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 /// reverse 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 if 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 EnableCheckForCFGHazards("enable-objc-arc-checkforcfghazards",
717 /// This function appends a unique ARCAnnotationProvenanceSourceMDKind id to an
718 /// instruction so that we can track backwards when post processing via the llvm
719 /// arc annotation processor tool. If the function is an
720 static MDString *AppendMDNodeToSourcePtr(unsigned NodeId,
724 // If pointer is a result of an instruction and it does not have a source
725 // MDNode it, attach a new MDNode onto it. If pointer is a result of
726 // an instruction and does have a source MDNode attached to it, return a
727 // reference to said Node. Otherwise just return 0.
728 if (Instruction *Inst = dyn_cast<Instruction>(Ptr)) {
730 if (!(Node = Inst->getMetadata(NodeId))) {
731 // We do not have any node. Generate and attatch the hash MDString to the
734 // We just use an MDString to ensure that this metadata gets written out
735 // of line at the module level and to provide a very simple format
736 // encoding the information herein. Both of these makes it simpler to
737 // parse the annotations by a simple external program.
739 raw_string_ostream os(Str);
740 os << "(" << Inst->getParent()->getParent()->getName() << ",%"
741 << Inst->getName() << ")";
743 Hash = MDString::get(Inst->getContext(), os.str());
744 Inst->setMetadata(NodeId, MDNode::get(Inst->getContext(),Hash));
746 // We have a node. Grab its hash and return it.
747 assert(Node->getNumOperands() == 1 &&
748 "An ARCAnnotationProvenanceSourceMDKind can only have 1 operand.");
749 Hash = cast<MDString>(Node->getOperand(0));
751 } else if (Argument *Arg = dyn_cast<Argument>(Ptr)) {
753 raw_string_ostream os(str);
754 os << "(" << Arg->getParent()->getName() << ",%" << Arg->getName()
756 Hash = MDString::get(Arg->getContext(), os.str());
762 static std::string SequenceToString(Sequence A) {
764 raw_string_ostream os(str);
769 /// Helper function to change a Sequence into a String object using our overload
770 /// for raw_ostream so we only have printing code in one location.
771 static MDString *SequenceToMDString(LLVMContext &Context,
773 return MDString::get(Context, SequenceToString(A));
776 /// A simple function to generate a MDNode which describes the change in state
777 /// for Value *Ptr caused by Instruction *Inst.
778 static void AppendMDNodeToInstForPtr(unsigned NodeId,
781 MDString *PtrSourceMDNodeID,
785 Value *tmp[3] = {PtrSourceMDNodeID,
786 SequenceToMDString(Inst->getContext(),
788 SequenceToMDString(Inst->getContext(),
790 Node = MDNode::get(Inst->getContext(),
791 ArrayRef<Value*>(tmp, 3));
793 Inst->setMetadata(NodeId, Node);
796 /// Add to the beginning of the basic block llvm.ptr.annotations which show the
797 /// state of a pointer at the entrance to a basic block.
798 static void GenerateARCBBEntranceAnnotation(const char *Name, BasicBlock *BB,
799 Value *Ptr, Sequence Seq) {
800 Module *M = BB->getParent()->getParent();
801 LLVMContext &C = M->getContext();
802 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
803 Type *I8XX = PointerType::getUnqual(I8X);
804 Type *Params[] = {I8XX, I8XX};
805 FunctionType *FTy = FunctionType::get(Type::getVoidTy(C),
806 ArrayRef<Type*>(Params, 2),
808 Constant *Callee = M->getOrInsertFunction(Name, FTy);
810 IRBuilder<> Builder(BB, BB->getFirstInsertionPt());
813 StringRef Tmp = Ptr->getName();
814 if (0 == (PtrName = M->getGlobalVariable(Tmp, true))) {
815 Value *ActualPtrName = Builder.CreateGlobalStringPtr(Tmp,
817 PtrName = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
818 cast<Constant>(ActualPtrName), Tmp);
822 std::string SeqStr = SequenceToString(Seq);
823 if (0 == (S = M->getGlobalVariable(SeqStr, true))) {
824 Value *ActualPtrName = Builder.CreateGlobalStringPtr(SeqStr,
826 S = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
827 cast<Constant>(ActualPtrName), SeqStr);
830 Builder.CreateCall2(Callee, PtrName, S);
833 /// Add to the end of the basic block llvm.ptr.annotations which show the state
834 /// of the pointer at the bottom of the basic block.
835 static void GenerateARCBBTerminatorAnnotation(const char *Name, BasicBlock *BB,
836 Value *Ptr, Sequence Seq) {
837 Module *M = BB->getParent()->getParent();
838 LLVMContext &C = M->getContext();
839 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
840 Type *I8XX = PointerType::getUnqual(I8X);
841 Type *Params[] = {I8XX, I8XX};
842 FunctionType *FTy = FunctionType::get(Type::getVoidTy(C),
843 ArrayRef<Type*>(Params, 2),
845 Constant *Callee = M->getOrInsertFunction(Name, FTy);
847 IRBuilder<> Builder(BB, llvm::prior(BB->end()));
850 StringRef Tmp = Ptr->getName();
851 if (0 == (PtrName = M->getGlobalVariable(Tmp, true))) {
852 Value *ActualPtrName = Builder.CreateGlobalStringPtr(Tmp,
854 PtrName = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
855 cast<Constant>(ActualPtrName), Tmp);
859 std::string SeqStr = SequenceToString(Seq);
860 if (0 == (S = M->getGlobalVariable(SeqStr, true))) {
861 Value *ActualPtrName = Builder.CreateGlobalStringPtr(SeqStr,
863 S = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
864 cast<Constant>(ActualPtrName), SeqStr);
866 Builder.CreateCall2(Callee, PtrName, S);
869 /// Adds a source annotation to pointer and a state change annotation to Inst
870 /// referencing the source annotation and the old/new state of pointer.
871 static void GenerateARCAnnotation(unsigned InstMDId,
877 if (EnableARCAnnotations) {
878 // First generate the source annotation on our pointer. This will return an
879 // MDString* if Ptr actually comes from an instruction implying we can put
880 // in a source annotation. If AppendMDNodeToSourcePtr returns 0 (i.e. NULL),
881 // then we know that our pointer is from an Argument so we put a reference
882 // to the argument number.
884 // The point of this is to make it easy for the
885 // llvm-arc-annotation-processor tool to cross reference where the source
886 // pointer is in the LLVM IR since the LLVM IR parser does not submit such
887 // information via debug info for backends to use (since why would anyone
888 // need such a thing from LLVM IR besides in non standard cases
890 MDString *SourcePtrMDNode =
891 AppendMDNodeToSourcePtr(PtrMDId, Ptr);
892 AppendMDNodeToInstForPtr(InstMDId, Inst, Ptr, SourcePtrMDNode, OldSeq,
897 // The actual interface for accessing the above functionality is defined via
898 // some simple macros which are defined below. We do this so that the user does
899 // not need to pass in what metadata id is needed resulting in cleaner code and
900 // additionally since it provides an easy way to conditionally no-op all
901 // annotation support in a non-debug build.
903 /// Use this macro to annotate a sequence state change when processing
904 /// instructions bottom up,
905 #define ANNOTATE_BOTTOMUP(inst, ptr, old, new) \
906 GenerateARCAnnotation(ARCAnnotationBottomUpMDKind, \
907 ARCAnnotationProvenanceSourceMDKind, (inst), \
908 const_cast<Value*>(ptr), (old), (new))
909 /// Use this macro to annotate a sequence state change when processing
910 /// instructions top down.
911 #define ANNOTATE_TOPDOWN(inst, ptr, old, new) \
912 GenerateARCAnnotation(ARCAnnotationTopDownMDKind, \
913 ARCAnnotationProvenanceSourceMDKind, (inst), \
914 const_cast<Value*>(ptr), (old), (new))
916 #define ANNOTATE_BB(_states, _bb, _name, _type, _direction) \
918 if (EnableARCAnnotations) { \
919 for(BBState::ptr_const_iterator I = (_states)._direction##_ptr_begin(), \
920 E = (_states)._direction##_ptr_end(); I != E; ++I) { \
921 Value *Ptr = const_cast<Value*>(I->first); \
922 Sequence Seq = I->second.GetSeq(); \
923 GenerateARCBB ## _type ## Annotation(_name, (_bb), Ptr, Seq); \
928 #define ANNOTATE_BOTTOMUP_BBSTART(_states, _basicblock) \
929 ANNOTATE_BB(_states, _basicblock, "llvm.arc.annotation.bottomup.bbstart", \
931 #define ANNOTATE_BOTTOMUP_BBEND(_states, _basicblock) \
932 ANNOTATE_BB(_states, _basicblock, "llvm.arc.annotation.bottomup.bbend", \
933 Terminator, bottom_up)
934 #define ANNOTATE_TOPDOWN_BBSTART(_states, _basicblock) \
935 ANNOTATE_BB(_states, _basicblock, "llvm.arc.annotation.topdown.bbstart", \
937 #define ANNOTATE_TOPDOWN_BBEND(_states, _basicblock) \
938 ANNOTATE_BB(_states, _basicblock, "llvm.arc.annotation.topdown.bbend", \
939 Terminator, top_down)
941 #else // !ARC_ANNOTATION
942 // If annotations are off, noop.
943 #define ANNOTATE_BOTTOMUP(inst, ptr, old, new)
944 #define ANNOTATE_TOPDOWN(inst, ptr, old, new)
945 #define ANNOTATE_BOTTOMUP_BBSTART(states, basicblock)
946 #define ANNOTATE_BOTTOMUP_BBEND(states, basicblock)
947 #define ANNOTATE_TOPDOWN_BBSTART(states, basicblock)
948 #define ANNOTATE_TOPDOWN_BBEND(states, basicblock)
949 #endif // !ARC_ANNOTATION
952 /// \brief The main ARC optimization pass.
