• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- ObjCARC.cpp - ObjC ARC Optimization --------------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
13 //
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
18 // simplifications.
19 //
20 // This file also defines a simple ARC-aware AliasAnalysis.
21 //
22 // WARNING: This file knows about certain library functions. It recognizes them
23 // by name, and hardwires knowedge of their semantics.
24 //
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.
28 //
29 //===----------------------------------------------------------------------===//
30 
31 #define DEBUG_TYPE "objc-arc"
32 #include "llvm/Function.h"
33 #include "llvm/Intrinsics.h"
34 #include "llvm/GlobalVariable.h"
35 #include "llvm/DerivedTypes.h"
36 #include "llvm/Module.h"
37 #include "llvm/Analysis/ValueTracking.h"
38 #include "llvm/Transforms/Utils/Local.h"
39 #include "llvm/Support/CallSite.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/ADT/StringSwitch.h"
42 #include "llvm/ADT/DenseMap.h"
43 #include "llvm/ADT/STLExtras.h"
44 using namespace llvm;
45 
46 // A handy option to enable/disable all optimizations in this file.
47 static cl::opt<bool> EnableARCOpts("enable-objc-arc-opts", cl::init(true));
48 
49 //===----------------------------------------------------------------------===//
50 // Misc. Utilities
51 //===----------------------------------------------------------------------===//
52 
53 namespace {
54   /// MapVector - An associative container with fast insertion-order
55   /// (deterministic) iteration over its elements. Plus the special
56   /// blot operation.
57   template<class KeyT, class ValueT>
58   class MapVector {
59     /// Map - Map keys to indices in Vector.
60     typedef DenseMap<KeyT, size_t> MapTy;
61     MapTy Map;
62 
63     /// Vector - Keys and values.
64     typedef std::vector<std::pair<KeyT, ValueT> > VectorTy;
65     VectorTy Vector;
66 
67   public:
68     typedef typename VectorTy::iterator iterator;
69     typedef typename VectorTy::const_iterator const_iterator;
begin()70     iterator begin() { return Vector.begin(); }
end()71     iterator end() { return Vector.end(); }
begin() const72     const_iterator begin() const { return Vector.begin(); }
end() const73     const_iterator end() const { return Vector.end(); }
74 
75 #ifdef XDEBUG
~MapVector()76     ~MapVector() {
77       assert(Vector.size() >= Map.size()); // May differ due to blotting.
78       for (typename MapTy::const_iterator I = Map.begin(), E = Map.end();
79            I != E; ++I) {
80         assert(I->second < Vector.size());
81         assert(Vector[I->second].first == I->first);
82       }
83       for (typename VectorTy::const_iterator I = Vector.begin(),
84            E = Vector.end(); I != E; ++I)
85         assert(!I->first ||
86                (Map.count(I->first) &&
87                 Map[I->first] == size_t(I - Vector.begin())));
88     }
89 #endif
90 
operator [](KeyT Arg)91     ValueT &operator[](KeyT Arg) {
92       std::pair<typename MapTy::iterator, bool> Pair =
93         Map.insert(std::make_pair(Arg, size_t(0)));
94       if (Pair.second) {
95         Pair.first->second = Vector.size();
96         Vector.push_back(std::make_pair(Arg, ValueT()));
97         return Vector.back().second;
98       }
99       return Vector[Pair.first->second].second;
100     }
101 
102     std::pair<iterator, bool>
insert(const std::pair<KeyT,ValueT> & InsertPair)103     insert(const std::pair<KeyT, ValueT> &InsertPair) {
104       std::pair<typename MapTy::iterator, bool> Pair =
105         Map.insert(std::make_pair(InsertPair.first, size_t(0)));
106       if (Pair.second) {
107         Pair.first->second = Vector.size();
108         Vector.push_back(InsertPair);
109         return std::make_pair(llvm::prior(Vector.end()), true);
110       }
111       return std::make_pair(Vector.begin() + Pair.first->second, false);
112     }
113 
find(KeyT Key) const114     const_iterator find(KeyT Key) const {
115       typename MapTy::const_iterator It = Map.find(Key);
116       if (It == Map.end()) return Vector.end();
117       return Vector.begin() + It->second;
118     }
119 
120     /// blot - This is similar to erase, but instead of removing the element
121     /// from the vector, it just zeros out the key in the vector. This leaves
122     /// iterators intact, but clients must be prepared for zeroed-out keys when
123     /// iterating.
blot(KeyT Key)124     void blot(KeyT Key) {
125       typename MapTy::iterator It = Map.find(Key);
126       if (It == Map.end()) return;
127       Vector[It->second].first = KeyT();
128       Map.erase(It);
129     }
130 
clear()131     void clear() {
132       Map.clear();
133       Vector.clear();
134     }
135   };
136 }
137 
138 //===----------------------------------------------------------------------===//
139 // ARC Utilities.
140 //===----------------------------------------------------------------------===//
141 
142 namespace {
143   /// InstructionClass - A simple classification for instructions.
144   enum InstructionClass {
145     IC_Retain,              ///< objc_retain
146     IC_RetainRV,            ///< objc_retainAutoreleasedReturnValue
147     IC_RetainBlock,         ///< objc_retainBlock
148     IC_Release,             ///< objc_release
149     IC_Autorelease,         ///< objc_autorelease
150     IC_AutoreleaseRV,       ///< objc_autoreleaseReturnValue
151     IC_AutoreleasepoolPush, ///< objc_autoreleasePoolPush
152     IC_AutoreleasepoolPop,  ///< objc_autoreleasePoolPop
153     IC_NoopCast,            ///< objc_retainedObject, etc.
154     IC_FusedRetainAutorelease, ///< objc_retainAutorelease
155     IC_FusedRetainAutoreleaseRV, ///< objc_retainAutoreleaseReturnValue
156     IC_LoadWeakRetained,    ///< objc_loadWeakRetained (primitive)
157     IC_StoreWeak,           ///< objc_storeWeak (primitive)
158     IC_InitWeak,            ///< objc_initWeak (derived)
159     IC_LoadWeak,            ///< objc_loadWeak (derived)
160     IC_MoveWeak,            ///< objc_moveWeak (derived)
161     IC_CopyWeak,            ///< objc_copyWeak (derived)
162     IC_DestroyWeak,         ///< objc_destroyWeak (derived)
163     IC_CallOrUser,          ///< could call objc_release and/or "use" pointers
164     IC_Call,                ///< could call objc_release
165     IC_User,                ///< could "use" a pointer
166     IC_None                 ///< anything else
167   };
168 }
169 
170 /// IsPotentialUse - Test whether the given value is possible a
171 /// reference-counted pointer.
IsPotentialUse(const Value * Op)172 static bool IsPotentialUse(const Value *Op) {
173   // Pointers to static or stack storage are not reference-counted pointers.
174   if (isa<Constant>(Op) || isa<AllocaInst>(Op))
175     return false;
176   // Special arguments are not reference-counted.
177   if (const Argument *Arg = dyn_cast<Argument>(Op))
178     if (Arg->hasByValAttr() ||
179         Arg->hasNestAttr() ||
180         Arg->hasStructRetAttr())
181       return false;
182   // Only consider values with pointer types, and not function pointers.
183   PointerType *Ty = dyn_cast<PointerType>(Op->getType());
184   if (!Ty || isa<FunctionType>(Ty->getElementType()))
185     return false;
186   // Conservatively assume anything else is a potential use.
187   return true;
188 }
189 
190 /// GetCallSiteClass - Helper for GetInstructionClass. Determines what kind
191 /// of construct CS is.
GetCallSiteClass(ImmutableCallSite CS)192 static InstructionClass GetCallSiteClass(ImmutableCallSite CS) {
193   for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
194        I != E; ++I)
195     if (IsPotentialUse(*I))
196       return CS.onlyReadsMemory() ? IC_User : IC_CallOrUser;
197 
198   return CS.onlyReadsMemory() ? IC_None : IC_Call;
199 }
200 
201 /// GetFunctionClass - Determine if F is one of the special known Functions.
202 /// If it isn't, return IC_CallOrUser.
GetFunctionClass(const Function * F)203 static InstructionClass GetFunctionClass(const Function *F) {
204   Function::const_arg_iterator AI = F->arg_begin(), AE = F->arg_end();
205 
206   // No arguments.
207   if (AI == AE)
208     return StringSwitch<InstructionClass>(F->getName())
209       .Case("objc_autoreleasePoolPush",  IC_AutoreleasepoolPush)
210       .Default(IC_CallOrUser);
211 
212   // One argument.
213   const Argument *A0 = AI++;
214   if (AI == AE)
215     // Argument is a pointer.
216     if (PointerType *PTy = dyn_cast<PointerType>(A0->getType())) {
217       Type *ETy = PTy->getElementType();
218       // Argument is i8*.
219       if (ETy->isIntegerTy(8))
220         return StringSwitch<InstructionClass>(F->getName())
221           .Case("objc_retain",                IC_Retain)
222           .Case("objc_retainAutoreleasedReturnValue", IC_RetainRV)
223           .Case("objc_retainBlock",           IC_RetainBlock)
224           .Case("objc_release",               IC_Release)
225           .Case("objc_autorelease",           IC_Autorelease)
226           .Case("objc_autoreleaseReturnValue", IC_AutoreleaseRV)
227           .Case("objc_autoreleasePoolPop",    IC_AutoreleasepoolPop)
228           .Case("objc_retainedObject",        IC_NoopCast)
229           .Case("objc_unretainedObject",      IC_NoopCast)
230           .Case("objc_unretainedPointer",     IC_NoopCast)
231           .Case("objc_retain_autorelease",    IC_FusedRetainAutorelease)
232           .Case("objc_retainAutorelease",     IC_FusedRetainAutorelease)
233           .Case("objc_retainAutoreleaseReturnValue",IC_FusedRetainAutoreleaseRV)
234           .Default(IC_CallOrUser);
235 
236       // Argument is i8**
237       if (PointerType *Pte = dyn_cast<PointerType>(ETy))
238         if (Pte->getElementType()->isIntegerTy(8))
239           return StringSwitch<InstructionClass>(F->getName())
240             .Case("objc_loadWeakRetained",      IC_LoadWeakRetained)
241             .Case("objc_loadWeak",              IC_LoadWeak)
242             .Case("objc_destroyWeak",           IC_DestroyWeak)
243             .Default(IC_CallOrUser);
244     }
245 
246   // Two arguments, first is i8**.
247   const Argument *A1 = AI++;
248   if (AI == AE)
249     if (PointerType *PTy = dyn_cast<PointerType>(A0->getType()))
250       if (PointerType *Pte = dyn_cast<PointerType>(PTy->getElementType()))
251         if (Pte->getElementType()->isIntegerTy(8))
252           if (PointerType *PTy1 = dyn_cast<PointerType>(A1->getType())) {
253             Type *ETy1 = PTy1->getElementType();
254             // Second argument is i8*
255             if (ETy1->isIntegerTy(8))
256               return StringSwitch<InstructionClass>(F->getName())
257                      .Case("objc_storeWeak",             IC_StoreWeak)
258                      .Case("objc_initWeak",              IC_InitWeak)
259                      .Default(IC_CallOrUser);
260             // Second argument is i8**.
261             if (PointerType *Pte1 = dyn_cast<PointerType>(ETy1))
262               if (Pte1->getElementType()->isIntegerTy(8))
263                 return StringSwitch<InstructionClass>(F->getName())
264                        .Case("objc_moveWeak",              IC_MoveWeak)
265                        .Case("objc_copyWeak",              IC_CopyWeak)
266                        .Default(IC_CallOrUser);
267           }
268 
269   // Anything else.
270   return IC_CallOrUser;
271 }
272 
273 /// GetInstructionClass - Determine what kind of construct V is.
GetInstructionClass(const Value * V)274 static InstructionClass GetInstructionClass(const Value *V) {
275   if (const Instruction *I = dyn_cast<Instruction>(V)) {
276     // Any instruction other than bitcast and gep with a pointer operand have a
277     // use of an objc pointer. Bitcasts, GEPs, Selects, PHIs transfer a pointer
278     // to a subsequent use, rather than using it themselves, in this sense.
279     // As a short cut, several other opcodes are known to have no pointer
280     // operands of interest. And ret is never followed by a release, so it's
281     // not interesting to examine.
282     switch (I->getOpcode()) {
283     case Instruction::Call: {
284       const CallInst *CI = cast<CallInst>(I);
285       // Check for calls to special functions.
286       if (const Function *F = CI->getCalledFunction()) {
287         InstructionClass Class = GetFunctionClass(F);
288         if (Class != IC_CallOrUser)
289           return Class;
290 
291         // None of the intrinsic functions do objc_release. For intrinsics, the
292         // only question is whether or not they may be users.
293         switch (F->getIntrinsicID()) {
294         case 0: break;
295         case Intrinsic::bswap: case Intrinsic::ctpop:
296         case Intrinsic::ctlz: case Intrinsic::cttz:
297         case Intrinsic::returnaddress: case Intrinsic::frameaddress:
298         case Intrinsic::stacksave: case Intrinsic::stackrestore:
299         case Intrinsic::vastart: case Intrinsic::vacopy: case Intrinsic::vaend:
300         // Don't let dbg info affect our results.
301         case Intrinsic::dbg_declare: case Intrinsic::dbg_value:
302           // Short cut: Some intrinsics obviously don't use ObjC pointers.
303           return IC_None;
304         default:
305           for (Function::const_arg_iterator AI = F->arg_begin(),
306                AE = F->arg_end(); AI != AE; ++AI)
307             if (IsPotentialUse(AI))
308               return IC_User;
309           return IC_None;
310         }
311       }
312       return GetCallSiteClass(CI);
313     }
314     case Instruction::Invoke:
315       return GetCallSiteClass(cast<InvokeInst>(I));
316     case Instruction::BitCast:
317     case Instruction::GetElementPtr:
318     case Instruction::Select: case Instruction::PHI:
319     case Instruction::Ret: case Instruction::Br:
320     case Instruction::Switch: case Instruction::IndirectBr:
321     case Instruction::Alloca: case Instruction::VAArg:
322     case Instruction::Add: case Instruction::FAdd:
323     case Instruction::Sub: case Instruction::FSub:
324     case Instruction::Mul: case Instruction::FMul:
325     case Instruction::SDiv: case Instruction::UDiv: case Instruction::FDiv:
326     case Instruction::SRem: case Instruction::URem: case Instruction::FRem:
327     case Instruction::Shl: case Instruction::LShr: case Instruction::AShr:
328     case Instruction::And: case Instruction::Or: case Instruction::Xor:
329     case Instruction::SExt: case Instruction::ZExt: case Instruction::Trunc:
330     case Instruction::IntToPtr: case Instruction::FCmp:
331     case Instruction::FPTrunc: case Instruction::FPExt:
332     case Instruction::FPToUI: case Instruction::FPToSI:
333     case Instruction::UIToFP: case Instruction::SIToFP:
334     case Instruction::InsertElement: case Instruction::ExtractElement:
335     case Instruction::ShuffleVector:
336     case Instruction::ExtractValue:
337       break;
338     case Instruction::ICmp:
339       // Comparing a pointer with null, or any other constant, isn't an
340       // interesting use, because we don't care what the pointer points to, or
341       // about the values of any other dynamic reference-counted pointers.
342       if (IsPotentialUse(I->getOperand(1)))
343         return IC_User;
344       break;
345     default:
346       // For anything else, check all the operands.
347       // Note that this includes both operands of a Store: while the first
348       // operand isn't actually being dereferenced, it is being stored to
349       // memory where we can no longer track who might read it and dereference
350       // it, so we have to consider it potentially used.
351       for (User::const_op_iterator OI = I->op_begin(), OE = I->op_end();
352            OI != OE; ++OI)
353         if (IsPotentialUse(*OI))
354           return IC_User;
355     }
356   }
357 
358   // Otherwise, it's totally inert for ARC purposes.
359   return IC_None;
360 }
361 
362 /// GetBasicInstructionClass - Determine what kind of construct V is. This is
363 /// similar to GetInstructionClass except that it only detects objc runtine
364 /// calls. This allows it to be faster.
GetBasicInstructionClass(const Value * V)365 static InstructionClass GetBasicInstructionClass(const Value *V) {
366   if (const CallInst *CI = dyn_cast<CallInst>(V)) {
367     if (const Function *F = CI->getCalledFunction())
368       return GetFunctionClass(F);
369     // Otherwise, be conservative.
370     return IC_CallOrUser;
371   }
372 
373   // Otherwise, be conservative.
374   return IC_User;
375 }
376 
377 /// IsRetain - Test if the the given class is objc_retain or
378 /// equivalent.
IsRetain(InstructionClass Class)379 static bool IsRetain(InstructionClass Class) {
380   return Class == IC_Retain ||
381          Class == IC_RetainRV;
382 }
383 
384 /// IsAutorelease - Test if the the given class is objc_autorelease or
385 /// equivalent.
IsAutorelease(InstructionClass Class)386 static bool IsAutorelease(InstructionClass Class) {
387   return Class == IC_Autorelease ||
388          Class == IC_AutoreleaseRV;
389 }
390 
391 /// IsForwarding - Test if the given class represents instructions which return
392 /// their argument verbatim.
IsForwarding(InstructionClass Class)393 static bool IsForwarding(InstructionClass Class) {
394   // objc_retainBlock technically doesn't always return its argument
395   // verbatim, but it doesn't matter for our purposes here.
396   return Class == IC_Retain ||
397          Class == IC_RetainRV ||
398          Class == IC_Autorelease ||
399          Class == IC_AutoreleaseRV ||
400          Class == IC_RetainBlock ||
401          Class == IC_NoopCast;
402 }
403 
404 /// IsNoopOnNull - Test if the given class represents instructions which do
405 /// nothing if passed a null pointer.
IsNoopOnNull(InstructionClass Class)406 static bool IsNoopOnNull(InstructionClass Class) {
407   return Class == IC_Retain ||
408          Class == IC_RetainRV ||
409          Class == IC_Release ||
410          Class == IC_Autorelease ||
411          Class == IC_AutoreleaseRV ||
412          Class == IC_RetainBlock;
413 }
414 
415 /// IsAlwaysTail - Test if the given class represents instructions which are
416 /// always safe to mark with the "tail" keyword.
IsAlwaysTail(InstructionClass Class)417 static bool IsAlwaysTail(InstructionClass Class) {
418   // IC_RetainBlock may be given a stack argument.
419   return Class == IC_Retain ||
420          Class == IC_RetainRV ||
421          Class == IC_Autorelease ||
422          Class == IC_AutoreleaseRV;
423 }
424 
425 /// IsNoThrow - Test if the given class represents instructions which are always
426 /// safe to mark with the nounwind attribute..
IsNoThrow(InstructionClass Class)427 static bool IsNoThrow(InstructionClass Class) {
428   // objc_retainBlock is not nounwind because it calls user copy constructors
429   // which could theoretically throw.
