1 //===- ObjCARC.cpp - ObjC ARC Optimization --------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines ObjC ARC optimizations. ARC stands for
11 // Automatic Reference Counting and is a system for managing reference counts
12 // for objects in Objective C.
14 // The optimizations performed include elimination of redundant, partially
15 // redundant, and inconsequential reference count operations, elimination of
16 // redundant weak pointer operations, pattern-matching and replacement of
17 // low-level operations into higher-level operations, and numerous minor
20 // This file also defines a simple ARC-aware AliasAnalysis.
22 // WARNING: This file knows about certain library functions. It recognizes them
23 // by name, and hardwires 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/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/raw_ostream.h"
38 // A handy option to enable/disable all optimizations in this file.
39 static cl::opt<bool> EnableARCOpts("enable-objc-arc-opts", cl::init(true));
41 //===----------------------------------------------------------------------===//
43 //===----------------------------------------------------------------------===//
46 /// MapVector - An associative container with fast insertion-order
47 /// (deterministic) iteration over its elements. Plus the special
49 template<class KeyT, class ValueT>
51 /// Map - Map keys to indices in Vector.
52 typedef DenseMap<KeyT, size_t> MapTy;
55 /// Vector - Keys and values.
56 typedef std::vector<std::pair<KeyT, ValueT> > VectorTy;
60 typedef typename VectorTy::iterator iterator;
61 typedef typename VectorTy::const_iterator const_iterator;
62 iterator begin() { return Vector.begin(); }
63 iterator end() { return Vector.end(); }
64 const_iterator begin() const { return Vector.begin(); }
65 const_iterator end() const { return Vector.end(); }
69 assert(Vector.size() >= Map.size()); // May differ due to blotting.
70 for (typename MapTy::const_iterator I = Map.begin(), E = Map.end();
72 assert(I->second < Vector.size());
73 assert(Vector[I->second].first == I->first);
75 for (typename VectorTy::const_iterator I = Vector.begin(),
76 E = Vector.end(); I != E; ++I)
78 (Map.count(I->first) &&
79 Map[I->first] == size_t(I - Vector.begin())));
83 ValueT &operator[](const KeyT &Arg) {
84 std::pair<typename MapTy::iterator, bool> Pair =
85 Map.insert(std::make_pair(Arg, size_t(0)));
87 size_t Num = Vector.size();
88 Pair.first->second = Num;
89 Vector.push_back(std::make_pair(Arg, ValueT()));
90 return Vector[Num].second;
92 return Vector[Pair.first->second].second;
95 std::pair<iterator, bool>
96 insert(const std::pair<KeyT, ValueT> &InsertPair) {
97 std::pair<typename MapTy::iterator, bool> Pair =
98 Map.insert(std::make_pair(InsertPair.first, size_t(0)));
100 size_t Num = Vector.size();
101 Pair.first->second = Num;
102 Vector.push_back(InsertPair);
103 return std::make_pair(Vector.begin() + Num, true);
105 return std::make_pair(Vector.begin() + Pair.first->second, false);
108 const_iterator find(const KeyT &Key) const {
109 typename MapTy::const_iterator It = Map.find(Key);
110 if (It == Map.end()) return Vector.end();
111 return Vector.begin() + It->second;
114 /// blot - This is similar to erase, but instead of removing the element
115 /// from the vector, it just zeros out the key in the vector. This leaves
116 /// iterators intact, but clients must be prepared for zeroed-out keys when
118 void blot(const KeyT &Key) {
119 typename MapTy::iterator It = Map.find(Key);
120 if (It == Map.end()) return;
121 Vector[It->second].first = KeyT();
132 //===----------------------------------------------------------------------===//
134 //===----------------------------------------------------------------------===//
136 #include "llvm/ADT/StringSwitch.h"
137 #include "llvm/Analysis/ValueTracking.h"
138 #include "llvm/IR/Intrinsics.h"
139 #include "llvm/IR/Module.h"
140 #include "llvm/Support/CallSite.h"
141 #include "llvm/Transforms/Utils/Local.h"
144 /// InstructionClass - 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
172 /// IsPotentialUse - Test whether the given value is possible a
173 /// reference-counted pointer.
174 static bool IsPotentialUse(const Value *Op) {
175 // Pointers to static or stack storage are not reference-counted pointers.
176 if (isa<Constant>(Op) || isa<AllocaInst>(Op))
178 // Special arguments are not reference-counted.
179 if (const Argument *Arg = dyn_cast<Argument>(Op))
180 if (Arg->hasByValAttr() ||
181 Arg->hasNestAttr() ||
182 Arg->hasStructRetAttr())
184 // Only consider values with pointer types.
185 // It seemes intuitive to exclude function pointer types as well, since
186 // functions are never reference-counted, however clang occasionally
187 // bitcasts reference-counted pointers to function-pointer type
189 PointerType *Ty = dyn_cast<PointerType>(Op->getType());
192 // Conservatively assume anything else is a potential use.
196 /// GetCallSiteClass - Helper for GetInstructionClass. Determines what kind
197 /// of construct CS is.
198 static InstructionClass GetCallSiteClass(ImmutableCallSite CS) {
199 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
201 if (IsPotentialUse(*I))
202 return CS.onlyReadsMemory() ? IC_User : IC_CallOrUser;
204 return CS.onlyReadsMemory() ? IC_None : IC_Call;
207 /// GetFunctionClass - Determine if F is one of the special known Functions.
208 /// If it isn't, return IC_CallOrUser.
209 static InstructionClass GetFunctionClass(const Function *F) {
210 Function::const_arg_iterator AI = F->arg_begin(), AE = F->arg_end();
214 return StringSwitch<InstructionClass>(F->getName())
215 .Case("objc_autoreleasePoolPush", IC_AutoreleasepoolPush)
216 .Default(IC_CallOrUser);
219 const Argument *A0 = AI++;
221 // Argument is a pointer.
222 if (PointerType *PTy = dyn_cast<PointerType>(A0->getType())) {
223 Type *ETy = PTy->getElementType();
225 if (ETy->isIntegerTy(8))
226 return StringSwitch<InstructionClass>(F->getName())
227 .Case("objc_retain", IC_Retain)
228 .Case("objc_retainAutoreleasedReturnValue", IC_RetainRV)
229 .Case("objc_retainBlock", IC_RetainBlock)
230 .Case("objc_release", IC_Release)
231 .Case("objc_autorelease", IC_Autorelease)
232 .Case("objc_autoreleaseReturnValue", IC_AutoreleaseRV)
233 .Case("objc_autoreleasePoolPop", IC_AutoreleasepoolPop)
234 .Case("objc_retainedObject", IC_NoopCast)
235 .Case("objc_unretainedObject", IC_NoopCast)
236 .Case("objc_unretainedPointer", IC_NoopCast)
237 .Case("objc_retain_autorelease", IC_FusedRetainAutorelease)
238 .Case("objc_retainAutorelease", IC_FusedRetainAutorelease)
239 .Case("objc_retainAutoreleaseReturnValue",IC_FusedRetainAutoreleaseRV)
240 .Default(IC_CallOrUser);
243 if (PointerType *Pte = dyn_cast<PointerType>(ETy))
244 if (Pte->getElementType()->isIntegerTy(8))
245 return StringSwitch<InstructionClass>(F->getName())
246 .Case("objc_loadWeakRetained", IC_LoadWeakRetained)
247 .Case("objc_loadWeak", IC_LoadWeak)
248 .Case("objc_destroyWeak", IC_DestroyWeak)
249 .Default(IC_CallOrUser);
252 // Two arguments, first is i8**.
253 const Argument *A1 = AI++;
255 if (PointerType *PTy = dyn_cast<PointerType>(A0->getType()))
256 if (PointerType *Pte = dyn_cast<PointerType>(PTy->getElementType()))
257 if (Pte->getElementType()->isIntegerTy(8))
258 if (PointerType *PTy1 = dyn_cast<PointerType>(A1->getType())) {
259 Type *ETy1 = PTy1->getElementType();
260 // Second argument is i8*
261 if (ETy1->isIntegerTy(8))
262 return StringSwitch<InstructionClass>(F->getName())
263 .Case("objc_storeWeak", IC_StoreWeak)
264 .Case("objc_initWeak", IC_InitWeak)
265 .Case("objc_storeStrong", IC_StoreStrong)
266 .Default(IC_CallOrUser);
267 // Second argument is i8**.
268 if (PointerType *Pte1 = dyn_cast<PointerType>(ETy1))
269 if (Pte1->getElementType()->isIntegerTy(8))
270 return StringSwitch<InstructionClass>(F->getName())
271 .Case("objc_moveWeak", IC_MoveWeak)
272 .Case("objc_copyWeak", IC_CopyWeak)
273 .Default(IC_CallOrUser);
277 return IC_CallOrUser;
280 /// GetInstructionClass - Determine what kind of construct V is.
281 static InstructionClass GetInstructionClass(const Value *V) {
282 if (const Instruction *I = dyn_cast<Instruction>(V)) {
283 // Any instruction other than bitcast and gep with a pointer operand have a
284 // use of an objc pointer. Bitcasts, GEPs, Selects, PHIs transfer a pointer
285 // to a subsequent use, rather than using it themselves, in this sense.
286 // As a short cut, several other opcodes are known to have no pointer
287 // operands of interest. And ret is never followed by a release, so it's
288 // not interesting to examine.
289 switch (I->getOpcode()) {
290 case Instruction::Call: {
291 const CallInst *CI = cast<CallInst>(I);
292 // Check for calls to special functions.
293 if (const Function *F = CI->getCalledFunction()) {
294 InstructionClass Class = GetFunctionClass(F);
295 if (Class != IC_CallOrUser)
298 // None of the intrinsic functions do objc_release. For intrinsics, the
299 // only question is whether or not they may be users.
300 switch (F->getIntrinsicID()) {
301 case Intrinsic::returnaddress: case Intrinsic::frameaddress:
302 case Intrinsic::stacksave: case Intrinsic::stackrestore:
303 case Intrinsic::vastart: case Intrinsic::vacopy: case Intrinsic::vaend:
304 case Intrinsic::objectsize: case Intrinsic::prefetch:
305 case Intrinsic::stackprotector:
306 case Intrinsic::eh_return_i32: case Intrinsic::eh_return_i64:
307 case Intrinsic::eh_typeid_for: case Intrinsic::eh_dwarf_cfa:
308 case Intrinsic::eh_sjlj_lsda: case Intrinsic::eh_sjlj_functioncontext:
309 case Intrinsic::init_trampoline: case Intrinsic::adjust_trampoline:
310 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
311 case Intrinsic::invariant_start: case Intrinsic::invariant_end:
312 // Don't let dbg info affect our results.
313 case Intrinsic::dbg_declare: case Intrinsic::dbg_value:
314 // Short cut: Some intrinsics obviously don't use ObjC pointers.
320 return GetCallSiteClass(CI);
322 case Instruction::Invoke:
323 return GetCallSiteClass(cast<InvokeInst>(I));
324 case Instruction::BitCast:
325 case Instruction::GetElementPtr:
326 case Instruction::Select: case Instruction::PHI:
327 case Instruction::Ret: case Instruction::Br:
328 case Instruction::Switch: case Instruction::IndirectBr:
329 case Instruction::Alloca: case Instruction::VAArg:
330 case Instruction::Add: case Instruction::FAdd:
331 case Instruction::Sub: case Instruction::FSub:
332 case Instruction::Mul: case Instruction::FMul:
333 case Instruction::SDiv: case Instruction::UDiv: case Instruction::FDiv:
334 case Instruction::SRem: case Instruction::URem: case Instruction::FRem:
335 case Instruction::Shl: case Instruction::LShr: case Instruction::AShr:
336 case Instruction::And: case Instruction::Or: case Instruction::Xor:
337 case Instruction::SExt: case Instruction::ZExt: case Instruction::Trunc:
338 case Instruction::IntToPtr: case Instruction::FCmp:
339 case Instruction::FPTrunc: case Instruction::FPExt:
340 case Instruction::FPToUI: case Instruction::FPToSI:
341 case Instruction::UIToFP: case Instruction::SIToFP:
342 case Instruction::InsertElement: case Instruction::ExtractElement:
343 case Instruction::ShuffleVector:
344 case Instruction::ExtractValue:
346 case Instruction::ICmp:
347 // Comparing a pointer with null, or any other constant, isn't an
348 // interesting use, because we don't care what the pointer points to, or
349 // about the values of any other dynamic reference-counted pointers.
350 if (IsPotentialUse(I->getOperand(1)))
354 // For anything else, check all the operands.
355 // Note that this includes both operands of a Store: while the first
356 // operand isn't actually being dereferenced, it is being stored to
357 // memory where we can no longer track who might read it and dereference
358 // it, so we have to consider it potentially used.
359 for (User::const_op_iterator OI = I->op_begin(), OE = I->op_end();
361 if (IsPotentialUse(*OI))
366 // Otherwise, it's totally inert for ARC purposes.
370 /// GetBasicInstructionClass - Determine what kind of construct V is. This is
371 /// similar to GetInstructionClass except that it only detects objc runtine
372 /// calls. This allows it to be faster.
373 static InstructionClass GetBasicInstructionClass(const Value *V) {
374 if (const CallInst *CI = dyn_cast<CallInst>(V)) {
375 if (const Function *F = CI->getCalledFunction())
376 return GetFunctionClass(F);
377 // Otherwise, be conservative.
378 return IC_CallOrUser;
381 // Otherwise, be conservative.
382 return isa<InvokeInst>(V) ? IC_CallOrUser : IC_User;
385 /// IsRetain - Test if the given class is objc_retain or
387 static bool IsRetain(InstructionClass Class) {
388 return Class == IC_Retain ||
389 Class == IC_RetainRV;
392 /// IsAutorelease - Test if the given class is objc_autorelease or
394 static bool IsAutorelease(InstructionClass Class) {
395 return Class == IC_Autorelease ||
396 Class == IC_AutoreleaseRV;
399 /// IsForwarding - Test if the given class represents instructions which return
400 /// their argument verbatim.
401 static bool IsForwarding(InstructionClass Class) {
402 // objc_retainBlock technically doesn't always return its argument
403 // verbatim, but it doesn't matter for our purposes here.
404 return Class == IC_Retain ||
405 Class == IC_RetainRV ||
406 Class == IC_Autorelease ||
407 Class == IC_AutoreleaseRV ||
408 Class == IC_RetainBlock ||
409 Class == IC_NoopCast;
412 /// IsNoopOnNull - Test if the given class represents instructions which do
413 /// nothing if passed a null pointer.
414 static bool IsNoopOnNull(InstructionClass Class) {
415 return Class == IC_Retain ||
416 Class == IC_RetainRV ||
417 Class == IC_Release ||
418 Class == IC_Autorelease ||
419 Class == IC_AutoreleaseRV ||
420 Class == IC_RetainBlock;
423 /// IsAlwaysTail - Test if the given class represents instructions which are
424 /// always safe to mark with the "tail" keyword.
425 static bool IsAlwaysTail(InstructionClass Class) {
426 // IC_RetainBlock may be given a stack argument.
427 return Class == IC_Retain ||
428 Class == IC_RetainRV ||
429 Class == IC_Autorelease ||
430 Class == IC_AutoreleaseRV;
433 /// IsNoThrow - Test if the given class represents instructions which are always
434 /// safe to mark with the nounwind attribute..
435 static bool IsNoThrow(InstructionClass Class) {
436 // objc_retainBlock is not nounwind because it calls user copy constructors
437 // which could theoretically throw.
438 return Class == IC_Retain ||
439 Class == IC_RetainRV ||
440 Class == IC_Release ||
441 Class == IC_Autorelease ||
442 Class == IC_AutoreleaseRV ||
443 Class == IC_AutoreleasepoolPush ||
444 Class == IC_AutoreleasepoolPop;
447 /// EraseInstruction - Erase the given instruction. Many ObjC calls return their
448 /// argument verbatim, so if it's such a call and the return value has users,
449 /// replace them with the argument value.
450 static void EraseInstruction(Instruction *CI) {
451 Value *OldArg = cast<CallInst>(CI)->getArgOperand(0);
453 bool Unused = CI->use_empty();
456 // Replace the return value with the argument.
457 assert(IsForwarding(GetBasicInstructionClass(CI)) &&
458 "Can't delete non-forwarding instruction with users!");
459 CI->replaceAllUsesWith(OldArg);
462 CI->eraseFromParent();
465 RecursivelyDeleteTriviallyDeadInstructions(OldArg);
468 /// GetUnderlyingObjCPtr - This is a wrapper around getUnderlyingObject which
469 /// also knows how to look through objc_retain and objc_autorelease calls, which
470 /// we know to return their argument verbatim.
471 static const Value *GetUnderlyingObjCPtr(const Value *V) {
473 V = GetUnderlyingObject(V);
474 if (!IsForwarding(GetBasicInstructionClass(V)))
476 V = cast<CallInst>(V)->getArgOperand(0);
482 /// StripPointerCastsAndObjCCalls - This is a wrapper around
483 /// Value::stripPointerCasts which also knows how to look through objc_retain
484 /// and objc_autorelease calls, which we know to return their argument verbatim.
485 static const Value *StripPointerCastsAndObjCCalls(const Value *V) {
487 V = V->stripPointerCasts();
488 if (!IsForwarding(GetBasicInstructionClass(V)))
490 V = cast<CallInst>(V)->getArgOperand(0);
495 /// StripPointerCastsAndObjCCalls - This is a wrapper around
496 /// Value::stripPointerCasts which also knows how to look through objc_retain
497 /// and objc_autorelease calls, which we know to return their argument verbatim.
