1 //===- ObjCARCOpts.cpp - ObjC ARC Optimization ----------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file defines ObjC ARC optimizations. ARC stands for Automatic
11 /// Reference Counting and is a system for managing reference counts for objects
14 /// The optimizations performed include elimination of redundant, partially
15 /// redundant, and inconsequential reference count operations, elimination of
16 /// redundant weak pointer operations, and numerous minor simplifications.
18 /// WARNING: This file knows about certain library functions. It recognizes them
19 /// by name, and hardwires knowledge of their semantics.
21 /// WARNING: This file knows about how certain Objective-C library functions are
22 /// used. Naive LLVM IR transformations which would otherwise be
23 /// behavior-preserving may break these assumptions.
25 //===----------------------------------------------------------------------===//
27 #define DEBUG_TYPE "objc-arc-opts"
29 #include "DependencyAnalysis.h"
30 #include "ObjCARCAliasAnalysis.h"
31 #include "ProvenanceAnalysis.h"
32 #include "llvm/ADT/DenseMap.h"
33 #include "llvm/ADT/STLExtras.h"
34 #include "llvm/ADT/SmallPtrSet.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/IR/IRBuilder.h"
37 #include "llvm/IR/LLVMContext.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/raw_ostream.h"
43 using namespace llvm::objcarc;
45 /// \defgroup MiscUtils Miscellaneous utilities that are not ARC specific.
49 /// \brief An associative container with fast insertion-order (deterministic)
50 /// iteration over its elements. Plus the special blot operation.
51 template<class KeyT, class ValueT>
53 /// Map keys to indices in Vector.
54 typedef DenseMap<KeyT, size_t> MapTy;
57 typedef std::vector<std::pair<KeyT, ValueT> > VectorTy;
62 typedef typename VectorTy::iterator iterator;
63 typedef typename VectorTy::const_iterator const_iterator;
64 iterator begin() { return Vector.begin(); }
65 iterator end() { return Vector.end(); }
66 const_iterator begin() const { return Vector.begin(); }
67 const_iterator end() const { return Vector.end(); }
71 assert(Vector.size() >= Map.size()); // May differ due to blotting.
72 for (typename MapTy::const_iterator I = Map.begin(), E = Map.end();
74 assert(I->second < Vector.size());
75 assert(Vector[I->second].first == I->first);
77 for (typename VectorTy::const_iterator I = Vector.begin(),
78 E = Vector.end(); I != E; ++I)
80 (Map.count(I->first) &&
81 Map[I->first] == size_t(I - Vector.begin())));
85 ValueT &operator[](const KeyT &Arg) {
86 std::pair<typename MapTy::iterator, bool> Pair =
87 Map.insert(std::make_pair(Arg, size_t(0)));
89 size_t Num = Vector.size();
90 Pair.first->second = Num;
91 Vector.push_back(std::make_pair(Arg, ValueT()));
92 return Vector[Num].second;
94 return Vector[Pair.first->second].second;
97 std::pair<iterator, bool>
98 insert(const std::pair<KeyT, ValueT> &InsertPair) {
99 std::pair<typename MapTy::iterator, bool> Pair =
100 Map.insert(std::make_pair(InsertPair.first, size_t(0)));
102 size_t Num = Vector.size();
103 Pair.first->second = Num;
104 Vector.push_back(InsertPair);
105 return std::make_pair(Vector.begin() + Num, true);
107 return std::make_pair(Vector.begin() + Pair.first->second, false);
110 const_iterator find(const KeyT &Key) const {
111 typename MapTy::const_iterator It = Map.find(Key);
112 if (It == Map.end()) return Vector.end();
113 return Vector.begin() + It->second;
116 /// This is similar to erase, but instead of removing the element from the
117 /// vector, it just zeros out the key in the vector. This leaves iterators
118 /// intact, but clients must be prepared for zeroed-out keys when iterating.
119 void blot(const KeyT &Key) {
120 typename MapTy::iterator It = Map.find(Key);
121 if (It == Map.end()) return;
122 Vector[It->second].first = KeyT();
135 /// \defgroup ARCUtilities Utility declarations/definitions specific to ARC.
138 /// \brief This is similar to StripPointerCastsAndObjCCalls but it stops as soon
139 /// as it finds a value with multiple uses.
140 static const Value *FindSingleUseIdentifiedObject(const Value *Arg) {
141 if (Arg->hasOneUse()) {
142 if (const BitCastInst *BC = dyn_cast<BitCastInst>(Arg))
143 return FindSingleUseIdentifiedObject(BC->getOperand(0));
144 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Arg))
145 if (GEP->hasAllZeroIndices())
146 return FindSingleUseIdentifiedObject(GEP->getPointerOperand());
147 if (IsForwarding(GetBasicInstructionClass(Arg)))
148 return FindSingleUseIdentifiedObject(
149 cast<CallInst>(Arg)->getArgOperand(0));
150 if (!IsObjCIdentifiedObject(Arg))
155 // If we found an identifiable object but it has multiple uses, but they are
156 // trivial uses, we can still consider this to be a single-use value.
157 if (IsObjCIdentifiedObject(Arg)) {
158 for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
161 if (!U->use_empty() || StripPointerCastsAndObjCCalls(U) != Arg)
171 /// \brief Test whether the given retainable object pointer escapes.
173 /// This differs from regular escape analysis in that a use as an
174 /// argument to a call is not considered an escape.
176 static bool DoesRetainableObjPtrEscape(const User *Ptr) {
177 DEBUG(dbgs() << "DoesRetainableObjPtrEscape: Target: " << *Ptr << "\n");
179 // Walk the def-use chains.
180 SmallVector<const Value *, 4> Worklist;
181 Worklist.push_back(Ptr);
182 // If Ptr has any operands add them as well.
183 for (User::const_op_iterator I = Ptr->op_begin(), E = Ptr->op_end(); I != E;
185 Worklist.push_back(*I);
188 // Ensure we do not visit any value twice.
189 SmallPtrSet<const Value *, 8> VisitedSet;
192 const Value *V = Worklist.pop_back_val();
194 DEBUG(dbgs() << "Visiting: " << *V << "\n");
196 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
198 const User *UUser = *UI;
200 DEBUG(dbgs() << "User: " << *UUser << "\n");
202 // Special - Use by a call (callee or argument) is not considered
204 switch (GetBasicInstructionClass(UUser)) {
209 case IC_AutoreleaseRV: {
210 DEBUG(dbgs() << "User copies pointer arguments. Pointer Escapes!\n");
211 // These special functions make copies of their pointer arguments.
214 case IC_IntrinsicUser:
215 // Use by the use intrinsic is not an escape.
219 // Use by an instruction which copies the value is an escape if the
220 // result is an escape.
221 if (isa<BitCastInst>(UUser) || isa<GetElementPtrInst>(UUser) ||
222 isa<PHINode>(UUser) || isa<SelectInst>(UUser)) {
224 if (VisitedSet.insert(UUser)) {
225 DEBUG(dbgs() << "User copies value. Ptr escapes if result escapes."
226 " Adding to list.\n");
227 Worklist.push_back(UUser);
229 DEBUG(dbgs() << "Already visited node.\n");
233 // Use by a load is not an escape.
234 if (isa<LoadInst>(UUser))
236 // Use by a store is not an escape if the use is the address.
237 if (const StoreInst *SI = dyn_cast<StoreInst>(UUser))
238 if (V != SI->getValueOperand())
242 // Regular calls and other stuff are not considered escapes.
245 // Otherwise, conservatively assume an escape.
246 DEBUG(dbgs() << "Assuming ptr escapes.\n");
249 } while (!Worklist.empty());
252 DEBUG(dbgs() << "Ptr does not escape.\n");
258 /// \defgroup ARCOpt ARC Optimization.
261 // TODO: On code like this:
264 // stuff_that_cannot_release()
265 // objc_autorelease(%x)
266 // stuff_that_cannot_release()
268 // stuff_that_cannot_release()
269 // objc_autorelease(%x)
271 // The second retain and autorelease can be deleted.
273 // TODO: It should be possible to delete
274 // objc_autoreleasePoolPush and objc_autoreleasePoolPop
275 // pairs if nothing is actually autoreleased between them. Also, autorelease
276 // calls followed by objc_autoreleasePoolPop calls (perhaps in ObjC++ code
277 // after inlining) can be turned into plain release calls.
279 // TODO: Critical-edge splitting. If the optimial insertion point is
280 // a critical edge, the current algorithm has to fail, because it doesn't
281 // know how to split edges. It should be possible to make the optimizer
282 // think in terms of edges, rather than blocks, and then split critical
285 // TODO: OptimizeSequences could generalized to be Interprocedural.
287 // TODO: Recognize that a bunch of other objc runtime calls have
288 // non-escaping arguments and non-releasing arguments, and may be
289 // non-autoreleasing.
291 // TODO: Sink autorelease calls as far as possible. Unfortunately we
292 // usually can't sink them past other calls, which would be the main
293 // case where it would be useful.
295 // TODO: The pointer returned from objc_loadWeakRetained is retained.
297 // TODO: Delete release+retain pairs (rare).
299 STATISTIC(NumNoops, "Number of no-op objc calls eliminated");
300 STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated");
301 STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases");
302 STATISTIC(NumRets, "Number of return value forwarding "
303 "retain+autoreleaes eliminated");
304 STATISTIC(NumRRs, "Number of retain+release paths eliminated");
305 STATISTIC(NumPeeps, "Number of calls peephole-optimized");
310 /// \brief A sequence of states that a pointer may go through in which an
311 /// objc_retain and objc_release are actually needed.
314 S_Retain, ///< objc_retain(x).
315 S_CanRelease, ///< foo(x) -- x could possibly see a ref count decrement.
316 S_Use, ///< any use of x.
317 S_Stop, ///< like S_Release, but code motion is stopped.
318 S_Release, ///< objc_release(x).
319 S_MovableRelease ///< objc_release(x), !clang.imprecise_release.
322 raw_ostream &operator<<(raw_ostream &OS, const Sequence S)
323 LLVM_ATTRIBUTE_UNUSED;
324 raw_ostream &operator<<(raw_ostream &OS, const Sequence S) {
327 return OS << "S_None";
329 return OS << "S_Retain";
331 return OS << "S_CanRelease";
333 return OS << "S_Use";
335 return OS << "S_Release";
336 case S_MovableRelease:
337 return OS << "S_MovableRelease";
339 return OS << "S_Stop";
341 llvm_unreachable("Unknown sequence type.");
345 static Sequence MergeSeqs(Sequence A, Sequence B, bool TopDown) {
349 if (A == S_None || B == S_None)
352 if (A > B) std::swap(A, B);
354 // Choose the side which is further along in the sequence.
355 if ((A == S_Retain || A == S_CanRelease) &&
356 (B == S_CanRelease || B == S_Use))
359 // Choose the side which is further along in the sequence.
360 if ((A == S_Use || A == S_CanRelease) &&
361 (B == S_Use || B == S_Release || B == S_Stop || B == S_MovableRelease))
363 // If both sides are releases, choose the more conservative one.
364 if (A == S_Stop && (B == S_Release || B == S_MovableRelease))
366 if (A == S_Release && B == S_MovableRelease)
374 /// \brief Unidirectional information about either a
375 /// retain-decrement-use-release sequence or release-use-decrement-retain
376 /// reverese sequence.
378 /// After an objc_retain, the reference count of the referenced
379 /// object is known to be positive. Similarly, before an objc_release, the
380 /// reference count of the referenced object is known to be positive. If
381 /// there are retain-release pairs in code regions where the retain count
382 /// is known to be positive, they can be eliminated, regardless of any side
383 /// effects between them.
385 /// Also, a retain+release pair nested within another retain+release
386 /// pair all on the known same pointer value can be eliminated, regardless
387 /// of any intervening side effects.
389 /// KnownSafe is true when either of these conditions is satisfied.
392 /// True of the objc_release calls are all marked with the "tail" keyword.
393 bool IsTailCallRelease;
395 /// If the Calls are objc_release calls and they all have a
396 /// clang.imprecise_release tag, this is the metadata tag.
397 MDNode *ReleaseMetadata;
399 /// For a top-down sequence, the set of objc_retains or
400 /// objc_retainBlocks. For bottom-up, the set of objc_releases.
