#include "ObjCARCAliasAnalysis.h"
#include "ProvenanceAnalysis.h"
#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
return std::make_pair(Vector.begin() + Pair.first->second, false);
}
+ iterator find(const KeyT &Key) {
+ typename MapTy::iterator It = Map.find(Key);
+ if (It == Map.end()) return Vector.end();
+ return Vector.begin() + It->second;
+ }
+
const_iterator find(const KeyT &Key) const {
typename MapTy::const_iterator It = Map.find(Key);
if (It == Map.end()) return Vector.end();
return false;
}
+/// This is a wrapper around getUnderlyingObjCPtr along the lines of
+/// GetUnderlyingObjects except that it returns early when it sees the first
+/// alloca.
+static inline bool AreAnyUnderlyingObjectsAnAlloca(const Value *V) {
+ SmallPtrSet<const Value *, 4> Visited;
+ SmallVector<const Value *, 4> Worklist;
+ Worklist.push_back(V);
+ do {
+ const Value *P = Worklist.pop_back_val();
+ P = GetUnderlyingObjCPtr(P);
+
+ if (isa<AllocaInst>(P))
+ return true;
+
+ if (!Visited.insert(P))
+ continue;
+
+ if (const SelectInst *SI = dyn_cast<const SelectInst>(P)) {
+ Worklist.push_back(SI->getTrueValue());
+ Worklist.push_back(SI->getFalseValue());
+ continue;
+ }
+
+ if (const PHINode *PN = dyn_cast<const PHINode>(P)) {
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+ Worklist.push_back(PN->getIncomingValue(i));
+ continue;
+ }
+ } while (!Worklist.empty());
+
+ return false;
+}
+
+
/// @}
///
/// \defgroup ARCOpt ARC Optimization.
STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated");
STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases");
STATISTIC(NumRets, "Number of return value forwarding "
- "retain+autoreleaes eliminated");
+ "retain+autoreleases eliminated");
STATISTIC(NumRRs, "Number of retain+release paths eliminated");
STATISTIC(NumPeeps, "Number of calls peephole-optimized");
+#ifndef NDEBUG
+STATISTIC(NumRetainsBeforeOpt,
+ "Number of retains before optimization");
+STATISTIC(NumReleasesBeforeOpt,
+ "Number of releases before optimization");
+STATISTIC(NumRetainsAfterOpt,
+ "Number of retains after optimization");
+STATISTIC(NumReleasesAfterOpt,
+ "Number of releases after optimization");
+#endif
namespace {
/// \enum Sequence
/// sequence.
SmallPtrSet<Instruction *, 2> ReverseInsertPts;
+ /// If this is true, we cannot perform code motion but can still remove
+ /// retain/release pairs.
+ bool CFGHazardAfflicted;
+
RRInfo() :
- KnownSafe(false), IsTailCallRelease(false), ReleaseMetadata(0) {}
+ KnownSafe(false), IsTailCallRelease(false), ReleaseMetadata(0),
+ CFGHazardAfflicted(false) {}
void clear();
- bool IsTrackingImpreciseReleases() {
- return ReleaseMetadata != 0;
- }
+ /// Conservatively merge the two RRInfo. Returns true if a partial merge has
+ /// occured, false otherwise.
+ bool Merge(const RRInfo &Other);
+
};
}
ReleaseMetadata = 0;
Calls.clear();
ReverseInsertPts.clear();
+ CFGHazardAfflicted = false;
+}
+
+bool RRInfo::Merge(const RRInfo &Other) {
+ // Conservatively merge the ReleaseMetadata information.
+ if (ReleaseMetadata != Other.ReleaseMetadata)
+ ReleaseMetadata = 0;
+
+ // Conservatively merge the boolean state.
+ KnownSafe &= Other.KnownSafe;
+ IsTailCallRelease &= Other.IsTailCallRelease;
+ CFGHazardAfflicted |= Other.CFGHazardAfflicted;
+
+ // Merge the call sets.
+ Calls.insert(Other.Calls.begin(), Other.Calls.end());
+
+ // Merge the insert point sets. If there are any differences,
+ // that makes this a partial merge.
+ bool Partial = ReverseInsertPts.size() != Other.ReverseInsertPts.size();
+ for (SmallPtrSet<Instruction *, 2>::const_iterator
+ I = Other.ReverseInsertPts.begin(),
+ E = Other.ReverseInsertPts.end(); I != E; ++I)
+ Partial |= ReverseInsertPts.insert(*I);
+ return Partial;
}
namespace {
/// The current position in the sequence.
Sequence Seq : 8;
- public:
/// Unidirectional information about the current sequence.
- ///
- /// TODO: Encapsulate this better.