498 static Value *StripPointerCastsAndObjCCalls(Value *V) {
500 V = V->stripPointerCasts();
501 if (!IsForwarding(GetBasicInstructionClass(V)))
503 V = cast<CallInst>(V)->getArgOperand(0);
508 /// GetObjCArg - Assuming the given instruction is one of the special calls such
509 /// as objc_retain or objc_release, return the argument value, stripped of no-op
510 /// casts and forwarding calls.
511 static Value *GetObjCArg(Value *Inst) {
512 return StripPointerCastsAndObjCCalls(cast<CallInst>(Inst)->getArgOperand(0));
515 /// IsObjCIdentifiedObject - This is similar to AliasAnalysis'
516 /// isObjCIdentifiedObject, except that it uses special knowledge of
517 /// ObjC conventions...
518 static bool IsObjCIdentifiedObject(const Value *V) {
519 // Assume that call results and arguments have their own "provenance".
520 // Constants (including GlobalVariables) and Allocas are never
521 // reference-counted.
522 if (isa<CallInst>(V) || isa<InvokeInst>(V) ||
523 isa<Argument>(V) || isa<Constant>(V) ||
527 if (const LoadInst *LI = dyn_cast<LoadInst>(V)) {
528 const Value *Pointer =
529 StripPointerCastsAndObjCCalls(LI->getPointerOperand());
530 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Pointer)) {
531 // A constant pointer can't be pointing to an object on the heap. It may
532 // be reference-counted, but it won't be deleted.
533 if (GV->isConstant())
535 StringRef Name = GV->getName();
536 // These special variables are known to hold values which are not
537 // reference-counted pointers.
538 if (Name.startswith("\01L_OBJC_SELECTOR_REFERENCES_") ||
539 Name.startswith("\01L_OBJC_CLASSLIST_REFERENCES_") ||
540 Name.startswith("\01L_OBJC_CLASSLIST_SUP_REFS_$_") ||
541 Name.startswith("\01L_OBJC_METH_VAR_NAME_") ||
542 Name.startswith("\01l_objc_msgSend_fixup_"))
550 /// FindSingleUseIdentifiedObject - This is similar to
551 /// StripPointerCastsAndObjCCalls but it stops as soon as it finds a value
552 /// with multiple uses.
553 static const Value *FindSingleUseIdentifiedObject(const Value *Arg) {
554 if (Arg->hasOneUse()) {
555 if (const BitCastInst *BC = dyn_cast<BitCastInst>(Arg))
556 return FindSingleUseIdentifiedObject(BC->getOperand(0));
557 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Arg))
558 if (GEP->hasAllZeroIndices())
559 return FindSingleUseIdentifiedObject(GEP->getPointerOperand());
560 if (IsForwarding(GetBasicInstructionClass(Arg)))
561 return FindSingleUseIdentifiedObject(
562 cast<CallInst>(Arg)->getArgOperand(0));
563 if (!IsObjCIdentifiedObject(Arg))
568 // If we found an identifiable object but it has multiple uses, but they are
569 // trivial uses, we can still consider this to be a single-use value.
570 if (IsObjCIdentifiedObject(Arg)) {
571 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
574 if (!U->use_empty() || StripPointerCastsAndObjCCalls(U) != Arg)
584 /// ModuleHasARC - Test if the given module looks interesting to run ARC
586 static bool ModuleHasARC(const Module &M) {
588 M.getNamedValue("objc_retain") ||
589 M.getNamedValue("objc_release") ||
590 M.getNamedValue("objc_autorelease") ||
591 M.getNamedValue("objc_retainAutoreleasedReturnValue") ||
592 M.getNamedValue("objc_retainBlock") ||
593 M.getNamedValue("objc_autoreleaseReturnValue") ||
594 M.getNamedValue("objc_autoreleasePoolPush") ||
595 M.getNamedValue("objc_loadWeakRetained") ||
596 M.getNamedValue("objc_loadWeak") ||
597 M.getNamedValue("objc_destroyWeak") ||
598 M.getNamedValue("objc_storeWeak") ||
599 M.getNamedValue("objc_initWeak") ||
600 M.getNamedValue("objc_moveWeak") ||
601 M.getNamedValue("objc_copyWeak") ||
602 M.getNamedValue("objc_retainedObject") ||
603 M.getNamedValue("objc_unretainedObject") ||
604 M.getNamedValue("objc_unretainedPointer");
607 /// DoesObjCBlockEscape - Test whether the given pointer, which is an
608 /// Objective C block pointer, does not "escape". This differs from regular
609 /// escape analysis in that a use as an argument to a call is not considered
611 static bool DoesObjCBlockEscape(const Value *BlockPtr) {
612 // Walk the def-use chains.
613 SmallVector<const Value *, 4> Worklist;
614 Worklist.push_back(BlockPtr);
616 const Value *V = Worklist.pop_back_val();
617 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
619 const User *UUser = *UI;
620 // Special - Use by a call (callee or argument) is not considered
622 switch (GetBasicInstructionClass(UUser)) {
627 case IC_AutoreleaseRV:
628 // These special functions make copies of their pointer arguments.
632 // Use by an instruction which copies the value is an escape if the
633 // result is an escape.
634 if (isa<BitCastInst>(UUser) || isa<GetElementPtrInst>(UUser) ||
635 isa<PHINode>(UUser) || isa<SelectInst>(UUser)) {
636 Worklist.push_back(UUser);
639 // Use by a load is not an escape.
640 if (isa<LoadInst>(UUser))
642 // Use by a store is not an escape if the use is the address.
643 if (const StoreInst *SI = dyn_cast<StoreInst>(UUser))
644 if (V != SI->getValueOperand())
648 // Regular calls and other stuff are not considered escapes.
651 // Otherwise, conservatively assume an escape.
654 } while (!Worklist.empty());
660 //===----------------------------------------------------------------------===//
661 // ARC AliasAnalysis.
662 //===----------------------------------------------------------------------===//
664 #include "llvm/Analysis/AliasAnalysis.h"
665 #include "llvm/Analysis/Passes.h"
666 #include "llvm/Pass.h"
669 /// ObjCARCAliasAnalysis - This is a simple alias analysis
670 /// implementation that uses knowledge of ARC constructs to answer queries.
672 /// TODO: This class could be generalized to know about other ObjC-specific
673 /// tricks. Such as knowing that ivars in the non-fragile ABI are non-aliasing
674 /// even though their offsets are dynamic.
675 class ObjCARCAliasAnalysis : public ImmutablePass,
676 public AliasAnalysis {
678 static char ID; // Class identification, replacement for typeinfo
679 ObjCARCAliasAnalysis() : ImmutablePass(ID) {
680 initializeObjCARCAliasAnalysisPass(*PassRegistry::getPassRegistry());
684 virtual void initializePass() {
685 InitializeAliasAnalysis(this);
688 /// getAdjustedAnalysisPointer - This method is used when a pass implements
689 /// an analysis interface through multiple inheritance. If needed, it
690 /// should override this to adjust the this pointer as needed for the
691 /// specified pass info.
692 virtual void *getAdjustedAnalysisPointer(const void *PI) {
693 if (PI == &AliasAnalysis::ID)
694 return static_cast<AliasAnalysis *>(this);
698 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
699 virtual AliasResult alias(const Location &LocA, const Location &LocB);
700 virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
701 virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
702 virtual ModRefBehavior getModRefBehavior(const Function *F);
703 virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
704 const Location &Loc);
705 virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
706 ImmutableCallSite CS2);
708 } // End of anonymous namespace
710 // Register this pass...
711 char ObjCARCAliasAnalysis::ID = 0;
712 INITIALIZE_AG_PASS(ObjCARCAliasAnalysis, AliasAnalysis, "objc-arc-aa",
713 "ObjC-ARC-Based Alias Analysis", false, true, false)
715 ImmutablePass *llvm::createObjCARCAliasAnalysisPass() {
716 return new ObjCARCAliasAnalysis();
720 ObjCARCAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
721 AU.setPreservesAll();
722 AliasAnalysis::getAnalysisUsage(AU);
725 AliasAnalysis::AliasResult
726 ObjCARCAliasAnalysis::alias(const Location &LocA, const Location &LocB) {
728 return AliasAnalysis::alias(LocA, LocB);
730 // First, strip off no-ops, including ObjC-specific no-ops, and try making a
731 // precise alias query.
732 const Value *SA = StripPointerCastsAndObjCCalls(LocA.Ptr);
733 const Value *SB = StripPointerCastsAndObjCCalls(LocB.Ptr);
735 AliasAnalysis::alias(Location(SA, LocA.Size, LocA.TBAATag),
736 Location(SB, LocB.Size, LocB.TBAATag));
737 if (Result != MayAlias)
740 // If that failed, climb to the underlying object, including climbing through
741 // ObjC-specific no-ops, and try making an imprecise alias query.
742 const Value *UA = GetUnderlyingObjCPtr(SA);
743 const Value *UB = GetUnderlyingObjCPtr(SB);
744 if (UA != SA || UB != SB) {
745 Result = AliasAnalysis::alias(Location(UA), Location(UB));
746 // We can't use MustAlias or PartialAlias results here because
747 // GetUnderlyingObjCPtr may return an offsetted pointer value.
748 if (Result == NoAlias)
752 // If that failed, fail. We don't need to chain here, since that's covered
753 // by the earlier precise query.
758 ObjCARCAliasAnalysis::pointsToConstantMemory(const Location &Loc,
761 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
763 // First, strip off no-ops, including ObjC-specific no-ops, and try making
764 // a precise alias query.
765 const Value *S = StripPointerCastsAndObjCCalls(Loc.Ptr);
766 if (AliasAnalysis::pointsToConstantMemory(Location(S, Loc.Size, Loc.TBAATag),
770 // If that failed, climb to the underlying object, including climbing through
771 // ObjC-specific no-ops, and try making an imprecise alias query.
772 const Value *U = GetUnderlyingObjCPtr(S);
774 return AliasAnalysis::pointsToConstantMemory(Location(U), OrLocal);
776 // If that failed, fail. We don't need to chain here, since that's covered
777 // by the earlier precise query.
781 AliasAnalysis::ModRefBehavior
782 ObjCARCAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
783 // We have nothing to do. Just chain to the next AliasAnalysis.
784 return AliasAnalysis::getModRefBehavior(CS);
787 AliasAnalysis::ModRefBehavior
788 ObjCARCAliasAnalysis::getModRefBehavior(const Function *F) {
790 return AliasAnalysis::getModRefBehavior(F);
792 switch (GetFunctionClass(F)) {
794 return DoesNotAccessMemory;
799 return AliasAnalysis::getModRefBehavior(F);
802 AliasAnalysis::ModRefResult
803 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS, const Location &Loc) {
805 return AliasAnalysis::getModRefInfo(CS, Loc);
807 switch (GetBasicInstructionClass(CS.getInstruction())) {
811 case IC_AutoreleaseRV:
813 case IC_AutoreleasepoolPush:
814 case IC_FusedRetainAutorelease:
815 case IC_FusedRetainAutoreleaseRV:
816 // These functions don't access any memory visible to the compiler.
817 // Note that this doesn't include objc_retainBlock, because it updates
818 // pointers when it copies block data.
824 return AliasAnalysis::getModRefInfo(CS, Loc);
827 AliasAnalysis::ModRefResult
828 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS1,
829 ImmutableCallSite CS2) {
830 // TODO: Theoretically we could check for dependencies between objc_* calls
831 // and OnlyAccessesArgumentPointees calls or other well-behaved calls.
832 return AliasAnalysis::getModRefInfo(CS1, CS2);
835 //===----------------------------------------------------------------------===//
837 //===----------------------------------------------------------------------===//
839 #include "llvm/Support/InstIterator.h"
840 #include "llvm/Transforms/Scalar.h"
843 /// ObjCARCExpand - Early ARC transformations.
844 class ObjCARCExpand : public FunctionPass {
845 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
846 virtual bool doInitialization(Module &M);
847 virtual bool runOnFunction(Function &F);
849 /// Run - A flag indicating whether this optimization pass should run.
854 ObjCARCExpand() : FunctionPass(ID) {
855 initializeObjCARCExpandPass(*PassRegistry::getPassRegistry());
860 char ObjCARCExpand::ID = 0;
861 INITIALIZE_PASS(ObjCARCExpand,
862 "objc-arc-expand", "ObjC ARC expansion", false, false)
864 Pass *llvm::createObjCARCExpandPass() {
865 return new ObjCARCExpand();
868 void ObjCARCExpand::getAnalysisUsage(AnalysisUsage &AU) const {
869 AU.setPreservesCFG();
872 bool ObjCARCExpand::doInitialization(Module &M) {
873 Run = ModuleHasARC(M);
877 bool ObjCARCExpand::runOnFunction(Function &F) {
881 // If nothing in the Module uses ARC, don't do anything.
885 bool Changed = false;
887 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I) {
888 Instruction *Inst = &*I;
890 DEBUG(dbgs() << "ObjCARCExpand: Visiting: " << *Inst << "\n");
892 switch (GetBasicInstructionClass(Inst)) {
896 case IC_AutoreleaseRV:
897 case IC_FusedRetainAutorelease:
898 case IC_FusedRetainAutoreleaseRV: {
899 // These calls return their argument verbatim, as a low-level
900 // optimization. However, this makes high-level optimizations
901 // harder. Undo any uses of this optimization that the front-end
902 // emitted here. We'll redo them in the contract pass.
904 Value *Value = cast<CallInst>(Inst)->getArgOperand(0);
905 DEBUG(dbgs() << "ObjCARCExpand: Old = " << *Inst << "\n"
906 " New = " << *Value << "\n");
907 Inst->replaceAllUsesWith(Value);
915 DEBUG(dbgs() << "ObjCARCExpand: Finished Queue.\n\n");
920 //===----------------------------------------------------------------------===//
921 // ARC autorelease pool elimination.
922 //===----------------------------------------------------------------------===//
924 #include "llvm/ADT/STLExtras.h"
925 #include "llvm/IR/Constants.h"
928 /// ObjCARCAPElim - Autorelease pool elimination.
929 class ObjCARCAPElim : public ModulePass {
930 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
931 virtual bool runOnModule(Module &M);
933 static bool MayAutorelease(ImmutableCallSite CS, unsigned Depth = 0);
934 static bool OptimizeBB(BasicBlock *BB);
938 ObjCARCAPElim() : ModulePass(ID) {
939 initializeObjCARCAPElimPass(*PassRegistry::getPassRegistry());
944 char ObjCARCAPElim::ID = 0;
945 INITIALIZE_PASS(ObjCARCAPElim,
947 "ObjC ARC autorelease pool elimination",
950 Pass *llvm::createObjCARCAPElimPass() {
951 return new ObjCARCAPElim();
954 void ObjCARCAPElim::getAnalysisUsage(AnalysisUsage &AU) const {
955 AU.setPreservesCFG();
958 /// MayAutorelease - Interprocedurally determine if calls made by the
959 /// given call site can possibly produce autoreleases.
960 bool ObjCARCAPElim::MayAutorelease(ImmutableCallSite CS, unsigned Depth) {
961 if (const Function *Callee = CS.getCalledFunction()) {
962 if (Callee->isDeclaration() || Callee->mayBeOverridden())
964 for (Function::const_iterator I = Callee->begin(), E = Callee->end();
966 const BasicBlock *BB = I;
967 for (BasicBlock::const_iterator J = BB->begin(), F = BB->end();
969 if (ImmutableCallSite JCS = ImmutableCallSite(J))
970 // This recursion depth limit is arbitrary. It's just great
971 // enough to cover known interesting testcases.
973 !JCS.onlyReadsMemory() &&
974 MayAutorelease(JCS, Depth + 1))
983 bool ObjCARCAPElim::OptimizeBB(BasicBlock *BB) {
984 bool Changed = false;
986 Instruction *Push = 0;
987 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
988 Instruction *Inst = I++;
989 switch (GetBasicInstructionClass(Inst)) {
990 case IC_AutoreleasepoolPush:
993 case IC_AutoreleasepoolPop:
994 // If this pop matches a push and nothing in between can autorelease,
996 if (Push && cast<CallInst>(Inst)->getArgOperand(0) == Push) {
998 Inst->eraseFromParent();
999 Push->eraseFromParent();
1004 if (MayAutorelease(ImmutableCallSite(Inst)))
1015 bool ObjCARCAPElim::runOnModule(Module &M) {
1019 // If nothing in the Module uses ARC, don't do anything.
1020 if (!ModuleHasARC(M))
1023 // Find the llvm.global_ctors variable, as the first step in
1024 // identifying the global constructors. In theory, unnecessary autorelease
1025 // pools could occur anywhere, but in practice it's pretty rare. Global
1026 // ctors are a place where autorelease pools get inserted automatically,
1027 // so it's pretty common for them to be unnecessary, and it's pretty
1028 // profitable to eliminate them.
1029 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1033 assert(GV->hasDefinitiveInitializer() &&
1034 "llvm.global_ctors is uncooperative!");
1036 bool Changed = false;
1038 // Dig the constructor functions out of GV's initializer.