401 SmallPtrSet<Instruction *, 2> Calls;
403 /// The set of optimal insert positions for moving calls in the opposite
405 SmallPtrSet<Instruction *, 2> ReverseInsertPts;
408 KnownSafe(false), IsTailCallRelease(false), ReleaseMetadata(0) {}
412 bool IsTrackingImpreciseReleases() {
413 return ReleaseMetadata != 0;
418 void RRInfo::clear() {
420 IsTailCallRelease = false;
423 ReverseInsertPts.clear();
427 /// \brief This class summarizes several per-pointer runtime properties which
428 /// are propogated through the flow graph.
430 /// True if the reference count is known to be incremented.
431 bool KnownPositiveRefCount;
433 /// True of we've seen an opportunity for partial RR elimination, such as
434 /// pushing calls into a CFG triangle or into one side of a CFG diamond.
437 /// The current position in the sequence.
441 /// Unidirectional information about the current sequence.
443 /// TODO: Encapsulate this better.
446 PtrState() : KnownPositiveRefCount(false), Partial(false),
449 void SetKnownPositiveRefCount() {
450 KnownPositiveRefCount = true;
453 void ClearKnownPositiveRefCount() {
454 KnownPositiveRefCount = false;
457 bool HasKnownPositiveRefCount() const {
458 return KnownPositiveRefCount;
461 void SetSeq(Sequence NewSeq) {
462 DEBUG(dbgs() << "Old: " << Seq << "; New: " << NewSeq << "\n");
466 Sequence GetSeq() const {
470 void ClearSequenceProgress() {
471 ResetSequenceProgress(S_None);
474 void ResetSequenceProgress(Sequence NewSeq) {
480 void Merge(const PtrState &Other, bool TopDown);
485 PtrState::Merge(const PtrState &Other, bool TopDown) {
486 Seq = MergeSeqs(Seq, Other.Seq, TopDown);
487 KnownPositiveRefCount = KnownPositiveRefCount && Other.KnownPositiveRefCount;
489 // If we're not in a sequence (anymore), drop all associated state.
493 } else if (Partial || Other.Partial) {
494 // If we're doing a merge on a path that's previously seen a partial
495 // merge, conservatively drop the sequence, to avoid doing partial
496 // RR elimination. If the branch predicates for the two merge differ,
497 // mixing them is unsafe.
498 ClearSequenceProgress();
500 // Conservatively merge the ReleaseMetadata information.
501 if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
502 RRI.ReleaseMetadata = 0;
504 RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe;
505 RRI.IsTailCallRelease = RRI.IsTailCallRelease &&
506 Other.RRI.IsTailCallRelease;
507 RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
509 // Merge the insert point sets. If there are any differences,
510 // that makes this a partial merge.
511 Partial = RRI.ReverseInsertPts.size() != Other.RRI.ReverseInsertPts.size();
512 for (SmallPtrSet<Instruction *, 2>::const_iterator
513 I = Other.RRI.ReverseInsertPts.begin(),
514 E = Other.RRI.ReverseInsertPts.end(); I != E; ++I)
515 Partial |= RRI.ReverseInsertPts.insert(*I);
520 /// \brief Per-BasicBlock state.
522 /// The number of unique control paths from the entry which can reach this
524 unsigned TopDownPathCount;
526 /// The number of unique control paths to exits from this block.
527 unsigned BottomUpPathCount;
529 /// A type for PerPtrTopDown and PerPtrBottomUp.
530 typedef MapVector<const Value *, PtrState> MapTy;
532 /// The top-down traversal uses this to record information known about a
533 /// pointer at the bottom of each block.
536 /// The bottom-up traversal uses this to record information known about a
537 /// pointer at the top of each block.
538 MapTy PerPtrBottomUp;
540 /// Effective predecessors of the current block ignoring ignorable edges and
541 /// ignored backedges.
542 SmallVector<BasicBlock *, 2> Preds;
543 /// Effective successors of the current block ignoring ignorable edges and
544 /// ignored backedges.
545 SmallVector<BasicBlock *, 2> Succs;
548 BBState() : TopDownPathCount(0), BottomUpPathCount(0) {}
550 typedef MapTy::iterator ptr_iterator;
551 typedef MapTy::const_iterator ptr_const_iterator;
553 ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); }
554 ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); }
555 ptr_const_iterator top_down_ptr_begin() const {
556 return PerPtrTopDown.begin();
558 ptr_const_iterator top_down_ptr_end() const {
559 return PerPtrTopDown.end();
562 ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); }
563 ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); }
564 ptr_const_iterator bottom_up_ptr_begin() const {
565 return PerPtrBottomUp.begin();
567 ptr_const_iterator bottom_up_ptr_end() const {
568 return PerPtrBottomUp.end();
571 /// Mark this block as being an entry block, which has one path from the
572 /// entry by definition.
573 void SetAsEntry() { TopDownPathCount = 1; }
575 /// Mark this block as being an exit block, which has one path to an exit by
577 void SetAsExit() { BottomUpPathCount = 1; }
579 PtrState &getPtrTopDownState(const Value *Arg) {
580 return PerPtrTopDown[Arg];
583 PtrState &getPtrBottomUpState(const Value *Arg) {
584 return PerPtrBottomUp[Arg];
587 void clearBottomUpPointers() {
588 PerPtrBottomUp.clear();
591 void clearTopDownPointers() {
592 PerPtrTopDown.clear();
595 void InitFromPred(const BBState &Other);
596 void InitFromSucc(const BBState &Other);
597 void MergePred(const BBState &Other);
598 void MergeSucc(const BBState &Other);
600 /// Return the number of possible unique paths from an entry to an exit
601 /// which pass through this block. This is only valid after both the
602 /// top-down and bottom-up traversals are complete.
603 unsigned GetAllPathCount() const {
604 assert(TopDownPathCount != 0);
605 assert(BottomUpPathCount != 0);
606 return TopDownPathCount * BottomUpPathCount;
609 // Specialized CFG utilities.
610 typedef SmallVectorImpl<BasicBlock *>::const_iterator edge_iterator;
611 edge_iterator pred_begin() { return Preds.begin(); }
612 edge_iterator pred_end() { return Preds.end(); }
613 edge_iterator succ_begin() { return Succs.begin(); }
614 edge_iterator succ_end() { return Succs.end(); }
616 void addSucc(BasicBlock *Succ) { Succs.push_back(Succ); }
617 void addPred(BasicBlock *Pred) { Preds.push_back(Pred); }
619 bool isExit() const { return Succs.empty(); }
623 void BBState::InitFromPred(const BBState &Other) {
624 PerPtrTopDown = Other.PerPtrTopDown;
625 TopDownPathCount = Other.TopDownPathCount;
628 void BBState::InitFromSucc(const BBState &Other) {
629 PerPtrBottomUp = Other.PerPtrBottomUp;
630 BottomUpPathCount = Other.BottomUpPathCount;
633 /// The top-down traversal uses this to merge information about predecessors to
634 /// form the initial state for a new block.
635 void BBState::MergePred(const BBState &Other) {
636 // Other.TopDownPathCount can be 0, in which case it is either dead or a
637 // loop backedge. Loop backedges are special.
638 TopDownPathCount += Other.TopDownPathCount;
640 // Check for overflow. If we have overflow, fall back to conservative
642 if (TopDownPathCount < Other.TopDownPathCount) {
643 clearTopDownPointers();
647 // For each entry in the other set, if our set has an entry with the same key,
648 // merge the entries. Otherwise, copy the entry and merge it with an empty
650 for (ptr_const_iterator MI = Other.top_down_ptr_begin(),
651 ME = Other.top_down_ptr_end(); MI != ME; ++MI) {
652 std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI);
653 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
657 // For each entry in our set, if the other set doesn't have an entry with the
658 // same key, force it to merge with an empty entry.
659 for (ptr_iterator MI = top_down_ptr_begin(),
660 ME = top_down_ptr_end(); MI != ME; ++MI)
661 if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end())
662 MI->second.Merge(PtrState(), /*TopDown=*/true);
665 /// The bottom-up traversal uses this to merge information about successors to
666 /// form the initial state for a new block.
667 void BBState::MergeSucc(const BBState &Other) {
668 // Other.BottomUpPathCount can be 0, in which case it is either dead or a
669 // loop backedge. Loop backedges are special.
670 BottomUpPathCount += Other.BottomUpPathCount;
672 // Check for overflow. If we have overflow, fall back to conservative
674 if (BottomUpPathCount < Other.BottomUpPathCount) {
675 clearBottomUpPointers();
679 // For each entry in the other set, if our set has an entry with the
680 // same key, merge the entries. Otherwise, copy the entry and merge
681 // it with an empty entry.
682 for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(),
683 ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) {
684 std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI);
685 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
689 // For each entry in our set, if the other set doesn't have an entry
690 // with the same key, force it to merge with an empty entry.
691 for (ptr_iterator MI = bottom_up_ptr_begin(),
692 ME = bottom_up_ptr_end(); MI != ME; ++MI)
693 if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end())
694 MI->second.Merge(PtrState(), /*TopDown=*/false);
697 // Only enable ARC Annotations if we are building a debug version of
700 #define ARC_ANNOTATIONS
703 // Define some macros along the lines of DEBUG and some helper functions to make
704 // it cleaner to create annotations in the source code and to no-op when not
705 // building in debug mode.
706 #ifdef ARC_ANNOTATIONS
708 #include "llvm/Support/CommandLine.h"
710 /// Enable/disable ARC sequence annotations.
712 EnableARCAnnotations("enable-objc-arc-annotations", cl::init(false));
714 /// This function appends a unique ARCAnnotationProvenanceSourceMDKind id to an
715 /// instruction so that we can track backwards when post processing via the llvm
716 /// arc annotation processor tool. If the function is an
717 static MDString *AppendMDNodeToSourcePtr(unsigned NodeId,
721 // If pointer is a result of an instruction and it does not have a source
722 // MDNode it, attach a new MDNode onto it. If pointer is a result of
723 // an instruction and does have a source MDNode attached to it, return a
724 // reference to said Node. Otherwise just return 0.
725 if (Instruction *Inst = dyn_cast<Instruction>(Ptr)) {
727 if (!(Node = Inst->getMetadata(NodeId))) {
728 // We do not have any node. Generate and attatch the hash MDString to the
731 // We just use an MDString to ensure that this metadata gets written out
732 // of line at the module level and to provide a very simple format
733 // encoding the information herein. Both of these makes it simpler to
734 // parse the annotations by a simple external program.
736 raw_string_ostream os(Str);
737 os << "(" << Inst->getParent()->getParent()->getName() << ",%"
738 << Inst->getName() << ")";
740 Hash = MDString::get(Inst->getContext(), os.str());
741 Inst->setMetadata(NodeId, MDNode::get(Inst->getContext(),Hash));
743 // We have a node. Grab its hash and return it.
744 assert(Node->getNumOperands() == 1 &&
745 "An ARCAnnotationProvenanceSourceMDKind can only have 1 operand.");
746 Hash = cast<MDString>(Node->getOperand(0));
748 } else if (Argument *Arg = dyn_cast<Argument>(Ptr)) {
750 raw_string_ostream os(str);
751 os << "(" << Arg->getParent()->getName() << ",%" << Arg->getName()
753 Hash = MDString::get(Arg->getContext(), os.str());
759 static std::string SequenceToString(Sequence A) {
761 raw_string_ostream os(str);
766 /// Helper function to change a Sequence into a String object using our overload
767 /// for raw_ostream so we only have printing code in one location.
768 static MDString *SequenceToMDString(LLVMContext &Context,
770 return MDString::get(Context, SequenceToString(A));
773 /// A simple function to generate a MDNode which describes the change in state
774 /// for Value *Ptr caused by Instruction *Inst.
775 static void AppendMDNodeToInstForPtr(unsigned NodeId,
778 MDString *PtrSourceMDNodeID,
782 Value *tmp[3] = {PtrSourceMDNodeID,
783 SequenceToMDString(Inst->getContext(),
785 SequenceToMDString(Inst->getContext(),
787 Node = MDNode::get(Inst->getContext(),
788 ArrayRef<Value*>(tmp, 3));
790 Inst->setMetadata(NodeId, Node);
793 /// Add to the beginning of the basic block llvm.ptr.annotations which show the
794 /// state of a pointer at the entrance to a basic block.