RRInfo RRI;
+ public:
PtrState() : KnownPositiveRefCount(false), Partial(false),
Seq(S_None) {}
+
+ bool IsKnownSafe() const {
+ return RRI.KnownSafe;
+ }
+
+ void SetKnownSafe(const bool NewValue) {
+ RRI.KnownSafe = NewValue;
+ }
+
+ bool IsTailCallRelease() const {
+ return RRI.IsTailCallRelease;
+ }
+
+ void SetTailCallRelease(const bool NewValue) {
+ RRI.IsTailCallRelease = NewValue;
+ }
+
+ bool IsTrackingImpreciseReleases() const {
+ return RRI.ReleaseMetadata != 0;
+ }
+
+ const MDNode *GetReleaseMetadata() const {
+ return RRI.ReleaseMetadata;
+ }
+
+ void SetReleaseMetadata(MDNode *NewValue) {
+ RRI.ReleaseMetadata = NewValue;
+ }
+
+ bool IsCFGHazardAfflicted() const {
+ return RRI.CFGHazardAfflicted;
+ }
+
+ void SetCFGHazardAfflicted(const bool NewValue) {
+ RRI.CFGHazardAfflicted = NewValue;
+ }
+
void SetKnownPositiveRefCount() {
+ DEBUG(dbgs() << "Setting Known Positive.\n");
KnownPositiveRefCount = true;
}
void ClearKnownPositiveRefCount() {
+ DEBUG(dbgs() << "Clearing Known Positive.\n");
KnownPositiveRefCount = false;
}
}
void ResetSequenceProgress(Sequence NewSeq) {
+ DEBUG(dbgs() << "Resetting sequence progress.\n");
SetSeq(NewSeq);
Partial = false;
RRI.clear();
}
void Merge(const PtrState &Other, bool TopDown);
+
+ void InsertCall(Instruction *I) {
+ RRI.Calls.insert(I);
+ }
+
+ void InsertReverseInsertPt(Instruction *I) {
+ RRI.ReverseInsertPts.insert(I);
+ }
+
+ void ClearReverseInsertPts() {
+ RRI.ReverseInsertPts.clear();
+ }
+
+ bool HasReverseInsertPts() const {
+ return !RRI.ReverseInsertPts.empty();
+ }
+
+ const RRInfo &GetRRInfo() const {
+ return RRI;
+ }
};
}
void
PtrState::Merge(const PtrState &Other, bool TopDown) {
Seq = MergeSeqs(Seq, Other.Seq, TopDown);
- KnownPositiveRefCount = KnownPositiveRefCount && Other.KnownPositiveRefCount;
+ KnownPositiveRefCount &= Other.KnownPositiveRefCount;
// If we're not in a sequence (anymore), drop all associated state.
if (Seq == S_None) {
// mixing them is unsafe.
ClearSequenceProgress();
} else {
- // Conservatively merge the ReleaseMetadata information.
- if (RRI.ReleaseMetadata != Other.RRI.ReleaseMetadata)
- RRI.ReleaseMetadata = 0;
-
- RRI.KnownSafe = RRI.KnownSafe && Other.RRI.KnownSafe;
- RRI.IsTailCallRelease = RRI.IsTailCallRelease &&
- Other.RRI.IsTailCallRelease;
- RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end());
-
- // Merge the insert point sets. If there are any differences,
- // that makes this a partial merge.
- Partial = RRI.ReverseInsertPts.size() != Other.RRI.ReverseInsertPts.size();
- for (SmallPtrSet<Instruction *, 2>::const_iterator
- I = Other.RRI.ReverseInsertPts.begin(),
- E = Other.RRI.ReverseInsertPts.end(); I != E; ++I)
- Partial |= RRI.ReverseInsertPts.insert(*I);
+ // Otherwise merge the other PtrState's RRInfo into our RRInfo. At this
+ // point, we know that currently we are not partial. Stash whether or not
+ // the merge operation caused us to undergo a partial merging of reverse
+ // insertion points.
+ Partial = RRI.Merge(Other.RRI);
}
}
/// definition.
void SetAsExit() { BottomUpPathCount = 1; }
+ /// Attempt to find the PtrState object describing the top down state for
+ /// pointer Arg. Return a new initialized PtrState describing the top down
+ /// state for Arg if we do not find one.
PtrState &getPtrTopDownState(const Value *Arg) {
return PerPtrTopDown[Arg];
}
+ /// Attempt to find the PtrState object describing the bottom up state for
+ /// pointer Arg. Return a new initialized PtrState describing the bottom up
+ /// state for Arg if we do not find one.
PtrState &getPtrBottomUpState(const Value *Arg) {
return PerPtrBottomUp[Arg];
}
+ /// Attempt to find the PtrState object describing the bottom up state for
+ /// pointer Arg.
+ ptr_iterator findPtrBottomUpState(const Value *Arg) {
+ return PerPtrBottomUp.find(Arg);
+ }
+
void clearBottomUpPointers() {
PerPtrBottomUp.clear();
}
void MergePred(const BBState &Other);
void MergeSucc(const BBState &Other);
- /// Return the number of possible unique paths from an entry to an exit
+ /// Compute the number of possible unique paths from an entry to an exit
/// which pass through this block. This is only valid after both the
/// top-down and bottom-up traversals are complete.
- unsigned GetAllPathCount() const {
+ ///
+ /// Returns true if overflow occured. Returns false if overflow did not
+ /// occur.
+ bool GetAllPathCountWithOverflow(unsigned &PathCount) const {
assert(TopDownPathCount != 0);
assert(BottomUpPathCount != 0);
- return TopDownPathCount * BottomUpPathCount;
+ unsigned long long Product =
+ (unsigned long long)TopDownPathCount*BottomUpPathCount;
+ PathCount = Product;
+ // Overflow occured if any of the upper bits of Product are set.
+ return Product >> 32;
}
// Specialized CFG utilities.
bool Changed;
ProvenanceAnalysis PA;
+ // This is used to track if a pointer is stored into an alloca.
+ DenseSet<const Value *> MultiOwnersSet;
+
/// A flag indicating whether this optimization pass should run.
bool Run;
/// them. These are initialized lazily to avoid cluttering up the Module
/// with unused declarations.
