IC_RetainBlock:
3043 // An objc_retainBlock call with just a use may need to be kept,
3044 // because it may be copying a block from the stack to the heap.
3045 if (!IsRetainBlockOptimizable(Inst))
3050 Arg = GetObjCArg(Inst);
3052 PtrState &S = MyStates.getPtrTopDownState(Arg);
3054 // Don't do retain+release tracking for IC_RetainRV, because it's
3055 // better to let it remain as the first instruction after a call.
3056 if (Class != IC_RetainRV) {
3057 // If we see two retains in a row on the same pointer. If so, make
3058 // a note, and we'll cicle back to revisit it after we've
3059 // hopefully eliminated the second retain, which may allow us to
3060 // eliminate the first retain too.
3061 // Theoretically we could implement removal of nested retain+release
3062 // pairs by making PtrState hold a stack of states, but this is
3063 // simple and avoids adding overhead for the non-nested case.
3064 if (S.GetSeq() == S_Retain)
3065 NestingDetected = true;
3067 S.ResetSequenceProgress(S_Retain);
3068 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
3069 S.RRI.KnownSafe = S.IsKnownIncremented();
3070 S.RRI.Calls.insert(Inst);
3073 S.SetKnownPositiveRefCount();
3075 // A retain can be a potential use; procede to the generic checking
3080 Arg = GetObjCArg(Inst);
3082 PtrState &S = MyStates.getPtrTopDownState(Arg);
3085 switch (S.GetSeq()) {
3088 S.RRI.ReverseInsertPts.clear();
3091 S.RRI.ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
3092 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
3093 Releases[Inst] = S.RRI;
3094 S.ClearSequenceProgress();
3100 case S_MovableRelease:
3101 llvm_unreachable("top-down pointer in release state!");
3105 case IC_AutoreleasepoolPop:
3106 // Conservatively, clear MyStates for all known pointers.
3107 MyStates.clearTopDownPointers();
3108 return NestingDetected;
3109 case IC_AutoreleasepoolPush:
3111 // These are irrelevant.
3112 return NestingDetected;
3117 // Consider any other possible effects of this instruction on each
3118 // pointer being tracked.
3119 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
3120 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
3121 const Value *Ptr = MI->first;
3123 continue; // Handled above.
3124 PtrState &S = MI->second;
3125 Sequence Seq = S.GetSeq();
3127 // Check for possible releases.
3128 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
3132 S.SetSeq(S_CanRelease);
3133 assert(S.RRI.ReverseInsertPts.empty());
3134 S.RRI.ReverseInsertPts.insert(Inst);
3136 // One call can't cause a transition from S_Retain to S_CanRelease
3137 // and S_CanRelease to S_Use. If we've made the first transition,
3146 case S_MovableRelease:
3147 llvm_unreachable("top-down pointer in release state!");
3151 // Check for possible direct uses.
3154 if (CanUse(Inst, Ptr, PA, Class))
3163 case S_MovableRelease:
3164 llvm_unreachable("top-down pointer in release state!");
3168 return NestingDetected;
3172 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
3173 DenseMap<const BasicBlock *, BBState> &BBStates,
3174 DenseMap<Value *, RRInfo> &Releases) {
3175 bool NestingDetected = false;
3176 BBState &MyStates = BBStates[BB];
3178 // Merge the states from each predecessor to compute the initial state
3179 // for the current block.
3180 BBState::edge_iterator PI(MyStates.pred_begin()),
3181 PE(MyStates.pred_end());
3183 const BasicBlock *Pred = *PI;
3184 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
3185 assert(I != BBStates.end());
3186 MyStates.InitFromPred(I->second);
3188 for (; PI != PE; ++PI) {
3190 I = BBStates.find(Pred);
3191 assert(I != BBStates.end());
3192 MyStates.MergePred(I->second);
3196 // Visit all the instructions, top-down.
3197 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3198 Instruction *Inst = I;
3200 DEBUG(dbgs() << "ObjCARCOpt::VisitTopDown: Visiting " << *Inst << "\n");
3202 NestingDetected |= VisitInstructionTopDown(Inst, Releases, MyStates);
3205 CheckForCFGHazards(BB, BBStates, MyStates);
3206 return NestingDetected;
3210 ComputePostOrders(Function &F,
3211 SmallVectorImpl<BasicBlock *> &PostOrder,
3212 SmallVectorImpl<BasicBlock *> &ReverseCFGPostOrder,
3213 unsigned NoObjCARCExceptionsMDKind,
3214 DenseMap<const BasicBlock *, BBState> &BBStates) {
3215 /// The visited set, for doing DFS walks.
3216 SmallPtrSet<BasicBlock *, 16> Visited;
3218 // Do DFS, computing the PostOrder.
3219 SmallPtrSet<BasicBlock *, 16> OnStack;
3220 SmallVector<std::pair<BasicBlock *, succ_iterator>, 16> SuccStack;
3222 // Functions always have exactly one entry block, and we don't have
3223 // any other block that we treat like an entry block.
3224 BasicBlock *EntryBB = &F.getEntryBlock();
3225 BBState &MyStates = BBStates[EntryBB];
3226 MyStates.SetAsEntry();
3227 TerminatorInst *EntryTI = cast<TerminatorInst>(&EntryBB->back());
3228 SuccStack.push_back(std::make_pair(EntryBB, succ_iterator(EntryTI)));
3229 Visited.insert(EntryBB);
3230 OnStack.insert(EntryBB);
3233 BasicBlock *CurrBB = SuccStack.back().first;
3234 TerminatorInst *TI = cast<TerminatorInst>(&CurrBB->back());
3235 succ_iterator SE(TI, false);
3237 while (SuccStack.back().second != SE) {
3238 BasicBlock *SuccBB = *SuccStack.back().second++;
3239 if (Visited.insert(SuccBB)) {
3240 TerminatorInst *TI = cast<TerminatorInst>(&SuccBB->back());
3241 SuccStack.push_back(std::make_pair(SuccBB, succ_iterator(TI)));
3242 BBStates[CurrBB].addSucc(SuccBB);
3243 BBState &SuccStates = BBStates[SuccBB];
3244 SuccStates.addPred(CurrBB);
3245 OnStack.insert(SuccBB);
3249 if (!OnStack.count(SuccBB)) {
3250 BBStates[CurrBB].addSucc(SuccBB);
3251 BBStates[SuccBB].addPred(CurrBB);
3254 OnStack.erase(CurrBB);
3255 PostOrder.push_back(CurrBB);
3256 SuccStack.pop_back();
3257 } while (!SuccStack.empty());
3261 // Do reverse-CFG DFS, computing the reverse-CFG PostOrder.