430   return Class == IC_Retain ||
431          Class == IC_RetainRV ||
432          Class == IC_Release ||
433          Class == IC_Autorelease ||
434          Class == IC_AutoreleaseRV ||
435          Class == IC_AutoreleasepoolPush ||
436          Class == IC_AutoreleasepoolPop;
437 }
438 
439 /// EraseInstruction - Erase the given instruction. ObjC calls return their
440 /// argument verbatim, so if it's such a call and the return value has users,
441 /// replace them with the argument value.
EraseInstruction(Instruction * CI)442 static void EraseInstruction(Instruction *CI) {
443   Value *OldArg = cast<CallInst>(CI)->getArgOperand(0);
444 
445   bool Unused = CI->use_empty();
446 
447   if (!Unused) {
448     // Replace the return value with the argument.
449     assert(IsForwarding(GetBasicInstructionClass(CI)) &&
450            "Can't delete non-forwarding instruction with users!");
451     CI->replaceAllUsesWith(OldArg);
452   }
453 
454   CI->eraseFromParent();
455 
456   if (Unused)
457     RecursivelyDeleteTriviallyDeadInstructions(OldArg);
458 }
459 
460 /// GetUnderlyingObjCPtr - This is a wrapper around getUnderlyingObject which
461 /// also knows how to look through objc_retain and objc_autorelease calls, which
462 /// we know to return their argument verbatim.
GetUnderlyingObjCPtr(const Value * V)463 static const Value *GetUnderlyingObjCPtr(const Value *V) {
464   for (;;) {
465     V = GetUnderlyingObject(V);
466     if (!IsForwarding(GetBasicInstructionClass(V)))
467       break;
468     V = cast<CallInst>(V)->getArgOperand(0);
469   }
470 
471   return V;
472 }
473 
474 /// StripPointerCastsAndObjCCalls - This is a wrapper around
475 /// Value::stripPointerCasts which also knows how to look through objc_retain
476 /// and objc_autorelease calls, which we know to return their argument verbatim.
StripPointerCastsAndObjCCalls(const Value * V)477 static const Value *StripPointerCastsAndObjCCalls(const Value *V) {
478   for (;;) {
479     V = V->stripPointerCasts();
480     if (!IsForwarding(GetBasicInstructionClass(V)))
481       break;
482     V = cast<CallInst>(V)->getArgOperand(0);
483   }
484   return V;
485 }
486 
487 /// StripPointerCastsAndObjCCalls - This is a wrapper around
488 /// Value::stripPointerCasts which also knows how to look through objc_retain
489 /// and objc_autorelease calls, which we know to return their argument verbatim.
StripPointerCastsAndObjCCalls(Value * V)490 static Value *StripPointerCastsAndObjCCalls(Value *V) {
491   for (;;) {
492     V = V->stripPointerCasts();
493     if (!IsForwarding(GetBasicInstructionClass(V)))
494       break;
495     V = cast<CallInst>(V)->getArgOperand(0);
496   }
497   return V;
498 }
499 
500 /// GetObjCArg - Assuming the given instruction is one of the special calls such
501 /// as objc_retain or objc_release, return the argument value, stripped of no-op
502 /// casts and forwarding calls.
GetObjCArg(Value * Inst)503 static Value *GetObjCArg(Value *Inst) {
504   return StripPointerCastsAndObjCCalls(cast<CallInst>(Inst)->getArgOperand(0));
505 }
506 
507 /// IsObjCIdentifiedObject - This is similar to AliasAnalysis'
508 /// isObjCIdentifiedObject, except that it uses special knowledge of
509 /// ObjC conventions...
IsObjCIdentifiedObject(const Value * V)510 static bool IsObjCIdentifiedObject(const Value *V) {
511   // Assume that call results and arguments have their own "provenance".
512   // Constants (including GlobalVariables) and Allocas are never
513   // reference-counted.
514   if (isa<CallInst>(V) || isa<InvokeInst>(V) ||
515       isa<Argument>(V) || isa<Constant>(V) ||
516       isa<AllocaInst>(V))
517     return true;
518 
519   if (const LoadInst *LI = dyn_cast<LoadInst>(V)) {
520     const Value *Pointer =
521       StripPointerCastsAndObjCCalls(LI->getPointerOperand());
522     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Pointer)) {
523       // A constant pointer can't be pointing to an object on the heap. It may
524       // be reference-counted, but it won't be deleted.
525       if (GV->isConstant())
526         return true;
527       StringRef Name = GV->getName();
528       // These special variables are known to hold values which are not
529       // reference-counted pointers.
530       if (Name.startswith("\01L_OBJC_SELECTOR_REFERENCES_") ||
531           Name.startswith("\01L_OBJC_CLASSLIST_REFERENCES_") ||
532           Name.startswith("\01L_OBJC_CLASSLIST_SUP_REFS_$_") ||
533           Name.startswith("\01L_OBJC_METH_VAR_NAME_") ||
534           Name.startswith("\01l_objc_msgSend_fixup_"))
535         return true;
536     }
537   }
538 
539   return false;
540 }
541 
542 /// FindSingleUseIdentifiedObject - This is similar to
543 /// StripPointerCastsAndObjCCalls but it stops as soon as it finds a value
544 /// with multiple uses.
FindSingleUseIdentifiedObject(const Value * Arg)545 static const Value *FindSingleUseIdentifiedObject(const Value *Arg) {
546   if (Arg->hasOneUse()) {
547     if (const BitCastInst *BC = dyn_cast<BitCastInst>(Arg))
548       return FindSingleUseIdentifiedObject(BC->getOperand(0));
549     if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Arg))
550       if (GEP->hasAllZeroIndices())
551         return FindSingleUseIdentifiedObject(GEP->getPointerOperand());
552     if (IsForwarding(GetBasicInstructionClass(Arg)))
553       return FindSingleUseIdentifiedObject(
554                cast<CallInst>(Arg)->getArgOperand(0));
555     if (!IsObjCIdentifiedObject(Arg))
556       return 0;
557     return Arg;
558   }
559 
560   // If we found an identifiable object but it has multiple uses, but they
561   // are trivial uses, we can still consider this to be a single-use
562   // value.
563   if (IsObjCIdentifiedObject(Arg)) {
564     for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
565          UI != UE; ++UI) {
566       const User *U = *UI;
567       if (!U->use_empty() || StripPointerCastsAndObjCCalls(U) != Arg)
568          return 0;
569     }
570 
571     return Arg;
572   }
573 
574   return 0;
575 }
576 
577 /// ModuleHasARC - Test if the given module looks interesting to run ARC
578 /// optimization on.
ModuleHasARC(const Module & M)579 static bool ModuleHasARC(const Module &M) {
580   return
581     M.getNamedValue("objc_retain") ||
582     M.getNamedValue("objc_release") ||
583     M.getNamedValue("objc_autorelease") ||
584     M.getNamedValue("objc_retainAutoreleasedReturnValue") ||
585     M.getNamedValue("objc_retainBlock") ||
586     M.getNamedValue("objc_autoreleaseReturnValue") ||
587     M.getNamedValue("objc_autoreleasePoolPush") ||
588     M.getNamedValue("objc_loadWeakRetained") ||
589     M.getNamedValue("objc_loadWeak") ||
590     M.getNamedValue("objc_destroyWeak") ||
591     M.getNamedValue("objc_storeWeak") ||
592     M.getNamedValue("objc_initWeak") ||
593     M.getNamedValue("objc_moveWeak") ||
594     M.getNamedValue("objc_copyWeak") ||
595     M.getNamedValue("objc_retainedObject") ||
596     M.getNamedValue("objc_unretainedObject") ||
597     M.getNamedValue("objc_unretainedPointer");
598 }
599 
600 //===----------------------------------------------------------------------===//
601 // ARC AliasAnalysis.
602 //===----------------------------------------------------------------------===//
603 
604 #include "llvm/Pass.h"
605 #include "llvm/Analysis/AliasAnalysis.h"
606 #include "llvm/Analysis/Passes.h"
607 
608 namespace {
609   /// ObjCARCAliasAnalysis - This is a simple alias analysis
610   /// implementation that uses knowledge of ARC constructs to answer queries.
611   ///
612   /// TODO: This class could be generalized to know about other ObjC-specific
613   /// tricks. Such as knowing that ivars in the non-fragile ABI are non-aliasing
614   /// even though their offsets are dynamic.
615   class ObjCARCAliasAnalysis : public ImmutablePass,
616                                public AliasAnalysis {
617   public:
618     static char ID; // Class identification, replacement for typeinfo
ObjCARCAliasAnalysis()619     ObjCARCAliasAnalysis() : ImmutablePass(ID) {
620       initializeObjCARCAliasAnalysisPass(*PassRegistry::getPassRegistry());
621     }
622 
623   private:
initializePass()624     virtual void initializePass() {
625       InitializeAliasAnalysis(this);
626     }
627 
628     /// getAdjustedAnalysisPointer - This method is used when a pass implements
629     /// an analysis interface through multiple inheritance.  If needed, it
630     /// should override this to adjust the this pointer as needed for the
631     /// specified pass info.
getAdjustedAnalysisPointer(const void * PI)632     virtual void *getAdjustedAnalysisPointer(const void *PI) {
633       if (PI == &AliasAnalysis::ID)
634         return (AliasAnalysis*)this;
635       return this;
636     }
637 
638     virtual void getAnalysisUsage(AnalysisUsage &AU) const;
639     virtual AliasResult alias(const Location &LocA, const Location &LocB);
640     virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
641     virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
642     virtual ModRefBehavior getModRefBehavior(const Function *F);
643     virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
644                                        const Location &Loc);
645     virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
646                                        ImmutableCallSite CS2);
647   };
648 }  // End of anonymous namespace
649 
650 // Register this pass...
651 char ObjCARCAliasAnalysis::ID = 0;
652 INITIALIZE_AG_PASS(ObjCARCAliasAnalysis, AliasAnalysis, "objc-arc-aa",
653                    "ObjC-ARC-Based Alias Analysis", false, true, false)
654 
createObjCARCAliasAnalysisPass()655 ImmutablePass *llvm::createObjCARCAliasAnalysisPass() {
656   return new ObjCARCAliasAnalysis();
657 }
658 
659 void
getAnalysisUsage(AnalysisUsage & AU) const660 ObjCARCAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
661   AU.setPreservesAll();
662   AliasAnalysis::getAnalysisUsage(AU);
663 }
664 
665 AliasAnalysis::AliasResult
alias(const Location & LocA,const Location & LocB)666 ObjCARCAliasAnalysis::alias(const Location &LocA, const Location &LocB) {
667   if (!EnableARCOpts)
668     return AliasAnalysis::alias(LocA, LocB);
669 
670   // First, strip off no-ops, including ObjC-specific no-ops, and try making a
671   // precise alias query.
672   const Value *SA = StripPointerCastsAndObjCCalls(LocA.Ptr);
673   const Value *SB = StripPointerCastsAndObjCCalls(LocB.Ptr);
674   AliasResult Result =
675     AliasAnalysis::alias(Location(SA, LocA.Size, LocA.TBAATag),
676                          Location(SB, LocB.Size, LocB.TBAATag));
677   if (Result != MayAlias)
678     return Result;
679 
680   // If that failed, climb to the underlying object, including climbing through
681   // ObjC-specific no-ops, and try making an imprecise alias query.
682   const Value *UA = GetUnderlyingObjCPtr(SA);
683   const Value *UB = GetUnderlyingObjCPtr(SB);
684   if (UA != SA || UB != SB) {
685     Result = AliasAnalysis::alias(Location(UA), Location(UB));
686     // We can't use MustAlias or PartialAlias results here because
687     // GetUnderlyingObjCPtr may return an offsetted pointer value.
688     if (Result == NoAlias)
689       return NoAlias;
690   }
691 
692   // If that failed, fail. We don't need to chain here, since that's covered
693   // by the earlier precise query.
694   return MayAlias;
695 }
696 
697 bool
pointsToConstantMemory(const Location & Loc,bool OrLocal)698 ObjCARCAliasAnalysis::pointsToConstantMemory(const Location &Loc,
699                                              bool OrLocal) {
700   if (!EnableARCOpts)
701     return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
702 
703   // First, strip off no-ops, including ObjC-specific no-ops, and try making
704   // a precise alias query.
705   const Value *S = StripPointerCastsAndObjCCalls(Loc.Ptr);
706   if (AliasAnalysis::pointsToConstantMemory(Location(S, Loc.Size, Loc.TBAATag),
707                                             OrLocal))
708     return true;
709 
710   // If that failed, climb to the underlying object, including climbing through
711   // ObjC-specific no-ops, and try making an imprecise alias query.
712   const Value *U = GetUnderlyingObjCPtr(S);
713   if (U != S)
714     return AliasAnalysis::pointsToConstantMemory(Location(U), OrLocal);
715 
716   // If that failed, fail. We don't need to chain here, since that's covered
717   // by the earlier precise query.
718   return false;
719 }
720 
721 AliasAnalysis::ModRefBehavior
getModRefBehavior(ImmutableCallSite CS)722 ObjCARCAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
723   // We have nothing to do. Just chain to the next AliasAnalysis.
724   return AliasAnalysis::getModRefBehavior(CS);
725 }
726 
727 AliasAnalysis::ModRefBehavior
getModRefBehavior(const Function * F)728 ObjCARCAliasAnalysis::getModRefBehavior(const Function *F) {
729   if (!EnableARCOpts)
730     return AliasAnalysis::getModRefBehavior(F);
731 
732   switch (GetFunctionClass(F)) {
733   case IC_NoopCast:
734     return DoesNotAccessMemory;
735   default:
736     break;
737   }
738 
739   return AliasAnalysis::getModRefBehavior(F);
740 }
741 
742 AliasAnalysis::ModRefResult
getModRefInfo(ImmutableCallSite CS,const Location & Loc)743 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS, const Location &Loc) {
744   if (!EnableARCOpts)
745     return AliasAnalysis::getModRefInfo(CS, Loc);
746 
747   switch (GetBasicInstructionClass(CS.getInstruction())) {
748   case IC_Retain:
749   case IC_RetainRV:
750   case IC_Autorelease:
751   case IC_AutoreleaseRV:
752   case IC_NoopCast:
753   case IC_AutoreleasepoolPush:
754   case IC_FusedRetainAutorelease:
755   case IC_FusedRetainAutoreleaseRV:
756     // These functions don't access any memory visible to the compiler.
757     // Note that this doesn't include objc_retainBlock, becuase it updates
758     // pointers when it copies block data.
759     return NoModRef;
760   default:
761     break;
762   }
763 
764   return AliasAnalysis::getModRefInfo(CS, Loc);
765 }
766 
767 AliasAnalysis::ModRefResult
getModRefInfo(ImmutableCallSite CS1,ImmutableCallSite CS2)768 ObjCARCAliasAnalysis::getModRefInfo(ImmutableCallSite CS1,
769                                     ImmutableCallSite CS2) {
770   // TODO: Theoretically we could check for dependencies between objc_* calls
771   // and OnlyAccessesArgumentPointees calls or other well-behaved calls.
772   return AliasAnalysis::getModRefInfo(CS1, CS2);
773 }
774 
775 //===----------------------------------------------------------------------===//
776 // ARC expansion.
777 //===----------------------------------------------------------------------===//
778 
779 #include "llvm/Support/InstIterator.h"
780 #include "llvm/Transforms/Scalar.h"
781 
782 namespace {
783   /// ObjCARCExpand - Early ARC transformations.
784   class ObjCARCExpand : public FunctionPass {
785     virtual void getAnalysisUsage(AnalysisUsage &AU) const;
786     virtual bool doInitialization(Module &M);
787     virtual bool runOnFunction(Function &F);
788 
789     /// Run - A flag indicating whether this optimization pass should run.
790     bool Run;
791 
792   public:
793     static char ID;
ObjCARCExpand()794     ObjCARCExpand() : FunctionPass(ID) {
795       initializeObjCARCExpandPass(*PassRegistry::getPassRegistry());
796     }
797   };
798 }
799 
800 char ObjCARCExpand::ID = 0;
801 INITIALIZE_PASS(ObjCARCExpand,
802                 "objc-arc-expand", "ObjC ARC expansion", false, false)
803 
createObjCARCExpandPass()804 Pass *llvm::createObjCARCExpandPass() {
805   return new ObjCARCExpand();
806 }
807 
getAnalysisUsage(AnalysisUsage & AU) const808 void ObjCARCExpand::getAnalysisUsage(AnalysisUsage &AU) const {
809   AU.setPreservesCFG();
810 }
811 
doInitialization(Module & M)812 bool ObjCARCExpand::doInitialization(Module &M) {
813   Run = ModuleHasARC(M);
814   return false;
815 }
816 
runOnFunction(Function & F)817 bool ObjCARCExpand::runOnFunction(Function &F) {
818   if (!EnableARCOpts)
819     return false;
820 
821   // If nothing in the Module uses ARC, don't do anything.
822   if (!Run)
823     return false;
824 
825   bool Changed = false;
826 
827   for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I) {
828     Instruction *Inst = &*I;
829 
830     switch (GetBasicInstructionClass(Inst)) {
831     case IC_Retain:
832     case IC_RetainRV:
833     case IC_Autorelease:
834     case IC_AutoreleaseRV:
835     case IC_FusedRetainAutorelease:
836     case IC_FusedRetainAutoreleaseRV:
837       // These calls return their argument verbatim, as a low-level
838       // optimization. However, this makes high-level optimizations
839       // harder. Undo any uses of this optimization that the front-end
840       // emitted here. We'll redo them in a later pass.
841       Changed = true;
842       Inst->replaceAllUsesWith(cast<CallInst>(Inst)->getArgOperand(0));
843       break;
844     default:
845       break;
846     }
847   }
848 
849   return Changed;
850 }
851 
852 //===----------------------------------------------------------------------===//
853 // ARC optimization.
854 //===----------------------------------------------------------------------===//
855 
856 // TODO: On code like this:
857 //
858 // objc_retain(%x)
859 // stuff_that_cannot_release()
860 // objc_autorelease(%x)
861 // stuff_that_cannot_release()
862 // objc_retain(%x)
863 // stuff_that_cannot_release()
864 // objc_autorelease(%x)
865 //
866 // The second retain and autorelease can be deleted.
867 
868 // TODO: It should be possible to delete
869 // objc_autoreleasePoolPush and objc_autoreleasePoolPop
870 // pairs if nothing is actually autoreleased between them. Also, autorelease
871 // calls followed by objc_autoreleasePoolPop calls (perhaps in ObjC++ code
872 // after inlining) can be turned into plain release calls.
873 
874 // TODO: Critical-edge splitting. If the optimial insertion point is
875 // a critical edge, the current algorithm has to fail, because it doesn't
876 // know how to split edges. It should be possible to make the optimizer
877 // think in terms of edges, rather than blocks, and then split critical
878 // edges on demand.
879 
880 // TODO: OptimizeSequences could generalized to be Interprocedural.
881 
882 // TODO: Recognize that a bunch of other objc runtime calls have
883 // non-escaping arguments and non-releasing arguments, and may be
884 // non-autoreleasing.