- /// Declaration for ObjC runtime function
- /// objc_retainAutoreleasedReturnValue.
- Constant *RetainRVCallee;
/// Declaration for ObjC runtime function objc_autoreleaseReturnValue.
Constant *AutoreleaseRVCallee;
/// Declaration for ObjC runtime function objc_release.
unsigned ARCAnnotationProvenanceSourceMDKind;
#endif // ARC_ANNOATIONS
- Constant *getRetainRVCallee(Module *M);
Constant *getAutoreleaseRVCallee(Module *M);
Constant *getReleaseCallee(Module *M);
Constant *getRetainCallee(Module *M);
bool IsRetainBlockOptimizable(const Instruction *Inst);
- void OptimizeRetainCall(Function &F, Instruction *Retain);
bool OptimizeRetainRVCall(Function &F, Instruction *RetainRV);
void OptimizeAutoreleaseRVCall(Function &F, Instruction *AutoreleaseRV,
InstructionClass &Class);
void OptimizeReturns(Function &F);
+#ifndef NDEBUG
+ void GatherStatistics(Function &F, bool AfterOptimization = false);
+#endif
+
virtual void getAnalysisUsage(AnalysisUsage &AU) const;
virtual bool doInitialization(Module &M);
virtual bool runOnFunction(Function &F);
return true;
}
-Constant *ObjCARCOpt::getRetainRVCallee(Module *M) {
- if (!RetainRVCallee) {
- LLVMContext &C = M->getContext();
- Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
- Type *Params[] = { I8X };
- FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false);
- AttributeSet Attribute =
- AttributeSet().addAttribute(M->getContext(), AttributeSet::FunctionIndex,
- Attribute::NoUnwind);
- RetainRVCallee =
- M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy,
- Attribute);
- }
- return RetainRVCallee;
-}
-
Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) {
if (!AutoreleaseRVCallee) {
LLVMContext &C = M->getContext();
return AutoreleaseCallee;
}
-/// Turn objc_retain into objc_retainAutoreleasedReturnValue if the operand is a
-/// return value.
-void
-ObjCARCOpt::OptimizeRetainCall(Function &F, Instruction *Retain) {
- ImmutableCallSite CS(GetObjCArg(Retain));
- const Instruction *Call = CS.getInstruction();
- if (!Call) return;
- if (Call->getParent() != Retain->getParent()) return;
-
- // Check that the call is next to the retain.
- BasicBlock::const_iterator I = Call;
- ++I;
- while (IsNoopInstruction(I)) ++I;
- if (&*I != Retain)
- return;
-
- // Turn it to an objc_retainAutoreleasedReturnValue..
- Changed = true;
- ++NumPeeps;
-
- DEBUG(dbgs() << "Transforming objc_retain => "
- "objc_retainAutoreleasedReturnValue since the operand is a "
- "return value.\nOld: "<< *Retain << "\n");
-
- cast<CallInst>(Retain)->setCalledFunction(getRetainRVCallee(F.getParent()));
-
- DEBUG(dbgs() << "New: " << *Retain << "\n");
-}
-
/// Turn objc_retainAutoreleasedReturnValue into objc_retain if the operand is
/// not a return value. Or, if it can be paired with an
/// objc_autoreleaseReturnValue, delete the pair and return true.
if (!IsRetainBlockOptimizable(Inst))
return false;
+ Changed = true;
+ ++NumPeeps;
+
+ DEBUG(dbgs() << "Strength reduced retainBlock => retain.\n");
+ DEBUG(dbgs() << "Old: " << *Inst << "\n");
CallInst *RetainBlock = cast<CallInst>(Inst);
RetainBlock->setCalledFunction(getRetainCallee(F.getParent()));
// Remove copy_on_escape metadata.
RetainBlock->setMetadata(CopyOnEscapeMDKind, 0);
Class = IC_Retain;
-
+ DEBUG(dbgs() << "New: " << *Inst << "\n");
return true;
}
break;
}
case IC_RetainBlock:
- // If we strength reduce an objc_retainBlock to amn objc_retain, continue
+ // If we strength reduce an objc_retainBlock to an objc_retain, continue
// onto the objc_retain peephole optimizations. Otherwise break.
- if (!OptimizeRetainBlockCall(F, Inst, Class))
- break;
- // FALLTHROUGH
- case IC_Retain:
- OptimizeRetainCall(F, Inst);
+ OptimizeRetainBlockCall(F, Inst, Class);
break;
case IC_RetainRV:
if (OptimizeRetainRVCall(F, Inst))
CallInst *NewCall =
CallInst::Create(getReleaseCallee(F.getParent()),
Call->getArgOperand(0), "", Call);
- NewCall->setMetadata(ImpreciseReleaseMDKind,
- MDNode::get(C, ArrayRef<Value *>()));
+ NewCall->setMetadata(ImpreciseReleaseMDKind, MDNode::get(C, None));
DEBUG(dbgs() << "Replacing autorelease{,RV}(x) with objc_release(x) "
"since x is otherwise unused.\nOld: " << *Call << "\nNew: "
}
}
+/// If we have a top down pointer in the S_Use state, make sure that there are
+/// no CFG hazards by checking the states of various bottom up pointers.
