1 //===- ObjCARCContract.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 late ObjC ARC optimizations. ARC stands for Automatic
11 /// Reference Counting and is a system for managing reference counts for objects
14 /// This specific file mainly deals with ``contracting'' multiple lower level
15 /// operations into singular higher level operations through pattern matching.
17 /// WARNING: This file knows about certain library functions. It recognizes them
18 /// by name, and hardwires knowledge of their semantics.
20 /// WARNING: This file knows about how certain Objective-C library functions are
21 /// used. Naive LLVM IR transformations which would otherwise be
22 /// behavior-preserving may break these assumptions.
24 //===----------------------------------------------------------------------===//
26 // TODO: ObjCARCContract could insert PHI nodes when uses aren't
27 // dominated by single calls.
29 #include "ARCRuntimeEntryPoints.h"
30 #include "DependencyAnalysis.h"
32 #include "ProvenanceAnalysis.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/InlineAsm.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/raw_ostream.h"
41 using namespace llvm::objcarc;
43 #define DEBUG_TYPE "objc-arc-contract"
45 STATISTIC(NumPeeps, "Number of calls peephole-optimized");
46 STATISTIC(NumStoreStrongs, "Number objc_storeStrong calls formed");
48 //===----------------------------------------------------------------------===//
50 //===----------------------------------------------------------------------===//
53 /// \brief Late ARC optimizations
55 /// These change the IR in a way that makes it difficult to be analyzed by
56 /// ObjCARCOpt, so it's run late.
57 class ObjCARCContract : public FunctionPass {
61 ProvenanceAnalysis PA;
62 ARCRuntimeEntryPoints EP;
64 /// A flag indicating whether this optimization pass should run.
67 /// The inline asm string to insert between calls and RetainRV calls to make
68 /// the optimization work on targets which need it.
69 const MDString *RVInstMarker;
71 /// The set of inserted objc_storeStrong calls. If at the end of walking the
72 /// function we have found no alloca instructions, these calls can be marked
74 SmallPtrSet<CallInst *, 8> StoreStrongCalls;
76 /// Returns true if we eliminated Inst.
77 bool tryToPeepholeInstruction(Function &F, Instruction *Inst,
79 SmallPtrSetImpl<Instruction *> &DepInsts,
80 SmallPtrSetImpl<const BasicBlock *> &Visited,
81 bool &TailOkForStoreStrong);
83 bool optimizeRetainCall(Function &F, Instruction *Retain);
86 contractAutorelease(Function &F, Instruction *Autorelease,
88 SmallPtrSetImpl<Instruction *> &DependingInstructions,
89 SmallPtrSetImpl<const BasicBlock *> &Visited);
91 void tryToContractReleaseIntoStoreStrong(Instruction *Release,
94 void getAnalysisUsage(AnalysisUsage &AU) const override;
95 bool doInitialization(Module &M) override;
96 bool runOnFunction(Function &F) override;
100 ObjCARCContract() : FunctionPass(ID) {
101 initializeObjCARCContractPass(*PassRegistry::getPassRegistry());
106 //===----------------------------------------------------------------------===//
108 //===----------------------------------------------------------------------===//
110 /// Turn objc_retain into objc_retainAutoreleasedReturnValue if the operand is a
111 /// return value. We do this late so we do not disrupt the dataflow analysis in
113 bool ObjCARCContract::optimizeRetainCall(Function &F, Instruction *Retain) {
114 ImmutableCallSite CS(GetArgRCIdentityRoot(Retain));
115 const Instruction *Call = CS.getInstruction();
118 if (Call->getParent() != Retain->getParent())
121 // Check that the call is next to the retain.
122 BasicBlock::const_iterator I = ++Call->getIterator();
123 while (IsNoopInstruction(&*I))
128 // Turn it to an objc_retainAutoreleasedReturnValue.
132 DEBUG(dbgs() << "Transforming objc_retain => "
133 "objc_retainAutoreleasedReturnValue since the operand is a "
134 "return value.\nOld: "<< *Retain << "\n");
136 // We do not have to worry about tail calls/does not throw since
137 // retain/retainRV have the same properties.