885 
886 // TODO: Sink autorelease calls as far as possible. Unfortunately we
887 // usually can't sink them past other calls, which would be the main
888 // case where it would be useful.
889 
890 // TODO: The pointer returned from objc_loadWeakRetained is retained.
891 
892 // TODO: Delete release+retain pairs (rare).
893 
894 #include "llvm/GlobalAlias.h"
895 #include "llvm/Constants.h"
896 #include "llvm/LLVMContext.h"
897 #include "llvm/Support/ErrorHandling.h"
898 #include "llvm/Support/CFG.h"
899 #include "llvm/ADT/PostOrderIterator.h"
900 #include "llvm/ADT/Statistic.h"
901 
902 STATISTIC(NumNoops,       "Number of no-op objc calls eliminated");
903 STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated");
904 STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases");
905 STATISTIC(NumRets,        "Number of return value forwarding "
906                           "retain+autoreleaes eliminated");
907 STATISTIC(NumRRs,         "Number of retain+release paths eliminated");
908 STATISTIC(NumPeeps,       "Number of calls peephole-optimized");
909 
910 namespace {
911   /// ProvenanceAnalysis - This is similar to BasicAliasAnalysis, and it
912   /// uses many of the same techniques, except it uses special ObjC-specific
913   /// reasoning about pointer relationships.
914   class ProvenanceAnalysis {
915     AliasAnalysis *AA;
916 
917     typedef std::pair<const Value *, const Value *> ValuePairTy;
918     typedef DenseMap<ValuePairTy, bool> CachedResultsTy;
919     CachedResultsTy CachedResults;
920 
921     bool relatedCheck(const Value *A, const Value *B);
922     bool relatedSelect(const SelectInst *A, const Value *B);
923     bool relatedPHI(const PHINode *A, const Value *B);
924 
925     // Do not implement.
926     void operator=(const ProvenanceAnalysis &);
927     ProvenanceAnalysis(const ProvenanceAnalysis &);
928 
929   public:
ProvenanceAnalysis()930     ProvenanceAnalysis() {}
931 
setAA(AliasAnalysis * aa)932     void setAA(AliasAnalysis *aa) { AA = aa; }
933 
getAA() const934     AliasAnalysis *getAA() const { return AA; }
935 
936     bool related(const Value *A, const Value *B);
937 
clear()938     void clear() {
939       CachedResults.clear();
940     }
941   };
942 }
943 
relatedSelect(const SelectInst * A,const Value * B)944 bool ProvenanceAnalysis::relatedSelect(const SelectInst *A, const Value *B) {
945   // If the values are Selects with the same condition, we can do a more precise
946   // check: just check for relations between the values on corresponding arms.
947   if (const SelectInst *SB = dyn_cast<SelectInst>(B))
948     if (A->getCondition() == SB->getCondition()) {
949       if (related(A->getTrueValue(), SB->getTrueValue()))
950         return true;
951       if (related(A->getFalseValue(), SB->getFalseValue()))
952         return true;
953       return false;
954     }
955 
956   // Check both arms of the Select node individually.
957   if (related(A->getTrueValue(), B))
958     return true;
959   if (related(A->getFalseValue(), B))
960     return true;
961 
962   // The arms both checked out.
963   return false;
964 }
965 
relatedPHI(const PHINode * A,const Value * B)966 bool ProvenanceAnalysis::relatedPHI(const PHINode *A, const Value *B) {
967   // If the values are PHIs in the same block, we can do a more precise as well
968   // as efficient check: just check for relations between the values on
969   // corresponding edges.
970   if (const PHINode *PNB = dyn_cast<PHINode>(B))
971     if (PNB->getParent() == A->getParent()) {
972       for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i)
973         if (related(A->getIncomingValue(i),
974                     PNB->getIncomingValueForBlock(A->getIncomingBlock(i))))
975           return true;
976       return false;
977     }
978 
979   // Check each unique source of the PHI node against B.
980   SmallPtrSet<const Value *, 4> UniqueSrc;
981   for (unsigned i = 0, e = A->getNumIncomingValues(); i != e; ++i) {
982     const Value *PV1 = A->getIncomingValue(i);
983     if (UniqueSrc.insert(PV1) && related(PV1, B))
984       return true;
985   }
986 
987   // All of the arms checked out.
988   return false;
989 }
990 
991 /// isStoredObjCPointer - Test if the value of P, or any value covered by its
992 /// provenance, is ever stored within the function (not counting callees).
isStoredObjCPointer(const Value * P)993 static bool isStoredObjCPointer(const Value *P) {
994   SmallPtrSet<const Value *, 8> Visited;
995   SmallVector<const Value *, 8> Worklist;
996   Worklist.push_back(P);
997   Visited.insert(P);
998   do {
999     P = Worklist.pop_back_val();
1000     for (Value::const_use_iterator UI = P->use_begin(), UE = P->use_end();
1001          UI != UE; ++UI) {
1002       const User *Ur = *UI;
1003       if (isa<StoreInst>(Ur)) {
1004         if (UI.getOperandNo() == 0)
1005           // The pointer is stored.
1006           return true;
1007         // The pointed is stored through.
1008         continue;
1009       }
1010       if (isa<CallInst>(Ur))
1011         // The pointer is passed as an argument, ignore this.
1012         continue;
1013       if (isa<PtrToIntInst>(P))
1014         // Assume the worst.
1015         return true;
1016       if (Visited.insert(Ur))
1017         Worklist.push_back(Ur);
1018     }
1019   } while (!Worklist.empty());
1020 
1021   // Everything checked out.
1022   return false;
1023 }
1024 
relatedCheck(const Value * A,const Value * B)1025 bool ProvenanceAnalysis::relatedCheck(const Value *A, const Value *B) {
1026   // Skip past provenance pass-throughs.
1027   A = GetUnderlyingObjCPtr(A);
1028   B = GetUnderlyingObjCPtr(B);
1029 
1030   // Quick check.
1031   if (A == B)
1032     return true;
1033 
1034   // Ask regular AliasAnalysis, for a first approximation.
1035   switch (AA->alias(A, B)) {
1036   case AliasAnalysis::NoAlias:
1037     return false;
1038   case AliasAnalysis::MustAlias:
1039   case AliasAnalysis::PartialAlias:
1040     return true;
1041   case AliasAnalysis::MayAlias:
1042     break;
1043   }
1044 
1045   bool AIsIdentified = IsObjCIdentifiedObject(A);
1046   bool BIsIdentified = IsObjCIdentifiedObject(B);
1047 
1048   // An ObjC-Identified object can't alias a load if it is never locally stored.
1049   if (AIsIdentified) {
1050     if (BIsIdentified) {
1051       // If both pointers have provenance, they can be directly compared.
1052       if (A != B)
1053         return false;
1054     } else {
1055       if (isa<LoadInst>(B))
1056         return isStoredObjCPointer(A);
1057     }
1058   } else {
1059     if (BIsIdentified && isa<LoadInst>(A))
1060       return isStoredObjCPointer(B);
1061   }
1062 
1063    // Special handling for PHI and Select.
1064   if (const PHINode *PN = dyn_cast<PHINode>(A))
1065     return relatedPHI(PN, B);
1066   if (const PHINode *PN = dyn_cast<PHINode>(B))
1067     return relatedPHI(PN, A);
1068   if (const SelectInst *S = dyn_cast<SelectInst>(A))
1069     return relatedSelect(S, B);
1070   if (const SelectInst *S = dyn_cast<SelectInst>(B))
1071     return relatedSelect(S, A);
1072 
1073   // Conservative.
1074   return true;
1075 }
1076 
related(const Value * A,const Value * B)1077 bool ProvenanceAnalysis::related(const Value *A, const Value *B) {
1078   // Begin by inserting a conservative value into the map. If the insertion
1079   // fails, we have the answer already. If it succeeds, leave it there until we
1080   // compute the real answer to guard against recursive queries.
1081   if (A > B) std::swap(A, B);
1082   std::pair<CachedResultsTy::iterator, bool> Pair =
1083     CachedResults.insert(std::make_pair(ValuePairTy(A, B), true));
1084   if (!Pair.second)
1085     return Pair.first->second;
1086 
1087   bool Result = relatedCheck(A, B);
1088   CachedResults[ValuePairTy(A, B)] = Result;
1089   return Result;
1090 }
1091 
1092 namespace {
1093   // Sequence - A sequence of states that a pointer may go through in which an
1094   // objc_retain and objc_release are actually needed.
1095   enum Sequence {
1096     S_None,
1097     S_Retain,         ///< objc_retain(x)
1098     S_CanRelease,     ///< foo(x) -- x could possibly see a ref count decrement
1099     S_Use,            ///< any use of x
1100     S_Stop,           ///< like S_Release, but code motion is stopped
1101     S_Release,        ///< objc_release(x)
1102     S_MovableRelease  ///< objc_release(x), !clang.imprecise_release
1103   };
1104 }
1105 
MergeSeqs(Sequence A,Sequence B,bool TopDown)1106 static Sequence MergeSeqs(Sequence A, Sequence B, bool TopDown) {
1107   // The easy cases.
1108   if (A == B)
1109     return A;
1110   if (A == S_None || B == S_None)
1111     return S_None;
1112 
1113   if (A > B) std::swap(A, B);
1114   if (TopDown) {
1115     // Choose the side which is further along in the sequence.
1116     if ((A == S_Retain || A == S_CanRelease) &&
1117         (B == S_CanRelease || B == S_Use))
1118       return B;
1119   } else {
1120     // Choose the side which is further along in the sequence.
1121     if ((A == S_Use || A == S_CanRelease) &&
1122         (B == S_Use || B == S_Release || B == S_Stop || B == S_MovableRelease))
1123       return A;
1124     // If both sides are releases, choose the more conservative one.
1125     if (A == S_Stop && (B == S_Release || B == S_MovableRelease))
1126       return A;
1127     if (A == S_Release && B == S_MovableRelease)
1128       return A;
1129   }
1130 
1131   return S_None;
1132 }
1133 
1134 namespace {
1135   /// RRInfo - Unidirectional information about either a
1136   /// retain-decrement-use-release sequence or release-use-decrement-retain
1137   /// reverese sequence.
1138   struct RRInfo {
1139     /// KnownSafe - After an objc_retain, the reference count of the referenced
1140     /// object is known to be positive. Similarly, before an objc_release, the
1141     /// reference count of the referenced object is known to be positive. If
1142     /// there are retain-release pairs in code regions where the retain count
1143     /// is known to be positive, they can be eliminated, regardless of any side
1144     /// effects between them.
1145     ///
1146     /// Also, a retain+release pair nested within another retain+release
1147     /// pair all on the known same pointer value can be eliminated, regardless
1148     /// of any intervening side effects.
1149     ///
1150     /// KnownSafe is true when either of these conditions is satisfied.
1151     bool KnownSafe;
1152 
1153     /// IsRetainBlock - True if the Calls are objc_retainBlock calls (as
1154     /// opposed to objc_retain calls).
1155     bool IsRetainBlock;
1156 
1157     /// IsTailCallRelease - True of the objc_release calls are all marked
1158     /// with the "tail" keyword.
1159     bool IsTailCallRelease;
1160 
1161     /// ReleaseMetadata - If the Calls are objc_release calls and they all have
1162     /// a clang.imprecise_release tag, this is the metadata tag.
1163     MDNode *ReleaseMetadata;
1164 
1165     /// Calls - For a top-down sequence, the set of objc_retains or
1166     /// objc_retainBlocks. For bottom-up, the set of objc_releases.
1167     SmallPtrSet<Instruction *, 2> Calls;
1168 
1169     /// ReverseInsertPts - The set of optimal insert positions for
1170     /// moving calls in the opposite sequence.
1171     SmallPtrSet<Instruction *, 2> ReverseInsertPts;
1172 
RRInfo__anon1f24f85e0711::RRInfo1173     RRInfo() :
1174       KnownSafe(false), IsRetainBlock(false), IsTailCallRelease(false),
1175       ReleaseMetadata(0) {}
1176 
1177     void clear();
1178   };
1179 }
1180 
clear()1181 void RRInfo::clear() {
1182   KnownSafe = false;
1183   IsRetainBlock = false;
1184   IsTailCallRelease = false;
1185   ReleaseMetadata = 0;
1186   Calls.clear();
1187   ReverseInsertPts.clear();
1188 }
1189 
1190 namespace {
1191   /// PtrState - This class summarizes several per-pointer runtime properties
1192   /// which are propogated through the flow graph.
1193   class PtrState {
1194     /// RefCount - The known minimum number of reference count increments.
1195     unsigned RefCount;
1196 
1197     /// NestCount - The known minimum level of retain+release nesting.
1198     unsigned NestCount;
1199 
1200     /// Seq - The current position in the sequence.
1201     Sequence Seq;
1202 
1203   public:
1204     /// RRI - Unidirectional information about the current sequence.
1205     /// TODO: Encapsulate this better.
1206     RRInfo RRI;
1207 
PtrState()1208     PtrState() : RefCount(0), NestCount(0), Seq(S_None) {}
1209 
SetAtLeastOneRefCount()1210     void SetAtLeastOneRefCount()  {
1211       if (RefCount == 0) RefCount = 1;
1212     }
1213 
IncrementRefCount()1214     void IncrementRefCount() {
1215       if (RefCount != UINT_MAX) ++RefCount;
1216     }
1217 
DecrementRefCount()1218     void DecrementRefCount() {
1219       if (RefCount != 0) --RefCount;
1220     }
1221 
IsKnownIncremented() const1222     bool IsKnownIncremented() const {
1223       return RefCount > 0;
1224     }
1225 
IncrementNestCount()1226     void IncrementNestCount() {
1227       if (NestCount != UINT_MAX) ++NestCount;
1228     }
1229 
DecrementNestCount()1230     void DecrementNestCount() {
1231       if (NestCount != 0) --NestCount;
1232     }
1233 
IsKnownNested() const1234     bool IsKnownNested() const {
1235       return NestCount > 0;
1236     }
1237 
SetSeq(Sequence NewSeq)1238     void SetSeq(Sequence NewSeq) {
1239       Seq = NewSeq;
1240     }
1241 
SetSeqToRelease(MDNode * M)1242     void SetSeqToRelease(MDNode *M) {
1243       if (Seq == S_None || Seq == S_Use) {
1244         Seq = M ? S_MovableRelease : S_Release;
1245         RRI.ReleaseMetadata = M;
1246       } else if (Seq != S_MovableRelease || RRI.ReleaseMetadata != M) {
1247         Seq = S_Release;
1248         RRI.ReleaseMetadata = 0;
1249       }
1250     }
1251 
GetSeq() const1252     Sequence GetSeq() const {
1253       return Seq;
1254     }
1255 
ClearSequenceProgress()1256     void ClearSequenceProgress() {
1257       Seq = S_None;
1258       RRI.clear();
1259     }
1260 
1261     void Merge(const PtrState &Other, bool TopDown);
1262   };
1263 }
1264 
1265 void
Merge(const PtrState & Other,bool TopDown)1266 PtrState::Merge(const PtrState &Other, bool TopDown) {
1267   Seq = MergeSeqs(Seq, Other.Seq, TopDown);
1268   RefCount = std::min(RefCount, Other.RefCount);
1269   NestCount = std::min(NestCount, Other.NestCount);
1270 
1271   // We can't merge a plain objc_retain with an objc_retainBlock.
1272   if (RRI.IsRetainBlock != Other.RRI.IsRetainBlock)
1273     Seq = S_None;
1274 
1275   if (Seq == S_None) {
1276     RRI.clear();
1277   } else {
1278     // Conservatively merge the ReleaseMetadata information.
1279     if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
1280       RRI.ReleaseMetadata = 0;
1281 
1282     RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe;
1283     RRI.IsTailCallRelease = RRI.IsTailCallRelease && Other.RRI.IsTailCallRelease;
1284     RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
1285     RRI.ReverseInsertPts.insert(Other.RRI.ReverseInsertPts.begin(),
1286                                 Other.RRI.ReverseInsertPts.end());
1287   }
1288 }
1289 
1290 namespace {
1291   /// BBState - Per-BasicBlock state.
1292   class BBState {
1293     /// TopDownPathCount - The number of unique control paths from the entry
1294     /// which can reach this block.
1295     unsigned TopDownPathCount;
1296 
1297     /// BottomUpPathCount - The number of unique control paths to exits
1298     /// from this block.
1299     unsigned BottomUpPathCount;
1300 
1301     /// MapTy - A type for PerPtrTopDown and PerPtrBottomUp.
1302     typedef MapVector<const Value *, PtrState> MapTy;
1303 
1304     /// PerPtrTopDown - The top-down traversal uses this to record information
1305     /// known about a pointer at the bottom of each block.
1306     MapTy PerPtrTopDown;
1307 
1308     /// PerPtrBottomUp - The bottom-up traversal uses this to record information
1309     /// known about a pointer at the top of each block.
1310     MapTy PerPtrBottomUp;
1311 
1312   public:
BBState()1313     BBState() : TopDownPathCount(0), BottomUpPathCount(0) {}
1314 
1315     typedef MapTy::iterator ptr_iterator;
1316     typedef MapTy::const_iterator ptr_const_iterator;
1317 
top_down_ptr_begin()1318     ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); }
top_down_ptr_end()1319     ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); }
top_down_ptr_begin() const1320     ptr_const_iterator top_down_ptr_begin() const {
1321       return PerPtrTopDown.begin();
1322     }
top_down_ptr_end() const1323     ptr_const_iterator top_down_ptr_end() const {
1324       return PerPtrTopDown.end();
1325     }
1326 
bottom_up_ptr_begin()1327     ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); }
bottom_up_ptr_end()1328     ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); }
bottom_up_ptr_begin() const1329     ptr_const_iterator bottom_up_ptr_begin() const {
1330       return PerPtrBottomUp.begin();
1331     }
bottom_up_ptr_end() const1332     ptr_const_iterator bottom_up_ptr_end() const {
1333       return PerPtrBottomUp.end();
1334     }
1335 
1336     /// SetAsEntry - Mark this block as being an entry block, which has one
1337     /// path from the entry by definition.
SetAsEntry()1338     void SetAsEntry() { TopDownPathCount = 1; }
1339 
1340     /// SetAsExit - Mark this block as being an exit block, which has one
1341     /// path to an exit by definition.
SetAsExit()1342     void SetAsExit()  { BottomUpPathCount = 1; }
1343 
getPtrTopDownState(const Value * Arg)1344     PtrState &getPtrTopDownState(const Value *Arg) {
1345       return PerPtrTopDown[Arg];
1346     }
1347 
getPtrBottomUpState(const Value * Arg)1348     PtrState &getPtrBottomUpState(const Value *Arg) {
1349       return PerPtrBottomUp[Arg];
1350     }
1351 
clearBottomUpPointers()1352     void clearBottomUpPointers() {
1353       PerPtrBottomUp.clear();
1354     }
1355 
clearTopDownPointers()1356     void clearTopDownPointers() {
1357       PerPtrTopDown.clear();
1358     }
1359 
1360     void InitFromPred(const BBState &Other);
1361     void InitFromSucc(const BBState &Other);
1362     void MergePred(const BBState &Other);
1363     void MergeSucc(const BBState &Other);
1364 
1365     /// GetAllPathCount - Return the number of possible unique paths from an
1366     /// entry to an exit which pass through this block. This is only valid
1367     /// after both the top-down and bottom-up traversals are complete.