+static void CheckForUseCFGHazard(const Sequence SuccSSeq,
+ const bool SuccSRRIKnownSafe,
+ PtrState &S,
+ bool &SomeSuccHasSame,
+ bool &AllSuccsHaveSame,
+ bool &NotAllSeqEqualButKnownSafe,
+ bool &ShouldContinue) {
+ switch (SuccSSeq) {
+ case S_CanRelease: {
+ if (!S.IsKnownSafe() && !SuccSRRIKnownSafe) {
+ S.ClearSequenceProgress();
+ break;
+ }
+ S.SetCFGHazardAfflicted(true);
+ ShouldContinue = true;
+ break;
+ }
+ case S_Use:
+ SomeSuccHasSame = true;
+ break;
+ case S_Stop:
+ case S_Release:
+ case S_MovableRelease:
+ if (!S.IsKnownSafe() && !SuccSRRIKnownSafe)
+ AllSuccsHaveSame = false;
+ else
+ NotAllSeqEqualButKnownSafe = true;
+ break;
+ case S_Retain:
+ llvm_unreachable("bottom-up pointer in retain state!");
+ case S_None:
+ llvm_unreachable("This should have been handled earlier.");
+ }
+}
+
+/// If we have a Top Down pointer in the S_CanRelease state, make sure that
+/// there are no CFG hazards by checking the states of various bottom up
+/// pointers.
+static void CheckForCanReleaseCFGHazard(const Sequence SuccSSeq,
+ const bool SuccSRRIKnownSafe,
+ PtrState &S,
+ bool &SomeSuccHasSame,
+ bool &AllSuccsHaveSame,
+ bool &NotAllSeqEqualButKnownSafe) {
+ switch (SuccSSeq) {
+ case S_CanRelease:
+ SomeSuccHasSame = true;
+ break;
+ case S_Stop:
+ case S_Release:
+ case S_MovableRelease:
+ case S_Use:
+ if (!S.IsKnownSafe() && !SuccSRRIKnownSafe)
+ AllSuccsHaveSame = false;
+ else
+ NotAllSeqEqualButKnownSafe = true;
+ break;
+ case S_Retain:
+ llvm_unreachable("bottom-up pointer in retain state!");
+ case S_None:
+ llvm_unreachable("This should have been handled earlier.");
+ }
+}
+
/// Check for critical edges, loop boundaries, irreducible control flow, or
/// other CFG structures where moving code across the edge would result in it
/// being executed more.
// If any top-down local-use or possible-dec has a succ which is earlier in
// the sequence, forget it.
for (BBState::ptr_iterator I = MyStates.top_down_ptr_begin(),
- E = MyStates.top_down_ptr_end(); I != E; ++I)
- switch (I->second.GetSeq()) {
- default: break;
- case S_Use: {
- const Value *Arg = I->first;
- const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
- bool SomeSuccHasSame = false;
- bool AllSuccsHaveSame = true;
- PtrState &S = I->second;
- succ_const_iterator SI(TI), SE(TI, false);
-
- for (; SI != SE; ++SI) {
- Sequence SuccSSeq = S_None;
- bool SuccSRRIKnownSafe = false;
- // If VisitBottomUp has pointer information for this successor, take
- // what we know about it.
- DenseMap<const BasicBlock *, BBState>::iterator BBI =
- BBStates.find(*SI);
- assert(BBI != BBStates.end());
- const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
- SuccSSeq = SuccS.GetSeq();
- SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
- switch (SuccSSeq) {
- case S_None:
- case S_CanRelease: {
- if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
- S.ClearSequenceProgress();
- break;
- }
- continue;
- }
- case S_Use:
- SomeSuccHasSame = true;
- break;
- case S_Stop:
- case S_Release:
- case S_MovableRelease:
- if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
- AllSuccsHaveSame = false;
- break;
- case S_Retain:
- llvm_unreachable("bottom-up pointer in retain state!");
- }
- }
- // If the state at the other end of any of the successor edges
- // matches the current state, require all edges to match. This
- // guards against loops in the middle of a sequence.
- if (SomeSuccHasSame && !AllSuccsHaveSame)
+ E = MyStates.top_down_ptr_end(); I != E; ++I) {
+ PtrState &S = I->second;
+ const Sequence Seq = I->second.GetSeq();
+
+ // We only care about S_Retain, S_CanRelease, and S_Use.
+ if (Seq == S_None)
+ continue;
+
+ // Make sure that if extra top down states are added in the future that this
+ // code is updated to handle it.
+ assert((Seq == S_Retain || Seq == S_CanRelease || Seq == S_Use) &&
+ "Unknown top down sequence state.");
+
+ const Value *Arg = I->first;
+ const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
+ bool SomeSuccHasSame = false;
+ bool AllSuccsHaveSame = true;
+ bool NotAllSeqEqualButKnownSafe = false;
+
+ succ_const_iterator SI(TI), SE(TI, false);
+
+ for (; SI != SE; ++SI) {
+ // If VisitBottomUp has pointer information for this successor, take
+ // what we know about it.
+ const DenseMap<const BasicBlock *, BBState>::iterator BBI =
+ BBStates.find(*SI);
+ assert(BBI != BBStates.end());
+ const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
+ const Sequence SuccSSeq = SuccS.GetSeq();
+
+ // If bottom up, the pointer is in an S_None state, clear the sequence
+ // progress since the sequence in the bottom up state finished
+ // suggesting a mismatch in between retains/releases. This is true for
+ // all three cases that we are handling here: S_Retain, S_Use, and
+ // S_CanRelease.