3262 // Functions may have many exits, and there also blocks which we treat
3263 // as exits due to ignored edges.
3264 SmallVector<std::pair<BasicBlock *, BBState::edge_iterator>, 16> PredStack;
3265 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
3266 BasicBlock *ExitBB = I;
3267 BBState &MyStates = BBStates[ExitBB];
3268 if (!MyStates.isExit())
3271 MyStates.SetAsExit();
3273 PredStack.push_back(std::make_pair(ExitBB, MyStates.pred_begin()));
3274 Visited.insert(ExitBB);
3275 while (!PredStack.empty()) {
3276 reverse_dfs_next_succ:
3277 BBState::edge_iterator PE = BBStates[PredStack.back().first].pred_end();
3278 while (PredStack.back().second != PE) {
3279 BasicBlock *BB = *PredStack.back().second++;
3280 if (Visited.insert(BB)) {
3281 PredStack.push_back(std::make_pair(BB, BBStates[BB].pred_begin()));
3282 goto reverse_dfs_next_succ;
3285 ReverseCFGPostOrder.push_back(PredStack.pop_back_val().first);
3290 // Visit the function both top-down and bottom-up.
3292 ObjCARCOpt::Visit(Function &F,
3293 DenseMap<const BasicBlock *, BBState> &BBStates,
3294 MapVector<Value *, RRInfo> &Retains,
3295 DenseMap<Value *, RRInfo> &Releases) {
3297 // Use reverse-postorder traversals, because we magically know that loops
3298 // will be well behaved, i.e. they won't repeatedly call retain on a single
3299 // pointer without doing a release. We can't use the ReversePostOrderTraversal
3300 // class here because we want the reverse-CFG postorder to consider each
3301 // function exit point, and we want to ignore selected cycle edges.
3302 SmallVector<BasicBlock *, 16> PostOrder;
3303 SmallVector<BasicBlock *, 16> ReverseCFGPostOrder;
3304 ComputePostOrders(F, PostOrder, ReverseCFGPostOrder,
3305 NoObjCARCExceptionsMDKind,
3308 // Use reverse-postorder on the reverse CFG for bottom-up.
3309 bool BottomUpNestingDetected = false;
3310 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
3311 ReverseCFGPostOrder.rbegin(), E = ReverseCFGPostOrder.rend();
3313 BottomUpNestingDetected |= VisitBottomUp(*I, BBStates, Retains);
3315 // Use reverse-postorder for top-down.
3316 bool TopDownNestingDetected = false;
3317 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
3318 PostOrder.rbegin(), E = PostOrder.rend();
3320 TopDownNestingDetected |= VisitTopDown(*I, BBStates, Releases);
3322 return TopDownNestingDetected && BottomUpNestingDetected;
3325 /// Move the calls in RetainsToMove and ReleasesToMove.
3326 void ObjCARCOpt::MoveCalls(Value *Arg,
3327 RRInfo &RetainsToMove,
3328 RRInfo &ReleasesToMove,
3329 MapVector<Value *, RRInfo> &Retains,
3330 DenseMap<Value *, RRInfo> &Releases,
3331 SmallVectorImpl<Instruction *> &DeadInsts,
3333 Type *ArgTy = Arg->getType();
3334 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
3336 // Insert the new retain and release calls.
3337 for (SmallPtrSet<Instruction *, 2>::const_iterator
3338 PI = ReleasesToMove.ReverseInsertPts.begin(),
3339 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
3340 Instruction *InsertPt = *PI;
3341 Value *MyArg = ArgTy == ParamTy ? Arg :
3342 new BitCastInst(Arg, ParamTy, "", InsertPt);
3344 CallInst::Create(RetainsToMove.IsRetainBlock ?
3345 getRetainBlockCallee(M) : getRetainCallee(M),
3346 MyArg, "", InsertPt);
3347 Call->setDoesNotThrow();
3348 if (RetainsToMove.IsRetainBlock)
3349 Call->setMetadata(CopyOnEscapeMDKind,
3350 MDNode::get(M->getContext(), ArrayRef<Value *>()));
3352 Call->setTailCall();
3354 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Inserting new Release: " << *Call
3356 " At insertion point: " << *InsertPt
3359 for (SmallPtrSet<Instruction *, 2>::const_iterator
3360 PI = RetainsToMove.ReverseInsertPts.begin(),
3361 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
3362 Instruction *InsertPt = *PI;
3363 Value *MyArg = ArgTy == ParamTy ? Arg :
3364 new BitCastInst(Arg, ParamTy, "", InsertPt);
3365 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
3367 // Attach a clang.imprecise_release metadata tag, if appropriate.
3368 if (MDNode *M = ReleasesToMove.ReleaseMetadata)
3369 Call->setMetadata(ImpreciseReleaseMDKind, M);
3370 Call->setDoesNotThrow();
3371 if (ReleasesToMove.IsTailCallRelease)
3372 Call->setTailCall();
3374 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Inserting new Retain: " << *Call
3376 " At insertion point: " << *InsertPt
3380 // Delete the original retain and release calls.
3381 for (SmallPtrSet<Instruction *, 2>::const_iterator
3382 AI = RetainsToMove.Calls.begin(),
3383 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
3384 Instruction *OrigRetain = *AI;
3385 Retains.blot(OrigRetain);
3386 DeadInsts.push_back(OrigRetain);
3387 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Deleting retain: " << *OrigRetain <<
3390 for (SmallPtrSet<Instruction *, 2>::const_iterator
3391 AI = ReleasesToMove.Calls.begin(),
3392 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
3393 Instruction *OrigRelease = *AI;
3394 Releases.erase(OrigRelease);
3395 DeadInsts.push_back(OrigRelease);
3396 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Deleting release: " << *OrigRelease
3401 /// Identify pairings between the retains and releases, and delete and/or move
3404 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
3406 MapVector<Value *, RRInfo> &Retains,
3407 DenseMap<Value *, RRInfo> &Releases,
3409 bool AnyPairsCompletelyEliminated = false;
3410 RRInfo RetainsToMove;
3411 RRInfo ReleasesToMove;
3412 SmallVector<Instruction *, 4> NewRetains;
3413 SmallVector<Instruction *, 4> NewReleases;
3414 SmallVector<Instruction *, 8> DeadInsts;
3416 // Visit each retain.