953 class ObjCARCOpt : public FunctionPass {
955 ProvenanceAnalysis PA;
957 /// A flag indicating whether this optimization pass should run.
960 /// Declarations for ObjC runtime functions, for use in creating calls to
961 /// them. These are initialized lazily to avoid cluttering up the Module
962 /// with unused declarations.
964 /// Declaration for ObjC runtime function
965 /// objc_retainAutoreleasedReturnValue.
966 Constant *RetainRVCallee;
967 /// Declaration for ObjC runtime function objc_autoreleaseReturnValue.
968 Constant *AutoreleaseRVCallee;
969 /// Declaration for ObjC runtime function objc_release.
970 Constant *ReleaseCallee;
971 /// Declaration for ObjC runtime function objc_retain.
972 Constant *RetainCallee;
973 /// Declaration for ObjC runtime function objc_retainBlock.
974 Constant *RetainBlockCallee;
975 /// Declaration for ObjC runtime function objc_autorelease.
976 Constant *AutoreleaseCallee;
978 /// Flags which determine whether each of the interesting runtine functions
979 /// is in fact used in the current function.
980 unsigned UsedInThisFunction;
982 /// The Metadata Kind for clang.imprecise_release metadata.
983 unsigned ImpreciseReleaseMDKind;
985 /// The Metadata Kind for clang.arc.copy_on_escape metadata.
986 unsigned CopyOnEscapeMDKind;
988 /// The Metadata Kind for clang.arc.no_objc_arc_exceptions metadata.
989 unsigned NoObjCARCExceptionsMDKind;
991 #ifdef ARC_ANNOTATIONS
992 /// The Metadata Kind for llvm.arc.annotation.bottomup metadata.
993 unsigned ARCAnnotationBottomUpMDKind;
994 /// The Metadata Kind for llvm.arc.annotation.topdown metadata.
995 unsigned ARCAnnotationTopDownMDKind;
996 /// The Metadata Kind for llvm.arc.annotation.provenancesource metadata.
997 unsigned ARCAnnotationProvenanceSourceMDKind;
998 #endif // ARC_ANNOATIONS
1000 Constant *getRetainRVCallee(Module *M);
1001 Constant *getAutoreleaseRVCallee(Module *M);
1002 Constant *getReleaseCallee(Module *M);
1003 Constant *getRetainCallee(Module *M);
1004 Constant *getRetainBlockCallee(Module *M);
1005 Constant *getAutoreleaseCallee(Module *M);
1007 bool IsRetainBlockOptimizable(const Instruction *Inst);
1009 void OptimizeRetainCall(Function &F, Instruction *Retain);
1010 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
1011 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
1012 InstructionClass &Class);
1013 bool OptimizeRetainBlockCall(Function &F, Instruction *RetainBlock,
1014 InstructionClass &Class);
1015 void OptimizeIndividualCalls(Function &F);
1017 void CheckForCFGHazards(const BasicBlock *BB,
1018 DenseMap<const BasicBlock *, BBState> &BBStates,
1019 BBState &MyStates) const;
1020 bool VisitInstructionBottomUp(Instruction *Inst,
1022 MapVector<Value *, RRInfo> &Retains,
1024 bool VisitBottomUp(BasicBlock *BB,
1025 DenseMap<const BasicBlock *, BBState> &BBStates,
1026 MapVector<Value *, RRInfo> &Retains);
1027 bool VisitInstructionTopDown(Instruction *Inst,
1028 DenseMap<Value *, RRInfo> &Releases,
1030 bool VisitTopDown(BasicBlock *BB,
1031 DenseMap<const BasicBlock *, BBState> &BBStates,
1032 DenseMap<Value *, RRInfo> &Releases);
1033 bool Visit(Function &F,
1034 DenseMap<const BasicBlock *, BBState> &BBStates,
1035 MapVector<Value *, RRInfo> &Retains,
1036 DenseMap<Value *, RRInfo> &Releases);
1038 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1039 MapVector<Value *, RRInfo> &Retains,
1040 DenseMap<Value *, RRInfo> &Releases,
1041 SmallVectorImpl<Instruction *> &DeadInsts,
1044 bool ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState> &BBStates,
1045 MapVector<Value *, RRInfo> &Retains,
1046 DenseMap<Value *, RRInfo> &Releases,
1048 SmallVector<Instruction *, 4> &NewRetains,
1049 SmallVector<Instruction *, 4> &NewReleases,
1050 SmallVector<Instruction *, 8> &DeadInsts,
1051 RRInfo &RetainsToMove,
1052 RRInfo &ReleasesToMove,
1055 bool &AnyPairsCompletelyEliminated);
1057 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1058 MapVector<Value *, RRInfo> &Retains,
1059 DenseMap<Value *, RRInfo> &Releases,
1062 void OptimizeWeakCalls(Function &F);
1064 bool OptimizeSequences(Function &F);
1066 void OptimizeReturns(Function &F);
1068 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1069 virtual bool doInitialization(Module &M);
1070 virtual bool runOnFunction(Function &F);
1071 virtual void releaseMemory();
1075 ObjCARCOpt() : FunctionPass(ID) {
1076 initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1081 char ObjCARCOpt::ID = 0;
1082 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1083 "objc-arc", "ObjC ARC optimization", false, false)
1084 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1085 INITIALIZE_PASS_END(ObjCARCOpt,
1086 "objc-arc", "ObjC ARC optimization", false, false)
1088 Pass *llvm::createObjCARCOptPass() {
1089 return new ObjCARCOpt();
1092 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1093 AU.addRequired<ObjCARCAliasAnalysis>();
1094 AU.addRequired<AliasAnalysis>();
1095 // ARC optimization doesn't currently split critical edges.
1096 AU.setPreservesCFG();
1099 bool ObjCARCOpt::IsRetainBlockOptimizable(const Instruction *Inst) {
1100 // Without the magic metadata tag, we have to assume this might be an
1101 // objc_retainBlock call inserted to convert a block pointer to an id,
1102 // in which case it really is needed.
1103 if (!Inst->getMetadata(CopyOnEscapeMDKind))
1106 // If the pointer "escapes" (not including being used in a call),
1107 // the copy may be needed.
1108 if (DoesRetainableObjPtrEscape(Inst))
1111 // Otherwise, it's not needed.
1115 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1116 if (!RetainRVCallee) {
1117 LLVMContext &C = M->getContext();
1118 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1119 Type *Params[] = { I8X };
1120 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1121 AttributeSet Attribute =
1122 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1123 Attribute::NoUnwind);
1125 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1128 return RetainRVCallee;
1131 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1132 if (!AutoreleaseRVCallee) {
1133 LLVMContext &C = M->getContext();
1134 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1135 Type *Params[] = { I8X };
1136 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1137 AttributeSet Attribute =
1138 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1139 Attribute::NoUnwind);
1140 AutoreleaseRVCallee =
1141 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1144 return AutoreleaseRVCallee;
1147 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1148 if (!ReleaseCallee) {
1149 LLVMContext &C = M->getContext();
1150 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1151 AttributeSet Attribute =
1152 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1153 Attribute::NoUnwind);
1155 M->getOrInsertFunction(
1157 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1160 return ReleaseCallee;
1163 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1164 if (!RetainCallee) {
1165 LLVMContext &C = M->getContext();
1166 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1167 AttributeSet Attribute =
1168 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1169 Attribute::NoUnwind);
1171 M->getOrInsertFunction(
1173 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1176 return RetainCallee;
1179 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1180 if (!RetainBlockCallee) {
1181 LLVMContext &C = M->getContext();
1182 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1183 // objc_retainBlock is not nounwind because it calls user copy constructors
1184 // which could theoretically throw.
1186 M->getOrInsertFunction(
1188 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1191 return RetainBlockCallee;
1194 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1195 if (!AutoreleaseCallee) {
1196 LLVMContext &C = M->getContext();
1197 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1198 AttributeSet Attribute =
1199 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1200 Attribute::NoUnwind);
1202 M->getOrInsertFunction(
1204 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1207 return AutoreleaseCallee;
1210 /// Turn objc_retain into objc_retainAutoreleasedReturnValue if the operand is a
1213 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
1214 ImmutableCallSite CS(GetObjCArg(Retain));
1215 const Instruction *Call = CS.getInstruction();
1217 if (Call->getParent() != Retain->getParent()) return;
1219 // Check that the call is next to the retain.
1220 BasicBlock::const_iterator I = Call;
1222 while (IsNoopInstruction(I)) ++I;
1226 // Turn it to an objc_retainAutoreleasedReturnValue..
1230 DEBUG(dbgs() << "Transforming objc_retain => "
1231 "objc_retainAutoreleasedReturnValue since the operand is a "
1232 "return value.\nOld: "<< *Retain << "\n");
1234 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
1236 DEBUG(dbgs() << "New: " << *Retain << "\n");
1239 /// Turn objc_retainAutoreleasedReturnValue into objc_retain if the operand is
1240 /// not a return value. Or, if it can be paired with an
1241 /// objc_autoreleaseReturnValue, delete the pair and return true.
1243 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
1244 // Check for the argument being from an immediately preceding call or invoke.
1245 const Value *Arg = GetObjCArg(RetainRV);
1246 ImmutableCallSite CS(Arg);
1247 if (const Instruction *Call = CS.getInstruction()) {
1248 if (Call->getParent() == RetainRV->getParent()) {
1249 BasicBlock::const_iterator I = Call;
1251 while (IsNoopInstruction(I)) ++I;
1252 if (&*I == RetainRV)
1254 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
1255 BasicBlock *RetainRVParent = RetainRV->getParent();
1256 if (II->getNormalDest() == RetainRVParent) {
1257 BasicBlock::const_iterator I = RetainRVParent->begin();
1258 while (IsNoopInstruction(I)) ++I;
1259 if (&*I == RetainRV)
1265 // Check for being preceded by an objc_autoreleaseReturnValue on the same
1266 // pointer. In this case, we can delete the pair.