1039 ConstantArray *Init = cast<ConstantArray>(GV->getInitializer());
1040 for (User::op_iterator OI = Init->op_begin(), OE = Init->op_end();
1043 // llvm.global_ctors is an array of pairs where the second members
1044 // are constructor functions.
1045 Function *F = dyn_cast<Function>(cast<ConstantStruct>(Op)->getOperand(1));
1046 // If the user used a constructor function with the wrong signature and
1047 // it got bitcasted or whatever, look the other way.
1050 // Only look at function definitions.
1051 if (F->isDeclaration())
1053 // Only look at functions with one basic block.
1054 if (llvm::next(F->begin()) != F->end())
1056 // Ok, a single-block constructor function definition. Try to optimize it.
1057 Changed |= OptimizeBB(F->begin());
1063 //===----------------------------------------------------------------------===//
1064 // ARC optimization.
1065 //===----------------------------------------------------------------------===//
1067 // TODO: On code like this:
1070 // stuff_that_cannot_release()
1071 // objc_autorelease(%x)
1072 // stuff_that_cannot_release()
1074 // stuff_that_cannot_release()
1075 // objc_autorelease(%x)
1077 // The second retain and autorelease can be deleted.
1079 // TODO: It should be possible to delete
1080 // objc_autoreleasePoolPush and objc_autoreleasePoolPop
1081 // pairs if nothing is actually autoreleased between them. Also, autorelease
1082 // calls followed by objc_autoreleasePoolPop calls (perhaps in ObjC++ code
1083 // after inlining) can be turned into plain release calls.
1085 // TODO: Critical-edge splitting. If the optimial insertion point is
1086 // a critical edge, the current algorithm has to fail, because it doesn't
1087 // know how to split edges. It should be possible to make the optimizer
1088 // think in terms of edges, rather than blocks, and then split critical
1091 // TODO: OptimizeSequences could generalized to be Interprocedural.
1093 // TODO: Recognize that a bunch of other objc runtime calls have
1094 // non-escaping arguments and non-releasing arguments, and may be
1095 // non-autoreleasing.
1097 // TODO: Sink autorelease calls as far as possible. Unfortunately we
1098 // usually can't sink them past other calls, which would be the main
1099 // case where it would be useful.
1101 // TODO: The pointer returned from objc_loadWeakRetained is retained.
1103 // TODO: Delete release+retain pairs (rare).
1105 #include "llvm/ADT/SmallPtrSet.h"
1106 #include "llvm/ADT/Statistic.h"
1107 #include "llvm/IR/LLVMContext.h"
1108 #include "llvm/Support/CFG.h"
1110 STATISTIC(NumNoops, "Number of no-op objc calls eliminated");
1111 STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated");
1112 STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases");
1113 STATISTIC(NumRets, "Number of return value forwarding "
1114 "retain+autoreleaes eliminated");
1115 STATISTIC(NumRRs, "Number of retain+release paths eliminated");
1116 STATISTIC(NumPeeps, "Number of calls peephole-optimized");
1119 /// ProvenanceAnalysis - This is similar to BasicAliasAnalysis, and it
1120 /// uses many of the same techniques, except it uses special ObjC-specific
1121 /// reasoning about pointer relationships.
1122 class ProvenanceAnalysis {
1125 typedef std::pair<const Value *, const Value *> ValuePairTy;
1126 typedef DenseMap<ValuePairTy, bool> CachedResultsTy;
1127 CachedResultsTy CachedResults;
1129 bool relatedCheck(const Value *A, const Value *B);
1130 bool relatedSelect(const SelectInst *A, const Value *B);
1131 bool relatedPHI(const PHINode *A, const Value *B);
1133 void operator=(const ProvenanceAnalysis &) LLVM_DELETED_FUNCTION;
1134 ProvenanceAnalysis(const ProvenanceAnalysis &) LLVM_DELETED_FUNCTION;
1137 ProvenanceAnalysis() {}
1139 void setAA(AliasAnalysis *aa) { AA = aa; }
1141 AliasAnalysis *getAA() const { return AA; }
1143 bool related(const Value *A, const Value *B);
1146 CachedResults.clear();
1151 bool ProvenanceAnalysis::relatedSelect(const SelectInst *A, const Value *B) {
1152 // If the values are Selects with the same condition, we can do a more precise
1153 // check: just check for relations between the values on corresponding arms.
1154 if (const SelectInst *SB = dyn_cast<SelectInst>(B))
1155 if (A->getCondition() == SB->getCondition())
1156 return related(A->getTrueValue(), SB->getTrueValue()) ||
1157 related(A->getFalseValue(), SB->getFalseValue());
1159 // Check both arms of the Select node individually.
1160 return related(A->getTrueValue(), B) ||
1161 related(A->getFalseValue(), B);
1164 bool ProvenanceAnalysis::relatedPHI(const PHINode *A, const Value *B) {
1165 // If the values are PHIs in the same block, we can do a more precise as well
1166 // as efficient check: just check for relations between the values on
1167 // corresponding edges.
1168 if (const PHINode *PNB = dyn_cast<PHINode>(B))
1169 if (PNB->getParent() == A->getParent()) {
1170 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i)
1171 if (related(A->getIncomingValue(i),
1172 PNB->getIncomingValueForBlock(A->getIncomingBlock(i))))
1177 // Check each unique source of the PHI node against B.
1178 SmallPtrSet<const Value *, 4> UniqueSrc;
1179 for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i) {
1180 const Value *PV1 = A->getIncomingValue(i);
1181 if (UniqueSrc.insert(PV1) && related(PV1, B))
1185 // All of the arms checked out.
1189 /// isStoredObjCPointer - Test if the value of P, or any value covered by its
1190 /// provenance, is ever stored within the function (not counting callees).
1191 static bool isStoredObjCPointer(const Value *P) {
1192 SmallPtrSet<const Value *, 8> Visited;
1193 SmallVector<const Value *, 8> Worklist;
1194 Worklist.push_back(P);
1197 P = Worklist.pop_back_val();
1198 for (Value::const_use_iterator UI = P->use_begin(), UE = P->use_end();
1200 const User *Ur = *UI;
1201 if (isa<StoreInst>(Ur)) {
1202 if (UI.getOperandNo() == 0)
1203 // The pointer is stored.
1205 // The pointed is stored through.
1208 if (isa<CallInst>(Ur))
1209 // The pointer is passed as an argument, ignore this.
1211 if (isa<PtrToIntInst>(P))
1212 // Assume the worst.
1214 if (Visited.insert(Ur))
1215 Worklist.push_back(Ur);
1217 } while (!Worklist.empty());
1219 // Everything checked out.
1223 bool ProvenanceAnalysis::relatedCheck(const Value *A, const Value *B) {
1224 // Skip past provenance pass-throughs.
1225 A = GetUnderlyingObjCPtr(A);
1226 B = GetUnderlyingObjCPtr(B);
1232 // Ask regular AliasAnalysis, for a first approximation.
1233 switch (AA->alias(A, B)) {
1234 case AliasAnalysis::NoAlias:
1236 case AliasAnalysis::MustAlias:
1237 case AliasAnalysis::PartialAlias:
1239 case AliasAnalysis::MayAlias:
1243 bool AIsIdentified = IsObjCIdentifiedObject(A);
1244 bool BIsIdentified = IsObjCIdentifiedObject(B);
1246 // An ObjC-Identified object can't alias a load if it is never locally stored.
1247 if (AIsIdentified) {
1248 // Check for an obvious escape.
1249 if (isa<LoadInst>(B))
1250 return isStoredObjCPointer(A);
1251 if (BIsIdentified) {
1252 // Check for an obvious escape.
1253 if (isa<LoadInst>(A))
1254 return isStoredObjCPointer(B);
1255 // Both pointers are identified and escapes aren't an evident problem.
1258 } else if (BIsIdentified) {
1259 // Check for an obvious escape.
1260 if (isa<LoadInst>(A))
1261 return isStoredObjCPointer(B);
1264 // Special handling for PHI and Select.
1265 if (const PHINode *PN = dyn_cast<PHINode>(A))
1266 return relatedPHI(PN, B);
1267 if (const PHINode *PN = dyn_cast<PHINode>(B))
1268 return relatedPHI(PN, A);
1269 if (const SelectInst *S = dyn_cast<SelectInst>(A))
1270 return relatedSelect(S, B);
1271 if (const SelectInst *S = dyn_cast<SelectInst>(B))
1272 return relatedSelect(S, A);
1278 bool ProvenanceAnalysis::related(const Value *A, const Value *B) {
1279 // Begin by inserting a conservative value into the map. If the insertion
1280 // fails, we have the answer already. If it succeeds, leave it there until we
1281 // compute the real answer to guard against recursive queries.
1282 if (A > B) std::swap(A, B);
1283 std::pair<CachedResultsTy::iterator, bool> Pair =
1284 CachedResults.insert(std::make_pair(ValuePairTy(A, B), true));
1286 return Pair.first->second;
1288 bool Result = relatedCheck(A, B);
1289 CachedResults[ValuePairTy(A, B)] = Result;
1294 // Sequence - A sequence of states that a pointer may go through in which an
1295 // objc_retain and objc_release are actually needed.
1298 S_Retain, ///< objc_retain(x)
1299 S_CanRelease, ///< foo(x) -- x could possibly see a ref count decrement
1300 S_Use, ///< any use of x
1301 S_Stop, ///< like S_Release, but code motion is stopped
1302 S_Release, ///< objc_release(x)
1303 S_MovableRelease ///< objc_release(x), !clang.imprecise_release
1307 static Sequence MergeSeqs(Sequence A, Sequence B, bool TopDown) {
1311 if (A == S_None || B == S_None)
1314 if (A > B) std::swap(A, B);
1316 // Choose the side which is further along in the sequence.
1317 if ((A == S_Retain || A == S_CanRelease) &&
1318 (B == S_CanRelease || B == S_Use))
1321 // Choose the side which is further along in the sequence.
1322 if ((A == S_Use || A == S_CanRelease) &&
1323 (B == S_Use || B == S_Release || B == S_Stop || B == S_MovableRelease))
1325 // If both sides are releases, choose the more conservative one.
1326 if (A == S_Stop && (B == S_Release || B == S_MovableRelease))
1328 if (A == S_Release && B == S_MovableRelease)
1336 /// RRInfo - Unidirectional information about either a
1337 /// retain-decrement-use-release sequence or release-use-decrement-retain
1338 /// reverese sequence.
1340 /// KnownSafe - After an objc_retain, the reference count of the referenced
1341 /// object is known to be positive. Similarly, before an objc_release, the
1342 /// reference count of the referenced object is known to be positive. If
1343 /// there are retain-release pairs in code regions where the retain count
1344 /// is known to be positive, they can be eliminated, regardless of any side
1345 /// effects between them.
1347 /// Also, a retain+release pair nested within another retain+release
1348 /// pair all on the known same pointer value can be eliminated, regardless
1349 /// of any intervening side effects.
1351 /// KnownSafe is true when either of these conditions is satisfied.
1354 /// IsRetainBlock - True if the Calls are objc_retainBlock calls (as
1355 /// opposed to objc_retain calls).
1358 /// IsTailCallRelease - True of the objc_release calls are all marked
1359 /// with the "tail" keyword.
1360 bool IsTailCallRelease;
1362 /// ReleaseMetadata - If the Calls are objc_release calls and they all have
1363 /// a clang.imprecise_release tag, this is the metadata tag.
1364 MDNode *ReleaseMetadata;
1366 /// Calls - For a top-down sequence, the set of objc_retains or
1367 /// objc_retainBlocks. For bottom-up, the set of objc_releases.
1368 SmallPtrSet<Instruction *, 2> Calls;
1370 /// ReverseInsertPts - The set of optimal insert positions for
1371 /// moving calls in the opposite sequence.
1372 SmallPtrSet<Instruction *, 2> ReverseInsertPts;
1375 KnownSafe(false), IsRetainBlock(false),
1376 IsTailCallRelease(false),
1377 ReleaseMetadata(0) {}
1383 void RRInfo::clear() {
1385 IsRetainBlock = false;
1386 IsTailCallRelease = false;
1387 ReleaseMetadata = 0;
1389 ReverseInsertPts.clear();
1393 /// PtrState - This class summarizes several per-pointer runtime properties
1394 /// which are propogated through the flow graph.
1396 /// KnownPositiveRefCount - True if the reference count is known to
1398 bool KnownPositiveRefCount;
1400 /// Partial - True of we've seen an opportunity for partial RR elimination,
1401 /// such as pushing calls into a CFG triangle or into one side of a
1405 /// Seq - The current position in the sequence.
1409 /// RRI - Unidirectional information about the current sequence.
1410 /// TODO: Encapsulate this better.
1413 PtrState() : KnownPositiveRefCount(false), Partial(false),
1416 void SetKnownPositiveRefCount() {
1417 KnownPositiveRefCount = true;
1420 void ClearRefCount() {
1421 KnownPositiveRefCount = false;
1424 bool IsKnownIncremented() const {
1425 return KnownPositiveRefCount;
1428 void SetSeq(Sequence NewSeq) {
1432 Sequence GetSeq() const {
1436 void ClearSequenceProgress() {
1437 ResetSequenceProgress(S_None);
1440 void ResetSequenceProgress(Sequence NewSeq) {
1446 void Merge(const PtrState &Other, bool TopDown);
1451 PtrState::Merge(const PtrState &Other, bool TopDown) {
1452 Seq = MergeSeqs(Seq, Other.Seq, TopDown);
1453 KnownPositiveRefCount = KnownPositiveRefCount && Other.KnownPositiveRefCount;
1455 // We can't merge a plain objc_retain with an objc_retainBlock.
1456 if (RRI.IsRetainBlock != Other.RRI.IsRetainBlock)
1459 // If we're not in a sequence (anymore), drop all associated state.
1460 if (Seq == S_None) {
1463 } else if (Partial || Other.Partial) {
1464 // If we're doing a merge on a path that's previously seen a partial
1465 // merge, conservatively drop the sequence, to avoid doing partial
1466 // RR elimination. If the branch predicates for the two merge differ,
1467 // mixing them is unsafe.
1468 ClearSequenceProgress();
1470 // Conservatively merge the ReleaseMetadata information.
1471 if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
1472 RRI.ReleaseMetadata = 0;
1474 RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe;
1475 RRI.IsTailCallRelease = RRI.IsTailCallRelease &&
1476 Other.RRI.IsTailCallRelease;
1477 RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
1479 // Merge the insert point sets. If there are any differences,
1480 // that makes this a partial merge.
1481 Partial = RRI.ReverseInsertPts.size() != Other.RRI.ReverseInsertPts.size();
1482 for (SmallPtrSet<Instruction *, 2>::const_iterator
1483 I = Other.RRI.ReverseInsertPts.begin(),
1484 E = Other.RRI.ReverseInsertPts.end(); I != E; ++I)
1485 Partial |= RRI.ReverseInsertPts.insert(*I);
1490 /// BBState - Per-BasicBlock state.
1492 /// TopDownPathCount - The number of unique control paths from the entry
1493 /// which can reach this block.
1494 unsigned TopDownPathCount;
1496 /// BottomUpPathCount - The number of unique control paths to exits
1497 /// from this block.
1498 unsigned BottomUpPathCount;
1500 /// MapTy - A type for PerPtrTopDown and PerPtrBottomUp.
1501 typedef MapVector<const Value *, PtrState> MapTy;
1503 /// PerPtrTopDown - The top-down traversal uses this to record information
1504 /// known about a pointer at the bottom of each block.
1505 MapTy PerPtrTopDown;
1507 /// PerPtrBottomUp - The bottom-up traversal uses this to record information
1508 /// known about a pointer at the top of each block.
1509 MapTy PerPtrBottomUp;
1511 /// Preds, Succs - Effective successors and predecessors of the current
1512 /// block (this ignores ignorable edges and ignored backedges).
1513 SmallVector<BasicBlock *, 2> Preds;
1514 SmallVector<BasicBlock *, 2> Succs;
1517 BBState() : TopDownPathCount(0), BottomUpPathCount(0) {}
1519 typedef MapTy::iterator ptr_iterator;
1520 typedef MapTy::const_iterator ptr_const_iterator;
1522 ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); }
1523 ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); }
1524 ptr_const_iterator top_down_ptr_begin() const {
1525 return PerPtrTopDown.begin();
1527 ptr_const_iterator top_down_ptr_end() const {
1528 return PerPtrTopDown.end();
1531 ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); }
1532 ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); }
1533 ptr_const_iterator bottom_up_ptr_begin() const {
1534 return PerPtrBottomUp.begin();
1536 ptr_const_iterator bottom_up_ptr_end() const {
1537 return PerPtrBottomUp.end();
1540 /// SetAsEntry - Mark this block as being an entry block, which has one
1541 /// path from the entry by definition.
1542 void SetAsEntry() { TopDownPathCount = 1; }
1544 /// SetAsExit - Mark this block as being an exit block, which has one
1545 /// path to an exit by definition.
1546 void SetAsExit() { BottomUpPathCount = 1; }
1548 PtrState &getPtrTopDownState(const Value *Arg) {
1549 return PerPtrTopDown[Arg];
1552 PtrState &getPtrBottomUpState(const Value *Arg) {
1553 return PerPtrBottomUp[Arg];
1556 void clearBottomUpPointers() {
1557 PerPtrBottomUp.clear();
1560 void clearTopDownPointers() {
1561 PerPtrTopDown.clear();
1564 void InitFromPred(const BBState &Other);
1565 void InitFromSucc(const BBState &Other);
1566 void MergePred(const BBState &Other);
1567 void MergeSucc(const BBState &Other);
1569 /// GetAllPathCount - Return the number of possible unique paths from an
1570 /// entry to an exit which pass through this block. This is only valid
1571 /// after both the top-down and bottom-up traversals are complete.