138 Constant *Decl = EP.get(ARCRuntimeEntryPointKind::RetainRV);
139 cast<CallInst>(Retain)->setCalledFunction(Decl);
141 DEBUG(dbgs() << "New: " << *Retain << "\n");
145 /// Merge an autorelease with a retain into a fused call.
146 bool ObjCARCContract::contractAutorelease(
147 Function &F, Instruction *Autorelease, ARCInstKind Class,
148 SmallPtrSetImpl<Instruction *> &DependingInstructions,
149 SmallPtrSetImpl<const BasicBlock *> &Visited) {
150 const Value *Arg = GetArgRCIdentityRoot(Autorelease);
152 // Check that there are no instructions between the retain and the autorelease
153 // (such as an autorelease_pop) which may change the count.
154 CallInst *Retain = nullptr;
155 if (Class == ARCInstKind::AutoreleaseRV)
156 FindDependencies(RetainAutoreleaseRVDep, Arg,
157 Autorelease->getParent(), Autorelease,
158 DependingInstructions, Visited, PA);
160 FindDependencies(RetainAutoreleaseDep, Arg,
161 Autorelease->getParent(), Autorelease,
162 DependingInstructions, Visited, PA);
165 if (DependingInstructions.size() != 1) {
166 DependingInstructions.clear();
170 Retain = dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
171 DependingInstructions.clear();
173 if (!Retain || GetBasicARCInstKind(Retain) != ARCInstKind::Retain ||
174 GetArgRCIdentityRoot(Retain) != Arg)
180 DEBUG(dbgs() << " Fusing retain/autorelease!\n"
181 " Autorelease:" << *Autorelease << "\n"
182 " Retain: " << *Retain << "\n");
184 Constant *Decl = EP.get(Class == ARCInstKind::AutoreleaseRV
185 ? ARCRuntimeEntryPointKind::RetainAutoreleaseRV
186 : ARCRuntimeEntryPointKind::RetainAutorelease);
187 Retain->setCalledFunction(Decl);
189 DEBUG(dbgs() << " New RetainAutorelease: " << *Retain << "\n");
191 EraseInstruction(Autorelease);
195 static StoreInst *findSafeStoreForStoreStrongContraction(LoadInst *Load,
196 Instruction *Release,
197 ProvenanceAnalysis &PA,
199 StoreInst *Store = nullptr;
200 bool SawRelease = false;
202 // Get the location associated with Load.
203 MemoryLocation Loc = MemoryLocation::get(Load);
204 auto *LocPtr = Loc.Ptr->stripPointerCasts();
206 // Walk down to find the store and the release, which may be in either order.
207 for (auto I = std::next(BasicBlock::iterator(Load)),
208 E = Load->getParent()->end();
210 // If we found the store we were looking for and saw the release,
211 // break. There is no more work to be done.
212 if (Store && SawRelease)
215 // Now we know that we have not seen either the store or the release. If I
216 // is the release, mark that we saw the release and continue.
217 Instruction *Inst = &*I;
218 if (Inst == Release) {
223 // Otherwise, we check if Inst is a "good" store. Grab the instruction class
225 ARCInstKind Class = GetBasicARCInstKind(Inst);
227 // If Inst is an unrelated retain, we don't care about it.
229 // TODO: This is one area where the optimization could be made more
234 // If we have seen the store, but not the release...
236 // We need to make sure that it is safe to move the release from its
237 // current position to the store. This implies proving that any
238 // instruction in between Store and the Release conservatively can not use
239 // the RCIdentityRoot of Release. If we can prove we can ignore Inst, so
241 if (!CanUse(Inst, Load, PA, Class)) {
245 // Otherwise, be conservative and return nullptr.
249 // Ok, now we know we have not seen a store yet. See if Inst can write to
250 // our load location, if it can not, just ignore the instruction.
251 if (!isModSet(AA->getModRefInfo(Inst, Loc)))
254 Store = dyn_cast<StoreInst>(Inst);
256 // If Inst can, then check if Inst is a simple store. If Inst is not a
257 // store or a store that is not simple, then we have some we do not
258 // understand writing to this memory implying we can not move the load
259 // over the write to any subsequent store that we may find.
260 if (!Store || !Store->isSimple())
263 // Then make sure that the pointer we are storing to is Ptr. If so, we
265 if (Store->getPointerOperand()->stripPointerCasts() == LocPtr)
268 // Otherwise, we have an unknown store to some other ptr that clobbers
273 // If we did not find the store or did not see the release, fail.