3417 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
3418 E = Retains.end(); I != E; ++I) {
3419 Value *V = I->first;
3420 if (!V) continue; // blotted
3422 Instruction *Retain = cast<Instruction>(V);
3424 DEBUG(dbgs() << "ObjCARCOpt::PerformCodePlacement: Visiting: " << *Retain
3427 Value *Arg = GetObjCArg(Retain);
3429 // If the object being released is in static or stack storage, we know it's
3430 // not being managed by ObjC reference counting, so we can delete pairs
3431 // regardless of what possible decrements or uses lie between them.
3432 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg);
3434 // A constant pointer can't be pointing to an object on the heap. It may
3435 // be reference-counted, but it won't be deleted.
3436 if (const LoadInst *LI = dyn_cast<LoadInst>(Arg))
3437 if (const GlobalVariable *GV =
3438 dyn_cast<GlobalVariable>(
3439 StripPointerCastsAndObjCCalls(LI->getPointerOperand())))
3440 if (GV->isConstant())
3443 // If a pair happens in a region where it is known that the reference count
3444 // is already incremented, we can similarly ignore possible decrements.
3445 bool KnownSafeTD = true, KnownSafeBU = true;
3447 // Connect the dots between the top-down-collected RetainsToMove and
3448 // bottom-up-collected ReleasesToMove to form sets of related calls.
3449 // This is an iterative process so that we connect multiple releases
3450 // to multiple retains if needed.
3451 unsigned OldDelta = 0;
3452 unsigned NewDelta = 0;
3453 unsigned OldCount = 0;
3454 unsigned NewCount = 0;
3455 bool FirstRelease = true;
3456 bool FirstRetain = true;
3457 NewRetains.push_back(Retain);
3459 for (SmallVectorImpl<Instruction *>::const_iterator
3460 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
3461 Instruction *NewRetain = *NI;
3462 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
3463 assert(It != Retains.end());
3464 const RRInfo &NewRetainRRI = It->second;
3465 KnownSafeTD &= NewRetainRRI.KnownSafe;
3466 for (SmallPtrSet<Instruction *, 2>::const_iterator
3467 LI = NewRetainRRI.Calls.begin(),
3468 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
3469 Instruction *NewRetainRelease = *LI;
3470 DenseMap<Value *, RRInfo>::const_iterator Jt =
3471 Releases.find(NewRetainRelease);
3472 if (Jt == Releases.end())
3474 const RRInfo &NewRetainReleaseRRI = Jt->second;
3475 assert(NewRetainReleaseRRI.Calls.count(NewRetain));
3476 if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
3478 BBStates[NewRetainRelease->getParent()].GetAllPathCount();
3480 // Merge the ReleaseMetadata and IsTailCallRelease values.
3482 ReleasesToMove.ReleaseMetadata =
3483 NewRetainReleaseRRI.ReleaseMetadata;
3484 ReleasesToMove.IsTailCallRelease =
3485 NewRetainReleaseRRI.IsTailCallRelease;
3486 FirstRelease = false;
3488 if (ReleasesToMove.ReleaseMetadata !=
3489 NewRetainReleaseRRI.ReleaseMetadata)
3490 ReleasesToMove.ReleaseMetadata = 0;
3491 if (ReleasesToMove.IsTailCallRelease !=
3492 NewRetainReleaseRRI.IsTailCallRelease)
3493 ReleasesToMove.IsTailCallRelease = false;
3496 // Collect the optimal insertion points.
3498 for (SmallPtrSet<Instruction *, 2>::const_iterator
3499 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
3500 RE = NewRetainReleaseRRI.ReverseInsertPts.end();
3502 Instruction *RIP = *RI;
3503 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
3504 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
3506 NewReleases.push_back(NewRetainRelease);
3511 if (NewReleases.empty()) break;
3513 // Back the other way.
3514 for (SmallVectorImpl<Instruction *>::const_iterator
3515 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
3516 Instruction *NewRelease = *NI;
3517 DenseMap<Value *, RRInfo>::const_iterator It =
3518 Releases.find(NewRelease);
3519 assert(It != Releases.end());
3520 const RRInfo &NewReleaseRRI = It->second;
3521 KnownSafeBU &= NewReleaseRRI.KnownSafe;
3522 for (SmallPtrSet<Instruction *, 2>::const_iterator
3523 LI = NewReleaseRRI.Calls.begin(),
3524 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
3525 Instruction *NewReleaseRetain = *LI;
3526 MapVector<Value *, RRInfo>::const_iterator Jt =
3527 Retains.find(NewReleaseRetain);
3528 if (Jt == Retains.end())
3530 const RRInfo &NewReleaseRetainRRI = Jt->second;
3531 assert(NewReleaseRetainRRI.Calls.count(NewRelease));
3532 if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
3533 unsigned PathCount =
3534 BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
3535 OldDelta += PathCount;
3536 OldCount += PathCount;
3538 // Merge the IsRetainBlock values.
3540 RetainsToMove.IsRetainBlock = NewReleaseRetainRRI.IsRetainBlock;
3541 FirstRetain = false;
3542 } else if (ReleasesToMove.IsRetainBlock !=
3543 NewReleaseRetainRRI.IsRetainBlock)
3544 // It's not possible to merge the sequences if one uses
3545 // objc_retain and the other uses objc_retainBlock.
3548 // Collect the optimal insertion points.
3550 for (SmallPtrSet<Instruction *, 2>::const_iterator
3551 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
3552 RE = NewReleaseRetainRRI.ReverseInsertPts.end();
3554 Instruction *RIP = *RI;
3555 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
3556 PathCount = BBStates[RIP->getParent()].GetAllPathCount();
3557 NewDelta += PathCount;
3558 NewCount += PathCount;
3561 NewRetains.push_back(NewReleaseRetain);
3565 NewReleases.clear();
3566 if (NewRetains.empty()) break;
3569 // If the pointer is known incremented or nested, we can safely delete the
3570 // pair regardless of what's between them.