1267 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
1269 do --I; while (I != Begin && IsNoopInstruction(I));
1270 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
1271 GetObjCArg(I) == Arg) {
1275 DEBUG(dbgs() << "Erasing autoreleaseRV,retainRV pair: " << *I << "\n"
1276 << "Erasing " << *RetainRV << "\n");
1278 EraseInstruction(I);
1279 EraseInstruction(RetainRV);
1284 // Turn it to a plain objc_retain.
1288 DEBUG(dbgs() << "Transforming objc_retainAutoreleasedReturnValue => "
1289 "objc_retain since the operand is not a return value.\n"
1290 "Old = " << *RetainRV << "\n");
1292 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
1294 DEBUG(dbgs() << "New = " << *RetainRV << "\n");
1299 /// Turn objc_autoreleaseReturnValue into objc_autorelease if the result is not
1300 /// used as a return value.
1302 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
1303 InstructionClass &Class) {
1304 // Check for a return of the pointer value.
1305 const Value *Ptr = GetObjCArg(AutoreleaseRV);
1306 SmallVector<const Value *, 2> Users;
1307 Users.push_back(Ptr);
1309 Ptr = Users.pop_back_val();
1310 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
1312 const User *I = *UI;
1313 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
1315 if (isa<BitCastInst>(I))
1318 } while (!Users.empty());
1323 DEBUG(dbgs() << "Transforming objc_autoreleaseReturnValue => "
1324 "objc_autorelease since its operand is not used as a return "
1326 "Old = " << *AutoreleaseRV << "\n");
1328 CallInst *AutoreleaseRVCI = cast<CallInst>(AutoreleaseRV);
1330 setCalledFunction(getAutoreleaseCallee(F.getParent()));
1331 AutoreleaseRVCI->setTailCall(false); // Never tail call objc_autorelease.
1332 Class = IC_Autorelease;
1334 DEBUG(dbgs() << "New: " << *AutoreleaseRV << "\n");
1338 // \brief Attempt to strength reduce objc_retainBlock calls to objc_retain
1341 // Specifically: If an objc_retainBlock call has the copy_on_escape metadata and
1342 // does not escape (following the rules of block escaping), strength reduce the
1343 // objc_retainBlock to an objc_retain.
1345 // TODO: If an objc_retainBlock call is dominated period by a previous
1346 // objc_retainBlock call, strength reduce the objc_retainBlock to an
1349 ObjCARCOpt::OptimizeRetainBlockCall(Function &F, Instruction *Inst,
1350 InstructionClass &Class) {
1351 assert(GetBasicInstructionClass(Inst) == Class);
1352 assert(IC_RetainBlock == Class);
1354 // If we can not optimize Inst, return false.
1355 if (!IsRetainBlockOptimizable(Inst))
1358 CallInst *RetainBlock = cast<CallInst>(Inst);
1359 RetainBlock->setCalledFunction(getRetainCallee(F.getParent()));
1360 // Remove copy_on_escape metadata.
1361 RetainBlock->setMetadata(CopyOnEscapeMDKind, 0);
1367 /// Visit each call, one at a time, and make simplifications without doing any
1368 /// additional analysis.
1369 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
1370 DEBUG(dbgs() << "\n== ObjCARCOpt::OptimizeIndividualCalls ==\n");
1371 // Reset all the flags in preparation for recomputing them.
1372 UsedInThisFunction = 0;
1374 // Visit all objc_* calls in F.
1375 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
1376 Instruction *Inst = &*I++;
1378 InstructionClass Class = GetBasicInstructionClass(Inst);
1380 DEBUG(dbgs() << "Visiting: Class: " << Class << "; " << *Inst << "\n");
1385 // Delete no-op casts. These function calls have special semantics, but
1386 // the semantics are entirely implemented via lowering in the front-end,
1387 // so by the time they reach the optimizer, they are just no-op calls
1388 // which return their argument.
1390 // There are gray areas here, as the ability to cast reference-counted
1391 // pointers to raw void* and back allows code to break ARC assumptions,
1392 // however these are currently considered to be unimportant.
1396 DEBUG(dbgs() << "Erasing no-op cast: " << *Inst << "\n");
1397 EraseInstruction(Inst);
1400 // If the pointer-to-weak-pointer is null, it's undefined behavior.
1403 case IC_LoadWeakRetained:
1405 case IC_DestroyWeak: {
1406 CallInst *CI = cast<CallInst>(Inst);
1407 if (IsNullOrUndef(CI->getArgOperand(0))) {
1409 Type *Ty = CI->getArgOperand(0)->getType();
1410 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1411 Constant::getNullValue(Ty),
1413 llvm::Value *NewValue = UndefValue::get(CI->getType());
1414 DEBUG(dbgs() << "A null pointer-to-weak-pointer is undefined behavior."
1415 "\nOld = " << *CI << "\nNew = " << *NewValue << "\n");
1416 CI->replaceAllUsesWith(NewValue);
1417 CI->eraseFromParent();
1424 CallInst *CI = cast<CallInst>(Inst);
1425 if (IsNullOrUndef(CI->getArgOperand(0)) ||
1426 IsNullOrUndef(CI->getArgOperand(1))) {
1428 Type *Ty = CI->getArgOperand(0)->getType();
1429 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1430 Constant::getNullValue(Ty),
1433 llvm::Value *NewValue = UndefValue::get(CI->getType());
1434 DEBUG(dbgs() << "A null pointer-to-weak-pointer is undefined behavior."
1435 "\nOld = " << *CI << "\nNew = " << *NewValue << "\n");
1437 CI->replaceAllUsesWith(NewValue);
1438 CI->eraseFromParent();
1443 case IC_RetainBlock:
1444 // If we strength reduce an objc_retainBlock to amn objc_retain, continue
1445 // onto the objc_retain peephole optimizations. Otherwise break.
1446 if (!OptimizeRetainBlockCall(F, Inst, Class))
1450 OptimizeRetainCall(F, Inst);
1453 if (OptimizeRetainRVCall(F, Inst))
1456 case IC_AutoreleaseRV:
1457 OptimizeAutoreleaseRVCall(F, Inst, Class);
1461 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
1462 if (IsAutorelease(Class) && Inst->use_empty()) {
1463 CallInst *Call = cast<CallInst>(Inst);
1464 const Value *Arg = Call->getArgOperand(0);
1465 Arg = FindSingleUseIdentifiedObject(Arg);
1470 // Create the declaration lazily.
1471 LLVMContext &C = Inst->getContext();
1473 CallInst::Create(getReleaseCallee(F.getParent()),
1474 Call->getArgOperand(0), "", Call);
1475 NewCall->setMetadata(ImpreciseReleaseMDKind,
1476 MDNode::get(C, ArrayRef<Value *>()));
1478 DEBUG(dbgs() << "Replacing autorelease{,RV}(x) with objc_release(x) "
1479 "since x is otherwise unused.\nOld: " << *Call << "\nNew: "
1480 << *NewCall << "\n");
1482 EraseInstruction(Call);
1488 // For functions which can never be passed stack arguments, add
1490 if (IsAlwaysTail(Class)) {
1492 DEBUG(dbgs() << "Adding tail keyword to function since it can never be "
1493 "passed stack args: " << *Inst << "\n");
1494 cast<CallInst>(Inst)->setTailCall();
1497 // Ensure that functions that can never have a "tail" keyword due to the
1498 // semantics of ARC truly do not do so.
1499 if (IsNeverTail(Class)) {
1501 DEBUG(dbgs() << "Removing tail keyword from function: " << *Inst <<
1503 cast<CallInst>(Inst)->setTailCall(false);
1506 // Set nounwind as needed.
1507 if (IsNoThrow(Class)) {
1509 DEBUG(dbgs() << "Found no throw class. Setting nounwind on: " << *Inst
1511 cast<CallInst>(Inst)->setDoesNotThrow();
1514 if (!IsNoopOnNull(Class)) {
1515 UsedInThisFunction |= 1 << Class;
1519 const Value *Arg = GetObjCArg(Inst);
1521 // ARC calls with null are no-ops. Delete them.
1522 if (IsNullOrUndef(Arg)) {
1525 DEBUG(dbgs() << "ARC calls with null are no-ops. Erasing: " << *Inst
1527 EraseInstruction(Inst);
1531 // Keep track of which of retain, release, autorelease, and retain_block
1532 // are actually present in this function.
1533 UsedInThisFunction |= 1 << Class;
1535 // If Arg is a PHI, and one or more incoming values to the
1536 // PHI are null, and the call is control-equivalent to the PHI, and there
1537 // are no relevant side effects between the PHI and the call, the call
1538 // could be pushed up to just those paths with non-null incoming values.
1539 // For now, don't bother splitting critical edges for this.
1540 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
1541 Worklist.push_back(std::make_pair(Inst, Arg));
1543 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
1547 const PHINode *PN = dyn_cast<PHINode>(Arg);
1550 // Determine if the PHI has any null operands, or any incoming
1552 bool HasNull = false;
1553 bool HasCriticalEdges = false;
1554 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1556 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
1557 if (IsNullOrUndef(Incoming))
1559 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
1560 .getNumSuccessors() != 1) {
1561 HasCriticalEdges = true;
1565 // If we have null operands and no critical edges, optimize.
