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 if the Calls are objc_retainBlock calls (as opposed to objc_retain
398 /// True of the objc_release calls are all marked with the "tail" keyword.
399 bool IsTailCallRelease;
401 /// If the Calls are objc_release calls and they all have a
402 /// clang.imprecise_release tag, this is the metadata tag.
403 MDNode *ReleaseMetadata;
405 /// For a top-down sequence, the set of objc_retains or
406 /// objc_retainBlocks. For bottom-up, the set of objc_releases.
407 SmallPtrSet<Instruction *, 2> Calls;
409 /// The set of optimal insert positions for moving calls in the opposite
411 SmallPtrSet<Instruction *, 2> ReverseInsertPts;
414 KnownSafe(false), IsRetainBlock(false),
415 IsTailCallRelease(false),
416 ReleaseMetadata(0) {}
422 void RRInfo::clear() {
424 IsRetainBlock = false;
425 IsTailCallRelease = false;
428 ReverseInsertPts.clear();
432 /// \brief This class summarizes several per-pointer runtime properties which
433 /// are propogated through the flow graph.
435 /// True if the reference count is known to be incremented.
436 bool KnownPositiveRefCount;
438 /// True of we've seen an opportunity for partial RR elimination, such as
439 /// pushing calls into a CFG triangle or into one side of a CFG diamond.
442 /// The current position in the sequence.
446 /// Unidirectional information about the current sequence.
448 /// TODO: Encapsulate this better.
451 PtrState() : KnownPositiveRefCount(false), Partial(false),
454 void SetKnownPositiveRefCount() {
455 KnownPositiveRefCount = true;
458 void ClearKnownPositiveRefCount() {
459 KnownPositiveRefCount = false;
462 bool HasKnownPositiveRefCount() const {
463 return KnownPositiveRefCount;
466 void SetSeq(Sequence NewSeq) {
470 Sequence GetSeq() const {
474 void ClearSequenceProgress() {
475 ResetSequenceProgress(S_None);
478 void ResetSequenceProgress(Sequence NewSeq) {
484 void Merge(const PtrState &Other, bool TopDown);
489 PtrState::Merge(const PtrState &Other, bool TopDown) {
490 Seq = MergeSeqs(Seq, Other.Seq, TopDown);
491 KnownPositiveRefCount = KnownPositiveRefCount && Other.KnownPositiveRefCount;
493 // We can't merge a plain objc_retain with an objc_retainBlock.
494 if (RRI.IsRetainBlock != Other.RRI.IsRetainBlock)
497 // If we're not in a sequence (anymore), drop all associated state.
501 } else if (Partial || Other.Partial) {
502 // If we're doing a merge on a path that's previously seen a partial
503 // merge, conservatively drop the sequence, to avoid doing partial
504 // RR elimination. If the branch predicates for the two merge differ,
505 // mixing them is unsafe.
506 ClearSequenceProgress();
508 // Conservatively merge the ReleaseMetadata information.
509 if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
510 RRI.ReleaseMetadata = 0;
512 RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe;
513 RRI.IsTailCallRelease = RRI.IsTailCallRelease &&
514 Other.RRI.IsTailCallRelease;
515 RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
517 // Merge the insert point sets. If there are any differences,
518 // that makes this a partial merge.
519 Partial = RRI.ReverseInsertPts.size() != Other.RRI.ReverseInsertPts.size();
520 for (SmallPtrSet<Instruction *, 2>::const_iterator
521 I = Other.RRI.ReverseInsertPts.begin(),
522 E = Other.RRI.ReverseInsertPts.end(); I != E; ++I)
523 Partial |= RRI.ReverseInsertPts.insert(*I);
528 /// \brief Per-BasicBlock state.
530 /// The number of unique control paths from the entry which can reach this
532 unsigned TopDownPathCount;
534 /// The number of unique control paths to exits from this block.
535 unsigned BottomUpPathCount;
537 /// A type for PerPtrTopDown and PerPtrBottomUp.
538 typedef MapVector<const Value *, PtrState> MapTy;
540 /// The top-down traversal uses this to record information known about a
541 /// pointer at the bottom of each block.
544 /// The bottom-up traversal uses this to record information known about a
545 /// pointer at the top of each block.
546 MapTy PerPtrBottomUp;
548 /// Effective predecessors of the current block ignoring ignorable edges and
549 /// ignored backedges.
550 SmallVector<BasicBlock *, 2> Preds;
551 /// Effective successors of the current block ignoring ignorable edges and
552 /// ignored backedges.
553 SmallVector<BasicBlock *, 2> Succs;
556 BBState() : TopDownPathCount(0), BottomUpPathCount(0) {}
558 typedef MapTy::iterator ptr_iterator;
559 typedef MapTy::const_iterator ptr_const_iterator;
561 ptr_iterator top_down_ptr_begin() { return PerPtrTopDown.begin(); }
562 ptr_iterator top_down_ptr_end() { return PerPtrTopDown.end(); }
563 ptr_const_iterator top_down_ptr_begin() const {
564 return PerPtrTopDown.begin();
566 ptr_const_iterator top_down_ptr_end() const {
567 return PerPtrTopDown.end();
570 ptr_iterator bottom_up_ptr_begin() { return PerPtrBottomUp.begin(); }
571 ptr_iterator bottom_up_ptr_end() { return PerPtrBottomUp.end(); }
572 ptr_const_iterator bottom_up_ptr_begin() const {
573 return PerPtrBottomUp.begin();
575 ptr_const_iterator bottom_up_ptr_end() const {
576 return PerPtrBottomUp.end();
579 /// Mark this block as being an entry block, which has one path from the
580 /// entry by definition.
581 void SetAsEntry() { TopDownPathCount = 1; }
583 /// Mark this block as being an exit block, which has one path to an exit by
585 void SetAsExit() { BottomUpPathCount = 1; }
587 PtrState &getPtrTopDownState(const Value *Arg) {
588 return PerPtrTopDown[Arg];
591 PtrState &getPtrBottomUpState(const Value *Arg) {
592 return PerPtrBottomUp[Arg];
595 void clearBottomUpPointers() {
596 PerPtrBottomUp.clear();
599 void clearTopDownPointers() {
600 PerPtrTopDown.clear();
603 void InitFromPred(const BBState &Other);
604 void InitFromSucc(const BBState &Other);
605 void MergePred(const BBState &Other);
606 void MergeSucc(const BBState &Other);
608 /// Return the number of possible unique paths from an entry to an exit
609 /// which pass through this block. This is only valid after both the
610 /// top-down and bottom-up traversals are complete.
611 unsigned GetAllPathCount() const {
612 assert(TopDownPathCount != 0);
613 assert(BottomUpPathCount != 0);
614 return TopDownPathCount * BottomUpPathCount;
617 // Specialized CFG utilities.
618 typedef SmallVectorImpl<BasicBlock *>::const_iterator edge_iterator;
619 edge_iterator pred_begin() { return Preds.begin(); }
620 edge_iterator pred_end() { return Preds.end(); }
621 edge_iterator succ_begin() { return Succs.begin(); }
622 edge_iterator succ_end() { return Succs.end(); }
624 void addSucc(BasicBlock *Succ) { Succs.push_back(Succ); }
625 void addPred(BasicBlock *Pred) { Preds.push_back(Pred); }
627 bool isExit() const { return Succs.empty(); }
631 void BBState::InitFromPred(const BBState &Other) {
632 PerPtrTopDown = Other.PerPtrTopDown;
633 TopDownPathCount = Other.TopDownPathCount;
636 void BBState::InitFromSucc(const BBState &Other) {
637 PerPtrBottomUp = Other.PerPtrBottomUp;
638 BottomUpPathCount = Other.BottomUpPathCount;
641 /// The top-down traversal uses this to merge information about predecessors to
642 /// form the initial state for a new block.
643 void BBState::MergePred(const BBState &Other) {
644 // Other.TopDownPathCount can be 0, in which case it is either dead or a
645 // loop backedge. Loop backedges are special.
646 TopDownPathCount += Other.TopDownPathCount;
648 // Check for overflow. If we have overflow, fall back to conservative
650 if (TopDownPathCount < Other.TopDownPathCount) {
651 clearTopDownPointers();
655 // For each entry in the other set, if our set has an entry with the same key,
656 // merge the entries. Otherwise, copy the entry and merge it with an empty
658 for (ptr_const_iterator MI = Other.top_down_ptr_begin(),
659 ME = Other.top_down_ptr_end(); MI != ME; ++MI) {
660 std::pair<ptr_iterator, bool> Pair = PerPtrTopDown.insert(*MI);
661 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
665 // For each entry in our set, if the other set doesn't have an entry with the
666 // same key, force it to merge with an empty entry.
667 for (ptr_iterator MI = top_down_ptr_begin(),
668 ME = top_down_ptr_end(); MI != ME; ++MI)
669 if (Other.PerPtrTopDown.find(MI->first) == Other.PerPtrTopDown.end())
670 MI->second.Merge(PtrState(), /*TopDown=*/true);
673 /// The bottom-up traversal uses this to merge information about successors to
674 /// form the initial state for a new block.
675 void BBState::MergeSucc(const BBState &Other) {
676 // Other.BottomUpPathCount can be 0, in which case it is either dead or a
677 // loop backedge. Loop backedges are special.
678 BottomUpPathCount += Other.BottomUpPathCount;
680 // Check for overflow. If we have overflow, fall back to conservative
682 if (BottomUpPathCount < Other.BottomUpPathCount) {
683 clearBottomUpPointers();
687 // For each entry in the other set, if our set has an entry with the
688 // same key, merge the entries. Otherwise, copy the entry and merge
689 // it with an empty entry.
690 for (ptr_const_iterator MI = Other.bottom_up_ptr_begin(),
691 ME = Other.bottom_up_ptr_end(); MI != ME; ++MI) {
692 std::pair<ptr_iterator, bool> Pair = PerPtrBottomUp.insert(*MI);
693 Pair.first->second.Merge(Pair.second ? PtrState() : MI->second,
697 // For each entry in our set, if the other set doesn't have an entry
698 // with the same key, force it to merge with an empty entry.
699 for (ptr_iterator MI = bottom_up_ptr_begin(),
700 ME = bottom_up_ptr_end(); MI != ME; ++MI)
701 if (Other.PerPtrBottomUp.find(MI->first) == Other.PerPtrBottomUp.end())
702 MI->second.Merge(PtrState(), /*TopDown=*/false);
705 // Only enable ARC Annotations if we are building a debug version of
708 #define ARC_ANNOTATIONS
711 // Define some macros along the lines of DEBUG and some helper functions to make
712 // it cleaner to create annotations in the source code and to no-op when not
713 // building in debug mode.
714 #ifdef ARC_ANNOTATIONS
716 #include "llvm/Support/CommandLine.h"
718 /// Enable/disable ARC sequence annotations.
720 EnableARCAnnotations("enable-objc-arc-annotations", cl::init(false));
722 /// This function appends a unique ARCAnnotationProvenanceSourceMDKind id to an
723 /// instruction so that we can track backwards when post processing via the llvm
724 /// arc annotation processor tool. If the function is an
725 static MDString *AppendMDNodeToSourcePtr(unsigned NodeId,
729 // If pointer is a result of an instruction and it does not have a source
730 // MDNode it, attach a new MDNode onto it. If pointer is a result of
731 // an instruction and does have a source MDNode attached to it, return a
732 // reference to said Node. Otherwise just return 0.
