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() << "DoesRetainableObjPtrEscape: 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() << "DoesRetainableObjPtrEscape: 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() << "DoesRetainableObjPtrEscape: User copies pointer "
211 "arguments. Pointer Escapes!\n");
212 // These special functions make copies of their pointer arguments.
215 case IC_IntrinsicUser:
216 // Use by the use intrinsic is not an escape.
220 // Use by an instruction which copies the value is an escape if the
221 // result is an escape.
222 if (isa<BitCastInst>(UUser) || isa<GetElementPtrInst>(UUser) ||
223 isa<PHINode>(UUser) || isa<SelectInst>(UUser)) {
225 if (VisitedSet.insert(UUser)) {
226 DEBUG(dbgs() << "DoesRetainableObjPtrEscape: User copies value. "
227 "Ptr escapes if result escapes. Adding to list.\n");
228 Worklist.push_back(UUser);
230 DEBUG(dbgs() << "DoesRetainableObjPtrEscape: Already visited node."
235 // Use by a load is not an escape.
236 if (isa<LoadInst>(UUser))
238 // Use by a store is not an escape if the use is the address.
239 if (const StoreInst *SI = dyn_cast<StoreInst>(UUser))
240 if (V != SI->getValueOperand())
244 // Regular calls and other stuff are not considered escapes.
247 // Otherwise, conservatively assume an escape.
248 DEBUG(dbgs() << "DoesRetainableObjPtrEscape: Assuming ptr escapes.\n");
251 } while (!Worklist.empty());
254 DEBUG(dbgs() << "DoesRetainableObjPtrEscape: Ptr does not escape.\n");
260 /// \defgroup ARCOpt ARC Optimization.
263 // TODO: On code like this:
266 // stuff_that_cannot_release()
267 // objc_autorelease(%x)
268 // stuff_that_cannot_release()
270 // stuff_that_cannot_release()
271 // objc_autorelease(%x)
273 // The second retain and autorelease can be deleted.
275 // TODO: It should be possible to delete
276 // objc_autoreleasePoolPush and objc_autoreleasePoolPop
277 // pairs if nothing is actually autoreleased between them. Also, autorelease
278 // calls followed by objc_autoreleasePoolPop calls (perhaps in ObjC++ code
279 // after inlining) can be turned into plain release calls.
281 // TODO: Critical-edge splitting. If the optimial insertion point is
282 // a critical edge, the current algorithm has to fail, because it doesn't
283 // know how to split edges. It should be possible to make the optimizer
284 // think in terms of edges, rather than blocks, and then split critical
287 // TODO: OptimizeSequences could generalized to be Interprocedural.
289 // TODO: Recognize that a bunch of other objc runtime calls have
290 // non-escaping arguments and non-releasing arguments, and may be
291 // non-autoreleasing.
293 // TODO: Sink autorelease calls as far as possible. Unfortunately we
294 // usually can't sink them past other calls, which would be the main
295 // case where it would be useful.
297 // TODO: The pointer returned from objc_loadWeakRetained is retained.
299 // TODO: Delete release+retain pairs (rare).
301 STATISTIC(NumNoops, "Number of no-op objc calls eliminated");
302 STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated");
303 STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases");
304 STATISTIC(NumRets, "Number of return value forwarding "
305 "retain+autoreleaes eliminated");
306 STATISTIC(NumRRs, "Number of retain+release paths eliminated");
307 STATISTIC(NumPeeps, "Number of calls peephole-optimized");
312 /// \brief A sequence of states that a pointer may go through in which an
313 /// objc_retain and objc_release are actually needed.
316 S_Retain, ///< objc_retain(x).
317 S_CanRelease, ///< foo(x) -- x could possibly see a ref count decrement.
318 S_Use, ///< any use of x.
319 S_Stop, ///< like S_Release, but code motion is stopped.
320 S_Release, ///< objc_release(x).
321 S_MovableRelease ///< objc_release(x), !clang.imprecise_release.
324 raw_ostream &operator<<(raw_ostream &OS, const Sequence S)
325 LLVM_ATTRIBUTE_UNUSED;
326 raw_ostream &operator<<(raw_ostream &OS, const Sequence S) {
329 return OS << "S_None";
331 return OS << "S_Retain";
333 return OS << "S_CanRelease";
335 return OS << "S_Use";
337 return OS << "S_Release";
338 case S_MovableRelease:
339 return OS << "S_MovableRelease";
341 return OS << "S_Stop";
343 llvm_unreachable("Unknown sequence type.");
347 static Sequence MergeSeqs(Sequence A, Sequence B, bool TopDown) {
351 if (A == S_None || B == S_None)
354 if (A > B) std::swap(A, B);
356 // Choose the side which is further along in the sequence.
357 if ((A == S_Retain || A == S_CanRelease) &&
358 (B == S_CanRelease || B == S_Use))
361 // Choose the side which is further along in the sequence.
362 if ((A == S_Use || A == S_CanRelease) &&
363 (B == S_Use || B == S_Release || B == S_Stop || B == S_MovableRelease))
365 // If both sides are releases, choose the more conservative one.
366 if (A == S_Stop && (B == S_Release || B == S_MovableRelease))
368 if (A == S_Release && B == S_MovableRelease)
376 /// \brief Unidirectional information about either a
377 /// retain-decrement-use-release sequence or release-use-decrement-retain
378 /// reverese sequence.
380 /// After an objc_retain, the reference count of the referenced
381 /// object is known to be positive. Similarly, before an objc_release, the
382 /// reference count of the referenced object is known to be positive. If
383 /// there are retain-release pairs in code regions where the retain count
384 /// is known to be positive, they can be eliminated, regardless of any side
385 /// effects between them.
387 /// Also, a retain+release pair nested within another retain+release
388 /// pair all on the known same pointer value can be eliminated, regardless
389 /// of any intervening side effects.
391 /// KnownSafe is true when either of these conditions is satisfied.
394 /// True of the objc_release calls are all marked with the "tail" keyword.
395 bool IsTailCallRelease;
397 /// If the Calls are objc_release calls and they all have a
398 /// clang.imprecise_release tag, this is the metadata tag.
399 MDNode *ReleaseMetadata;
401 /// For a top-down sequence, the set of objc_retains or
402 /// objc_retainBlocks. For bottom-up, the set of objc_releases.
403 SmallPtrSet<Instruction *, 2> Calls;
405 /// The set of optimal insert positions for moving calls in the opposite
407 SmallPtrSet<Instruction *, 2> ReverseInsertPts;
410 KnownSafe(false), IsTailCallRelease(false), ReleaseMetadata(0) {}
416 void RRInfo::clear() {
418 IsTailCallRelease = false;
421 ReverseInsertPts.clear();
425 /// \brief This class summarizes several per-pointer runtime properties which
426 /// are propogated through the flow graph.
428 /// True if the reference count is known to be incremented.
429 bool KnownPositiveRefCount;
431 /// True of we've seen an opportunity for partial RR elimination, such as
432 /// pushing calls into a CFG triangle or into one side of a CFG diamond.
435 /// The current position in the sequence.
439 /// Unidirectional information about the current sequence.
441 /// TODO: Encapsulate this better.
444 PtrState() : KnownPositiveRefCount(false), Partial(false),
447 void SetKnownPositiveRefCount() {
448 KnownPositiveRefCount = true;
451 void ClearKnownPositiveRefCount() {
452 KnownPositiveRefCount = false;
455 bool HasKnownPositiveRefCount() const {
456 return KnownPositiveRefCount;
459 void SetSeq(Sequence NewSeq) {
463 Sequence GetSeq() const {
467 void ClearSequenceProgress() {
468 ResetSequenceProgress(S_None);
471 void ResetSequenceProgress(Sequence NewSeq) {
477 void Merge(const PtrState &Other, bool TopDown);
482 PtrState::Merge(const PtrState &Other, bool TopDown) {
483 Seq = MergeSeqs(Seq, Other.Seq, TopDown);
484 KnownPositiveRefCount = KnownPositiveRefCount && Other.KnownPositiveRefCount;
486 // If we're not in a sequence (anymore), drop all associated state.
490 } else if (Partial || Other.Partial) {
491 // If we're doing a merge on a path that's previously seen a partial
492 // merge, conservatively drop the sequence, to avoid doing partial
493 // RR elimination. If the branch predicates for the two merge differ,
494 // mixing them is unsafe.
495 ClearSequenceProgress();
497 // Conservatively merge the ReleaseMetadata information.
498 if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
499 RRI.ReleaseMetadata = 0;
501 RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe;
502 RRI.IsTailCallRelease = RRI.IsTailCallRelease &&
503 Other.RRI.IsTailCallRelease;
504 RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
506 // Merge the insert point sets. If there are any differences,
507 // that makes this a partial merge.
508 Partial = RRI.ReverseInsertPts.size() != Other.RRI.ReverseInsertPts.size();
509 for (SmallPtrSet<Instruction *, 2>::const_iterator
510 I = Other.RRI.ReverseInsertPts.begin(),
511 E = Other.RRI.ReverseInsertPts.end(); I != E; ++I)
512 Partial |= RRI.ReverseInsertPts.insert(*I);
517 /// \brief Per-BasicBlock state.
519 /// The number of unique control paths from the entry which can reach this
521 unsigned TopDownPathCount;
523 /// The number of unique control paths to exits from this block.
524 unsigned BottomUpPathCount;
526 /// A type for PerPtrTopDown and PerPtrBottomUp.
527 typedef MapVector<const Value *, PtrState> MapTy;
529 /// The top-down traversal uses this to record information known about a
530 /// pointer at the bottom of each block.
533 /// The bottom-up traversal uses this to record information known about a
534 /// pointer at the top of each block.
535 MapTy PerPtrBottomUp;
537 /// Effective predecessors of the current block ignoring ignorable edges and
538 /// ignored backedges.
539 SmallVector<BasicBlock *, 2> Preds;
540 /// Effective successors of the current block ignoring ignorable edges and
541 /// ignored backedges.
542 SmallVector<BasicBlock *, 2> Succs;
545 BBState() : TopDownPathCount(0), BottomUpPathCount(0) {}
547 typedef MapTy::iterator ptr_iterator;
548 typedef MapTy::const_iterator ptr_const_iterator;
550 ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); }
551 ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); }
552 ptr_const_iterator top_down_ptr_begin() const {
553 return PerPtrTopDown.begin();
555 ptr_const_iterator top_down_ptr_end() const {
556 return PerPtrTopDown.end();
559 ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); }
560 ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); }
561 ptr_const_iterator bottom_up_ptr_begin() const {
562 return PerPtrBottomUp.begin();
564 ptr_const_iterator bottom_up_ptr_end() const {
565 return PerPtrBottomUp.end();
568 /// Mark this block as being an entry block, which has one path from the
569 /// entry by definition.