274 if (!Store || !SawRelease)
282 findRetainForStoreStrongContraction(Value *New, StoreInst *Store,
283 Instruction *Release,
284 ProvenanceAnalysis &PA) {
285 // Walk up from the Store to find the retain.
286 BasicBlock::iterator I = Store->getIterator();
287 BasicBlock::iterator Begin = Store->getParent()->begin();
288 while (I != Begin && GetBasicARCInstKind(&*I) != ARCInstKind::Retain) {
289 Instruction *Inst = &*I;
291 // It is only safe to move the retain to the store if we can prove
292 // conservatively that nothing besides the release can decrement reference
293 // counts in between the retain and the store.
294 if (CanDecrementRefCount(Inst, New, PA) && Inst != Release)
298 Instruction *Retain = &*I;
299 if (GetBasicARCInstKind(Retain) != ARCInstKind::Retain)
301 if (GetArgRCIdentityRoot(Retain) != New)
306 /// Attempt to merge an objc_release with a store, load, and objc_retain to form
307 /// an objc_storeStrong. An objc_storeStrong:
309 /// objc_storeStrong(i8** %old_ptr, i8* new_value)
311 /// is equivalent to the following IR sequence:
313 /// ; Load old value.
314 /// %old_value = load i8** %old_ptr (1)
316 /// ; Increment the new value and then release the old value. This must occur
317 /// ; in order in case old_value releases new_value in its destructor causing
318 /// ; us to potentially have a dangling ptr.
319 /// tail call i8* @objc_retain(i8* %new_value) (2)
320 /// tail call void @objc_release(i8* %old_value) (3)
322 /// ; Store the new_value into old_ptr
323 /// store i8* %new_value, i8** %old_ptr (4)
325 /// The safety of this optimization is based around the following
328 /// 1. We are forming the store strong at the store. Thus to perform this
329 /// optimization it must be safe to move the retain, load, and release to
331 /// 2. We need to make sure that any re-orderings of (1), (2), (3), (4) are
333 void ObjCARCContract::tryToContractReleaseIntoStoreStrong(Instruction *Release,
334 inst_iterator &Iter) {
335 // See if we are releasing something that we just loaded.
336 auto *Load = dyn_cast<LoadInst>(GetArgRCIdentityRoot(Release));
337 if (!Load || !Load->isSimple())
340 // For now, require everything to be in one basic block.
341 BasicBlock *BB = Release->getParent();
342 if (Load->getParent() != BB)
345 // First scan down the BB from Load, looking for a store of the RCIdentityRoot
348 findSafeStoreForStoreStrongContraction(Load, Release, PA, AA);
353 // Then find what new_value's RCIdentity Root is.
354 Value *New = GetRCIdentityRoot(Store->getValueOperand());
356 // Then walk up the BB and look for a retain on New without any intervening
357 // instructions which conservatively might decrement ref counts.
358 Instruction *Retain =
359 findRetainForStoreStrongContraction(New, Store, Release, PA);
369 llvm::dbgs() << " Contracting retain, release into objc_storeStrong.\n"
371 << " Store: " << *Store << "\n"
372 << " Release: " << *Release << "\n"
373 << " Retain: " << *Retain << "\n"
374 << " Load: " << *Load << "\n");
376 LLVMContext &C = Release->getContext();
377 Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
378 Type *I8XX = PointerType::getUnqual(I8X);
380 Value *Args[] = { Load->getPointerOperand(), New };
381 if (Args[0]->getType() != I8XX)
382 Args[0] = new BitCastInst(Args[0], I8XX, "", Store);
383 if (Args[1]->getType() != I8X)
384 Args[1] = new BitCastInst(Args[1], I8X, "", Store);
385 Constant *Decl = EP.get(ARCRuntimeEntryPointKind::StoreStrong);
386 CallInst *StoreStrong = CallInst::Create(Decl, Args, "", Store);
387 StoreStrong->setDoesNotThrow();
388 StoreStrong->setDebugLoc(Store->getDebugLoc());
390 // We can't set the tail flag yet, because we haven't yet determined
391 // whether there are any escaping allocas. Remember this call, so that
392 // we can set the tail flag once we know it's safe.