795 static void GenerateARCBBEntranceAnnotation(const char *Name, BasicBlock *BB,
796 Value *Ptr, Sequence Seq) {
797 Module *M = BB->getParent()->getParent();
798 LLVMContext &C = M->getContext();
799 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
800 Type *I8XX = PointerType::getUnqual(I8X);
801 Type *Params[] = {I8XX, I8XX};
802 FunctionType *FTy = FunctionType::get(Type::getVoidTy(C),
803 ArrayRef<Type*>(Params, 2),
805 Constant *Callee = M->getOrInsertFunction(Name, FTy);
807 IRBuilder<> Builder(BB, BB->getFirstInsertionPt());
810 StringRef Tmp = Ptr->getName();
811 if (0 == (PtrName = M->getGlobalVariable(Tmp, true))) {
812 Value *ActualPtrName = Builder.CreateGlobalStringPtr(Tmp,
814 PtrName = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
815 cast<Constant>(ActualPtrName), Tmp);
819 std::string SeqStr = SequenceToString(Seq);
820 if (0 == (S = M->getGlobalVariable(SeqStr, true))) {
821 Value *ActualPtrName = Builder.CreateGlobalStringPtr(SeqStr,
823 S = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
824 cast<Constant>(ActualPtrName), SeqStr);
827 Builder.CreateCall2(Callee, PtrName, S);
830 /// Add to the end of the basic block llvm.ptr.annotations which show the state
831 /// of the pointer at the bottom of the basic block.
832 static void GenerateARCBBTerminatorAnnotation(const char *Name, BasicBlock *BB,
833 Value *Ptr, Sequence Seq) {
834 Module *M = BB->getParent()->getParent();
835 LLVMContext &C = M->getContext();
836 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
837 Type *I8XX = PointerType::getUnqual(I8X);
838 Type *Params[] = {I8XX, I8XX};
839 FunctionType *FTy = FunctionType::get(Type::getVoidTy(C),
840 ArrayRef<Type*>(Params, 2),
842 Constant *Callee = M->getOrInsertFunction(Name, FTy);
844 IRBuilder<> Builder(BB, llvm::prior(BB->end()));
847 StringRef Tmp = Ptr->getName();
848 if (0 == (PtrName = M->getGlobalVariable(Tmp, true))) {
849 Value *ActualPtrName = Builder.CreateGlobalStringPtr(Tmp,
851 PtrName = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
852 cast<Constant>(ActualPtrName), Tmp);
856 std::string SeqStr = SequenceToString(Seq);
857 if (0 == (S = M->getGlobalVariable(SeqStr, true))) {
858 Value *ActualPtrName = Builder.CreateGlobalStringPtr(SeqStr,
860 S = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
861 cast<Constant>(ActualPtrName), SeqStr);
863 Builder.CreateCall2(Callee, PtrName, S);
866 /// Adds a source annotation to pointer and a state change annotation to Inst
867 /// referencing the source annotation and the old/new state of pointer.
868 static void GenerateARCAnnotation(unsigned InstMDId,
874 if (EnableARCAnnotations) {
875 // First generate the source annotation on our pointer. This will return an
876 // MDString* if Ptr actually comes from an instruction implying we can put
877 // in a source annotation. If AppendMDNodeToSourcePtr returns 0 (i.e. NULL),
878 // then we know that our pointer is from an Argument so we put a reference
879 // to the argument number.
881 // The point of this is to make it easy for the
882 // llvm-arc-annotation-processor tool to cross reference where the source
883 // pointer is in the LLVM IR since the LLVM IR parser does not submit such
884 // information via debug info for backends to use (since why would anyone
885 // need such a thing from LLVM IR besides in non standard cases
887 MDString *SourcePtrMDNode =
888 AppendMDNodeToSourcePtr(PtrMDId, Ptr);
889 AppendMDNodeToInstForPtr(InstMDId, Inst, Ptr, SourcePtrMDNode, OldSeq,
894 // The actual interface for accessing the above functionality is defined via
895 // some simple macros which are defined below. We do this so that the user does
896 // not need to pass in what metadata id is needed resulting in cleaner code and
897 // additionally since it provides an easy way to conditionally no-op all
898 // annotation support in a non-debug build.
900 /// Use this macro to annotate a sequence state change when processing
901 /// instructions bottom up,
902 #define ANNOTATE_BOTTOMUP(inst, ptr, old, new) \
903 GenerateARCAnnotation(ARCAnnotationBottomUpMDKind, \
904 ARCAnnotationProvenanceSourceMDKind, (inst), \
905 const_cast<Value*>(ptr), (old), (new))
906 /// Use this macro to annotate a sequence state change when processing
907 /// instructions top down.
908 #define ANNOTATE_TOPDOWN(inst, ptr, old, new) \
909 GenerateARCAnnotation(ARCAnnotationTopDownMDKind, \
910 ARCAnnotationProvenanceSourceMDKind, (inst), \
911 const_cast<Value*>(ptr), (old), (new))
913 #define ANNOTATE_BB(_states, _bb, _name, _type, _direction) \
915 if (EnableARCAnnotations) { \
916 for(BBState::ptr_const_iterator I = (_states)._direction##_ptr_begin(), \
917 E = (_states)._direction##_ptr_end(); I != E; ++I) { \
918 Value *Ptr = const_cast<Value*>(I->first); \
919 Sequence Seq = I->second.GetSeq(); \
920 GenerateARCBB ## _type ## Annotation(_name, (_bb), Ptr, Seq); \
925 #define ANNOTATE_BOTTOMUP_BBSTART(_states, _basicblock) \
926 ANNOTATE_BB(_states, _basicblock, "llvm.arc.annotation.bottomup.bbstart", \
928 #define ANNOTATE_BOTTOMUP_BBEND(_states, _basicblock) \
929 ANNOTATE_BB(_states, _basicblock, "llvm.arc.annotation.bottomup.bbend", \
930 Terminator, bottom_up)
931 #define ANNOTATE_TOPDOWN_BBSTART(_states, _basicblock) \
932 ANNOTATE_BB(_states, _basicblock, "llvm.arc.annotation.topdown.bbstart", \
934 #define ANNOTATE_TOPDOWN_BBEND(_states, _basicblock) \
935 ANNOTATE_BB(_states, _basicblock, "llvm.arc.annotation.topdown.bbend", \
936 Terminator, top_down)
938 #else // !ARC_ANNOTATION
939 // If annotations are off, noop.
940 #define ANNOTATE_BOTTOMUP(inst, ptr, old, new)
941 #define ANNOTATE_TOPDOWN(inst, ptr, old, new)
942 #define ANNOTATE_BOTTOMUP_BBSTART(states, basicblock)
943 #define ANNOTATE_BOTTOMUP_BBEND(states, basicblock)
944 #define ANNOTATE_TOPDOWN_BBSTART(states, basicblock)
945 #define ANNOTATE_TOPDOWN_BBEND(states, basicblock)
946 #endif // !ARC_ANNOTATION
949 /// \brief The main ARC optimization pass.
950 class ObjCARCOpt : public FunctionPass {
952 ProvenanceAnalysis PA;
954 /// A flag indicating whether this optimization pass should run.
957 /// Declarations for ObjC runtime functions, for use in creating calls to
958 /// them. These are initialized lazily to avoid cluttering up the Module
959 /// with unused declarations.
961 /// Declaration for ObjC runtime function
962 /// objc_retainAutoreleasedReturnValue.
963 Constant *RetainRVCallee;
964 /// Declaration for ObjC runtime function objc_autoreleaseReturnValue.
965 Constant *AutoreleaseRVCallee;
966 /// Declaration for ObjC runtime function objc_release.
967 Constant *ReleaseCallee;
968 /// Declaration for ObjC runtime function objc_retain.
969 Constant *RetainCallee;
970 /// Declaration for ObjC runtime function objc_retainBlock.
971 Constant *RetainBlockCallee;
972 /// Declaration for ObjC runtime function objc_autorelease.
973 Constant *AutoreleaseCallee;
975 /// Flags which determine whether each of the interesting runtine functions
976 /// is in fact used in the current function.
977 unsigned UsedInThisFunction;
979 /// The Metadata Kind for clang.imprecise_release metadata.
980 unsigned ImpreciseReleaseMDKind;
982 /// The Metadata Kind for clang.arc.copy_on_escape metadata.
983 unsigned CopyOnEscapeMDKind;
985 /// The Metadata Kind for clang.arc.no_objc_arc_exceptions metadata.
986 unsigned NoObjCARCExceptionsMDKind;
988 #ifdef ARC_ANNOTATIONS
989 /// The Metadata Kind for llvm.arc.annotation.bottomup metadata.
990 unsigned ARCAnnotationBottomUpMDKind;
991 /// The Metadata Kind for llvm.arc.annotation.topdown metadata.
992 unsigned ARCAnnotationTopDownMDKind;
993 /// The Metadata Kind for llvm.arc.annotation.provenancesource metadata.
994 unsigned ARCAnnotationProvenanceSourceMDKind;
995 #endif // ARC_ANNOATIONS
997 Constant *getRetainRVCallee(Module *M);
998 Constant *getAutoreleaseRVCallee(Module *M);
999 Constant *getReleaseCallee(Module *M);
1000 Constant *getRetainCallee(Module *M);
1001 Constant *getRetainBlockCallee(Module *M);
1002 Constant *getAutoreleaseCallee(Module *M);
1004 bool IsRetainBlockOptimizable(const Instruction *Inst);
1006 void OptimizeRetainCall(Function &F, Instruction *Retain);
1007 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
1008 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
1009 InstructionClass &Class);
1010 bool OptimizeRetainBlockCall(Function &F, Instruction *RetainBlock,
1011 InstructionClass &Class);
1012 void OptimizeIndividualCalls(Function &F);
1014 void CheckForCFGHazards(const BasicBlock *BB,
1015 DenseMap<const BasicBlock *, BBState> &BBStates,
1016 BBState &MyStates) const;
1017 bool VisitInstructionBottomUp(Instruction *Inst,
1019 MapVector<Value *, RRInfo> &Retains,
1021 bool VisitBottomUp(BasicBlock *BB,
1022 DenseMap<const BasicBlock *, BBState> &BBStates,
1023 MapVector<Value *, RRInfo> &Retains);
1024 bool VisitInstructionTopDown(Instruction *Inst,
1025 DenseMap<Value *, RRInfo> &Releases,
1027 bool VisitTopDown(BasicBlock *BB,
1028 DenseMap<const BasicBlock *, BBState> &BBStates,
1029 DenseMap<Value *, RRInfo> &Releases);
1030 bool Visit(Function &F,
1031 DenseMap<const BasicBlock *, BBState> &BBStates,
1032 MapVector<Value *, RRInfo> &Retains,
1033 DenseMap<Value *, RRInfo> &Releases);
1035 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1036 MapVector<Value *, RRInfo> &Retains,
1037 DenseMap<Value *, RRInfo> &Releases,
1038 SmallVectorImpl<Instruction *> &DeadInsts,
1041 bool ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState> &BBStates,
1042 MapVector<Value *, RRInfo> &Retains,
1043 DenseMap<Value *, RRInfo> &Releases,
1045 SmallVector<Instruction *, 4> &NewRetains,
1046 SmallVector<Instruction *, 4> &NewReleases,
1047 SmallVector<Instruction *, 8> &DeadInsts,
1048 RRInfo &RetainsToMove,
1049 RRInfo &ReleasesToMove,
1052 bool &AnyPairsCompletelyEliminated);
1054 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1055 MapVector<Value *, RRInfo> &Retains,
1056 DenseMap<Value *, RRInfo> &Releases,
1059 void OptimizeWeakCalls(Function &F);
1061 bool OptimizeSequences(Function &F);
1063 void OptimizeReturns(Function &F);
1065 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1066 virtual bool doInitialization(Module &M);
1067 virtual bool runOnFunction(Function &F);
1068 virtual void releaseMemory();
1072 ObjCARCOpt() : FunctionPass(ID) {
1073 initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1078 char ObjCARCOpt::ID = 0;
1079 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1080 "objc-arc", "ObjC ARC optimization", false, false)
1081 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1082 INITIALIZE_PASS_END(ObjCARCOpt,
1083 "objc-arc", "ObjC ARC optimization", false, false)
1085 Pass *llvm::createObjCARCOptPass() {
1086 return new ObjCARCOpt();
1089 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1090 AU.addRequired<ObjCARCAliasAnalysis>();
1091 AU.addRequired<AliasAnalysis>();
1092 // ARC optimization doesn't currently split critical edges.