570 void SetAsEntry() { TopDownPathCount = 1; }
572 /// Mark this block as being an exit block, which has one path to an exit by
574 void SetAsExit() { BottomUpPathCount = 1; }
576 PtrState &getPtrTopDownState(const Value *Arg) {
577 return PerPtrTopDown[Arg];
580 PtrState &getPtrBottomUpState(const Value *Arg) {
581 return PerPtrBottomUp[Arg];
584 void clearBottomUpPointers() {
585 PerPtrBottomUp.clear();
588 void clearTopDownPointers() {
589 PerPtrTopDown.clear();
592 void InitFromPred(const BBState &Other);
593 void InitFromSucc(const BBState &Other);
594 void MergePred(const BBState &Other);
595 void MergeSucc(const BBState &Other);
597 /// Return the number of possible unique paths from an entry to an exit
598 /// which pass through this block. This is only valid after both the
599 /// top-down and bottom-up traversals are complete.
600 unsigned GetAllPathCount() const {
601 assert(TopDownPathCount != 0);
602 assert(BottomUpPathCount != 0);
603 return TopDownPathCount * BottomUpPathCount;
606 // Specialized CFG utilities.
607 typedef SmallVectorImpl<BasicBlock *>::const_iterator edge_iterator;
608 edge_iterator pred_begin() { return Preds.begin(); }
609 edge_iterator pred_end() { return Preds.end(); }
610 edge_iterator succ_begin() { return Succs.begin(); }
611 edge_iterator succ_end() { return Succs.end(); }
613 void addSucc(BasicBlock *Succ) { Succs.push_back(Succ); }
614 void addPred(BasicBlock *Pred) { Preds.push_back(Pred); }
616 bool isExit() const { return Succs.empty(); }
620 void BBState::InitFromPred(const BBState &Other) {
621 PerPtrTopDown = Other.PerPtrTopDown;
622 TopDownPathCount = Other.TopDownPathCount;
625 void BBState::InitFromSucc(const BBState &Other) {
626 PerPtrBottomUp = Other.PerPtrBottomUp;
627 BottomUpPathCount = Other.BottomUpPathCount;
630 /// The top-down traversal uses this to merge information about predecessors to
631 /// form the initial state for a new block.
632 void BBState::MergePred(const BBState &Other) {
633 // Other.TopDownPathCount can be 0, in which case it is either dead or a
634 // loop backedge. Loop backedges are special.
635 TopDownPathCount += Other.TopDownPathCount;
637 // Check for overflow. If we have overflow, fall back to conservative
639 if (TopDownPathCount < Other.TopDownPathCount) {
640 clearTopDownPointers();
644 // For each entry in the other set, if our set has an entry with the same key,
645 // merge the entries. Otherwise, copy the entry and merge it with an empty
647 for (ptr_const_iterator MI = Other.top_down_ptr_begin(),
648 ME = Other.top_down_ptr_end(); MI != ME; ++MI) {
649 std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI);
650 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
654 // For each entry in our set, if the other set doesn't have an entry with the
655 // same key, force it to merge with an empty entry.
656 for (ptr_iterator MI = top_down_ptr_begin(),
657 ME = top_down_ptr_end(); MI != ME; ++MI)
658 if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end())
659 MI->second.Merge(PtrState(), /*TopDown=*/true);
662 /// The bottom-up traversal uses this to merge information about successors to
663 /// form the initial state for a new block.
664 void BBState::MergeSucc(const BBState &Other) {
665 // Other.BottomUpPathCount can be 0, in which case it is either dead or a
666 // loop backedge. Loop backedges are special.
667 BottomUpPathCount += Other.BottomUpPathCount;
669 // Check for overflow. If we have overflow, fall back to conservative
671 if (BottomUpPathCount < Other.BottomUpPathCount) {
672 clearBottomUpPointers();
676 // For each entry in the other set, if our set has an entry with the
677 // same key, merge the entries. Otherwise, copy the entry and merge
678 // it with an empty entry.
679 for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(),
680 ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) {
681 std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI);
682 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
686 // For each entry in our set, if the other set doesn't have an entry
687 // with the same key, force it to merge with an empty entry.
688 for (ptr_iterator MI = bottom_up_ptr_begin(),
689 ME = bottom_up_ptr_end(); MI != ME; ++MI)
690 if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end())
691 MI->second.Merge(PtrState(), /*TopDown=*/false);
694 // Only enable ARC Annotations if we are building a debug version of
697 #define ARC_ANNOTATIONS
700 // Define some macros along the lines of DEBUG and some helper functions to make
701 // it cleaner to create annotations in the source code and to no-op when not
702 // building in debug mode.
703 #ifdef ARC_ANNOTATIONS
705 #include "llvm/Support/CommandLine.h"
707 /// Enable/disable ARC sequence annotations.
709 EnableARCAnnotations("enable-objc-arc-annotations", cl::init(false));
711 /// This function appends a unique ARCAnnotationProvenanceSourceMDKind id to an
712 /// instruction so that we can track backwards when post processing via the llvm
713 /// arc annotation processor tool. If the function is an
714 static MDString *AppendMDNodeToSourcePtr(unsigned NodeId,
718 // If pointer is a result of an instruction and it does not have a source
719 // MDNode it, attach a new MDNode onto it. If pointer is a result of
720 // an instruction and does have a source MDNode attached to it, return a
721 // reference to said Node. Otherwise just return 0.
722 if (Instruction *Inst = dyn_cast<Instruction>(Ptr)) {
724 if (!(Node = Inst->getMetadata(NodeId))) {
725 // We do not have any node. Generate and attatch the hash MDString to the
728 // We just use an MDString to ensure that this metadata gets written out
729 // of line at the module level and to provide a very simple format
730 // encoding the information herein. Both of these makes it simpler to
731 // parse the annotations by a simple external program.
733 raw_string_ostream os(Str);
734 os << "(" << Inst->getParent()->getParent()->getName() << ",%"
735 << Inst->getName() << ")";
737 Hash = MDString::get(Inst->getContext(), os.str());
738 Inst->setMetadata(NodeId, MDNode::get(Inst->getContext(),Hash));
740 // We have a node. Grab its hash and return it.
741 assert(Node->getNumOperands() == 1 &&
742 "An ARCAnnotationProvenanceSourceMDKind can only have 1 operand.");
743 Hash = cast<MDString>(Node->getOperand(0));
745 } else if (Argument *Arg = dyn_cast<Argument>(Ptr)) {
747 raw_string_ostream os(str);
748 os << "(" << Arg->getParent()->getName() << ",%" << Arg->getName()
750 Hash = MDString::get(Arg->getContext(), os.str());
756 static std::string SequenceToString(Sequence A) {
758 raw_string_ostream os(str);
763 /// Helper function to change a Sequence into a String object using our overload
764 /// for raw_ostream so we only have printing code in one location.
765 static MDString *SequenceToMDString(LLVMContext &Context,
767 return MDString::get(Context, SequenceToString(A));
770 /// A simple function to generate a MDNode which describes the change in state
771 /// for Value *Ptr caused by Instruction *Inst.
772 static void AppendMDNodeToInstForPtr(unsigned NodeId,
775 MDString *PtrSourceMDNodeID,
779 Value *tmp[3] = {PtrSourceMDNodeID,
780 SequenceToMDString(Inst->getContext(),
782 SequenceToMDString(Inst->getContext(),
784 Node = MDNode::get(Inst->getContext(),
785 ArrayRef<Value*>(tmp, 3));
787 Inst->setMetadata(NodeId, Node);
790 /// Add to the beginning of the basic block llvm.ptr.annotations which show the
791 /// state of a pointer at the entrance to a basic block.
792 static void GenerateARCBBEntranceAnnotation(const char *Name, BasicBlock *BB,
793 Value *Ptr, Sequence Seq) {
794 Module *M = BB->getParent()->getParent();
795 LLVMContext &C = M->getContext();
796 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
797 Type *I8XX = PointerType::getUnqual(I8X);
798 Type *Params[] = {I8XX, I8XX};
799 FunctionType *FTy = FunctionType::get(Type::getVoidTy(C),
800 ArrayRef<Type*>(Params, 2),
802 Constant *Callee = M->getOrInsertFunction(Name, FTy);
804 IRBuilder<> Builder(BB, BB->getFirstInsertionPt());
807 StringRef Tmp = Ptr->getName();
808 if (0 == (PtrName = M->getGlobalVariable(Tmp, true))) {
809 Value *ActualPtrName = Builder.CreateGlobalStringPtr(Tmp,
811 PtrName = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
812 cast<Constant>(ActualPtrName), Tmp);
816 std::string SeqStr = SequenceToString(Seq);
817 if (0 == (S = M->getGlobalVariable(SeqStr, true))) {
818 Value *ActualPtrName = Builder.CreateGlobalStringPtr(SeqStr,
820 S = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
821 cast<Constant>(ActualPtrName), SeqStr);
824 Builder.CreateCall2(Callee, PtrName, S);
827 /// Add to the end of the basic block llvm.ptr.annotations which show the state
828 /// of the pointer at the bottom of the basic block.
829 static void GenerateARCBBTerminatorAnnotation(const char *Name, BasicBlock *BB,
830 Value *Ptr, Sequence Seq) {
831 Module *M = BB->getParent()->getParent();
832 LLVMContext &C = M->getContext();
833 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
834 Type *I8XX = PointerType::getUnqual(I8X);
835 Type *Params[] = {I8XX, I8XX};
836 FunctionType *FTy = FunctionType::get(Type::getVoidTy(C),
837 ArrayRef<Type*>(Params, 2),
839 Constant *Callee = M->getOrInsertFunction(Name, FTy);
841 IRBuilder<> Builder(BB, llvm::prior(BB->end()));
844 StringRef Tmp = Ptr->getName();
845 if (0 == (PtrName = M->getGlobalVariable(Tmp, true))) {
846 Value *ActualPtrName = Builder.CreateGlobalStringPtr(Tmp,
848 PtrName = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
849 cast<Constant>(ActualPtrName), Tmp);
853 std::string SeqStr = SequenceToString(Seq);
854 if (0 == (S = M->getGlobalVariable(SeqStr, true))) {
855 Value *ActualPtrName = Builder.CreateGlobalStringPtr(SeqStr,
857 S = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
858 cast<Constant>(ActualPtrName), SeqStr);
860 Builder.CreateCall2(Callee, PtrName, S);
863 /// Adds a source annotation to pointer and a state change annotation to Inst
864 /// referencing the source annotation and the old/new state of pointer.
865 static void GenerateARCAnnotation(unsigned InstMDId,
871 if (EnableARCAnnotations) {
872 // First generate the source annotation on our pointer. This will return an
873 // MDString* if Ptr actually comes from an instruction implying we can put
874 // in a source annotation. If AppendMDNodeToSourcePtr returns 0 (i.e. NULL),
875 // then we know that our pointer is from an Argument so we put a reference
876 // to the argument number.
878 // The point of this is to make it easy for the
879 // llvm-arc-annotation-processor tool to cross reference where the source
880 // pointer is in the LLVM IR since the LLVM IR parser does not submit such
881 // information via debug info for backends to use (since why would anyone
882 // need such a thing from LLVM IR besides in non standard cases
884 MDString *SourcePtrMDNode =
885 AppendMDNodeToSourcePtr(PtrMDId, Ptr);
886 AppendMDNodeToInstForPtr(InstMDId, Inst, Ptr, SourcePtrMDNode, OldSeq,
891 // The actual interface for accessing the above functionality is defined via
892 // some simple macros which are defined below. We do this so that the user does
893 // not need to pass in what metadata id is needed resulting in cleaner code and
894 // additionally since it provides an easy way to conditionally no-op all
895 // annotation support in a non-debug build.