393 StoreStrongCalls.insert(StoreStrong);
395 DEBUG(llvm::dbgs() << " New Store Strong: " << *StoreStrong << "\n");
397 if (&*Iter == Retain) ++Iter;
398 if (&*Iter == Store) ++Iter;
399 Store->eraseFromParent();
400 Release->eraseFromParent();
401 EraseInstruction(Retain);
402 if (Load->use_empty())
403 Load->eraseFromParent();
406 bool ObjCARCContract::tryToPeepholeInstruction(
407 Function &F, Instruction *Inst, inst_iterator &Iter,
408 SmallPtrSetImpl<Instruction *> &DependingInsts,
409 SmallPtrSetImpl<const BasicBlock *> &Visited,
410 bool &TailOkForStoreStrongs) {
411 // Only these library routines return their argument. In particular,
412 // objc_retainBlock does not necessarily return its argument.
413 ARCInstKind Class = GetBasicARCInstKind(Inst);
415 case ARCInstKind::FusedRetainAutorelease:
416 case ARCInstKind::FusedRetainAutoreleaseRV:
418 case ARCInstKind::Autorelease:
419 case ARCInstKind::AutoreleaseRV:
420 return contractAutorelease(F, Inst, Class, DependingInsts, Visited);
421 case ARCInstKind::Retain:
422 // Attempt to convert retains to retainrvs if they are next to function
424 if (!optimizeRetainCall(F, Inst))
426 // If we succeed in our optimization, fall through.
428 case ARCInstKind::RetainRV:
429 case ARCInstKind::ClaimRV: {
430 // If we're compiling for a target which needs a special inline-asm
431 // marker to do the return value optimization, insert it now.
434 BasicBlock::iterator BBI = Inst->getIterator();
435 BasicBlock *InstParent = Inst->getParent();
437 // Step up to see if the call immediately precedes the RV call.
438 // If it's an invoke, we have to cross a block boundary. And we have
439 // to carefully dodge no-op instructions.
441 if (BBI == InstParent->begin()) {
442 BasicBlock *Pred = InstParent->getSinglePredecessor();
444 goto decline_rv_optimization;
445 BBI = Pred->getTerminator()->getIterator();
449 } while (IsNoopInstruction(&*BBI));
451 if (&*BBI == GetArgRCIdentityRoot(Inst)) {
452 DEBUG(dbgs() << "Adding inline asm marker for the return value "
455 InlineAsm *IA = InlineAsm::get(
456 FunctionType::get(Type::getVoidTy(Inst->getContext()),
458 RVInstMarker->getString(),
459 /*Constraints=*/"", /*hasSideEffects=*/true);
460 CallInst::Create(IA, "", Inst);
462 decline_rv_optimization:
465 case ARCInstKind::InitWeak: {
466 // objc_initWeak(p, null) => *p = null
467 CallInst *CI = cast<CallInst>(Inst);
468 if (IsNullOrUndef(CI->getArgOperand(1))) {
470 ConstantPointerNull::get(cast<PointerType>(CI->getType()));
472 new StoreInst(Null, CI->getArgOperand(0), CI);
474 DEBUG(dbgs() << "OBJCARCContract: Old = " << *CI << "\n"
475 << " New = " << *Null << "\n");
477 CI->replaceAllUsesWith(Null);
478 CI->eraseFromParent();
482 case ARCInstKind::Release:
483 // Try to form an objc store strong from our release. If we fail, there is
484 // nothing further to do below, so continue.
485 tryToContractReleaseIntoStoreStrong(Inst, Iter);
487 case ARCInstKind::User:
488 // Be conservative if the function has any alloca instructions.
489 // Technically we only care about escaping alloca instructions,
490 // but this is sufficient to handle some interesting cases.
491 if (isa<AllocaInst>(Inst))
492 TailOkForStoreStrongs = false;
494 case ARCInstKind::IntrinsicUser:
495 // Remove calls to @clang.arc.use(...).
496 Inst->eraseFromParent();
503 //===----------------------------------------------------------------------===//
505 //===----------------------------------------------------------------------===//
507 bool ObjCARCContract::runOnFunction(Function &F) {
511 // If nothing in the Module uses ARC, don't do anything.