+ if (SuccSSeq == S_None) {
S.ClearSequenceProgress();
- break;
- }
- case S_CanRelease: {
- const Value *Arg = I->first;
- const TerminatorInst *TI = cast<TerminatorInst>(&BB->back());
- bool SomeSuccHasSame = false;
- bool AllSuccsHaveSame = true;
- PtrState &S = I->second;
- succ_const_iterator SI(TI), SE(TI, false);
-
- for (; SI != SE; ++SI) {
- Sequence SuccSSeq = S_None;
- bool SuccSRRIKnownSafe = false;
- // If VisitBottomUp has pointer information for this successor, take
- // what we know about it.
- DenseMap<const BasicBlock *, BBState>::iterator BBI =
- BBStates.find(*SI);
- assert(BBI != BBStates.end());
- const PtrState &SuccS = BBI->second.getPtrBottomUpState(Arg);
- SuccSSeq = SuccS.GetSeq();
- SuccSRRIKnownSafe = SuccS.RRI.KnownSafe;
- switch (SuccSSeq) {
- case S_None: {
- if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe) {
- S.ClearSequenceProgress();
- break;
- }
+ continue;
+ }
+
+ // If we have S_Use or S_CanRelease, perform our check for cfg hazard
+ // checks.
+ const bool SuccSRRIKnownSafe = SuccS.IsKnownSafe();
+
+ // *NOTE* We do not use Seq from above here since we are allowing for
+ // S.GetSeq() to change while we are visiting basic blocks.
+ switch(S.GetSeq()) {
+ case S_Use: {
+ bool ShouldContinue = false;
+ CheckForUseCFGHazard(SuccSSeq, SuccSRRIKnownSafe, S, SomeSuccHasSame,
+ AllSuccsHaveSame, NotAllSeqEqualButKnownSafe,
+ ShouldContinue);
+ if (ShouldContinue)
continue;
- }
- case S_CanRelease:
- SomeSuccHasSame = true;
- break;
- case S_Stop:
- case S_Release:
- case S_MovableRelease:
- case S_Use:
- if (!S.RRI.KnownSafe && !SuccSRRIKnownSafe)
- AllSuccsHaveSame = false;
- break;
- case S_Retain:
- llvm_unreachable("bottom-up pointer in retain state!");
- }
+ break;
+ }
+ case S_CanRelease: {
+ CheckForCanReleaseCFGHazard(SuccSSeq, SuccSRRIKnownSafe, S,
+ SomeSuccHasSame, AllSuccsHaveSame,
+ NotAllSeqEqualButKnownSafe);
+ break;
+ }
+ case S_Retain:
+ case S_None:
+ case S_Stop:
+ case S_Release:
+ case S_MovableRelease:
+ break;
}
- // If the state at the other end of any of the successor edges
- // matches the current state, require all edges to match. This
- // guards against loops in the middle of a sequence.
- if (SomeSuccHasSame && !AllSuccsHaveSame)
- S.ClearSequenceProgress();
- break;
}
+
+ // If the state at the other end of any of the successor edges
+ // matches the current state, require all edges to match. This
+ // guards against loops in the middle of a sequence.
+ if (SomeSuccHasSame && !AllSuccsHaveSame) {
+ S.ClearSequenceProgress();
+ } else if (NotAllSeqEqualButKnownSafe) {
+ // If we would have cleared the state foregoing the fact that we are known
+ // safe, stop code motion. This is because whether or not it is safe to
+ // remove RR pairs via KnownSafe is an orthogonal concept to whether we
+ // are allowed to perform code motion.
+ S.SetCFGHazardAfflicted(true);
}
+ }
}
bool
Sequence NewSeq = ReleaseMetadata ? S_MovableRelease : S_Release;
ANNOTATE_BOTTOMUP(Inst, Arg, S.GetSeq(), NewSeq);
S.ResetSequenceProgress(NewSeq);
- S.RRI.ReleaseMetadata = ReleaseMetadata;
- S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
- S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
- S.RRI.Calls.insert(Inst);
+ S.SetReleaseMetadata(ReleaseMetadata);
+ S.SetKnownSafe(S.HasKnownPositiveRefCount());
+ S.SetTailCallRelease(cast<CallInst>(Inst)->isTailCall());
+ S.InsertCall(Inst);
S.SetKnownPositiveRefCount();
break;
}
case S_Use:
// If OldSeq is not S_Use or OldSeq is S_Use and we are tracking an
// imprecise release, clear our reverse insertion points.
- if (OldSeq != S_Use || S.RRI.IsTrackingImpreciseReleases())
- S.RRI.ReverseInsertPts.clear();
+ if (OldSeq != S_Use || S.IsTrackingImpreciseReleases())
+ S.ClearReverseInsertPts();
// FALL THROUGH
case S_CanRelease:
// Don't do retain+release tracking for IC_RetainRV, because it's
// better to let it remain as the first instruction after a call.
if (Class != IC_RetainRV)
- Retains[Inst] = S.RRI;
+ Retains[Inst] = S.GetRRInfo();
S.ClearSequenceProgress();
break;
case S_None:
case IC_None:
// These are irrelevant.
return NestingDetected;
+ case IC_User:
+ // If we have a store into an alloca of a pointer we are tracking, the
+ // pointer has multiple owners implying that we must be more conservative.
+ //
+ // This comes up in the context of a pointer being ``KnownSafe''. In the
+ // presense of a block being initialized, the frontend will emit the
+ // objc_retain on the original pointer and the release on the pointer loaded
+ // from the alloca. The optimizer will through the provenance analysis
+ // realize that the two are related, but since we only require KnownSafe in
+ // one direction, will match the inner retain on the original pointer with
+ // the guard release on the original pointer. This is fixed by ensuring that
+ // in the presense of allocas we only unconditionally remove pointers if
+ // both our retain and our release are KnownSafe.