1093 AU.setPreservesCFG();
1096 bool ObjCARCOpt::IsRetainBlockOptimizable(const Instruction *Inst) {
1097 // Without the magic metadata tag, we have to assume this might be an
1098 // objc_retainBlock call inserted to convert a block pointer to an id,
1099 // in which case it really is needed.
1100 if (!Inst->getMetadata(CopyOnEscapeMDKind))
1103 // If the pointer "escapes" (not including being used in a call),
1104 // the copy may be needed.
1105 if (DoesRetainableObjPtrEscape(Inst))
1108 // Otherwise, it's not needed.
1112 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1113 if (!RetainRVCallee) {
1114 LLVMContext &C = M->getContext();
1115 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1116 Type *Params[] = { I8X };
1117 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1118 AttributeSet Attribute =
1119 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1120 Attribute::NoUnwind);
1122 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1125 return RetainRVCallee;
1128 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1129 if (!AutoreleaseRVCallee) {
1130 LLVMContext &C = M->getContext();
1131 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1132 Type *Params[] = { I8X };
1133 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1134 AttributeSet Attribute =
1135 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1136 Attribute::NoUnwind);
1137 AutoreleaseRVCallee =
1138 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1141 return AutoreleaseRVCallee;
1144 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1145 if (!ReleaseCallee) {
1146 LLVMContext &C = M->getContext();
1147 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1148 AttributeSet Attribute =
1149 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1150 Attribute::NoUnwind);
1152 M->getOrInsertFunction(
1154 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1157 return ReleaseCallee;
1160 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1161 if (!RetainCallee) {
1162 LLVMContext &C = M->getContext();
1163 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1164 AttributeSet Attribute =
1165 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1166 Attribute::NoUnwind);
1168 M->getOrInsertFunction(
1170 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1173 return RetainCallee;
1176 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1177 if (!RetainBlockCallee) {
1178 LLVMContext &C = M->getContext();
1179 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1180 // objc_retainBlock is not nounwind because it calls user copy constructors
1181 // which could theoretically throw.
1183 M->getOrInsertFunction(
1185 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1188 return RetainBlockCallee;
1191 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1192 if (!AutoreleaseCallee) {
1193 LLVMContext &C = M->getContext();
1194 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1195 AttributeSet Attribute =
1196 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1197 Attribute::NoUnwind);
1199 M->getOrInsertFunction(
1201 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1204 return AutoreleaseCallee;
1207 /// Turn objc_retain into objc_retainAutoreleasedReturnValue if the operand is a
1210 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
1211 ImmutableCallSite CS(GetObjCArg(Retain));
1212 const Instruction *Call = CS.getInstruction();
1214 if (Call->getParent() != Retain->getParent()) return;
1216 // Check that the call is next to the retain.
1217 BasicBlock::const_iterator I = Call;
1219 while (IsNoopInstruction(I)) ++I;
1223 // Turn it to an objc_retainAutoreleasedReturnValue..
1227 DEBUG(dbgs() << "Transforming objc_retain => "
1228 "objc_retainAutoreleasedReturnValue since the operand is a "
1229 "return value.\nOld: "<< *Retain << "\n");
1231 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
1233 DEBUG(dbgs() << "New: " << *Retain << "\n");
1236 /// Turn objc_retainAutoreleasedReturnValue into objc_retain if the operand is
1237 /// not a return value. Or, if it can be paired with an
1238 /// objc_autoreleaseReturnValue, delete the pair and return true.
1240 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
1241 // Check for the argument being from an immediately preceding call or invoke.
1242 const Value *Arg = GetObjCArg(RetainRV);
1243 ImmutableCallSite CS(Arg);
1244 if (const Instruction *Call = CS.getInstruction()) {
1245 if (Call->getParent() == RetainRV->getParent()) {
1246 BasicBlock::const_iterator I = Call;
1248 while (IsNoopInstruction(I)) ++I;
1249 if (&*I == RetainRV)
1251 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
1252 BasicBlock *RetainRVParent = RetainRV->getParent();
1253 if (II->getNormalDest() == RetainRVParent) {
1254 BasicBlock::const_iterator I = RetainRVParent->begin();
1255 while (IsNoopInstruction(I)) ++I;
1256 if (&*I == RetainRV)
1262 // Check for being preceded by an objc_autoreleaseReturnValue on the same
1263 // pointer. In this case, we can delete the pair.
1264 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
1266 do --I; while (I != Begin && IsNoopInstruction(I));
1267 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
1268 GetObjCArg(I) == Arg) {
1272 DEBUG(dbgs() << "Erasing autoreleaseRV,retainRV pair: " << *I << "\n"
1273 << "Erasing " << *RetainRV << "\n");
1275 EraseInstruction(I);
1276 EraseInstruction(RetainRV);
1281 // Turn it to a plain objc_retain.
1285 DEBUG(dbgs() << "Transforming objc_retainAutoreleasedReturnValue => "
1286 "objc_retain since the operand is not a return value.\n"
1287 "Old = " << *RetainRV << "\n");
1289 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
1291 DEBUG(dbgs() << "New = " << *RetainRV << "\n");
1296 /// Turn objc_autoreleaseReturnValue into objc_autorelease if the result is not
1297 /// used as a return value.
1299 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
1300 InstructionClass &Class) {
1301 // Check for a return of the pointer value.
1302 const Value *Ptr = GetObjCArg(AutoreleaseRV);
1303 SmallVector<const Value *, 2> Users;
1304 Users.push_back(Ptr);
1306 Ptr = Users.pop_back_val();
1307 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
1309 const User *I = *UI;
1310 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
1312 if (isa<BitCastInst>(I))
1315 } while (!Users.empty());
1320 DEBUG(dbgs() << "Transforming objc_autoreleaseReturnValue => "
1321 "objc_autorelease since its operand is not used as a return "
1323 "Old = " << *AutoreleaseRV << "\n");
1325 CallInst *AutoreleaseRVCI = cast<CallInst>(AutoreleaseRV);
1327 setCalledFunction(getAutoreleaseCallee(F.getParent()));
1328 AutoreleaseRVCI->setTailCall(false); // Never tail call objc_autorelease.
1329 Class = IC_Autorelease;
1331 DEBUG(dbgs() << "New: " << *AutoreleaseRV << "\n");
1335 // \brief Attempt to strength reduce objc_retainBlock calls to objc_retain
1338 // Specifically: If an objc_retainBlock call has the copy_on_escape metadata and
1339 // does not escape (following the rules of block escaping), strength reduce the
1340 // objc_retainBlock to an objc_retain.
1342 // TODO: If an objc_retainBlock call is dominated period by a previous
1343 // objc_retainBlock call, strength reduce the objc_retainBlock to an
1346 ObjCARCOpt::OptimizeRetainBlockCall(Function &F, Instruction *Inst,
1347 InstructionClass &Class) {
1348 assert(GetBasicInstructionClass(Inst) == Class);
1349 assert(IC_RetainBlock == Class);
1351 // If we can not optimize Inst, return false.
1352 if (!IsRetainBlockOptimizable(Inst))
1355 CallInst *RetainBlock = cast<CallInst>(Inst);
1356 RetainBlock->setCalledFunction(getRetainCallee(F.getParent()));
1357 // Remove copy_on_escape metadata.
1358 RetainBlock->setMetadata(CopyOnEscapeMDKind, 0);
1364 /// Visit each call, one at a time, and make simplifications without doing any
1365 /// additional analysis.
1366 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
1367 DEBUG(dbgs() << "\n== ObjCARCOpt::OptimizeIndividualCalls ==\n");
1368 // Reset all the flags in preparation for recomputing them.
1369 UsedInThisFunction = 0;
1371 // Visit all objc_* calls in F.
1372 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
1373 Instruction *Inst = &*I++;
1375 InstructionClass Class = GetBasicInstructionClass(Inst);
1377 DEBUG(dbgs() << "Visiting: Class: " << Class << "; " << *Inst << "\n");
1382 // Delete no-op casts. These function calls have special semantics, but
1383 // the semantics are entirely implemented via lowering in the front-end,
1384 // so by the time they reach the optimizer, they are just no-op calls
1385 // which return their argument.
1387 // There are gray areas here, as the ability to cast reference-counted
1388 // pointers to raw void* and back allows code to break ARC assumptions,
1389 // however these are currently considered to be unimportant.
1393 DEBUG(dbgs() << "Erasing no-op cast: " << *Inst << "\n");
1394 EraseInstruction(Inst);
1397 // If the pointer-to-weak-pointer is null, it's undefined behavior.
1400 case IC_LoadWeakRetained:
1402 case IC_DestroyWeak: {
1403 CallInst *CI = cast<CallInst>(Inst);
1404 if (IsNullOrUndef(CI->getArgOperand(0))) {
1406 Type *Ty = CI->getArgOperand(0)->getType();
1407 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1408 Constant::getNullValue(Ty),
1410 llvm::Value *NewValue = UndefValue::get(CI->getType());
1411 DEBUG(dbgs() << "A null pointer-to-weak-pointer is undefined behavior."
1412 "\nOld = " << *CI << "\nNew = " << *NewValue << "\n");
1413 CI->replaceAllUsesWith(NewValue);
1414 CI->eraseFromParent();
1421 CallInst *CI = cast<CallInst>(Inst);
1422 if (IsNullOrUndef(CI->getArgOperand(0)) ||
1423 IsNullOrUndef(CI->getArgOperand(1))) {
1425 Type *Ty = CI->getArgOperand(0)->getType();
1426 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1427 Constant::getNullValue(Ty),
1430 llvm::Value *NewValue = UndefValue::get(CI->getType());
1431 DEBUG(dbgs() << "A null pointer-to-weak-pointer is undefined behavior."
1432 "\nOld = " << *CI << "\nNew = " << *NewValue << "\n");
1434 CI->replaceAllUsesWith(NewValue);
1435 CI->eraseFromParent();
1440 case IC_RetainBlock:
1441 // If we strength reduce an objc_retainBlock to amn objc_retain, continue
1442 // onto the objc_retain peephole optimizations. Otherwise break.
1443 if (!OptimizeRetainBlockCall(F, Inst, Class))
1447 OptimizeRetainCall(F, Inst);
1450 if (OptimizeRetainRVCall(F, Inst))
1453 case IC_AutoreleaseRV:
1454 OptimizeAutoreleaseRVCall(F, Inst, Class);
1458 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
1459 if (IsAutorelease(Class) && Inst->use_empty()) {
1460 CallInst *Call = cast<CallInst>(Inst);
1461 const Value *Arg = Call->getArgOperand(0);
1462 Arg = FindSingleUseIdentifiedObject(Arg);
1467 // Create the declaration lazily.
1468 LLVMContext &C = Inst->getContext();
1470 CallInst::Create(getReleaseCallee(F.getParent()),
1471 Call->getArgOperand(0), "", Call);
1472 NewCall->setMetadata(ImpreciseReleaseMDKind,
1473 MDNode::get(C, ArrayRef<Value *>()));
1475 DEBUG(dbgs() << "Replacing autorelease{,RV}(x) with objc_release(x) "
1476 "since x is otherwise unused.\nOld: " << *Call << "\nNew: "
1477 << *NewCall << "\n");
1479 EraseInstruction(Call);
1485 // For functions which can never be passed stack arguments, add
1487 if (IsAlwaysTail(Class)) {
1489 DEBUG(dbgs() << "Adding tail keyword to function since it can never be "
1490 "passed stack args: " << *Inst << "\n");
1491 cast<CallInst>(Inst)->setTailCall();
1494 // Ensure that functions that can never have a "tail" keyword due to the
1495 // semantics of ARC truly do not do so.
1496 if (IsNeverTail(Class)) {
1498 DEBUG(dbgs() << "Removing tail keyword from function: " << *Inst <<
1500 cast<CallInst>(Inst)->setTailCall(false);
1503 // Set nounwind as needed.
1504 if (IsNoThrow(Class)) {
1506 DEBUG(dbgs() << "Found no throw class. Setting nounwind on: " << *Inst
1508 cast<CallInst>(Inst)->setDoesNotThrow();
1511 if (!IsNoopOnNull(Class)) {
1512 UsedInThisFunction |= 1 << Class;
1516 const Value *Arg = GetObjCArg(Inst);
1518 // ARC calls with null are no-ops. Delete them.