733 if (Instruction *Inst = dyn_cast<Instruction>(Ptr)) {
735 if (!(Node = Inst->getMetadata(NodeId))) {
736 // We do not have any node. Generate and attatch the hash MDString to the
739 // We just use an MDString to ensure that this metadata gets written out
740 // of line at the module level and to provide a very simple format
741 // encoding the information herein. Both of these makes it simpler to
742 // parse the annotations by a simple external program.
744 raw_string_ostream os(Str);
745 os << "(" << Inst->getParent()->getParent()->getName() << ",%"
746 << Inst->getName() << ")";
748 Hash = MDString::get(Inst->getContext(), os.str());
749 Inst->setMetadata(NodeId, MDNode::get(Inst->getContext(),Hash));
751 // We have a node. Grab its hash and return it.
752 assert(Node->getNumOperands() == 1 &&
753 "An ARCAnnotationProvenanceSourceMDKind can only have 1 operand.");
754 Hash = cast<MDString>(Node->getOperand(0));
756 } else if (Argument *Arg = dyn_cast<Argument>(Ptr)) {
758 raw_string_ostream os(str);
759 os << "(" << Arg->getParent()->getName() << ",%" << Arg->getName()
761 Hash = MDString::get(Arg->getContext(), os.str());
767 static std::string SequenceToString(Sequence A) {
769 raw_string_ostream os(str);
774 /// Helper function to change a Sequence into a String object using our overload
775 /// for raw_ostream so we only have printing code in one location.
776 static MDString *SequenceToMDString(LLVMContext &Context,
778 return MDString::get(Context, SequenceToString(A));
781 /// A simple function to generate a MDNode which describes the change in state
782 /// for Value *Ptr caused by Instruction *Inst.
783 static void AppendMDNodeToInstForPtr(unsigned NodeId,
786 MDString *PtrSourceMDNodeID,
790 Value *tmp[3] = {PtrSourceMDNodeID,
791 SequenceToMDString(Inst->getContext(),
793 SequenceToMDString(Inst->getContext(),
795 Node = MDNode::get(Inst->getContext(),
796 ArrayRef<Value*>(tmp, 3));
798 Inst->setMetadata(NodeId, Node);
801 /// Add to the beginning of the basic block llvm.ptr.annotations which show the
802 /// state of a pointer at the entrance to a basic block.
803 static void GenerateARCBBEntranceAnnotation(const char *Name, BasicBlock *BB,
804 Value *Ptr, Sequence Seq) {
805 Module *M = BB->getParent()->getParent();
806 LLVMContext &C = M->getContext();
807 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
808 Type *I8XX = PointerType::getUnqual(I8X);
809 Type *Params[] = {I8XX, I8XX};
810 FunctionType *FTy = FunctionType::get(Type::getVoidTy(C),
811 ArrayRef<Type*>(Params, 2),
813 Constant *Callee = M->getOrInsertFunction(Name, FTy);
815 IRBuilder<> Builder(BB, BB->getFirstInsertionPt());
818 StringRef Tmp = Ptr->getName();
819 if (0 == (PtrName = M->getGlobalVariable(Tmp, true))) {
820 Value *ActualPtrName = Builder.CreateGlobalStringPtr(Tmp,
822 PtrName = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
823 cast<Constant>(ActualPtrName), Tmp);
827 std::string SeqStr = SequenceToString(Seq);
828 if (0 == (S = M->getGlobalVariable(SeqStr, true))) {
829 Value *ActualPtrName = Builder.CreateGlobalStringPtr(SeqStr,
831 S = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
832 cast<Constant>(ActualPtrName), SeqStr);
835 Builder.CreateCall2(Callee, PtrName, S);
838 /// Add to the end of the basic block llvm.ptr.annotations which show the state
839 /// of the pointer at the bottom of the basic block.
840 static void GenerateARCBBTerminatorAnnotation(const char *Name, BasicBlock *BB,
841 Value *Ptr, Sequence Seq) {
842 Module *M = BB->getParent()->getParent();
843 LLVMContext &C = M->getContext();
844 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
845 Type *I8XX = PointerType::getUnqual(I8X);
846 Type *Params[] = {I8XX, I8XX};
847 FunctionType *FTy = FunctionType::get(Type::getVoidTy(C),
848 ArrayRef<Type*>(Params, 2),
850 Constant *Callee = M->getOrInsertFunction(Name, FTy);
852 IRBuilder<> Builder(BB, llvm::prior(BB->end()));
855 StringRef Tmp = Ptr->getName();
856 if (0 == (PtrName = M->getGlobalVariable(Tmp, true))) {
857 Value *ActualPtrName = Builder.CreateGlobalStringPtr(Tmp,
859 PtrName = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
860 cast<Constant>(ActualPtrName), Tmp);
864 std::string SeqStr = SequenceToString(Seq);
865 if (0 == (S = M->getGlobalVariable(SeqStr, true))) {
866 Value *ActualPtrName = Builder.CreateGlobalStringPtr(SeqStr,
868 S = new GlobalVariable(*M, I8X, true, GlobalVariable::InternalLinkage,
869 cast<Constant>(ActualPtrName), SeqStr);
871 Builder.CreateCall2(Callee, PtrName, S);
874 /// Adds a source annotation to pointer and a state change annotation to Inst
875 /// referencing the source annotation and the old/new state of pointer.
876 static void GenerateARCAnnotation(unsigned InstMDId,
882 if (EnableARCAnnotations) {
883 // First generate the source annotation on our pointer. This will return an
884 // MDString* if Ptr actually comes from an instruction implying we can put
885 // in a source annotation. If AppendMDNodeToSourcePtr returns 0 (i.e. NULL),
886 // then we know that our pointer is from an Argument so we put a reference
887 // to the argument number.
889 // The point of this is to make it easy for the
890 // llvm-arc-annotation-processor tool to cross reference where the source
891 // pointer is in the LLVM IR since the LLVM IR parser does not submit such
892 // information via debug info for backends to use (since why would anyone
893 // need such a thing from LLVM IR besides in non standard cases
895 MDString *SourcePtrMDNode =
896 AppendMDNodeToSourcePtr(PtrMDId, Ptr);
897 AppendMDNodeToInstForPtr(InstMDId, Inst, Ptr, SourcePtrMDNode, OldSeq,
902 // The actual interface for accessing the above functionality is defined via
903 // some simple macros which are defined below. We do this so that the user does
904 // not need to pass in what metadata id is needed resulting in cleaner code and
905 // additionally since it provides an easy way to conditionally no-op all
906 // annotation support in a non-debug build.
908 /// Use this macro to annotate a sequence state change when processing
909 /// instructions bottom up,
910 #define ANNOTATE_BOTTOMUP(inst, ptr, old, new) \
911 GenerateARCAnnotation(ARCAnnotationBottomUpMDKind, \
912 ARCAnnotationProvenanceSourceMDKind, (inst), \
913 const_cast<Value*>(ptr), (old), (new))
914 /// Use this macro to annotate a sequence state change when processing
915 /// instructions top down.
916 #define ANNOTATE_TOPDOWN(inst, ptr, old, new) \
917 GenerateARCAnnotation(ARCAnnotationTopDownMDKind, \
918 ARCAnnotationProvenanceSourceMDKind, (inst), \
919 const_cast<Value*>(ptr), (old), (new))
921 #else // !ARC_ANNOTATION
922 // If annotations are off, noop.
923 #define ANNOTATE_BOTTOMUP(inst, ptr, old, new)
924 #define ANNOTATE_TOPDOWN(inst, ptr, old, new)
925 #endif // !ARC_ANNOTATION
928 /// \brief The main ARC optimization pass.
929 class ObjCARCOpt : public FunctionPass {
931 ProvenanceAnalysis PA;
933 /// A flag indicating whether this optimization pass should run.
936 /// Declarations for ObjC runtime functions, for use in creating calls to
937 /// them. These are initialized lazily to avoid cluttering up the Module
938 /// with unused declarations.
940 /// Declaration for ObjC runtime function
941 /// objc_retainAutoreleasedReturnValue.
942 Constant *RetainRVCallee;
943 /// Declaration for ObjC runtime function objc_autoreleaseReturnValue.
944 Constant *AutoreleaseRVCallee;
945 /// Declaration for ObjC runtime function objc_release.
946 Constant *ReleaseCallee;
947 /// Declaration for ObjC runtime function objc_retain.
948 Constant *RetainCallee;
949 /// Declaration for ObjC runtime function objc_retainBlock.
950 Constant *RetainBlockCallee;
951 /// Declaration for ObjC runtime function objc_autorelease.
952 Constant *AutoreleaseCallee;
954 /// Flags which determine whether each of the interesting runtine functions
955 /// is in fact used in the current function.
956 unsigned UsedInThisFunction;
958 /// The Metadata Kind for clang.imprecise_release metadata.
959 unsigned ImpreciseReleaseMDKind;
961 /// The Metadata Kind for clang.arc.copy_on_escape metadata.
962 unsigned CopyOnEscapeMDKind;
964 /// The Metadata Kind for clang.arc.no_objc_arc_exceptions metadata.
965 unsigned NoObjCARCExceptionsMDKind;
967 #ifdef ARC_ANNOTATIONS
968 /// The Metadata Kind for llvm.arc.annotation.bottomup metadata.
969 unsigned ARCAnnotationBottomUpMDKind;
970 /// The Metadata Kind for llvm.arc.annotation.topdown metadata.
971 unsigned ARCAnnotationTopDownMDKind;
972 /// The Metadata Kind for llvm.arc.annotation.provenancesource metadata.