1566 if (!HasCriticalEdges && HasNull) {
1567 SmallPtrSet<Instruction *, 4> DependingInstructions;
1568 SmallPtrSet<const BasicBlock *, 4> Visited;
1570 // Check that there is nothing that cares about the reference
1571 // count between the call and the phi.
1574 case IC_RetainBlock:
1575 // These can always be moved up.
1578 // These can't be moved across things that care about the retain
1580 FindDependencies(NeedsPositiveRetainCount, Arg,
1581 Inst->getParent(), Inst,
1582 DependingInstructions, Visited, PA);
1584 case IC_Autorelease:
1585 // These can't be moved across autorelease pool scope boundaries.
1586 FindDependencies(AutoreleasePoolBoundary, Arg,
1587 Inst->getParent(), Inst,
1588 DependingInstructions, Visited, PA);
1591 case IC_AutoreleaseRV:
1592 // Don't move these; the RV optimization depends on the autoreleaseRV
1593 // being tail called, and the retainRV being immediately after a call
1594 // (which might still happen if we get lucky with codegen layout, but
1595 // it's not worth taking the chance).
1598 llvm_unreachable("Invalid dependence flavor");
1601 if (DependingInstructions.size() == 1 &&
1602 *DependingInstructions.begin() == PN) {
1605 // Clone the call into each predecessor that has a non-null value.
1606 CallInst *CInst = cast<CallInst>(Inst);
1607 Type *ParamTy = CInst->getArgOperand(0)->getType();
1608 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1610 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
1611 if (!IsNullOrUndef(Incoming)) {
1612 CallInst *Clone = cast<CallInst>(CInst->clone());
1613 Value *Op = PN->getIncomingValue(i);
1614 Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
1615 if (Op->getType() != ParamTy)
1616 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
1617 Clone->setArgOperand(0, Op);
1618 Clone->insertBefore(InsertPos);
1620 DEBUG(dbgs() << "Cloning "
1622 "And inserting clone at " << *InsertPos << "\n");
1623 Worklist.push_back(std::make_pair(Clone, Incoming));
1626 // Erase the original call.
1627 DEBUG(dbgs() << "Erasing: " << *CInst << "\n");
1628 EraseInstruction(CInst);
1632 } while (!Worklist.empty());
1636 /// Check for critical edges, loop boundaries, irreducible control flow, or
1637 /// other CFG structures where moving code across the edge would result in it
1638 /// being executed more.
1640 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
1641 DenseMap<const BasicBlock *, BBState> &BBStates,
1642 BBState &MyStates) const {
1643 // If any top-down local-use or possible-dec has a succ which is earlier in
1644 // the sequence, forget it.
1645 for (BBState::ptr_iterator I = MyStates.top_down_ptr_begin(),
1646 E = MyStates.top_down_ptr_end(); I != E; ++I)
1647 switch (I->second.GetSeq()) {
1650 const Value *Arg = I->first;
1651 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1652 bool SomeSuccHasSame = false;
1653 bool AllSuccsHaveSame = true;
1654 PtrState &S = I->second;
1655 succ_const_iterator SI(TI), SE(TI, false);
1657 for (; SI != SE; ++SI) {
1658 Sequence SuccSSeq = S_None;
1659 bool SuccSRRIKnownSafe = false;
1660 // If VisitBottomUp has pointer information for this successor, take
1661 // what we know about it.
1662 DenseMap<const BasicBlock *, BBState>::iterator BBI =
1664 assert(BBI != BBStates.end());
1665 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
1666 SuccSSeq = SuccS.GetSeq();
1667 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
1670 case S_CanRelease: {
1671 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
1672 S.ClearSequenceProgress();
1678 SomeSuccHasSame = true;
1682 case S_MovableRelease:
1683 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
1684 AllSuccsHaveSame = false;
1687 llvm_unreachable("bottom-up pointer in retain state!");
1690 // If the state at the other end of any of the successor edges
1691 // matches the current state, require all edges to match. This
1692 // guards against loops in the middle of a sequence.
1693 if (SomeSuccHasSame && !AllSuccsHaveSame)
1694 S.ClearSequenceProgress();
1697 case S_CanRelease: {
1698 const Value *Arg = I->first;
1699 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1700 bool SomeSuccHasSame = false;
1701 bool AllSuccsHaveSame = true;
1702 PtrState &S = I->second;
1703 succ_const_iterator SI(TI), SE(TI, false);
1705 for (; SI != SE; ++SI) {
1706 Sequence SuccSSeq = S_None;
1707 bool SuccSRRIKnownSafe = false;
1708 // If VisitBottomUp has pointer information for this successor, take
1709 // what we know about it.
1710 DenseMap<const BasicBlock *, BBState>::iterator BBI =
1712 assert(BBI != BBStates.end());
1713 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
1714 SuccSSeq = SuccS.GetSeq();
1715 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
1718 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
1719 S.ClearSequenceProgress();
1725 SomeSuccHasSame = true;
1729 case S_MovableRelease:
1731 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
1732 AllSuccsHaveSame = false;
1735 llvm_unreachable("bottom-up pointer in retain state!");
1738 // If the state at the other end of any of the successor edges
1739 // matches the current state, require all edges to match. This
1740 // guards against loops in the middle of a sequence.
1741 if (SomeSuccHasSame && !AllSuccsHaveSame)
1742 S.ClearSequenceProgress();
1749 ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst,
1751 MapVector<Value *, RRInfo> &Retains,
1752 BBState &MyStates) {
1753 bool NestingDetected = false;
1754 InstructionClass Class = GetInstructionClass(Inst);
1755 const Value *Arg = 0;
1757 DEBUG(dbgs() << "Class: " << Class << "\n");
1761 Arg = GetObjCArg(Inst);
1763 PtrState &S = MyStates.getPtrBottomUpState(Arg);
1765 // If we see two releases in a row on the same pointer. If so, make
1766 // a note, and we'll cicle back to revisit it after we've
1767 // hopefully eliminated the second release, which may allow us to
1768 // eliminate the first release too.
1769 // Theoretically we could implement removal of nested retain+release
1770 // pairs by making PtrState hold a stack of states, but this is
1771 // simple and avoids adding overhead for the non-nested case.
1772 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease) {
1773 DEBUG(dbgs() << "Found nested releases (i.e. a release pair)\n");
1774 NestingDetected = true;
1777 MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
1778 Sequence NewSeq = ReleaseMetadata ? S_MovableRelease : S_Release;
1779 ANNOTATE_BOTTOMUP(Inst, Arg, S.GetSeq(), NewSeq);
1780 S.ResetSequenceProgress(NewSeq);
1781 S.RRI.ReleaseMetadata = ReleaseMetadata;
1782 S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
1783 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
1784 S.RRI.Calls.insert(Inst);
1785 S.SetKnownPositiveRefCount();
1788 case IC_RetainBlock:
1789 // In OptimizeIndividualCalls, we have strength reduced all optimizable
1790 // objc_retainBlocks to objc_retains. Thus at this point any
1791 // objc_retainBlocks that we see are not optimizable.
1795 Arg = GetObjCArg(Inst);
1797 PtrState &S = MyStates.getPtrBottomUpState(Arg);
1798 S.SetKnownPositiveRefCount();
1800 Sequence OldSeq = S.GetSeq();
1804 case S_MovableRelease:
1806 // If OldSeq is not S_Use or OldSeq is S_Use and we are tracking an
1807 // imprecise release, clear our reverse insertion points.
1808 if (OldSeq != S_Use || S.RRI.IsTrackingImpreciseReleases())
1809 S.RRI.ReverseInsertPts.clear();
1812 // Don't do retain+release tracking for IC_RetainRV, because it's
1813 // better to let it remain as the first instruction after a call.
1814 if (Class != IC_RetainRV)
1815 Retains[Inst] = S.RRI;
1816 S.ClearSequenceProgress();
1821 llvm_unreachable("bottom-up pointer in retain state!");
1823 ANNOTATE_BOTTOMUP(Inst, Arg, OldSeq, S.GetSeq());
1824 // A retain moving bottom up can be a use.
1827 case IC_AutoreleasepoolPop:
1828 // Conservatively, clear MyStates for all known pointers.
1829 MyStates.clearBottomUpPointers();
1830 return NestingDetected;
1831 case IC_AutoreleasepoolPush:
1833 // These are irrelevant.
1834 return NestingDetected;
1839 // Consider any other possible effects of this instruction on each
1840 // pointer being tracked.
1841 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
1842 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
1843 const Value *Ptr = MI->first;
1845 continue; // Handled above.
1846 PtrState &S = MI->second;
1847 Sequence Seq = S.GetSeq();
1849 // Check for possible releases.
1850 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
1851 DEBUG(dbgs() << "CanAlterRefCount: Seq: " << Seq << "; " << *Ptr
1853 S.ClearKnownPositiveRefCount();
1856 S.SetSeq(S_CanRelease);
1857 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S.GetSeq());
1861 case S_MovableRelease:
1866 llvm_unreachable("bottom-up pointer in retain state!");
1870 // Check for possible direct uses.
1873 case S_MovableRelease:
1874 if (CanUse(Inst, Ptr, PA, Class)) {
1875 DEBUG(dbgs() << "CanUse: Seq: " << Seq << "; " << *Ptr
1877 assert(S.RRI.ReverseInsertPts.empty());
1878 // If this is an invoke instruction, we're scanning it as part of
1879 // one of its successor blocks, since we can't insert code after it
1880 // in its own block, and we don't want to split critical edges.
1881 if (isa<InvokeInst>(Inst))
1882 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
1884 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
1886 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S_Use);
1887 } else if (Seq == S_Release && IsUser(Class)) {
1888 DEBUG(dbgs() << "PreciseReleaseUse: Seq: " << Seq << "; " << *Ptr
1890 // Non-movable releases depend on any possible objc pointer use.