897 /// Use this macro to annotate a sequence state change when processing
898 /// instructions bottom up,
899 #define ANNOTATE_BOTTOMUP(inst, ptr, old, new) \
900 GenerateARCAnnotation(ARCAnnotationBottomUpMDKind, \
901 ARCAnnotationProvenanceSourceMDKind, (inst), \
902 const_cast<Value*>(ptr), (old), (new))
903 /// Use this macro to annotate a sequence state change when processing
904 /// instructions top down.
905 #define ANNOTATE_TOPDOWN(inst, ptr, old, new) \
906 GenerateARCAnnotation(ARCAnnotationTopDownMDKind, \
907 ARCAnnotationProvenanceSourceMDKind, (inst), \
908 const_cast<Value*>(ptr), (old), (new))
910 #else // !ARC_ANNOTATION
911 // If annotations are off, noop.
912 #define ANNOTATE_BOTTOMUP(inst, ptr, old, new)
913 #define ANNOTATE_TOPDOWN(inst, ptr, old, new)
914 #endif // !ARC_ANNOTATION
917 /// \brief The main ARC optimization pass.
918 class ObjCARCOpt : public FunctionPass {
920 ProvenanceAnalysis PA;
922 /// A flag indicating whether this optimization pass should run.
925 /// Declarations for ObjC runtime functions, for use in creating calls to
926 /// them. These are initialized lazily to avoid cluttering up the Module
927 /// with unused declarations.
929 /// Declaration for ObjC runtime function
930 /// objc_retainAutoreleasedReturnValue.
931 Constant *RetainRVCallee;
932 /// Declaration for ObjC runtime function objc_autoreleaseReturnValue.
933 Constant *AutoreleaseRVCallee;
934 /// Declaration for ObjC runtime function objc_release.
935 Constant *ReleaseCallee;
936 /// Declaration for ObjC runtime function objc_retain.
937 Constant *RetainCallee;
938 /// Declaration for ObjC runtime function objc_retainBlock.
939 Constant *RetainBlockCallee;
940 /// Declaration for ObjC runtime function objc_autorelease.
941 Constant *AutoreleaseCallee;
943 /// Flags which determine whether each of the interesting runtine functions
944 /// is in fact used in the current function.
945 unsigned UsedInThisFunction;
947 /// The Metadata Kind for clang.imprecise_release metadata.
948 unsigned ImpreciseReleaseMDKind;
950 /// The Metadata Kind for clang.arc.copy_on_escape metadata.
951 unsigned CopyOnEscapeMDKind;
953 /// The Metadata Kind for clang.arc.no_objc_arc_exceptions metadata.
954 unsigned NoObjCARCExceptionsMDKind;
956 #ifdef ARC_ANNOTATIONS
957 /// The Metadata Kind for llvm.arc.annotation.bottomup metadata.
958 unsigned ARCAnnotationBottomUpMDKind;
959 /// The Metadata Kind for llvm.arc.annotation.topdown metadata.
960 unsigned ARCAnnotationTopDownMDKind;
961 /// The Metadata Kind for llvm.arc.annotation.provenancesource metadata.
962 unsigned ARCAnnotationProvenanceSourceMDKind;
963 #endif // ARC_ANNOATIONS
965 Constant *getRetainRVCallee(Module *M);
966 Constant *getAutoreleaseRVCallee(Module *M);
967 Constant *getReleaseCallee(Module *M);
968 Constant *getRetainCallee(Module *M);
969 Constant *getRetainBlockCallee(Module *M);
970 Constant *getAutoreleaseCallee(Module *M);
972 bool IsRetainBlockOptimizable(const Instruction *Inst);
974 void OptimizeRetainCall(Function &F, Instruction *Retain);
975 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
976 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
977 InstructionClass &Class);
978 bool OptimizeRetainBlockCall(Function &F, Instruction *RetainBlock,
979 InstructionClass &Class);
980 void OptimizeIndividualCalls(Function &F);
982 void CheckForCFGHazards(const BasicBlock *BB,
983 DenseMap<const BasicBlock *, BBState> &BBStates,
984 BBState &MyStates) const;
985 bool VisitInstructionBottomUp(Instruction *Inst,
987 MapVector<Value *, RRInfo> &Retains,
989 bool VisitBottomUp(BasicBlock *BB,
990 DenseMap<const BasicBlock *, BBState> &BBStates,
991 MapVector<Value *, RRInfo> &Retains);
992 bool VisitInstructionTopDown(Instruction *Inst,
993 DenseMap<Value *, RRInfo> &Releases,
995 bool VisitTopDown(BasicBlock *BB,
996 DenseMap<const BasicBlock *, BBState> &BBStates,
997 DenseMap<Value *, RRInfo> &Releases);
998 bool Visit(Function &F,
999 DenseMap<const BasicBlock *, BBState> &BBStates,
1000 MapVector<Value *, RRInfo> &Retains,
1001 DenseMap<Value *, RRInfo> &Releases);
1003 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1004 MapVector<Value *, RRInfo> &Retains,
1005 DenseMap<Value *, RRInfo> &Releases,
1006 SmallVectorImpl<Instruction *> &DeadInsts,
1009 bool ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState> &BBStates,
1010 MapVector<Value *, RRInfo> &Retains,
1011 DenseMap<Value *, RRInfo> &Releases,
1013 SmallVector<Instruction *, 4> &NewRetains,
1014 SmallVector<Instruction *, 4> &NewReleases,
1015 SmallVector<Instruction *, 8> &DeadInsts,
1016 RRInfo &RetainsToMove,
1017 RRInfo &ReleasesToMove,
1020 bool &AnyPairsCompletelyEliminated);
1022 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1023 MapVector<Value *, RRInfo> &Retains,
1024 DenseMap<Value *, RRInfo> &Releases,
1027 void OptimizeWeakCalls(Function &F);
1029 bool OptimizeSequences(Function &F);
1031 void OptimizeReturns(Function &F);
1033 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1034 virtual bool doInitialization(Module &M);
1035 virtual bool runOnFunction(Function &F);
1036 virtual void releaseMemory();
1040 ObjCARCOpt() : FunctionPass(ID) {
1041 initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1046 char ObjCARCOpt::ID = 0;
1047 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1048 "objc-arc", "ObjC ARC optimization", false, false)
1049 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1050 INITIALIZE_PASS_END(ObjCARCOpt,
1051 "objc-arc", "ObjC ARC optimization", false, false)
1053 Pass *llvm::createObjCARCOptPass() {
1054 return new ObjCARCOpt();
1057 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1058 AU.addRequired<ObjCARCAliasAnalysis>();
1059 AU.addRequired<AliasAnalysis>();
1060 // ARC optimization doesn't currently split critical edges.
1061 AU.setPreservesCFG();
1064 bool ObjCARCOpt::IsRetainBlockOptimizable(const Instruction *Inst) {
1065 // Without the magic metadata tag, we have to assume this might be an
1066 // objc_retainBlock call inserted to convert a block pointer to an id,
1067 // in which case it really is needed.
1068 if (!Inst->getMetadata(CopyOnEscapeMDKind))
1071 // If the pointer "escapes" (not including being used in a call),
1072 // the copy may be needed.
1073 if (DoesRetainableObjPtrEscape(Inst))
1076 // Otherwise, it's not needed.
1080 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1081 if (!RetainRVCallee) {
1082 LLVMContext &C = M->getContext();
1083 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1084 Type *Params[] = { I8X };
1085 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1086 AttributeSet Attribute =
1087 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1088 Attribute::NoUnwind);
1090 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1093 return RetainRVCallee;
1096 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1097 if (!AutoreleaseRVCallee) {
1098 LLVMContext &C = M->getContext();
1099 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1100 Type *Params[] = { I8X };
1101 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1102 AttributeSet Attribute =
1103 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1104 Attribute::NoUnwind);
1105 AutoreleaseRVCallee =
1106 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1109 return AutoreleaseRVCallee;
1112 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1113 if (!ReleaseCallee) {
1114 LLVMContext &C = M->getContext();
1115 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1116 AttributeSet Attribute =
1117 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1118 Attribute::NoUnwind);
1120 M->getOrInsertFunction(
1122 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1125 return ReleaseCallee;
1128 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1129 if (!RetainCallee) {
1130 LLVMContext &C = M->getContext();
1131 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1132 AttributeSet Attribute =
1133 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1134 Attribute::NoUnwind);
1136 M->getOrInsertFunction(
1138 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1141 return RetainCallee;
1144 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1145 if (!RetainBlockCallee) {
1146 LLVMContext &C = M->getContext();
1147 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1148 // objc_retainBlock is not nounwind because it calls user copy constructors
1149 // which could theoretically throw.
1151 M->getOrInsertFunction(
1153 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1156 return RetainBlockCallee;
1159 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1160 if (!AutoreleaseCallee) {
1161 LLVMContext &C = M->getContext();
1162 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1163 AttributeSet Attribute =
1164 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1165 Attribute::NoUnwind);
1167 M->getOrInsertFunction(
1169 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1172 return AutoreleaseCallee;
1175 /// Turn objc_retain into objc_retainAutoreleasedReturnValue if the operand is a
1178 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
1179 ImmutableCallSite CS(GetObjCArg(Retain));
1180 const Instruction *Call = CS.getInstruction();
1182 if (Call->getParent() != Retain->getParent()) return;
1184 // Check that the call is next to the retain.
1185 BasicBlock::const_iterator I = Call;
1187 while (IsNoopInstruction(I)) ++I;
1191 // Turn it to an objc_retainAutoreleasedReturnValue..
1195 DEBUG(dbgs() << "ObjCARCOpt::OptimizeRetainCall: Transforming "
1196 "objc_retain => objc_retainAutoreleasedReturnValue"
1197 " since the operand is a return value.\n"
1199 << *Retain << "\n");
1201 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
1203 DEBUG(dbgs() << " New: "
1204 << *Retain << "\n");
1207 /// Turn objc_retainAutoreleasedReturnValue into objc_retain if the operand is
1208 /// not a return value. Or, if it can be paired with an
1209 /// objc_autoreleaseReturnValue, delete the pair and return true.
1211 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
1212 // Check for the argument being from an immediately preceding call or invoke.
1213 const Value *Arg = GetObjCArg(RetainRV);
1214 ImmutableCallSite CS(Arg);
1215 if (const Instruction *Call = CS.getInstruction()) {
1216 if (Call->getParent() == RetainRV->getParent()) {
1217 BasicBlock::const_iterator I = Call;
1219 while (IsNoopInstruction(I)) ++I;
1220 if (&*I == RetainRV)
1222 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
1223 BasicBlock *RetainRVParent = RetainRV->getParent();
1224 if (II->getNormalDest() == RetainRVParent) {
1225 BasicBlock::const_iterator I = RetainRVParent->begin();
1226 while (IsNoopInstruction(I)) ++I;
1227 if (&*I == RetainRV)
1233 // Check for being preceded by an objc_autoreleaseReturnValue on the same
1234 // pointer. In this case, we can delete the pair.