973 unsigned ARCAnnotationProvenanceSourceMDKind;
974 #endif // ARC_ANNOATIONS
976 Constant *getRetainRVCallee(Module *M);
977 Constant *getAutoreleaseRVCallee(Module *M);
978 Constant *getReleaseCallee(Module *M);
979 Constant *getRetainCallee(Module *M);
980 Constant *getRetainBlockCallee(Module *M);
981 Constant *getAutoreleaseCallee(Module *M);
983 bool IsRetainBlockOptimizable(const Instruction *Inst);
985 void OptimizeRetainCall(Function &F, Instruction *Retain);
986 bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
987 void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
988 InstructionClass &Class);
989 void OptimizeIndividualCalls(Function &F);
991 void CheckForCFGHazards(const BasicBlock *BB,
992 DenseMap<const BasicBlock *, BBState> &BBStates,
993 BBState &MyStates) const;
994 bool VisitInstructionBottomUp(Instruction *Inst,
996 MapVector<Value *, RRInfo> &Retains,
998 bool VisitBottomUp(BasicBlock *BB,
999 DenseMap<const BasicBlock *, BBState> &BBStates,
1000 MapVector<Value *, RRInfo> &Retains);
1001 bool VisitInstructionTopDown(Instruction *Inst,
1002 DenseMap<Value *, RRInfo> &Releases,
1004 bool VisitTopDown(BasicBlock *BB,
1005 DenseMap<const BasicBlock *, BBState> &BBStates,
1006 DenseMap<Value *, RRInfo> &Releases);
1007 bool Visit(Function &F,
1008 DenseMap<const BasicBlock *, BBState> &BBStates,
1009 MapVector<Value *, RRInfo> &Retains,
1010 DenseMap<Value *, RRInfo> &Releases);
1012 void MoveCalls(Value *Arg, RRInfo &RetainsToMove, RRInfo &ReleasesToMove,
1013 MapVector<Value *, RRInfo> &Retains,
1014 DenseMap<Value *, RRInfo> &Releases,
1015 SmallVectorImpl<Instruction *> &DeadInsts,
1018 bool ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState> &BBStates,
1019 MapVector<Value *, RRInfo> &Retains,
1020 DenseMap<Value *, RRInfo> &Releases,
1022 SmallVector<Instruction *, 4> &NewRetains,
1023 SmallVector<Instruction *, 4> &NewReleases,
1024 SmallVector<Instruction *, 8> &DeadInsts,
1025 RRInfo &RetainsToMove,
1026 RRInfo &ReleasesToMove,
1029 bool &AnyPairsCompletelyEliminated);
1031 bool PerformCodePlacement(DenseMap<const BasicBlock *, BBState> &BBStates,
1032 MapVector<Value *, RRInfo> &Retains,
1033 DenseMap<Value *, RRInfo> &Releases,
1036 void OptimizeWeakCalls(Function &F);
1038 bool OptimizeSequences(Function &F);
1040 void OptimizeReturns(Function &F);
1042 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
1043 virtual bool doInitialization(Module &M);
1044 virtual bool runOnFunction(Function &F);
1045 virtual void releaseMemory();
1049 ObjCARCOpt() : FunctionPass(ID) {
1050 initializeObjCARCOptPass(*PassRegistry::getPassRegistry());
1055 char ObjCARCOpt::ID = 0;
1056 INITIALIZE_PASS_BEGIN(ObjCARCOpt,
1057 "objc-arc", "ObjC ARC optimization", false, false)
1058 INITIALIZE_PASS_DEPENDENCY(ObjCARCAliasAnalysis)
1059 INITIALIZE_PASS_END(ObjCARCOpt,
1060 "objc-arc", "ObjC ARC optimization", false, false)
1062 Pass *llvm::createObjCARCOptPass() {
1063 return new ObjCARCOpt();
1066 void ObjCARCOpt::getAnalysisUsage(AnalysisUsage &AU) const {
1067 AU.addRequired<ObjCARCAliasAnalysis>();
1068 AU.addRequired<AliasAnalysis>();
1069 // ARC optimization doesn't currently split critical edges.
1070 AU.setPreservesCFG();
1073 bool ObjCARCOpt::IsRetainBlockOptimizable(const Instruction *Inst) {
1074 // Without the magic metadata tag, we have to assume this might be an
1075 // objc_retainBlock call inserted to convert a block pointer to an id,
1076 // in which case it really is needed.
1077 if (!Inst->getMetadata(CopyOnEscapeMDKind))
1080 // If the pointer "escapes" (not including being used in a call),
1081 // the copy may be needed.
1082 if (DoesRetainableObjPtrEscape(Inst))
1085 // Otherwise, it's not needed.
1089 Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
1090 if (!RetainRVCallee) {
1091 LLVMContext &C = M->getContext();
1092 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1093 Type *Params[] = { I8X };
1094 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1095 AttributeSet Attribute =
1096 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1097 Attribute::NoUnwind);
1099 M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
1102 return RetainRVCallee;
1105 Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
1106 if (!AutoreleaseRVCallee) {
1107 LLVMContext &C = M->getContext();
1108 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
1109 Type *Params[] = { I8X };
1110 FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
1111 AttributeSet Attribute =
1112 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1113 Attribute::NoUnwind);
1114 AutoreleaseRVCallee =
1115 M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy,
1118 return AutoreleaseRVCallee;
1121 Constant *ObjCARCOpt::getReleaseCallee(Module *M) {
1122 if (!ReleaseCallee) {
1123 LLVMContext &C = M->getContext();
1124 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1125 AttributeSet Attribute =
1126 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1127 Attribute::NoUnwind);
1129 M->getOrInsertFunction(
1131 FunctionType::get(Type::getVoidTy(C), Params, /*isVarArg=*/false),
1134 return ReleaseCallee;
1137 Constant *ObjCARCOpt::getRetainCallee(Module *M) {
1138 if (!RetainCallee) {
1139 LLVMContext &C = M->getContext();
1140 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1141 AttributeSet Attribute =
1142 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1143 Attribute::NoUnwind);
1145 M->getOrInsertFunction(
1147 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1150 return RetainCallee;
1153 Constant *ObjCARCOpt::getRetainBlockCallee(Module *M) {
1154 if (!RetainBlockCallee) {
1155 LLVMContext &C = M->getContext();
1156 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1157 // objc_retainBlock is not nounwind because it calls user copy constructors
1158 // which could theoretically throw.
1160 M->getOrInsertFunction(
1162 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1165 return RetainBlockCallee;
1168 Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) {
1169 if (!AutoreleaseCallee) {
1170 LLVMContext &C = M->getContext();
1171 Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) };
1172 AttributeSet Attribute =
1173 AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
1174 Attribute::NoUnwind);
1176 M->getOrInsertFunction(
1178 FunctionType::get(Params[0], Params, /*isVarArg=*/false),
1181 return AutoreleaseCallee;
1184 /// Turn objc_retain into objc_retainAutoreleasedReturnValue if the operand is a
1187 ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
1188 ImmutableCallSite CS(GetObjCArg(Retain));
1189 const Instruction *Call = CS.getInstruction();
1191 if (Call->getParent() != Retain->getParent()) return;
1193 // Check that the call is next to the retain.
1194 BasicBlock::const_iterator I = Call;
1196 while (IsNoopInstruction(I)) ++I;
1200 // Turn it to an objc_retainAutoreleasedReturnValue..
1204 DEBUG(dbgs() << "ObjCARCOpt::OptimizeRetainCall: Transforming "
1205 "objc_retain => objc_retainAutoreleasedReturnValue"
1206 " since the operand is a return value.\n"
1208 << *Retain << "\n");
1210 cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
1212 DEBUG(dbgs() << " New: "
1213 << *Retain << "\n");
1216 /// Turn objc_retainAutoreleasedReturnValue into objc_retain if the operand is
1217 /// not a return value. Or, if it can be paired with an
1218 /// objc_autoreleaseReturnValue, delete the pair and return true.
1220 ObjCARCOpt::OptimizeRetainRVCall(Function &F, Instruction *RetainRV) {
1221 // Check for the argument being from an immediately preceding call or invoke.
1222 const Value *Arg = GetObjCArg(RetainRV);
1223 ImmutableCallSite CS(Arg);
1224 if (const Instruction *Call = CS.getInstruction()) {
1225 if (Call->getParent() == RetainRV->getParent()) {
1226 BasicBlock::const_iterator I = Call;
1228 while (IsNoopInstruction(I)) ++I;
1229 if (&*I == RetainRV)
1231 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
1232 BasicBlock *RetainRVParent = RetainRV->getParent();
1233 if (II->getNormalDest() == RetainRVParent) {
1234 BasicBlock::const_iterator I = RetainRVParent->begin();
1235 while (IsNoopInstruction(I)) ++I;
1236 if (&*I == RetainRV)
1242 // Check for being preceded by an objc_autoreleaseReturnValue on the same
1243 // pointer. In this case, we can delete the pair.
1244 BasicBlock::iterator I = RetainRV, Begin = RetainRV->getParent()->begin();
1246 do --I; while (I != Begin && IsNoopInstruction(I));
1247 if (GetBasicInstructionClass(I) == IC_AutoreleaseRV &&
1248 GetObjCArg(I) == Arg) {
1252 DEBUG(dbgs() << "ObjCARCOpt::OptimizeRetainRVCall: Erasing " << *I << "\n"
1253 << " Erasing " << *RetainRV
1256 EraseInstruction(I);
1257 EraseInstruction(RetainRV);
1262 // Turn it to a plain objc_retain.
1266 DEBUG(dbgs() << "ObjCARCOpt::OptimizeRetainRVCall: Transforming "
1267 "objc_retainAutoreleasedReturnValue => "
1268 "objc_retain since the operand is not a return value.\n"
1270 << *RetainRV << "\n");
1272 cast<CallInst>(RetainRV)->setCalledFunction(getRetainCallee(F.getParent()));
1274 DEBUG(dbgs() << " New: "
1275 << *RetainRV << "\n");
1280 /// Turn objc_autoreleaseReturnValue into objc_autorelease if the result is not
1281 /// used as a return value.
1283 ObjCARCOpt::OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
1284 InstructionClass &Class) {
1285 // Check for a return of the pointer value.
1286 const Value *Ptr = GetObjCArg(AutoreleaseRV);
1287 SmallVector<const Value *, 2> Users;
1288 Users.push_back(Ptr);
1290 Ptr = Users.pop_back_val();
1291 for (Value::const_use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end();
1293 const User *I = *UI;
1294 if (isa<ReturnInst>(I) || GetBasicInstructionClass(I) == IC_RetainRV)
1296 if (isa<BitCastInst>(I))
1299 } while (!Users.empty());
1304 DEBUG(dbgs() << "ObjCARCOpt::OptimizeAutoreleaseRVCall: Transforming "
1305 "objc_autoreleaseReturnValue => "
1306 "objc_autorelease since its operand is not used as a return "
1309 << *AutoreleaseRV << "\n");
1311 CallInst *AutoreleaseRVCI = cast<CallInst>(AutoreleaseRV);
1313 setCalledFunction(getAutoreleaseCallee(F.getParent()));
1314 AutoreleaseRVCI->setTailCall(false); // Never tail call objc_autorelease.
1315 Class = IC_Autorelease;
1317 DEBUG(dbgs() << " New: "
1318 << *AutoreleaseRV << "\n");
1322 /// Visit each call, one at a time, and make simplifications without doing any
1323 /// additional analysis.
1324 void ObjCARCOpt::OptimizeIndividualCalls(Function &F) {
1325 // Reset all the flags in preparation for recomputing them.
1326 UsedInThisFunction = 0;
1328 // Visit all objc_* calls in F.
1329 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
1330 Instruction *Inst = &*I++;
1332 InstructionClass Class = GetBasicInstructionClass(Inst);
1334 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Visiting: Class: "
1335 << Class << "; " << *Inst << "\n");
1340 // Delete no-op casts. These function calls have special semantics, but
1341 // the semantics are entirely implemented via lowering in the front-end,
1342 // so by the time they reach the optimizer, they are just no-op calls
1343 // which return their argument.
1345 // There are gray areas here, as the ability to cast reference-counted
1346 // pointers to raw void* and back allows code to break ARC assumptions,
1347 // however these are currently considered to be unimportant.
1351 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Erasing no-op cast:"
1352 " " << *Inst << "\n");
1353 EraseInstruction(Inst);
1356 // If the pointer-to-weak-pointer is null, it's undefined behavior.