1892 ANNOTATE_BOTTOMUP(Inst, Ptr, S_Release, S_Stop);
1893 assert(S.RRI.ReverseInsertPts.empty());
1894 // As above; handle invoke specially.
1895 if (isa<InvokeInst>(Inst))
1896 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
1898 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
1902 if (CanUse(Inst, Ptr, PA, Class)) {
1903 DEBUG(dbgs() << "PreciseStopUse: Seq: " << Seq << "; " << *Ptr
1906 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S_Use);
1914 llvm_unreachable("bottom-up pointer in retain state!");
1918 return NestingDetected;
1922 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
1923 DenseMap<const BasicBlock *, BBState> &BBStates,
1924 MapVector<Value *, RRInfo> &Retains) {
1926 DEBUG(dbgs() << "\n== ObjCARCOpt::VisitBottomUp ==\n");
1928 bool NestingDetected = false;
1929 BBState &MyStates = BBStates[BB];
1931 // Merge the states from each successor to compute the initial state
1932 // for the current block.
1933 BBState::edge_iterator SI(MyStates.succ_begin()),
1934 SE(MyStates.succ_end());
1936 const BasicBlock *Succ = *SI;
1937 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
1938 assert(I != BBStates.end());
1939 MyStates.InitFromSucc(I->second);
1941 for (; SI != SE; ++SI) {
1943 I = BBStates.find(Succ);
1944 assert(I != BBStates.end());
1945 MyStates.MergeSucc(I->second);
1949 // If ARC Annotations are enabled, output the current state of pointers at the
1950 // bottom of the basic block.
1951 ANNOTATE_BOTTOMUP_BBEND(MyStates, BB);
1953 // Visit all the instructions, bottom-up.
1954 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
1955 Instruction *Inst = llvm::prior(I);
1957 // Invoke instructions are visited as part of their successors (below).
1958 if (isa<InvokeInst>(Inst))
1961 DEBUG(dbgs() << "Visiting " << *Inst << "\n");
1963 NestingDetected |= VisitInstructionBottomUp(Inst, BB, Retains, MyStates);
1966 // If there's a predecessor with an invoke, visit the invoke as if it were
1967 // part of this block, since we can't insert code after an invoke in its own
1968 // block, and we don't want to split critical edges.
1969 for (BBState::edge_iterator PI(MyStates.pred_begin()),
1970 PE(MyStates.pred_end()); PI != PE; ++PI) {
1971 BasicBlock *Pred = *PI;
1972 if (InvokeInst *II = dyn_cast<InvokeInst>(&Pred->back()))
1973 NestingDetected |= VisitInstructionBottomUp(II, BB, Retains, MyStates);
1976 // If ARC Annotations are enabled, output the current state of pointers at the
1977 // top of the basic block.
1978 ANNOTATE_BOTTOMUP_BBSTART(MyStates, BB);
1980 return NestingDetected;
1984 ObjCARCOpt::VisitInstructionTopDown(Instruction *Inst,
1985 DenseMap<Value *, RRInfo> &Releases,
1986 BBState &MyStates) {
1987 bool NestingDetected = false;
1988 InstructionClass Class = GetInstructionClass(Inst);
1989 const Value *Arg = 0;
1992 case IC_RetainBlock:
1993 // In OptimizeIndividualCalls, we have strength reduced all optimizable
1994 // objc_retainBlocks to objc_retains. Thus at this point any
1995 // objc_retainBlocks that we see are not optimizable.
1999 Arg = GetObjCArg(Inst);
2001 PtrState &S = MyStates.getPtrTopDownState(Arg);
2003 // Don't do retain+release tracking for IC_RetainRV, because it's
2004 // better to let it remain as the first instruction after a call.
2005 if (Class != IC_RetainRV) {
2006 // If we see two retains in a row on the same pointer. If so, make
2007 // a note, and we'll cicle back to revisit it after we've
2008 // hopefully eliminated the second retain, which may allow us to
2009 // eliminate the first retain too.
2010 // Theoretically we could implement removal of nested retain+release
2011 // pairs by making PtrState hold a stack of states, but this is
2012 // simple and avoids adding overhead for the non-nested case.
2013 if (S.GetSeq() == S_Retain)
2014 NestingDetected = true;
2016 ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_Retain);
2017 S.ResetSequenceProgress(S_Retain);
2018 S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
2019 S.RRI.Calls.insert(Inst);
2022 S.SetKnownPositiveRefCount();
2024 // A retain can be a potential use; procede to the generic checking
2029 Arg = GetObjCArg(Inst);
2031 PtrState &S = MyStates.getPtrTopDownState(Arg);
2032 S.ClearKnownPositiveRefCount();
2034 Sequence OldSeq = S.GetSeq();
2036 MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2041 if (OldSeq == S_Retain || ReleaseMetadata != 0)
2042 S.RRI.ReverseInsertPts.clear();
2045 S.RRI.ReleaseMetadata = ReleaseMetadata;
2046 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2047 Releases[Inst] = S.RRI;
2048 ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_None);
2049 S.ClearSequenceProgress();
2055 case S_MovableRelease:
2056 llvm_unreachable("top-down pointer in release state!");
2060 case IC_AutoreleasepoolPop:
2061 // Conservatively, clear MyStates for all known pointers.
2062 MyStates.clearTopDownPointers();
2063 return NestingDetected;
2064 case IC_AutoreleasepoolPush:
2066 // These are irrelevant.
2067 return NestingDetected;
2072 // Consider any other possible effects of this instruction on each
2073 // pointer being tracked.
2074 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
2075 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
2076 const Value *Ptr = MI->first;
2078 continue; // Handled above.
2079 PtrState &S = MI->second;
2080 Sequence Seq = S.GetSeq();
2082 // Check for possible releases.
2083 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2084 DEBUG(dbgs() << "CanAlterRefCount: Seq: " << Seq << "; " << *Ptr
2086 S.ClearKnownPositiveRefCount();
2089 S.SetSeq(S_CanRelease);
2090 ANNOTATE_TOPDOWN(Inst, Ptr, Seq, S_CanRelease);
2091 assert(S.RRI.ReverseInsertPts.empty());
2092 S.RRI.ReverseInsertPts.insert(Inst);
2094 // One call can't cause a transition from S_Retain to S_CanRelease
2095 // and S_CanRelease to S_Use. If we've made the first transition,
2104 case S_MovableRelease:
2105 llvm_unreachable("top-down pointer in release state!");
2109 // Check for possible direct uses.
2112 if (CanUse(Inst, Ptr, PA, Class)) {
2113 DEBUG(dbgs() << "CanUse: Seq: " << Seq << "; " << *Ptr
2116 ANNOTATE_TOPDOWN(Inst, Ptr, Seq, S_Use);
2125 case S_MovableRelease:
2126 llvm_unreachable("top-down pointer in release state!");
2130 return NestingDetected;
2134 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
2135 DenseMap<const BasicBlock *, BBState> &BBStates,
2136 DenseMap<Value *, RRInfo> &Releases) {
2137 DEBUG(dbgs() << "\n== ObjCARCOpt::VisitTopDown ==\n");
2138 bool NestingDetected = false;
2139 BBState &MyStates = BBStates[BB];
2141 // Merge the states from each predecessor to compute the initial state
2142 // for the current block.
2143 BBState::edge_iterator PI(MyStates.pred_begin()),
2144 PE(MyStates.pred_end());
2146 const BasicBlock *Pred = *PI;
2147 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
2148 assert(I != BBStates.end());
2149 MyStates.InitFromPred(I->second);
2151 for (; PI != PE; ++PI) {
2153 I = BBStates.find(Pred);
2154 assert(I != BBStates.end());
2155 MyStates.MergePred(I->second);
2159 // If ARC Annotations are enabled, output the current state of pointers at the
2160 // top of the basic block.
2161 ANNOTATE_TOPDOWN_BBSTART(MyStates, BB);
2163 // Visit all the instructions, top-down.
2164 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2165 Instruction *Inst = I;
2167 DEBUG(dbgs() << "Visiting " << *Inst << "\n");
2169 NestingDetected |= VisitInstructionTopDown(Inst, Releases, MyStates);
2172 // If ARC Annotations are enabled, output the current state of pointers at the
2173 // bottom of the basic block.
2174 ANNOTATE_TOPDOWN_BBEND(MyStates, BB);
2176 #ifdef ARC_ANNOTATIONS
2177 if (EnableARCAnnotations && EnableCheckForCFGHazards)
2179 CheckForCFGHazards(BB, BBStates, MyStates);
2180 return NestingDetected;
2184 ComputePostOrders(Function &F,
2185 SmallVectorImpl<BasicBlock *> &PostOrder,
2186 SmallVectorImpl<BasicBlock *> &ReverseCFGPostOrder,
2187 unsigned NoObjCARCExceptionsMDKind,
2188 DenseMap<const BasicBlock *, BBState> &BBStates) {
2189 /// The visited set, for doing DFS walks.
2190 SmallPtrSet<BasicBlock *, 16> Visited;
2192 // Do DFS, computing the PostOrder.
2193 SmallPtrSet<BasicBlock *, 16> OnStack;
2194 SmallVector<std::pair<BasicBlock *, succ_iterator>, 16> SuccStack;
2196 // Functions always have exactly one entry block, and we don't have
2197 // any other block that we treat like an entry block.