+ if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
+ if (AreAnyUnderlyingObjectsAnAlloca(SI->getPointerOperand())) {
+ BBState::ptr_iterator I = MyStates.findPtrBottomUpState(
+ StripPointerCastsAndObjCCalls(SI->getValueOperand()));
+ if (I != MyStates.bottom_up_ptr_end())
+ MultiOwnersSet.insert(I->first);
+ }
+ }
+ break;
default:
break;
}
if (CanUse(Inst, Ptr, PA, Class)) {
DEBUG(dbgs() << "CanUse: Seq: " << Seq << "; " << *Ptr
<< "\n");
- assert(S.RRI.ReverseInsertPts.empty());
+ assert(!S.HasReverseInsertPts());
// If this is an invoke instruction, we're scanning it as part of
// one of its successor blocks, since we can't insert code after it
// in its own block, and we don't want to split critical edges.
if (isa<InvokeInst>(Inst))
- S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
+ S.InsertReverseInsertPt(BB->getFirstInsertionPt());
else
- S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
+ S.InsertReverseInsertPt(llvm::next(BasicBlock::iterator(Inst)));
S.SetSeq(S_Use);
ANNOTATE_BOTTOMUP(Inst, Ptr, Seq, S_Use);
} else if (Seq == S_Release && IsUser(Class)) {
// Non-movable releases depend on any possible objc pointer use.
S.SetSeq(S_Stop);
ANNOTATE_BOTTOMUP(Inst, Ptr, S_Release, S_Stop);
- assert(S.RRI.ReverseInsertPts.empty());
+ assert(!S.HasReverseInsertPts());
// As above; handle invoke specially.
if (isa<InvokeInst>(Inst))
- S.RRI.ReverseInsertPts.insert(BB->getFirstInsertionPt());
+ S.InsertReverseInsertPt(BB->getFirstInsertionPt());
else
- S.RRI.ReverseInsertPts.insert(llvm::next(BasicBlock::iterator(Inst)));
+ S.InsertReverseInsertPt(llvm::next(BasicBlock::iterator(Inst)));
}
break;
case S_Stop:
ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_Retain);
S.ResetSequenceProgress(S_Retain);
- S.RRI.KnownSafe = S.HasKnownPositiveRefCount();
- S.RRI.Calls.insert(Inst);
+ S.SetKnownSafe(S.HasKnownPositiveRefCount());
+ S.InsertCall(Inst);
}
S.SetKnownPositiveRefCount();
case S_Retain:
case S_CanRelease:
if (OldSeq == S_Retain || ReleaseMetadata != 0)
- S.RRI.ReverseInsertPts.clear();
+ S.ClearReverseInsertPts();
// FALL THROUGH
case S_Use:
- S.RRI.ReleaseMetadata = ReleaseMetadata;
- S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall();
- Releases[Inst] = S.RRI;
+ S.SetReleaseMetadata(ReleaseMetadata);
+ S.SetTailCallRelease(cast<CallInst>(Inst)->isTailCall());
+ Releases[Inst] = S.GetRRInfo();
ANNOTATE_TOPDOWN(Inst, Arg, S.GetSeq(), S_None);
S.ClearSequenceProgress();
break;
case S_Retain:
S.SetSeq(S_CanRelease);
ANNOTATE_TOPDOWN(Inst, Ptr, Seq, S_CanRelease);
- assert(S.RRI.ReverseInsertPts.empty());
- S.RRI.ReverseInsertPts.insert(Inst);
+ assert(!S.HasReverseInsertPts());
+ S.InsertReverseInsertPt(Inst);
// One call can't cause a transition from S_Retain to S_CanRelease
// and S_CanRelease to S_Use. If we've made the first transition,
Call->setDoesNotThrow();
Call->setTailCall();
- DEBUG(dbgs() << "Inserting new Release: " << *Call << "\n"
+ DEBUG(dbgs() << "Inserting new Retain: " << *Call << "\n"
"At insertion point: " << *InsertPt << "\n");
}
for (SmallPtrSet<Instruction *, 2>::const_iterator
bool KnownSafe,
bool &AnyPairsCompletelyEliminated) {
// If a pair happens in a region where it is known that the reference count
- // is already incremented, we can similarly ignore possible decrements.
+ // is already incremented, we can similarly ignore possible decrements unless
+ // we are dealing with a retainable object with multiple provenance sources.
bool KnownSafeTD = true, KnownSafeBU = true;
+ bool MultipleOwners = false;
+ bool CFGHazardAfflicted = false;
// Connect the dots between the top-down-collected RetainsToMove and
// bottom-up-collected ReleasesToMove to form sets of related calls.
assert(It != Retains.end());
const RRInfo &NewRetainRRI = It->second;
KnownSafeTD &= NewRetainRRI.KnownSafe;
+ MultipleOwners =
+ MultipleOwners || MultiOwnersSet.count(GetObjCArg(NewRetain));
for (SmallPtrSet<Instruction *, 2>::const_iterator
LI = NewRetainRRI.Calls.begin(),
LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) {
const RRInfo &NewRetainReleaseRRI = Jt->second;
assert(NewRetainReleaseRRI.Calls.count(NewRetain));
if (ReleasesToMove.Calls.insert(NewRetainRelease)) {
- OldDelta -=
- BBStates[NewRetainRelease->getParent()].GetAllPathCount();
+
+ // If we overflow when we compute the path count, don't remove/move
+ // anything.