1572 unsigned GetAllPathCount() const {
1573 assert(TopDownPathCount != 0);
1574 assert(BottomUpPathCount != 0);
1575 return TopDownPathCount * BottomUpPathCount;
1578 // Specialized CFG utilities.
1579 typedef SmallVectorImpl<BasicBlock *>::const_iterator edge_iterator;
1580 edge_iterator pred_begin() { return Preds.begin(); }
1581 edge_iterator pred_end() { return Preds.end(); }
1582 edge_iterator succ_begin() { return Succs.begin(); }
1583 edge_iterator succ_end() { return Succs.end(); }
1585 void addSucc(BasicBlock *Succ) { Succs.push_back(Succ); }
1586 void addPred(BasicBlock *Pred) { Preds.push_back(Pred); }
1588 bool isExit() const { return Succs.empty(); }
1592 void BBState::InitFromPred(const BBState &Other) {
1593 PerPtrTopDown = Other.PerPtrTopDown;
1594 TopDownPathCount = Other.TopDownPathCount;
1597 void BBState::InitFromSucc(const BBState &Other) {
1598 PerPtrBottomUp = Other.PerPtrBottomUp;
1599 BottomUpPathCount = Other.BottomUpPathCount;
1602 /// MergePred - The top-down traversal uses this to merge information about
1603 /// predecessors to form the initial state for a new block.
1604 void BBState::MergePred(const BBState &Other) {
1605 // Other.TopDownPathCount can be 0, in which case it is either dead or a
1606 // loop backedge. Loop backedges are special.
1607 TopDownPathCount += Other.TopDownPathCount;
1609 // Check for overflow. If we have overflow, fall back to conservative behavior.
1610 if (TopDownPathCount < Other.TopDownPathCount) {
1611 clearTopDownPointers();
1615 // For each entry in the other set, if our set has an entry with the same key,
1616 // merge the entries. Otherwise, copy the entry and merge it with an empty
1618 for (ptr_const_iterator MI = Other.top_down_ptr_begin(),
1619 ME = Other.top_down_ptr_end(); MI != ME; ++MI) {
1620 std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI);
1621 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1625 // For each entry in our set, if the other set doesn't have an entry with the
1626 // same key, force it to merge with an empty entry.
1627 for (ptr_iterator MI = top_down_ptr_begin(),
1628 ME = top_down_ptr_end(); MI != ME; ++MI)
1629 if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end())
1630 MI->second.Merge(PtrState(), /*TopDown=*/true);
1633 /// MergeSucc - The bottom-up traversal uses this to merge information about
1634 /// successors to form the initial state for a new block.
1635 void BBState::MergeSucc(const BBState &Other) {
1636 // Other.BottomUpPathCount can be 0, in which case it is either dead or a
1637 // loop backedge. Loop backedges are special.
1638 BottomUpPathCount += Other.BottomUpPathCount;
1640 // Check for overflow. If we have overflow, fall back to conservative behavior.
1641 if (BottomUpPathCount < Other.BottomUpPathCount) {
1642 clearBottomUpPointers();
1646 // For each entry in the other set, if our set has an entry with the
1647 // same key, merge the entries. Otherwise, copy the entry and merge
1648 // it with an empty entry.
1649 for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(),
1650 ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) {
1651 std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI);
1652 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1656 // For each entry in our set, if the other set doesn't have an entry
1657 // with the same key, force it to merge with an empty entry.
1658 for (ptr_iterator MI = bottom_up_ptr_begin(),
1659 ME = bottom_up_ptr_end(); MI != ME; ++MI)
1660 if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end())
1661 MI->second.Merge(PtrState(), /*TopDown=*/false);
1665 /// ObjCARCOpt - The main ARC optimization pass.
1666 class ObjCARCOpt : public FunctionPass {
1668 ProvenanceAnalysis PA;
1670 /// Run - A flag indicating whether this optimization pass should run.
1673 /// RetainRVCallee, etc. - Declarations for ObjC runtime
1674 /// functions, for use in creating calls to them. These are initialized
1675 /// lazily to avoid cluttering up the Module with unused declarations.
1676 Constant *RetainRVCallee, *AutoreleaseRVCallee, *ReleaseCallee,
1677 *RetainCallee, *RetainBlockCallee, *AutoreleaseCallee;
1679 /// UsedInThisFunciton - Flags which determine whether each of the
1680 /// interesting runtine functions is in fact used in the current function.
1681 unsigned UsedInThisFunction;
1683 /// ImpreciseReleaseMDKind - The Metadata Kind for clang.imprecise_release
1685 unsigned ImpreciseReleaseMDKind;
1687 /// CopyOnEscapeMDKind - The Metadata Kind for clang.arc.copy_on_escape
1689 unsigned CopyOnEscapeMDKind;
1691 /// NoObjCARCExceptionsMDKind - The Metadata Kind for
1692 /// clang.arc.no_objc_arc_exceptions metadata.
1693 unsigned NoObjCARCExceptionsMDKind;
1695 Constant *getRetainRVCallee(Module *M);
1696 Constant *getAutoreleaseRVCallee(Module *M);
1697 Constant *getReleaseCallee(Module *M);
1698 Constant *getRetainCallee(Module *M);
1699 Constant *getRetainBlockCallee(Module *M);
1700 Constant *getAutoreleaseCallee(Module *M);
1702 bool IsRetainBlockOptimizable(const Instruction *Inst);
1704 void OptimizeRetainCall(Function &F, Instruction *Retain);
1705 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
1706 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV);
1707 void OptimizeIndividualCalls(Function &F);
1709 void CheckForCFGHazards(const BasicBlock *BB,
1710 DenseMap<const BasicBlock *, BBState> &BBStates,
1711 BBState &MyStates) const;
1712 bool VisitInstructionBottomUp(Instruction *Inst,
1714 MapVector<Value *, RRInfo> &Retains,
1716 bool VisitBottomUp(BasicBlock *BB,
1717 DenseMap<const BasicBlock *, BBState> &BBStates,
1718 MapVector<Value *, RRInfo> &Retains);
1719 bool VisitInstructionTopDown(Instruction *Inst,
1720 DenseMap<Value *, RRInfo> &Releases,
1722 bool VisitTopDown(BasicBlock *BB,
1723 DenseMap<const BasicBlock *, BBState> &BBStates,
1724 DenseMap<Value *, RRInfo> &Releases);
1725 bool Visit(Function &F,
1726 DenseMap<const BasicBlock *, BBState> &BBStates,
1727 MapVector<Value *, RRInfo> &Retains,
1728 DenseMap<Value *, RRInfo> &Releases);
1730 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1731 MapVector<Value *, RRInfo> &Retains,
1732 DenseMap<Value *, RRInfo> &Releases,
1733 SmallVectorImpl<Instruction *> &DeadInsts,
1736 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1737 MapVector<Value *, RRInfo> &Retains,
1738 DenseMap<Value *, RRInfo> &Releases,
1741 void OptimizeWeakCalls(Function &F);
1743 bool OptimizeSequences(Function &F);
1745 void OptimizeReturns(Function &F);
1747 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1748 virtual bool doInitialization(Module &M);
1749 virtual bool runOnFunction(Function &F);
1750 virtual void releaseMemory();
1754 ObjCARCOpt() : FunctionPass(ID) {
1755 initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1760 char ObjCARCOpt::ID = 0;
1761 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1762 "objc-arc", "ObjC ARC optimization", false, false)
1763 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1764 INITIALIZE_PASS_END(ObjCARCOpt,
1765 "objc-arc", "ObjC ARC optimization", false, false)
1767 Pass *llvm::createObjCARCOptPass() {
1768 return new ObjCARCOpt();
1771 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1772 AU.addRequired<ObjCARCAliasAnalysis>();
1773 AU.addRequired<AliasAnalysis>();
1774 // ARC optimization doesn't currently split critical edges.
1775 AU.setPreservesCFG();
1778 bool ObjCARCOpt::IsRetainBlockOptimizable(const Instruction *Inst) {
1779 // Without the magic metadata tag, we have to assume this might be an
1780 // objc_retainBlock call inserted to convert a block pointer to an id,
1781 // in which case it really is needed.
1782 if (!Inst->getMetadata(CopyOnEscapeMDKind))
1785 // If the pointer "escapes" (not including being used in a call),
1786 // the copy may be needed.
1787 if (DoesObjCBlockEscape(Inst))
1790 // Otherwise, it's not needed.
1794 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1795 if (!RetainRVCallee) {
1796 LLVMContext &C = M->getContext();
1797 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1798 Type *Params[] = { I8X };
1799 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1800 AttributeSet Attribute =
1801 AttributeSet().addAttr(M->getContext(), AttributeSet::FunctionIndex,
1802 Attribute::get(C, Attribute::NoUnwind));
1804 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1807 return RetainRVCallee;
1810 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1811 if (!AutoreleaseRVCallee) {
1812 LLVMContext &C = M->getContext();
1813 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1814 Type *Params[] = { I8X };
1815 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1816 AttributeSet Attribute =
1817 AttributeSet().addAttr(M->getContext(), AttributeSet::FunctionIndex,
1818 Attribute::get(C, Attribute::NoUnwind));
1819 AutoreleaseRVCallee =
1820 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1823 return AutoreleaseRVCallee;
1826 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1827 if (!ReleaseCallee) {
1828 LLVMContext &C = M->getContext();
1829 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1830 AttributeSet Attribute =
1831 AttributeSet().addAttr(M->getContext(), AttributeSet::FunctionIndex,
1832 Attribute::get(C, Attribute::NoUnwind));
1834 M->getOrInsertFunction(
1836 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1839 return ReleaseCallee;
1842 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1843 if (!RetainCallee) {
1844 LLVMContext &C = M->getContext();
1845 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1846 AttributeSet Attribute =
1847 AttributeSet().addAttr(M->getContext(), AttributeSet::FunctionIndex,
1848 Attribute::get(C, Attribute::NoUnwind));
1850 M->getOrInsertFunction(
1852 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1855 return RetainCallee;
1858 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1859 if (!RetainBlockCallee) {
1860 LLVMContext &C = M->getContext();
1861 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1862 // objc_retainBlock is not nounwind because it calls user copy constructors
1863 // which could theoretically throw.
1865 M->getOrInsertFunction(
1867 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1870 return RetainBlockCallee;
1873 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1874 if (!AutoreleaseCallee) {
1875 LLVMContext &C = M->getContext();
1876 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1877 AttributeSet Attribute =
1878 AttributeSet().addAttr(M->getContext(), AttributeSet::FunctionIndex,
1879 Attribute::get(C, Attribute::NoUnwind));
1881 M->getOrInsertFunction(
1883 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1886 return AutoreleaseCallee;
1889 /// IsPotentialUse - Test whether the given value is possible a
1890 /// reference-counted pointer, including tests which utilize AliasAnalysis.
1891 static bool IsPotentialUse(const Value *Op, AliasAnalysis &AA) {
1892 // First make the rudimentary check.
1893 if (!IsPotentialUse(Op))
1896 // Objects in constant memory are not reference-counted.
1897 if (AA.pointsToConstantMemory(Op))
1900 // Pointers in constant memory are not pointing to reference-counted objects.
1901 if (const LoadInst *LI = dyn_cast<LoadInst>(Op))
1902 if (AA.pointsToConstantMemory(LI->getPointerOperand()))
1905 // Otherwise assume the worst.
1909 /// CanAlterRefCount - Test whether the given instruction can result in a
1910 /// reference count modification (positive or negative) for the pointer's
1913 CanAlterRefCount(const Instruction *Inst, const Value *Ptr,
1914 ProvenanceAnalysis &PA, InstructionClass Class) {
1916 case IC_Autorelease:
1917 case IC_AutoreleaseRV:
1919 // These operations never directly modify a reference count.
1924 ImmutableCallSite CS = static_cast<const Value *>(Inst);
1925 assert(CS && "Only calls can alter reference counts!");
1927 // See if AliasAnalysis can help us with the call.
1928 AliasAnalysis::ModRefBehavior MRB = PA.getAA()->getModRefBehavior(CS);
1929 if (AliasAnalysis::onlyReadsMemory(MRB))
1931 if (AliasAnalysis::onlyAccessesArgPointees(MRB)) {
1932 for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1934 const Value *Op = *I;
1935 if (IsPotentialUse(Op, *PA.getAA()) && PA.related(Ptr, Op))
1941 // Assume the worst.
1945 /// CanUse - Test whether the given instruction can "use" the given pointer's
1946 /// object in a way that requires the reference count to be positive.
1948 CanUse(const Instruction *Inst, const Value *Ptr, ProvenanceAnalysis &PA,
1949 InstructionClass Class) {
1950 // IC_Call operations (as opposed to IC_CallOrUser) never "use" objc pointers.
1951 if (Class == IC_Call)
1954 // Consider various instructions which may have pointer arguments which are
1956 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(Inst)) {
1957 // Comparing a pointer with null, or any other constant, isn't really a use,
1958 // because we don't care what the pointer points to, or about the values
1959 // of any other dynamic reference-counted pointers.
1960 if (!IsPotentialUse(ICI->getOperand(1), *PA.getAA()))
1962 } else if (ImmutableCallSite CS = static_cast<const Value *>(Inst)) {
1963 // For calls, just check the arguments (and not the callee operand).
1964 for (ImmutableCallSite::arg_iterator OI = CS.arg_begin(),
1965 OE = CS.arg_end(); OI != OE; ++OI) {
1966 const Value *Op = *OI;
1967 if (IsPotentialUse(Op, *PA.getAA()) && PA.related(Ptr, Op))
1971 } else if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1972 // Special-case stores, because we don't care about the stored value, just
1973 // the store address.
1974 const Value *Op = GetUnderlyingObjCPtr(SI->getPointerOperand());
1975 // If we can't tell what the underlying object was, assume there is a
1977 return IsPotentialUse(Op, *PA.getAA()) && PA.related(Op, Ptr);
1980 // Check each operand for a match.
1981 for (User::const_op_iterator OI = Inst->op_begin(), OE = Inst->op_end();
1983 const Value *Op = *OI;
1984 if (IsPotentialUse(Op, *PA.getAA()) && PA.related(Ptr, Op))
1990 /// CanInterruptRV - Test whether the given instruction can autorelease
1991 /// any pointer or cause an autoreleasepool pop.
1993 CanInterruptRV(InstructionClass Class) {
1995 case IC_AutoreleasepoolPop:
1998 case IC_Autorelease:
1999 case IC_AutoreleaseRV:
2000 case IC_FusedRetainAutorelease:
2001 case IC_FusedRetainAutoreleaseRV:
2009 /// DependenceKind - There are several kinds of dependence-like concepts in
2011 enum DependenceKind {
2012 NeedsPositiveRetainCount,
2013 AutoreleasePoolBoundary,
2014 CanChangeRetainCount,
2015 RetainAutoreleaseDep, ///< Blocks objc_retainAutorelease.
2016 RetainAutoreleaseRVDep, ///< Blocks objc_retainAutoreleaseReturnValue.
2017 RetainRVDep ///< Blocks objc_retainAutoreleasedReturnValue.
2021 /// Depends - Test if there can be dependencies on Inst through Arg. This
2022 /// function only tests dependencies relevant for removing pairs of calls.
2024 Depends(DependenceKind Flavor, Instruction *Inst, const Value *Arg,
2025 ProvenanceAnalysis &PA) {
2026 // If we've reached the definition of Arg, stop.
2031 case NeedsPositiveRetainCount: {
2032 InstructionClass Class = GetInstructionClass(Inst);
2034 case IC_AutoreleasepoolPop:
2035 case IC_AutoreleasepoolPush:
2039 return CanUse(Inst, Arg, PA, Class);
2043 case AutoreleasePoolBoundary: {
2044 InstructionClass Class = GetInstructionClass(Inst);
2046 case IC_AutoreleasepoolPop:
2047 case IC_AutoreleasepoolPush:
2048 // These mark the end and begin of an autorelease pool scope.
2051 // Nothing else does this.
2056 case CanChangeRetainCount: {
2057 InstructionClass Class = GetInstructionClass(Inst);
2059 case IC_AutoreleasepoolPop:
2060 // Conservatively assume this can decrement any count.
2062 case IC_AutoreleasepoolPush:
2066 return CanAlterRefCount(Inst, Arg, PA, Class);
2070 case RetainAutoreleaseDep:
2071 switch (GetBasicInstructionClass(Inst)) {
2072 case IC_AutoreleasepoolPop:
2073 case IC_AutoreleasepoolPush:
2074 // Don't merge an objc_autorelease with an objc_retain inside a different
2075 // autoreleasepool scope.
2079 // Check for a retain of the same pointer for merging.
2080 return GetObjCArg(Inst) == Arg;
2082 // Nothing else matters for objc_retainAutorelease formation.
2086 case RetainAutoreleaseRVDep: {
2087 InstructionClass Class = GetBasicInstructionClass(Inst);
2091 // Check for a retain of the same pointer for merging.
2092 return GetObjCArg(Inst) == Arg;
2094 // Anything that can autorelease interrupts
2095 // retainAutoreleaseReturnValue formation.