2198 BasicBlock *EntryBB = &F.getEntryBlock();
2199 BBState &MyStates = BBStates[EntryBB];
2200 MyStates.SetAsEntry();
2201 TerminatorInst *EntryTI = cast<TerminatorInst>(&EntryBB->back());
2202 SuccStack.push_back(std::make_pair(EntryBB, succ_iterator(EntryTI)));
2203 Visited.insert(EntryBB);
2204 OnStack.insert(EntryBB);
2207 BasicBlock *CurrBB = SuccStack.back().first;
2208 TerminatorInst *TI = cast<TerminatorInst>(&CurrBB->back());
2209 succ_iterator SE(TI, false);
2211 while (SuccStack.back().second != SE) {
2212 BasicBlock *SuccBB = *SuccStack.back().second++;
2213 if (Visited.insert(SuccBB)) {
2214 TerminatorInst *TI = cast<TerminatorInst>(&SuccBB->back());
2215 SuccStack.push_back(std::make_pair(SuccBB, succ_iterator(TI)));
2216 BBStates[CurrBB].addSucc(SuccBB);
2217 BBState &SuccStates = BBStates[SuccBB];
2218 SuccStates.addPred(CurrBB);
2219 OnStack.insert(SuccBB);
2223 if (!OnStack.count(SuccBB)) {
2224 BBStates[CurrBB].addSucc(SuccBB);
2225 BBStates[SuccBB].addPred(CurrBB);
2228 OnStack.erase(CurrBB);
2229 PostOrder.push_back(CurrBB);
2230 SuccStack.pop_back();
2231 } while (!SuccStack.empty());
2235 // Do reverse-CFG DFS, computing the reverse-CFG PostOrder.
2236 // Functions may have many exits, and there also blocks which we treat
2237 // as exits due to ignored edges.
2238 SmallVector<std::pair<BasicBlock *, BBState::edge_iterator>, 16> PredStack;
2239 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2240 BasicBlock *ExitBB = I;
2241 BBState &MyStates = BBStates[ExitBB];
2242 if (!MyStates.isExit())
2245 MyStates.SetAsExit();
2247 PredStack.push_back(std::make_pair(ExitBB, MyStates.pred_begin()));
2248 Visited.insert(ExitBB);
2249 while (!PredStack.empty()) {
2250 reverse_dfs_next_succ:
2251 BBState::edge_iterator PE = BBStates[PredStack.back().first].pred_end();
2252 while (PredStack.back().second != PE) {
2253 BasicBlock *BB = *PredStack.back().second++;
2254 if (Visited.insert(BB)) {
2255 PredStack.push_back(std::make_pair(BB, BBStates[BB].pred_begin()));
2256 goto reverse_dfs_next_succ;
2259 ReverseCFGPostOrder.push_back(PredStack.pop_back_val().first);
2264 // Visit the function both top-down and bottom-up.
2266 ObjCARCOpt::Visit(Function &F,
2267 DenseMap<const BasicBlock *, BBState> &BBStates,
2268 MapVector<Value *, RRInfo> &Retains,
2269 DenseMap<Value *, RRInfo> &Releases) {
2271 // Use reverse-postorder traversals, because we magically know that loops
2272 // will be well behaved, i.e. they won't repeatedly call retain on a single
2273 // pointer without doing a release. We can't use the ReversePostOrderTraversal
2274 // class here because we want the reverse-CFG postorder to consider each
2275 // function exit point, and we want to ignore selected cycle edges.
2276 SmallVector<BasicBlock *, 16> PostOrder;
2277 SmallVector<BasicBlock *, 16> ReverseCFGPostOrder;
2278 ComputePostOrders(F, PostOrder, ReverseCFGPostOrder,
2279 NoObjCARCExceptionsMDKind,
2282 // Use reverse-postorder on the reverse CFG for bottom-up.
2283 bool BottomUpNestingDetected = false;
2284 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2285 ReverseCFGPostOrder.rbegin(), E = ReverseCFGPostOrder.rend();
2287 BottomUpNestingDetected |= VisitBottomUp(*I, BBStates, Retains);
2289 // Use reverse-postorder for top-down.
2290 bool TopDownNestingDetected = false;
2291 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2292 PostOrder.rbegin(), E = PostOrder.rend();
2294 TopDownNestingDetected |= VisitTopDown(*I, BBStates, Releases);
2296 return TopDownNestingDetected && BottomUpNestingDetected;
2299 /// Move the calls in RetainsToMove and ReleasesToMove.
2300 void ObjCARCOpt::MoveCalls(Value *Arg,
2301 RRInfo &RetainsToMove,
2302 RRInfo &ReleasesToMove,
2303 MapVector<Value *, RRInfo> &Retains,
2304 DenseMap<Value *, RRInfo> &Releases,
2305 SmallVectorImpl<Instruction *> &DeadInsts,
2307 Type *ArgTy = Arg->getType();
2308 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
2310 DEBUG(dbgs() << "== ObjCARCOpt::MoveCalls ==\n");
2312 // Insert the new retain and release calls.
2313 for (SmallPtrSet<Instruction *, 2>::const_iterator
2314 PI = ReleasesToMove.ReverseInsertPts.begin(),
2315 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2316 Instruction *InsertPt = *PI;
2317 Value *MyArg = ArgTy == ParamTy ? Arg :
2318 new BitCastInst(Arg, ParamTy, "", InsertPt);
2320 CallInst::Create(getRetainCallee(M), MyArg, "", InsertPt);
2321 Call->setDoesNotThrow();
2322 Call->setTailCall();
2324 DEBUG(dbgs() << "Inserting new Release: " << *Call << "\n"
2325 "At insertion point: " << *InsertPt << "\n");
2327 for (SmallPtrSet<Instruction *, 2>::const_iterator
2328 PI = RetainsToMove.ReverseInsertPts.begin(),
2329 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2330 Instruction *InsertPt = *PI;
2331 Value *MyArg = ArgTy == ParamTy ? Arg :
2332 new BitCastInst(Arg, ParamTy, "", InsertPt);
2333 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
2335 // Attach a clang.imprecise_release metadata tag, if appropriate.
2336 if (MDNode *M = ReleasesToMove.ReleaseMetadata)
2337 Call->setMetadata(ImpreciseReleaseMDKind, M);
2338 Call->setDoesNotThrow();
2339 if (ReleasesToMove.IsTailCallRelease)
2340 Call->setTailCall();
2342 DEBUG(dbgs() << "Inserting new Release: " << *Call << "\n"
2343 "At insertion point: " << *InsertPt << "\n");
2346 // Delete the original retain and release calls.
2347 for (SmallPtrSet<Instruction *, 2>::const_iterator
2348 AI = RetainsToMove.Calls.begin(),
2349 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
2350 Instruction *OrigRetain = *AI;
2351 Retains.blot(OrigRetain);
2352 DeadInsts.push_back(OrigRetain);
2353 DEBUG(dbgs() << "Deleting retain: " << *OrigRetain << "\n");
2355 for (SmallPtrSet<Instruction *, 2>::const_iterator
2356 AI = ReleasesToMove.Calls.begin(),
2357 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
2358 Instruction *OrigRelease = *AI;
2359 Releases.erase(OrigRelease);
2360 DeadInsts.push_back(OrigRelease);
2361 DEBUG(dbgs() << "Deleting release: " << *OrigRelease << "\n");
2367 ObjCARCOpt::ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState>
2369 MapVector<Value *, RRInfo> &Retains,
2370 DenseMap<Value *, RRInfo> &Releases,
2372 SmallVector<Instruction *, 4> &NewRetains,
2373 SmallVector<Instruction *, 4> &NewReleases,
2374 SmallVector<Instruction *, 8> &DeadInsts,
2375 RRInfo &RetainsToMove,
2376 RRInfo &ReleasesToMove,
2379 bool &AnyPairsCompletelyEliminated) {
2380 // If a pair happens in a region where it is known that the reference count
2381 // is already incremented, we can similarly ignore possible decrements.
2382 bool KnownSafeTD = true, KnownSafeBU = true;
2384 // Connect the dots between the top-down-collected RetainsToMove and
2385 // bottom-up-collected ReleasesToMove to form sets of related calls.
2386 // This is an iterative process so that we connect multiple releases
2387 // to multiple retains if needed.
2388 unsigned OldDelta = 0;
2389 unsigned NewDelta = 0;
2390 unsigned OldCount = 0;
2391 unsigned NewCount = 0;
2392 bool FirstRelease = true;
2394 for (SmallVectorImpl<Instruction *>::const_iterator
2395 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
2396 Instruction *NewRetain = *NI;
2397 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
2398 assert(It != Retains.end());
2399 const RRInfo &NewRetainRRI = It->second;
2400 KnownSafeTD &= NewRetainRRI.KnownSafe;
2401 for (SmallPtrSet<Instruction *, 2>::const_iterator
2402 LI = NewRetainRRI.Calls.begin(),
2403 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
2404 Instruction *NewRetainRelease = *LI;
2405 DenseMap<Value *, RRInfo>::const_iterator Jt =
2406 Releases.find(NewRetainRelease);
2407 if (Jt == Releases.end())
2409 const RRInfo &NewRetainReleaseRRI = Jt->second;
2410 assert(NewRetainReleaseRRI.Calls.count(NewRetain));
2411 if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
2413 BBStates[NewRetainRelease->getParent()].GetAllPathCount();
2415 // Merge the ReleaseMetadata and IsTailCallRelease values.
2417 ReleasesToMove.ReleaseMetadata =
2418 NewRetainReleaseRRI.ReleaseMetadata;
2419 ReleasesToMove.IsTailCallRelease =
2420 NewRetainReleaseRRI.IsTailCallRelease;
2421 FirstRelease = false;
2423 if (ReleasesToMove.ReleaseMetadata !=
2424 NewRetainReleaseRRI.ReleaseMetadata)
2425 ReleasesToMove.ReleaseMetadata = 0;
2426 if (ReleasesToMove.IsTailCallRelease !=
2427 NewRetainReleaseRRI.IsTailCallRelease)
2428 ReleasesToMove.IsTailCallRelease = false;
2431 // Collect the optimal insertion points.