1519 if (IsNullOrUndef(Arg)) {
1522 DEBUG(dbgs() << "ARC calls with null are no-ops. Erasing: " << *Inst
1524 EraseInstruction(Inst);
1528 // Keep track of which of retain, release, autorelease, and retain_block
1529 // are actually present in this function.
1530 UsedInThisFunction |= 1 << Class;
1532 // If Arg is a PHI, and one or more incoming values to the
1533 // PHI are null, and the call is control-equivalent to the PHI, and there
1534 // are no relevant side effects between the PHI and the call, the call
1535 // could be pushed up to just those paths with non-null incoming values.
1536 // For now, don't bother splitting critical edges for this.
1537 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
1538 Worklist.push_back(std::make_pair(Inst, Arg));
1540 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
1544 const PHINode *PN = dyn_cast<PHINode>(Arg);
1547 // Determine if the PHI has any null operands, or any incoming
1549 bool HasNull = false;
1550 bool HasCriticalEdges = false;
1551 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1553 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
1554 if (IsNullOrUndef(Incoming))
1556 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
1557 .getNumSuccessors() != 1) {
1558 HasCriticalEdges = true;
1562 // If we have null operands and no critical edges, optimize.
1563 if (!HasCriticalEdges && HasNull) {
1564 SmallPtrSet<Instruction *, 4> DependingInstructions;
1565 SmallPtrSet<const BasicBlock *, 4> Visited;
1567 // Check that there is nothing that cares about the reference
1568 // count between the call and the phi.
1571 case IC_RetainBlock:
1572 // These can always be moved up.
1575 // These can't be moved across things that care about the retain
1577 FindDependencies(NeedsPositiveRetainCount, Arg,
1578 Inst->getParent(), Inst,
1579 DependingInstructions, Visited, PA);
1581 case IC_Autorelease:
1582 // These can't be moved across autorelease pool scope boundaries.
1583 FindDependencies(AutoreleasePoolBoundary, Arg,
1584 Inst->getParent(), Inst,
1585 DependingInstructions, Visited, PA);
1588 case IC_AutoreleaseRV:
1589 // Don't move these; the RV optimization depends on the autoreleaseRV
1590 // being tail called, and the retainRV being immediately after a call
1591 // (which might still happen if we get lucky with codegen layout, but
1592 // it's not worth taking the chance).
1595 llvm_unreachable("Invalid dependence flavor");
1598 if (DependingInstructions.size() == 1 &&
1599 *DependingInstructions.begin() == PN) {
1602 // Clone the call into each predecessor that has a non-null value.
1603 CallInst *CInst = cast<CallInst>(Inst);
1604 Type *ParamTy = CInst->getArgOperand(0)->getType();
1605 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1607 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
1608 if (!IsNullOrUndef(Incoming)) {
1609 CallInst *Clone = cast<CallInst>(CInst->clone());
1610 Value *Op = PN->getIncomingValue(i);
1611 Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
1612 if (Op->getType() != ParamTy)
1613 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
1614 Clone->setArgOperand(0, Op);
1615 Clone->insertBefore(InsertPos);
1617 DEBUG(dbgs() << "Cloning "
1619 "And inserting clone at " << *InsertPos << "\n");
1620 Worklist.push_back(std::make_pair(Clone, Incoming));
1623 // Erase the original call.
1624 DEBUG(dbgs() << "Erasing: " << *CInst << "\n");
1625 EraseInstruction(CInst);
1629 } while (!Worklist.empty());
1633 /// Check for critical edges, loop boundaries, irreducible control flow, or
1634 /// other CFG structures where moving code across the edge would result in it
1635 /// being executed more.
1637 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
1638 DenseMap<const BasicBlock *, BBState> &BBStates,
1639 BBState &MyStates) const {
1640 // If any top-down local-use or possible-dec has a succ which is earlier in
1641 // the sequence, forget it.
1642 for (BBState::ptr_iterator I = MyStates.top_down_ptr_begin(),
1643 E = MyStates.top_down_ptr_end(); I != E; ++I)
1644 switch (I->second.GetSeq()) {
1647 const Value *Arg = I->first;
1648 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1649 bool SomeSuccHasSame = false;
1650 bool AllSuccsHaveSame = true;
1651 PtrState &S = I->second;
1652 succ_const_iterator SI(TI), SE(TI, false);
1654 for (; SI != SE; ++SI) {
1655 Sequence SuccSSeq = S_None;
1656 bool SuccSRRIKnownSafe = false;
1657 // If VisitBottomUp has pointer information for this successor, take
1658 // what we know about it.
1659 DenseMap<const BasicBlock *, BBState>::iterator BBI =
1661 assert(BBI != BBStates.end());
1662 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
1663 SuccSSeq = SuccS.GetSeq();
1664 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
1667 case S_CanRelease: {
1668 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
1669 S.ClearSequenceProgress();
1675 SomeSuccHasSame = true;
1679 case S_MovableRelease:
1680 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
1681 AllSuccsHaveSame = false;
1684 llvm_unreachable("bottom-up pointer in retain state!");
1687 // If the state at the other end of any of the successor edges
1688 // matches the current state, require all edges to match. This
1689 // guards against loops in the middle of a sequence.
1690 if (SomeSuccHasSame && !AllSuccsHaveSame)
1691 S.ClearSequenceProgress();
1694 case S_CanRelease: {
1695 const Value *Arg = I->first;
1696 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1697 bool SomeSuccHasSame = false;
1698 bool AllSuccsHaveSame = true;
1699 PtrState &S = I->second;
1700 succ_const_iterator SI(TI), SE(TI, false);
1702 for (; SI != SE; ++SI) {
1703 Sequence SuccSSeq = S_None;
1704 bool SuccSRRIKnownSafe = false;
1705 // If VisitBottomUp has pointer information for this successor, take
1706 // what we know about it.
1707 DenseMap<const BasicBlock *, BBState>::iterator BBI =
1709 assert(BBI != BBStates.end());
1710 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
1711 SuccSSeq = SuccS.GetSeq();
1712 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
1715 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
1716 S.ClearSequenceProgress();
1722 SomeSuccHasSame = true;
1726 case S_MovableRelease:
1728 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
1729 AllSuccsHaveSame = false;
1732 llvm_unreachable("bottom-up pointer in retain state!");
1735 // If the state at the other end of any of the successor edges
1736 // matches the current state, require all edges to match. This
1737 // guards against loops in the middle of a sequence.
1738 if (SomeSuccHasSame && !AllSuccsHaveSame)
1739 S.ClearSequenceProgress();
1746 ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst,
1748 MapVector<Value *, RRInfo> &Retains,
1749 BBState &MyStates) {
1750 bool NestingDetected = false;
1751 InstructionClass Class = GetInstructionClass(Inst);
1752 const Value *Arg = 0;
1754 DEBUG(dbgs() << "Class: " << Class << "\n");
1758 Arg = GetObjCArg(Inst);
1760 PtrState &S = MyStates.getPtrBottomUpState(Arg);
1762 // If we see two releases in a row on the same pointer. If so, make
1763 // a note, and we'll cicle back to revisit it after we've
1764 // hopefully eliminated the second release, which may allow us to
1765 // eliminate the first release too.
1766 // Theoretically we could implement removal of nested retain+release
1767 // pairs by making PtrState hold a stack of states, but this is
1768 // simple and avoids adding overhead for the non-nested case.
1769 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease) {
1770 DEBUG(dbgs() << "Found nested releases (i.e. a release pair)\n");
1771 NestingDetected = true;
1774 MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
1775 Sequence NewSeq = ReleaseMetadata ? S_MovableRelease : S_Release;
1776 ANNOTATE_BOTTOMUP(Inst, Arg, S.GetSeq(), NewSeq);
1777 S.ResetSequenceProgress(NewSeq);
1778 S.RRI.ReleaseMetadata = ReleaseMetadata;
1779 S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
1780 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
1781 S.RRI.Calls.insert(Inst);
1782 S.SetKnownPositiveRefCount();
1785 case IC_RetainBlock:
1786 // In OptimizeIndividualCalls, we have strength reduced all optimizable
1787 // objc_retainBlocks to objc_retains. Thus at this point any
1788 // objc_retainBlocks that we see are not optimizable.
1792 Arg = GetObjCArg(Inst);
1794 PtrState &S = MyStates.getPtrBottomUpState(Arg);
1795 S.SetKnownPositiveRefCount();
1797 Sequence OldSeq = S.GetSeq();
1801 case S_MovableRelease:
1803 // If OldSeq is not S_Use or OldSeq is S_Use and we are tracking an
1804 // imprecise release, clear our reverse insertion points.
1805 if (OldSeq != S_Use || S.RRI.IsTrackingImpreciseReleases())
1806 S.RRI.ReverseInsertPts.clear();
1809 // Don't do retain+release tracking for IC_RetainRV, because it's
1810 // better to let it remain as the first instruction after a call.
1811 if (Class != IC_RetainRV)
1812 Retains[Inst] = S.RRI;
1813 S.ClearSequenceProgress();
1818 llvm_unreachable("bottom-up pointer in retain state!");
1820 ANNOTATE_BOTTOMUP(Inst, Arg, OldSeq, S.GetSeq());
1821 // A retain moving bottom up can be a use.
1824 case IC_AutoreleasepoolPop:
1825 // Conservatively, clear MyStates for all known pointers.
1826 MyStates.clearBottomUpPointers();
1827 return NestingDetected;
1828 case IC_AutoreleasepoolPush:
1830 // These are irrelevant.
1831 return NestingDetected;
1836 // Consider any other possible effects of this instruction on each
1837 // pointer being tracked.
1838 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
1839 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
1840 const Value *Ptr = MI->first;
1842 continue; // Handled above.
1843 PtrState &S = MI->second;
1844 Sequence Seq = S.GetSeq();
1846 // Check for possible releases.
1847 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
1848 DEBUG(dbgs() << "CanAlterRefCount: Seq: " << Seq << "; " << *Ptr
1850 S.ClearKnownPositiveRefCount();
1853 S.SetSeq(S_CanRelease);
1854 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S.GetSeq());
1858 case S_MovableRelease:
1863 llvm_unreachable("bottom-up pointer in retain state!");
1867 // Check for possible direct uses.
1870 case S_MovableRelease:
1871 if (CanUse(Inst, Ptr, PA, Class)) {
1872 DEBUG(dbgs() << "CanUse: Seq: " << Seq << "; " << *Ptr
1874 assert(S.RRI.ReverseInsertPts.empty());
1875 // If this is an invoke instruction, we're scanning it as part of
1876 // one of its successor blocks, since we can't insert code after it
1877 // in its own block, and we don't want to split critical edges.
1878 if (isa<InvokeInst>(Inst))
1879 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
1881 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
1883 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S_Use);
1884 } else if (Seq == S_Release && IsUser(Class)) {
1885 DEBUG(dbgs() << "PreciseReleaseUse: Seq: " << Seq << "; " << *Ptr
1887 // Non-movable releases depend on any possible objc pointer use.
1889 ANNOTATE_BOTTOMUP(Inst, Ptr, S_Release, S_Stop);
1890 assert(S.RRI.ReverseInsertPts.empty());
1891 // As above; handle invoke specially.
1892 if (isa<InvokeInst>(Inst))
1893 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
1895 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
1899 if (CanUse(Inst, Ptr, PA, Class)) {
1900 DEBUG(dbgs() << "PreciseStopUse: Seq: " << Seq << "; " << *Ptr
1903 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S_Use);
1911 llvm_unreachable("bottom-up pointer in retain state!");
1915 return NestingDetected;
1919 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
1920 DenseMap<const BasicBlock *, BBState> &BBStates,
1921 MapVector<Value *, RRInfo> &Retains) {
1923 DEBUG(dbgs() << "\n== ObjCARCOpt::VisitBottomUp ==\n");
1925 bool NestingDetected = false;
1926 BBState &MyStates = BBStates[BB];
1928 // Merge the states from each successor to compute the initial state
1929 // for the current block.
1930 BBState::edge_iterator SI(MyStates.succ_begin()),
1931 SE(MyStates.succ_end());
1933 const BasicBlock *Succ = *SI;
1934 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
1935 assert(I != BBStates.end());
1936 MyStates.InitFromSucc(I->second);
1938 for (; SI != SE; ++SI) {
1940 I = BBStates.find(Succ);
1941 assert(I != BBStates.end());
1942 MyStates.MergeSucc(I->second);
1946 // If ARC Annotations are enabled, output the current state of pointers at the
1947 // bottom of the basic block.