2096 return CanInterruptRV(Class);
2101 return CanInterruptRV(GetBasicInstructionClass(Inst));
2104 llvm_unreachable("Invalid dependence flavor");
2107 /// FindDependencies - Walk up the CFG from StartPos (which is in StartBB) and
2108 /// find local and non-local dependencies on Arg.
2109 /// TODO: Cache results?
2111 FindDependencies(DependenceKind Flavor,
2113 BasicBlock *StartBB, Instruction *StartInst,
2114 SmallPtrSet<Instruction *, 4> &DependingInstructions,
2115 SmallPtrSet<const BasicBlock *, 4> &Visited,
2116 ProvenanceAnalysis &PA) {
2117 BasicBlock::iterator StartPos = StartInst;
2119 SmallVector<std::pair<BasicBlock *, BasicBlock::iterator>, 4> Worklist;
2120 Worklist.push_back(std::make_pair(StartBB, StartPos));
2122 std::pair<BasicBlock *, BasicBlock::iterator> Pair =
2123 Worklist.pop_back_val();
2124 BasicBlock *LocalStartBB = Pair.first;
2125 BasicBlock::iterator LocalStartPos = Pair.second;
2126 BasicBlock::iterator StartBBBegin = LocalStartBB->begin();
2128 if (LocalStartPos == StartBBBegin) {
2129 pred_iterator PI(LocalStartBB), PE(LocalStartBB, false);
2131 // If we've reached the function entry, produce a null dependence.
2132 DependingInstructions.insert(0);
2134 // Add the predecessors to the worklist.
2136 BasicBlock *PredBB = *PI;
2137 if (Visited.insert(PredBB))
2138 Worklist.push_back(std::make_pair(PredBB, PredBB->end()));
2139 } while (++PI != PE);
2143 Instruction *Inst = --LocalStartPos;
2144 if (Depends(Flavor, Inst, Arg, PA)) {
2145 DependingInstructions.insert(Inst);
2149 } while (!Worklist.empty());
2151 // Determine whether the original StartBB post-dominates all of the blocks we
2152 // visited. If not, insert a sentinal indicating that most optimizations are
2154 for (SmallPtrSet<const BasicBlock *, 4>::const_iterator I = Visited.begin(),
2155 E = Visited.end(); I != E; ++I) {
2156 const BasicBlock *BB = *I;
2159 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2160 for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
2161 const BasicBlock *Succ = *SI;
2162 if (Succ != StartBB && !Visited.count(Succ)) {
2163 DependingInstructions.insert(reinterpret_cast<Instruction *>(-1));
2170 static bool isNullOrUndef(const Value *V) {
2171 return isa<ConstantPointerNull>(V) || isa<UndefValue>(V);
2174 static bool isNoopInstruction(const Instruction *I) {
2175 return isa<BitCastInst>(I) ||
2176 (isa<GetElementPtrInst>(I) &&
2177 cast<GetElementPtrInst>(I)->hasAllZeroIndices());
2180 /// OptimizeRetainCall - Turn objc_retain into
2181 /// objc_retainAutoreleasedReturnValue if the operand is a return value.
2183 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
2184 ImmutableCallSite CS(GetObjCArg(Retain));
2185 const Instruction *Call = CS.getInstruction();
2187 if (Call->getParent() != Retain->getParent()) return;
2189 // Check that the call is next to the retain.
2190 BasicBlock::const_iterator I = Call;
2192 while (isNoopInstruction(I)) ++I;
2196 // Turn it to an objc_retainAutoreleasedReturnValue..
2199 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
2202 /// OptimizeRetainRVCall - Turn objc_retainAutoreleasedReturnValue into
2203 /// objc_retain if the operand is not a return value. Or, if it can be paired
2204 /// with an objc_autoreleaseReturnValue, delete the pair and return true.
2206 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
2207 // Check for the argument being from an immediately preceding call or invoke.
2208 const Value *Arg = GetObjCArg(RetainRV);
2209 ImmutableCallSite CS(Arg);
2210 if (const Instruction *Call = CS.getInstruction()) {
2211 if (Call->getParent() == RetainRV->getParent()) {
2212 BasicBlock::const_iterator I = Call;
2214 while (isNoopInstruction(I)) ++I;
2215 if (&*I == RetainRV)
2217 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
2218 BasicBlock *RetainRVParent = RetainRV->getParent();
2219 if (II->getNormalDest() == RetainRVParent) {
2220 BasicBlock::const_iterator I = RetainRVParent->begin();
2221 while (isNoopInstruction(I)) ++I;
2222 if (&*I == RetainRV)
2228 // Check for being preceded by an objc_autoreleaseReturnValue on the same
2229 // pointer. In this case, we can delete the pair.
2230 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
2232 do --I; while (I != Begin && isNoopInstruction(I));
2233 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
2234 GetObjCArg(I) == Arg) {
2237 EraseInstruction(I);
2238 EraseInstruction(RetainRV);
2243 // Turn it to a plain objc_retain.
2246 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
2250 /// OptimizeAutoreleaseRVCall - Turn objc_autoreleaseReturnValue into
2251 /// objc_autorelease if the result is not used as a return value.
2253 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV) {
2254 // Check for a return of the pointer value.
2255 const Value *Ptr = GetObjCArg(AutoreleaseRV);
2256 SmallVector<const Value *, 2> Users;
2257 Users.push_back(Ptr);
2259 Ptr = Users.pop_back_val();
2260 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
2262 const User *I = *UI;
2263 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
2265 if (isa<BitCastInst>(I))
2268 } while (!Users.empty());
2272 cast<CallInst>(AutoreleaseRV)->
2273 setCalledFunction(getAutoreleaseCallee(F.getParent()));
2276 /// OptimizeIndividualCalls - Visit each call, one at a time, and make
2277 /// simplifications without doing any additional analysis.
2278 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
2279 // Reset all the flags in preparation for recomputing them.
2280 UsedInThisFunction = 0;
2282 // Visit all objc_* calls in F.
2283 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2284 Instruction *Inst = &*I++;
2286 DEBUG(dbgs() << "ObjCARCOpt: OptimizeIndividualCalls: Visiting: " <<
2289 InstructionClass Class = GetBasicInstructionClass(Inst);
2294 // Delete no-op casts. These function calls have special semantics, but
2295 // the semantics are entirely implemented via lowering in the front-end,
2296 // so by the time they reach the optimizer, they are just no-op calls
2297 // which return their argument.
2299 // There are gray areas here, as the ability to cast reference-counted
2300 // pointers to raw void* and back allows code to break ARC assumptions,
2301 // however these are currently considered to be unimportant.
2305 EraseInstruction(Inst);
2308 // If the pointer-to-weak-pointer is null, it's undefined behavior.
2311 case IC_LoadWeakRetained:
2313 case IC_DestroyWeak: {
2314 CallInst *CI = cast<CallInst>(Inst);
2315 if (isNullOrUndef(CI->getArgOperand(0))) {
2317 Type *Ty = CI->getArgOperand(0)->getType();
2318 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2319 Constant::getNullValue(Ty),
2321 CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
2322 CI->eraseFromParent();
2329 CallInst *CI = cast<CallInst>(Inst);
2330 if (isNullOrUndef(CI->getArgOperand(0)) ||
2331 isNullOrUndef(CI->getArgOperand(1))) {
2333 Type *Ty = CI->getArgOperand(0)->getType();
2334 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2335 Constant::getNullValue(Ty),
2337 CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
2338 CI->eraseFromParent();
2344 OptimizeRetainCall(F, Inst);
2347 if (OptimizeRetainRVCall(F, Inst))
2350 case IC_AutoreleaseRV:
2351 OptimizeAutoreleaseRVCall(F, Inst);
2355 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
2356 if (IsAutorelease(Class) && Inst->use_empty()) {
2357 CallInst *Call = cast<CallInst>(Inst);
2358 const Value *Arg = Call->getArgOperand(0);
2359 Arg = FindSingleUseIdentifiedObject(Arg);
2364 // Create the declaration lazily.
2365 LLVMContext &C = Inst->getContext();
2367 CallInst::Create(getReleaseCallee(F.getParent()),
2368 Call->getArgOperand(0), "", Call);
2369 NewCall->setMetadata(ImpreciseReleaseMDKind,
2370 MDNode::get(C, ArrayRef<Value *>()));
2371 EraseInstruction(Call);
2377 // For functions which can never be passed stack arguments, add
2379 if (IsAlwaysTail(Class)) {
2381 cast<CallInst>(Inst)->setTailCall();
2384 // Set nounwind as needed.
2385 if (IsNoThrow(Class)) {
2387 cast<CallInst>(Inst)->setDoesNotThrow();
2390 if (!IsNoopOnNull(Class)) {
2391 UsedInThisFunction |= 1 << Class;
2395 const Value *Arg = GetObjCArg(Inst);
2397 // ARC calls with null are no-ops. Delete them.
2398 if (isNullOrUndef(Arg)) {
2401 EraseInstruction(Inst);
2405 // Keep track of which of retain, release, autorelease, and retain_block
2406 // are actually present in this function.
2407 UsedInThisFunction |= 1 << Class;
2409 // If Arg is a PHI, and one or more incoming values to the
2410 // PHI are null, and the call is control-equivalent to the PHI, and there
2411 // are no relevant side effects between the PHI and the call, the call
2412 // could be pushed up to just those paths with non-null incoming values.
2413 // For now, don't bother splitting critical edges for this.
2414 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
2415 Worklist.push_back(std::make_pair(Inst, Arg));
2417 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
2421 const PHINode *PN = dyn_cast<PHINode>(Arg);
2424 // Determine if the PHI has any null operands, or any incoming
2426 bool HasNull = false;
2427 bool HasCriticalEdges = false;
2428 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2430 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2431 if (isNullOrUndef(Incoming))
2433 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
2434 .getNumSuccessors() != 1) {
2435 HasCriticalEdges = true;
2439 // If we have null operands and no critical edges, optimize.
2440 if (!HasCriticalEdges && HasNull) {
2441 SmallPtrSet<Instruction *, 4> DependingInstructions;
2442 SmallPtrSet<const BasicBlock *, 4> Visited;
2444 // Check that there is nothing that cares about the reference
2445 // count between the call and the phi.
2448 case IC_RetainBlock:
2449 // These can always be moved up.
2452 // These can't be moved across things that care about the retain
2454 FindDependencies(NeedsPositiveRetainCount, Arg,
2455 Inst->getParent(), Inst,
2456 DependingInstructions, Visited, PA);
2458 case IC_Autorelease:
2459 // These can't be moved across autorelease pool scope boundaries.
2460 FindDependencies(AutoreleasePoolBoundary, Arg,
2461 Inst->getParent(), Inst,
2462 DependingInstructions, Visited, PA);
2465 case IC_AutoreleaseRV:
2466 // Don't move these; the RV optimization depends on the autoreleaseRV
2467 // being tail called, and the retainRV being immediately after a call
2468 // (which might still happen if we get lucky with codegen layout, but
2469 // it's not worth taking the chance).
2472 llvm_unreachable("Invalid dependence flavor");
2475 if (DependingInstructions.size() == 1 &&
2476 *DependingInstructions.begin() == PN) {
2479 // Clone the call into each predecessor that has a non-null value.
2480 CallInst *CInst = cast<CallInst>(Inst);
2481 Type *ParamTy = CInst->getArgOperand(0)->getType();
2482 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2484 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2485 if (!isNullOrUndef(Incoming)) {
2486 CallInst *Clone = cast<CallInst>(CInst->clone());
2487 Value *Op = PN->getIncomingValue(i);
2488 Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
2489 if (Op->getType() != ParamTy)
2490 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
2491 Clone->setArgOperand(0, Op);
2492 Clone->insertBefore(InsertPos);
2493 Worklist.push_back(std::make_pair(Clone, Incoming));
2496 // Erase the original call.
2497 EraseInstruction(CInst);
2501 } while (!Worklist.empty());
2503 DEBUG(dbgs() << "ObjCARCOpt: Finished Individual Call Queue.\n\n");
2508 /// CheckForCFGHazards - Check for critical edges, loop boundaries, irreducible
2509 /// control flow, or other CFG structures where moving code across the edge
2510 /// would result in it being executed more.
2512 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
2513 DenseMap<const BasicBlock *, BBState> &BBStates,
2514 BBState &MyStates) const {
2515 // If any top-down local-use or possible-dec has a succ which is earlier in
2516 // the sequence, forget it.
2517 for (BBState::ptr_iterator I = MyStates.top_down_ptr_begin(),
2518 E = MyStates.top_down_ptr_end(); I != E; ++I)
2519 switch (I->second.GetSeq()) {
2522 const Value *Arg = I->first;
2523 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2524 bool SomeSuccHasSame = false;
2525 bool AllSuccsHaveSame = true;
2526 PtrState &S = I->second;
2527 succ_const_iterator SI(TI), SE(TI, false);
2529 // If the terminator is an invoke marked with the
2530 // clang.arc.no_objc_arc_exceptions metadata, the unwind edge can be
2531 // ignored, for ARC purposes.
2532 if (isa<InvokeInst>(TI) && TI->getMetadata(NoObjCARCExceptionsMDKind))
2535 for (; SI != SE; ++SI) {
2536 Sequence SuccSSeq = S_None;
2537 bool SuccSRRIKnownSafe = false;
2538 // If VisitBottomUp has pointer information for this successor, take
2539 // what we know about it.
2540 DenseMap<const BasicBlock *, BBState>::iterator BBI =
2542 assert(BBI != BBStates.end());
2543 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
2544 SuccSSeq = SuccS.GetSeq();
2545 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
2548 case S_CanRelease: {
2549 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
2550 S.ClearSequenceProgress();
2556 SomeSuccHasSame = true;
2560 case S_MovableRelease:
2561 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
2562 AllSuccsHaveSame = false;
2565 llvm_unreachable("bottom-up pointer in retain state!");
2568 // If the state at the other end of any of the successor edges
2569 // matches the current state, require all edges to match. This
2570 // guards against loops in the middle of a sequence.
2571 if (SomeSuccHasSame && !AllSuccsHaveSame)
2572 S.ClearSequenceProgress();
2575 case S_CanRelease: {
2576 const Value *Arg = I->first;
2577 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2578 bool SomeSuccHasSame = false;
2579 bool AllSuccsHaveSame = true;
2580 PtrState &S = I->second;
2581 succ_const_iterator SI(TI), SE(TI, false);
2583 // If the terminator is an invoke marked with the
2584 // clang.arc.no_objc_arc_exceptions metadata, the unwind edge can be
2585 // ignored, for ARC purposes.
2586 if (isa<InvokeInst>(TI) && TI->getMetadata(NoObjCARCExceptionsMDKind))
2589 for (; SI != SE; ++SI) {
2590 Sequence SuccSSeq = S_None;
2591 bool SuccSRRIKnownSafe = false;
2592 // If VisitBottomUp has pointer information for this successor, take
2593 // what we know about it.
2594 DenseMap<const BasicBlock *, BBState>::iterator BBI =
2596 assert(BBI != BBStates.end());
2597 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
2598 SuccSSeq = SuccS.GetSeq();
2599 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
2602 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
2603 S.ClearSequenceProgress();
2609 SomeSuccHasSame = true;
2613 case S_MovableRelease:
2615 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
2616 AllSuccsHaveSame = false;
2619 llvm_unreachable("bottom-up pointer in retain state!");
2622 // If the state at the other end of any of the successor edges
2623 // matches the current state, require all edges to match. This
2624 // guards against loops in the middle of a sequence.
2625 if (SomeSuccHasSame && !AllSuccsHaveSame)
2626 S.ClearSequenceProgress();
2633 ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst,
2635 MapVector<Value *, RRInfo> &Retains,
2636 BBState &MyStates) {
2637 bool NestingDetected = false;
2638 InstructionClass Class = GetInstructionClass(Inst);
2639 const Value *Arg = 0;
2643 Arg = GetObjCArg(Inst);
2645 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2647 // If we see two releases in a row on the same pointer. If so, make
2648 // a note, and we'll cicle back to revisit it after we've
2649 // hopefully eliminated the second release, which may allow us to
2650 // eliminate the first release too.
2651 // Theoretically we could implement removal of nested retain+release
2652 // pairs by making PtrState hold a stack of states, but this is
2653 // simple and avoids adding overhead for the non-nested case.
2654 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease)
2655 NestingDetected = true;
2657 MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2658 S.ResetSequenceProgress(ReleaseMetadata ? S_MovableRelease : S_Release);
2659 S.RRI.ReleaseMetadata = ReleaseMetadata;
2660 S.RRI.KnownSafe = S.IsKnownIncremented();
2661 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2662 S.RRI.Calls.insert(Inst);
2664 S.SetKnownPositiveRefCount();
2667 case IC_RetainBlock:
2668 // An objc_retainBlock call with just a use may need to be kept,
2669 // because it may be copying a block from the stack to the heap.
2670 if (!IsRetainBlockOptimizable(Inst))
2675 Arg = GetObjCArg(Inst);
2677 PtrState &S = MyStates.getPtrBottomUpState(Arg);
2678 S.SetKnownPositiveRefCount();
2680 switch (S.GetSeq()) {
2683 case S_MovableRelease:
2685 S.RRI.ReverseInsertPts.clear();
2688 // Don't do retain+release tracking for IC_RetainRV, because it's
2689 // better to let it remain as the first instruction after a call.