2433 for (SmallPtrSet<Instruction *, 2>::const_iterator
2434 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
2435 RE = NewRetainReleaseRRI.ReverseInsertPts.end();
2437 Instruction *RIP = *RI;
2438 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
2439 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
2441 NewReleases.push_back(NewRetainRelease);
2446 if (NewReleases.empty()) break;
2448 // Back the other way.
2449 for (SmallVectorImpl<Instruction *>::const_iterator
2450 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
2451 Instruction *NewRelease = *NI;
2452 DenseMap<Value *, RRInfo>::const_iterator It =
2453 Releases.find(NewRelease);
2454 assert(It != Releases.end());
2455 const RRInfo &NewReleaseRRI = It->second;
2456 KnownSafeBU &= NewReleaseRRI.KnownSafe;
2457 for (SmallPtrSet<Instruction *, 2>::const_iterator
2458 LI = NewReleaseRRI.Calls.begin(),
2459 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
2460 Instruction *NewReleaseRetain = *LI;
2461 MapVector<Value *, RRInfo>::const_iterator Jt =
2462 Retains.find(NewReleaseRetain);
2463 if (Jt == Retains.end())
2465 const RRInfo &NewReleaseRetainRRI = Jt->second;
2466 assert(NewReleaseRetainRRI.Calls.count(NewRelease));
2467 if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
2468 unsigned PathCount =
2469 BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
2470 OldDelta += PathCount;
2471 OldCount += PathCount;
2473 // Collect the optimal insertion points.
2475 for (SmallPtrSet<Instruction *, 2>::const_iterator
2476 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
2477 RE = NewReleaseRetainRRI.ReverseInsertPts.end();
2479 Instruction *RIP = *RI;
2480 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
2481 PathCount = BBStates[RIP->getParent()].GetAllPathCount();
2482 NewDelta += PathCount;
2483 NewCount += PathCount;
2486 NewRetains.push_back(NewReleaseRetain);
2490 NewReleases.clear();
2491 if (NewRetains.empty()) break;
2494 // If the pointer is known incremented or nested, we can safely delete the
2495 // pair regardless of what's between them.
2496 if (KnownSafeTD || KnownSafeBU) {
2497 RetainsToMove.ReverseInsertPts.clear();
2498 ReleasesToMove.ReverseInsertPts.clear();
2501 // Determine whether the new insertion points we computed preserve the
2502 // balance of retain and release calls through the program.
2503 // TODO: If the fully aggressive solution isn't valid, try to find a
2504 // less aggressive solution which is.
2509 // Determine whether the original call points are balanced in the retain and
2510 // release calls through the program. If not, conservatively don't touch
2512 // TODO: It's theoretically possible to do code motion in this case, as
2513 // long as the existing imbalances are maintained.
2518 assert(OldCount != 0 && "Unreachable code?");
2519 NumRRs += OldCount - NewCount;
2520 // Set to true if we completely removed any RR pairs.
2521 AnyPairsCompletelyEliminated = NewCount == 0;
2523 // We can move calls!
2527 /// Identify pairings between the retains and releases, and delete and/or move
2530 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
2532 MapVector<Value *, RRInfo> &Retains,
2533 DenseMap<Value *, RRInfo> &Releases,
2535 DEBUG(dbgs() << "\n== ObjCARCOpt::PerformCodePlacement ==\n");
2537 bool AnyPairsCompletelyEliminated = false;
2538 RRInfo RetainsToMove;
2539 RRInfo ReleasesToMove;
2540 SmallVector<Instruction *, 4> NewRetains;
2541 SmallVector<Instruction *, 4> NewReleases;
2542 SmallVector<Instruction *, 8> DeadInsts;
2544 // Visit each retain.
2545 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
2546 E = Retains.end(); I != E; ++I) {
2547 Value *V = I->first;
2548 if (!V) continue; // blotted
2550 Instruction *Retain = cast<Instruction>(V);
2552 DEBUG(dbgs() << "Visiting: " << *Retain << "\n");
2554 Value *Arg = GetObjCArg(Retain);
2556 // If the object being released is in static or stack storage, we know it's
2557 // not being managed by ObjC reference counting, so we can delete pairs
2558 // regardless of what possible decrements or uses lie between them.
2559 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg);
2561 // A constant pointer can't be pointing to an object on the heap. It may
2562 // be reference-counted, but it won't be deleted.
2563 if (const LoadInst *LI = dyn_cast<LoadInst>(Arg))
2564 if (const GlobalVariable *GV =
2565 dyn_cast<GlobalVariable>(
2566 StripPointerCastsAndObjCCalls(LI->getPointerOperand())))
2567 if (GV->isConstant())
2570 // Connect the dots between the top-down-collected RetainsToMove and
2571 // bottom-up-collected ReleasesToMove to form sets of related calls.
2572 NewRetains.push_back(Retain);
2573 bool PerformMoveCalls =
2574 ConnectTDBUTraversals(BBStates, Retains, Releases, M, NewRetains,
2575 NewReleases, DeadInsts, RetainsToMove,
2576 ReleasesToMove, Arg, KnownSafe,
2577 AnyPairsCompletelyEliminated);
2579 #ifdef ARC_ANNOTATIONS
2580 // Do not move calls if ARC annotations are requested. If we were to move
2581 // calls in this case, we would not be able
2582 PerformMoveCalls = PerformMoveCalls && !EnableARCAnnotations;
2583 #endif // ARC_ANNOTATIONS
2585 if (PerformMoveCalls) {
2586 // Ok, everything checks out and we're all set. Let's move/delete some
2588 MoveCalls(Arg, RetainsToMove, ReleasesToMove,
2589 Retains, Releases, DeadInsts, M);
2592 // Clean up state for next retain.
2593 NewReleases.clear();
2595 RetainsToMove.clear();
2596 ReleasesToMove.clear();
2599 // Now that we're done moving everything, we can delete the newly dead
2600 // instructions, as we no longer need them as insert points.
2601 while (!DeadInsts.empty())
2602 EraseInstruction(DeadInsts.pop_back_val());
2604 return AnyPairsCompletelyEliminated;
2607 /// Weak pointer optimizations.
2608 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
2609 DEBUG(dbgs() << "\n== ObjCARCOpt::OptimizeWeakCalls ==\n");
2611 // First, do memdep-style RLE and S2L optimizations. We can't use memdep
2612 // itself because it uses AliasAnalysis and we need to do provenance
2614 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2615 Instruction *Inst = &*I++;
2617 DEBUG(dbgs() << "Visiting: " << *Inst << "\n");
2619 InstructionClass Class = GetBasicInstructionClass(Inst);
2620 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
2623 // Delete objc_loadWeak calls with no users.
2624 if (Class == IC_LoadWeak && Inst->use_empty()) {
2625 Inst->eraseFromParent();
2629 // TODO: For now, just look for an earlier available version of this value
2630 // within the same block. Theoretically, we could do memdep-style non-local
2631 // analysis too, but that would want caching. A better approach would be to
2632 // use the technique that EarlyCSE uses.
2633 inst_iterator Current = llvm::prior(I);
2634 BasicBlock *CurrentBB = Current.getBasicBlockIterator();
2635 for (BasicBlock::iterator B = CurrentBB->begin(),
2636 J = Current.getInstructionIterator();
2638 Instruction *EarlierInst = &*llvm::prior(J);
2639 InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
2640 switch (EarlierClass) {
2642 case IC_LoadWeakRetained: {
2643 // If this is loading from the same pointer, replace this load's value
2645 CallInst *Call = cast<CallInst>(Inst);
2646 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2647 Value *Arg = Call->getArgOperand(0);
2648 Value *EarlierArg = EarlierCall->getArgOperand(0);
2649 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2650 case AliasAnalysis::MustAlias:
2652 // If the load has a builtin retain, insert a plain retain for it.
2653 if (Class == IC_LoadWeakRetained) {
2655 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2659 // Zap the fully redundant load.
2660 Call->replaceAllUsesWith(EarlierCall);
2661 Call->eraseFromParent();
2663 case AliasAnalysis::MayAlias:
2664 case AliasAnalysis::PartialAlias:
2666 case AliasAnalysis::NoAlias:
2673 // If this is storing to the same pointer and has the same size etc.
2674 // replace this load's value with the stored value.
2675 CallInst *Call = cast<CallInst>(Inst);
2676 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2677 Value *Arg = Call->getArgOperand(0);
2678 Value *EarlierArg = EarlierCall->getArgOperand(0);
2679 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2680 case AliasAnalysis::MustAlias:
2682 // If the load has a builtin retain, insert a plain retain for it.
2683 if (Class == IC_LoadWeakRetained) {
2685 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2689 // Zap the fully redundant load.
2690 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
2691 Call->eraseFromParent();
2693 case AliasAnalysis::MayAlias:
2694 case AliasAnalysis::PartialAlias:
2696 case AliasAnalysis::NoAlias:
2703 // TOOD: Grab the copied value.
2705 case IC_AutoreleasepoolPush:
2707 case IC_IntrinsicUser:
2709 // Weak pointers are only modified through the weak entry points
2710 // (and arbitrary calls, which could call the weak entry points).
2713 // Anything else could modify the weak pointer.
2720 // Then, for each destroyWeak with an alloca operand, check to see if
2721 // the alloca and all its users can be zapped.