GetAllPathCount() const1368     unsigned GetAllPathCount() const {
1369       return TopDownPathCount * BottomUpPathCount;
1370     }
1371 
1372     /// IsVisitedTopDown - Test whether the block for this BBState has been
1373     /// visited by the top-down portion of the algorithm.
isVisitedTopDown() const1374     bool isVisitedTopDown() const {
1375       return TopDownPathCount != 0;
1376     }
1377   };
1378 }
1379 
InitFromPred(const BBState & Other)1380 void BBState::InitFromPred(const BBState &Other) {
1381   PerPtrTopDown = Other.PerPtrTopDown;
1382   TopDownPathCount = Other.TopDownPathCount;
1383 }
1384 
InitFromSucc(const BBState & Other)1385 void BBState::InitFromSucc(const BBState &Other) {
1386   PerPtrBottomUp = Other.PerPtrBottomUp;
1387   BottomUpPathCount = Other.BottomUpPathCount;
1388 }
1389 
1390 /// MergePred - The top-down traversal uses this to merge information about
1391 /// predecessors to form the initial state for a new block.
MergePred(const BBState & Other)1392 void BBState::MergePred(const BBState &Other) {
1393   // Other.TopDownPathCount can be 0, in which case it is either dead or a
1394   // loop backedge. Loop backedges are special.
1395   TopDownPathCount += Other.TopDownPathCount;
1396 
1397   // For each entry in the other set, if our set has an entry with the same key,
1398   // merge the entries. Otherwise, copy the entry and merge it with an empty
1399   // entry.
1400   for (ptr_const_iterator MI = Other.top_down_ptr_begin(),
1401        ME = Other.top_down_ptr_end(); MI != ME; ++MI) {
1402     std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI);
1403     Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1404                              /*TopDown=*/true);
1405   }
1406 
1407   // For each entry in our set, if the other set doesn't have an entry with the
1408   // same key, force it to merge with an empty entry.
1409   for (ptr_iterator MI = top_down_ptr_begin(),
1410        ME = top_down_ptr_end(); MI != ME; ++MI)
1411     if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end())
1412       MI->second.Merge(PtrState(), /*TopDown=*/true);
1413 }
1414 
1415 /// MergeSucc - The bottom-up traversal uses this to merge information about
1416 /// successors to form the initial state for a new block.
MergeSucc(const BBState & Other)1417 void BBState::MergeSucc(const BBState &Other) {
1418   // Other.BottomUpPathCount can be 0, in which case it is either dead or a
1419   // loop backedge. Loop backedges are special.
1420   BottomUpPathCount += Other.BottomUpPathCount;
1421 
1422   // For each entry in the other set, if our set has an entry with the
1423   // same key, merge the entries. Otherwise, copy the entry and merge
1424   // it with an empty entry.
1425   for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(),
1426        ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) {
1427     std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI);
1428     Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
1429                              /*TopDown=*/false);
1430   }
1431 
1432   // For each entry in our set, if the other set doesn't have an entry
1433   // with the same key, force it to merge with an empty entry.
1434   for (ptr_iterator MI = bottom_up_ptr_begin(),
1435        ME = bottom_up_ptr_end(); MI != ME; ++MI)
1436     if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end())
1437       MI->second.Merge(PtrState(), /*TopDown=*/false);
1438 }
1439 
1440 namespace {
1441   /// ObjCARCOpt - The main ARC optimization pass.
1442   class ObjCARCOpt : public FunctionPass {
1443     bool Changed;
1444     ProvenanceAnalysis PA;
1445 
1446     /// Run - A flag indicating whether this optimization pass should run.
1447     bool Run;
1448 
1449     /// RetainRVCallee, etc. - Declarations for ObjC runtime
1450     /// functions, for use in creating calls to them. These are initialized
1451     /// lazily to avoid cluttering up the Module with unused declarations.
1452     Constant *RetainRVCallee, *AutoreleaseRVCallee, *ReleaseCallee,
1453              *RetainCallee, *RetainBlockCallee, *AutoreleaseCallee;
1454 
1455     /// UsedInThisFunciton - Flags which determine whether each of the
1456     /// interesting runtine functions is in fact used in the current function.
1457     unsigned UsedInThisFunction;
1458 
1459     /// ImpreciseReleaseMDKind - The Metadata Kind for clang.imprecise_release
1460     /// metadata.
1461     unsigned ImpreciseReleaseMDKind;
1462 
1463     Constant *getRetainRVCallee(Module *M);
1464     Constant *getAutoreleaseRVCallee(Module *M);
1465     Constant *getReleaseCallee(Module *M);
1466     Constant *getRetainCallee(Module *M);
1467     Constant *getRetainBlockCallee(Module *M);
1468     Constant *getAutoreleaseCallee(Module *M);
1469 
1470     void OptimizeRetainCall(Function &F, Instruction *Retain);
1471     bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
1472     void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV);
1473     void OptimizeIndividualCalls(Function &F);
1474 
1475     void CheckForCFGHazards(const BasicBlock *BB,
1476                             DenseMap<const BasicBlock *, BBState> &BBStates,
1477                             BBState &MyStates) const;
1478     bool VisitBottomUp(BasicBlock *BB,
1479                        DenseMap<const BasicBlock *, BBState> &BBStates,
1480                        MapVector<Value *, RRInfo> &Retains);
1481     bool VisitTopDown(BasicBlock *BB,
1482                       DenseMap<const BasicBlock *, BBState> &BBStates,
1483                       DenseMap<Value *, RRInfo> &Releases);
1484     bool Visit(Function &F,
1485                DenseMap<const BasicBlock *, BBState> &BBStates,
1486                MapVector<Value *, RRInfo> &Retains,
1487                DenseMap<Value *, RRInfo> &Releases);
1488 
1489     void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1490                    MapVector<Value *, RRInfo> &Retains,
1491                    DenseMap<Value *, RRInfo> &Releases,
1492                    SmallVectorImpl<Instruction *> &DeadInsts,
1493                    Module *M);
1494 
1495     bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1496                               MapVector<Value *, RRInfo> &Retains,
1497                               DenseMap<Value *, RRInfo> &Releases,
1498                               Module *M);
1499 
1500     void OptimizeWeakCalls(Function &F);
1501 
1502     bool OptimizeSequences(Function &F);
1503 
1504     void OptimizeReturns(Function &F);
1505 
1506     virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1507     virtual bool doInitialization(Module &M);
1508     virtual bool runOnFunction(Function &F);
1509     virtual void releaseMemory();
1510 
1511   public:
1512     static char ID;
ObjCARCOpt()1513     ObjCARCOpt() : FunctionPass(ID) {
1514       initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1515     }
1516   };
1517 }
1518 
1519 char ObjCARCOpt::ID = 0;
1520 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1521                       "objc-arc", "ObjC ARC optimization", false, false)
INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)1522 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1523 INITIALIZE_PASS_END(ObjCARCOpt,
1524                     "objc-arc", "ObjC ARC optimization", false, false)
1525 
1526 Pass *llvm::createObjCARCOptPass() {
1527   return new ObjCARCOpt();
1528 }
1529 
getAnalysisUsage(AnalysisUsage & AU) const1530 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1531   AU.addRequired<ObjCARCAliasAnalysis>();
1532   AU.addRequired<AliasAnalysis>();
1533   // ARC optimization doesn't currently split critical edges.
1534   AU.setPreservesCFG();
1535 }
1536 
getRetainRVCallee(Module * M)1537 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1538   if (!RetainRVCallee) {
1539     LLVMContext &C = M->getContext();
1540     Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1541     std::vector<Type *> Params;
1542     Params.push_back(I8X);
1543     FunctionType *FTy =
1544       FunctionType::get(I8X, Params, /*isVarArg=*/false);
1545     AttrListPtr Attributes;
1546     Attributes.addAttr(~0u, Attribute::NoUnwind);
1547     RetainRVCallee =
1548       M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1549                              Attributes);
1550   }
1551   return RetainRVCallee;
1552 }
1553 
getAutoreleaseRVCallee(Module * M)1554 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1555   if (!AutoreleaseRVCallee) {
1556     LLVMContext &C = M->getContext();
1557     Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1558     std::vector<Type *> Params;
1559     Params.push_back(I8X);
1560     FunctionType *FTy =
1561       FunctionType::get(I8X, Params, /*isVarArg=*/false);
1562     AttrListPtr Attributes;
1563     Attributes.addAttr(~0u, Attribute::NoUnwind);
1564     AutoreleaseRVCallee =
1565       M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1566                              Attributes);
1567   }
1568   return AutoreleaseRVCallee;
1569 }
1570 
getReleaseCallee(Module * M)1571 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1572   if (!ReleaseCallee) {
1573     LLVMContext &C = M->getContext();
1574     std::vector<Type *> Params;
1575     Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1576     AttrListPtr Attributes;
1577     Attributes.addAttr(~0u, Attribute::NoUnwind);
1578     ReleaseCallee =
1579       M->getOrInsertFunction(
1580         "objc_release",
1581         FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1582         Attributes);
1583   }
1584   return ReleaseCallee;
1585 }
1586 
getRetainCallee(Module * M)1587 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1588   if (!RetainCallee) {
1589     LLVMContext &C = M->getContext();
1590     std::vector<Type *> Params;
1591     Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1592     AttrListPtr Attributes;
1593     Attributes.addAttr(~0u, Attribute::NoUnwind);
1594     RetainCallee =
1595       M->getOrInsertFunction(
1596         "objc_retain",
1597         FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1598         Attributes);
1599   }
1600   return RetainCallee;
1601 }
1602 
getRetainBlockCallee(Module * M)1603 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1604   if (!RetainBlockCallee) {
1605     LLVMContext &C = M->getContext();
1606     std::vector<Type *> Params;
1607     Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1608     AttrListPtr Attributes;
1609     // objc_retainBlock is not nounwind because it calls user copy constructors
1610     // which could theoretically throw.
1611     RetainBlockCallee =
1612       M->getOrInsertFunction(
1613         "objc_retainBlock",
1614         FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1615         Attributes);
1616   }
1617   return RetainBlockCallee;
1618 }
1619 
getAutoreleaseCallee(Module * M)1620 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1621   if (!AutoreleaseCallee) {
1622     LLVMContext &C = M->getContext();
1623     std::vector<Type *> Params;
1624     Params.push_back(PointerType::getUnqual(Type::getInt8Ty(C)));
1625     AttrListPtr Attributes;
1626     Attributes.addAttr(~0u, Attribute::NoUnwind);
1627     AutoreleaseCallee =
1628       M->getOrInsertFunction(
1629         "objc_autorelease",
1630         FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1631         Attributes);
1632   }
1633   return AutoreleaseCallee;
1634 }
1635 
1636 /// CanAlterRefCount - Test whether the given instruction can result in a
1637 /// reference count modification (positive or negative) for the pointer's
1638 /// object.
1639 static bool
CanAlterRefCount(const Instruction * Inst,const Value * Ptr,ProvenanceAnalysis & PA,InstructionClass Class)1640 CanAlterRefCount(const Instruction *Inst, const Value *Ptr,
1641                  ProvenanceAnalysis &PA, InstructionClass Class) {
1642   switch (Class) {
1643   case IC_Autorelease:
1644   case IC_AutoreleaseRV:
1645   case IC_User:
1646     // These operations never directly modify a reference count.
1647     return false;
1648   default: break;
1649   }
1650 
1651   ImmutableCallSite CS = static_cast<const Value *>(Inst);
1652   assert(CS && "Only calls can alter reference counts!");
1653 
1654   // See if AliasAnalysis can help us with the call.
1655   AliasAnalysis::ModRefBehavior MRB = PA.getAA()->getModRefBehavior(CS);
1656   if (AliasAnalysis::onlyReadsMemory(MRB))
1657     return false;
1658   if (AliasAnalysis::onlyAccessesArgPointees(MRB)) {
1659     for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1660          I != E; ++I) {
1661       const Value *Op = *I;
1662       if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1663         return true;
1664     }
1665     return false;
1666   }
1667 
1668   // Assume the worst.
1669   return true;
1670 }
1671 
1672 /// CanUse - Test whether the given instruction can "use" the given pointer's
1673 /// object in a way that requires the reference count to be positive.
1674 static bool
CanUse(const Instruction * Inst,const Value * Ptr,ProvenanceAnalysis & PA,InstructionClass Class)1675 CanUse(const Instruction *Inst, const Value *Ptr, ProvenanceAnalysis &PA,
1676        InstructionClass Class) {
1677   // IC_Call operations (as opposed to IC_CallOrUser) never "use" objc pointers.
1678   if (Class == IC_Call)
1679     return false;
1680 
1681   // Consider various instructions which may have pointer arguments which are
1682   // not "uses".
1683   if (const ICmpInst *ICI = dyn_cast<ICmpInst>(Inst)) {
1684     // Comparing a pointer with null, or any other constant, isn't really a use,
1685     // because we don't care what the pointer points to, or about the values
1686     // of any other dynamic reference-counted pointers.
1687     if (!IsPotentialUse(ICI->getOperand(1)))
1688       return false;
1689   } else if (ImmutableCallSite CS = static_cast<const Value *>(Inst)) {
1690     // For calls, just check the arguments (and not the callee operand).
1691     for (ImmutableCallSite::arg_iterator OI = CS.arg_begin(),
1692          OE = CS.arg_end(); OI != OE; ++OI) {
1693       const Value *Op = *OI;
1694       if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1695         return true;
1696     }
1697     return false;
1698   } else if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1699     // Special-case stores, because we don't care about the stored value, just
1700     // the store address.
1701     const Value *Op = GetUnderlyingObjCPtr(SI->getPointerOperand());
1702     // If we can't tell what the underlying object was, assume there is a
1703     // dependence.
1704     return IsPotentialUse(Op) && PA.related(Op, Ptr);
1705   }
1706 
1707   // Check each operand for a match.
1708   for (User::const_op_iterator OI = Inst->op_begin(), OE = Inst->op_end();
1709        OI != OE; ++OI) {
1710     const Value *Op = *OI;
1711     if (IsPotentialUse(Op) && PA.related(Ptr, Op))
1712       return true;
1713   }
1714   return false;
1715 }
1716 
1717 /// CanInterruptRV - Test whether the given instruction can autorelease
1718 /// any pointer or cause an autoreleasepool pop.
1719 static bool
CanInterruptRV(InstructionClass Class)1720 CanInterruptRV(InstructionClass Class) {
1721   switch (Class) {
1722   case IC_AutoreleasepoolPop:
1723   case IC_CallOrUser:
1724   case IC_Call:
1725   case IC_Autorelease:
1726   case IC_AutoreleaseRV:
1727   case IC_FusedRetainAutorelease:
1728   case IC_FusedRetainAutoreleaseRV:
1729     return true;
1730   default:
1731     return false;
1732   }
1733 }
1734 
1735 namespace {
1736   /// DependenceKind - There are several kinds of dependence-like concepts in
1737   /// use here.
1738   enum DependenceKind {
1739     NeedsPositiveRetainCount,
1740     CanChangeRetainCount,
1741     RetainAutoreleaseDep,       ///< Blocks objc_retainAutorelease.
1742     RetainAutoreleaseRVDep,     ///< Blocks objc_retainAutoreleaseReturnValue.
1743     RetainRVDep                 ///< Blocks objc_retainAutoreleasedReturnValue.
1744   };
1745 }
1746 
1747 /// Depends - Test if there can be dependencies on Inst through Arg. This
1748 /// function only tests dependencies relevant for removing pairs of calls.
1749 static bool
Depends(DependenceKind Flavor,Instruction * Inst,const Value * Arg,ProvenanceAnalysis & PA)1750 Depends(DependenceKind Flavor, Instruction *Inst, const Value *Arg,
1751         ProvenanceAnalysis &PA) {
1752   // If we've reached the definition of Arg, stop.
1753   if (Inst == Arg)
1754     return true;
1755 
1756   switch (Flavor) {
1757   case NeedsPositiveRetainCount: {
1758     InstructionClass Class = GetInstructionClass(Inst);
1759     switch (Class) {
1760     case IC_AutoreleasepoolPop:
1761     case IC_AutoreleasepoolPush:
1762     case IC_None:
1763       return false;
1764     default:
1765       return CanUse(Inst, Arg, PA, Class);
1766     }
1767   }
1768 
1769   case CanChangeRetainCount: {
1770     InstructionClass Class = GetInstructionClass(Inst);
1771     switch (Class) {
1772     case IC_AutoreleasepoolPop:
1773       // Conservatively assume this can decrement any count.
1774       return true;
1775     case IC_AutoreleasepoolPush:
1776     case IC_None:
1777       return false;
1778     default:
1779       return CanAlterRefCount(Inst, Arg, PA, Class);
1780     }
1781   }
1782 
1783   case RetainAutoreleaseDep:
1784     switch (GetBasicInstructionClass(Inst)) {
1785     case IC_AutoreleasepoolPop:
1786       // Don't merge an objc_autorelease with an objc_retain inside a different
1787       // autoreleasepool scope.
1788       return true;
1789     case IC_Retain:
1790     case IC_RetainRV:
1791       // Check for a retain of the same pointer for merging.
1792       return GetObjCArg(Inst) == Arg;
1793     default:
1794       // Nothing else matters for objc_retainAutorelease formation.
1795       return false;
1796     }
1797     break;
1798 
1799   case RetainAutoreleaseRVDep: {
1800     InstructionClass Class = GetBasicInstructionClass(Inst);
1801     switch (Class) {
1802     case IC_Retain:
1803     case IC_RetainRV:
1804       // Check for a retain of the same pointer for merging.
1805       return GetObjCArg(Inst) == Arg;
1806     default:
1807       // Anything that can autorelease interrupts
1808       // retainAutoreleaseReturnValue formation.
1809       return CanInterruptRV(Class);
1810     }
1811     break;
1812   }
1813 
1814   case RetainRVDep:
1815     return CanInterruptRV(GetBasicInstructionClass(Inst));
1816   }
1817 
1818   llvm_unreachable("Invalid dependence flavor");
1819   return true;
1820 }
1821 
1822 /// FindDependencies - Walk up the CFG from StartPos (which is in StartBB) and
1823 /// find local and non-local dependencies on Arg.
1824 /// TODO: Cache results?