1359 case IC_LoadWeakRetained:
1361 case IC_DestroyWeak: {
1362 CallInst *CI = cast<CallInst>(Inst);
1363 if (IsNullOrUndef(CI->getArgOperand(0))) {
1365 Type *Ty = CI->getArgOperand(0)->getType();
1366 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1367 Constant::getNullValue(Ty),
1369 llvm::Value *NewValue = UndefValue::get(CI->getType());
1370 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: A null "
1371 "pointer-to-weak-pointer is undefined behavior.\n"
1375 CI->replaceAllUsesWith(NewValue);
1376 CI->eraseFromParent();
1383 CallInst *CI = cast<CallInst>(Inst);
1384 if (IsNullOrUndef(CI->getArgOperand(0)) ||
1385 IsNullOrUndef(CI->getArgOperand(1))) {
1387 Type *Ty = CI->getArgOperand(0)->getType();
1388 new StoreInst(UndefValue::get(cast<PointerType>(Ty)->getElementType()),
1389 Constant::getNullValue(Ty),
1392 llvm::Value *NewValue = UndefValue::get(CI->getType());
1393 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: A null "
1394 "pointer-to-weak-pointer is undefined behavior.\n"
1399 CI->replaceAllUsesWith(NewValue);
1400 CI->eraseFromParent();
1406 OptimizeRetainCall(F, Inst);
1409 if (OptimizeRetainRVCall(F, Inst))
1412 case IC_AutoreleaseRV:
1413 OptimizeAutoreleaseRVCall(F, Inst, Class);
1417 // objc_autorelease(x) -> objc_release(x) if x is otherwise unused.
1418 if (IsAutorelease(Class) && Inst->use_empty()) {
1419 CallInst *Call = cast<CallInst>(Inst);
1420 const Value *Arg = Call->getArgOperand(0);
1421 Arg = FindSingleUseIdentifiedObject(Arg);
1426 // Create the declaration lazily.
1427 LLVMContext &C = Inst->getContext();
1429 CallInst::Create(getReleaseCallee(F.getParent()),
1430 Call->getArgOperand(0), "", Call);
1431 NewCall->setMetadata(ImpreciseReleaseMDKind,
1432 MDNode::get(C, ArrayRef<Value *>()));
1434 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Replacing "
1435 "objc_autorelease(x) with objc_release(x) since x is "
1436 "otherwise unused.\n"
1437 " Old: " << *Call <<
1441 EraseInstruction(Call);
1447 // For functions which can never be passed stack arguments, add
1449 if (IsAlwaysTail(Class)) {
1451 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Adding tail keyword"
1452 " to function since it can never be passed stack args: " << *Inst <<
1454 cast<CallInst>(Inst)->setTailCall();
1457 // Ensure that functions that can never have a "tail" keyword due to the
1458 // semantics of ARC truly do not do so.
1459 if (IsNeverTail(Class)) {
1461 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Removing tail "
1462 "keyword from function: " << *Inst <<
1464 cast<CallInst>(Inst)->setTailCall(false);
1467 // Set nounwind as needed.
1468 if (IsNoThrow(Class)) {
1470 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Found no throw"
1471 " class. Setting nounwind on: " << *Inst << "\n");
1472 cast<CallInst>(Inst)->setDoesNotThrow();
1475 if (!IsNoopOnNull(Class)) {
1476 UsedInThisFunction |= 1 << Class;
1480 const Value *Arg = GetObjCArg(Inst);
1482 // ARC calls with null are no-ops. Delete them.
1483 if (IsNullOrUndef(Arg)) {
1486 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: ARC calls with "
1487 " null are no-ops. Erasing: " << *Inst << "\n");
1488 EraseInstruction(Inst);
1492 // Keep track of which of retain, release, autorelease, and retain_block
1493 // are actually present in this function.
1494 UsedInThisFunction |= 1 << Class;
1496 // If Arg is a PHI, and one or more incoming values to the
1497 // PHI are null, and the call is control-equivalent to the PHI, and there
1498 // are no relevant side effects between the PHI and the call, the call
1499 // could be pushed up to just those paths with non-null incoming values.
1500 // For now, don't bother splitting critical edges for this.
1501 SmallVector<std::pair<Instruction *, const Value *>, 4> Worklist;
1502 Worklist.push_back(std::make_pair(Inst, Arg));
1504 std::pair<Instruction *, const Value *> Pair = Worklist.pop_back_val();
1508 const PHINode *PN = dyn_cast<PHINode>(Arg);
1511 // Determine if the PHI has any null operands, or any incoming
1513 bool HasNull = false;
1514 bool HasCriticalEdges = false;
1515 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1517 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
1518 if (IsNullOrUndef(Incoming))
1520 else if (cast<TerminatorInst>(PN->getIncomingBlock(i)->back())
1521 .getNumSuccessors() != 1) {
1522 HasCriticalEdges = true;
1526 // If we have null operands and no critical edges, optimize.
1527 if (!HasCriticalEdges && HasNull) {
1528 SmallPtrSet<Instruction *, 4> DependingInstructions;
1529 SmallPtrSet<const BasicBlock *, 4> Visited;
1531 // Check that there is nothing that cares about the reference
1532 // count between the call and the phi.
1535 case IC_RetainBlock:
1536 // These can always be moved up.
1539 // These can't be moved across things that care about the retain
1541 FindDependencies(NeedsPositiveRetainCount, Arg,
1542 Inst->getParent(), Inst,
1543 DependingInstructions, Visited, PA);
1545 case IC_Autorelease:
1546 // These can't be moved across autorelease pool scope boundaries.
1547 FindDependencies(AutoreleasePoolBoundary, Arg,
1548 Inst->getParent(), Inst,
1549 DependingInstructions, Visited, PA);
1552 case IC_AutoreleaseRV:
1553 // Don't move these; the RV optimization depends on the autoreleaseRV
1554 // being tail called, and the retainRV being immediately after a call
1555 // (which might still happen if we get lucky with codegen layout, but
1556 // it's not worth taking the chance).
1559 llvm_unreachable("Invalid dependence flavor");
1562 if (DependingInstructions.size() == 1 &&
1563 *DependingInstructions.begin() == PN) {
1566 // Clone the call into each predecessor that has a non-null value.
1567 CallInst *CInst = cast<CallInst>(Inst);
1568 Type *ParamTy = CInst->getArgOperand(0)->getType();
1569 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1571 StripPointerCastsAndObjCCalls(PN->getIncomingValue(i));
1572 if (!IsNullOrUndef(Incoming)) {
1573 CallInst *Clone = cast<CallInst>(CInst->clone());
1574 Value *Op = PN->getIncomingValue(i);
1575 Instruction *InsertPos = &PN->getIncomingBlock(i)->back();
1576 if (Op->getType() != ParamTy)
1577 Op = new BitCastInst(Op, ParamTy, "", InsertPos);
1578 Clone->setArgOperand(0, Op);
1579 Clone->insertBefore(InsertPos);
1581 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Cloning "
1584 "clone at " << *InsertPos << "\n");
1585 Worklist.push_back(std::make_pair(Clone, Incoming));
1588 // Erase the original call.
1589 DEBUG(dbgs() << "Erasing: " << *CInst << "\n");
1590 EraseInstruction(CInst);
1594 } while (!Worklist.empty());
1596 DEBUG(dbgs() << "ObjCARCOpt::OptimizeIndividualCalls: Finished List.\n");
1599 /// Check for critical edges, loop boundaries, irreducible control flow, or
1600 /// other CFG structures where moving code across the edge would result in it
1601 /// being executed more.
1603 ObjCARCOpt::CheckForCFGHazards(const BasicBlock *BB,
1604 DenseMap<const BasicBlock *, BBState> &BBStates,
1605 BBState &MyStates) const {
1606 // If any top-down local-use or possible-dec has a succ which is earlier in
1607 // the sequence, forget it.
1608 for (BBState::ptr_iterator I = MyStates.top_down_ptr_begin(),
1609 E = MyStates.top_down_ptr_end(); I != E; ++I)
1610 switch (I->second.GetSeq()) {
1613 const Value *Arg = I->first;
1614 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1615 bool SomeSuccHasSame = false;
1616 bool AllSuccsHaveSame = true;
1617 PtrState &S = I->second;
1618 succ_const_iterator SI(TI), SE(TI, false);
1620 for (; SI != SE; ++SI) {
1621 Sequence SuccSSeq = S_None;
1622 bool SuccSRRIKnownSafe = false;
1623 // If VisitBottomUp has pointer information for this successor, take
1624 // what we know about it.
1625 DenseMap<const BasicBlock *, BBState>::iterator BBI =
1627 assert(BBI != BBStates.end());
1628 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
1629 SuccSSeq = SuccS.GetSeq();
1630 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
1633 case S_CanRelease: {
1634 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
1635 S.ClearSequenceProgress();
1641 SomeSuccHasSame = true;
1645 case S_MovableRelease:
1646 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
1647 AllSuccsHaveSame = false;
1650 llvm_unreachable("bottom-up pointer in retain state!");
1653 // If the state at the other end of any of the successor edges
1654 // matches the current state, require all edges to match. This
1655 // guards against loops in the middle of a sequence.
1656 if (SomeSuccHasSame && !AllSuccsHaveSame)
1657 S.ClearSequenceProgress();
1660 case S_CanRelease: {
1661 const Value *Arg = I->first;
1662 const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
1663 bool SomeSuccHasSame = false;
1664 bool AllSuccsHaveSame = true;
1665 PtrState &S = I->second;
1666 succ_const_iterator SI(TI), SE(TI, false);
1668 for (; SI != SE; ++SI) {
1669 Sequence SuccSSeq = S_None;
1670 bool SuccSRRIKnownSafe = false;
1671 // If VisitBottomUp has pointer information for this successor, take
1672 // what we know about it.
1673 DenseMap<const BasicBlock *, BBState>::iterator BBI =
1675 assert(BBI != BBStates.end());
1676 const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
1677 SuccSSeq = SuccS.GetSeq();
1678 SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
1681 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
1682 S.ClearSequenceProgress();
1688 SomeSuccHasSame = true;
1692 case S_MovableRelease:
1694 if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
1695 AllSuccsHaveSame = false;
1698 llvm_unreachable("bottom-up pointer in retain state!");
1701 // If the state at the other end of any of the successor edges
1702 // matches the current state, require all edges to match. This
1703 // guards against loops in the middle of a sequence.
1704 if (SomeSuccHasSame && !AllSuccsHaveSame)
1705 S.ClearSequenceProgress();
1712 ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst,
1714 MapVector<Value *, RRInfo> &Retains,
1715 BBState &MyStates) {
1716 bool NestingDetected = false;
1717 InstructionClass Class = GetInstructionClass(Inst);
1718 const Value *Arg = 0;
1722 Arg = GetObjCArg(Inst);
1724 PtrState &S = MyStates.getPtrBottomUpState(Arg);
1726 // If we see two releases in a row on the same pointer. If so, make
1727 // a note, and we'll cicle back to revisit it after we've
1728 // hopefully eliminated the second release, which may allow us to
1729 // eliminate the first release too.
1730 // Theoretically we could implement removal of nested retain+release
1731 // pairs by making PtrState hold a stack of states, but this is
1732 // simple and avoids adding overhead for the non-nested case.