1235 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
1237 do --I; while (I != Begin && IsNoopInstruction(I));
1238 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
1239 GetObjCArg(I) == Arg) {
1243 DEBUG(dbgs() << "ObjCARCOpt::OptimizeRetainRVCall: Erasing " << *I << "\n"
1244 << " Erasing " << *RetainRV
1247 EraseInstruction(I);
1248 EraseInstruction(RetainRV);
1253 // Turn it to a plain objc_retain.
1257 DEBUG(dbgs() << "ObjCARCOpt::OptimizeRetainRVCall: Transforming "
1258 "objc_retainAutoreleasedReturnValue => "
1259 "objc_retain since the operand is not a return value.\n"
1261 << *RetainRV << "\n");
1263 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
1265 DEBUG(dbgs() << " New: "
1266 << *RetainRV << "\n");
1271 /// Turn objc_autoreleaseReturnValue into objc_autorelease if the result is not
1272 /// used as a return value.
1274 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
1275 InstructionClass &Class) {
1276 // Check for a return of the pointer value.
1277 const Value *Ptr = GetObjCArg(AutoreleaseRV);
1278 SmallVector<const Value *, 2> Users;
1279 Users.push_back(Ptr);
1281 Ptr = Users.pop_back_val();
1282 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
1284 const User *I = *UI;
1285 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
1287 if (isa<BitCastInst>(I))
1290 } while (!Users.empty());
1295 DEBUG(dbgs() << "ObjCARCOpt::OptimizeAutoreleaseRVCall: Transforming "
1296 "objc_autoreleaseReturnValue => "
1297 "objc_autorelease since its operand is not used as a return "
1300 << *AutoreleaseRV << "\n");
1302 CallInst *AutoreleaseRVCI = cast<CallInst>(AutoreleaseRV);
1304 setCalledFunction(getAutoreleaseCallee(F.getParent()));
1305 AutoreleaseRVCI->setTailCall(false); // Never tail call objc_autorelease.
1306 Class = IC_Autorelease;
1308 DEBUG(dbgs() << " New: "
1309 << *AutoreleaseRV << "\n");
1313 // \brief Attempt to strength reduce objc_retainBlock calls to objc_retain
1316 // Specifically: If an objc_retainBlock call has the copy_on_escape metadata and
1317 // does not escape (following the rules of block escaping), strength reduce the
1318 // objc_retainBlock to an objc_retain.
1320 // TODO: If an objc_retainBlock call is dominated period by a previous
1321 // objc_retainBlock call, strength reduce the objc_retainBlock to an
1324 ObjCARCOpt::OptimizeRetainBlockCall(Function &F, Instruction *Inst,
1325 InstructionClass &Class) {
1326 assert(GetBasicInstructionClass(Inst) == Class);
1327 assert(IC_RetainBlock == Class);
1329 // If we can not optimize Inst, return false.
1330 if (!IsRetainBlockOptimizable(Inst))
1333 CallInst *RetainBlock = cast<CallInst>(Inst);
1334 RetainBlock->setCalledFunction(getRetainCallee(F.getParent()));
1335 // Remove copy_on_escape metadata.
1336 RetainBlock->setMetadata(CopyOnEscapeMDKind, 0);
1342 /// Visit each call, one at a time, and make simplifications without doing any
1343 /// additional analysis.
1344 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
1345 // Reset all the flags in preparation for recomputing them.
1346 UsedInThisFunction = 0;
1348 // Visit all objc_* calls in F.
1349 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
1350 Instruction *Inst = &*I++;
1352 InstructionClass Class = GetBasicInstructionClass(Inst);
1354 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Visiting: Class: "
1355 << Class << "; " << *Inst << "\n");
1360 // Delete no-op casts. These function calls have special semantics, but
1361 // the semantics are entirely implemented via lowering in the front-end,
1362 // so by the time they reach the optimizer, they are just no-op calls
1363 // which return their argument.
1365 // There are gray areas here, as the ability to cast reference-counted
1366 // pointers to raw void* and back allows code to break ARC assumptions,
1367 // however these are currently considered to be unimportant.
1371 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Erasing no-op cast:"
1372 " " << *Inst << "\n");
1373 EraseInstruction(Inst);
1376 // If the pointer-to-weak-pointer is null, it's undefined behavior.
1379 case IC_LoadWeakRetained:
1381 case IC_DestroyWeak: {
1382 CallInst *CI = cast<CallInst>(Inst);
1383 if (IsNullOrUndef(CI->getArgOperand(0))) {
1385 Type *Ty = CI->getArgOperand(0)->getType();
1386 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1387 Constant::getNullValue(Ty),
1389 llvm::Value *NewValue = UndefValue::get(CI->getType());
1390 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: A null "
1391 "pointer-to-weak-pointer is undefined behavior.\n"
1395 CI->replaceAllUsesWith(NewValue);
1396 CI->eraseFromParent();
1403 CallInst *CI = cast<CallInst>(Inst);
1404 if (IsNullOrUndef(CI->getArgOperand(0)) ||
1405 IsNullOrUndef(CI->getArgOperand(1))) {
1407 Type *Ty = CI->getArgOperand(0)->getType();
1408 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1409 Constant::getNullValue(Ty),
1412 llvm::Value *NewValue = UndefValue::get(CI->getType());
1413 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: A null "
1414 "pointer-to-weak-pointer is undefined behavior.\n"
1419 CI->replaceAllUsesWith(NewValue);
1420 CI->eraseFromParent();
1425 case IC_RetainBlock:
1426 // If we strength reduce an objc_retainBlock to amn objc_retain, continue
1427 // onto the objc_retain peephole optimizations. Otherwise break.
1428 if (!OptimizeRetainBlockCall(F, Inst, Class))
1432 OptimizeRetainCall(F, Inst);
1435 if (OptimizeRetainRVCall(F, Inst))
1438 case IC_AutoreleaseRV:
1439 OptimizeAutoreleaseRVCall(F, Inst, Class);
1443 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
1444 if (IsAutorelease(Class) && Inst->use_empty()) {
1445 CallInst *Call = cast<CallInst>(Inst);
1446 const Value *Arg = Call->getArgOperand(0);
1447 Arg = FindSingleUseIdentifiedObject(Arg);
1452 // Create the declaration lazily.
1453 LLVMContext &C = Inst->getContext();
1455 CallInst::Create(getReleaseCallee(F.getParent()),
1456 Call->getArgOperand(0), "", Call);
1457 NewCall->setMetadata(ImpreciseReleaseMDKind,
1458 MDNode::get(C, ArrayRef<Value *>()));
1460 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Replacing "
1461 "objc_autorelease(x) with objc_release(x) since x is "
1462 "otherwise unused.\n"
1463 " Old: " << *Call <<
1467 EraseInstruction(Call);
1473 // For functions which can never be passed stack arguments, add
1475 if (IsAlwaysTail(Class)) {
1477 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Adding tail keyword"
1478 " to function since it can never be passed stack args: " << *Inst <<
1480 cast<CallInst>(Inst)->setTailCall();
1483 // Ensure that functions that can never have a "tail" keyword due to the
1484 // semantics of ARC truly do not do so.
1485 if (IsNeverTail(Class)) {
1487 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Removing tail "
1488 "keyword from function: " << *Inst <<
1490 cast<CallInst>(Inst)->setTailCall(false);
1493 // Set nounwind as needed.
1494 if (IsNoThrow(Class)) {
1496 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Found no throw"
1497 " class. Setting nounwind on: " << *Inst << "\n");
1498 cast<CallInst>(Inst)->setDoesNotThrow();
1501 if (!IsNoopOnNull(Class)) {
1502 UsedInThisFunction |= 1 << Class;
1506 const Value *Arg = GetObjCArg(Inst);
1508 // ARC calls with null are no-ops. Delete them.
1509 if (IsNullOrUndef(Arg)) {
1512 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: ARC calls with "
1513 " null are no-ops. Erasing: " << *Inst << "\n");
1514 EraseInstruction(Inst);
1518 // Keep track of which of retain, release, autorelease, and retain_block
1519 // are actually present in this function.
1520 UsedInThisFunction |= 1 << Class;
1522 // If Arg is a PHI, and one or more incoming values to the
1523 // PHI are null, and the call is control-equivalent to the PHI, and there
1524 // are no relevant side effects between the PHI and the call, the call
1525 // could be pushed up to just those paths with non-null incoming values.
1526 // For now, don't bother splitting critical edges for this.
1527 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
1528 Worklist.push_back(std::make_pair(Inst, Arg));
1530 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
1534 const PHINode *PN = dyn_cast<PHINode>(Arg);
1537 // Determine if the PHI has any null operands, or any incoming
1539 bool HasNull = false;
1540 bool HasCriticalEdges = false;
1541 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1543 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
1544 if (IsNullOrUndef(Incoming))
1546 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
1547 .getNumSuccessors() != 1) {
1548 HasCriticalEdges = true;
1552 // If we have null operands and no critical edges, optimize.
1553 if (!HasCriticalEdges && HasNull) {
1554 SmallPtrSet<Instruction *, 4> DependingInstructions;
1555 SmallPtrSet<const BasicBlock *, 4> Visited;
1557 // Check that there is nothing that cares about the reference
1558 // count between the call and the phi.
1561 case IC_RetainBlock:
1562 // These can always be moved up.
1565 // These can't be moved across things that care about the retain
1567 FindDependencies(NeedsPositiveRetainCount, Arg,
1568 Inst->getParent(), Inst,
1569 DependingInstructions, Visited, PA);
1571 case IC_Autorelease:
1572 // These can't be moved across autorelease pool scope boundaries.
1573 FindDependencies(AutoreleasePoolBoundary, Arg,
1574 Inst->getParent(), Inst,
1575 DependingInstructions, Visited, PA);
1578 case IC_AutoreleaseRV:
1579 // Don't move these; the RV optimization depends on the autoreleaseRV
1580 // being tail called, and the retainRV being immediately after a call
1581 // (which might still happen if we get lucky with codegen layout, but
1582 // it's not worth taking the chance).
1585 llvm_unreachable("Invalid dependence flavor");
1588 if (DependingInstructions.size() == 1 &&
1589 *DependingInstructions.begin() == PN) {
1592 // Clone the call into each predecessor that has a non-null value.
1593 CallInst *CInst = cast<CallInst>(Inst);
1594 Type *ParamTy = CInst->getArgOperand(0)->getType();
1595 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1597 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
1598 if (!IsNullOrUndef(Incoming)) {
1599 CallInst *Clone = cast<CallInst>(CInst->clone());
1600 Value *Op = PN->getIncomingValue(i);
1601 Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
1602 if (Op->getType() != ParamTy)
1603 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
1604 Clone->setArgOperand(0, Op);
1605 Clone->insertBefore(InsertPos);
1607 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Cloning "
1610 "clone at " << *InsertPos << "\n");
1611 Worklist.push_back(std::make_pair(Clone, Incoming));
1614 // Erase the original call.
1615 DEBUG(dbgs() << "Erasing: " << *CInst << "\n");
1616 EraseInstruction(CInst);
1620 } while (!Worklist.empty());
1622 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Finished List.\n");
1625 /// Check for critical edges, loop boundaries, irreducible control flow, or
1626 /// other CFG structures where moving code across the edge would result in it
1627 /// being executed more.