+ const BBState &NRRBBState = BBStates[NewRetainRelease->getParent()];
+ unsigned PathCount;
+ if (NRRBBState.GetAllPathCountWithOverflow(PathCount))
+ return false;
+ OldDelta -= PathCount;
// Merge the ReleaseMetadata and IsTailCallRelease values.
if (FirstRelease) {
RE = NewRetainReleaseRRI.ReverseInsertPts.end();
RI != RE; ++RI) {
Instruction *RIP = *RI;
- if (ReleasesToMove.ReverseInsertPts.insert(RIP))
- NewDelta -= BBStates[RIP->getParent()].GetAllPathCount();
+ if (ReleasesToMove.ReverseInsertPts.insert(RIP)) {
+ // If we overflow when we compute the path count, don't
+ // remove/move anything.
+ const BBState &RIPBBState = BBStates[RIP->getParent()];
+ if (RIPBBState.GetAllPathCountWithOverflow(PathCount))
+ return false;
+ NewDelta -= PathCount;
+ }
}
NewReleases.push_back(NewRetainRelease);
}
assert(It != Releases.end());
const RRInfo &NewReleaseRRI = It->second;
KnownSafeBU &= NewReleaseRRI.KnownSafe;
+ CFGHazardAfflicted |= NewReleaseRRI.CFGHazardAfflicted;
for (SmallPtrSet<Instruction *, 2>::const_iterator
LI = NewReleaseRRI.Calls.begin(),
LE = NewReleaseRRI.Calls.end(); LI != LE; ++LI) {
const RRInfo &NewReleaseRetainRRI = Jt->second;
assert(NewReleaseRetainRRI.Calls.count(NewRelease));
if (RetainsToMove.Calls.insert(NewReleaseRetain)) {
- unsigned PathCount =
- BBStates[NewReleaseRetain->getParent()].GetAllPathCount();
+
+ // If we overflow when we compute the path count, don't remove/move
+ // anything.
+ const BBState &NRRBBState = BBStates[NewReleaseRetain->getParent()];
+ unsigned PathCount;
+ if (NRRBBState.GetAllPathCountWithOverflow(PathCount))
+ return false;
OldDelta += PathCount;
OldCount += PathCount;
RI != RE; ++RI) {
Instruction *RIP = *RI;
if (RetainsToMove.ReverseInsertPts.insert(RIP)) {
- PathCount = BBStates[RIP->getParent()].GetAllPathCount();
+ // If we overflow when we compute the path count, don't
+ // remove/move anything.
+ const BBState &RIPBBState = BBStates[RIP->getParent()];
+ if (RIPBBState.GetAllPathCountWithOverflow(PathCount))
+ return false;
NewDelta += PathCount;
NewCount += PathCount;
}
if (NewRetains.empty()) break;
}
- // If the pointer is known incremented or nested, we can safely delete the
- // pair regardless of what's between them.
- if (KnownSafeTD || KnownSafeBU) {
+ // If the pointer is known incremented in 1 direction and we do not have
+ // MultipleOwners, we can safely remove the retain/releases. Otherwise we need
+ // to be known safe in both directions.
+ bool UnconditionallySafe = (KnownSafeTD && KnownSafeBU) ||
+ ((KnownSafeTD || KnownSafeBU) && !MultipleOwners);
+ if (UnconditionallySafe) {
RetainsToMove.ReverseInsertPts.clear();
ReleasesToMove.ReverseInsertPts.clear();
NewCount = 0;
// less aggressive solution which is.
if (NewDelta != 0)
return false;
+
+ // At this point, we are not going to remove any RR pairs, but we still are
+ // able to move RR pairs. If one of our pointers is afflicted with
+ // CFGHazards, we cannot perform such code motion so exit early.
+ const bool WillPerformCodeMotion = RetainsToMove.ReverseInsertPts.size() ||
+ ReleasesToMove.ReverseInsertPts.size();
+ if (CFGHazardAfflicted && WillPerformCodeMotion)
+ return false;
}
// Determine whether the original call points are balanced in the retain and
if (OldDelta != 0)
return false;
+#ifdef ARC_ANNOTATIONS
+ // Do not move calls if ARC annotations are requested.
+ if (EnableARCAnnotations)
+ return false;
+#endif // ARC_ANNOTATIONS
+
Changed = true;
assert(OldCount != 0 && "Unreachable code?");
NumRRs += OldCount - NewCount;
ReleasesToMove, Arg, KnownSafe,
AnyPairsCompletelyEliminated);
-#ifdef ARC_ANNOTATIONS
- // Do not move calls if ARC annotations are requested. If we were to move
- // calls in this case, we would not be able
- PerformMoveCalls = PerformMoveCalls && !EnableARCAnnotations;
-#endif // ARC_ANNOTATIONS
-
if (PerformMoveCalls) {
// Ok, everything checks out and we're all set. Let's move/delete some
// code!
/// Identify program paths which execute sequences of retains and releases which
/// can be eliminated.
bool ObjCARCOpt::OptimizeSequences(Function &F) {
- /// Releases, Retains - These are used to store the results of the main flow
- /// analysis. These use Value* as the key instead of Instruction* so that the
- /// map stays valid when we get around to rewriting code and calls get
- /// replaced by arguments.