3571 if (KnownSafeTD || KnownSafeBU) {
3572 RetainsToMove.ReverseInsertPts.clear();
3573 ReleasesToMove.ReverseInsertPts.clear();
3576 // Determine whether the new insertion points we computed preserve the
3577 // balance of retain and release calls through the program.
3578 // TODO: If the fully aggressive solution isn't valid, try to find a
3579 // less aggressive solution which is.
3584 // Determine whether the original call points are balanced in the retain and
3585 // release calls through the program. If not, conservatively don't touch
3587 // TODO: It's theoretically possible to do code motion in this case, as
3588 // long as the existing imbalances are maintained.
3592 // Ok, everything checks out and we're all set. Let's move some code!
3594 assert(OldCount != 0 && "Unreachable code?");
3595 AnyPairsCompletelyEliminated = NewCount == 0;
3596 NumRRs += OldCount - NewCount;
3597 MoveCalls(Arg, RetainsToMove, ReleasesToMove,
3598 Retains, Releases, DeadInsts, M);
3601 NewReleases.clear();
3603 RetainsToMove.clear();
3604 ReleasesToMove.clear();
3607 // Now that we're done moving everything, we can delete the newly dead
3608 // instructions, as we no longer need them as insert points.
3609 while (!DeadInsts.empty())
3610 EraseInstruction(DeadInsts.pop_back_val());
3612 return AnyPairsCompletelyEliminated;
3615 /// Weak pointer optimizations.
3616 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
3617 // First, do memdep-style RLE and S2L optimizations. We can't use memdep
3618 // itself because it uses AliasAnalysis and we need to do provenance
3620 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3621 Instruction *Inst = &*I++;
3623 DEBUG(dbgs() << "ObjCARCOpt::OptimizeWeakCalls: Visiting: " << *Inst <<
3626 InstructionClass Class = GetBasicInstructionClass(Inst);
3627 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
3630 // Delete objc_loadWeak calls with no users.
3631 if (Class == IC_LoadWeak && Inst->use_empty()) {
3632 Inst->eraseFromParent();
3636 // TODO: For now, just look for an earlier available version of this value
3637 // within the same block. Theoretically, we could do memdep-style non-local
3638 // analysis too, but that would want caching. A better approach would be to
3639 // use the technique that EarlyCSE uses.
3640 inst_iterator Current = llvm::prior(I);
3641 BasicBlock *CurrentBB = Current.getBasicBlockIterator();
3642 for (BasicBlock::iterator B = CurrentBB->begin(),
3643 J = Current.getInstructionIterator();
3645 Instruction *EarlierInst = &*llvm::prior(J);
3646 InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
3647 switch (EarlierClass) {
3649 case IC_LoadWeakRetained: {
3650 // If this is loading from the same pointer, replace this load's value
3652 CallInst *Call = cast<CallInst>(Inst);
3653 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
3654 Value *Arg = Call->getArgOperand(0);
3655 Value *EarlierArg = EarlierCall->getArgOperand(0);
3656 switch (PA.getAA()->alias(Arg, EarlierArg)) {
3657 case AliasAnalysis::MustAlias:
3659 // If the load has a builtin retain, insert a plain retain for it.
3660 if (Class == IC_LoadWeakRetained) {
3662 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
3666 // Zap the fully redundant load.
3667 Call->replaceAllUsesWith(EarlierCall);
3668 Call->eraseFromParent();
3670 case AliasAnalysis::MayAlias:
3671 case AliasAnalysis::PartialAlias:
3673 case AliasAnalysis::NoAlias:
3680 // If this is storing to the same pointer and has the same size etc.
3681 // replace this load's value with the stored value.
3682 CallInst *Call = cast<CallInst>(Inst);
3683 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
3684 Value *Arg = Call->getArgOperand(0);
3685 Value *EarlierArg = EarlierCall->getArgOperand(0);
3686 switch (PA.getAA()->alias(Arg, EarlierArg)) {
3687 case AliasAnalysis::MustAlias:
3689 // If the load has a builtin retain, insert a plain retain for it.
3690 if (Class == IC_LoadWeakRetained) {
3692 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
3696 // Zap the fully redundant load.
3697 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
3698 Call->eraseFromParent();
3700 case AliasAnalysis::MayAlias:
3701 case AliasAnalysis::PartialAlias:
3703 case AliasAnalysis::NoAlias:
3710 // TOOD: Grab the copied value.
3712 case IC_AutoreleasepoolPush:
3715 // Weak pointers are only modified through the weak entry points
3716 // (and arbitrary calls, which could call the weak entry points).
3719 // Anything else could modify the weak pointer.
3726 // Then, for each destroyWeak with an alloca operand, check to see if
3727 // the alloca and all its users can be zapped.
3728 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3729 Instruction *Inst = &*I++;
3730 InstructionClass Class = GetBasicInstructionClass(Inst);
3731 if (Class != IC_DestroyWeak)
3734 CallInst *Call = cast<CallInst>(Inst);
3735 Value *Arg = Call->getArgOperand(0);
3736 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
3737 for (Value::use_iterator UI = Alloca->use_begin(),
3738 UE = Alloca->use_end(); UI != UE; ++UI) {
3739 const Instruction *UserInst = cast<Instruction>(*UI);
3740 switch (GetBasicInstructionClass(UserInst)) {
3743 case IC_DestroyWeak:
3750 for (Value::use_iterator UI = Alloca->use_begin(),
3751 UE = Alloca->use_end(); UI != UE; ) {
3752 CallInst *UserInst = cast<CallInst>(*UI++);
3753 switch (GetBasicInstructionClass(UserInst)) {
3756 // These functions return their second argument.
3757 UserInst->replaceAllUsesWith(UserInst->getArgOperand(1));
3759 case IC_DestroyWeak:
3763 llvm_unreachable("alloca really is used!");
3765 UserInst->eraseFromParent();
3767 Alloca->eraseFromParent();
3772 DEBUG(dbgs() << "ObjCARCOpt::OptimizeWeakCalls: Finished List.\n\n");
3776 /// Identify program paths which execute sequences of retains and releases which
3777 /// can be eliminated.