1629 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
1630 DenseMap<const BasicBlock *, BBState> &BBStates,
1631 BBState &MyStates) const {
1632 // If any top-down local-use or possible-dec has a succ which is earlier in
1633 // the sequence, forget it.
1634 for (BBState::ptr_iterator I = MyStates.top_down_ptr_begin(),
1635 E = MyStates.top_down_ptr_end(); I != E; ++I)
1636 switch (I->second.GetSeq()) {
1639 const Value *Arg = I->first;
1640 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1641 bool SomeSuccHasSame = false;
1642 bool AllSuccsHaveSame = true;
1643 PtrState &S = I->second;
1644 succ_const_iterator SI(TI), SE(TI, false);
1646 for (; SI != SE; ++SI) {
1647 Sequence SuccSSeq = S_None;
1648 bool SuccSRRIKnownSafe = false;
1649 // If VisitBottomUp has pointer information for this successor, take
1650 // what we know about it.
1651 DenseMap<const BasicBlock *, BBState>::iterator BBI =
1653 assert(BBI != BBStates.end());
1654 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
1655 SuccSSeq = SuccS.GetSeq();
1656 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
1659 case S_CanRelease: {
1660 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
1661 S.ClearSequenceProgress();
1667 SomeSuccHasSame = true;
1671 case S_MovableRelease:
1672 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
1673 AllSuccsHaveSame = false;
1676 llvm_unreachable("bottom-up pointer in retain state!");
1679 // If the state at the other end of any of the successor edges
1680 // matches the current state, require all edges to match. This
1681 // guards against loops in the middle of a sequence.
1682 if (SomeSuccHasSame && !AllSuccsHaveSame)
1683 S.ClearSequenceProgress();
1686 case S_CanRelease: {
1687 const Value *Arg = I->first;
1688 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1689 bool SomeSuccHasSame = false;
1690 bool AllSuccsHaveSame = true;
1691 PtrState &S = I->second;
1692 succ_const_iterator SI(TI), SE(TI, false);
1694 for (; SI != SE; ++SI) {
1695 Sequence SuccSSeq = S_None;
1696 bool SuccSRRIKnownSafe = false;
1697 // If VisitBottomUp has pointer information for this successor, take
1698 // what we know about it.
1699 DenseMap<const BasicBlock *, BBState>::iterator BBI =
1701 assert(BBI != BBStates.end());
1702 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
1703 SuccSSeq = SuccS.GetSeq();
1704 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
1707 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
1708 S.ClearSequenceProgress();
1714 SomeSuccHasSame = true;
1718 case S_MovableRelease:
1720 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
1721 AllSuccsHaveSame = false;
1724 llvm_unreachable("bottom-up pointer in retain state!");
1727 // If the state at the other end of any of the successor edges
1728 // matches the current state, require all edges to match. This
1729 // guards against loops in the middle of a sequence.
1730 if (SomeSuccHasSame && !AllSuccsHaveSame)
1731 S.ClearSequenceProgress();
1738 ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst,
1740 MapVector<Value *, RRInfo> &Retains,
1741 BBState &MyStates) {
1742 bool NestingDetected = false;
1743 InstructionClass Class = GetInstructionClass(Inst);
1744 const Value *Arg = 0;
1748 Arg = GetObjCArg(Inst);
1750 PtrState &S = MyStates.getPtrBottomUpState(Arg);
1752 // If we see two releases in a row on the same pointer. If so, make
1753 // a note, and we'll cicle back to revisit it after we've
1754 // hopefully eliminated the second release, which may allow us to
1755 // eliminate the first release too.
1756 // Theoretically we could implement removal of nested retain+release
1757 // pairs by making PtrState hold a stack of states, but this is
1758 // simple and avoids adding overhead for the non-nested case.
1759 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease) {
1760 DEBUG(dbgs() << "ObjCARCOpt::VisitInstructionBottomUp: Found nested "
1761 "releases (i.e. a release pair)\n");
1762 NestingDetected = true;
1765 MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
1766 Sequence NewSeq = ReleaseMetadata ? S_MovableRelease : S_Release;
1767 ANNOTATE_BOTTOMUP(Inst, Arg, S.GetSeq(), NewSeq);
1768 S.ResetSequenceProgress(NewSeq);
1769 S.RRI.ReleaseMetadata = ReleaseMetadata;
1770 S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
1771 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
1772 S.RRI.Calls.insert(Inst);
1773 S.SetKnownPositiveRefCount();
1776 case IC_RetainBlock:
1777 // In OptimizeIndividualCalls, we have strength reduced all optimizable
1778 // objc_retainBlocks to objc_retains. Thus at this point any
1779 // objc_retainBlocks that we see are not optimizable.
1783 Arg = GetObjCArg(Inst);
1785 PtrState &S = MyStates.getPtrBottomUpState(Arg);
1786 S.SetKnownPositiveRefCount();
1788 Sequence OldSeq = S.GetSeq();
1792 case S_MovableRelease:
1794 S.RRI.ReverseInsertPts.clear();
1797 // Don't do retain+release tracking for IC_RetainRV, because it's
1798 // better to let it remain as the first instruction after a call.
1799 if (Class != IC_RetainRV)
1800 Retains[Inst] = S.RRI;
1801 S.ClearSequenceProgress();
1806 llvm_unreachable("bottom-up pointer in retain state!");
1808 ANNOTATE_BOTTOMUP(Inst, Arg, OldSeq, S.GetSeq());
1809 return NestingDetected;
1811 case IC_AutoreleasepoolPop:
1812 // Conservatively, clear MyStates for all known pointers.
1813 MyStates.clearBottomUpPointers();
1814 return NestingDetected;
1815 case IC_AutoreleasepoolPush:
1817 // These are irrelevant.
1818 return NestingDetected;
1823 // Consider any other possible effects of this instruction on each
1824 // pointer being tracked.
1825 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
1826 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
1827 const Value *Ptr = MI->first;
1829 continue; // Handled above.
1830 PtrState &S = MI->second;
1831 Sequence Seq = S.GetSeq();
1833 // Check for possible releases.
1834 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
1835 S.ClearKnownPositiveRefCount();
1838 S.SetSeq(S_CanRelease);
1839 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S.GetSeq());
1843 case S_MovableRelease:
1848 llvm_unreachable("bottom-up pointer in retain state!");
1852 // Check for possible direct uses.
1855 case S_MovableRelease:
1856 if (CanUse(Inst, Ptr, PA, Class)) {
1857 assert(S.RRI.ReverseInsertPts.empty());
1858 // If this is an invoke instruction, we're scanning it as part of
1859 // one of its successor blocks, since we can't insert code after it
1860 // in its own block, and we don't want to split critical edges.
1861 if (isa<InvokeInst>(Inst))
1862 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
1864 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
1866 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S_Use);
1867 } else if (Seq == S_Release && IsUser(Class)) {
1868 // Non-movable releases depend on any possible objc pointer use.
1870 ANNOTATE_BOTTOMUP(Inst, Ptr, S_Release, S_Stop);
1871 assert(S.RRI.ReverseInsertPts.empty());
1872 // As above; handle invoke specially.
1873 if (isa<InvokeInst>(Inst))
1874 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
1876 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
1880 if (CanUse(Inst, Ptr, PA, Class)) {
1882 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S_Use);
1890 llvm_unreachable("bottom-up pointer in retain state!");
1894 return NestingDetected;
1898 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
1899 DenseMap<const BasicBlock *, BBState> &BBStates,
1900 MapVector<Value *, RRInfo> &Retains) {
1901 bool NestingDetected = false;
1902 BBState &MyStates = BBStates[BB];
1904 // Merge the states from each successor to compute the initial state
1905 // for the current block.
1906 BBState::edge_iterator SI(MyStates.succ_begin()),
1907 SE(MyStates.succ_end());
1909 const BasicBlock *Succ = *SI;
1910 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
1911 assert(I != BBStates.end());
1912 MyStates.InitFromSucc(I->second);
1914 for (; SI != SE; ++SI) {
1916 I = BBStates.find(Succ);
1917 assert(I != BBStates.end());
1918 MyStates.MergeSucc(I->second);
1922 #ifdef ARC_ANNOTATIONS
1923 if (EnableARCAnnotations) {
1924 // If ARC Annotations are enabled, output the current state of pointers at the
1925 // bottom of the basic block.
1926 for(BBState::ptr_const_iterator I = MyStates.bottom_up_ptr_begin(),
1927 E = MyStates.bottom_up_ptr_end(); I != E; ++I) {
1928 Value *Ptr = const_cast<Value*>(I->first);
1929 Sequence Seq = I->second.GetSeq();
1930 GenerateARCBBTerminatorAnnotation("llvm.arc.annotation.bottomup.bbend",
1937 // Visit all the instructions, bottom-up.
1938 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
1939 Instruction *Inst = llvm::prior(I);
1941 // Invoke instructions are visited as part of their successors (below).
1942 if (isa<InvokeInst>(Inst))
1945 DEBUG(dbgs() << "ObjCARCOpt::VisitButtonUp: Visiting " << *Inst << "\n");
1947 NestingDetected |= VisitInstructionBottomUp(Inst, BB, Retains, MyStates);
1950 // If there's a predecessor with an invoke, visit the invoke as if it were
1951 // part of this block, since we can't insert code after an invoke in its own
1952 // block, and we don't want to split critical edges.
1953 for (BBState::edge_iterator PI(MyStates.pred_begin()),
1954 PE(MyStates.pred_end()); PI != PE; ++PI) {
1955 BasicBlock *Pred = *PI;
1956 if (InvokeInst *II = dyn_cast<InvokeInst>(&Pred->back()))
1957 NestingDetected |= VisitInstructionBottomUp(II, BB, Retains, MyStates);
1960 #ifdef ARC_ANNOTATIONS
1961 if (EnableARCAnnotations) {
1962 // If ARC Annotations are enabled, output the current state of pointers at the
1963 // top of the basic block.
1964 for(BBState::ptr_const_iterator I = MyStates.bottom_up_ptr_begin(),
1965 E = MyStates.bottom_up_ptr_end(); I != E; ++I) {
1966 Value *Ptr = const_cast<Value*>(I->first);
1967 Sequence Seq = I->second.GetSeq();
1968 GenerateARCBBEntranceAnnotation("llvm.arc.annotation.bottomup.bbstart",
1974 return NestingDetected;
1978 ObjCARCOpt::VisitInstructionTopDown(Instruction *Inst,
1979 DenseMap<Value *, RRInfo> &Releases,
1980 BBState &MyStates) {
1981 bool NestingDetected = false;
1982 InstructionClass Class = GetInstructionClass(Inst);
1983 const Value *Arg = 0;
1986 case IC_RetainBlock:
1987 // In OptimizeIndividualCalls, we have strength reduced all optimizable
1988 // objc_retainBlocks to objc_retains. Thus at this point any
1989 // objc_retainBlocks that we see are not optimizable.
1993 Arg = GetObjCArg(Inst);
1995 PtrState &S = MyStates.getPtrTopDownState(Arg);
1997 // Don't do retain+release tracking for IC_RetainRV, because it's
1998 // better to let it remain as the first instruction after a call.