2690 if (Class != IC_RetainRV) {
2691 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2692 Retains[Inst] = S.RRI;
2694 S.ClearSequenceProgress();
2699 llvm_unreachable("bottom-up pointer in retain state!");
2701 return NestingDetected;
2703 case IC_AutoreleasepoolPop:
2704 // Conservatively, clear MyStates for all known pointers.
2705 MyStates.clearBottomUpPointers();
2706 return NestingDetected;
2707 case IC_AutoreleasepoolPush:
2709 // These are irrelevant.
2710 return NestingDetected;
2715 // Consider any other possible effects of this instruction on each
2716 // pointer being tracked.
2717 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
2718 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
2719 const Value *Ptr = MI->first;
2721 continue; // Handled above.
2722 PtrState &S = MI->second;
2723 Sequence Seq = S.GetSeq();
2725 // Check for possible releases.
2726 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2730 S.SetSeq(S_CanRelease);
2734 case S_MovableRelease:
2739 llvm_unreachable("bottom-up pointer in retain state!");
2743 // Check for possible direct uses.
2746 case S_MovableRelease:
2747 if (CanUse(Inst, Ptr, PA, Class)) {
2748 assert(S.RRI.ReverseInsertPts.empty());
2749 // If this is an invoke instruction, we're scanning it as part of
2750 // one of its successor blocks, since we can't insert code after it
2751 // in its own block, and we don't want to split critical edges.
2752 if (isa<InvokeInst>(Inst))
2753 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
2755 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
2757 } else if (Seq == S_Release &&
2758 (Class == IC_User || Class == IC_CallOrUser)) {
2759 // Non-movable releases depend on any possible objc pointer use.
2761 assert(S.RRI.ReverseInsertPts.empty());
2762 // As above; handle invoke specially.
2763 if (isa<InvokeInst>(Inst))
2764 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
2766 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
2770 if (CanUse(Inst, Ptr, PA, Class))
2778 llvm_unreachable("bottom-up pointer in retain state!");
2782 return NestingDetected;
2786 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
2787 DenseMap<const BasicBlock *, BBState> &BBStates,
2788 MapVector<Value *, RRInfo> &Retains) {
2789 bool NestingDetected = false;
2790 BBState &MyStates = BBStates[BB];
2792 // Merge the states from each successor to compute the initial state
2793 // for the current block.
2794 BBState::edge_iterator SI(MyStates.succ_begin()),
2795 SE(MyStates.succ_end());
2797 const BasicBlock *Succ = *SI;
2798 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
2799 assert(I != BBStates.end());
2800 MyStates.InitFromSucc(I->second);
2802 for (; SI != SE; ++SI) {
2804 I = BBStates.find(Succ);
2805 assert(I != BBStates.end());
2806 MyStates.MergeSucc(I->second);
2810 // Visit all the instructions, bottom-up.
2811 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
2812 Instruction *Inst = llvm::prior(I);
2814 // Invoke instructions are visited as part of their successors (below).
2815 if (isa<InvokeInst>(Inst))
2818 NestingDetected |= VisitInstructionBottomUp(Inst, BB, Retains, MyStates);
2821 // If there's a predecessor with an invoke, visit the invoke as if it were
2822 // part of this block, since we can't insert code after an invoke in its own
2823 // block, and we don't want to split critical edges.
2824 for (BBState::edge_iterator PI(MyStates.pred_begin()),
2825 PE(MyStates.pred_end()); PI != PE; ++PI) {
2826 BasicBlock *Pred = *PI;
2827 if (InvokeInst *II = dyn_cast<InvokeInst>(&Pred->back()))
2828 NestingDetected |= VisitInstructionBottomUp(II, BB, Retains, MyStates);
2831 return NestingDetected;
2835 ObjCARCOpt::VisitInstructionTopDown(Instruction *Inst,
2836 DenseMap<Value *, RRInfo> &Releases,
2837 BBState &MyStates) {
2838 bool NestingDetected = false;
2839 InstructionClass Class = GetInstructionClass(Inst);
2840 const Value *Arg = 0;
2843 case IC_RetainBlock:
2844 // An objc_retainBlock call with just a use may need to be kept,
2845 // because it may be copying a block from the stack to the heap.
2846 if (!IsRetainBlockOptimizable(Inst))
2851 Arg = GetObjCArg(Inst);
2853 PtrState &S = MyStates.getPtrTopDownState(Arg);
2855 // Don't do retain+release tracking for IC_RetainRV, because it's
2856 // better to let it remain as the first instruction after a call.
2857 if (Class != IC_RetainRV) {
2858 // If we see two retains in a row on the same pointer. If so, make
2859 // a note, and we'll cicle back to revisit it after we've
2860 // hopefully eliminated the second retain, which may allow us to
2861 // eliminate the first retain too.
2862 // Theoretically we could implement removal of nested retain+release
2863 // pairs by making PtrState hold a stack of states, but this is
2864 // simple and avoids adding overhead for the non-nested case.
2865 if (S.GetSeq() == S_Retain)
2866 NestingDetected = true;
2868 S.ResetSequenceProgress(S_Retain);
2869 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2870 S.RRI.KnownSafe = S.IsKnownIncremented();
2871 S.RRI.Calls.insert(Inst);
2874 S.SetKnownPositiveRefCount();
2876 // A retain can be a potential use; procede to the generic checking
2881 Arg = GetObjCArg(Inst);
2883 PtrState &S = MyStates.getPtrTopDownState(Arg);
2886 switch (S.GetSeq()) {
2889 S.RRI.ReverseInsertPts.clear();
2892 S.RRI.ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2893 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2894 Releases[Inst] = S.RRI;
2895 S.ClearSequenceProgress();
2901 case S_MovableRelease:
2902 llvm_unreachable("top-down pointer in release state!");
2906 case IC_AutoreleasepoolPop:
2907 // Conservatively, clear MyStates for all known pointers.
2908 MyStates.clearTopDownPointers();
2909 return NestingDetected;
2910 case IC_AutoreleasepoolPush:
2912 // These are irrelevant.
2913 return NestingDetected;
2918 // Consider any other possible effects of this instruction on each
2919 // pointer being tracked.
2920 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
2921 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
2922 const Value *Ptr = MI->first;
2924 continue; // Handled above.
2925 PtrState &S = MI->second;
2926 Sequence Seq = S.GetSeq();
2928 // Check for possible releases.
2929 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2933 S.SetSeq(S_CanRelease);
2934 assert(S.RRI.ReverseInsertPts.empty());
2935 S.RRI.ReverseInsertPts.insert(Inst);
2937 // One call can't cause a transition from S_Retain to S_CanRelease
2938 // and S_CanRelease to S_Use. If we've made the first transition,
2947 case S_MovableRelease:
2948 llvm_unreachable("top-down pointer in release state!");
2952 // Check for possible direct uses.
2955 if (CanUse(Inst, Ptr, PA, Class))
2964 case S_MovableRelease:
2965 llvm_unreachable("top-down pointer in release state!");
2969 return NestingDetected;
2973 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
2974 DenseMap<const BasicBlock *, BBState> &BBStates,
2975 DenseMap<Value *, RRInfo> &Releases) {
2976 bool NestingDetected = false;
2977 BBState &MyStates = BBStates[BB];
2979 // Merge the states from each predecessor to compute the initial state
2980 // for the current block.
2981 BBState::edge_iterator PI(MyStates.pred_begin()),
2982 PE(MyStates.pred_end());
2984 const BasicBlock *Pred = *PI;
2985 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
2986 assert(I != BBStates.end());
2987 MyStates.InitFromPred(I->second);
2989 for (; PI != PE; ++PI) {
2991 I = BBStates.find(Pred);
2992 assert(I != BBStates.end());
2993 MyStates.MergePred(I->second);
2997 // Visit all the instructions, top-down.
2998 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2999 Instruction *Inst = I;
3000 NestingDetected |= VisitInstructionTopDown(Inst, Releases, MyStates);
3003 CheckForCFGHazards(BB, BBStates, MyStates);
3004 return NestingDetected;
3008 ComputePostOrders(Function &F,
3009 SmallVectorImpl<BasicBlock *> &PostOrder,
3010 SmallVectorImpl<BasicBlock *> &ReverseCFGPostOrder,
3011 unsigned NoObjCARCExceptionsMDKind,
3012 DenseMap<const BasicBlock *, BBState> &BBStates) {
3013 /// Visited - The visited set, for doing DFS walks.
3014 SmallPtrSet<BasicBlock *, 16> Visited;
3016 // Do DFS, computing the PostOrder.
3017 SmallPtrSet<BasicBlock *, 16> OnStack;
3018 SmallVector<std::pair<BasicBlock *, succ_iterator>, 16> SuccStack;
3020 // Functions always have exactly one entry block, and we don't have
3021 // any other block that we treat like an entry block.
3022 BasicBlock *EntryBB = &F.getEntryBlock();
3023 BBState &MyStates = BBStates[EntryBB];
3024 MyStates.SetAsEntry();
3025 TerminatorInst *EntryTI = cast<TerminatorInst>(&EntryBB->back());
3026 SuccStack.push_back(std::make_pair(EntryBB, succ_iterator(EntryTI)));
3027 Visited.insert(EntryBB);
3028 OnStack.insert(EntryBB);
3031 BasicBlock *CurrBB = SuccStack.back().first;
3032 TerminatorInst *TI = cast<TerminatorInst>(&CurrBB->back());
3033 succ_iterator SE(TI, false);
3035 // If the terminator is an invoke marked with the
3036 // clang.arc.no_objc_arc_exceptions metadata, the unwind edge can be
3037 // ignored, for ARC purposes.
3038 if (isa<InvokeInst>(TI) && TI->getMetadata(NoObjCARCExceptionsMDKind))
3041 while (SuccStack.back().second != SE) {
3042 BasicBlock *SuccBB = *SuccStack.back().second++;
3043 if (Visited.insert(SuccBB)) {
3044 TerminatorInst *TI = cast<TerminatorInst>(&SuccBB->back());
3045 SuccStack.push_back(std::make_pair(SuccBB, succ_iterator(TI)));
3046 BBStates[CurrBB].addSucc(SuccBB);
3047 BBState &SuccStates = BBStates[SuccBB];
3048 SuccStates.addPred(CurrBB);
3049 OnStack.insert(SuccBB);
3053 if (!OnStack.count(SuccBB)) {
3054 BBStates[CurrBB].addSucc(SuccBB);
3055 BBStates[SuccBB].addPred(CurrBB);
3058 OnStack.erase(CurrBB);
3059 PostOrder.push_back(CurrBB);
3060 SuccStack.pop_back();
3061 } while (!SuccStack.empty());
3065 // Do reverse-CFG DFS, computing the reverse-CFG PostOrder.
3066 // Functions may have many exits, and there also blocks which we treat
3067 // as exits due to ignored edges.
3068 SmallVector<std::pair<BasicBlock *, BBState::edge_iterator>, 16> PredStack;
3069 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
3070 BasicBlock *ExitBB = I;
3071 BBState &MyStates = BBStates[ExitBB];
3072 if (!MyStates.isExit())
3075 MyStates.SetAsExit();
3077 PredStack.push_back(std::make_pair(ExitBB, MyStates.pred_begin()));
3078 Visited.insert(ExitBB);
3079 while (!PredStack.empty()) {
3080 reverse_dfs_next_succ:
3081 BBState::edge_iterator PE = BBStates[PredStack.back().first].pred_end();
3082 while (PredStack.back().second != PE) {
3083 BasicBlock *BB = *PredStack.back().second++;
3084 if (Visited.insert(BB)) {
3085 PredStack.push_back(std::make_pair(BB, BBStates[BB].pred_begin()));
3086 goto reverse_dfs_next_succ;
3089 ReverseCFGPostOrder.push_back(PredStack.pop_back_val().first);
3094 // Visit - Visit the function both top-down and bottom-up.
3096 ObjCARCOpt::Visit(Function &F,
3097 DenseMap<const BasicBlock *, BBState> &BBStates,
3098 MapVector<Value *, RRInfo> &Retains,
3099 DenseMap<Value *, RRInfo> &Releases) {
3101 // Use reverse-postorder traversals, because we magically know that loops
3102 // will be well behaved, i.e. they won't repeatedly call retain on a single
3103 // pointer without doing a release. We can't use the ReversePostOrderTraversal
3104 // class here because we want the reverse-CFG postorder to consider each
3105 // function exit point, and we want to ignore selected cycle edges.
3106 SmallVector<BasicBlock *, 16> PostOrder;
3107 SmallVector<BasicBlock *, 16> ReverseCFGPostOrder;
3108 ComputePostOrders(F, PostOrder, ReverseCFGPostOrder,
3109 NoObjCARCExceptionsMDKind,
3112 // Use reverse-postorder on the reverse CFG for bottom-up.
3113 bool BottomUpNestingDetected = false;
3114 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
3115 ReverseCFGPostOrder.rbegin(), E = ReverseCFGPostOrder.rend();
3117 BottomUpNestingDetected |= VisitBottomUp(*I, BBStates, Retains);
3119 // Use reverse-postorder for top-down.
3120 bool TopDownNestingDetected = false;
3121 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
3122 PostOrder.rbegin(), E = PostOrder.rend();
3124 TopDownNestingDetected |= VisitTopDown(*I, BBStates, Releases);
3126 return TopDownNestingDetected && BottomUpNestingDetected;
3129 /// MoveCalls - Move the calls in RetainsToMove and ReleasesToMove.
3130 void ObjCARCOpt::MoveCalls(Value *Arg,
3131 RRInfo &RetainsToMove,
3132 RRInfo &ReleasesToMove,
3133 MapVector<Value *, RRInfo> &Retains,
3134 DenseMap<Value *, RRInfo> &Releases,
3135 SmallVectorImpl<Instruction *> &DeadInsts,
3137 Type *ArgTy = Arg->getType();
3138 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
3140 // Insert the new retain and release calls.
3141 for (SmallPtrSet<Instruction *, 2>::const_iterator
3142 PI = ReleasesToMove.ReverseInsertPts.begin(),
3143 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
3144 Instruction *InsertPt = *PI;
3145 Value *MyArg = ArgTy == ParamTy ? Arg :
3146 new BitCastInst(Arg, ParamTy, "", InsertPt);
3148 CallInst::Create(RetainsToMove.IsRetainBlock ?
3149 getRetainBlockCallee(M) : getRetainCallee(M),
3150 MyArg, "", InsertPt);
3151 Call->setDoesNotThrow();
3152 if (RetainsToMove.IsRetainBlock)
3153 Call->setMetadata(CopyOnEscapeMDKind,
3154 MDNode::get(M->getContext(), ArrayRef<Value *>()));
3156 Call->setTailCall();
3158 for (SmallPtrSet<Instruction *, 2>::const_iterator
3159 PI = RetainsToMove.ReverseInsertPts.begin(),
3160 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
3161 Instruction *InsertPt = *PI;
3162 Value *MyArg = ArgTy == ParamTy ? Arg :
3163 new BitCastInst(Arg, ParamTy, "", InsertPt);
3164 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
3166 // Attach a clang.imprecise_release metadata tag, if appropriate.
3167 if (MDNode *M = ReleasesToMove.ReleaseMetadata)
3168 Call->setMetadata(ImpreciseReleaseMDKind, M);
3169 Call->setDoesNotThrow();
3170 if (ReleasesToMove.IsTailCallRelease)
3171 Call->setTailCall();
3174 // Delete the original retain and release calls.
3175 for (SmallPtrSet<Instruction *, 2>::const_iterator
3176 AI = RetainsToMove.Calls.begin(),
3177 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
3178 Instruction *OrigRetain = *AI;
3179 Retains.blot(OrigRetain);
3180 DeadInsts.push_back(OrigRetain);
3182 for (SmallPtrSet<Instruction *, 2>::const_iterator
3183 AI = ReleasesToMove.Calls.begin(),
3184 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
3185 Instruction *OrigRelease = *AI;
3186 Releases.erase(OrigRelease);
3187 DeadInsts.push_back(OrigRelease);
3191 /// PerformCodePlacement - Identify pairings between the retains and releases,
3192 /// and delete and/or move them.
3194 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
3196 MapVector<Value *, RRInfo> &Retains,
3197 DenseMap<Value *, RRInfo> &Releases,
3199 bool AnyPairsCompletelyEliminated = false;
3200 RRInfo RetainsToMove;
3201 RRInfo ReleasesToMove;
3202 SmallVector<Instruction *, 4> NewRetains;
3203 SmallVector<Instruction *, 4> NewReleases;
3204 SmallVector<Instruction *, 8> DeadInsts;
3206 // Visit each retain.
3207 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
3208 E = Retains.end(); I != E; ++I) {
3209 Value *V = I->first;
3210 if (!V) continue; // blotted
3212 Instruction *Retain = cast<Instruction>(V);
3213 Value *Arg = GetObjCArg(Retain);
3215 // If the object being released is in static or stack storage, we know it's
3216 // not being managed by ObjC reference counting, so we can delete pairs
3217 // regardless of what possible decrements or uses lie between them.
3218 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg);
3220 // A constant pointer can't be pointing to an object on the heap. It may
3221 // be reference-counted, but it won't be deleted.
3222 if (const LoadInst *LI = dyn_cast<LoadInst>(Arg))
3223 if (const GlobalVariable *GV =
3224 dyn_cast<GlobalVariable>(
3225 StripPointerCastsAndObjCCalls(LI->getPointerOperand())))
3226 if (GV->isConstant())
3229 // If a pair happens in a region where it is known that the reference count
3230 // is already incremented, we can similarly ignore possible decrements.