3778 bool ObjCARCOpt::OptimizeSequences(Function &F) {
3779 /// Releases, Retains - These are used to store the results of the main flow
3780 /// analysis. These use Value* as the key instead of Instruction* so that the
3781 /// map stays valid when we get around to rewriting code and calls get
3782 /// replaced by arguments.
3783 DenseMap<Value *, RRInfo> Releases;
3784 MapVector<Value *, RRInfo> Retains;
3786 /// This is used during the traversal of the function to track the
3787 /// states for each identified object at each block.
3788 DenseMap<const BasicBlock *, BBState> BBStates;
3790 // Analyze the CFG of the function, and all instructions.
3791 bool NestingDetected = Visit(F, BBStates, Retains, Releases);
3794 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
3798 /// Look for this pattern:
3800 /// %call = call i8* @something(...)
3801 /// %2 = call i8* @objc_retain(i8* %call)
3802 /// %3 = call i8* @objc_autorelease(i8* %2)
3805 /// And delete the retain and autorelease.
3807 /// Otherwise if it's just this:
3809 /// %3 = call i8* @objc_autorelease(i8* %2)
3812 /// convert the autorelease to autoreleaseRV.
3813 void ObjCARCOpt::OptimizeReturns(Function &F) {
3814 if (!F.getReturnType()->isPointerTy())
3817 SmallPtrSet<Instruction *, 4> DependingInstructions;
3818 SmallPtrSet<const BasicBlock *, 4> Visited;
3819 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
3820 BasicBlock *BB = FI;
3821 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
3823 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Visiting: " << *Ret << "\n");
3827 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
3828 FindDependencies(NeedsPositiveRetainCount, Arg,
3829 BB, Ret, DependingInstructions, Visited, PA);
3830 if (DependingInstructions.size() != 1)
3834 CallInst *Autorelease =
3835 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3838 InstructionClass AutoreleaseClass = GetBasicInstructionClass(Autorelease);
3839 if (!IsAutorelease(AutoreleaseClass))
3841 if (GetObjCArg(Autorelease) != Arg)
3844 DependingInstructions.clear();
3847 // Check that there is nothing that can affect the reference
3848 // count between the autorelease and the retain.
3849 FindDependencies(CanChangeRetainCount, Arg,
3850 BB, Autorelease, DependingInstructions, Visited, PA);
3851 if (DependingInstructions.size() != 1)
3856 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3858 // Check that we found a retain with the same argument.
3860 !IsRetain(GetBasicInstructionClass(Retain)) ||
3861 GetObjCArg(Retain) != Arg)
3864 DependingInstructions.clear();
3867 // Convert the autorelease to an autoreleaseRV, since it's
3868 // returning the value.
3869 if (AutoreleaseClass == IC_Autorelease) {
3870 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Converting autorelease "
3871 "=> autoreleaseRV since it's returning a value.\n"
3872 " In: " << *Autorelease
3874 Autorelease->setCalledFunction(getAutoreleaseRVCallee(F.getParent()));
3875 DEBUG(dbgs() << " Out: " << *Autorelease
3877 Autorelease->setTailCall(); // Always tail call autoreleaseRV.
3878 AutoreleaseClass = IC_AutoreleaseRV;
3881 // Check that there is nothing that can affect the reference
3882 // count between the retain and the call.
3883 // Note that Retain need not be in BB.
3884 FindDependencies(CanChangeRetainCount, Arg, Retain->getParent(), Retain,
3885 DependingInstructions, Visited, PA);
3886 if (DependingInstructions.size() != 1)
3891 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3893 // Check that the pointer is the return value of the call.
3894 if (!Call || Arg != Call)
3897 // Check that the call is a regular call.
3898 InstructionClass Class = GetBasicInstructionClass(Call);
3899 if (Class != IC_CallOrUser && Class != IC_Call)
3902 // If so, we can zap the retain and autorelease.
3905 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Erasing: " << *Retain
3907 << *Autorelease << "\n");
3908 EraseInstruction(Retain);
3909 EraseInstruction(Autorelease);
3915 DependingInstructions.clear();
3919 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Finished List.\n\n");
3923 bool ObjCARCOpt::doInitialization(Module &M) {
3927 // If nothing in the Module uses ARC, don't do anything.
3928 Run = ModuleHasARC(M);
3932 // Identify the imprecise release metadata kind.
3933 ImpreciseReleaseMDKind =
3934 M.getContext().getMDKindID("clang.imprecise_release");
3935 CopyOnEscapeMDKind =
3936 M.getContext().getMDKindID("clang.arc.copy_on_escape");
3937 NoObjCARCExceptionsMDKind =
3938 M.getContext().getMDKindID("clang.arc.no_objc_arc_exceptions");
3940 // Intuitively, objc_retain and others are nocapture, however in practice
3941 // they are not, because they return their argument value. And objc_release
3942 // calls finalizers which can have arbitrary side effects.
3944 // These are initialized lazily.
3946 AutoreleaseRVCallee = 0;
3949 RetainBlockCallee = 0;
3950 AutoreleaseCallee = 0;
3955 bool ObjCARCOpt::runOnFunction(Function &F) {
3959 // If nothing in the Module uses ARC, don't do anything.
3965 DEBUG(dbgs() << "ObjCARCOpt: Visiting Function: " << F.getName() << "\n");
3967 PA.setAA(&getAnalysis<AliasAnalysis>());
3969 // This pass performs several distinct transformations. As a compile-time aid
3970 // when compiling code that isn't ObjC, skip these if the relevant ObjC
3971 // library functions aren't declared.
3973 // Preliminary optimizations. This also computs UsedInThisFunction.
3974 OptimizeIndividualCalls(F);
3976 // Optimizations for weak pointers.
3977 if (UsedInThisFunction & ((1 << IC_LoadWeak) |
3978 (1 << IC_LoadWeakRetained) |
3979 (1 << IC_StoreWeak) |
3980 (1 << IC_InitWeak) |
3981 (1 << IC_CopyWeak) |
3982 (1 << IC_MoveWeak) |
3983 (1 << IC_DestroyWeak)))
3984 OptimizeWeakCalls(F);
3986 // Optimizations for retain+release pairs.