1948 ANNOTATE_BOTTOMUP_BBEND(MyStates, BB);
1950 // Visit all the instructions, bottom-up.
1951 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
1952 Instruction *Inst = llvm::prior(I);
1954 // Invoke instructions are visited as part of their successors (below).
1955 if (isa<InvokeInst>(Inst))
1958 DEBUG(dbgs() << "Visiting " << *Inst << "\n");
1960 NestingDetected |= VisitInstructionBottomUp(Inst, BB, Retains, MyStates);
1963 // If there's a predecessor with an invoke, visit the invoke as if it were
1964 // part of this block, since we can't insert code after an invoke in its own
1965 // block, and we don't want to split critical edges.
1966 for (BBState::edge_iterator PI(MyStates.pred_begin()),
1967 PE(MyStates.pred_end()); PI != PE; ++PI) {
1968 BasicBlock *Pred = *PI;
1969 if (InvokeInst *II = dyn_cast<InvokeInst>(&Pred->back()))
1970 NestingDetected |= VisitInstructionBottomUp(II, BB, Retains, MyStates);
1973 // If ARC Annotations are enabled, output the current state of pointers at the
1974 // top of the basic block.
1975 ANNOTATE_BOTTOMUP_BBSTART(MyStates, BB);
1977 return NestingDetected;
1981 ObjCARCOpt::VisitInstructionTopDown(Instruction *Inst,
1982 DenseMap<Value *, RRInfo> &Releases,
1983 BBState &MyStates) {
1984 bool NestingDetected = false;
1985 InstructionClass Class = GetInstructionClass(Inst);
1986 const Value *Arg = 0;
1989 case IC_RetainBlock:
1990 // In OptimizeIndividualCalls, we have strength reduced all optimizable
1991 // objc_retainBlocks to objc_retains. Thus at this point any
1992 // objc_retainBlocks that we see are not optimizable.
1996 Arg = GetObjCArg(Inst);
1998 PtrState &S = MyStates.getPtrTopDownState(Arg);
2000 // Don't do retain+release tracking for IC_RetainRV, because it's
2001 // better to let it remain as the first instruction after a call.
2002 if (Class != IC_RetainRV) {
2003 // If we see two retains in a row on the same pointer. If so, make
2004 // a note, and we'll cicle back to revisit it after we've
2005 // hopefully eliminated the second retain, which may allow us to
2006 // eliminate the first retain too.
2007 // Theoretically we could implement removal of nested retain+release
2008 // pairs by making PtrState hold a stack of states, but this is
2009 // simple and avoids adding overhead for the non-nested case.
2010 if (S.GetSeq() == S_Retain)
2011 NestingDetected = true;
2013 ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_Retain);
2014 S.ResetSequenceProgress(S_Retain);
2015 S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
2016 S.RRI.Calls.insert(Inst);
2019 S.SetKnownPositiveRefCount();
2021 // A retain can be a potential use; procede to the generic checking
2026 Arg = GetObjCArg(Inst);
2028 PtrState &S = MyStates.getPtrTopDownState(Arg);
2029 S.ClearKnownPositiveRefCount();
2031 Sequence OldSeq = S.GetSeq();
2033 MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2038 if (OldSeq == S_Retain || ReleaseMetadata != 0)
2039 S.RRI.ReverseInsertPts.clear();
2042 S.RRI.ReleaseMetadata = ReleaseMetadata;
2043 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2044 Releases[Inst] = S.RRI;
2045 ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_None);
2046 S.ClearSequenceProgress();
2052 case S_MovableRelease:
2053 llvm_unreachable("top-down pointer in release state!");
2057 case IC_AutoreleasepoolPop:
2058 // Conservatively, clear MyStates for all known pointers.
2059 MyStates.clearTopDownPointers();
2060 return NestingDetected;
2061 case IC_AutoreleasepoolPush:
2063 // These are irrelevant.
2064 return NestingDetected;
2069 // Consider any other possible effects of this instruction on each
2070 // pointer being tracked.
2071 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
2072 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
2073 const Value *Ptr = MI->first;
2075 continue; // Handled above.
2076 PtrState &S = MI->second;
2077 Sequence Seq = S.GetSeq();
2079 // Check for possible releases.
2080 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2081 DEBUG(dbgs() << "CanAlterRefCount: Seq: " << Seq << "; " << *Ptr
2083 S.ClearKnownPositiveRefCount();
2086 S.SetSeq(S_CanRelease);
2087 ANNOTATE_TOPDOWN(Inst, Ptr, Seq, S_CanRelease);
2088 assert(S.RRI.ReverseInsertPts.empty());
2089 S.RRI.ReverseInsertPts.insert(Inst);
2091 // One call can't cause a transition from S_Retain to S_CanRelease
2092 // and S_CanRelease to S_Use. If we've made the first transition,
2101 case S_MovableRelease:
2102 llvm_unreachable("top-down pointer in release state!");
2106 // Check for possible direct uses.
2109 if (CanUse(Inst, Ptr, PA, Class)) {
2110 DEBUG(dbgs() << "CanUse: Seq: " << Seq << "; " << *Ptr
2113 ANNOTATE_TOPDOWN(Inst, Ptr, Seq, S_Use);
2122 case S_MovableRelease:
2123 llvm_unreachable("top-down pointer in release state!");
2127 return NestingDetected;
2131 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
2132 DenseMap<const BasicBlock *, BBState> &BBStates,
2133 DenseMap<Value *, RRInfo> &Releases) {
2134 DEBUG(dbgs() << "\n== ObjCARCOpt::VisitTopDown ==\n");
2135 bool NestingDetected = false;
2136 BBState &MyStates = BBStates[BB];
2138 // Merge the states from each predecessor to compute the initial state
2139 // for the current block.
2140 BBState::edge_iterator PI(MyStates.pred_begin()),
2141 PE(MyStates.pred_end());
2143 const BasicBlock *Pred = *PI;
2144 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
2145 assert(I != BBStates.end());
2146 MyStates.InitFromPred(I->second);
2148 for (; PI != PE; ++PI) {
2150 I = BBStates.find(Pred);
2151 assert(I != BBStates.end());
2152 MyStates.MergePred(I->second);
2156 // If ARC Annotations are enabled, output the current state of pointers at the
2157 // top of the basic block.
2158 ANNOTATE_TOPDOWN_BBSTART(MyStates, BB);
2160 // Visit all the instructions, top-down.
2161 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2162 Instruction *Inst = I;
2164 DEBUG(dbgs() << "Visiting " << *Inst << "\n");
2166 NestingDetected |= VisitInstructionTopDown(Inst, Releases, MyStates);
2169 // If ARC Annotations are enabled, output the current state of pointers at the
2170 // bottom of the basic block.
2171 ANNOTATE_TOPDOWN_BBEND(MyStates, BB);
2173 CheckForCFGHazards(BB, BBStates, MyStates);
2174 return NestingDetected;
2178 ComputePostOrders(Function &F,
2179 SmallVectorImpl<BasicBlock *> &PostOrder,
2180 SmallVectorImpl<BasicBlock *> &ReverseCFGPostOrder,
2181 unsigned NoObjCARCExceptionsMDKind,
2182 DenseMap<const BasicBlock *, BBState> &BBStates) {
2183 /// The visited set, for doing DFS walks.
2184 SmallPtrSet<BasicBlock *, 16> Visited;
2186 // Do DFS, computing the PostOrder.
2187 SmallPtrSet<BasicBlock *, 16> OnStack;
2188 SmallVector<std::pair<BasicBlock *, succ_iterator>, 16> SuccStack;
2190 // Functions always have exactly one entry block, and we don't have
2191 // any other block that we treat like an entry block.
2192 BasicBlock *EntryBB = &F.getEntryBlock();
2193 BBState &MyStates = BBStates[EntryBB];
2194 MyStates.SetAsEntry();
2195 TerminatorInst *EntryTI = cast<TerminatorInst>(&EntryBB->back());
2196 SuccStack.push_back(std::make_pair(EntryBB, succ_iterator(EntryTI)));
2197 Visited.insert(EntryBB);
2198 OnStack.insert(EntryBB);
2201 BasicBlock *CurrBB = SuccStack.back().first;
2202 TerminatorInst *TI = cast<TerminatorInst>(&CurrBB->back());
2203 succ_iterator SE(TI, false);
2205 while (SuccStack.back().second != SE) {
2206 BasicBlock *SuccBB = *SuccStack.back().second++;
2207 if (Visited.insert(SuccBB)) {
2208 TerminatorInst *TI = cast<TerminatorInst>(&SuccBB->back());
2209 SuccStack.push_back(std::make_pair(SuccBB, succ_iterator(TI)));
2210 BBStates[CurrBB].addSucc(SuccBB);
2211 BBState &SuccStates = BBStates[SuccBB];
2212 SuccStates.addPred(CurrBB);
2213 OnStack.insert(SuccBB);
2217 if (!OnStack.count(SuccBB)) {
2218 BBStates[CurrBB].addSucc(SuccBB);
2219 BBStates[SuccBB].addPred(CurrBB);
2222 OnStack.erase(CurrBB);
2223 PostOrder.push_back(CurrBB);
2224 SuccStack.pop_back();
2225 } while (!SuccStack.empty());
2229 // Do reverse-CFG DFS, computing the reverse-CFG PostOrder.
2230 // Functions may have many exits, and there also blocks which we treat
2231 // as exits due to ignored edges.
2232 SmallVector<std::pair<BasicBlock *, BBState::edge_iterator>, 16> PredStack;
2233 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2234 BasicBlock *ExitBB = I;
2235 BBState &MyStates = BBStates[ExitBB];
2236 if (!MyStates.isExit())
2239 MyStates.SetAsExit();
2241 PredStack.push_back(std::make_pair(ExitBB, MyStates.pred_begin()));
2242 Visited.insert(ExitBB);
2243 while (!PredStack.empty()) {
2244 reverse_dfs_next_succ:
2245 BBState::edge_iterator PE = BBStates[PredStack.back().first].pred_end();
2246 while (PredStack.back().second != PE) {
2247 BasicBlock *BB = *PredStack.back().second++;
2248 if (Visited.insert(BB)) {
2249 PredStack.push_back(std::make_pair(BB, BBStates[BB].pred_begin()));
2250 goto reverse_dfs_next_succ;
2253 ReverseCFGPostOrder.push_back(PredStack.pop_back_val().first);
2258 // Visit the function both top-down and bottom-up.
2260 ObjCARCOpt::Visit(Function &F,
2261 DenseMap<const BasicBlock *, BBState> &BBStates,
2262 MapVector<Value *, RRInfo> &Retains,
2263 DenseMap<Value *, RRInfo> &Releases) {
2265 // Use reverse-postorder traversals, because we magically know that loops
2266 // will be well behaved, i.e. they won't repeatedly call retain on a single
2267 // pointer without doing a release. We can't use the ReversePostOrderTraversal
2268 // class here because we want the reverse-CFG postorder to consider each
2269 // function exit point, and we want to ignore selected cycle edges.
2270 SmallVector<BasicBlock *, 16> PostOrder;
2271 SmallVector<BasicBlock *, 16> ReverseCFGPostOrder;
2272 ComputePostOrders(F, PostOrder, ReverseCFGPostOrder,
2273 NoObjCARCExceptionsMDKind,
2276 // Use reverse-postorder on the reverse CFG for bottom-up.
2277 bool BottomUpNestingDetected = false;
2278 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2279 ReverseCFGPostOrder.rbegin(), E = ReverseCFGPostOrder.rend();
2281 BottomUpNestingDetected |= VisitBottomUp(*I, BBStates, Retains);
2283 // Use reverse-postorder for top-down.
2284 bool TopDownNestingDetected = false;
2285 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2286 PostOrder.rbegin(), E = PostOrder.rend();
2288 TopDownNestingDetected |= VisitTopDown(*I, BBStates, Releases);
2290 return TopDownNestingDetected && BottomUpNestingDetected;
2293 /// Move the calls in RetainsToMove and ReleasesToMove.