1999 if (Class != IC_RetainRV) {
2000 // If we see two retains in a row on the same pointer. If so, make
2001 // a note, and we'll cicle back to revisit it after we've
2002 // hopefully eliminated the second retain, which may allow us to
2003 // eliminate the first retain too.
2004 // Theoretically we could implement removal of nested retain+release
2005 // pairs by making PtrState hold a stack of states, but this is
2006 // simple and avoids adding overhead for the non-nested case.
2007 if (S.GetSeq() == S_Retain)
2008 NestingDetected = true;
2010 ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_Retain);
2011 S.ResetSequenceProgress(S_Retain);
2012 S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
2013 S.RRI.Calls.insert(Inst);
2016 S.SetKnownPositiveRefCount();
2018 // A retain can be a potential use; procede to the generic checking
2023 Arg = GetObjCArg(Inst);
2025 PtrState &S = MyStates.getPtrTopDownState(Arg);
2026 S.ClearKnownPositiveRefCount();
2028 switch (S.GetSeq()) {
2031 S.RRI.ReverseInsertPts.clear();
2034 S.RRI.ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2035 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2036 Releases[Inst] = S.RRI;
2037 ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_None);
2038 S.ClearSequenceProgress();
2044 case S_MovableRelease:
2045 llvm_unreachable("top-down pointer in release state!");
2049 case IC_AutoreleasepoolPop:
2050 // Conservatively, clear MyStates for all known pointers.
2051 MyStates.clearTopDownPointers();
2052 return NestingDetected;
2053 case IC_AutoreleasepoolPush:
2055 // These are irrelevant.
2056 return NestingDetected;
2061 // Consider any other possible effects of this instruction on each
2062 // pointer being tracked.
2063 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
2064 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
2065 const Value *Ptr = MI->first;
2067 continue; // Handled above.
2068 PtrState &S = MI->second;
2069 Sequence Seq = S.GetSeq();
2071 // Check for possible releases.
2072 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2073 S.ClearKnownPositiveRefCount();
2076 S.SetSeq(S_CanRelease);
2077 ANNOTATE_TOPDOWN(Inst, Ptr, Seq, S_CanRelease);
2078 assert(S.RRI.ReverseInsertPts.empty());
2079 S.RRI.ReverseInsertPts.insert(Inst);
2081 // One call can't cause a transition from S_Retain to S_CanRelease
2082 // and S_CanRelease to S_Use. If we've made the first transition,
2091 case S_MovableRelease:
2092 llvm_unreachable("top-down pointer in release state!");
2096 // Check for possible direct uses.
2099 if (CanUse(Inst, Ptr, PA, Class)) {
2101 ANNOTATE_TOPDOWN(Inst, Ptr, Seq, S_Use);
2110 case S_MovableRelease:
2111 llvm_unreachable("top-down pointer in release state!");
2115 return NestingDetected;
2119 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
2120 DenseMap<const BasicBlock *, BBState> &BBStates,
2121 DenseMap<Value *, RRInfo> &Releases) {
2122 bool NestingDetected = false;
2123 BBState &MyStates = BBStates[BB];
2125 // Merge the states from each predecessor to compute the initial state
2126 // for the current block.
2127 BBState::edge_iterator PI(MyStates.pred_begin()),
2128 PE(MyStates.pred_end());
2130 const BasicBlock *Pred = *PI;
2131 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
2132 assert(I != BBStates.end());
2133 MyStates.InitFromPred(I->second);
2135 for (; PI != PE; ++PI) {
2137 I = BBStates.find(Pred);
2138 assert(I != BBStates.end());
2139 MyStates.MergePred(I->second);
2143 #ifdef ARC_ANNOTATIONS
2144 if (EnableARCAnnotations) {
2145 // If ARC Annotations are enabled, output the current state of pointers at the
2146 // top of the basic block.
2147 for(BBState::ptr_const_iterator I = MyStates.top_down_ptr_begin(),
2148 E = MyStates.top_down_ptr_end(); I != E; ++I) {
2149 Value *Ptr = const_cast<Value*>(I->first);
2150 Sequence Seq = I->second.GetSeq();
2151 GenerateARCBBEntranceAnnotation("llvm.arc.annotation.topdown.bbstart",
2157 // Visit all the instructions, top-down.
2158 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2159 Instruction *Inst = I;
2161 DEBUG(dbgs() << "ObjCARCOpt::VisitTopDown: Visiting " << *Inst << "\n");
2163 NestingDetected |= VisitInstructionTopDown(Inst, Releases, MyStates);
2166 #ifdef ARC_ANNOTATIONS
2167 if (EnableARCAnnotations) {
2168 // If ARC Annotations are enabled, output the current state of pointers at the
2169 // bottom of the basic block.
2170 for(BBState::ptr_const_iterator I = MyStates.top_down_ptr_begin(),
2171 E = MyStates.top_down_ptr_end(); I != E; ++I) {
2172 Value *Ptr = const_cast<Value*>(I->first);
2173 Sequence Seq = I->second.GetSeq();
2174 GenerateARCBBTerminatorAnnotation("llvm.arc.annotation.topdown.bbend",
2180 CheckForCFGHazards(BB, BBStates, MyStates);
2181 return NestingDetected;
2185 ComputePostOrders(Function &F,
2186 SmallVectorImpl<BasicBlock *> &PostOrder,
2187 SmallVectorImpl<BasicBlock *> &ReverseCFGPostOrder,
2188 unsigned NoObjCARCExceptionsMDKind,
2189 DenseMap<const BasicBlock *, BBState> &BBStates) {
2190 /// The visited set, for doing DFS walks.
2191 SmallPtrSet<BasicBlock *, 16> Visited;
2193 // Do DFS, computing the PostOrder.
2194 SmallPtrSet<BasicBlock *, 16> OnStack;
2195 SmallVector<std::pair<BasicBlock *, succ_iterator>, 16> SuccStack;
2197 // Functions always have exactly one entry block, and we don't have
2198 // any other block that we treat like an entry block.
2199 BasicBlock *EntryBB = &F.getEntryBlock();
2200 BBState &MyStates = BBStates[EntryBB];
2201 MyStates.SetAsEntry();
2202 TerminatorInst *EntryTI = cast<TerminatorInst>(&EntryBB->back());
2203 SuccStack.push_back(std::make_pair(EntryBB, succ_iterator(EntryTI)));
2204 Visited.insert(EntryBB);
2205 OnStack.insert(EntryBB);
2208 BasicBlock *CurrBB = SuccStack.back().first;
2209 TerminatorInst *TI = cast<TerminatorInst>(&CurrBB->back());
2210 succ_iterator SE(TI, false);
2212 while (SuccStack.back().second != SE) {
2213 BasicBlock *SuccBB = *SuccStack.back().second++;
2214 if (Visited.insert(SuccBB)) {
2215 TerminatorInst *TI = cast<TerminatorInst>(&SuccBB->back());
2216 SuccStack.push_back(std::make_pair(SuccBB, succ_iterator(TI)));
2217 BBStates[CurrBB].addSucc(SuccBB);
2218 BBState &SuccStates = BBStates[SuccBB];
2219 SuccStates.addPred(CurrBB);
2220 OnStack.insert(SuccBB);
2224 if (!OnStack.count(SuccBB)) {
2225 BBStates[CurrBB].addSucc(SuccBB);
2226 BBStates[SuccBB].addPred(CurrBB);
2229 OnStack.erase(CurrBB);
2230 PostOrder.push_back(CurrBB);
2231 SuccStack.pop_back();
2232 } while (!SuccStack.empty());
2236 // Do reverse-CFG DFS, computing the reverse-CFG PostOrder.
2237 // Functions may have many exits, and there also blocks which we treat
2238 // as exits due to ignored edges.
2239 SmallVector<std::pair<BasicBlock *, BBState::edge_iterator>, 16> PredStack;
2240 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2241 BasicBlock *ExitBB = I;
2242 BBState &MyStates = BBStates[ExitBB];
2243 if (!MyStates.isExit())
2246 MyStates.SetAsExit();
2248 PredStack.push_back(std::make_pair(ExitBB, MyStates.pred_begin()));
2249 Visited.insert(ExitBB);
2250 while (!PredStack.empty()) {
2251 reverse_dfs_next_succ:
2252 BBState::edge_iterator PE = BBStates[PredStack.back().first].pred_end();
2253 while (PredStack.back().second != PE) {
2254 BasicBlock *BB = *PredStack.back().second++;
2255 if (Visited.insert(BB)) {
2256 PredStack.push_back(std::make_pair(BB, BBStates[BB].pred_begin()));
2257 goto reverse_dfs_next_succ;
2260 ReverseCFGPostOrder.push_back(PredStack.pop_back_val().first);
2265 // Visit the function both top-down and bottom-up.
2267 ObjCARCOpt::Visit(Function &F,
2268 DenseMap<const BasicBlock *, BBState> &BBStates,
2269 MapVector<Value *, RRInfo> &Retains,
2270 DenseMap<Value *, RRInfo> &Releases) {
2272 // Use reverse-postorder traversals, because we magically know that loops
2273 // will be well behaved, i.e. they won't repeatedly call retain on a single
2274 // pointer without doing a release. We can't use the ReversePostOrderTraversal
2275 // class here because we want the reverse-CFG postorder to consider each
2276 // function exit point, and we want to ignore selected cycle edges.
2277 SmallVector<BasicBlock *, 16> PostOrder;
2278 SmallVector<BasicBlock *, 16> ReverseCFGPostOrder;
2279 ComputePostOrders(F, PostOrder, ReverseCFGPostOrder,
2280 NoObjCARCExceptionsMDKind,
2283 // Use reverse-postorder on the reverse CFG for bottom-up.
2284 bool BottomUpNestingDetected = false;
2285 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2286 ReverseCFGPostOrder.rbegin(), E = ReverseCFGPostOrder.rend();
2288 BottomUpNestingDetected |= VisitBottomUp(*I, BBStates, Retains);
2290 // Use reverse-postorder for top-down.
2291 bool TopDownNestingDetected = false;
2292 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2293 PostOrder.rbegin(), E = PostOrder.rend();
2295 TopDownNestingDetected |= VisitTopDown(*I, BBStates, Releases);
2297 return TopDownNestingDetected && BottomUpNestingDetected;
2300 /// Move the calls in RetainsToMove and ReleasesToMove.
2301 void ObjCARCOpt::MoveCalls(Value *Arg,
2302 RRInfo &RetainsToMove,
2303 RRInfo &ReleasesToMove,
2304 MapVector<Value *, RRInfo> &Retains,
2305 DenseMap<Value *, RRInfo> &Releases,
2306 SmallVectorImpl<Instruction *> &DeadInsts,
2308 Type *ArgTy = Arg->getType();
2309 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
2311 // Insert the new retain and release calls.