2722 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2723 Instruction *Inst = &*I++;
2724 InstructionClass Class = GetBasicInstructionClass(Inst);
2725 if (Class != IC_DestroyWeak)
2728 CallInst *Call = cast<CallInst>(Inst);
2729 Value *Arg = Call->getArgOperand(0);
2730 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
2731 for (Value::use_iterator UI = Alloca->use_begin(),
2732 UE = Alloca->use_end(); UI != UE; ++UI) {
2733 const Instruction *UserInst = cast<Instruction>(*UI);
2734 switch (GetBasicInstructionClass(UserInst)) {
2737 case IC_DestroyWeak:
2744 for (Value::use_iterator UI = Alloca->use_begin(),
2745 UE = Alloca->use_end(); UI != UE; ) {
2746 CallInst *UserInst = cast<CallInst>(*UI++);
2747 switch (GetBasicInstructionClass(UserInst)) {
2750 // These functions return their second argument.
2751 UserInst->replaceAllUsesWith(UserInst->getArgOperand(1));
2753 case IC_DestroyWeak:
2757 llvm_unreachable("alloca really is used!");
2759 UserInst->eraseFromParent();
2761 Alloca->eraseFromParent();
2767 /// Identify program paths which execute sequences of retains and releases which
2768 /// can be eliminated.
2769 bool ObjCARCOpt::OptimizeSequences(Function &F) {
2770 /// Releases, Retains - These are used to store the results of the main flow
2771 /// analysis. These use Value* as the key instead of Instruction* so that the
2772 /// map stays valid when we get around to rewriting code and calls get
2773 /// replaced by arguments.
2774 DenseMap<Value *, RRInfo> Releases;
2775 MapVector<Value *, RRInfo> Retains;
2777 /// This is used during the traversal of the function to track the
2778 /// states for each identified object at each block.
2779 DenseMap<const BasicBlock *, BBState> BBStates;
2781 // Analyze the CFG of the function, and all instructions.
2782 bool NestingDetected = Visit(F, BBStates, Retains, Releases);
2785 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
2789 /// Check if there is a dependent call earlier that does not have anything in
2790 /// between the Retain and the call that can affect the reference count of their
2791 /// shared pointer argument. Note that Retain need not be in BB.
2793 HasSafePathToPredecessorCall(const Value *Arg, Instruction *Retain,
2794 SmallPtrSet<Instruction *, 4> &DepInsts,
2795 SmallPtrSet<const BasicBlock *, 4> &Visited,
2796 ProvenanceAnalysis &PA) {
2797 FindDependencies(CanChangeRetainCount, Arg, Retain->getParent(), Retain,
2798 DepInsts, Visited, PA);
2799 if (DepInsts.size() != 1)
2803 dyn_cast_or_null<CallInst>(*DepInsts.begin());
2805 // Check that the pointer is the return value of the call.
2806 if (!Call || Arg != Call)
2809 // Check that the call is a regular call.
2810 InstructionClass Class = GetBasicInstructionClass(Call);
2811 if (Class != IC_CallOrUser && Class != IC_Call)
2817 /// Find a dependent retain that precedes the given autorelease for which there
2818 /// is nothing in between the two instructions that can affect the ref count of
2821 FindPredecessorRetainWithSafePath(const Value *Arg, BasicBlock *BB,
2822 Instruction *Autorelease,
2823 SmallPtrSet<Instruction *, 4> &DepInsts,
2824 SmallPtrSet<const BasicBlock *, 4> &Visited,
2825 ProvenanceAnalysis &PA) {
2826 FindDependencies(CanChangeRetainCount, Arg,
2827 BB, Autorelease, DepInsts, Visited, PA);
2828 if (DepInsts.size() != 1)
2832 dyn_cast_or_null<CallInst>(*DepInsts.begin());
2834 // Check that we found a retain with the same argument.
2836 !IsRetain(GetBasicInstructionClass(Retain)) ||
2837 GetObjCArg(Retain) != Arg) {
2844 /// Look for an ``autorelease'' instruction dependent on Arg such that there are
2845 /// no instructions dependent on Arg that need a positive ref count in between
2846 /// the autorelease and the ret.
2848 FindPredecessorAutoreleaseWithSafePath(const Value *Arg, BasicBlock *BB,
2850 SmallPtrSet<Instruction *, 4> &DepInsts,
2851 SmallPtrSet<const BasicBlock *, 4> &V,
2852 ProvenanceAnalysis &PA) {
2853 FindDependencies(NeedsPositiveRetainCount, Arg,
2854 BB, Ret, DepInsts, V, PA);
2855 if (DepInsts.size() != 1)
2858 CallInst *Autorelease =
2859 dyn_cast_or_null<CallInst>(*DepInsts.begin());
2862 InstructionClass AutoreleaseClass = GetBasicInstructionClass(Autorelease);
2863 if (!IsAutorelease(AutoreleaseClass))
2865 if (GetObjCArg(Autorelease) != Arg)
2871 /// Look for this pattern:
2873 /// %call = call i8* @something(...)
2874 /// %2 = call i8* @objc_retain(i8* %call)
2875 /// %3 = call i8* @objc_autorelease(i8* %2)
2878 /// And delete the retain and autorelease.
2879 void ObjCARCOpt::OptimizeReturns(Function &F) {
2880 if (!F.getReturnType()->isPointerTy())
2883 DEBUG(dbgs() << "\n== ObjCARCOpt::OptimizeReturns ==\n");
2885 SmallPtrSet<Instruction *, 4> DependingInstructions;
2886 SmallPtrSet<const BasicBlock *, 4> Visited;
2887 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
2888 BasicBlock *BB = FI;
2889 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
2891 DEBUG(dbgs() << "Visiting: " << *Ret << "\n");
2896 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
2898 // Look for an ``autorelease'' instruction that is a predecssor of Ret and
2899 // dependent on Arg such that there are no instructions dependent on Arg
2900 // that need a positive ref count in between the autorelease and Ret.
2901 CallInst *Autorelease =
2902 FindPredecessorAutoreleaseWithSafePath(Arg, BB, Ret,
2903 DependingInstructions, Visited,
2906 DependingInstructions.clear();
2910 FindPredecessorRetainWithSafePath(Arg, BB, Autorelease,
2911 DependingInstructions, Visited, PA);
2913 DependingInstructions.clear();
2916 // Check that there is nothing that can affect the reference count
2917 // between the retain and the call. Note that Retain need not be in BB.
2918 if (HasSafePathToPredecessorCall(Arg, Retain, DependingInstructions,
2920 // If so, we can zap the retain and autorelease.
2923 DEBUG(dbgs() << "Erasing: " << *Retain << "\nErasing: "
2924 << *Autorelease << "\n");
2925 EraseInstruction(Retain);
2926 EraseInstruction(Autorelease);
2931 DependingInstructions.clear();
2936 bool ObjCARCOpt::doInitialization(Module &M) {
2940 // If nothing in the Module uses ARC, don't do anything.
2941 Run = ModuleHasARC(M);
2945 // Identify the imprecise release metadata kind.
2946 ImpreciseReleaseMDKind =
2947 M.getContext().getMDKindID("clang.imprecise_release");
2948 CopyOnEscapeMDKind =
2949 M.getContext().getMDKindID("clang.arc.copy_on_escape");
2950 NoObjCARCExceptionsMDKind =
2951 M.getContext().getMDKindID("clang.arc.no_objc_arc_exceptions");
2952 #ifdef ARC_ANNOTATIONS
2953 ARCAnnotationBottomUpMDKind =
2954 M.getContext().getMDKindID("llvm.arc.annotation.bottomup");
2955 ARCAnnotationTopDownMDKind =
2956 M.getContext().getMDKindID("llvm.arc.annotation.topdown");
2957 ARCAnnotationProvenanceSourceMDKind =
2958 M.getContext().getMDKindID("llvm.arc.annotation.provenancesource");
2959 #endif // ARC_ANNOTATIONS
2961 // Intuitively, objc_retain and others are nocapture, however in practice
2962 // they are not, because they return their argument value. And objc_release
2963 // calls finalizers which can have arbitrary side effects.
2965 // These are initialized lazily.
2967 AutoreleaseRVCallee = 0;
2970 RetainBlockCallee = 0;
2971 AutoreleaseCallee = 0;
2976 bool ObjCARCOpt::runOnFunction(Function &F) {
2980 // If nothing in the Module uses ARC, don't do anything.
2986 DEBUG(dbgs() << "<<< ObjCARCOpt: Visiting Function: " << F.getName() << " >>>"
2989 PA.setAA(&getAnalysis<AliasAnalysis>());
2991 // This pass performs several distinct transformations. As a compile-time aid
2992 // when compiling code that isn't ObjC, skip these if the relevant ObjC
2993 // library functions aren't declared.
2995 // Preliminary optimizations. This also computs UsedInThisFunction.
2996 OptimizeIndividualCalls(F);
2998 // Optimizations for weak pointers.
2999 if (UsedInThisFunction & ((1 << IC_LoadWeak) |
3000 (1 << IC_LoadWeakRetained) |
3001 (1 << IC_StoreWeak) |
3002 (1 << IC_InitWeak) |
3003 (1 << IC_CopyWeak) |
3004 (1 << IC_MoveWeak) |
3005 (1 << IC_DestroyWeak)))
3006 OptimizeWeakCalls(F);
3008 // Optimizations for retain+release pairs.
3009 if (UsedInThisFunction & ((1 << IC_Retain) |
3010 (1 << IC_RetainRV) |
3011 (1 << IC_RetainBlock)))
3012 if (UsedInThisFunction & (1 << IC_Release))
3013 // Run OptimizeSequences until it either stops making changes or
3014 // no retain+release pair nesting is detected.
3015 while (OptimizeSequences(F)) {}
3017 // Optimizations if objc_autorelease is used.
3018 if (UsedInThisFunction & ((1 << IC_Autorelease) |
3019 (1 << IC_AutoreleaseRV)))
3022 DEBUG(dbgs() << "\n");
3027 void ObjCARCOpt::releaseMemory() {