1825 static void
FindDependencies(DependenceKind Flavor,const Value * Arg,BasicBlock * StartBB,Instruction * StartInst,SmallPtrSet<Instruction *,4> & DependingInstructions,SmallPtrSet<const BasicBlock *,4> & Visited,ProvenanceAnalysis & PA)1826 FindDependencies(DependenceKind Flavor,
1827                  const Value *Arg,
1828                  BasicBlock *StartBB, Instruction *StartInst,
1829                  SmallPtrSet<Instruction *, 4> &DependingInstructions,
1830                  SmallPtrSet<const BasicBlock *, 4> &Visited,
1831                  ProvenanceAnalysis &PA) {
1832   BasicBlock::iterator StartPos = StartInst;
1833 
1834   SmallVector<std::pair<BasicBlock *, BasicBlock::iterator>, 4> Worklist;
1835   Worklist.push_back(std::make_pair(StartBB, StartPos));
1836   do {
1837     std::pair<BasicBlock *, BasicBlock::iterator> Pair =
1838       Worklist.pop_back_val();
1839     BasicBlock *LocalStartBB = Pair.first;
1840     BasicBlock::iterator LocalStartPos = Pair.second;
1841     BasicBlock::iterator StartBBBegin = LocalStartBB->begin();
1842     for (;;) {
1843       if (LocalStartPos == StartBBBegin) {
1844         pred_iterator PI(LocalStartBB), PE(LocalStartBB, false);
1845         if (PI == PE)
1846           // If we've reached the function entry, produce a null dependence.
1847           DependingInstructions.insert(0);
1848         else
1849           // Add the predecessors to the worklist.
1850           do {
1851             BasicBlock *PredBB = *PI;
1852             if (Visited.insert(PredBB))
1853               Worklist.push_back(std::make_pair(PredBB, PredBB->end()));
1854           } while (++PI != PE);
1855         break;
1856       }
1857 
1858       Instruction *Inst = --LocalStartPos;
1859       if (Depends(Flavor, Inst, Arg, PA)) {
1860         DependingInstructions.insert(Inst);
1861         break;
1862       }
1863     }
1864   } while (!Worklist.empty());
1865 
1866   // Determine whether the original StartBB post-dominates all of the blocks we
1867   // visited. If not, insert a sentinal indicating that most optimizations are
1868   // not safe.
1869   for (SmallPtrSet<const BasicBlock *, 4>::const_iterator I = Visited.begin(),
1870        E = Visited.end(); I != E; ++I) {
1871     const BasicBlock *BB = *I;
1872     if (BB == StartBB)
1873       continue;
1874     const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1875     for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
1876       const BasicBlock *Succ = *SI;
1877       if (Succ != StartBB && !Visited.count(Succ)) {
1878         DependingInstructions.insert(reinterpret_cast<Instruction *>(-1));
1879         return;
1880       }
1881     }
1882   }
1883 }
1884 
isNullOrUndef(const Value * V)1885 static bool isNullOrUndef(const Value *V) {
1886   return isa<ConstantPointerNull>(V) || isa<UndefValue>(V);
1887 }
1888 
isNoopInstruction(const Instruction * I)1889 static bool isNoopInstruction(const Instruction *I) {
1890   return isa<BitCastInst>(I) ||
1891          (isa<GetElementPtrInst>(I) &&
1892           cast<GetElementPtrInst>(I)->hasAllZeroIndices());
1893 }
1894 
1895 /// OptimizeRetainCall - Turn objc_retain into
1896 /// objc_retainAutoreleasedReturnValue if the operand is a return value.
1897 void
OptimizeRetainCall(Function & F,Instruction * Retain)1898 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
1899   CallSite CS(GetObjCArg(Retain));
1900   Instruction *Call = CS.getInstruction();
1901   if (!Call) return;
1902   if (Call->getParent() != Retain->getParent()) return;
1903 
1904   // Check that the call is next to the retain.
1905   BasicBlock::iterator I = Call;
1906   ++I;
1907   while (isNoopInstruction(I)) ++I;
1908   if (&*I != Retain)
1909     return;
1910 
1911   // Turn it to an objc_retainAutoreleasedReturnValue..
1912   Changed = true;
1913   ++NumPeeps;
1914   cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
1915 }
1916 
1917 /// OptimizeRetainRVCall - Turn objc_retainAutoreleasedReturnValue into
1918 /// objc_retain if the operand is not a return value.  Or, if it can be
1919 /// paired with an objc_autoreleaseReturnValue, delete the pair and
1920 /// return true.
1921 bool
OptimizeRetainRVCall(Function & F,Instruction * RetainRV)1922 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
1923   // Check for the argument being from an immediately preceding call.
1924   Value *Arg = GetObjCArg(RetainRV);
1925   CallSite CS(Arg);
1926   if (Instruction *Call = CS.getInstruction())
1927     if (Call->getParent() == RetainRV->getParent()) {
1928       BasicBlock::iterator I = Call;
1929       ++I;
1930       while (isNoopInstruction(I)) ++I;
1931       if (&*I == RetainRV)
1932         return false;
1933     }
1934 
1935   // Check for being preceded by an objc_autoreleaseReturnValue on the same
1936   // pointer. In this case, we can delete the pair.
1937   BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
1938   if (I != Begin) {
1939     do --I; while (I != Begin && isNoopInstruction(I));
1940     if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
1941         GetObjCArg(I) == Arg) {
1942       Changed = true;
1943       ++NumPeeps;
1944       EraseInstruction(I);
1945       EraseInstruction(RetainRV);
1946       return true;
1947     }
1948   }
1949 
1950   // Turn it to a plain objc_retain.
1951   Changed = true;
1952   ++NumPeeps;
1953   cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
1954   return false;
1955 }
1956 
1957 /// OptimizeAutoreleaseRVCall - Turn objc_autoreleaseReturnValue into
1958 /// objc_autorelease if the result is not used as a return value.
1959 void
OptimizeAutoreleaseRVCall(Function & F,Instruction * AutoreleaseRV)1960 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV) {
1961   // Check for a return of the pointer value.
1962   const Value *Ptr = GetObjCArg(AutoreleaseRV);
1963   SmallVector<const Value *, 2> Users;
1964   Users.push_back(Ptr);
1965   do {
1966     Ptr = Users.pop_back_val();
1967     for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
1968          UI != UE; ++UI) {
1969       const User *I = *UI;
1970       if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
1971         return;
1972       if (isa<BitCastInst>(I))
1973         Users.push_back(I);
1974     }
1975   } while (!Users.empty());
1976 
1977   Changed = true;
1978   ++NumPeeps;
1979   cast<CallInst>(AutoreleaseRV)->
1980     setCalledFunction(getAutoreleaseCallee(F.getParent()));
1981 }
1982 
1983 /// OptimizeIndividualCalls - Visit each call, one at a time, and make
1984 /// simplifications without doing any additional analysis.
OptimizeIndividualCalls(Function & F)1985 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
1986   // Reset all the flags in preparation for recomputing them.
1987   UsedInThisFunction = 0;
1988 
1989   // Visit all objc_* calls in F.
1990   for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
1991     Instruction *Inst = &*I++;
1992     InstructionClass Class = GetBasicInstructionClass(Inst);
1993 
1994     switch (Class) {
1995     default: break;
1996 
1997     // Delete no-op casts. These function calls have special semantics, but
1998     // the semantics are entirely implemented via lowering in the front-end,
1999     // so by the time they reach the optimizer, they are just no-op calls
2000     // which return their argument.
2001     //
2002     // There are gray areas here, as the ability to cast reference-counted
2003     // pointers to raw void* and back allows code to break ARC assumptions,
2004     // however these are currently considered to be unimportant.
2005     case IC_NoopCast:
2006       Changed = true;
2007       ++NumNoops;
2008       EraseInstruction(Inst);
2009       continue;
2010 
2011     // If the pointer-to-weak-pointer is null, it's undefined behavior.
2012     case IC_StoreWeak:
2013     case IC_LoadWeak:
2014     case IC_LoadWeakRetained:
2015     case IC_InitWeak:
2016     case IC_DestroyWeak: {
2017       CallInst *CI = cast<CallInst>(Inst);
2018       if (isNullOrUndef(CI->getArgOperand(0))) {
2019         Type *Ty = CI->getArgOperand(0)->getType();
2020         new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2021                       Constant::getNullValue(Ty),
2022                       CI);
2023         CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
2024         CI->eraseFromParent();
2025         continue;
2026       }
2027       break;
2028     }
2029     case IC_CopyWeak:
2030     case IC_MoveWeak: {
2031       CallInst *CI = cast<CallInst>(Inst);
2032       if (isNullOrUndef(CI->getArgOperand(0)) ||
2033           isNullOrUndef(CI->getArgOperand(1))) {
2034         Type *Ty = CI->getArgOperand(0)->getType();
2035         new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
2036                       Constant::getNullValue(Ty),
2037                       CI);
2038         CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
2039         CI->eraseFromParent();
2040         continue;
2041       }
2042       break;
2043     }
2044     case IC_Retain:
2045       OptimizeRetainCall(F, Inst);
2046       break;
2047     case IC_RetainRV:
2048       if (OptimizeRetainRVCall(F, Inst))
2049         continue;
2050       break;
2051     case IC_AutoreleaseRV:
2052       OptimizeAutoreleaseRVCall(F, Inst);
2053       break;
2054     }
2055 
2056     // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
2057     if (IsAutorelease(Class) && Inst->use_empty()) {
2058       CallInst *Call = cast<CallInst>(Inst);
2059       const Value *Arg = Call->getArgOperand(0);
2060       Arg = FindSingleUseIdentifiedObject(Arg);
2061       if (Arg) {
2062         Changed = true;
2063         ++NumAutoreleases;
2064 
2065         // Create the declaration lazily.
2066         LLVMContext &C = Inst->getContext();
2067         CallInst *NewCall =
2068           CallInst::Create(getReleaseCallee(F.getParent()),
2069                            Call->getArgOperand(0), "", Call);
2070         NewCall->setMetadata(ImpreciseReleaseMDKind,
2071                              MDNode::get(C, ArrayRef<Value *>()));
2072         EraseInstruction(Call);
2073         Inst = NewCall;
2074         Class = IC_Release;
2075       }
2076     }
2077 
2078     // For functions which can never be passed stack arguments, add
2079     // a tail keyword.
2080     if (IsAlwaysTail(Class)) {
2081       Changed = true;
2082       cast<CallInst>(Inst)->setTailCall();
2083     }
2084 
2085     // Set nounwind as needed.
2086     if (IsNoThrow(Class)) {
2087       Changed = true;
2088       cast<CallInst>(Inst)->setDoesNotThrow();
2089     }
2090 
2091     if (!IsNoopOnNull(Class)) {
2092       UsedInThisFunction |= 1 << Class;
2093       continue;
2094     }
2095 
2096     const Value *Arg = GetObjCArg(Inst);
2097 
2098     // ARC calls with null are no-ops. Delete them.
2099     if (isNullOrUndef(Arg)) {
2100       Changed = true;
2101       ++NumNoops;
2102       EraseInstruction(Inst);
2103       continue;
2104     }
2105 
2106     // Keep track of which of retain, release, autorelease, and retain_block
2107     // are actually present in this function.
2108     UsedInThisFunction |= 1 << Class;
2109 
2110     // If Arg is a PHI, and one or more incoming values to the
2111     // PHI are null, and the call is control-equivalent to the PHI, and there
2112     // are no relevant side effects between the PHI and the call, the call
2113     // could be pushed up to just those paths with non-null incoming values.
2114     // For now, don't bother splitting critical edges for this.
2115     SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
2116     Worklist.push_back(std::make_pair(Inst, Arg));
2117     do {
2118       std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
2119       Inst = Pair.first;
2120       Arg = Pair.second;
2121 
2122       const PHINode *PN = dyn_cast<PHINode>(Arg);
2123       if (!PN) continue;
2124 
2125       // Determine if the PHI has any null operands, or any incoming
2126       // critical edges.
2127       bool HasNull = false;
2128       bool HasCriticalEdges = false;
2129       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2130         Value *Incoming =
2131           StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2132         if (isNullOrUndef(Incoming))
2133           HasNull = true;
2134         else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
2135                    .getNumSuccessors() != 1) {
2136           HasCriticalEdges = true;
2137           break;
2138         }
2139       }
2140       // If we have null operands and no critical edges, optimize.
2141       if (!HasCriticalEdges && HasNull) {
2142         SmallPtrSet<Instruction *, 4> DependingInstructions;
2143         SmallPtrSet<const BasicBlock *, 4> Visited;
2144 
2145         // Check that there is nothing that cares about the reference
2146         // count between the call and the phi.
2147         FindDependencies(NeedsPositiveRetainCount, Arg,
2148                          Inst->getParent(), Inst,
2149                          DependingInstructions, Visited, PA);
2150         if (DependingInstructions.size() == 1 &&
2151             *DependingInstructions.begin() == PN) {
2152           Changed = true;
2153           ++NumPartialNoops;
2154           // Clone the call into each predecessor that has a non-null value.
2155           CallInst *CInst = cast<CallInst>(Inst);
2156           Type *ParamTy = CInst->getArgOperand(0)->getType();
2157           for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2158             Value *Incoming =
2159               StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
2160             if (!isNullOrUndef(Incoming)) {
2161               CallInst *Clone = cast<CallInst>(CInst->clone());
2162               Value *Op = PN->getIncomingValue(i);
2163               Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
2164               if (Op->getType() != ParamTy)
2165                 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
2166               Clone->setArgOperand(0, Op);
2167               Clone->insertBefore(InsertPos);
2168               Worklist.push_back(std::make_pair(Clone, Incoming));
2169             }
2170           }
2171           // Erase the original call.
2172           EraseInstruction(CInst);
2173           continue;
2174         }
2175       }
2176     } while (!Worklist.empty());
2177   }
2178 }
2179 
2180 /// CheckForCFGHazards - Check for critical edges, loop boundaries, irreducible
2181 /// control flow, or other CFG structures where moving code across the edge
2182 /// would result in it being executed more.
2183 void
CheckForCFGHazards(const BasicBlock * BB,DenseMap<const BasicBlock *,BBState> & BBStates,BBState & MyStates) const2184 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
2185                                DenseMap<const BasicBlock *, BBState> &BBStates,
2186                                BBState &MyStates) const {
2187   // If any top-down local-use or possible-dec has a succ which is earlier in
2188   // the sequence, forget it.
2189   for (BBState::ptr_const_iterator I = MyStates.top_down_ptr_begin(),
2190        E = MyStates.top_down_ptr_end(); I != E; ++I)
2191     switch (I->second.GetSeq()) {
2192     default: break;
2193     case S_Use: {
2194       const Value *Arg = I->first;
2195       const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2196       bool SomeSuccHasSame = false;
2197       bool AllSuccsHaveSame = true;
2198       PtrState &S = MyStates.getPtrTopDownState(Arg);
2199       for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
2200         PtrState &SuccS = BBStates[*SI].getPtrBottomUpState(Arg);
2201         switch (SuccS.GetSeq()) {
2202         case S_None:
2203         case S_CanRelease: {
2204           if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2205             S.ClearSequenceProgress();
2206           continue;
2207         }
2208         case S_Use:
2209           SomeSuccHasSame = true;
2210           break;
2211         case S_Stop:
2212         case S_Release:
2213         case S_MovableRelease:
2214           if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2215             AllSuccsHaveSame = false;
2216           break;
2217         case S_Retain:
2218           llvm_unreachable("bottom-up pointer in retain state!");
2219         }
2220       }
2221       // If the state at the other end of any of the successor edges
2222       // matches the current state, require all edges to match. This
2223       // guards against loops in the middle of a sequence.
2224       if (SomeSuccHasSame && !AllSuccsHaveSame)
2225         S.ClearSequenceProgress();
2226     }
2227     case S_CanRelease: {
2228       const Value *Arg = I->first;
2229       const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2230       bool SomeSuccHasSame = false;
2231       bool AllSuccsHaveSame = true;
2232       PtrState &S = MyStates.getPtrTopDownState(Arg);
2233       for (succ_const_iterator SI(TI), SE(TI, false); SI != SE; ++SI) {
2234         PtrState &SuccS = BBStates[*SI].getPtrBottomUpState(Arg);
2235         switch (SuccS.GetSeq()) {
2236         case S_None: {
2237           if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2238             S.ClearSequenceProgress();
2239           continue;
2240         }
2241         case S_CanRelease:
2242           SomeSuccHasSame = true;
2243           break;
2244         case S_Stop:
2245         case S_Release:
2246         case S_MovableRelease:
2247         case S_Use:
2248           if (!S.RRI.KnownSafe && !SuccS.RRI.KnownSafe)
2249             AllSuccsHaveSame = false;
2250           break;
2251         case S_Retain:
2252           llvm_unreachable("bottom-up pointer in retain state!");
2253         }
2254       }
2255       // If the state at the other end of any of the successor edges
2256       // matches the current state, require all edges to match. This
2257       // guards against loops in the middle of a sequence.
2258       if (SomeSuccHasSame && !AllSuccsHaveSame)
2259         S.ClearSequenceProgress();
2260     }
2261     }
2262 }
2263 
2264 bool
VisitBottomUp(BasicBlock * BB,DenseMap<const BasicBlock *,BBState> & BBStates,MapVector<Value *,RRInfo> & Retains)2265 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
2266                           DenseMap<const BasicBlock *, BBState> &BBStates,
2267                           MapVector<Value *, RRInfo> &Retains) {
2268   bool NestingDetected = false;
2269   BBState &MyStates = BBStates[BB];
2270 
2271   // Merge the states from each successor to compute the initial state
2272   // for the current block.
2273   const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
2274   succ_const_iterator SI(TI), SE(TI, false);
2275   if (SI == SE)
2276     MyStates.SetAsExit();
2277   else
2278     do {
2279       const BasicBlock *Succ = *SI++;
2280       if (Succ == BB)
2281         continue;
2282       DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
2283       // If we haven't seen this node yet, then we've found a CFG cycle.
2284       // Be optimistic here; it's CheckForCFGHazards' job detect trouble.
2285       if (I == BBStates.end())
2286         continue;
2287       MyStates.InitFromSucc(I->second);
2288       while (SI != SE) {
2289         Succ = *SI++;
2290         if (Succ != BB) {
2291           I = BBStates.find(Succ);
2292           if (I != BBStates.end())
2293             MyStates.MergeSucc(I->second);
2294         }
2295       }
2296       break;
2297     } while (SI != SE);
2298 
2299   // Visit all the instructions, bottom-up.
2300   for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
2301     Instruction *Inst = llvm::prior(I);
2302     InstructionClass Class = GetInstructionClass(Inst);
2303     const Value *Arg = 0;
2304 
2305     switch (Class) {
2306     case IC_Release: {
2307       Arg = GetObjCArg(Inst);
2308 
2309       PtrState &S = MyStates.getPtrBottomUpState(Arg);
2310 
2311       // If we see two releases in a row on the same pointer. If so, make
2312       // a note, and we'll cicle back to revisit it after we've
2313       // hopefully eliminated the second release, which may allow us to
2314       // eliminate the first release too.
2315       // Theoretically we could implement removal of nested retain+release
2316       // pairs by making PtrState hold a stack of states, but this is
2317       // simple and avoids adding overhead for the non-nested case.