1733 if (S.GetSeq() == S_Release || S.GetSeq() == S_MovableRelease) {
1734 DEBUG(dbgs() << "ObjCARCOpt::VisitInstructionBottomUp: Found nested "
1735 "releases (i.e. a release pair)\n");
1736 NestingDetected = true;
1739 MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
1740 Sequence NewSeq = ReleaseMetadata ? S_MovableRelease : S_Release;
1741 ANNOTATE_BOTTOMUP(Inst, Arg, S.GetSeq(), NewSeq);
1742 S.ResetSequenceProgress(NewSeq);
1743 S.RRI.ReleaseMetadata = ReleaseMetadata;
1744 S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
1745 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
1746 S.RRI.Calls.insert(Inst);
1747 S.SetKnownPositiveRefCount();
1750 case IC_RetainBlock:
1751 // An objc_retainBlock call with just a use may need to be kept,
1752 // because it may be copying a block from the stack to the heap.
1753 if (!IsRetainBlockOptimizable(Inst))
1758 Arg = GetObjCArg(Inst);
1760 PtrState &S = MyStates.getPtrBottomUpState(Arg);
1761 S.SetKnownPositiveRefCount();
1763 Sequence OldSeq = S.GetSeq();
1767 case S_MovableRelease:
1769 S.RRI.ReverseInsertPts.clear();
1772 // Don't do retain+release tracking for IC_RetainRV, because it's
1773 // better to let it remain as the first instruction after a call.
1774 if (Class != IC_RetainRV) {
1775 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
1776 Retains[Inst] = S.RRI;
1778 S.ClearSequenceProgress();
1783 llvm_unreachable("bottom-up pointer in retain state!");
1785 ANNOTATE_BOTTOMUP(Inst, Arg, OldSeq, S.GetSeq());
1786 return NestingDetected;
1788 case IC_AutoreleasepoolPop:
1789 // Conservatively, clear MyStates for all known pointers.
1790 MyStates.clearBottomUpPointers();
1791 return NestingDetected;
1792 case IC_AutoreleasepoolPush:
1794 // These are irrelevant.
1795 return NestingDetected;
1800 // Consider any other possible effects of this instruction on each
1801 // pointer being tracked.
1802 for (BBState::ptr_iterator MI = MyStates.bottom_up_ptr_begin(),
1803 ME = MyStates.bottom_up_ptr_end(); MI != ME; ++MI) {
1804 const Value *Ptr = MI->first;
1806 continue; // Handled above.
1807 PtrState &S = MI->second;
1808 Sequence Seq = S.GetSeq();
1810 // Check for possible releases.
1811 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
1812 S.ClearKnownPositiveRefCount();
1815 S.SetSeq(S_CanRelease);
1816 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S.GetSeq());
1820 case S_MovableRelease:
1825 llvm_unreachable("bottom-up pointer in retain state!");
1829 // Check for possible direct uses.
1832 case S_MovableRelease:
1833 if (CanUse(Inst, Ptr, PA, Class)) {
1834 assert(S.RRI.ReverseInsertPts.empty());
1835 // If this is an invoke instruction, we're scanning it as part of
1836 // one of its successor blocks, since we can't insert code after it
1837 // in its own block, and we don't want to split critical edges.
1838 if (isa<InvokeInst>(Inst))
1839 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
1841 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
1843 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S_Use);
1844 } else if (Seq == S_Release && IsUser(Class)) {
1845 // Non-movable releases depend on any possible objc pointer use.
1847 ANNOTATE_BOTTOMUP(Inst, Ptr, S_Release, S_Stop);
1848 assert(S.RRI.ReverseInsertPts.empty());
1849 // As above; handle invoke specially.
1850 if (isa<InvokeInst>(Inst))
1851 S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
1853 S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
1857 if (CanUse(Inst, Ptr, PA, Class)) {
1859 ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S_Use);
1867 llvm_unreachable("bottom-up pointer in retain state!");
1871 return NestingDetected;
1875 ObjCARCOpt::VisitBottomUp(BasicBlock *BB,
1876 DenseMap<const BasicBlock *, BBState> &BBStates,
1877 MapVector<Value *, RRInfo> &Retains) {
1878 bool NestingDetected = false;
1879 BBState &MyStates = BBStates[BB];
1881 // Merge the states from each successor to compute the initial state
1882 // for the current block.
1883 BBState::edge_iterator SI(MyStates.succ_begin()),
1884 SE(MyStates.succ_end());
1886 const BasicBlock *Succ = *SI;
1887 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ);
1888 assert(I != BBStates.end());
1889 MyStates.InitFromSucc(I->second);
1891 for (; SI != SE; ++SI) {
1893 I = BBStates.find(Succ);
1894 assert(I != BBStates.end());
1895 MyStates.MergeSucc(I->second);
1899 #ifdef ARC_ANNOTATIONS
1900 if (EnableARCAnnotations) {
1901 // If ARC Annotations are enabled, output the current state of pointers at the
1902 // bottom of the basic block.
1903 for(BBState::ptr_const_iterator I = MyStates.bottom_up_ptr_begin(),
1904 E = MyStates.bottom_up_ptr_end(); I != E; ++I) {
1905 Value *Ptr = const_cast<Value*>(I->first);
1906 Sequence Seq = I->second.GetSeq();
1907 GenerateARCBBTerminatorAnnotation("llvm.arc.annotation.bottomup.bbend",
1914 // Visit all the instructions, bottom-up.
1915 for (BasicBlock::iterator I = BB->end(), E = BB->begin(); I != E; --I) {
1916 Instruction *Inst = llvm::prior(I);
1918 // Invoke instructions are visited as part of their successors (below).
1919 if (isa<InvokeInst>(Inst))
1922 DEBUG(dbgs() << "ObjCARCOpt::VisitButtonUp: Visiting " << *Inst << "\n");
1924 NestingDetected |= VisitInstructionBottomUp(Inst, BB, Retains, MyStates);
1927 // If there's a predecessor with an invoke, visit the invoke as if it were
1928 // part of this block, since we can't insert code after an invoke in its own
1929 // block, and we don't want to split critical edges.
1930 for (BBState::edge_iterator PI(MyStates.pred_begin()),
1931 PE(MyStates.pred_end()); PI != PE; ++PI) {
1932 BasicBlock *Pred = *PI;
1933 if (InvokeInst *II = dyn_cast<InvokeInst>(&Pred->back()))
1934 NestingDetected |= VisitInstructionBottomUp(II, BB, Retains, MyStates);
1937 #ifdef ARC_ANNOTATIONS
1938 if (EnableARCAnnotations) {
1939 // If ARC Annotations are enabled, output the current state of pointers at the
1940 // top of the basic block.
1941 for(BBState::ptr_const_iterator I = MyStates.bottom_up_ptr_begin(),
1942 E = MyStates.bottom_up_ptr_end(); I != E; ++I) {
1943 Value *Ptr = const_cast<Value*>(I->first);
1944 Sequence Seq = I->second.GetSeq();
1945 GenerateARCBBEntranceAnnotation("llvm.arc.annotation.bottomup.bbstart",
1951 return NestingDetected;
1955 ObjCARCOpt::VisitInstructionTopDown(Instruction *Inst,
1956 DenseMap<Value *, RRInfo> &Releases,
1957 BBState &MyStates) {
1958 bool NestingDetected = false;
1959 InstructionClass Class = GetInstructionClass(Inst);
1960 const Value *Arg = 0;
1963 case IC_RetainBlock:
1964 // An objc_retainBlock call with just a use may need to be kept,
1965 // because it may be copying a block from the stack to the heap.
1966 if (!IsRetainBlockOptimizable(Inst))
1971 Arg = GetObjCArg(Inst);
1973 PtrState &S = MyStates.getPtrTopDownState(Arg);
1975 // Don't do retain+release tracking for IC_RetainRV, because it's
1976 // better to let it remain as the first instruction after a call.
1977 if (Class != IC_RetainRV) {
1978 // If we see two retains in a row on the same pointer. If so, make
1979 // a note, and we'll cicle back to revisit it after we've
1980 // hopefully eliminated the second retain, which may allow us to
1981 // eliminate the first retain too.
1982 // Theoretically we could implement removal of nested retain+release
1983 // pairs by making PtrState hold a stack of states, but this is
1984 // simple and avoids adding overhead for the non-nested case.
1985 if (S.GetSeq() == S_Retain)
1986 NestingDetected = true;
1988 ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_Retain);
1989 S.ResetSequenceProgress(S_Retain);
1990 S.RRI.IsRetainBlock = Class == IC_RetainBlock;
1991 S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
1992 S.RRI.Calls.insert(Inst);
1995 S.SetKnownPositiveRefCount();
1997 // A retain can be a potential use; procede to the generic checking
2002 Arg = GetObjCArg(Inst);
2004 PtrState &S = MyStates.getPtrTopDownState(Arg);
2005 S.ClearKnownPositiveRefCount();
2007 switch (S.GetSeq()) {
2010 S.RRI.ReverseInsertPts.clear();
2013 S.RRI.ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind);
2014 S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
2015 Releases[Inst] = S.RRI;
2016 ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_None);
2017 S.ClearSequenceProgress();
2023 case S_MovableRelease:
2024 llvm_unreachable("top-down pointer in release state!");
2028 case IC_AutoreleasepoolPop:
2029 // Conservatively, clear MyStates for all known pointers.
2030 MyStates.clearTopDownPointers();
2031 return NestingDetected;
2032 case IC_AutoreleasepoolPush:
2034 // These are irrelevant.
2035 return NestingDetected;
2040 // Consider any other possible effects of this instruction on each
2041 // pointer being tracked.
2042 for (BBState::ptr_iterator MI = MyStates.top_down_ptr_begin(),
2043 ME = MyStates.top_down_ptr_end(); MI != ME; ++MI) {
2044 const Value *Ptr = MI->first;
2046 continue; // Handled above.
2047 PtrState &S = MI->second;
2048 Sequence Seq = S.GetSeq();
2050 // Check for possible releases.
2051 if (CanAlterRefCount(Inst, Ptr, PA, Class)) {
2052 S.ClearKnownPositiveRefCount();
2055 S.SetSeq(S_CanRelease);
2056 ANNOTATE_TOPDOWN(Inst, Ptr, Seq, S_CanRelease);
2057 assert(S.RRI.ReverseInsertPts.empty());
2058 S.RRI.ReverseInsertPts.insert(Inst);
2060 // One call can't cause a transition from S_Retain to S_CanRelease
2061 // and S_CanRelease to S_Use. If we've made the first transition,
2070 case S_MovableRelease:
2071 llvm_unreachable("top-down pointer in release state!");
2075 // Check for possible direct uses.
2078 if (CanUse(Inst, Ptr, PA, Class)) {
2080 ANNOTATE_TOPDOWN(Inst, Ptr, Seq, S_Use);
2089 case S_MovableRelease:
2090 llvm_unreachable("top-down pointer in release state!");
2094 return NestingDetected;
2098 ObjCARCOpt::VisitTopDown(BasicBlock *BB,
2099 DenseMap<const BasicBlock *, BBState> &BBStates,
2100 DenseMap<Value *, RRInfo> &Releases) {
2101 bool NestingDetected = false;
2102 BBState &MyStates = BBStates[BB];
2104 // Merge the states from each predecessor to compute the initial state
2105 // for the current block.