2312 for (SmallPtrSet<Instruction *, 2>::const_iterator
2313 PI = ReleasesToMove.ReverseInsertPts.begin(),
2314 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2315 Instruction *InsertPt = *PI;
2316 Value *MyArg = ArgTy == ParamTy ? Arg :
2317 new BitCastInst(Arg, ParamTy, "", InsertPt);
2319 CallInst::Create(getRetainCallee(M), MyArg, "", InsertPt);
2320 Call->setDoesNotThrow();
2321 Call->setTailCall();
2323 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Inserting new Release: " << *Call
2325 " At insertion point: " << *InsertPt
2328 for (SmallPtrSet<Instruction *, 2>::const_iterator
2329 PI = RetainsToMove.ReverseInsertPts.begin(),
2330 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2331 Instruction *InsertPt = *PI;
2332 Value *MyArg = ArgTy == ParamTy ? Arg :
2333 new BitCastInst(Arg, ParamTy, "", InsertPt);
2334 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
2336 // Attach a clang.imprecise_release metadata tag, if appropriate.
2337 if (MDNode *M = ReleasesToMove.ReleaseMetadata)
2338 Call->setMetadata(ImpreciseReleaseMDKind, M);
2339 Call->setDoesNotThrow();
2340 if (ReleasesToMove.IsTailCallRelease)
2341 Call->setTailCall();
2343 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Inserting new Retain: " << *Call
2345 " At insertion point: " << *InsertPt
2349 // Delete the original retain and release calls.
2350 for (SmallPtrSet<Instruction *, 2>::const_iterator
2351 AI = RetainsToMove.Calls.begin(),
2352 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
2353 Instruction *OrigRetain = *AI;
2354 Retains.blot(OrigRetain);
2355 DeadInsts.push_back(OrigRetain);
2356 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Deleting retain: " << *OrigRetain <<
2359 for (SmallPtrSet<Instruction *, 2>::const_iterator
2360 AI = ReleasesToMove.Calls.begin(),
2361 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
2362 Instruction *OrigRelease = *AI;
2363 Releases.erase(OrigRelease);
2364 DeadInsts.push_back(OrigRelease);
2365 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Deleting release: " << *OrigRelease
2371 ObjCARCOpt::ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState>
2373 MapVector<Value *, RRInfo> &Retains,
2374 DenseMap<Value *, RRInfo> &Releases,
2376 SmallVector<Instruction *, 4> &NewRetains,
2377 SmallVector<Instruction *, 4> &NewReleases,
2378 SmallVector<Instruction *, 8> &DeadInsts,
2379 RRInfo &RetainsToMove,
2380 RRInfo &ReleasesToMove,
2383 bool &AnyPairsCompletelyEliminated) {
2384 // If a pair happens in a region where it is known that the reference count
2385 // is already incremented, we can similarly ignore possible decrements.
2386 bool KnownSafeTD = true, KnownSafeBU = true;
2388 // Connect the dots between the top-down-collected RetainsToMove and
2389 // bottom-up-collected ReleasesToMove to form sets of related calls.
2390 // This is an iterative process so that we connect multiple releases
2391 // to multiple retains if needed.
2392 unsigned OldDelta = 0;
2393 unsigned NewDelta = 0;
2394 unsigned OldCount = 0;
2395 unsigned NewCount = 0;
2396 bool FirstRelease = true;
2398 for (SmallVectorImpl<Instruction *>::const_iterator
2399 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
2400 Instruction *NewRetain = *NI;
2401 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
2402 assert(It != Retains.end());
2403 const RRInfo &NewRetainRRI = It->second;
2404 KnownSafeTD &= NewRetainRRI.KnownSafe;
2405 for (SmallPtrSet<Instruction *, 2>::const_iterator
2406 LI = NewRetainRRI.Calls.begin(),
2407 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
2408 Instruction *NewRetainRelease = *LI;
2409 DenseMap<Value *, RRInfo>::const_iterator Jt =
2410 Releases.find(NewRetainRelease);
2411 if (Jt == Releases.end())
2413 const RRInfo &NewRetainReleaseRRI = Jt->second;
2414 assert(NewRetainReleaseRRI.Calls.count(NewRetain));
2415 if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
2417 BBStates[NewRetainRelease->getParent()].GetAllPathCount();
2419 // Merge the ReleaseMetadata and IsTailCallRelease values.
2421 ReleasesToMove.ReleaseMetadata =
2422 NewRetainReleaseRRI.ReleaseMetadata;
2423 ReleasesToMove.IsTailCallRelease =
2424 NewRetainReleaseRRI.IsTailCallRelease;
2425 FirstRelease = false;
2427 if (ReleasesToMove.ReleaseMetadata !=
2428 NewRetainReleaseRRI.ReleaseMetadata)
2429 ReleasesToMove.ReleaseMetadata = 0;
2430 if (ReleasesToMove.IsTailCallRelease !=
2431 NewRetainReleaseRRI.IsTailCallRelease)
2432 ReleasesToMove.IsTailCallRelease = false;
2435 // Collect the optimal insertion points.
2437 for (SmallPtrSet<Instruction *, 2>::const_iterator
2438 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
2439 RE = NewRetainReleaseRRI.ReverseInsertPts.end();
2441 Instruction *RIP = *RI;
2442 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
2443 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
2445 NewReleases.push_back(NewRetainRelease);
2450 if (NewReleases.empty()) break;
2452 // Back the other way.
2453 for (SmallVectorImpl<Instruction *>::const_iterator
2454 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
2455 Instruction *NewRelease = *NI;
2456 DenseMap<Value *, RRInfo>::const_iterator It =
2457 Releases.find(NewRelease);
2458 assert(It != Releases.end());
2459 const RRInfo &NewReleaseRRI = It->second;
2460 KnownSafeBU &= NewReleaseRRI.KnownSafe;
2461 for (SmallPtrSet<Instruction *, 2>::const_iterator
2462 LI = NewReleaseRRI.Calls.begin(),
2463 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
2464 Instruction *NewReleaseRetain = *LI;
2465 MapVector<Value *, RRInfo>::const_iterator Jt =
2466 Retains.find(NewReleaseRetain);
2467 if (Jt == Retains.end())
2469 const RRInfo &NewReleaseRetainRRI = Jt->second;
2470 assert(NewReleaseRetainRRI.Calls.count(NewRelease));
2471 if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
2472 unsigned PathCount =
2473 BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
2474 OldDelta += PathCount;
2475 OldCount += PathCount;
2477 // Collect the optimal insertion points.
2479 for (SmallPtrSet<Instruction *, 2>::const_iterator
2480 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
2481 RE = NewReleaseRetainRRI.ReverseInsertPts.end();
2483 Instruction *RIP = *RI;
2484 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
2485 PathCount = BBStates[RIP->getParent()].GetAllPathCount();
2486 NewDelta += PathCount;
2487 NewCount += PathCount;
2490 NewRetains.push_back(NewReleaseRetain);
2494 NewReleases.clear();
2495 if (NewRetains.empty()) break;
2498 // If the pointer is known incremented or nested, we can safely delete the
2499 // pair regardless of what's between them.
2500 if (KnownSafeTD || KnownSafeBU) {
2501 RetainsToMove.ReverseInsertPts.clear();
2502 ReleasesToMove.ReverseInsertPts.clear();
2505 // Determine whether the new insertion points we computed preserve the
2506 // balance of retain and release calls through the program.
2507 // TODO: If the fully aggressive solution isn't valid, try to find a
2508 // less aggressive solution which is.
2513 // Determine whether the original call points are balanced in the retain and
2514 // release calls through the program. If not, conservatively don't touch
2516 // TODO: It's theoretically possible to do code motion in this case, as
2517 // long as the existing imbalances are maintained.
2522 assert(OldCount != 0 && "Unreachable code?");
2523 NumRRs += OldCount - NewCount;
2524 // Set to true if we completely removed any RR pairs.
2525 AnyPairsCompletelyEliminated = NewCount == 0;
2527 // We can move calls!
2531 /// Identify pairings between the retains and releases, and delete and/or move
2534 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
2536 MapVector<Value *, RRInfo> &Retains,
2537 DenseMap<Value *, RRInfo> &Releases,
2539 bool AnyPairsCompletelyEliminated = false;
2540 RRInfo RetainsToMove;
2541 RRInfo ReleasesToMove;
2542 SmallVector<Instruction *, 4> NewRetains;
2543 SmallVector<Instruction *, 4> NewReleases;
2544 SmallVector<Instruction *, 8> DeadInsts;
2546 // Visit each retain.
2547 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
2548 E = Retains.end(); I != E; ++I) {
2549 Value *V = I->first;
2550 if (!V) continue; // blotted
2552 Instruction *Retain = cast<Instruction>(V);
2554 DEBUG(dbgs() << "ObjCARCOpt::PerformCodePlacement: Visiting: " << *Retain
2557 Value *Arg = GetObjCArg(Retain);
2559 // If the object being released is in static or stack storage, we know it's
2560 // not being managed by ObjC reference counting, so we can delete pairs
2561 // regardless of what possible decrements or uses lie between them.
2562 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg);
2564 // A constant pointer can't be pointing to an object on the heap. It may
2565 // be reference-counted, but it won't be deleted.
2566 if (const LoadInst *LI = dyn_cast<LoadInst>(Arg))
2567 if (const GlobalVariable *GV =
2568 dyn_cast<GlobalVariable>(
2569 StripPointerCastsAndObjCCalls(LI->getPointerOperand())))
2570 if (GV->isConstant())
2573 // Connect the dots between the top-down-collected RetainsToMove and
2574 // bottom-up-collected ReleasesToMove to form sets of related calls.
2575 NewRetains.push_back(Retain);
2576 bool PerformMoveCalls =
2577 ConnectTDBUTraversals(BBStates, Retains, Releases, M, NewRetains,
2578 NewReleases, DeadInsts, RetainsToMove,
2579 ReleasesToMove, Arg, KnownSafe,
2580 AnyPairsCompletelyEliminated);
2582 #ifdef ARC_ANNOTATIONS
2583 // Do not move calls if ARC annotations are requested. If we were to move
2584 // calls in this case, we would not be able
2585 PerformMoveCalls = PerformMoveCalls && !EnableARCAnnotations;
2586 #endif // ARC_ANNOTATIONS
2588 if (PerformMoveCalls) {
2589 // Ok, everything checks out and we're all set. Let's move/delete some
2591 MoveCalls(Arg, RetainsToMove, ReleasesToMove,
2592 Retains, Releases, DeadInsts, M);
2595 // Clean up state for next retain.
2596 NewReleases.clear();
2598 RetainsToMove.clear();
2599 ReleasesToMove.clear();
2602 // Now that we're done moving everything, we can delete the newly dead
2603 // instructions, as we no longer need them as insert points.
2604 while (!DeadInsts.empty())
2605 EraseInstruction(DeadInsts.pop_back_val());
2607 return AnyPairsCompletelyEliminated;
2610 /// Weak pointer optimizations.
2611 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
2612 // First, do memdep-style RLE and S2L optimizations. We can't use memdep
2613 // itself because it uses AliasAnalysis and we need to do provenance
2615 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2616 Instruction *Inst = &*I++;
2618 DEBUG(dbgs() << "ObjCARCOpt::OptimizeWeakCalls: Visiting: " << *Inst <<
2621 InstructionClass Class = GetBasicInstructionClass(Inst);
2622 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
2625 // Delete objc_loadWeak calls with no users.