2318       if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease)
2319         NestingDetected = true;
2320 
2321       S.SetSeqToRelease(Inst->getMetadata(ImpreciseReleaseMDKind));
2322       S.RRI.clear();
2323       S.RRI.KnownSafe = S.IsKnownNested() || S.IsKnownIncremented();
2324       S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2325       S.RRI.Calls.insert(Inst);
2326 
2327       S.IncrementRefCount();
2328       S.IncrementNestCount();
2329       break;
2330     }
2331     case IC_RetainBlock:
2332     case IC_Retain:
2333     case IC_RetainRV: {
2334       Arg = GetObjCArg(Inst);
2335 
2336       PtrState &S = MyStates.getPtrBottomUpState(Arg);
2337       S.DecrementRefCount();
2338       S.SetAtLeastOneRefCount();
2339       S.DecrementNestCount();
2340 
2341       // An objc_retainBlock call with just a use still needs to be kept,
2342       // because it may be copying a block from the stack to the heap.
2343       if (Class == IC_RetainBlock && S.GetSeq() == S_Use)
2344         S.SetSeq(S_CanRelease);
2345 
2346       switch (S.GetSeq()) {
2347       case S_Stop:
2348       case S_Release:
2349       case S_MovableRelease:
2350       case S_Use:
2351         S.RRI.ReverseInsertPts.clear();
2352         // FALL THROUGH
2353       case S_CanRelease:
2354         // Don't do retain+release tracking for IC_RetainRV, because it's
2355         // better to let it remain as the first instruction after a call.
2356         if (Class != IC_RetainRV) {
2357           S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2358           Retains[Inst] = S.RRI;
2359         }
2360         S.ClearSequenceProgress();
2361         break;
2362       case S_None:
2363         break;
2364       case S_Retain:
2365         llvm_unreachable("bottom-up pointer in retain state!");
2366       }
2367       continue;
2368     }
2369     case IC_AutoreleasepoolPop:
2370       // Conservatively, clear MyStates for all known pointers.
2371       MyStates.clearBottomUpPointers();
2372       continue;
2373     case IC_AutoreleasepoolPush:
2374     case IC_None:
2375       // These are irrelevant.
2376       continue;
2377     default:
2378       break;
2379     }
2380 
2381     // Consider any other possible effects of this instruction on each
2382     // pointer being tracked.
2383     for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
2384          ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
2385       const Value *Ptr = MI->first;
2386       if (Ptr == Arg)
2387         continue; // Handled above.
2388       PtrState &S = MI->second;
2389       Sequence Seq = S.GetSeq();
2390 
2391       // Check for possible releases.
2392       if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2393         S.DecrementRefCount();
2394         switch (Seq) {
2395         case S_Use:
2396           S.SetSeq(S_CanRelease);
2397           continue;
2398         case S_CanRelease:
2399         case S_Release:
2400         case S_MovableRelease:
2401         case S_Stop:
2402         case S_None:
2403           break;
2404         case S_Retain:
2405           llvm_unreachable("bottom-up pointer in retain state!");
2406         }
2407       }
2408 
2409       // Check for possible direct uses.
2410       switch (Seq) {
2411       case S_Release:
2412       case S_MovableRelease:
2413         if (CanUse(Inst, Ptr, PA, Class)) {
2414           assert(S.RRI.ReverseInsertPts.empty());
2415           S.RRI.ReverseInsertPts.insert(Inst);
2416           S.SetSeq(S_Use);
2417         } else if (Seq == S_Release &&
2418                    (Class == IC_User || Class == IC_CallOrUser)) {
2419           // Non-movable releases depend on any possible objc pointer use.
2420           S.SetSeq(S_Stop);
2421           assert(S.RRI.ReverseInsertPts.empty());
2422           S.RRI.ReverseInsertPts.insert(Inst);
2423         }
2424         break;
2425       case S_Stop:
2426         if (CanUse(Inst, Ptr, PA, Class))
2427           S.SetSeq(S_Use);
2428         break;
2429       case S_CanRelease:
2430       case S_Use:
2431       case S_None:
2432         break;
2433       case S_Retain:
2434         llvm_unreachable("bottom-up pointer in retain state!");
2435       }
2436     }
2437   }
2438 
2439   return NestingDetected;
2440 }
2441 
2442 bool
VisitTopDown(BasicBlock * BB,DenseMap<const BasicBlock *,BBState> & BBStates,DenseMap<Value *,RRInfo> & Releases)2443 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
2444                          DenseMap<const BasicBlock *, BBState> &BBStates,
2445                          DenseMap<Value *, RRInfo> &Releases) {
2446   bool NestingDetected = false;
2447   BBState &MyStates = BBStates[BB];
2448 
2449   // Merge the states from each predecessor to compute the initial state
2450   // for the current block.
2451   const_pred_iterator PI(BB), PE(BB, false);
2452   if (PI == PE)
2453     MyStates.SetAsEntry();
2454   else
2455     do {
2456       const BasicBlock *Pred = *PI++;
2457       if (Pred == BB)
2458         continue;
2459       DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
2460       assert(I != BBStates.end());
2461       // If we haven't seen this node yet, then we've found a CFG cycle.
2462       // Be optimistic here; it's CheckForCFGHazards' job detect trouble.
2463       if (!I->second.isVisitedTopDown())
2464         continue;
2465       MyStates.InitFromPred(I->second);
2466       while (PI != PE) {
2467         Pred = *PI++;
2468         if (Pred != BB) {
2469           I = BBStates.find(Pred);
2470           assert(I != BBStates.end());
2471           if (I->second.isVisitedTopDown())
2472             MyStates.MergePred(I->second);
2473         }
2474       }
2475       break;
2476     } while (PI != PE);
2477 
2478   // Visit all the instructions, top-down.
2479   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2480     Instruction *Inst = I;
2481     InstructionClass Class = GetInstructionClass(Inst);
2482     const Value *Arg = 0;
2483 
2484     switch (Class) {
2485     case IC_RetainBlock:
2486     case IC_Retain:
2487     case IC_RetainRV: {
2488       Arg = GetObjCArg(Inst);
2489 
2490       PtrState &S = MyStates.getPtrTopDownState(Arg);
2491 
2492       // Don't do retain+release tracking for IC_RetainRV, because it's
2493       // better to let it remain as the first instruction after a call.
2494       if (Class != IC_RetainRV) {
2495         // If we see two retains in a row on the same pointer. If so, make
2496         // a note, and we'll cicle back to revisit it after we've
2497         // hopefully eliminated the second retain, which may allow us to
2498         // eliminate the first retain too.
2499         // Theoretically we could implement removal of nested retain+release
2500         // pairs by making PtrState hold a stack of states, but this is
2501         // simple and avoids adding overhead for the non-nested case.
2502         if (S.GetSeq() == S_Retain)
2503           NestingDetected = true;
2504 
2505         S.SetSeq(S_Retain);
2506         S.RRI.clear();
2507         S.RRI.IsRetainBlock = Class == IC_RetainBlock;
2508         // Don't check S.IsKnownIncremented() here because it's not
2509         // sufficient.
2510         S.RRI.KnownSafe = S.IsKnownNested();
2511         S.RRI.Calls.insert(Inst);
2512       }
2513 
2514       S.SetAtLeastOneRefCount();
2515       S.IncrementRefCount();
2516       S.IncrementNestCount();
2517       continue;
2518     }
2519     case IC_Release: {
2520       Arg = GetObjCArg(Inst);
2521 
2522       PtrState &S = MyStates.getPtrTopDownState(Arg);
2523       S.DecrementRefCount();
2524       S.DecrementNestCount();
2525 
2526       switch (S.GetSeq()) {
2527       case S_Retain:
2528       case S_CanRelease:
2529         S.RRI.ReverseInsertPts.clear();
2530         // FALL THROUGH
2531       case S_Use:
2532         S.RRI.ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2533         S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2534         Releases[Inst] = S.RRI;
2535         S.ClearSequenceProgress();
2536         break;
2537       case S_None:
2538         break;
2539       case S_Stop:
2540       case S_Release:
2541       case S_MovableRelease:
2542         llvm_unreachable("top-down pointer in release state!");
2543       }
2544       break;
2545     }
2546     case IC_AutoreleasepoolPop:
2547       // Conservatively, clear MyStates for all known pointers.
2548       MyStates.clearTopDownPointers();
2549       continue;
2550     case IC_AutoreleasepoolPush:
2551     case IC_None:
2552       // These are irrelevant.
2553       continue;
2554     default:
2555       break;
2556     }
2557 
2558     // Consider any other possible effects of this instruction on each
2559     // pointer being tracked.
2560     for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
2561          ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
2562       const Value *Ptr = MI->first;
2563       if (Ptr == Arg)
2564         continue; // Handled above.
2565       PtrState &S = MI->second;
2566       Sequence Seq = S.GetSeq();
2567 
2568       // Check for possible releases.
2569       if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2570         S.DecrementRefCount();
2571         switch (Seq) {
2572         case S_Retain:
2573           S.SetSeq(S_CanRelease);
2574           assert(S.RRI.ReverseInsertPts.empty());
2575           S.RRI.ReverseInsertPts.insert(Inst);
2576 
2577           // One call can't cause a transition from S_Retain to S_CanRelease
2578           // and S_CanRelease to S_Use. If we've made the first transition,
2579           // we're done.
2580           continue;
2581         case S_Use:
2582         case S_CanRelease:
2583         case S_None:
2584           break;
2585         case S_Stop:
2586         case S_Release:
2587         case S_MovableRelease:
2588           llvm_unreachable("top-down pointer in release state!");
2589         }
2590       }
2591 
2592       // Check for possible direct uses.
2593       switch (Seq) {
2594       case S_CanRelease:
2595         if (CanUse(Inst, Ptr, PA, Class))
2596           S.SetSeq(S_Use);
2597         break;
2598       case S_Retain:
2599         // An objc_retainBlock call may be responsible for copying the block
2600         // data from the stack to the heap. Model this by moving it straight
2601         // from S_Retain to S_Use.
2602         if (S.RRI.IsRetainBlock &&
2603             CanUse(Inst, Ptr, PA, Class)) {
2604           assert(S.RRI.ReverseInsertPts.empty());
2605           S.RRI.ReverseInsertPts.insert(Inst);
2606           S.SetSeq(S_Use);
2607         }
2608         break;
2609       case S_Use:
2610       case S_None:
2611         break;
2612       case S_Stop:
2613       case S_Release:
2614       case S_MovableRelease:
2615         llvm_unreachable("top-down pointer in release state!");
2616       }
2617     }
2618   }
2619 
2620   CheckForCFGHazards(BB, BBStates, MyStates);
2621   return NestingDetected;
2622 }
2623 
2624 // Visit - Visit the function both top-down and bottom-up.
2625 bool
Visit(Function & F,DenseMap<const BasicBlock *,BBState> & BBStates,MapVector<Value *,RRInfo> & Retains,DenseMap<Value *,RRInfo> & Releases)2626 ObjCARCOpt::Visit(Function &F,
2627                   DenseMap<const BasicBlock *, BBState> &BBStates,
2628                   MapVector<Value *, RRInfo> &Retains,
2629                   DenseMap<Value *, RRInfo> &Releases) {
2630   // Use reverse-postorder on the reverse CFG for bottom-up, because we
2631   // magically know that loops will be well behaved, i.e. they won't repeatedly
2632   // call retain on a single pointer without doing a release. We can't use
2633   // ReversePostOrderTraversal here because we want to walk up from each
2634   // function exit point.
2635   SmallPtrSet<BasicBlock *, 16> Visited;
2636   SmallVector<std::pair<BasicBlock *, pred_iterator>, 16> Stack;
2637   SmallVector<BasicBlock *, 16> Order;
2638   for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2639     BasicBlock *BB = I;
2640     if (BB->getTerminator()->getNumSuccessors() == 0)
2641       Stack.push_back(std::make_pair(BB, pred_begin(BB)));
2642   }
2643   while (!Stack.empty()) {
2644     pred_iterator End = pred_end(Stack.back().first);
2645     while (Stack.back().second != End) {
2646       BasicBlock *BB = *Stack.back().second++;
2647       if (Visited.insert(BB))
2648         Stack.push_back(std::make_pair(BB, pred_begin(BB)));
2649     }
2650     Order.push_back(Stack.pop_back_val().first);
2651   }
2652   bool BottomUpNestingDetected = false;
2653   for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2654          Order.rbegin(), E = Order.rend(); I != E; ++I) {
2655     BasicBlock *BB = *I;
2656     BottomUpNestingDetected |= VisitBottomUp(BB, BBStates, Retains);
2657   }
2658 
2659   // Use regular reverse-postorder for top-down.
2660   bool TopDownNestingDetected = false;
2661   typedef ReversePostOrderTraversal<Function *> RPOTType;
2662   RPOTType RPOT(&F);
2663   for (RPOTType::rpo_iterator I = RPOT.begin(), E = RPOT.end(); I != E; ++I) {
2664     BasicBlock *BB = *I;
2665     TopDownNestingDetected |= VisitTopDown(BB, BBStates, Releases);
2666   }
2667 
2668   return TopDownNestingDetected && BottomUpNestingDetected;
2669 }
2670 
2671 /// MoveCalls - Move the calls in RetainsToMove and ReleasesToMove.
MoveCalls(Value * Arg,RRInfo & RetainsToMove,RRInfo & ReleasesToMove,MapVector<Value *,RRInfo> & Retains,DenseMap<Value *,RRInfo> & Releases,SmallVectorImpl<Instruction * > & DeadInsts,Module * M)2672 void ObjCARCOpt::MoveCalls(Value *Arg,
2673                            RRInfo &RetainsToMove,
2674                            RRInfo &ReleasesToMove,
2675                            MapVector<Value *, RRInfo> &Retains,
2676                            DenseMap<Value *, RRInfo> &Releases,
2677                            SmallVectorImpl<Instruction *> &DeadInsts,
2678                            Module *M) {
2679   Type *ArgTy = Arg->getType();
2680   Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
2681 
2682   // Insert the new retain and release calls.
2683   for (SmallPtrSet<Instruction *, 2>::const_iterator
2684        PI = ReleasesToMove.ReverseInsertPts.begin(),
2685        PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2686     Instruction *InsertPt = *PI;
2687     Value *MyArg = ArgTy == ParamTy ? Arg :
2688                    new BitCastInst(Arg, ParamTy, "", InsertPt);
2689     CallInst *Call =
2690       CallInst::Create(RetainsToMove.IsRetainBlock ?
2691                          getRetainBlockCallee(M) : getRetainCallee(M),
2692                        MyArg, "", InsertPt);
2693     Call->setDoesNotThrow();
2694     if (!RetainsToMove.IsRetainBlock)
2695       Call->setTailCall();
2696   }
2697   for (SmallPtrSet<Instruction *, 2>::const_iterator
2698        PI = RetainsToMove.ReverseInsertPts.begin(),
2699        PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2700     Instruction *LastUse = *PI;
2701     Instruction *InsertPts[] = { 0, 0, 0 };
2702     if (InvokeInst *II = dyn_cast<InvokeInst>(LastUse)) {
2703       // We can't insert code immediately after an invoke instruction, so
2704       // insert code at the beginning of both successor blocks instead.
2705       // The invoke's return value isn't available in the unwind block,
2706       // but our releases will never depend on it, because they must be
2707       // paired with retains from before the invoke.
2708       InsertPts[0] = II->getNormalDest()->getFirstInsertionPt();
2709       InsertPts[1] = II->getUnwindDest()->getFirstInsertionPt();
2710     } else {
2711       // Insert code immediately after the last use.
2712       InsertPts[0] = llvm::next(BasicBlock::iterator(LastUse));
2713     }
2714 
2715     for (Instruction **I = InsertPts; *I; ++I) {
2716       Instruction *InsertPt = *I;
2717       Value *MyArg = ArgTy == ParamTy ? Arg :
2718                      new BitCastInst(Arg, ParamTy, "", InsertPt);
2719       CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
2720                                         "", InsertPt);
2721       // Attach a clang.imprecise_release metadata tag, if appropriate.
2722       if (MDNode *M = ReleasesToMove.ReleaseMetadata)
2723         Call->setMetadata(ImpreciseReleaseMDKind, M);
2724       Call->setDoesNotThrow();
2725       if (ReleasesToMove.IsTailCallRelease)
2726         Call->setTailCall();
2727     }
2728   }
2729 
2730   // Delete the original retain and release calls.
2731   for (SmallPtrSet<Instruction *, 2>::const_iterator
2732        AI = RetainsToMove.Calls.begin(),
2733        AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
2734     Instruction *OrigRetain = *AI;
2735     Retains.blot(OrigRetain);
2736     DeadInsts.push_back(OrigRetain);
2737   }
2738   for (SmallPtrSet<Instruction *, 2>::const_iterator
2739        AI = ReleasesToMove.Calls.begin(),
2740        AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
2741     Instruction *OrigRelease = *AI;
2742     Releases.erase(OrigRelease);
2743     DeadInsts.push_back(OrigRelease);
2744   }
2745 }
2746 
2747 bool
PerformCodePlacement(DenseMap<const BasicBlock *,BBState> & BBStates,MapVector<Value *,RRInfo> & Retains,DenseMap<Value *,RRInfo> & Releases,Module * M)2748 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
2749                                    &BBStates,
2750                                  MapVector<Value *, RRInfo> &Retains,
2751                                  DenseMap<Value *, RRInfo> &Releases,
2752                                  Module *M) {
2753   bool AnyPairsCompletelyEliminated = false;
2754   RRInfo RetainsToMove;
2755   RRInfo ReleasesToMove;
2756   SmallVector<Instruction *, 4> NewRetains;
2757   SmallVector<Instruction *, 4> NewReleases;
2758   SmallVector<Instruction *, 8> DeadInsts;
2759 
2760   for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
2761        E = Retains.end(); I != E; ++I) {
2762     Value *V = I->first;
2763     if (!V) continue; // blotted
2764 
2765     Instruction *Retain = cast<Instruction>(V);
2766     Value *Arg = GetObjCArg(Retain);
2767 
2768     // If the object being released is in static storage, we know it's
2769     // not being managed by ObjC reference counting, so we can delete pairs
2770     // regardless of what possible decrements or uses lie between them.
2771     bool KnownSafe = isa<Constant>(Arg);
2772 
2773     // Same for stack storage, unless this is an objc_retainBlock call,
2774     // which is responsible for copying the block data from the stack to
2775     // the heap.
2776     if (!I->second.IsRetainBlock && isa<AllocaInst>(Arg))
2777       KnownSafe = true;
2778 
2779     // A constant pointer can't be pointing to an object on the heap. It may
2780     // be reference-counted, but it won't be deleted.
2781     if (const LoadInst *LI = dyn_cast<LoadInst>(Arg))
2782       if (const GlobalVariable *GV =
2783             dyn_cast<GlobalVariable>(
2784               StripPointerCastsAndObjCCalls(LI->getPointerOperand())))
2785         if (GV->isConstant())
2786           KnownSafe = true;
2787 
2788     // If a pair happens in a region where it is known that the reference count
2789     // is already incremented, we can similarly ignore possible decrements.