2106 BBState::edge_iterator PI(MyStates.pred_begin()),
2107 PE(MyStates.pred_end());
2109 const BasicBlock *Pred = *PI;
2110 DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred);
2111 assert(I != BBStates.end());
2112 MyStates.InitFromPred(I->second);
2114 for (; PI != PE; ++PI) {
2116 I = BBStates.find(Pred);
2117 assert(I != BBStates.end());
2118 MyStates.MergePred(I->second);
2122 #ifdef ARC_ANNOTATIONS
2123 if (EnableARCAnnotations) {
2124 // If ARC Annotations are enabled, output the current state of pointers at the
2125 // top of the basic block.
2126 for(BBState::ptr_const_iterator I = MyStates.top_down_ptr_begin(),
2127 E = MyStates.top_down_ptr_end(); I != E; ++I) {
2128 Value *Ptr = const_cast<Value*>(I->first);
2129 Sequence Seq = I->second.GetSeq();
2130 GenerateARCBBEntranceAnnotation("llvm.arc.annotation.topdown.bbstart",
2136 // Visit all the instructions, top-down.
2137 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2138 Instruction *Inst = I;
2140 DEBUG(dbgs() << "ObjCARCOpt::VisitTopDown: Visiting " << *Inst << "\n");
2142 NestingDetected |= VisitInstructionTopDown(Inst, Releases, MyStates);
2145 #ifdef ARC_ANNOTATIONS
2146 if (EnableARCAnnotations) {
2147 // If ARC Annotations are enabled, output the current state of pointers at the
2148 // bottom of the basic block.
2149 for(BBState::ptr_const_iterator I = MyStates.top_down_ptr_begin(),
2150 E = MyStates.top_down_ptr_end(); I != E; ++I) {
2151 Value *Ptr = const_cast<Value*>(I->first);
2152 Sequence Seq = I->second.GetSeq();
2153 GenerateARCBBTerminatorAnnotation("llvm.arc.annotation.topdown.bbend",
2159 CheckForCFGHazards(BB, BBStates, MyStates);
2160 return NestingDetected;
2164 ComputePostOrders(Function &F,
2165 SmallVectorImpl<BasicBlock *> &PostOrder,
2166 SmallVectorImpl<BasicBlock *> &ReverseCFGPostOrder,
2167 unsigned NoObjCARCExceptionsMDKind,
2168 DenseMap<const BasicBlock *, BBState> &BBStates) {
2169 /// The visited set, for doing DFS walks.
2170 SmallPtrSet<BasicBlock *, 16> Visited;
2172 // Do DFS, computing the PostOrder.
2173 SmallPtrSet<BasicBlock *, 16> OnStack;
2174 SmallVector<std::pair<BasicBlock *, succ_iterator>, 16> SuccStack;
2176 // Functions always have exactly one entry block, and we don't have
2177 // any other block that we treat like an entry block.
2178 BasicBlock *EntryBB = &F.getEntryBlock();
2179 BBState &MyStates = BBStates[EntryBB];
2180 MyStates.SetAsEntry();
2181 TerminatorInst *EntryTI = cast<TerminatorInst>(&EntryBB->back());
2182 SuccStack.push_back(std::make_pair(EntryBB, succ_iterator(EntryTI)));
2183 Visited.insert(EntryBB);
2184 OnStack.insert(EntryBB);
2187 BasicBlock *CurrBB = SuccStack.back().first;
2188 TerminatorInst *TI = cast<TerminatorInst>(&CurrBB->back());
2189 succ_iterator SE(TI, false);
2191 while (SuccStack.back().second != SE) {
2192 BasicBlock *SuccBB = *SuccStack.back().second++;
2193 if (Visited.insert(SuccBB)) {
2194 TerminatorInst *TI = cast<TerminatorInst>(&SuccBB->back());
2195 SuccStack.push_back(std::make_pair(SuccBB, succ_iterator(TI)));
2196 BBStates[CurrBB].addSucc(SuccBB);
2197 BBState &SuccStates = BBStates[SuccBB];
2198 SuccStates.addPred(CurrBB);
2199 OnStack.insert(SuccBB);
2203 if (!OnStack.count(SuccBB)) {
2204 BBStates[CurrBB].addSucc(SuccBB);
2205 BBStates[SuccBB].addPred(CurrBB);
2208 OnStack.erase(CurrBB);
2209 PostOrder.push_back(CurrBB);
2210 SuccStack.pop_back();
2211 } while (!SuccStack.empty());
2215 // Do reverse-CFG DFS, computing the reverse-CFG PostOrder.
2216 // Functions may have many exits, and there also blocks which we treat
2217 // as exits due to ignored edges.
2218 SmallVector<std::pair<BasicBlock *, BBState::edge_iterator>, 16> PredStack;
2219 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2220 BasicBlock *ExitBB = I;
2221 BBState &MyStates = BBStates[ExitBB];
2222 if (!MyStates.isExit())
2225 MyStates.SetAsExit();
2227 PredStack.push_back(std::make_pair(ExitBB, MyStates.pred_begin()));
2228 Visited.insert(ExitBB);
2229 while (!PredStack.empty()) {
2230 reverse_dfs_next_succ:
2231 BBState::edge_iterator PE = BBStates[PredStack.back().first].pred_end();
2232 while (PredStack.back().second != PE) {
2233 BasicBlock *BB = *PredStack.back().second++;
2234 if (Visited.insert(BB)) {
2235 PredStack.push_back(std::make_pair(BB, BBStates[BB].pred_begin()));
2236 goto reverse_dfs_next_succ;
2239 ReverseCFGPostOrder.push_back(PredStack.pop_back_val().first);
2244 // Visit the function both top-down and bottom-up.
2246 ObjCARCOpt::Visit(Function &F,
2247 DenseMap<const BasicBlock *, BBState> &BBStates,
2248 MapVector<Value *, RRInfo> &Retains,
2249 DenseMap<Value *, RRInfo> &Releases) {
2251 // Use reverse-postorder traversals, because we magically know that loops
2252 // will be well behaved, i.e. they won't repeatedly call retain on a single
2253 // pointer without doing a release. We can't use the ReversePostOrderTraversal
2254 // class here because we want the reverse-CFG postorder to consider each
2255 // function exit point, and we want to ignore selected cycle edges.
2256 SmallVector<BasicBlock *, 16> PostOrder;
2257 SmallVector<BasicBlock *, 16> ReverseCFGPostOrder;
2258 ComputePostOrders(F, PostOrder, ReverseCFGPostOrder,
2259 NoObjCARCExceptionsMDKind,
2262 // Use reverse-postorder on the reverse CFG for bottom-up.
2263 bool BottomUpNestingDetected = false;
2264 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2265 ReverseCFGPostOrder.rbegin(), E = ReverseCFGPostOrder.rend();
2267 BottomUpNestingDetected |= VisitBottomUp(*I, BBStates, Retains);
2269 // Use reverse-postorder for top-down.
2270 bool TopDownNestingDetected = false;
2271 for (SmallVectorImpl<BasicBlock *>::const_reverse_iterator I =
2272 PostOrder.rbegin(), E = PostOrder.rend();
2274 TopDownNestingDetected |= VisitTopDown(*I, BBStates, Releases);
2276 return TopDownNestingDetected && BottomUpNestingDetected;
2279 /// Move the calls in RetainsToMove and ReleasesToMove.
2280 void ObjCARCOpt::MoveCalls(Value *Arg,
2281 RRInfo &RetainsToMove,
2282 RRInfo &ReleasesToMove,
2283 MapVector<Value *, RRInfo> &Retains,
2284 DenseMap<Value *, RRInfo> &Releases,
2285 SmallVectorImpl<Instruction *> &DeadInsts,
2287 Type *ArgTy = Arg->getType();
2288 Type *ParamTy = PointerType::getUnqual(Type::getInt8Ty(ArgTy->getContext()));
2290 // Insert the new retain and release calls.
2291 for (SmallPtrSet<Instruction *, 2>::const_iterator
2292 PI = ReleasesToMove.ReverseInsertPts.begin(),
2293 PE = ReleasesToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2294 Instruction *InsertPt = *PI;
2295 Value *MyArg = ArgTy == ParamTy ? Arg :
2296 new BitCastInst(Arg, ParamTy, "", InsertPt);
2298 CallInst::Create(RetainsToMove.IsRetainBlock ?
2299 getRetainBlockCallee(M) : getRetainCallee(M),
2300 MyArg, "", InsertPt);
2301 Call->setDoesNotThrow();
2302 if (RetainsToMove.IsRetainBlock)
2303 Call->setMetadata(CopyOnEscapeMDKind,
2304 MDNode::get(M->getContext(), ArrayRef<Value *>()));
2306 Call->setTailCall();
2308 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Inserting new Release: " << *Call
2310 " At insertion point: " << *InsertPt
2313 for (SmallPtrSet<Instruction *, 2>::const_iterator
2314 PI = RetainsToMove.ReverseInsertPts.begin(),
2315 PE = RetainsToMove.ReverseInsertPts.end(); PI != PE; ++PI) {
2316 Instruction *InsertPt = *PI;
2317 Value *MyArg = ArgTy == ParamTy ? Arg :
2318 new BitCastInst(Arg, ParamTy, "", InsertPt);
2319 CallInst *Call = CallInst::Create(getReleaseCallee(M), MyArg,
2321 // Attach a clang.imprecise_release metadata tag, if appropriate.
2322 if (MDNode *M = ReleasesToMove.ReleaseMetadata)
2323 Call->setMetadata(ImpreciseReleaseMDKind, M);
2324 Call->setDoesNotThrow();
2325 if (ReleasesToMove.IsTailCallRelease)
2326 Call->setTailCall();
2328 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Inserting new Retain: " << *Call
2330 " At insertion point: " << *InsertPt
2334 // Delete the original retain and release calls.
2335 for (SmallPtrSet<Instruction *, 2>::const_iterator
2336 AI = RetainsToMove.Calls.begin(),
2337 AE = RetainsToMove.Calls.end(); AI != AE; ++AI) {
2338 Instruction *OrigRetain = *AI;
2339 Retains.blot(OrigRetain);
2340 DeadInsts.push_back(OrigRetain);
2341 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Deleting retain: " << *OrigRetain <<
2344 for (SmallPtrSet<Instruction *, 2>::const_iterator
2345 AI = ReleasesToMove.Calls.begin(),
2346 AE = ReleasesToMove.Calls.end(); AI != AE; ++AI) {
2347 Instruction *OrigRelease = *AI;
2348 Releases.erase(OrigRelease);
2349 DeadInsts.push_back(OrigRelease);
2350 DEBUG(dbgs() << "ObjCARCOpt::MoveCalls: Deleting release: " << *OrigRelease
2356 ObjCARCOpt::ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState>
2358 MapVector<Value *, RRInfo> &Retains,
2359 DenseMap<Value *, RRInfo> &Releases,
2361 SmallVector<Instruction *, 4> &NewRetains,
2362 SmallVector<Instruction *, 4> &NewReleases,
2363 SmallVector<Instruction *, 8> &DeadInsts,
2364 RRInfo &RetainsToMove,
2365 RRInfo &ReleasesToMove,
2368 bool &AnyPairsCompletelyEliminated) {
2369 // If a pair happens in a region where it is known that the reference count
2370 // is already incremented, we can similarly ignore possible decrements.