3987 if (UsedInThisFunction & ((1 << IC_Retain) |
3988 (1 << IC_RetainRV) |
3989 (1 << IC_RetainBlock)))
3990 if (UsedInThisFunction & (1 << IC_Release))
3991 // Run OptimizeSequences until it either stops making changes or
3992 // no retain+release pair nesting is detected.
3993 while (OptimizeSequences(F)) {}
3995 // Optimizations if objc_autorelease is used.
3996 if (UsedInThisFunction & ((1 << IC_Autorelease) |
3997 (1 << IC_AutoreleaseRV)))
4000 DEBUG(dbgs() << "\n");
4005 void ObjCARCOpt::releaseMemory() {
4011 /// \defgroup ARCContract ARC Contraction.
4014 // TODO: ObjCARCContract could insert PHI nodes when uses aren't
4015 // dominated by single calls.
4017 #include "llvm/Analysis/Dominators.h"
4018 #include "llvm/IR/InlineAsm.h"
4019 #include "llvm/IR/Operator.h"
4021 STATISTIC(NumStoreStrongs, "Number objc_storeStrong calls formed");
4024 /// \brief Late ARC optimizations
4026 /// These change the IR in a way that makes it difficult to be analyzed by
4027 /// ObjCARCOpt, so it's run late.
4028 class ObjCARCContract : public FunctionPass {
4032 ProvenanceAnalysis PA;
4034 /// A flag indicating whether this optimization pass should run.
4037 /// Declarations for ObjC runtime functions, for use in creating calls to
4038 /// them. These are initialized lazily to avoid cluttering up the Module
4039 /// with unused declarations.
4041 /// Declaration for objc_storeStrong().
4042 Constant *StoreStrongCallee;
4043 /// Declaration for objc_retainAutorelease().
4044 Constant *RetainAutoreleaseCallee;
4045 /// Declaration for objc_retainAutoreleaseReturnValue().
4046 Constant *RetainAutoreleaseRVCallee;
4048 /// The inline asm string to insert between calls and RetainRV calls to make
4049 /// the optimization work on targets which need it.
4050 const MDString *RetainRVMarker;
4052 /// The set of inserted objc_storeStrong calls. If at the end of walking the
4053 /// function we have found no alloca instructions, these calls can be marked
4055 SmallPtrSet<CallInst *, 8> StoreStrongCalls;
4057 Constant *getStoreStrongCallee(Module *M);
4058 Constant *getRetainAutoreleaseCallee(Module *M);
4059 Constant *getRetainAutoreleaseRVCallee(Module *M);
4061 bool ContractAutorelease(Function &F, Instruction *Autorelease,
4062 InstructionClass Class,
4063 SmallPtrSet<Instruction *, 4>
4064 &DependingInstructions,
4065 SmallPtrSet<const BasicBlock *, 4>
4068 void ContractRelease(Instruction *Release,
4069 inst_iterator &Iter);
4071 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
4072 virtual bool doInitialization(Module &M);
4073 virtual bool runOnFunction(Function &F);
4077 ObjCARCContract() : FunctionPass(ID) {
4078 initializeObjCARCContractPass(*PassRegistry::getPassRegistry());
4083 char ObjCARCContract::ID = 0;
4084 INITIALIZE_PASS_BEGIN(ObjCARCContract,
4085 "objc-arc-contract", "ObjC ARC contraction", false, false)
4086 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
4087 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4088 INITIALIZE_PASS_END(ObjCARCContract,
4089 "objc-arc-contract", "ObjC ARC contraction", false, false)
4091 Pass *llvm::createObjCARCContractPass() {
4092 return new ObjCARCContract();
4095 void ObjCARCContract::getAnalysisUsage(AnalysisUsage &AU) const {
4096 AU.addRequired<AliasAnalysis>();
4097 AU.addRequired<DominatorTree>();
4098 AU.setPreservesCFG();
4101 Constant *ObjCARCContract::getStoreStrongCallee(Module *M) {
4102 if (!StoreStrongCallee) {
4103 LLVMContext &C = M->getContext();
4104 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
4105 Type *I8XX = PointerType::getUnqual(I8X);
4106 Type *Params[] = { I8XX, I8X };
4108 AttributeSet Attr = AttributeSet()
4109 .addAttribute(M->getContext(), AttributeSet::FunctionIndex,
4110 Attribute::NoUnwind)
4111 .addAttribute(M->getContext(), 1, Attribute::NoCapture);
4114 M->getOrInsertFunction(
4116 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
4119 return StoreStrongCallee;
4122 Constant *ObjCARCContract::getRetainAutoreleaseCallee(Module *M) {
4123 if (!RetainAutoreleaseCallee) {
4124 LLVMContext &C = M->getContext();
4125 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
4126 Type *Params[] = { I8X };
4127 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
4128 AttributeSet Attribute =
4129 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
4130 Attribute::NoUnwind);
4131 RetainAutoreleaseCallee =
4132 M->getOrInsertFunction("objc_retainAutorelease", FTy, Attribute);
4134 return RetainAutoreleaseCallee;
4137 Constant *ObjCARCContract::getRetainAutoreleaseRVCallee(Module *M) {
4138 if (!RetainAutoreleaseRVCallee) {
4139 LLVMContext &C = M->getContext();
4140 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
4141 Type *Params[] = { I8X };
4142 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
4143 AttributeSet Attribute =
4144 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
4145 Attribute::NoUnwind);
4146 RetainAutoreleaseRVCallee =
4147 M->getOrInsertFunction("objc_retainAutoreleaseReturnValue", FTy,
4150 return RetainAutoreleaseRVCallee;
4153 /// Merge an autorelease with a retain into a fused call.
4155 ObjCARCContract::ContractAutorelease(Function &F, Instruction *Autorelease,
4156 InstructionClass Class,
4157 SmallPtrSet<Instruction *, 4>
4158 &DependingInstructions,
4159 SmallPtrSet<const BasicBlock *, 4>
4161 const Value *Arg = GetObjCArg(Autorelease);
4163 // Check that there are no instructions between the retain and the autorelease
4164 // (such as an autorelease_pop) which may change the count.