2790     bool KnownSafeTD = true, KnownSafeBU = true;
2791 
2792     // Connect the dots between the top-down-collected RetainsToMove and
2793     // bottom-up-collected ReleasesToMove to form sets of related calls.
2794     // This is an iterative process so that we connect multiple releases
2795     // to multiple retains if needed.
2796     unsigned OldDelta = 0;
2797     unsigned NewDelta = 0;
2798     unsigned OldCount = 0;
2799     unsigned NewCount = 0;
2800     bool FirstRelease = true;
2801     bool FirstRetain = true;
2802     NewRetains.push_back(Retain);
2803     for (;;) {
2804       for (SmallVectorImpl<Instruction *>::const_iterator
2805            NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
2806         Instruction *NewRetain = *NI;
2807         MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
2808         assert(It != Retains.end());
2809         const RRInfo &NewRetainRRI = It->second;
2810         KnownSafeTD &= NewRetainRRI.KnownSafe;
2811         for (SmallPtrSet<Instruction *, 2>::const_iterator
2812              LI = NewRetainRRI.Calls.begin(),
2813              LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
2814           Instruction *NewRetainRelease = *LI;
2815           DenseMap<Value *, RRInfo>::const_iterator Jt =
2816             Releases.find(NewRetainRelease);
2817           if (Jt == Releases.end())
2818             goto next_retain;
2819           const RRInfo &NewRetainReleaseRRI = Jt->second;
2820           assert(NewRetainReleaseRRI.Calls.count(NewRetain));
2821           if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
2822             OldDelta -=
2823               BBStates[NewRetainRelease->getParent()].GetAllPathCount();
2824 
2825             // Merge the ReleaseMetadata and IsTailCallRelease values.
2826             if (FirstRelease) {
2827               ReleasesToMove.ReleaseMetadata =
2828                 NewRetainReleaseRRI.ReleaseMetadata;
2829               ReleasesToMove.IsTailCallRelease =
2830                 NewRetainReleaseRRI.IsTailCallRelease;
2831               FirstRelease = false;
2832             } else {
2833               if (ReleasesToMove.ReleaseMetadata !=
2834                     NewRetainReleaseRRI.ReleaseMetadata)
2835                 ReleasesToMove.ReleaseMetadata = 0;
2836               if (ReleasesToMove.IsTailCallRelease !=
2837                     NewRetainReleaseRRI.IsTailCallRelease)
2838                 ReleasesToMove.IsTailCallRelease = false;
2839             }
2840 
2841             // Collect the optimal insertion points.
2842             if (!KnownSafe)
2843               for (SmallPtrSet<Instruction *, 2>::const_iterator
2844                    RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
2845                    RE = NewRetainReleaseRRI.ReverseInsertPts.end();
2846                    RI != RE; ++RI) {
2847                 Instruction *RIP = *RI;
2848                 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
2849                   NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
2850               }
2851             NewReleases.push_back(NewRetainRelease);
2852           }
2853         }
2854       }
2855       NewRetains.clear();
2856       if (NewReleases.empty()) break;
2857 
2858       // Back the other way.
2859       for (SmallVectorImpl<Instruction *>::const_iterator
2860            NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
2861         Instruction *NewRelease = *NI;
2862         DenseMap<Value *, RRInfo>::const_iterator It =
2863           Releases.find(NewRelease);
2864         assert(It != Releases.end());
2865         const RRInfo &NewReleaseRRI = It->second;
2866         KnownSafeBU &= NewReleaseRRI.KnownSafe;
2867         for (SmallPtrSet<Instruction *, 2>::const_iterator
2868              LI = NewReleaseRRI.Calls.begin(),
2869              LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
2870           Instruction *NewReleaseRetain = *LI;
2871           MapVector<Value *, RRInfo>::const_iterator Jt =
2872             Retains.find(NewReleaseRetain);
2873           if (Jt == Retains.end())
2874             goto next_retain;
2875           const RRInfo &NewReleaseRetainRRI = Jt->second;
2876           assert(NewReleaseRetainRRI.Calls.count(NewRelease));
2877           if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
2878             unsigned PathCount =
2879               BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
2880             OldDelta += PathCount;
2881             OldCount += PathCount;
2882 
2883             // Merge the IsRetainBlock values.
2884             if (FirstRetain) {
2885               RetainsToMove.IsRetainBlock = NewReleaseRetainRRI.IsRetainBlock;
2886               FirstRetain = false;
2887             } else if (ReleasesToMove.IsRetainBlock !=
2888                        NewReleaseRetainRRI.IsRetainBlock)
2889               // It's not possible to merge the sequences if one uses
2890               // objc_retain and the other uses objc_retainBlock.
2891               goto next_retain;
2892 
2893             // Collect the optimal insertion points.
2894             if (!KnownSafe)
2895               for (SmallPtrSet<Instruction *, 2>::const_iterator
2896                    RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
2897                    RE = NewReleaseRetainRRI.ReverseInsertPts.end();
2898                    RI != RE; ++RI) {
2899                 Instruction *RIP = *RI;
2900                 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
2901                   PathCount = BBStates[RIP->getParent()].GetAllPathCount();
2902                   NewDelta += PathCount;
2903                   NewCount += PathCount;
2904                 }
2905               }
2906             NewRetains.push_back(NewReleaseRetain);
2907           }
2908         }
2909       }
2910       NewReleases.clear();
2911       if (NewRetains.empty()) break;
2912     }
2913 
2914     // If the pointer is known incremented or nested, we can safely delete the
2915     // pair regardless of what's between them.
2916     if (KnownSafeTD || KnownSafeBU) {
2917       RetainsToMove.ReverseInsertPts.clear();
2918       ReleasesToMove.ReverseInsertPts.clear();
2919       NewCount = 0;
2920     } else {
2921       // Determine whether the new insertion points we computed preserve the
2922       // balance of retain and release calls through the program.
2923       // TODO: If the fully aggressive solution isn't valid, try to find a
2924       // less aggressive solution which is.
2925       if (NewDelta != 0)
2926         goto next_retain;
2927     }
2928 
2929     // Determine whether the original call points are balanced in the retain and
2930     // release calls through the program. If not, conservatively don't touch
2931     // them.
2932     // TODO: It's theoretically possible to do code motion in this case, as
2933     // long as the existing imbalances are maintained.
2934     if (OldDelta != 0)
2935       goto next_retain;
2936 
2937     // Ok, everything checks out and we're all set. Let's move some code!
2938     Changed = true;
2939     AnyPairsCompletelyEliminated = NewCount == 0;
2940     NumRRs += OldCount - NewCount;
2941     MoveCalls(Arg, RetainsToMove, ReleasesToMove,
2942               Retains, Releases, DeadInsts, M);
2943 
2944   next_retain:
2945     NewReleases.clear();
2946     NewRetains.clear();
2947     RetainsToMove.clear();
2948     ReleasesToMove.clear();
2949   }
2950 
2951   // Now that we're done moving everything, we can delete the newly dead
2952   // instructions, as we no longer need them as insert points.
2953   while (!DeadInsts.empty())
2954     EraseInstruction(DeadInsts.pop_back_val());
2955 
2956   return AnyPairsCompletelyEliminated;
2957 }
2958 
2959 /// OptimizeWeakCalls - Weak pointer optimizations.
OptimizeWeakCalls(Function & F)2960 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
2961   // First, do memdep-style RLE and S2L optimizations. We can't use memdep
2962   // itself because it uses AliasAnalysis and we need to do provenance
2963   // queries instead.
2964   for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2965     Instruction *Inst = &*I++;
2966     InstructionClass Class = GetBasicInstructionClass(Inst);
2967     if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
2968       continue;
2969 
2970     // Delete objc_loadWeak calls with no users.
2971     if (Class == IC_LoadWeak && Inst->use_empty()) {
2972       Inst->eraseFromParent();
2973       continue;
2974     }
2975 
2976     // TODO: For now, just look for an earlier available version of this value
2977     // within the same block. Theoretically, we could do memdep-style non-local
2978     // analysis too, but that would want caching. A better approach would be to
2979     // use the technique that EarlyCSE uses.
2980     inst_iterator Current = llvm::prior(I);
2981     BasicBlock *CurrentBB = Current.getBasicBlockIterator();
2982     for (BasicBlock::iterator B = CurrentBB->begin(),
2983                               J = Current.getInstructionIterator();
2984          J != B; --J) {
2985       Instruction *EarlierInst = &*llvm::prior(J);
2986       InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
2987       switch (EarlierClass) {
2988       case IC_LoadWeak:
2989       case IC_LoadWeakRetained: {
2990         // If this is loading from the same pointer, replace this load's value
2991         // with that one.
2992         CallInst *Call = cast<CallInst>(Inst);
2993         CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2994         Value *Arg = Call->getArgOperand(0);
2995         Value *EarlierArg = EarlierCall->getArgOperand(0);
2996         switch (PA.getAA()->alias(Arg, EarlierArg)) {
2997         case AliasAnalysis::MustAlias:
2998           Changed = true;
2999           // If the load has a builtin retain, insert a plain retain for it.
3000           if (Class == IC_LoadWeakRetained) {
3001             CallInst *CI =
3002               CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
3003                                "", Call);
3004             CI->setTailCall();
3005           }
3006           // Zap the fully redundant load.
3007           Call->replaceAllUsesWith(EarlierCall);
3008           Call->eraseFromParent();
3009           goto clobbered;
3010         case AliasAnalysis::MayAlias:
3011         case AliasAnalysis::PartialAlias:
3012           goto clobbered;
3013         case AliasAnalysis::NoAlias:
3014           break;
3015         }
3016         break;
3017       }
3018       case IC_StoreWeak:
3019       case IC_InitWeak: {
3020         // If this is storing to the same pointer and has the same size etc.
3021         // replace this load's value with the stored value.
3022         CallInst *Call = cast<CallInst>(Inst);
3023         CallInst *EarlierCall = cast<CallInst>(EarlierInst);
3024         Value *Arg = Call->getArgOperand(0);
3025         Value *EarlierArg = EarlierCall->getArgOperand(0);
3026         switch (PA.getAA()->alias(Arg, EarlierArg)) {
3027         case AliasAnalysis::MustAlias:
3028           Changed = true;
3029           // If the load has a builtin retain, insert a plain retain for it.
3030           if (Class == IC_LoadWeakRetained) {
3031             CallInst *CI =
3032               CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
3033                                "", Call);
3034             CI->setTailCall();
3035           }
3036           // Zap the fully redundant load.
3037           Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
3038           Call->eraseFromParent();
3039           goto clobbered;
3040         case AliasAnalysis::MayAlias:
3041         case AliasAnalysis::PartialAlias:
3042           goto clobbered;
3043         case AliasAnalysis::NoAlias:
3044           break;
3045         }
3046         break;
3047       }
3048       case IC_MoveWeak:
3049       case IC_CopyWeak:
3050         // TOOD: Grab the copied value.
3051         goto clobbered;
3052       case IC_AutoreleasepoolPush:
3053       case IC_None:
3054       case IC_User:
3055         // Weak pointers are only modified through the weak entry points
3056         // (and arbitrary calls, which could call the weak entry points).
3057         break;
3058       default:
3059         // Anything else could modify the weak pointer.
3060         goto clobbered;
3061       }
3062     }
3063   clobbered:;
3064   }
3065 
3066   // Then, for each destroyWeak with an alloca operand, check to see if
3067   // the alloca and all its users can be zapped.
3068   for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3069     Instruction *Inst = &*I++;
3070     InstructionClass Class = GetBasicInstructionClass(Inst);
3071     if (Class != IC_DestroyWeak)
3072       continue;
3073 
3074     CallInst *Call = cast<CallInst>(Inst);
3075     Value *Arg = Call->getArgOperand(0);
3076     if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
3077       for (Value::use_iterator UI = Alloca->use_begin(),
3078            UE = Alloca->use_end(); UI != UE; ++UI) {
3079         Instruction *UserInst = cast<Instruction>(*UI);
3080         switch (GetBasicInstructionClass(UserInst)) {
3081         case IC_InitWeak:
3082         case IC_StoreWeak:
3083         case IC_DestroyWeak:
3084           continue;
3085         default:
3086           goto done;
3087         }
3088       }
3089       Changed = true;
3090       for (Value::use_iterator UI = Alloca->use_begin(),
3091            UE = Alloca->use_end(); UI != UE; ) {
3092         CallInst *UserInst = cast<CallInst>(*UI++);
3093         if (!UserInst->use_empty())
3094           UserInst->replaceAllUsesWith(UserInst->getOperand(1));
3095         UserInst->eraseFromParent();
3096       }
3097       Alloca->eraseFromParent();
3098     done:;
3099     }
3100   }
3101 }
3102 
3103 /// OptimizeSequences - Identify program paths which execute sequences of
3104 /// retains and releases which can be eliminated.
OptimizeSequences(Function & F)3105 bool ObjCARCOpt::OptimizeSequences(Function &F) {
3106   /// Releases, Retains - These are used to store the results of the main flow
3107   /// analysis. These use Value* as the key instead of Instruction* so that the
3108   /// map stays valid when we get around to rewriting code and calls get
3109   /// replaced by arguments.
3110   DenseMap<Value *, RRInfo> Releases;
3111   MapVector<Value *, RRInfo> Retains;
3112 
3113   /// BBStates, This is used during the traversal of the function to track the
3114   /// states for each identified object at each block.
3115   DenseMap<const BasicBlock *, BBState> BBStates;
3116 
3117   // Analyze the CFG of the function, and all instructions.
3118   bool NestingDetected = Visit(F, BBStates, Retains, Releases);
3119 
3120   // Transform.
3121   return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
3122          NestingDetected;
3123 }
3124 
3125 /// OptimizeReturns - Look for this pattern:
3126 ///
3127 ///    %call = call i8* @something(...)
3128 ///    %2 = call i8* @objc_retain(i8* %call)
3129 ///    %3 = call i8* @objc_autorelease(i8* %2)
3130 ///    ret i8* %3
3131 ///
3132 /// And delete the retain and autorelease.
3133 ///
3134 /// Otherwise if it's just this:
3135 ///
3136 ///    %3 = call i8* @objc_autorelease(i8* %2)
3137 ///    ret i8* %3
3138 ///
3139 /// convert the autorelease to autoreleaseRV.
OptimizeReturns(Function & F)3140 void ObjCARCOpt::OptimizeReturns(Function &F) {
3141   if (!F.getReturnType()->isPointerTy())
3142     return;
3143 
3144   SmallPtrSet<Instruction *, 4> DependingInstructions;
3145   SmallPtrSet<const BasicBlock *, 4> Visited;
3146   for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
3147     BasicBlock *BB = FI;
3148     ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
3149     if (!Ret) continue;
3150 
3151     const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
3152     FindDependencies(NeedsPositiveRetainCount, Arg,
3153                      BB, Ret, DependingInstructions, Visited, PA);
3154     if (DependingInstructions.size() != 1)
3155       goto next_block;
3156 
3157     {
3158       CallInst *Autorelease =
3159         dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3160       if (!Autorelease)
3161         goto next_block;
3162       InstructionClass AutoreleaseClass =
3163         GetBasicInstructionClass(Autorelease);
3164       if (!IsAutorelease(AutoreleaseClass))
3165         goto next_block;
3166       if (GetObjCArg(Autorelease) != Arg)
3167         goto next_block;
3168 
3169       DependingInstructions.clear();
3170       Visited.clear();
3171 
3172       // Check that there is nothing that can affect the reference
3173       // count between the autorelease and the retain.
3174       FindDependencies(CanChangeRetainCount, Arg,
3175                        BB, Autorelease, DependingInstructions, Visited, PA);
3176       if (DependingInstructions.size() != 1)
3177         goto next_block;
3178 
3179       {
3180         CallInst *Retain =
3181           dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3182 
3183         // Check that we found a retain with the same argument.
3184         if (!Retain ||
3185             !IsRetain(GetBasicInstructionClass(Retain)) ||
3186             GetObjCArg(Retain) != Arg)
3187           goto next_block;
3188 
3189         DependingInstructions.clear();
3190         Visited.clear();
3191 
3192         // Convert the autorelease to an autoreleaseRV, since it's
3193         // returning the value.
3194         if (AutoreleaseClass == IC_Autorelease) {
3195           Autorelease->setCalledFunction(getAutoreleaseRVCallee(F.getParent()));
3196           AutoreleaseClass = IC_AutoreleaseRV;
3197         }
3198 
3199         // Check that there is nothing that can affect the reference
3200         // count between the retain and the call.
3201         // Note that Retain need not be in BB.
3202         FindDependencies(CanChangeRetainCount, Arg, Retain->getParent(), Retain,
3203                          DependingInstructions, Visited, PA);
3204         if (DependingInstructions.size() != 1)
3205           goto next_block;
3206 
3207         {
3208           CallInst *Call =
3209             dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3210 
3211           // Check that the pointer is the return value of the call.
3212           if (!Call || Arg != Call)
3213             goto next_block;
3214 
3215           // Check that the call is a regular call.
3216           InstructionClass Class = GetBasicInstructionClass(Call);
3217           if (Class != IC_CallOrUser && Class != IC_Call)
3218             goto next_block;
3219 
3220           // If so, we can zap the retain and autorelease.
3221           Changed = true;
3222           ++NumRets;
3223           EraseInstruction(Retain);
3224           EraseInstruction(Autorelease);
3225         }
3226       }
3227     }
3228 
3229   next_block:
3230     DependingInstructions.clear();
3231     Visited.clear();
3232   }
3233 }
3234 
doInitialization(Module & M)3235 bool ObjCARCOpt::doInitialization(Module &M) {
3236   if (!EnableARCOpts)
3237     return false;
3238 
3239   Run = ModuleHasARC(M);
3240   if (!Run)
3241     return false;
3242 
3243   // Identify the imprecise release metadata kind.
3244   ImpreciseReleaseMDKind =
3245     M.getContext().getMDKindID("clang.imprecise_release");
3246 
3247   // Intuitively, objc_retain and others are nocapture, however in practice
3248   // they are not, because they return their argument value. And objc_release
3249   // calls finalizers.
3250 
3251   // These are initialized lazily.
3252   RetainRVCallee = 0;
3253   AutoreleaseRVCallee = 0;
3254   ReleaseCallee = 0;
3255   RetainCallee = 0;
3256   RetainBlockCallee = 0;
3257   AutoreleaseCallee = 0;
3258 
3259   return false;
3260 }
3261 
runOnFunction(Function & F)3262 bool ObjCARCOpt::runOnFunction(Function &F) {
3263   if (!EnableARCOpts)
3264     return false;
3265 
3266   // If nothing in the Module uses ARC, don't do anything.
3267   if (!Run)
3268     return false;
3269 
3270   Changed = false;
3271 
3272   PA.setAA(&getAnalysis<AliasAnalysis>());
3273 
3274   // This pass performs several distinct transformations. As a compile-time aid
3275   // when compiling code that isn't ObjC, skip these if the relevant ObjC
3276   // library functions aren't declared.