3231 bool KnownSafeTD = true, KnownSafeBU = true;
3233 // Connect the dots between the top-down-collected RetainsToMove and
3234 // bottom-up-collected ReleasesToMove to form sets of related calls.
3235 // This is an iterative process so that we connect multiple releases
3236 // to multiple retains if needed.
3237 unsigned OldDelta = 0;
3238 unsigned NewDelta = 0;
3239 unsigned OldCount = 0;
3240 unsigned NewCount = 0;
3241 bool FirstRelease = true;
3242 bool FirstRetain = true;
3243 NewRetains.push_back(Retain);
3245 for (SmallVectorImpl<Instruction *>::const_iterator
3246 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
3247 Instruction *NewRetain = *NI;
3248 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
3249 assert(It != Retains.end());
3250 const RRInfo &NewRetainRRI = It->second;
3251 KnownSafeTD &= NewRetainRRI.KnownSafe;
3252 for (SmallPtrSet<Instruction *, 2>::const_iterator
3253 LI = NewRetainRRI.Calls.begin(),
3254 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
3255 Instruction *NewRetainRelease = *LI;
3256 DenseMap<Value *, RRInfo>::const_iterator Jt =
3257 Releases.find(NewRetainRelease);
3258 if (Jt == Releases.end())
3260 const RRInfo &NewRetainReleaseRRI = Jt->second;
3261 assert(NewRetainReleaseRRI.Calls.count(NewRetain));
3262 if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
3264 BBStates[NewRetainRelease->getParent()].GetAllPathCount();
3266 // Merge the ReleaseMetadata and IsTailCallRelease values.
3268 ReleasesToMove.ReleaseMetadata =
3269 NewRetainReleaseRRI.ReleaseMetadata;
3270 ReleasesToMove.IsTailCallRelease =
3271 NewRetainReleaseRRI.IsTailCallRelease;
3272 FirstRelease = false;
3274 if (ReleasesToMove.ReleaseMetadata !=
3275 NewRetainReleaseRRI.ReleaseMetadata)
3276 ReleasesToMove.ReleaseMetadata = 0;
3277 if (ReleasesToMove.IsTailCallRelease !=
3278 NewRetainReleaseRRI.IsTailCallRelease)
3279 ReleasesToMove.IsTailCallRelease = false;
3282 // Collect the optimal insertion points.
3284 for (SmallPtrSet<Instruction *, 2>::const_iterator
3285 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
3286 RE = NewRetainReleaseRRI.ReverseInsertPts.end();
3288 Instruction *RIP = *RI;
3289 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
3290 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
3292 NewReleases.push_back(NewRetainRelease);
3297 if (NewReleases.empty()) break;
3299 // Back the other way.
3300 for (SmallVectorImpl<Instruction *>::const_iterator
3301 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
3302 Instruction *NewRelease = *NI;
3303 DenseMap<Value *, RRInfo>::const_iterator It =
3304 Releases.find(NewRelease);
3305 assert(It != Releases.end());
3306 const RRInfo &NewReleaseRRI = It->second;
3307 KnownSafeBU &= NewReleaseRRI.KnownSafe;
3308 for (SmallPtrSet<Instruction *, 2>::const_iterator
3309 LI = NewReleaseRRI.Calls.begin(),
3310 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
3311 Instruction *NewReleaseRetain = *LI;
3312 MapVector<Value *, RRInfo>::const_iterator Jt =
3313 Retains.find(NewReleaseRetain);
3314 if (Jt == Retains.end())
3316 const RRInfo &NewReleaseRetainRRI = Jt->second;
3317 assert(NewReleaseRetainRRI.Calls.count(NewRelease));
3318 if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
3319 unsigned PathCount =
3320 BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
3321 OldDelta += PathCount;
3322 OldCount += PathCount;
3324 // Merge the IsRetainBlock values.
3326 RetainsToMove.IsRetainBlock = NewReleaseRetainRRI.IsRetainBlock;
3327 FirstRetain = false;
3328 } else if (ReleasesToMove.IsRetainBlock !=
3329 NewReleaseRetainRRI.IsRetainBlock)
3330 // It's not possible to merge the sequences if one uses
3331 // objc_retain and the other uses objc_retainBlock.
3334 // Collect the optimal insertion points.
3336 for (SmallPtrSet<Instruction *, 2>::const_iterator
3337 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
3338 RE = NewReleaseRetainRRI.ReverseInsertPts.end();
3340 Instruction *RIP = *RI;
3341 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
3342 PathCount = BBStates[RIP->getParent()].GetAllPathCount();
3343 NewDelta += PathCount;
3344 NewCount += PathCount;
3347 NewRetains.push_back(NewReleaseRetain);
3351 NewReleases.clear();
3352 if (NewRetains.empty()) break;
3355 // If the pointer is known incremented or nested, we can safely delete the
3356 // pair regardless of what's between them.
3357 if (KnownSafeTD || KnownSafeBU) {
3358 RetainsToMove.ReverseInsertPts.clear();
3359 ReleasesToMove.ReverseInsertPts.clear();
3362 // Determine whether the new insertion points we computed preserve the
3363 // balance of retain and release calls through the program.
3364 // TODO: If the fully aggressive solution isn't valid, try to find a
3365 // less aggressive solution which is.
3370 // Determine whether the original call points are balanced in the retain and
3371 // release calls through the program. If not, conservatively don't touch
3373 // TODO: It's theoretically possible to do code motion in this case, as
3374 // long as the existing imbalances are maintained.
3378 // Ok, everything checks out and we're all set. Let's move some code!
3380 assert(OldCount != 0 && "Unreachable code?");
3381 AnyPairsCompletelyEliminated = NewCount == 0;
3382 NumRRs += OldCount - NewCount;
3383 MoveCalls(Arg, RetainsToMove, ReleasesToMove,
3384 Retains, Releases, DeadInsts, M);
3387 NewReleases.clear();
3389 RetainsToMove.clear();
3390 ReleasesToMove.clear();
3393 // Now that we're done moving everything, we can delete the newly dead
3394 // instructions, as we no longer need them as insert points.
3395 while (!DeadInsts.empty())
3396 EraseInstruction(DeadInsts.pop_back_val());
3398 return AnyPairsCompletelyEliminated;
3401 /// OptimizeWeakCalls - Weak pointer optimizations.
3402 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
3403 // First, do memdep-style RLE and S2L optimizations. We can't use memdep
3404 // itself because it uses AliasAnalysis and we need to do provenance
3406 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3407 Instruction *Inst = &*I++;
3409 DEBUG(dbgs() << "ObjCARCOpt: OptimizeWeakCalls: Visiting: " << *Inst <<
3412 InstructionClass Class = GetBasicInstructionClass(Inst);
3413 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
3416 // Delete objc_loadWeak calls with no users.
3417 if (Class == IC_LoadWeak && Inst->use_empty()) {
3418 Inst->eraseFromParent();
3422 // TODO: For now, just look for an earlier available version of this value
3423 // within the same block. Theoretically, we could do memdep-style non-local
3424 // analysis too, but that would want caching. A better approach would be to
3425 // use the technique that EarlyCSE uses.
3426 inst_iterator Current = llvm::prior(I);
3427 BasicBlock *CurrentBB = Current.getBasicBlockIterator();
3428 for (BasicBlock::iterator B = CurrentBB->begin(),
3429 J = Current.getInstructionIterator();
3431 Instruction *EarlierInst = &*llvm::prior(J);
3432 InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
3433 switch (EarlierClass) {
3435 case IC_LoadWeakRetained: {
3436 // If this is loading from the same pointer, replace this load's value
3438 CallInst *Call = cast<CallInst>(Inst);
3439 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
3440 Value *Arg = Call->getArgOperand(0);
3441 Value *EarlierArg = EarlierCall->getArgOperand(0);
3442 switch (PA.getAA()->alias(Arg, EarlierArg)) {
3443 case AliasAnalysis::MustAlias:
3445 // If the load has a builtin retain, insert a plain retain for it.
3446 if (Class == IC_LoadWeakRetained) {
3448 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
3452 // Zap the fully redundant load.
3453 Call->replaceAllUsesWith(EarlierCall);
3454 Call->eraseFromParent();
3456 case AliasAnalysis::MayAlias:
3457 case AliasAnalysis::PartialAlias:
3459 case AliasAnalysis::NoAlias:
3466 // If this is storing to the same pointer and has the same size etc.
3467 // replace this load's value with the stored value.
3468 CallInst *Call = cast<CallInst>(Inst);
3469 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
3470 Value *Arg = Call->getArgOperand(0);
3471 Value *EarlierArg = EarlierCall->getArgOperand(0);
3472 switch (PA.getAA()->alias(Arg, EarlierArg)) {
3473 case AliasAnalysis::MustAlias:
3475 // If the load has a builtin retain, insert a plain retain for it.
3476 if (Class == IC_LoadWeakRetained) {
3478 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
3482 // Zap the fully redundant load.
3483 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
3484 Call->eraseFromParent();
3486 case AliasAnalysis::MayAlias:
3487 case AliasAnalysis::PartialAlias:
3489 case AliasAnalysis::NoAlias:
3496 // TOOD: Grab the copied value.
3498 case IC_AutoreleasepoolPush:
3501 // Weak pointers are only modified through the weak entry points
3502 // (and arbitrary calls, which could call the weak entry points).
3505 // Anything else could modify the weak pointer.
3512 // Then, for each destroyWeak with an alloca operand, check to see if
3513 // the alloca and all its users can be zapped.
3514 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3515 Instruction *Inst = &*I++;
3516 InstructionClass Class = GetBasicInstructionClass(Inst);
3517 if (Class != IC_DestroyWeak)
3520 CallInst *Call = cast<CallInst>(Inst);
3521 Value *Arg = Call->getArgOperand(0);
3522 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
3523 for (Value::use_iterator UI = Alloca->use_begin(),
3524 UE = Alloca->use_end(); UI != UE; ++UI) {
3525 const Instruction *UserInst = cast<Instruction>(*UI);
3526 switch (GetBasicInstructionClass(UserInst)) {
3529 case IC_DestroyWeak:
3536 for (Value::use_iterator UI = Alloca->use_begin(),
3537 UE = Alloca->use_end(); UI != UE; ) {
3538 CallInst *UserInst = cast<CallInst>(*UI++);
3539 switch (GetBasicInstructionClass(UserInst)) {
3542 // These functions return their second argument.
3543 UserInst->replaceAllUsesWith(UserInst->getArgOperand(1));
3545 case IC_DestroyWeak:
3549 llvm_unreachable("alloca really is used!");
3551 UserInst->eraseFromParent();
3553 Alloca->eraseFromParent();
3558 DEBUG(dbgs() << "ObjCARCOpt: Finished visiting weak calls.\n\n");
3562 /// OptimizeSequences - Identify program paths which execute sequences of
3563 /// retains and releases which can be eliminated.
3564 bool ObjCARCOpt::OptimizeSequences(Function &F) {
3565 /// Releases, Retains - These are used to store the results of the main flow
3566 /// analysis. These use Value* as the key instead of Instruction* so that the
3567 /// map stays valid when we get around to rewriting code and calls get
3568 /// replaced by arguments.
3569 DenseMap<Value *, RRInfo> Releases;
3570 MapVector<Value *, RRInfo> Retains;
3572 /// BBStates, This is used during the traversal of the function to track the
3573 /// states for each identified object at each block.
3574 DenseMap<const BasicBlock *, BBState> BBStates;
3576 // Analyze the CFG of the function, and all instructions.
3577 bool NestingDetected = Visit(F, BBStates, Retains, Releases);
3580 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
3584 /// OptimizeReturns - Look for this pattern:
3586 /// %call = call i8* @something(...)
3587 /// %2 = call i8* @objc_retain(i8* %call)
3588 /// %3 = call i8* @objc_autorelease(i8* %2)
3591 /// And delete the retain and autorelease.
3593 /// Otherwise if it's just this:
3595 /// %3 = call i8* @objc_autorelease(i8* %2)
3598 /// convert the autorelease to autoreleaseRV.
3599 void ObjCARCOpt::OptimizeReturns(Function &F) {
3600 if (!F.getReturnType()->isPointerTy())
3603 SmallPtrSet<Instruction *, 4> DependingInstructions;
3604 SmallPtrSet<const BasicBlock *, 4> Visited;
3605 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
3606 BasicBlock *BB = FI;
3607 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
3609 DEBUG(dbgs() << "ObjCARCOpt: OptimizeReturns: Visiting: " << *Ret << "\n");
3613 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
3614 FindDependencies(NeedsPositiveRetainCount, Arg,
3615 BB, Ret, DependingInstructions, Visited, PA);
3616 if (DependingInstructions.size() != 1)
3620 CallInst *Autorelease =
3621 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3624 InstructionClass AutoreleaseClass = GetBasicInstructionClass(Autorelease);
3625 if (!IsAutorelease(AutoreleaseClass))
3627 if (GetObjCArg(Autorelease) != Arg)
3630 DependingInstructions.clear();
3633 // Check that there is nothing that can affect the reference
3634 // count between the autorelease and the retain.
3635 FindDependencies(CanChangeRetainCount, Arg,
3636 BB, Autorelease, DependingInstructions, Visited, PA);
3637 if (DependingInstructions.size() != 1)
3642 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3644 // Check that we found a retain with the same argument.
3646 !IsRetain(GetBasicInstructionClass(Retain)) ||
3647 GetObjCArg(Retain) != Arg)
3650 DependingInstructions.clear();
3653 // Convert the autorelease to an autoreleaseRV, since it's
3654 // returning the value.
3655 if (AutoreleaseClass == IC_Autorelease) {
3656 Autorelease->setCalledFunction(getAutoreleaseRVCallee(F.getParent()));
3657 AutoreleaseClass = IC_AutoreleaseRV;
3660 // Check that there is nothing that can affect the reference
3661 // count between the retain and the call.
3662 // Note that Retain need not be in BB.
3663 FindDependencies(CanChangeRetainCount, Arg, Retain->getParent(), Retain,
3664 DependingInstructions, Visited, PA);
3665 if (DependingInstructions.size() != 1)
3670 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3672 // Check that the pointer is the return value of the call.
3673 if (!Call || Arg != Call)
3676 // Check that the call is a regular call.
3677 InstructionClass Class = GetBasicInstructionClass(Call);
3678 if (Class != IC_CallOrUser && Class != IC_Call)
3681 // If so, we can zap the retain and autorelease.
3684 EraseInstruction(Retain);
3685 EraseInstruction(Autorelease);
3691 DependingInstructions.clear();
3695 DEBUG(dbgs() << "ObjCARCOpt: OptimizeReturns: Finished visiting returns.\n\n");
3699 bool ObjCARCOpt::doInitialization(Module &M) {
3703 // If nothing in the Module uses ARC, don't do anything.
3704 Run = ModuleHasARC(M);
3708 // Identify the imprecise release metadata kind.
3709 ImpreciseReleaseMDKind =
3710 M.getContext().getMDKindID("clang.imprecise_release");
3711 CopyOnEscapeMDKind =
3712 M.getContext().getMDKindID("clang.arc.copy_on_escape");
3713 NoObjCARCExceptionsMDKind =
3714 M.getContext().getMDKindID("clang.arc.no_objc_arc_exceptions");
3716 // Intuitively, objc_retain and others are nocapture, however in practice
3717 // they are not, because they return their argument value. And objc_release
3718 // calls finalizers which can have arbitrary side effects.
3720 // These are initialized lazily.
3722 AutoreleaseRVCallee = 0;
3725 RetainBlockCallee = 0;
3726 AutoreleaseCallee = 0;
3731 bool ObjCARCOpt::runOnFunction(Function &F) {
3735 // If nothing in the Module uses ARC, don't do anything.
3741 PA.setAA(&getAnalysis<AliasAnalysis>());
3743 // This pass performs several distinct transformations. As a compile-time aid
3744 // when compiling code that isn't ObjC, skip these if the relevant ObjC
3745 // library functions aren't declared.
3747 // Preliminary optimizations. This also computs UsedInThisFunction.
3748 OptimizeIndividualCalls(F);
3750 // Optimizations for weak pointers.
3751 if (UsedInThisFunction & ((1 << IC_LoadWeak) |
3752 (1 << IC_LoadWeakRetained) |
3753 (1 << IC_StoreWeak) |
3754 (1 << IC_InitWeak) |
3755 (1 << IC_CopyWeak) |
3756 (1 << IC_MoveWeak) |
3757 (1 << IC_DestroyWeak)))
3758 OptimizeWeakCalls(F);
3760 // Optimizations for retain+release pairs.
3761 if (UsedInThisFunction & ((1 << IC_Retain) |
3762 (1 << IC_RetainRV) |
3763 (1 << IC_RetainBlock)))
3764 if (UsedInThisFunction & (1 << IC_Release))
3765 // Run OptimizeSequences until it either stops making changes or
3766 // no retain+release pair nesting is detected.
3767 while (OptimizeSequences(F)) {}
3769 // Optimizations if objc_autorelease is used.
3770 if (UsedInThisFunction & ((1 << IC_Autorelease) |
3771 (1 << IC_AutoreleaseRV)))
3777 void ObjCARCOpt::releaseMemory() {
3781 //===----------------------------------------------------------------------===//
3783 //===----------------------------------------------------------------------===//
3785 // TODO: ObjCARCContract could insert PHI nodes when uses aren't
3786 // dominated by single calls.