4165 CallInst *Retain = 0;
4166 if (Class == IC_AutoreleaseRV)
4167 FindDependencies(RetainAutoreleaseRVDep, Arg,
4168 Autorelease->getParent(), Autorelease,
4169 DependingInstructions, Visited, PA);
4171 FindDependencies(RetainAutoreleaseDep, Arg,
4172 Autorelease->getParent(), Autorelease,
4173 DependingInstructions, Visited, PA);
4176 if (DependingInstructions.size() != 1) {
4177 DependingInstructions.clear();
4181 Retain = dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
4182 DependingInstructions.clear();
4185 GetBasicInstructionClass(Retain) != IC_Retain ||
4186 GetObjCArg(Retain) != Arg)
4192 DEBUG(dbgs() << "ObjCARCContract::ContractAutorelease: Fusing "
4193 "retain/autorelease. Erasing: " << *Autorelease << "\n"
4195 << *Retain << "\n");
4197 if (Class == IC_AutoreleaseRV)
4198 Retain->setCalledFunction(getRetainAutoreleaseRVCallee(F.getParent()));
4200 Retain->setCalledFunction(getRetainAutoreleaseCallee(F.getParent()));
4202 DEBUG(dbgs() << " New Retain: "
4203 << *Retain << "\n");
4205 EraseInstruction(Autorelease);
4209 /// Attempt to merge an objc_release with a store, load, and objc_retain to form
4210 /// an objc_storeStrong. This can be a little tricky because the instructions
4211 /// don't always appear in order, and there may be unrelated intervening
4213 void ObjCARCContract::ContractRelease(Instruction *Release,
4214 inst_iterator &Iter) {
4215 LoadInst *Load = dyn_cast<LoadInst>(GetObjCArg(Release));
4216 if (!Load || !Load->isSimple()) return;
4218 // For now, require everything to be in one basic block.
4219 BasicBlock *BB = Release->getParent();
4220 if (Load->getParent() != BB) return;
4222 // Walk down to find the store and the release, which may be in either order.
4223 BasicBlock::iterator I = Load, End = BB->end();
4225 AliasAnalysis::Location Loc = AA->getLocation(Load);
4226 StoreInst *Store = 0;
4227 bool SawRelease = false;
4228 for (; !Store || !SawRelease; ++I) {
4232 Instruction *Inst = I;
4233 if (Inst == Release) {
4238 InstructionClass Class = GetBasicInstructionClass(Inst);
4240 // Unrelated retains are harmless.
4241 if (IsRetain(Class))
4245 // The store is the point where we're going to put the objc_storeStrong,
4246 // so make sure there are no uses after it.
4247 if (CanUse(Inst, Load, PA, Class))
4249 } else if (AA->getModRefInfo(Inst, Loc) & AliasAnalysis::Mod) {
4250 // We are moving the load down to the store, so check for anything
4251 // else which writes to the memory between the load and the store.
4252 Store = dyn_cast<StoreInst>(Inst);
4253 if (!Store || !Store->isSimple()) return;
4254 if (Store->getPointerOperand() != Loc.Ptr) return;
4258 Value *New = StripPointerCastsAndObjCCalls(Store->getValueOperand());
4260 // Walk up to find the retain.
4262 BasicBlock::iterator Begin = BB->begin();
4263 while (I != Begin && GetBasicInstructionClass(I) != IC_Retain)
4265 Instruction *Retain = I;
4266 if (GetBasicInstructionClass(Retain) != IC_Retain) return;
4267 if (GetObjCArg(Retain) != New) return;
4272 LLVMContext &C = Release->getContext();
4273 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
4274 Type *I8XX = PointerType::getUnqual(I8X);
4276 Value *Args[] = { Load->getPointerOperand(), New };
4277 if (Args[0]->getType() != I8XX)
4278 Args[0] = new BitCastInst(Args[0], I8XX, "", Store);
4279 if (Args[1]->getType() != I8X)
4280 Args[1] = new BitCastInst(Args[1], I8X, "", Store);
4281 CallInst *StoreStrong =
4282 CallInst::Create(getStoreStrongCallee(BB->getParent()->getParent()),
4284 StoreStrong->setDoesNotThrow();
4285 StoreStrong->setDebugLoc(Store->getDebugLoc());
4287 // We can't set the tail flag yet, because we haven't yet determined
4288 // whether there are any escaping allocas. Remember this call, so that
4289 // we can set the tail flag once we know it's safe.
4290 StoreStrongCalls.insert(StoreStrong);
4292 if (&*Iter == Store) ++Iter;
4293 Store->eraseFromParent();
4294 Release->eraseFromParent();
4295 EraseInstruction(Retain);
4296 if (Load->use_empty())
4297 Load->eraseFromParent();
4300 bool ObjCARCContract::doInitialization(Module &M) {
4301 // If nothing in the Module uses ARC, don't do anything.
4302 Run = ModuleHasARC(M);
4306 // These are initialized lazily.
4307 StoreStrongCallee = 0;
4308 RetainAutoreleaseCallee = 0;
4309 RetainAutoreleaseRVCallee = 0;
4311 // Initialize RetainRVMarker.
4313 if (NamedMDNode *NMD =
4314 M.getNamedMetadata("clang.arc.retainAutoreleasedReturnValueMarker"))
4315 if (NMD->getNumOperands() == 1) {
4316 const MDNode *N = NMD->getOperand(0);
4317 if (N->getNumOperands() == 1)
4318 if (const MDString *S = dyn_cast<MDString>(N->getOperand(0)))
4325 bool ObjCARCContract::runOnFunction(Function &F) {
4329 // If nothing in the Module uses ARC, don't do anything.
4334 AA = &getAnalysis<AliasAnalysis>();
4335 DT = &getAnalysis<DominatorTree>();
4337 PA.setAA(&getAnalysis<AliasAnalysis>());
4339 // Track whether it's ok to mark objc_storeStrong calls with the "tail"
4340 // keyword. Be conservative if the function has variadic arguments.
4341 // It seems that functions which "return twice" are also unsafe for the
4342 // "tail" argument, because they are setjmp, which could need to
4343 // return to an earlier stack state.