3277 
3278   // Preliminary optimizations. This also computs UsedInThisFunction.
3279   OptimizeIndividualCalls(F);
3280 
3281   // Optimizations for weak pointers.
3282   if (UsedInThisFunction & ((1 << IC_LoadWeak) |
3283                             (1 << IC_LoadWeakRetained) |
3284                             (1 << IC_StoreWeak) |
3285                             (1 << IC_InitWeak) |
3286                             (1 << IC_CopyWeak) |
3287                             (1 << IC_MoveWeak) |
3288                             (1 << IC_DestroyWeak)))
3289     OptimizeWeakCalls(F);
3290 
3291   // Optimizations for retain+release pairs.
3292   if (UsedInThisFunction & ((1 << IC_Retain) |
3293                             (1 << IC_RetainRV) |
3294                             (1 << IC_RetainBlock)))
3295     if (UsedInThisFunction & (1 << IC_Release))
3296       // Run OptimizeSequences until it either stops making changes or
3297       // no retain+release pair nesting is detected.
3298       while (OptimizeSequences(F)) {}
3299 
3300   // Optimizations if objc_autorelease is used.
3301   if (UsedInThisFunction &
3302       ((1 << IC_Autorelease) | (1 << IC_AutoreleaseRV)))
3303     OptimizeReturns(F);
3304 
3305   return Changed;
3306 }
3307 
releaseMemory()3308 void ObjCARCOpt::releaseMemory() {
3309   PA.clear();
3310 }
3311 
3312 //===----------------------------------------------------------------------===//
3313 // ARC contraction.
3314 //===----------------------------------------------------------------------===//
3315 
3316 // TODO: ObjCARCContract could insert PHI nodes when uses aren't
3317 // dominated by single calls.
3318 
3319 #include "llvm/Operator.h"
3320 #include "llvm/InlineAsm.h"
3321 #include "llvm/Analysis/Dominators.h"
3322 
3323 STATISTIC(NumStoreStrongs, "Number objc_storeStrong calls formed");
3324 
3325 namespace {
3326   /// ObjCARCContract - Late ARC optimizations.  These change the IR in a way
3327   /// that makes it difficult to be analyzed by ObjCARCOpt, so it's run late.
3328   class ObjCARCContract : public FunctionPass {
3329     bool Changed;
3330     AliasAnalysis *AA;
3331     DominatorTree *DT;
3332     ProvenanceAnalysis PA;
3333 
3334     /// Run - A flag indicating whether this optimization pass should run.
3335     bool Run;
3336 
3337     /// StoreStrongCallee, etc. - Declarations for ObjC runtime
3338     /// functions, for use in creating calls to them. These are initialized
3339     /// lazily to avoid cluttering up the Module with unused declarations.
3340     Constant *StoreStrongCallee,
3341              *RetainAutoreleaseCallee, *RetainAutoreleaseRVCallee;
3342 
3343     /// RetainRVMarker - The inline asm string to insert between calls and
3344     /// RetainRV calls to make the optimization work on targets which need it.
3345     const MDString *RetainRVMarker;
3346 
3347     Constant *getStoreStrongCallee(Module *M);
3348     Constant *getRetainAutoreleaseCallee(Module *M);
3349     Constant *getRetainAutoreleaseRVCallee(Module *M);
3350 
3351     bool ContractAutorelease(Function &F, Instruction *Autorelease,
3352                              InstructionClass Class,
3353                              SmallPtrSet<Instruction *, 4>
3354                                &DependingInstructions,
3355                              SmallPtrSet<const BasicBlock *, 4>
3356                                &Visited);
3357 
3358     void ContractRelease(Instruction *Release,
3359                          inst_iterator &Iter);
3360 
3361     virtual void getAnalysisUsage(AnalysisUsage &AU) const;
3362     virtual bool doInitialization(Module &M);
3363     virtual bool runOnFunction(Function &F);
3364 
3365   public:
3366     static char ID;
ObjCARCContract()3367     ObjCARCContract() : FunctionPass(ID) {
3368       initializeObjCARCContractPass(*PassRegistry::getPassRegistry());
3369     }
3370   };
3371 }
3372 
3373 char ObjCARCContract::ID = 0;
3374 INITIALIZE_PASS_BEGIN(ObjCARCContract,
3375                       "objc-arc-contract", "ObjC ARC contraction", false, false)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)3376 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3377 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3378 INITIALIZE_PASS_END(ObjCARCContract,
3379                     "objc-arc-contract", "ObjC ARC contraction", false, false)
3380 
3381 Pass *llvm::createObjCARCContractPass() {
3382   return new ObjCARCContract();
3383 }
3384 
getAnalysisUsage(AnalysisUsage & AU) const3385 void ObjCARCContract::getAnalysisUsage(AnalysisUsage &AU) const {
3386   AU.addRequired<AliasAnalysis>();
3387   AU.addRequired<DominatorTree>();
3388   AU.setPreservesCFG();
3389 }
3390 
getStoreStrongCallee(Module * M)3391 Constant *ObjCARCContract::getStoreStrongCallee(Module *M) {
3392   if (!StoreStrongCallee) {
3393     LLVMContext &C = M->getContext();
3394     Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3395     Type *I8XX = PointerType::getUnqual(I8X);
3396     std::vector<Type *> Params;
3397     Params.push_back(I8XX);
3398     Params.push_back(I8X);
3399 
3400     AttrListPtr Attributes;
3401     Attributes.addAttr(~0u, Attribute::NoUnwind);
3402     Attributes.addAttr(1, Attribute::NoCapture);
3403 
3404     StoreStrongCallee =
3405       M->getOrInsertFunction(
3406         "objc_storeStrong",
3407         FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
3408         Attributes);
3409   }
3410   return StoreStrongCallee;
3411 }
3412 
getRetainAutoreleaseCallee(Module * M)3413 Constant *ObjCARCContract::getRetainAutoreleaseCallee(Module *M) {
3414   if (!RetainAutoreleaseCallee) {
3415     LLVMContext &C = M->getContext();
3416     Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3417     std::vector<Type *> Params;
3418     Params.push_back(I8X);
3419     FunctionType *FTy =
3420       FunctionType::get(I8X, Params, /*isVarArg=*/false);
3421     AttrListPtr Attributes;
3422     Attributes.addAttr(~0u, Attribute::NoUnwind);
3423     RetainAutoreleaseCallee =
3424       M->getOrInsertFunction("objc_retainAutorelease", FTy, Attributes);
3425   }
3426   return RetainAutoreleaseCallee;
3427 }
3428 
getRetainAutoreleaseRVCallee(Module * M)3429 Constant *ObjCARCContract::getRetainAutoreleaseRVCallee(Module *M) {
3430   if (!RetainAutoreleaseRVCallee) {
3431     LLVMContext &C = M->getContext();
3432     Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3433     std::vector<Type *> Params;
3434     Params.push_back(I8X);
3435     FunctionType *FTy =
3436       FunctionType::get(I8X, Params, /*isVarArg=*/false);
3437     AttrListPtr Attributes;
3438     Attributes.addAttr(~0u, Attribute::NoUnwind);
3439     RetainAutoreleaseRVCallee =
3440       M->getOrInsertFunction("objc_retainAutoreleaseReturnValue", FTy,
3441                              Attributes);
3442   }
3443   return RetainAutoreleaseRVCallee;
3444 }
3445 
3446 /// ContractAutorelease - Merge an autorelease with a retain into a fused
3447 /// call.
3448 bool
ContractAutorelease(Function & F,Instruction * Autorelease,InstructionClass Class,SmallPtrSet<Instruction *,4> & DependingInstructions,SmallPtrSet<const BasicBlock *,4> & Visited)3449 ObjCARCContract::ContractAutorelease(Function &F, Instruction *Autorelease,
3450                                      InstructionClass Class,
3451                                      SmallPtrSet<Instruction *, 4>
3452                                        &DependingInstructions,
3453                                      SmallPtrSet<const BasicBlock *, 4>
3454                                        &Visited) {
3455   const Value *Arg = GetObjCArg(Autorelease);
3456 
3457   // Check that there are no instructions between the retain and the autorelease
3458   // (such as an autorelease_pop) which may change the count.
3459   CallInst *Retain = 0;
3460   if (Class == IC_AutoreleaseRV)
3461     FindDependencies(RetainAutoreleaseRVDep, Arg,
3462                      Autorelease->getParent(), Autorelease,
3463                      DependingInstructions, Visited, PA);
3464   else
3465     FindDependencies(RetainAutoreleaseDep, Arg,
3466                      Autorelease->getParent(), Autorelease,
3467                      DependingInstructions, Visited, PA);
3468 
3469   Visited.clear();
3470   if (DependingInstructions.size() != 1) {
3471     DependingInstructions.clear();
3472     return false;
3473   }
3474 
3475   Retain = dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
3476   DependingInstructions.clear();
3477 
3478   if (!Retain ||
3479       GetBasicInstructionClass(Retain) != IC_Retain ||
3480       GetObjCArg(Retain) != Arg)
3481     return false;
3482 
3483   Changed = true;
3484   ++NumPeeps;
3485 
3486   if (Class == IC_AutoreleaseRV)
3487     Retain->setCalledFunction(getRetainAutoreleaseRVCallee(F.getParent()));
3488   else
3489     Retain->setCalledFunction(getRetainAutoreleaseCallee(F.getParent()));
3490 
3491   EraseInstruction(Autorelease);
3492   return true;
3493 }
3494 
3495 /// ContractRelease - Attempt to merge an objc_release with a store, load, and
3496 /// objc_retain to form an objc_storeStrong. This can be a little tricky because
3497 /// the instructions don't always appear in order, and there may be unrelated
3498 /// intervening instructions.
ContractRelease(Instruction * Release,inst_iterator & Iter)3499 void ObjCARCContract::ContractRelease(Instruction *Release,
3500                                       inst_iterator &Iter) {
3501   LoadInst *Load = dyn_cast<LoadInst>(GetObjCArg(Release));
3502   if (!Load || !Load->isSimple()) return;
3503 
3504   // For now, require everything to be in one basic block.
3505   BasicBlock *BB = Release->getParent();
3506   if (Load->getParent() != BB) return;
3507 
3508   // Walk down to find the store.
3509   BasicBlock::iterator I = Load, End = BB->end();
3510   ++I;
3511   AliasAnalysis::Location Loc = AA->getLocation(Load);
3512   while (I != End &&
3513          (&*I == Release ||
3514           IsRetain(GetBasicInstructionClass(I)) ||
3515           !(AA->getModRefInfo(I, Loc) & AliasAnalysis::Mod)))
3516     ++I;
3517   StoreInst *Store = dyn_cast<StoreInst>(I);
3518   if (!Store || !Store->isSimple()) return;
3519   if (Store->getPointerOperand() != Loc.Ptr) return;
3520 
3521   Value *New = StripPointerCastsAndObjCCalls(Store->getValueOperand());
3522 
3523   // Walk up to find the retain.
3524   I = Store;
3525   BasicBlock::iterator Begin = BB->begin();
3526   while (I != Begin && GetBasicInstructionClass(I) != IC_Retain)
3527     --I;
3528   Instruction *Retain = I;
3529   if (GetBasicInstructionClass(Retain) != IC_Retain) return;
3530   if (GetObjCArg(Retain) != New) return;
3531 
3532   Changed = true;
3533   ++NumStoreStrongs;
3534 
3535   LLVMContext &C = Release->getContext();
3536   Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
3537   Type *I8XX = PointerType::getUnqual(I8X);
3538 
3539   Value *Args[] = { Load->getPointerOperand(), New };
3540   if (Args[0]->getType() != I8XX)
3541     Args[0] = new BitCastInst(Args[0], I8XX, "", Store);
3542   if (Args[1]->getType() != I8X)
3543     Args[1] = new BitCastInst(Args[1], I8X, "", Store);
3544   CallInst *StoreStrong =
3545     CallInst::Create(getStoreStrongCallee(BB->getParent()->getParent()),
3546                      Args, "", Store);
3547   StoreStrong->setDoesNotThrow();
3548   StoreStrong->setDebugLoc(Store->getDebugLoc());
3549 
3550   if (&*Iter == Store) ++Iter;
3551   Store->eraseFromParent();
3552   Release->eraseFromParent();
3553   EraseInstruction(Retain);
3554   if (Load->use_empty())
3555     Load->eraseFromParent();
3556 }
3557 
doInitialization(Module & M)3558 bool ObjCARCContract::doInitialization(Module &M) {
3559   Run = ModuleHasARC(M);
3560   if (!Run)
3561     return false;
3562 
3563   // These are initialized lazily.
3564   StoreStrongCallee = 0;
3565   RetainAutoreleaseCallee = 0;
3566   RetainAutoreleaseRVCallee = 0;
3567 
3568   // Initialize RetainRVMarker.
3569   RetainRVMarker = 0;
3570   if (NamedMDNode *NMD =
3571         M.getNamedMetadata("clang.arc.retainAutoreleasedReturnValueMarker"))
3572     if (NMD->getNumOperands() == 1) {
3573       const MDNode *N = NMD->getOperand(0);
3574       if (N->getNumOperands() == 1)
3575         if (const MDString *S = dyn_cast<MDString>(N->getOperand(0)))
3576           RetainRVMarker = S;
3577     }
3578 
3579   return false;
3580 }
3581 
runOnFunction(Function & F)3582 bool ObjCARCContract::runOnFunction(Function &F) {
3583   if (!EnableARCOpts)
3584     return false;
3585 
3586   // If nothing in the Module uses ARC, don't do anything.
3587   if (!Run)
3588     return false;
3589 
3590   Changed = false;
3591   AA = &getAnalysis<AliasAnalysis>();
3592   DT = &getAnalysis<DominatorTree>();
3593 
3594   PA.setAA(&getAnalysis<AliasAnalysis>());
3595 
3596   // For ObjC library calls which return their argument, replace uses of the
3597   // argument with uses of the call return value, if it dominates the use. This
3598   // reduces register pressure.
3599   SmallPtrSet<Instruction *, 4> DependingInstructions;
3600   SmallPtrSet<const BasicBlock *, 4> Visited;
3601   for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
3602     Instruction *Inst = &*I++;
3603 
3604     // Only these library routines return their argument. In particular,
3605     // objc_retainBlock does not necessarily return its argument.
3606     InstructionClass Class = GetBasicInstructionClass(Inst);
3607     switch (Class) {
3608     case IC_Retain:
3609     case IC_FusedRetainAutorelease:
3610     case IC_FusedRetainAutoreleaseRV:
3611       break;
3612     case IC_Autorelease:
3613     case IC_AutoreleaseRV:
3614       if (ContractAutorelease(F, Inst, Class, DependingInstructions, Visited))
3615         continue;
3616       break;
3617     case IC_RetainRV: {
3618       // If we're compiling for a target which needs a special inline-asm
3619       // marker to do the retainAutoreleasedReturnValue optimization,
3620       // insert it now.
3621       if (!RetainRVMarker)
3622         break;
3623       BasicBlock::iterator BBI = Inst;
3624       --BBI;
3625       while (isNoopInstruction(BBI)) --BBI;
3626       if (&*BBI == GetObjCArg(Inst)) {
3627         InlineAsm *IA =
3628           InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
3629                                            /*isVarArg=*/false),
3630                          RetainRVMarker->getString(),
3631                          /*Constraints=*/"", /*hasSideEffects=*/true);
3632         CallInst::Create(IA, "", Inst);
3633       }
3634       break;
3635     }
3636     case IC_InitWeak: {
3637       // objc_initWeak(p, null) => *p = null
3638       CallInst *CI = cast<CallInst>(Inst);
3639       if (isNullOrUndef(CI->getArgOperand(1))) {
3640         Value *Null =
3641           ConstantPointerNull::get(cast<PointerType>(CI->getType()));
3642         Changed = true;
3643         new StoreInst(Null, CI->getArgOperand(0), CI);
3644         CI->replaceAllUsesWith(Null);
3645         CI->eraseFromParent();
3646       }
3647       continue;
3648     }
3649     case IC_Release:
3650       ContractRelease(Inst, I);
3651       continue;
3652     default:
3653       continue;
3654     }
3655 
3656     // Don't use GetObjCArg because we don't want to look through bitcasts
3657     // and such; to do the replacement, the argument must have type i8*.
3658     const Value *Arg = cast<CallInst>(Inst)->getArgOperand(0);
3659     for (;;) {
3660       // If we're compiling bugpointed code, don't get in trouble.
3661       if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))
3662         break;
3663       // Look through the uses of the pointer.
3664       for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
3665            UI != UE; ) {
3666         Use &U = UI.getUse();
3667         unsigned OperandNo = UI.getOperandNo();
3668         ++UI; // Increment UI now, because we may unlink its element.
3669         if (Instruction *UserInst = dyn_cast<Instruction>(U.getUser()))
3670           if (Inst != UserInst && DT->dominates(Inst, UserInst)) {
3671             Changed = true;
3672             Instruction *Replacement = Inst;
3673             Type *UseTy = U.get()->getType();
3674             if (PHINode *PHI = dyn_cast<PHINode>(UserInst)) {
3675               // For PHI nodes, insert the bitcast in the predecessor block.
3676               unsigned ValNo =
3677                 PHINode::getIncomingValueNumForOperand(OperandNo);
3678               BasicBlock *BB =
3679                 PHI->getIncomingBlock(ValNo);
3680               if (Replacement->getType() != UseTy)
3681                 Replacement = new BitCastInst(Replacement, UseTy, "",
3682                                               &BB->back());
3683               for (unsigned i = 0, e = PHI->getNumIncomingValues();
3684                    i != e; ++i)
3685                 if (PHI->getIncomingBlock(i) == BB) {
3686                   // Keep the UI iterator valid.
3687                   if (&PHI->getOperandUse(
3688                         PHINode::getOperandNumForIncomingValue(i)) ==
3689                         &UI.getUse())
3690                     ++UI;
3691                   PHI->setIncomingValue(i, Replacement);
3692                 }
3693             } else {
3694               if (Replacement->getType() != UseTy)
3695                 Replacement = new BitCastInst(Replacement, UseTy, "", UserInst);
3696               U.set(Replacement);
3697             }
3698           }
3699       }
3700 
3701       // If Arg is a no-op casted pointer, strip one level of casts and
3702       // iterate.
3703       if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg))
3704         Arg = BI->getOperand(0);
3705       else if (isa<GEPOperator>(Arg) &&
3706                cast<GEPOperator>(Arg)->hasAllZeroIndices())
3707         Arg = cast<GEPOperator>(Arg)->getPointerOperand();
3708       else if (isa<GlobalAlias>(Arg) &&
3709                !cast<GlobalAlias>(Arg)->mayBeOverridden())
3710         Arg = cast<GlobalAlias>(Arg)->getAliasee();
3711       else
3712         break;
3713     }
3714   }
3715 
3716   return Changed;
3717 }
3718