2294 void ObjCARCOpt::MoveCalls(Value *Arg,
2295 RRInfo &RetainsToMove,
2296 RRInfo &ReleasesToMove,
2297 MapVector<Value *, RRInfo> &Retains,
2298 DenseMap<Value *, RRInfo> &Releases,
2299 SmallVectorImpl<Instruction *> &DeadInsts,
2301 Type *ArgTy = Arg->getType();
2302 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
2304 DEBUG(dbgs() << "== ObjCARCOpt::MoveCalls ==\n");
2306 // Insert the new retain and release calls.
2307 for (SmallPtrSet<Instruction *, 2>::const_iterator
2308 PI = ReleasesToMove.ReverseInsertPts.begin(),
2309 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2310 Instruction *InsertPt = *PI;
2311 Value *MyArg = ArgTy == ParamTy ? Arg :
2312 new BitCastInst(Arg, ParamTy, "", InsertPt);
2314 CallInst::Create(getRetainCallee(M), MyArg, "", InsertPt);
2315 Call->setDoesNotThrow();
2316 Call->setTailCall();
2318 DEBUG(dbgs() << "Inserting new Release: " << *Call << "\n"
2319 "At insertion point: " << *InsertPt << "\n");
2321 for (SmallPtrSet<Instruction *, 2>::const_iterator
2322 PI = RetainsToMove.ReverseInsertPts.begin(),
2323 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2324 Instruction *InsertPt = *PI;
2325 Value *MyArg = ArgTy == ParamTy ? Arg :
2326 new BitCastInst(Arg, ParamTy, "", InsertPt);
2327 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
2329 // Attach a clang.imprecise_release metadata tag, if appropriate.
2330 if (MDNode *M = ReleasesToMove.ReleaseMetadata)
2331 Call->setMetadata(ImpreciseReleaseMDKind, M);
2332 Call->setDoesNotThrow();
2333 if (ReleasesToMove.IsTailCallRelease)
2334 Call->setTailCall();
2336 DEBUG(dbgs() << "Inserting new Release: " << *Call << "\n"
2337 "At insertion point: " << *InsertPt << "\n");
2340 // Delete the original retain and release calls.
2341 for (SmallPtrSet<Instruction *, 2>::const_iterator
2342 AI = RetainsToMove.Calls.begin(),
2343 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
2344 Instruction *OrigRetain = *AI;
2345 Retains.blot(OrigRetain);
2346 DeadInsts.push_back(OrigRetain);
2347 DEBUG(dbgs() << "Deleting retain: " << *OrigRetain << "\n");
2349 for (SmallPtrSet<Instruction *, 2>::const_iterator
2350 AI = ReleasesToMove.Calls.begin(),
2351 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
2352 Instruction *OrigRelease = *AI;
2353 Releases.erase(OrigRelease);
2354 DeadInsts.push_back(OrigRelease);
2355 DEBUG(dbgs() << "Deleting release: " << *OrigRelease << "\n");
2361 ObjCARCOpt::ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState>
2363 MapVector<Value *, RRInfo> &Retains,
2364 DenseMap<Value *, RRInfo> &Releases,
2366 SmallVector<Instruction *, 4> &NewRetains,
2367 SmallVector<Instruction *, 4> &NewReleases,
2368 SmallVector<Instruction *, 8> &DeadInsts,
2369 RRInfo &RetainsToMove,
2370 RRInfo &ReleasesToMove,
2373 bool &AnyPairsCompletelyEliminated) {
2374 // If a pair happens in a region where it is known that the reference count
2375 // is already incremented, we can similarly ignore possible decrements.
2376 bool KnownSafeTD = true, KnownSafeBU = true;
2378 // Connect the dots between the top-down-collected RetainsToMove and
2379 // bottom-up-collected ReleasesToMove to form sets of related calls.
2380 // This is an iterative process so that we connect multiple releases
2381 // to multiple retains if needed.
2382 unsigned OldDelta = 0;
2383 unsigned NewDelta = 0;
2384 unsigned OldCount = 0;
2385 unsigned NewCount = 0;
2386 bool FirstRelease = true;
2388 for (SmallVectorImpl<Instruction *>::const_iterator
2389 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
2390 Instruction *NewRetain = *NI;
2391 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
2392 assert(It != Retains.end());
2393 const RRInfo &NewRetainRRI = It->second;
2394 KnownSafeTD &= NewRetainRRI.KnownSafe;
2395 for (SmallPtrSet<Instruction *, 2>::const_iterator
2396 LI = NewRetainRRI.Calls.begin(),
2397 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
2398 Instruction *NewRetainRelease = *LI;
2399 DenseMap<Value *, RRInfo>::const_iterator Jt =
2400 Releases.find(NewRetainRelease);
2401 if (Jt == Releases.end())
2403 const RRInfo &NewRetainReleaseRRI = Jt->second;
2404 assert(NewRetainReleaseRRI.Calls.count(NewRetain));
2405 if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
2407 BBStates[NewRetainRelease->getParent()].GetAllPathCount();
2409 // Merge the ReleaseMetadata and IsTailCallRelease values.
2411 ReleasesToMove.ReleaseMetadata =
2412 NewRetainReleaseRRI.ReleaseMetadata;
2413 ReleasesToMove.IsTailCallRelease =
2414 NewRetainReleaseRRI.IsTailCallRelease;
2415 FirstRelease = false;
2417 if (ReleasesToMove.ReleaseMetadata !=
2418 NewRetainReleaseRRI.ReleaseMetadata)
2419 ReleasesToMove.ReleaseMetadata = 0;
2420 if (ReleasesToMove.IsTailCallRelease !=
2421 NewRetainReleaseRRI.IsTailCallRelease)
2422 ReleasesToMove.IsTailCallRelease = false;
2425 // Collect the optimal insertion points.
2427 for (SmallPtrSet<Instruction *, 2>::const_iterator
2428 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
2429 RE = NewRetainReleaseRRI.ReverseInsertPts.end();
2431 Instruction *RIP = *RI;
2432 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
2433 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
2435 NewReleases.push_back(NewRetainRelease);
2440 if (NewReleases.empty()) break;
2442 // Back the other way.
2443 for (SmallVectorImpl<Instruction *>::const_iterator
2444 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
2445 Instruction *NewRelease = *NI;
2446 DenseMap<Value *, RRInfo>::const_iterator It =
2447 Releases.find(NewRelease);
2448 assert(It != Releases.end());
2449 const RRInfo &NewReleaseRRI = It->second;
2450 KnownSafeBU &= NewReleaseRRI.KnownSafe;
2451 for (SmallPtrSet<Instruction *, 2>::const_iterator
2452 LI = NewReleaseRRI.Calls.begin(),
2453 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
2454 Instruction *NewReleaseRetain = *LI;
2455 MapVector<Value *, RRInfo>::const_iterator Jt =
2456 Retains.find(NewReleaseRetain);
2457 if (Jt == Retains.end())
2459 const RRInfo &NewReleaseRetainRRI = Jt->second;
2460 assert(NewReleaseRetainRRI.Calls.count(NewRelease));
2461 if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
2462 unsigned PathCount =
2463 BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
2464 OldDelta += PathCount;
2465 OldCount += PathCount;
2467 // Collect the optimal insertion points.
2469 for (SmallPtrSet<Instruction *, 2>::const_iterator
2470 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
2471 RE = NewReleaseRetainRRI.ReverseInsertPts.end();
2473 Instruction *RIP = *RI;
2474 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
2475 PathCount = BBStates[RIP->getParent()].GetAllPathCount();
2476 NewDelta += PathCount;
2477 NewCount += PathCount;
2480 NewRetains.push_back(NewReleaseRetain);
2484 NewReleases.clear();
2485 if (NewRetains.empty()) break;
2488 // If the pointer is known incremented or nested, we can safely delete the
2489 // pair regardless of what's between them.
2490 if (KnownSafeTD || KnownSafeBU) {
2491 RetainsToMove.ReverseInsertPts.clear();
2492 ReleasesToMove.ReverseInsertPts.clear();
2495 // Determine whether the new insertion points we computed preserve the
2496 // balance of retain and release calls through the program.
2497 // TODO: If the fully aggressive solution isn't valid, try to find a
2498 // less aggressive solution which is.
2503 // Determine whether the original call points are balanced in the retain and
2504 // release calls through the program. If not, conservatively don't touch
2506 // TODO: It's theoretically possible to do code motion in this case, as
2507 // long as the existing imbalances are maintained.
2512 assert(OldCount != 0 && "Unreachable code?");
2513 NumRRs += OldCount - NewCount;
2514 // Set to true if we completely removed any RR pairs.
2515 AnyPairsCompletelyEliminated = NewCount == 0;
2517 // We can move calls!
2521 /// Identify pairings between the retains and releases, and delete and/or move
2524 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
2526 MapVector<Value *, RRInfo> &Retains,
2527 DenseMap<Value *, RRInfo> &Releases,
2529 DEBUG(dbgs() << "\n== ObjCARCOpt::PerformCodePlacement ==\n");
2531 bool AnyPairsCompletelyEliminated = false;
2532 RRInfo RetainsToMove;
2533 RRInfo ReleasesToMove;
2534 SmallVector<Instruction *, 4> NewRetains;
2535 SmallVector<Instruction *, 4> NewReleases;
2536 SmallVector<Instruction *, 8> DeadInsts;
2538 // Visit each retain.
2539 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
2540 E = Retains.end(); I != E; ++I) {
2541 Value *V = I->first;
2542 if (!V) continue; // blotted
2544 Instruction *Retain = cast<Instruction>(V);
2546 DEBUG(dbgs() << "Visiting: " << *Retain << "\n");
2548 Value *Arg = GetObjCArg(Retain);
2550 // If the object being released is in static or stack storage, we know it's
2551 // not being managed by ObjC reference counting, so we can delete pairs
2552 // regardless of what possible decrements or uses lie between them.
2553 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg);
2555 // A constant pointer can't be pointing to an object on the heap. It may
2556 // be reference-counted, but it won't be deleted.
2557 if (const LoadInst *LI = dyn_cast<LoadInst>(Arg))
2558 if (const GlobalVariable *GV =
2559 dyn_cast<GlobalVariable>(
2560 StripPointerCastsAndObjCCalls(LI->getPointerOperand())))
2561 if (GV->isConstant())
2564 // Connect the dots between the top-down-collected RetainsToMove and
2565 // bottom-up-collected ReleasesToMove to form sets of related calls.
2566 NewRetains.push_back(Retain);
2567 bool PerformMoveCalls =
2568 ConnectTDBUTraversals(BBStates, Retains, Releases, M, NewRetains,
2569 NewReleases, DeadInsts, RetainsToMove,
2570 ReleasesToMove, Arg, KnownSafe,
2571 AnyPairsCompletelyEliminated);
2573 #ifdef ARC_ANNOTATIONS
2574 // Do not move calls if ARC annotations are requested. If we were to move
2575 // calls in this case, we would not be able
2576 PerformMoveCalls = PerformMoveCalls && !EnableARCAnnotations;
2577 #endif // ARC_ANNOTATIONS
2579 if (PerformMoveCalls) {
2580 // Ok, everything checks out and we're all set. Let's move/delete some
2582 MoveCalls(Arg, RetainsToMove, ReleasesToMove,
2583 Retains, Releases, DeadInsts, M);
2586 // Clean up state for next retain.
2587 NewReleases.clear();
2589 RetainsToMove.clear();
2590 ReleasesToMove.clear();
2593 // Now that we're done moving everything, we can delete the newly dead
2594 // instructions, as we no longer need them as insert points.
2595 while (!DeadInsts.empty())
2596 EraseInstruction(DeadInsts.pop_back_val());
2598 return AnyPairsCompletelyEliminated;
2601 /// Weak pointer optimizations.
2602 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
2603 DEBUG(dbgs() << "\n== ObjCARCOpt::OptimizeWeakCalls ==\n");
2605 // First, do memdep-style RLE and S2L optimizations. We can't use memdep
2606 // itself because it uses AliasAnalysis and we need to do provenance
2608 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2609 Instruction *Inst = &*I++;
2611 DEBUG(dbgs() << "Visiting: " << *Inst << "\n");
2613 InstructionClass Class = GetBasicInstructionClass(Inst);
2614 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
2617 // Delete objc_loadWeak calls with no users.
2618 if (Class == IC_LoadWeak && Inst->use_empty()) {
2619 Inst->eraseFromParent();
2623 // TODO: For now, just look for an earlier available version of this value
2624 // within the same block. Theoretically, we could do memdep-style non-local
2625 // analysis too, but that would want caching. A better approach would be to
2626 // use the technique that EarlyCSE uses.