4344 bool TailOkForStoreStrongs = !F.isVarArg() &&
4345 !F.callsFunctionThatReturnsTwice();
4347 // For ObjC library calls which return their argument, replace uses of the
4348 // argument with uses of the call return value, if it dominates the use. This
4349 // reduces register pressure.
4350 SmallPtrSet<Instruction *, 4> DependingInstructions;
4351 SmallPtrSet<const BasicBlock *, 4> Visited;
4352 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
4353 Instruction *Inst = &*I++;
4355 DEBUG(dbgs() << "ObjCARCContract: Visiting: " << *Inst << "\n");
4357 // Only these library routines return their argument. In particular,
4358 // objc_retainBlock does not necessarily return its argument.
4359 InstructionClass Class = GetBasicInstructionClass(Inst);
4362 case IC_FusedRetainAutorelease:
4363 case IC_FusedRetainAutoreleaseRV:
4365 case IC_Autorelease:
4366 case IC_AutoreleaseRV:
4367 if (ContractAutorelease(F, Inst, Class, DependingInstructions, Visited))
4371 // If we're compiling for a target which needs a special inline-asm
4372 // marker to do the retainAutoreleasedReturnValue optimization,
4374 if (!RetainRVMarker)
4376 BasicBlock::iterator BBI = Inst;
4377 BasicBlock *InstParent = Inst->getParent();
4379 // Step up to see if the call immediately precedes the RetainRV call.
4380 // If it's an invoke, we have to cross a block boundary. And we have
4381 // to carefully dodge no-op instructions.
4383 if (&*BBI == InstParent->begin()) {
4384 BasicBlock *Pred = InstParent->getSinglePredecessor();
4386 goto decline_rv_optimization;
4387 BBI = Pred->getTerminator();
4391 } while (isNoopInstruction(BBI));
4393 if (&*BBI == GetObjCArg(Inst)) {
4394 DEBUG(dbgs() << "ObjCARCContract: Adding inline asm marker for "
4395 "retainAutoreleasedReturnValue optimization.\n");
4398 InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
4399 /*isVarArg=*/false),
4400 RetainRVMarker->getString(),
4401 /*Constraints=*/"", /*hasSideEffects=*/true);
4402 CallInst::Create(IA, "", Inst);
4404 decline_rv_optimization:
4408 // objc_initWeak(p, null) => *p = null
4409 CallInst *CI = cast<CallInst>(Inst);
4410 if (isNullOrUndef(CI->getArgOperand(1))) {
4412 ConstantPointerNull::get(cast<PointerType>(CI->getType()));
4414 new StoreInst(Null, CI->getArgOperand(0), CI);
4416 DEBUG(dbgs() << "OBJCARCContract: Old = " << *CI << "\n"
4417 << " New = " << *Null << "\n");
4419 CI->replaceAllUsesWith(Null);
4420 CI->eraseFromParent();
4425 ContractRelease(Inst, I);
4428 // Be conservative if the function has any alloca instructions.
4429 // Technically we only care about escaping alloca instructions,
4430 // but this is sufficient to handle some interesting cases.
4431 if (isa<AllocaInst>(Inst))
4432 TailOkForStoreStrongs = false;
4438 DEBUG(dbgs() << "ObjCARCContract: Finished List.\n\n");
4440 // Don't use GetObjCArg because we don't want to look through bitcasts
4441 // and such; to do the replacement, the argument must have type i8*.
4442 const Value *Arg = cast<CallInst>(Inst)->getArgOperand(0);
4444 // If we're compiling bugpointed code, don't get in trouble.
4445 if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))
4447 // Look through the uses of the pointer.
4448 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
4450 Use &U = UI.getUse();
4451 unsigned OperandNo = UI.getOperandNo();
4452 ++UI; // Increment UI now, because we may unlink its element.
4454 // If the call's return value dominates a use of the call's argument
4455 // value, rewrite the use to use the return value. We check for
4456 // reachability here because an unreachable call is considered to
4457 // trivially dominate itself, which would lead us to rewriting its
4458 // argument in terms of its return value, which would lead to
4459 // infinite loops in GetObjCArg.
4460 if (DT->isReachableFromEntry(U) && DT->dominates(Inst, U)) {
4462 Instruction *Replacement = Inst;
4463 Type *UseTy = U.get()->getType();
4464 if (PHINode *PHI = dyn_cast<PHINode>(U.getUser())) {
4465 // For PHI nodes, insert the bitcast in the predecessor block.
4466 unsigned ValNo = PHINode::getIncomingValueNumForOperand(OperandNo);
4467 BasicBlock *BB = PHI->getIncomingBlock(ValNo);
4468 if (Replacement->getType() != UseTy)
4469 Replacement = new BitCastInst(Replacement, UseTy, "",
4471 // While we're here, rewrite all edges for this PHI, rather
4472 // than just one use at a time, to minimize the number of
4473 // bitcasts we emit.
4474 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4475 if (PHI->getIncomingBlock(i) == BB) {
4476 // Keep the UI iterator valid.
4477 if (&PHI->getOperandUse(
4478 PHINode::getOperandNumForIncomingValue(i)) ==
4481 PHI->setIncomingValue(i, Replacement);
4484 if (Replacement->getType() != UseTy)
4485 Replacement = new BitCastInst(Replacement, UseTy, "",
4486 cast<Instruction>(U.getUser()));
4492 // If Arg is a no-op casted pointer, strip one level of casts and iterate.
4493 if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg))
4494 Arg = BI->getOperand(0);
4495 else if (isa<GEPOperator>(Arg) &&
4496 cast<GEPOperator>(Arg)->hasAllZeroIndices())
4497 Arg = cast<GEPOperator>(Arg)->getPointerOperand();
4498 else if (isa<GlobalAlias>(Arg) &&
4499 !cast<GlobalAlias>(Arg)->mayBeOverridden())
4500 Arg = cast<GlobalAlias>(Arg)->getAliasee();
4506 // If this function has no escaping allocas or suspicious vararg usage,
4507 // objc_storeStrong calls can be marked with the "tail" keyword.
4508 if (TailOkForStoreStrongs)
4509 for (SmallPtrSet<CallInst *, 8>::iterator I = StoreStrongCalls.begin(),
4510 E = StoreStrongCalls.end(); I != E; ++I)
4511 (*I)->setTailCall();
4512 StoreStrongCalls.clear();