+ // Releases, Retains - These are used to store the results of the main flow
+ // analysis. These use Value* as the key instead of Instruction* so that the
+ // map stays valid when we get around to rewriting code and calls get
+ // replaced by arguments.
DenseMap<Value *, RRInfo> Releases;
MapVector<Value *, RRInfo> Retains;
- /// This is used during the traversal of the function to track the
- /// states for each identified object at each block.
+ // This is used during the traversal of the function to track the
+ // states for each identified object at each block.
DenseMap<const BasicBlock *, BBState> BBStates;
// Analyze the CFG of the function, and all instructions.
bool NestingDetected = Visit(F, BBStates, Retains, Releases);
// Transform.
- return PerformCodePlacement(BBStates, Retains, Releases, F.getParent()) &&
- NestingDetected;
+ bool AnyPairsCompletelyEliminated = PerformCodePlacement(BBStates, Retains,
+ Releases,
+ F.getParent());
+
+ // Cleanup.
+ MultiOwnersSet.clear();
+
+ return AnyPairsCompletelyEliminated && NestingDetected;
}
/// Check if there is a dependent call earlier that does not have anything in
const Value *Arg = StripPointerCastsAndObjCCalls(Ret->getOperand(0));
- // Look for an ``autorelease'' instruction that is a predecssor of Ret and
+ // Look for an ``autorelease'' instruction that is a predecessor of Ret and
// dependent on Arg such that there are no instructions dependent on Arg
// that need a positive ref count in between the autorelease and Ret.
CallInst *Autorelease =
FindPredecessorAutoreleaseWithSafePath(Arg, BB, Ret,
DependingInstructions, Visited,
PA);
- if (Autorelease) {
- DependingInstructions.clear();
- Visited.clear();
-
- CallInst *Retain =
- FindPredecessorRetainWithSafePath(Arg, BB, Autorelease,
- DependingInstructions, Visited, PA);
- if (Retain) {
- DependingInstructions.clear();
- Visited.clear();
-
- // Check that there is nothing that can affect the reference count
- // between the retain and the call. Note that Retain need not be in BB.
- if (HasSafePathToPredecessorCall(Arg, Retain, DependingInstructions,
- Visited, PA)) {
- // If so, we can zap the retain and autorelease.
- Changed = true;
- ++NumRets;
- DEBUG(dbgs() << "Erasing: " << *Retain << "\nErasing: "
- << *Autorelease << "\n");
- EraseInstruction(Retain);
- EraseInstruction(Autorelease);
- }
- }
- }
+ DependingInstructions.clear();
+ Visited.clear();
+
+ if (!Autorelease)
+ continue;
+ CallInst *Retain =
+ FindPredecessorRetainWithSafePath(Arg, BB, Autorelease,
+ DependingInstructions, Visited, PA);
DependingInstructions.clear();
Visited.clear();
+
+ if (!Retain)
+ continue;
+
+ // Check that there is nothing that can affect the reference count
+ // between the retain and the call. Note that Retain need not be in BB.
+ bool HasSafePathToCall = HasSafePathToPredecessorCall(Arg, Retain,
+ DependingInstructions,
+ Visited, PA);
+ DependingInstructions.clear();
+ Visited.clear();
+
+ if (!HasSafePathToCall)
+ continue;
+
+ // If so, we can zap the retain and autorelease.
+ Changed = true;
+ ++NumRets;
+ DEBUG(dbgs() << "Erasing: " << *Retain << "\nErasing: "
+ << *Autorelease << "\n");
+ EraseInstruction(Retain);
+ EraseInstruction(Autorelease);
}
}
+#ifndef NDEBUG
+void
+ObjCARCOpt::GatherStatistics(Function &F, bool AfterOptimization) {
+ llvm::Statistic &NumRetains =
+ AfterOptimization? NumRetainsAfterOpt : NumRetainsBeforeOpt;
+ llvm::Statistic &NumReleases =
+ AfterOptimization? NumReleasesAfterOpt : NumReleasesBeforeOpt;
+
+ for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
+ Instruction *Inst = &*I++;
+ switch (GetBasicInstructionClass(Inst)) {
+ default:
+ break;
+ case IC_Retain:
+ ++NumRetains;
+ break;
+ case IC_Release:
+ ++NumReleases;
+ break;
+ }
+ }
+}
+#endif
+
bool ObjCARCOpt::doInitialization(Module &M) {
if (!EnableARCOpts)
return false;
// calls finalizers which can have arbitrary side effects.
// These are initialized lazily.
- RetainRVCallee = 0;
AutoreleaseRVCallee = 0;
ReleaseCallee = 0;
RetainCallee = 0;
PA.setAA(&getAnalysis<AliasAnalysis>());
+#ifndef NDEBUG
+ if (AreStatisticsEnabled()) {
+ GatherStatistics(F, false);
+ }
+#endif
+
// This pass performs several distinct transformations. As a compile-time aid
// when compiling code that isn't ObjC, skip these if the relevant ObjC
// library functions aren't declared.
- // Preliminary optimizations. This also computs UsedInThisFunction.
+ // Preliminary optimizations. This also computes UsedInThisFunction.
OptimizeIndividualCalls(F);
// Optimizations for weak pointers.
(1 << IC_AutoreleaseRV)))
OptimizeReturns(F);
+ // Gather statistics after optimization.
+#ifndef NDEBUG
+ if (AreStatisticsEnabled()) {
+ GatherStatistics(F, true);
+ }
+#endif
+
DEBUG(dbgs() << "\n");
return Changed;
projects / oota-llvm.git / blobdiff