2627 inst_iterator Current = llvm::prior(I);
2628 BasicBlock *CurrentBB = Current.getBasicBlockIterator();
2629 for (BasicBlock::iterator B = CurrentBB->begin(),
2630 J = Current.getInstructionIterator();
2632 Instruction *EarlierInst = &*llvm::prior(J);
2633 InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
2634 switch (EarlierClass) {
2636 case IC_LoadWeakRetained: {
2637 // If this is loading from the same pointer, replace this load's value
2639 CallInst *Call = cast<CallInst>(Inst);
2640 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2641 Value *Arg = Call->getArgOperand(0);
2642 Value *EarlierArg = EarlierCall->getArgOperand(0);
2643 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2644 case AliasAnalysis::MustAlias:
2646 // If the load has a builtin retain, insert a plain retain for it.
2647 if (Class == IC_LoadWeakRetained) {
2649 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2653 // Zap the fully redundant load.
2654 Call->replaceAllUsesWith(EarlierCall);
2655 Call->eraseFromParent();
2657 case AliasAnalysis::MayAlias:
2658 case AliasAnalysis::PartialAlias:
2660 case AliasAnalysis::NoAlias:
2667 // If this is storing to the same pointer and has the same size etc.
2668 // replace this load's value with the stored value.
2669 CallInst *Call = cast<CallInst>(Inst);
2670 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2671 Value *Arg = Call->getArgOperand(0);
2672 Value *EarlierArg = EarlierCall->getArgOperand(0);
2673 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2674 case AliasAnalysis::MustAlias:
2676 // If the load has a builtin retain, insert a plain retain for it.
2677 if (Class == IC_LoadWeakRetained) {
2679 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2683 // Zap the fully redundant load.
2684 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
2685 Call->eraseFromParent();
2687 case AliasAnalysis::MayAlias:
2688 case AliasAnalysis::PartialAlias:
2690 case AliasAnalysis::NoAlias:
2697 // TOOD: Grab the copied value.
2699 case IC_AutoreleasepoolPush:
2701 case IC_IntrinsicUser:
2703 // Weak pointers are only modified through the weak entry points
2704 // (and arbitrary calls, which could call the weak entry points).
2707 // Anything else could modify the weak pointer.
2714 // Then, for each destroyWeak with an alloca operand, check to see if
2715 // the alloca and all its users can be zapped.
2716 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2717 Instruction *Inst = &*I++;
2718 InstructionClass Class = GetBasicInstructionClass(Inst);
2719 if (Class != IC_DestroyWeak)
2722 CallInst *Call = cast<CallInst>(Inst);
2723 Value *Arg = Call->getArgOperand(0);
2724 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
2725 for (Value::use_iterator UI = Alloca->use_begin(),
2726 UE = Alloca->use_end(); UI != UE; ++UI) {
2727 const Instruction *UserInst = cast<Instruction>(*UI);
2728 switch (GetBasicInstructionClass(UserInst)) {
2731 case IC_DestroyWeak:
2738 for (Value::use_iterator UI = Alloca->use_begin(),
2739 UE = Alloca->use_end(); UI != UE; ) {
2740 CallInst *UserInst = cast<CallInst>(*UI++);
2741 switch (GetBasicInstructionClass(UserInst)) {
2744 // These functions return their second argument.
2745 UserInst->replaceAllUsesWith(UserInst->getArgOperand(1));
2747 case IC_DestroyWeak:
2751 llvm_unreachable("alloca really is used!");
2753 UserInst->eraseFromParent();
2755 Alloca->eraseFromParent();
2761 /// Identify program paths which execute sequences of retains and releases which
2762 /// can be eliminated.
2763 bool ObjCARCOpt::OptimizeSequences(Function &F) {
2764 /// Releases, Retains - These are used to store the results of the main flow
2765 /// analysis. These use Value* as the key instead of Instruction* so that the
2766 /// map stays valid when we get around to rewriting code and calls get
2767 /// replaced by arguments.
2768 DenseMap<Value *, RRInfo> Releases;
2769 MapVector<Value *, RRInfo> Retains;
2771 /// This is used during the traversal of the function to track the
2772 /// states for each identified object at each block.
2773 DenseMap<const BasicBlock *, BBState> BBStates;
2775 // Analyze the CFG of the function, and all instructions.
2776 bool NestingDetected = Visit(F, BBStates, Retains, Releases);
2779 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
2783 /// Check if there is a dependent call earlier that does not have anything in
2784 /// between the Retain and the call that can affect the reference count of their
2785 /// shared pointer argument. Note that Retain need not be in BB.
2787 HasSafePathToPredecessorCall(const Value *Arg, Instruction *Retain,
2788 SmallPtrSet<Instruction *, 4> &DepInsts,
2789 SmallPtrSet<const BasicBlock *, 4> &Visited,
2790 ProvenanceAnalysis &PA) {
2791 FindDependencies(CanChangeRetainCount, Arg, Retain->getParent(), Retain,
2792 DepInsts, Visited, PA);
2793 if (DepInsts.size() != 1)
2797 dyn_cast_or_null<CallInst>(*DepInsts.begin());
2799 // Check that the pointer is the return value of the call.
2800 if (!Call || Arg != Call)
2803 // Check that the call is a regular call.
2804 InstructionClass Class = GetBasicInstructionClass(Call);
2805 if (Class != IC_CallOrUser && Class != IC_Call)
2811 /// Find a dependent retain that precedes the given autorelease for which there
2812 /// is nothing in between the two instructions that can affect the ref count of
2815 FindPredecessorRetainWithSafePath(const Value *Arg, BasicBlock *BB,
2816 Instruction *Autorelease,
2817 SmallPtrSet<Instruction *, 4> &DepInsts,
2818 SmallPtrSet<const BasicBlock *, 4> &Visited,
2819 ProvenanceAnalysis &PA) {
2820 FindDependencies(CanChangeRetainCount, Arg,
2821 BB, Autorelease, DepInsts, Visited, PA);
2822 if (DepInsts.size() != 1)
2826 dyn_cast_or_null<CallInst>(*DepInsts.begin());
2828 // Check that we found a retain with the same argument.
2830 !IsRetain(GetBasicInstructionClass(Retain)) ||
2831 GetObjCArg(Retain) != Arg) {
2838 /// Look for an ``autorelease'' instruction dependent on Arg such that there are
2839 /// no instructions dependent on Arg that need a positive ref count in between
2840 /// the autorelease and the ret.
2842 FindPredecessorAutoreleaseWithSafePath(const Value *Arg, BasicBlock *BB,
2844 SmallPtrSet<Instruction *, 4> &DepInsts,
2845 SmallPtrSet<const BasicBlock *, 4> &V,
2846 ProvenanceAnalysis &PA) {
2847 FindDependencies(NeedsPositiveRetainCount, Arg,
2848 BB, Ret, DepInsts, V, PA);
2849 if (DepInsts.size() != 1)
2852 CallInst *Autorelease =
2853 dyn_cast_or_null<CallInst>(*DepInsts.begin());
2856 InstructionClass AutoreleaseClass = GetBasicInstructionClass(Autorelease);
2857 if (!IsAutorelease(AutoreleaseClass))
2859 if (GetObjCArg(Autorelease) != Arg)
2865 /// Look for this pattern:
2867 /// %call = call i8* @something(...)
2868 /// %2 = call i8* @objc_retain(i8* %call)
2869 /// %3 = call i8* @objc_autorelease(i8* %2)
2872 /// And delete the retain and autorelease.
2873 void ObjCARCOpt::OptimizeReturns(Function &F) {
2874 if (!F.getReturnType()->isPointerTy())
2877 DEBUG(dbgs() << "\n== ObjCARCOpt::OptimizeReturns ==\n");
2879 SmallPtrSet<Instruction *, 4> DependingInstructions;
2880 SmallPtrSet<const BasicBlock *, 4> Visited;
2881 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
2882 BasicBlock *BB = FI;
2883 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
2885 DEBUG(dbgs() << "Visiting: " << *Ret << "\n");
2890 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
2892 // Look for an ``autorelease'' instruction that is a predecssor of Ret and
2893 // dependent on Arg such that there are no instructions dependent on Arg
2894 // that need a positive ref count in between the autorelease and Ret.
2895 CallInst *Autorelease =
2896 FindPredecessorAutoreleaseWithSafePath(Arg, BB, Ret,
2897 DependingInstructions, Visited,
2900 DependingInstructions.clear();
2904 FindPredecessorRetainWithSafePath(Arg, BB, Autorelease,
2905 DependingInstructions, Visited, PA);
2907 DependingInstructions.clear();
2910 // Check that there is nothing that can affect the reference count
2911 // between the retain and the call. Note that Retain need not be in BB.
2912 if (HasSafePathToPredecessorCall(Arg, Retain, DependingInstructions,
2914 // If so, we can zap the retain and autorelease.
2917 DEBUG(dbgs() << "Erasing: " << *Retain << "\nErasing: "
2918 << *Autorelease << "\n");
2919 EraseInstruction(Retain);
2920 EraseInstruction(Autorelease);
2925 DependingInstructions.clear();
2930 bool ObjCARCOpt::doInitialization(Module &M) {
2934 // If nothing in the Module uses ARC, don't do anything.
2935 Run = ModuleHasARC(M);
2939 // Identify the imprecise release metadata kind.
2940 ImpreciseReleaseMDKind =
2941 M.getContext().getMDKindID("clang.imprecise_release");
2942 CopyOnEscapeMDKind =
2943 M.getContext().getMDKindID("clang.arc.copy_on_escape");
2944 NoObjCARCExceptionsMDKind =
2945 M.getContext().getMDKindID("clang.arc.no_objc_arc_exceptions");
2946 #ifdef ARC_ANNOTATIONS
2947 ARCAnnotationBottomUpMDKind =
2948 M.getContext().getMDKindID("llvm.arc.annotation.bottomup");
2949 ARCAnnotationTopDownMDKind =
2950 M.getContext().getMDKindID("llvm.arc.annotation.topdown");
2951 ARCAnnotationProvenanceSourceMDKind =
2952 M.getContext().getMDKindID("llvm.arc.annotation.provenancesource");
2953 #endif // ARC_ANNOTATIONS
2955 // Intuitively, objc_retain and others are nocapture, however in practice
2956 // they are not, because they return their argument value. And objc_release
2957 // calls finalizers which can have arbitrary side effects.
2959 // These are initialized lazily.
2961 AutoreleaseRVCallee = 0;
2964 RetainBlockCallee = 0;
2965 AutoreleaseCallee = 0;
2970 bool ObjCARCOpt::runOnFunction(Function &F) {
2974 // If nothing in the Module uses ARC, don't do anything.
2980 DEBUG(dbgs() << "<<< ObjCARCOpt: Visiting Function: " << F.getName() << " >>>"
2983 PA.setAA(&getAnalysis<AliasAnalysis>());
2985 // This pass performs several distinct transformations. As a compile-time aid
2986 // when compiling code that isn't ObjC, skip these if the relevant ObjC
2987 // library functions aren't declared.
2989 // Preliminary optimizations. This also computs UsedInThisFunction.
2990 OptimizeIndividualCalls(F);
2992 // Optimizations for weak pointers.
2993 if (UsedInThisFunction & ((1 << IC_LoadWeak) |
2994 (1 << IC_LoadWeakRetained) |
2995 (1 << IC_StoreWeak) |
2996 (1 << IC_InitWeak) |
2997 (1 << IC_CopyWeak) |
2998 (1 << IC_MoveWeak) |
2999 (1 << IC_DestroyWeak)))
3000 OptimizeWeakCalls(F);
3002 // Optimizations for retain+release pairs.
3003 if (UsedInThisFunction & ((1 << IC_Retain) |
3004 (1 << IC_RetainRV) |
3005 (1 << IC_RetainBlock)))
3006 if (UsedInThisFunction & (1 << IC_Release))
3007 // Run OptimizeSequences until it either stops making changes or
3008 // no retain+release pair nesting is detected.
3009 while (OptimizeSequences(F)) {}
3011 // Optimizations if objc_autorelease is used.
3012 if (UsedInThisFunction & ((1 << IC_Autorelease) |
3013 (1 << IC_AutoreleaseRV)))
3016 DEBUG(dbgs() << "\n");
3021 void ObjCARCOpt::releaseMemory() {