2626 if (Class == IC_LoadWeak && Inst->use_empty()) {
2627 Inst->eraseFromParent();
2631 // TODO: For now, just look for an earlier available version of this value
2632 // within the same block. Theoretically, we could do memdep-style non-local
2633 // analysis too, but that would want caching. A better approach would be to
2634 // use the technique that EarlyCSE uses.
2635 inst_iterator Current = llvm::prior(I);
2636 BasicBlock *CurrentBB = Current.getBasicBlockIterator();
2637 for (BasicBlock::iterator B = CurrentBB->begin(),
2638 J = Current.getInstructionIterator();
2640 Instruction *EarlierInst = &*llvm::prior(J);
2641 InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
2642 switch (EarlierClass) {
2644 case IC_LoadWeakRetained: {
2645 // If this is loading from the same pointer, replace this load's value
2647 CallInst *Call = cast<CallInst>(Inst);
2648 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2649 Value *Arg = Call->getArgOperand(0);
2650 Value *EarlierArg = EarlierCall->getArgOperand(0);
2651 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2652 case AliasAnalysis::MustAlias:
2654 // If the load has a builtin retain, insert a plain retain for it.
2655 if (Class == IC_LoadWeakRetained) {
2657 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2661 // Zap the fully redundant load.
2662 Call->replaceAllUsesWith(EarlierCall);
2663 Call->eraseFromParent();
2665 case AliasAnalysis::MayAlias:
2666 case AliasAnalysis::PartialAlias:
2668 case AliasAnalysis::NoAlias:
2675 // If this is storing to the same pointer and has the same size etc.
2676 // replace this load's value with the stored value.
2677 CallInst *Call = cast<CallInst>(Inst);
2678 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2679 Value *Arg = Call->getArgOperand(0);
2680 Value *EarlierArg = EarlierCall->getArgOperand(0);
2681 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2682 case AliasAnalysis::MustAlias:
2684 // If the load has a builtin retain, insert a plain retain for it.
2685 if (Class == IC_LoadWeakRetained) {
2687 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2691 // Zap the fully redundant load.
2692 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
2693 Call->eraseFromParent();
2695 case AliasAnalysis::MayAlias:
2696 case AliasAnalysis::PartialAlias:
2698 case AliasAnalysis::NoAlias:
2705 // TOOD: Grab the copied value.
2707 case IC_AutoreleasepoolPush:
2709 case IC_IntrinsicUser:
2711 // Weak pointers are only modified through the weak entry points
2712 // (and arbitrary calls, which could call the weak entry points).
2715 // Anything else could modify the weak pointer.
2722 // Then, for each destroyWeak with an alloca operand, check to see if
2723 // the alloca and all its users can be zapped.
2724 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2725 Instruction *Inst = &*I++;
2726 InstructionClass Class = GetBasicInstructionClass(Inst);
2727 if (Class != IC_DestroyWeak)
2730 CallInst *Call = cast<CallInst>(Inst);
2731 Value *Arg = Call->getArgOperand(0);
2732 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
2733 for (Value::use_iterator UI = Alloca->use_begin(),
2734 UE = Alloca->use_end(); UI != UE; ++UI) {
2735 const Instruction *UserInst = cast<Instruction>(*UI);
2736 switch (GetBasicInstructionClass(UserInst)) {
2739 case IC_DestroyWeak:
2746 for (Value::use_iterator UI = Alloca->use_begin(),
2747 UE = Alloca->use_end(); UI != UE; ) {
2748 CallInst *UserInst = cast<CallInst>(*UI++);
2749 switch (GetBasicInstructionClass(UserInst)) {
2752 // These functions return their second argument.
2753 UserInst->replaceAllUsesWith(UserInst->getArgOperand(1));
2755 case IC_DestroyWeak:
2759 llvm_unreachable("alloca really is used!");
2761 UserInst->eraseFromParent();
2763 Alloca->eraseFromParent();
2768 DEBUG(dbgs() << "ObjCARCOpt::OptimizeWeakCalls: Finished List.\n\n");
2772 /// Identify program paths which execute sequences of retains and releases which
2773 /// can be eliminated.
2774 bool ObjCARCOpt::OptimizeSequences(Function &F) {
2775 /// Releases, Retains - These are used to store the results of the main flow
2776 /// analysis. These use Value* as the key instead of Instruction* so that the
2777 /// map stays valid when we get around to rewriting code and calls get
2778 /// replaced by arguments.
2779 DenseMap<Value *, RRInfo> Releases;
2780 MapVector<Value *, RRInfo> Retains;
2782 /// This is used during the traversal of the function to track the
2783 /// states for each identified object at each block.
2784 DenseMap<const BasicBlock *, BBState> BBStates;
2786 // Analyze the CFG of the function, and all instructions.
2787 bool NestingDetected = Visit(F, BBStates, Retains, Releases);
2790 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
2794 /// Look for this pattern:
2796 /// %call = call i8* @something(...)
2797 /// %2 = call i8* @objc_retain(i8* %call)
2798 /// %3 = call i8* @objc_autorelease(i8* %2)
2801 /// And delete the retain and autorelease.
2803 /// Otherwise if it's just this:
2805 /// %3 = call i8* @objc_autorelease(i8* %2)
2808 /// convert the autorelease to autoreleaseRV.
2809 void ObjCARCOpt::OptimizeReturns(Function &F) {
2810 if (!F.getReturnType()->isPointerTy())
2813 SmallPtrSet<Instruction *, 4> DependingInstructions;
2814 SmallPtrSet<const BasicBlock *, 4> Visited;
2815 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
2816 BasicBlock *BB = FI;
2817 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
2819 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Visiting: " << *Ret << "\n");
2823 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
2824 FindDependencies(NeedsPositiveRetainCount, Arg,
2825 BB, Ret, DependingInstructions, Visited, PA);
2826 if (DependingInstructions.size() != 1)
2830 CallInst *Autorelease =
2831 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
2834 InstructionClass AutoreleaseClass = GetBasicInstructionClass(Autorelease);
2835 if (!IsAutorelease(AutoreleaseClass))
2837 if (GetObjCArg(Autorelease) != Arg)
2840 DependingInstructions.clear();
2843 // Check that there is nothing that can affect the reference
2844 // count between the autorelease and the retain.
2845 FindDependencies(CanChangeRetainCount, Arg,
2846 BB, Autorelease, DependingInstructions, Visited, PA);
2847 if (DependingInstructions.size() != 1)
2852 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
2854 // Check that we found a retain with the same argument.
2856 !IsRetain(GetBasicInstructionClass(Retain)) ||
2857 GetObjCArg(Retain) != Arg)
2860 DependingInstructions.clear();
2863 // Convert the autorelease to an autoreleaseRV, since it's
2864 // returning the value.
2865 if (AutoreleaseClass == IC_Autorelease) {
2866 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Converting autorelease "
2867 "=> autoreleaseRV since it's returning a value.\n"
2868 " In: " << *Autorelease
2870 Autorelease->setCalledFunction(getAutoreleaseRVCallee(F.getParent()));
2871 DEBUG(dbgs() << " Out: " << *Autorelease
2873 Autorelease->setTailCall(); // Always tail call autoreleaseRV.
2874 AutoreleaseClass = IC_AutoreleaseRV;
2877 // Check that there is nothing that can affect the reference
2878 // count between the retain and the call.
2879 // Note that Retain need not be in BB.
2880 FindDependencies(CanChangeRetainCount, Arg, Retain->getParent(), Retain,
2881 DependingInstructions, Visited, PA);
2882 if (DependingInstructions.size() != 1)
2887 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
2889 // Check that the pointer is the return value of the call.
2890 if (!Call || Arg != Call)
2893 // Check that the call is a regular call.
2894 InstructionClass Class = GetBasicInstructionClass(Call);
2895 if (Class != IC_CallOrUser && Class != IC_Call)
2898 // If so, we can zap the retain and autorelease.
2901 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Erasing: " << *Retain
2903 << *Autorelease << "\n");
2904 EraseInstruction(Retain);
2905 EraseInstruction(Autorelease);
2911 DependingInstructions.clear();
2915 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Finished List.\n\n");
2919 bool ObjCARCOpt::doInitialization(Module &M) {
2923 // If nothing in the Module uses ARC, don't do anything.
2924 Run = ModuleHasARC(M);
2928 // Identify the imprecise release metadata kind.
2929 ImpreciseReleaseMDKind =
2930 M.getContext().getMDKindID("clang.imprecise_release");
2931 CopyOnEscapeMDKind =
2932 M.getContext().getMDKindID("clang.arc.copy_on_escape");
2933 NoObjCARCExceptionsMDKind =
2934 M.getContext().getMDKindID("clang.arc.no_objc_arc_exceptions");
2935 #ifdef ARC_ANNOTATIONS
2936 ARCAnnotationBottomUpMDKind =
2937 M.getContext().getMDKindID("llvm.arc.annotation.bottomup");
2938 ARCAnnotationTopDownMDKind =
2939 M.getContext().getMDKindID("llvm.arc.annotation.topdown");
2940 ARCAnnotationProvenanceSourceMDKind =
2941 M.getContext().getMDKindID("llvm.arc.annotation.provenancesource");
2942 #endif // ARC_ANNOTATIONS
2944 // Intuitively, objc_retain and others are nocapture, however in practice
2945 // they are not, because they return their argument value. And objc_release
2946 // calls finalizers which can have arbitrary side effects.
2948 // These are initialized lazily.
2950 AutoreleaseRVCallee = 0;
2953 RetainBlockCallee = 0;
2954 AutoreleaseCallee = 0;
2959 bool ObjCARCOpt::runOnFunction(Function &F) {
2963 // If nothing in the Module uses ARC, don't do anything.
2969 DEBUG(dbgs() << "ObjCARCOpt: Visiting Function: " << F.getName() << "\n");
2971 PA.setAA(&getAnalysis<AliasAnalysis>());
2973 // This pass performs several distinct transformations. As a compile-time aid
2974 // when compiling code that isn't ObjC, skip these if the relevant ObjC
2975 // library functions aren't declared.
2977 // Preliminary optimizations. This also computs UsedInThisFunction.
2978 OptimizeIndividualCalls(F);
2980 // Optimizations for weak pointers.
2981 if (UsedInThisFunction & ((1 << IC_LoadWeak) |
2982 (1 << IC_LoadWeakRetained) |
2983 (1 << IC_StoreWeak) |
2984 (1 << IC_InitWeak) |
2985 (1 << IC_CopyWeak) |
2986 (1 << IC_MoveWeak) |
2987 (1 << IC_DestroyWeak)))
2988 OptimizeWeakCalls(F);
2990 // Optimizations for retain+release pairs.
2991 if (UsedInThisFunction & ((1 << IC_Retain) |
2992 (1 << IC_RetainRV) |
2993 (1 << IC_RetainBlock)))
2994 if (UsedInThisFunction & (1 << IC_Release))
2995 // Run OptimizeSequences until it either stops making changes or
2996 // no retain+release pair nesting is detected.
2997 while (OptimizeSequences(F)) {}
2999 // Optimizations if objc_autorelease is used.
3000 if (UsedInThisFunction & ((1 << IC_Autorelease) |
3001 (1 << IC_AutoreleaseRV)))
3004 DEBUG(dbgs() << "\n");
3009 void ObjCARCOpt::releaseMemory() {