3788 #include "llvm/Analysis/Dominators.h"
3789 #include "llvm/IR/InlineAsm.h"
3790 #include "llvm/IR/Operator.h"
3792 STATISTIC(NumStoreStrongs, "Number objc_storeStrong calls formed");
3795 /// ObjCARCContract - Late ARC optimizations. These change the IR in a way
3796 /// that makes it difficult to be analyzed by ObjCARCOpt, so it's run late.
3797 class ObjCARCContract : public FunctionPass {
3801 ProvenanceAnalysis PA;
3803 /// Run - A flag indicating whether this optimization pass should run.
3806 /// StoreStrongCallee, etc. - Declarations for ObjC runtime
3807 /// functions, for use in creating calls to them. These are initialized
3808 /// lazily to avoid cluttering up the Module with unused declarations.
3809 Constant *StoreStrongCallee,
3810 *RetainAutoreleaseCallee, *RetainAutoreleaseRVCallee;
3812 /// RetainRVMarker - The inline asm string to insert between calls and
3813 /// RetainRV calls to make the optimization work on targets which need it.
3814 const MDString *RetainRVMarker;
3816 /// StoreStrongCalls - The set of inserted objc_storeStrong calls. If
3817 /// at the end of walking the function we have found no alloca
3818 /// instructions, these calls can be marked "tail".
3819 SmallPtrSet<CallInst *, 8> StoreStrongCalls;
3821 Constant *getStoreStrongCallee(Module *M);
3822 Constant *getRetainAutoreleaseCallee(Module *M);
3823 Constant *getRetainAutoreleaseRVCallee(Module *M);
3825 bool ContractAutorelease(Function &F, Instruction *Autorelease,
3826 InstructionClass Class,
3827 SmallPtrSet<Instruction *, 4>
3828 &DependingInstructions,
3829 SmallPtrSet<const BasicBlock *, 4>
3832 void ContractRelease(Instruction *Release,
3833 inst_iterator &Iter);
3835 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
3836 virtual bool doInitialization(Module &M);
3837 virtual bool runOnFunction(Function &F);
3841 ObjCARCContract() : FunctionPass(ID) {
3842 initializeObjCARCContractPass(*PassRegistry::getPassRegistry());
3847 char ObjCARCContract::ID = 0;
3848 INITIALIZE_PASS_BEGIN(ObjCARCContract,
3849 "objc-arc-contract", "ObjC ARC contraction", false, false)
3850 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3851 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3852 INITIALIZE_PASS_END(ObjCARCContract,
3853 "objc-arc-contract", "ObjC ARC contraction", false, false)
3855 Pass *llvm::createObjCARCContractPass() {
3856 return new ObjCARCContract();
3859 void ObjCARCContract::getAnalysisUsage(AnalysisUsage &AU) const {
3860 AU.addRequired<AliasAnalysis>();
3861 AU.addRequired<DominatorTree>();
3862 AU.setPreservesCFG();
3865 Constant *ObjCARCContract::getStoreStrongCallee(Module *M) {
3866 if (!StoreStrongCallee) {
3867 LLVMContext &C = M->getContext();
3868 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3869 Type *I8XX = PointerType::getUnqual(I8X);
3870 Type *Params[] = { I8XX, I8X };
3872 AttributeSet Attribute = AttributeSet()
3873 .addAttr(M->getContext(), AttributeSet::FunctionIndex,
3874 Attribute::get(C, Attribute::NoUnwind))
3875 .addAttr(M->getContext(), 1, Attribute::get(C, Attribute::NoCapture));
3878 M->getOrInsertFunction(
3880 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
3883 return StoreStrongCallee;
3886 Constant *ObjCARCContract::getRetainAutoreleaseCallee(Module *M) {
3887 if (!RetainAutoreleaseCallee) {
3888 LLVMContext &C = M->getContext();
3889 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3890 Type *Params[] = { I8X };
3891 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
3892 AttributeSet Attribute =
3893 AttributeSet().addAttr(M->getContext(), AttributeSet::FunctionIndex,
3894 Attribute::get(C, Attribute::NoUnwind));
3895 RetainAutoreleaseCallee =
3896 M->getOrInsertFunction("objc_retainAutorelease", FTy, Attribute);
3898 return RetainAutoreleaseCallee;
3901 Constant *ObjCARCContract::getRetainAutoreleaseRVCallee(Module *M) {
3902 if (!RetainAutoreleaseRVCallee) {
3903 LLVMContext &C = M->getContext();
3904 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3905 Type *Params[] = { I8X };
3906 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
3907 AttributeSet Attribute =
3908 AttributeSet().addAttr(M->getContext(), AttributeSet::FunctionIndex,
3909 Attribute::get(C, Attribute::NoUnwind));
3910 RetainAutoreleaseRVCallee =
3911 M->getOrInsertFunction("objc_retainAutoreleaseReturnValue", FTy,
3914 return RetainAutoreleaseRVCallee;
3917 /// ContractAutorelease - Merge an autorelease with a retain into a fused call.
3919 ObjCARCContract::ContractAutorelease(Function &F, Instruction *Autorelease,
3920 InstructionClass Class,
3921 SmallPtrSet<Instruction *, 4>
3922 &DependingInstructions,
3923 SmallPtrSet<const BasicBlock *, 4>
3925 const Value *Arg = GetObjCArg(Autorelease);
3927 // Check that there are no instructions between the retain and the autorelease
3928 // (such as an autorelease_pop) which may change the count.
3929 CallInst *Retain = 0;
3930 if (Class == IC_AutoreleaseRV)
3931 FindDependencies(RetainAutoreleaseRVDep, Arg,
3932 Autorelease->getParent(), Autorelease,
3933 DependingInstructions, Visited, PA);
3935 FindDependencies(RetainAutoreleaseDep, Arg,
3936 Autorelease->getParent(), Autorelease,
3937 DependingInstructions, Visited, PA);
3940 if (DependingInstructions.size() != 1) {
3941 DependingInstructions.clear();
3945 Retain = dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3946 DependingInstructions.clear();
3949 GetBasicInstructionClass(Retain) != IC_Retain ||
3950 GetObjCArg(Retain) != Arg)
3956 if (Class == IC_AutoreleaseRV)
3957 Retain->setCalledFunction(getRetainAutoreleaseRVCallee(F.getParent()));
3959 Retain->setCalledFunction(getRetainAutoreleaseCallee(F.getParent()));
3961 EraseInstruction(Autorelease);
3965 /// ContractRelease - Attempt to merge an objc_release with a store, load, and
3966 /// objc_retain to form an objc_storeStrong. This can be a little tricky because
3967 /// the instructions don't always appear in order, and there may be unrelated
3968 /// intervening instructions.
3969 void ObjCARCContract::ContractRelease(Instruction *Release,
3970 inst_iterator &Iter) {
3971 LoadInst *Load = dyn_cast<LoadInst>(GetObjCArg(Release));
3972 if (!Load || !Load->isSimple()) return;
3974 // For now, require everything to be in one basic block.
3975 BasicBlock *BB = Release->getParent();
3976 if (Load->getParent() != BB) return;
3978 // Walk down to find the store and the release, which may be in either order.
3979 BasicBlock::iterator I = Load, End = BB->end();
3981 AliasAnalysis::Location Loc = AA->getLocation(Load);
3982 StoreInst *Store = 0;
3983 bool SawRelease = false;
3984 for (; !Store || !SawRelease; ++I) {
3988 Instruction *Inst = I;
3989 if (Inst == Release) {
3994 InstructionClass Class = GetBasicInstructionClass(Inst);
3996 // Unrelated retains are harmless.
3997 if (IsRetain(Class))
4001 // The store is the point where we're going to put the objc_storeStrong,
4002 // so make sure there are no uses after it.
4003 if (CanUse(Inst, Load, PA, Class))
4005 } else if (AA->getModRefInfo(Inst, Loc) & AliasAnalysis::Mod) {
4006 // We are moving the load down to the store, so check for anything
4007 // else which writes to the memory between the load and the store.
4008 Store = dyn_cast<StoreInst>(Inst);
4009 if (!Store || !Store->isSimple()) return;
4010 if (Store->getPointerOperand() != Loc.Ptr) return;
4014 Value *New = StripPointerCastsAndObjCCalls(Store->getValueOperand());
4016 // Walk up to find the retain.
4018 BasicBlock::iterator Begin = BB->begin();
4019 while (I != Begin && GetBasicInstructionClass(I) != IC_Retain)
4021 Instruction *Retain = I;
4022 if (GetBasicInstructionClass(Retain) != IC_Retain) return;
4023 if (GetObjCArg(Retain) != New) return;
4028 LLVMContext &C = Release->getContext();
4029 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
4030 Type *I8XX = PointerType::getUnqual(I8X);
4032 Value *Args[] = { Load->getPointerOperand(), New };
4033 if (Args[0]->getType() != I8XX)
4034 Args[0] = new BitCastInst(Args[0], I8XX, "", Store);
4035 if (Args[1]->getType() != I8X)
4036 Args[1] = new BitCastInst(Args[1], I8X, "", Store);
4037 CallInst *StoreStrong =
4038 CallInst::Create(getStoreStrongCallee(BB->getParent()->getParent()),
4040 StoreStrong->setDoesNotThrow();
4041 StoreStrong->setDebugLoc(Store->getDebugLoc());
4043 // We can't set the tail flag yet, because we haven't yet determined
4044 // whether there are any escaping allocas. Remember this call, so that
4045 // we can set the tail flag once we know it's safe.
4046 StoreStrongCalls.insert(StoreStrong);
4048 if (&*Iter == Store) ++Iter;
4049 Store->eraseFromParent();
4050 Release->eraseFromParent();
4051 EraseInstruction(Retain);
4052 if (Load->use_empty())
4053 Load->eraseFromParent();
4056 bool ObjCARCContract::doInitialization(Module &M) {
4057 // If nothing in the Module uses ARC, don't do anything.
4058 Run = ModuleHasARC(M);
4062 // These are initialized lazily.
4063 StoreStrongCallee = 0;
4064 RetainAutoreleaseCallee = 0;
4065 RetainAutoreleaseRVCallee = 0;
4067 // Initialize RetainRVMarker.
4069 if (NamedMDNode *NMD =
4070 M.getNamedMetadata("clang.arc.retainAutoreleasedReturnValueMarker"))
4071 if (NMD->getNumOperands() == 1) {
4072 const MDNode *N = NMD->getOperand(0);
4073 if (N->getNumOperands() == 1)
4074 if (const MDString *S = dyn_cast<MDString>(N->getOperand(0)))
4081 bool ObjCARCContract::runOnFunction(Function &F) {
4085 // If nothing in the Module uses ARC, don't do anything.
4090 AA = &getAnalysis<AliasAnalysis>();
4091 DT = &getAnalysis<DominatorTree>();
4093 PA.setAA(&getAnalysis<AliasAnalysis>());
4095 // Track whether it's ok to mark objc_storeStrong calls with the "tail"
4096 // keyword. Be conservative if the function has variadic arguments.
4097 // It seems that functions which "return twice" are also unsafe for the
4098 // "tail" argument, because they are setjmp, which could need to
4099 // return to an earlier stack state.
4100 bool TailOkForStoreStrongs = !F.isVarArg() &&
4101 !F.callsFunctionThatReturnsTwice();
4103 // For ObjC library calls which return their argument, replace uses of the
4104 // argument with uses of the call return value, if it dominates the use. This
4105 // reduces register pressure.
4106 SmallPtrSet<Instruction *, 4> DependingInstructions;
4107 SmallPtrSet<const BasicBlock *, 4> Visited;
4108 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
4109 Instruction *Inst = &*I++;
4111 DEBUG(dbgs() << "ObjCARCContract: Visiting: " << *Inst << "\n");
4113 // Only these library routines return their argument. In particular,
4114 // objc_retainBlock does not necessarily return its argument.
4115 InstructionClass Class = GetBasicInstructionClass(Inst);
4118 case IC_FusedRetainAutorelease:
4119 case IC_FusedRetainAutoreleaseRV:
4121 case IC_Autorelease:
4122 case IC_AutoreleaseRV:
4123 if (ContractAutorelease(F, Inst, Class, DependingInstructions, Visited))
4127 // If we're compiling for a target which needs a special inline-asm
4128 // marker to do the retainAutoreleasedReturnValue optimization,
4130 if (!RetainRVMarker)
4132 BasicBlock::iterator BBI = Inst;
4133 BasicBlock *InstParent = Inst->getParent();
4135 // Step up to see if the call immediately precedes the RetainRV call.
4136 // If it's an invoke, we have to cross a block boundary. And we have
4137 // to carefully dodge no-op instructions.
4139 if (&*BBI == InstParent->begin()) {
4140 BasicBlock *Pred = InstParent->getSinglePredecessor();
4142 goto decline_rv_optimization;
4143 BBI = Pred->getTerminator();
4147 } while (isNoopInstruction(BBI));
4149 if (&*BBI == GetObjCArg(Inst)) {
4150 DEBUG(dbgs() << "ObjCARCContract: Adding inline asm marker for "
4151 "retainAutoreleasedReturnValue optimization.\n");
4154 InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
4155 /*isVarArg=*/false),
4156 RetainRVMarker->getString(),
4157 /*Constraints=*/"", /*hasSideEffects=*/true);
4158 CallInst::Create(IA, "", Inst);
4160 decline_rv_optimization:
4164 // objc_initWeak(p, null) => *p = null
4165 CallInst *CI = cast<CallInst>(Inst);
4166 if (isNullOrUndef(CI->getArgOperand(1))) {
4168 ConstantPointerNull::get(cast<PointerType>(CI->getType()));
4170 new StoreInst(Null, CI->getArgOperand(0), CI);
4171 CI->replaceAllUsesWith(Null);
4172 CI->eraseFromParent();
4177 ContractRelease(Inst, I);
4180 // Be conservative if the function has any alloca instructions.
4181 // Technically we only care about escaping alloca instructions,
4182 // but this is sufficient to handle some interesting cases.
4183 if (isa<AllocaInst>(Inst))
4184 TailOkForStoreStrongs = false;
4190 DEBUG(dbgs() << "ObjCARCContract: Finished Queue.\n\n");
4192 // Don't use GetObjCArg because we don't want to look through bitcasts
4193 // and such; to do the replacement, the argument must have type i8*.
4194 const Value *Arg = cast<CallInst>(Inst)->getArgOperand(0);
4196 // If we're compiling bugpointed code, don't get in trouble.
4197 if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))
4199 // Look through the uses of the pointer.
4200 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
4202 Use &U = UI.getUse();
4203 unsigned OperandNo = UI.getOperandNo();
4204 ++UI; // Increment UI now, because we may unlink its element.
4206 // If the call's return value dominates a use of the call's argument
4207 // value, rewrite the use to use the return value. We check for
4208 // reachability here because an unreachable call is considered to
4209 // trivially dominate itself, which would lead us to rewriting its
4210 // argument in terms of its return value, which would lead to
4211 // infinite loops in GetObjCArg.
4212 if (DT->isReachableFromEntry(U) && DT->dominates(Inst, U)) {
4214 Instruction *Replacement = Inst;
4215 Type *UseTy = U.get()->getType();
4216 if (PHINode *PHI = dyn_cast<PHINode>(U.getUser())) {
4217 // For PHI nodes, insert the bitcast in the predecessor block.
4218 unsigned ValNo = PHINode::getIncomingValueNumForOperand(OperandNo);
4219 BasicBlock *BB = PHI->getIncomingBlock(ValNo);
4220 if (Replacement->getType() != UseTy)
4221 Replacement = new BitCastInst(Replacement, UseTy, "",
4223 // While we're here, rewrite all edges for this PHI, rather
4224 // than just one use at a time, to minimize the number of
4225 // bitcasts we emit.
4226 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4227 if (PHI->getIncomingBlock(i) == BB) {
4228 // Keep the UI iterator valid.
4229 if (&PHI->getOperandUse(
4230 PHINode::getOperandNumForIncomingValue(i)) ==
4233 PHI->setIncomingValue(i, Replacement);
4236 if (Replacement->getType() != UseTy)
4237 Replacement = new BitCastInst(Replacement, UseTy, "",
4238 cast<Instruction>(U.getUser()));
4244 // If Arg is a no-op casted pointer, strip one level of casts and iterate.
4245 if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg))
4246 Arg = BI->getOperand(0);
4247 else if (isa<GEPOperator>(Arg) &&
4248 cast<GEPOperator>(Arg)->hasAllZeroIndices())
4249 Arg = cast<GEPOperator>(Arg)->getPointerOperand();
4250 else if (isa<GlobalAlias>(Arg) &&
4251 !cast<GlobalAlias>(Arg)->mayBeOverridden())
4252 Arg = cast<GlobalAlias>(Arg)->getAliasee();
4258 // If this function has no escaping allocas or suspicious vararg usage,
4259 // objc_storeStrong calls can be marked with the "tail" keyword.
4260 if (TailOkForStoreStrongs)
4261 for (SmallPtrSet<CallInst *, 8>::iterator I = StoreStrongCalls.begin(),
4262 E = StoreStrongCalls.end(); I != E; ++I)
4263 (*I)->setTailCall();
4264 StoreStrongCalls.clear();