2371 bool KnownSafeTD = true, KnownSafeBU = true;
2373 // Connect the dots between the top-down-collected RetainsToMove and
2374 // bottom-up-collected ReleasesToMove to form sets of related calls.
2375 // This is an iterative process so that we connect multiple releases
2376 // to multiple retains if needed.
2377 unsigned OldDelta = 0;
2378 unsigned NewDelta = 0;
2379 unsigned OldCount = 0;
2380 unsigned NewCount = 0;
2381 bool FirstRelease = true;
2382 bool FirstRetain = true;
2384 for (SmallVectorImpl<Instruction *>::const_iterator
2385 NI = NewRetains.begin(), NE = NewRetains.end(); NI != NE; ++NI) {
2386 Instruction *NewRetain = *NI;
2387 MapVector<Value *, RRInfo>::const_iterator It = Retains.find(NewRetain);
2388 assert(It != Retains.end());
2389 const RRInfo &NewRetainRRI = It->second;
2390 KnownSafeTD &= NewRetainRRI.KnownSafe;
2391 for (SmallPtrSet<Instruction *, 2>::const_iterator
2392 LI = NewRetainRRI.Calls.begin(),
2393 LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
2394 Instruction *NewRetainRelease = *LI;
2395 DenseMap<Value *, RRInfo>::const_iterator Jt =
2396 Releases.find(NewRetainRelease);
2397 if (Jt == Releases.end())
2399 const RRInfo &NewRetainReleaseRRI = Jt->second;
2400 assert(NewRetainReleaseRRI.Calls.count(NewRetain));
2401 if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
2403 BBStates[NewRetainRelease->getParent()].GetAllPathCount();
2405 // Merge the ReleaseMetadata and IsTailCallRelease values.
2407 ReleasesToMove.ReleaseMetadata =
2408 NewRetainReleaseRRI.ReleaseMetadata;
2409 ReleasesToMove.IsTailCallRelease =
2410 NewRetainReleaseRRI.IsTailCallRelease;
2411 FirstRelease = false;
2413 if (ReleasesToMove.ReleaseMetadata !=
2414 NewRetainReleaseRRI.ReleaseMetadata)
2415 ReleasesToMove.ReleaseMetadata = 0;
2416 if (ReleasesToMove.IsTailCallRelease !=
2417 NewRetainReleaseRRI.IsTailCallRelease)
2418 ReleasesToMove.IsTailCallRelease = false;
2421 // Collect the optimal insertion points.
2423 for (SmallPtrSet<Instruction *, 2>::const_iterator
2424 RI = NewRetainReleaseRRI.ReverseInsertPts.begin(),
2425 RE = NewRetainReleaseRRI.ReverseInsertPts.end();
2427 Instruction *RIP = *RI;
2428 if (ReleasesToMove.ReverseInsertPts.insert(RIP))
2429 NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
2431 NewReleases.push_back(NewRetainRelease);
2436 if (NewReleases.empty()) break;
2438 // Back the other way.
2439 for (SmallVectorImpl<Instruction *>::const_iterator
2440 NI = NewReleases.begin(), NE = NewReleases.end(); NI != NE; ++NI) {
2441 Instruction *NewRelease = *NI;
2442 DenseMap<Value *, RRInfo>::const_iterator It =
2443 Releases.find(NewRelease);
2444 assert(It != Releases.end());
2445 const RRInfo &NewReleaseRRI = It->second;
2446 KnownSafeBU &= NewReleaseRRI.KnownSafe;
2447 for (SmallPtrSet<Instruction *, 2>::const_iterator
2448 LI = NewReleaseRRI.Calls.begin(),
2449 LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
2450 Instruction *NewReleaseRetain = *LI;
2451 MapVector<Value *, RRInfo>::const_iterator Jt =
2452 Retains.find(NewReleaseRetain);
2453 if (Jt == Retains.end())
2455 const RRInfo &NewReleaseRetainRRI = Jt->second;
2456 assert(NewReleaseRetainRRI.Calls.count(NewRelease));
2457 if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
2458 unsigned PathCount =
2459 BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
2460 OldDelta += PathCount;
2461 OldCount += PathCount;
2463 // Merge the IsRetainBlock values.
2465 RetainsToMove.IsRetainBlock = NewReleaseRetainRRI.IsRetainBlock;
2466 FirstRetain = false;
2467 } else if (ReleasesToMove.IsRetainBlock !=
2468 NewReleaseRetainRRI.IsRetainBlock)
2469 // It's not possible to merge the sequences if one uses
2470 // objc_retain and the other uses objc_retainBlock.
2473 // Collect the optimal insertion points.
2475 for (SmallPtrSet<Instruction *, 2>::const_iterator
2476 RI = NewReleaseRetainRRI.ReverseInsertPts.begin(),
2477 RE = NewReleaseRetainRRI.ReverseInsertPts.end();
2479 Instruction *RIP = *RI;
2480 if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
2481 PathCount = BBStates[RIP->getParent()].GetAllPathCount();
2482 NewDelta += PathCount;
2483 NewCount += PathCount;
2486 NewRetains.push_back(NewReleaseRetain);
2490 NewReleases.clear();
2491 if (NewRetains.empty()) break;
2494 // If the pointer is known incremented or nested, we can safely delete the
2495 // pair regardless of what's between them.
2496 if (KnownSafeTD || KnownSafeBU) {
2497 RetainsToMove.ReverseInsertPts.clear();
2498 ReleasesToMove.ReverseInsertPts.clear();
2501 // Determine whether the new insertion points we computed preserve the
2502 // balance of retain and release calls through the program.
2503 // TODO: If the fully aggressive solution isn't valid, try to find a
2504 // less aggressive solution which is.
2509 // Determine whether the original call points are balanced in the retain and
2510 // release calls through the program. If not, conservatively don't touch
2512 // TODO: It's theoretically possible to do code motion in this case, as
2513 // long as the existing imbalances are maintained.
2518 assert(OldCount != 0 && "Unreachable code?");
2519 NumRRs += OldCount - NewCount;
2520 // Set to true if we completely removed any RR pairs.
2521 AnyPairsCompletelyEliminated = NewCount == 0;
2523 // We can move calls!
2527 /// Identify pairings between the retains and releases, and delete and/or move
2530 ObjCARCOpt::PerformCodePlacement(DenseMap<const BasicBlock *, BBState>
2532 MapVector<Value *, RRInfo> &Retains,
2533 DenseMap<Value *, RRInfo> &Releases,
2535 bool AnyPairsCompletelyEliminated = false;
2536 RRInfo RetainsToMove;
2537 RRInfo ReleasesToMove;
2538 SmallVector<Instruction *, 4> NewRetains;
2539 SmallVector<Instruction *, 4> NewReleases;
2540 SmallVector<Instruction *, 8> DeadInsts;
2542 // Visit each retain.
2543 for (MapVector<Value *, RRInfo>::const_iterator I = Retains.begin(),
2544 E = Retains.end(); I != E; ++I) {
2545 Value *V = I->first;
2546 if (!V) continue; // blotted
2548 Instruction *Retain = cast<Instruction>(V);
2550 DEBUG(dbgs() << "ObjCARCOpt::PerformCodePlacement: Visiting: " << *Retain
2553 Value *Arg = GetObjCArg(Retain);
2555 // If the object being released is in static or stack storage, we know it's
2556 // not being managed by ObjC reference counting, so we can delete pairs
2557 // regardless of what possible decrements or uses lie between them.
2558 bool KnownSafe = isa<Constant>(Arg) || isa<AllocaInst>(Arg);
2560 // A constant pointer can't be pointing to an object on the heap. It may
2561 // be reference-counted, but it won't be deleted.
2562 if (const LoadInst *LI = dyn_cast<LoadInst>(Arg))
2563 if (const GlobalVariable *GV =
2564 dyn_cast<GlobalVariable>(
2565 StripPointerCastsAndObjCCalls(LI->getPointerOperand())))
2566 if (GV->isConstant())
2569 // Connect the dots between the top-down-collected RetainsToMove and
2570 // bottom-up-collected ReleasesToMove to form sets of related calls.
2571 NewRetains.push_back(Retain);
2572 bool PerformMoveCalls =
2573 ConnectTDBUTraversals(BBStates, Retains, Releases, M, NewRetains,
2574 NewReleases, DeadInsts, RetainsToMove,
2575 ReleasesToMove, Arg, KnownSafe,
2576 AnyPairsCompletelyEliminated);
2578 #ifdef ARC_ANNOTATIONS
2579 // Do not move calls if ARC annotations are requested. If we were to move
2580 // calls in this case, we would not be able
2581 PerformMoveCalls = PerformMoveCalls && !EnableARCAnnotations;
2582 #endif // ARC_ANNOTATIONS
2584 if (PerformMoveCalls) {
2585 // Ok, everything checks out and we're all set. Let's move/delete some
2587 MoveCalls(Arg, RetainsToMove, ReleasesToMove,
2588 Retains, Releases, DeadInsts, M);
2591 // Clean up state for next retain.
2592 NewReleases.clear();
2594 RetainsToMove.clear();
2595 ReleasesToMove.clear();
2598 // Now that we're done moving everything, we can delete the newly dead
2599 // instructions, as we no longer need them as insert points.
2600 while (!DeadInsts.empty())
2601 EraseInstruction(DeadInsts.pop_back_val());
2603 return AnyPairsCompletelyEliminated;
2606 /// Weak pointer optimizations.
2607 void ObjCARCOpt::OptimizeWeakCalls(Function &F) {
2608 // First, do memdep-style RLE and S2L optimizations. We can't use memdep
2609 // itself because it uses AliasAnalysis and we need to do provenance
2611 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2612 Instruction *Inst = &*I++;
2614 DEBUG(dbgs() << "ObjCARCOpt::OptimizeWeakCalls: Visiting: " << *Inst <<
2617 InstructionClass Class = GetBasicInstructionClass(Inst);
2618 if (Class != IC_LoadWeak && Class != IC_LoadWeakRetained)
2621 // Delete objc_loadWeak calls with no users.
2622 if (Class == IC_LoadWeak && Inst->use_empty()) {
2623 Inst->eraseFromParent();
2627 // TODO: For now, just look for an earlier available version of this value
2628 // within the same block. Theoretically, we could do memdep-style non-local
2629 // analysis too, but that would want caching. A better approach would be to
2630 // use the technique that EarlyCSE uses.
2631 inst_iterator Current = llvm::prior(I);
2632 BasicBlock *CurrentBB = Current.getBasicBlockIterator();
2633 for (BasicBlock::iterator B = CurrentBB->begin(),
2634 J = Current.getInstructionIterator();
2636 Instruction *EarlierInst = &*llvm::prior(J);
2637 InstructionClass EarlierClass = GetInstructionClass(EarlierInst);
2638 switch (EarlierClass) {
2640 case IC_LoadWeakRetained: {
2641 // If this is loading from the same pointer, replace this load's value
2643 CallInst *Call = cast<CallInst>(Inst);
2644 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2645 Value *Arg = Call->getArgOperand(0);
2646 Value *EarlierArg = EarlierCall->getArgOperand(0);
2647 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2648 case AliasAnalysis::MustAlias:
2650 // If the load has a builtin retain, insert a plain retain for it.
2651 if (Class == IC_LoadWeakRetained) {
2653 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2657 // Zap the fully redundant load.
2658 Call->replaceAllUsesWith(EarlierCall);
2659 Call->eraseFromParent();
2661 case AliasAnalysis::MayAlias:
2662 case AliasAnalysis::PartialAlias:
2664 case AliasAnalysis::NoAlias:
2671 // If this is storing to the same pointer and has the same size etc.
2672 // replace this load's value with the stored value.
2673 CallInst *Call = cast<CallInst>(Inst);
2674 CallInst *EarlierCall = cast<CallInst>(EarlierInst);
2675 Value *Arg = Call->getArgOperand(0);
2676 Value *EarlierArg = EarlierCall->getArgOperand(0);
2677 switch (PA.getAA()->alias(Arg, EarlierArg)) {
2678 case AliasAnalysis::MustAlias:
2680 // If the load has a builtin retain, insert a plain retain for it.
2681 if (Class == IC_LoadWeakRetained) {
2683 CallInst::Create(getRetainCallee(F.getParent()), EarlierCall,
2687 // Zap the fully redundant load.
2688 Call->replaceAllUsesWith(EarlierCall->getArgOperand(1));
2689 Call->eraseFromParent();
2691 case AliasAnalysis::MayAlias:
2692 case AliasAnalysis::PartialAlias:
2694 case AliasAnalysis::NoAlias:
2701 // TOOD: Grab the copied value.
2703 case IC_AutoreleasepoolPush:
2705 case IC_IntrinsicUser:
2707 // Weak pointers are only modified through the weak entry points
2708 // (and arbitrary calls, which could call the weak entry points).
2711 // Anything else could modify the weak pointer.
2718 // Then, for each destroyWeak with an alloca operand, check to see if
2719 // the alloca and all its users can be zapped.
2720 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
2721 Instruction *Inst = &*I++;
2722 InstructionClass Class = GetBasicInstructionClass(Inst);
2723 if (Class != IC_DestroyWeak)
2726 CallInst *Call = cast<CallInst>(Inst);
2727 Value *Arg = Call->getArgOperand(0);
2728 if (AllocaInst *Alloca = dyn_cast<AllocaInst>(Arg)) {
2729 for (Value::use_iterator UI = Alloca->use_begin(),
2730 UE = Alloca->use_end(); UI != UE; ++UI) {
2731 const Instruction *UserInst = cast<Instruction>(*UI);
2732 switch (GetBasicInstructionClass(UserInst)) {
2735 case IC_DestroyWeak:
2742 for (Value::use_iterator UI = Alloca->use_begin(),
2743 UE = Alloca->use_end(); UI != UE; ) {
2744 CallInst *UserInst = cast<CallInst>(*UI++);
2745 switch (GetBasicInstructionClass(UserInst)) {
2748 // These functions return their second argument.
2749 UserInst->replaceAllUsesWith(UserInst->getArgOperand(1));
2751 case IC_DestroyWeak:
2755 llvm_unreachable("alloca really is used!");
2757 UserInst->eraseFromParent();
2759 Alloca->eraseFromParent();
2764 DEBUG(dbgs() << "ObjCARCOpt::OptimizeWeakCalls: Finished List.\n\n");
2768 /// Identify program paths which execute sequences of retains and releases which
2769 /// can be eliminated.
2770 bool ObjCARCOpt::OptimizeSequences(Function &F) {
2771 /// Releases, Retains - These are used to store the results of the main flow
2772 /// analysis. These use Value* as the key instead of Instruction* so that the
2773 /// map stays valid when we get around to rewriting code and calls get
2774 /// replaced by arguments.
2775 DenseMap<Value *, RRInfo> Releases;
2776 MapVector<Value *, RRInfo> Retains;
2778 /// This is used during the traversal of the function to track the
2779 /// states for each identified object at each block.
2780 DenseMap<const BasicBlock *, BBState> BBStates;
2782 // Analyze the CFG of the function, and all instructions.
2783 bool NestingDetected = Visit(F, BBStates, Retains, Releases);
2786 return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
2790 /// Look for this pattern:
2792 /// %call = call i8* @something(...)
2793 /// %2 = call i8* @objc_retain(i8* %call)
2794 /// %3 = call i8* @objc_autorelease(i8* %2)
2797 /// And delete the retain and autorelease.
2799 /// Otherwise if it's just this:
2801 /// %3 = call i8* @objc_autorelease(i8* %2)
2804 /// convert the autorelease to autoreleaseRV.
2805 void ObjCARCOpt::OptimizeReturns(Function &F) {
2806 if (!F.getReturnType()->isPointerTy())
2809 SmallPtrSet<Instruction *, 4> DependingInstructions;
2810 SmallPtrSet<const BasicBlock *, 4> Visited;
2811 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
2812 BasicBlock *BB = FI;
2813 ReturnInst *Ret = dyn_cast<ReturnInst>(&BB->back());
2815 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Visiting: " << *Ret << "\n");
2819 const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
2820 FindDependencies(NeedsPositiveRetainCount, Arg,
2821 BB, Ret, DependingInstructions, Visited, PA);
2822 if (DependingInstructions.size() != 1)
2826 CallInst *Autorelease =
2827 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
2830 InstructionClass AutoreleaseClass = GetBasicInstructionClass(Autorelease);
2831 if (!IsAutorelease(AutoreleaseClass))
2833 if (GetObjCArg(Autorelease) != Arg)
2836 DependingInstructions.clear();
2839 // Check that there is nothing that can affect the reference
2840 // count between the autorelease and the retain.
2841 FindDependencies(CanChangeRetainCount, Arg,
2842 BB, Autorelease, DependingInstructions, Visited, PA);
2843 if (DependingInstructions.size() != 1)
2848 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
2850 // Check that we found a retain with the same argument.
2852 !IsRetain(GetBasicInstructionClass(Retain)) ||
2853 GetObjCArg(Retain) != Arg)
2856 DependingInstructions.clear();
2859 // Convert the autorelease to an autoreleaseRV, since it's
2860 // returning the value.
2861 if (AutoreleaseClass == IC_Autorelease) {
2862 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Converting autorelease "
2863 "=> autoreleaseRV since it's returning a value.\n"
2864 " In: " << *Autorelease
2866 Autorelease->setCalledFunction(getAutoreleaseRVCallee(F.getParent()));
2867 DEBUG(dbgs() << " Out: " << *Autorelease
2869 Autorelease->setTailCall(); // Always tail call autoreleaseRV.
2870 AutoreleaseClass = IC_AutoreleaseRV;
2873 // Check that there is nothing that can affect the reference
2874 // count between the retain and the call.
2875 // Note that Retain need not be in BB.
2876 FindDependencies(CanChangeRetainCount, Arg, Retain->getParent(), Retain,
2877 DependingInstructions, Visited, PA);
2878 if (DependingInstructions.size() != 1)
2883 dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
2885 // Check that the pointer is the return value of the call.
2886 if (!Call || Arg != Call)
2889 // Check that the call is a regular call.
2890 InstructionClass Class = GetBasicInstructionClass(Call);
2891 if (Class != IC_CallOrUser && Class != IC_Call)
2894 // If so, we can zap the retain and autorelease.
2897 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Erasing: " << *Retain
2899 << *Autorelease << "\n");
2900 EraseInstruction(Retain);
2901 EraseInstruction(Autorelease);
2907 DependingInstructions.clear();
2911 DEBUG(dbgs() << "ObjCARCOpt::OptimizeReturns: Finished List.\n\n");
2915 bool ObjCARCOpt::doInitialization(Module &M) {
2919 // If nothing in the Module uses ARC, don't do anything.
2920 Run = ModuleHasARC(M);
2924 // Identify the imprecise release metadata kind.
2925 ImpreciseReleaseMDKind =
2926 M.getContext().getMDKindID("clang.imprecise_release");
2927 CopyOnEscapeMDKind =
2928 M.getContext().getMDKindID("clang.arc.copy_on_escape");
2929 NoObjCARCExceptionsMDKind =
2930 M.getContext().getMDKindID("clang.arc.no_objc_arc_exceptions");
2931 #ifdef ARC_ANNOTATIONS
2932 ARCAnnotationBottomUpMDKind =
2933 M.getContext().getMDKindID("llvm.arc.annotation.bottomup");
2934 ARCAnnotationTopDownMDKind =
2935 M.getContext().getMDKindID("llvm.arc.annotation.topdown");
2936 ARCAnnotationProvenanceSourceMDKind =
2937 M.getContext().getMDKindID("llvm.arc.annotation.provenancesource");
2938 #endif // ARC_ANNOTATIONS
2940 // Intuitively, objc_retain and others are nocapture, however in practice
2941 // they are not, because they return their argument value. And objc_release
2942 // calls finalizers which can have arbitrary side effects.
2944 // These are initialized lazily.
2946 AutoreleaseRVCallee = 0;
2949 RetainBlockCallee = 0;
2950 AutoreleaseCallee = 0;
2955 bool ObjCARCOpt::runOnFunction(Function &F) {
2959 // If nothing in the Module uses ARC, don't do anything.
2965 DEBUG(dbgs() << "ObjCARCOpt: Visiting Function: " << F.getName() << "\n");
2967 PA.setAA(&getAnalysis<AliasAnalysis>());
2969 // This pass performs several distinct transformations. As a compile-time aid
2970 // when compiling code that isn't ObjC, skip these if the relevant ObjC
2971 // library functions aren't declared.
2973 // Preliminary optimizations. This also computs UsedInThisFunction.
2974 OptimizeIndividualCalls(F);
2976 // Optimizations for weak pointers.
2977 if (UsedInThisFunction & ((1 << IC_LoadWeak) |
2978 (1 << IC_LoadWeakRetained) |
2979 (1 << IC_StoreWeak) |
2980 (1 << IC_InitWeak) |
2981 (1 << IC_CopyWeak) |
2982 (1 << IC_MoveWeak) |
2983 (1 << IC_DestroyWeak)))
2984 OptimizeWeakCalls(F);
2986 // Optimizations for retain+release pairs.
2987 if (UsedInThisFunction & ((1 << IC_Retain) |
2988 (1 << IC_RetainRV) |
2989 (1 << IC_RetainBlock)))
2990 if (UsedInThisFunction & (1 << IC_Release))
2991 // Run OptimizeSequences until it either stops making changes or
2992 // no retain+release pair nesting is detected.
2993 while (OptimizeSequences(F)) {}
2995 // Optimizations if objc_autorelease is used.
2996 if (UsedInThisFunction & ((1 << IC_Autorelease) |
2997 (1 << IC_AutoreleaseRV)))
3000 DEBUG(dbgs() << "\n");
3005 void ObjCARCOpt::releaseMemory() {