516 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
517 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
519 PA.setAA(&getAnalysis<AAResultsWrapperPass>().getAAResults());
521 DEBUG(llvm::dbgs() << "**** ObjCARC Contract ****\n");
523 // Track whether it's ok to mark objc_storeStrong calls with the "tail"
524 // keyword. Be conservative if the function has variadic arguments.
525 // It seems that functions which "return twice" are also unsafe for the
526 // "tail" argument, because they are setjmp, which could need to
527 // return to an earlier stack state.
528 bool TailOkForStoreStrongs =
529 !F.isVarArg() && !F.callsFunctionThatReturnsTwice();
531 // For ObjC library calls which return their argument, replace uses of the
532 // argument with uses of the call return value, if it dominates the use. This
533 // reduces register pressure.
534 SmallPtrSet<Instruction *, 4> DependingInstructions;
535 SmallPtrSet<const BasicBlock *, 4> Visited;
536 for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E;) {
537 Instruction *Inst = &*I++;
539 DEBUG(dbgs() << "Visiting: " << *Inst << "\n");
541 // First try to peephole Inst. If there is nothing further we can do in
542 // terms of undoing objc-arc-expand, process the next inst.
543 if (tryToPeepholeInstruction(F, Inst, I, DependingInstructions, Visited,
544 TailOkForStoreStrongs))
547 // Otherwise, try to undo objc-arc-expand.
549 // Don't use GetArgRCIdentityRoot because we don't want to look through bitcasts
550 // and such; to do the replacement, the argument must have type i8*.
552 // Function for replacing uses of Arg dominated by Inst.
553 auto ReplaceArgUses = [Inst, this](Value *Arg) {
554 // If we're compiling bugpointed code, don't get in trouble.
555 if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))
558 // Look through the uses of the pointer.
559 for (Value::use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
561 // Increment UI now, because we may unlink its element.
563 unsigned OperandNo = U.getOperandNo();
565 // If the call's return value dominates a use of the call's argument
566 // value, rewrite the use to use the return value. We check for
567 // reachability here because an unreachable call is considered to
568 // trivially dominate itself, which would lead us to rewriting its
569 // argument in terms of its return value, which would lead to
570 // infinite loops in GetArgRCIdentityRoot.
571 if (DT->isReachableFromEntry(U) && DT->dominates(Inst, U)) {
573 Instruction *Replacement = Inst;
574 Type *UseTy = U.get()->getType();
575 if (PHINode *PHI = dyn_cast<PHINode>(U.getUser())) {
576 // For PHI nodes, insert the bitcast in the predecessor block.
577 unsigned ValNo = PHINode::getIncomingValueNumForOperand(OperandNo);
578 BasicBlock *BB = PHI->getIncomingBlock(ValNo);
579 if (Replacement->getType() != UseTy)
580 Replacement = new BitCastInst(Replacement, UseTy, "",
582 // While we're here, rewrite all edges for this PHI, rather
583 // than just one use at a time, to minimize the number of
585 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
586 if (PHI->getIncomingBlock(i) == BB) {
587 // Keep the UI iterator valid.
590 PHINode::getOperandNumForIncomingValue(i)) == &*UI)
592 PHI->setIncomingValue(i, Replacement);
595 if (Replacement->getType() != UseTy)
596 Replacement = new BitCastInst(Replacement, UseTy, "",
597 cast<Instruction>(U.getUser()));
605 Value *Arg = cast<CallInst>(Inst)->getArgOperand(0);
606 Value *OrigArg = Arg;
608 // TODO: Change this to a do-while.
612 // If Arg is a no-op casted pointer, strip one level of casts and iterate.
613 if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg))
614 Arg = BI->getOperand(0);
615 else if (isa<GEPOperator>(Arg) &&
616 cast<GEPOperator>(Arg)->hasAllZeroIndices())
617 Arg = cast<GEPOperator>(Arg)->getPointerOperand();
618 else if (isa<GlobalAlias>(Arg) &&
619 !cast<GlobalAlias>(Arg)->isInterposable())
620 Arg = cast<GlobalAlias>(Arg)->getAliasee();
625 // Replace bitcast users of Arg that are dominated by Inst.
626 SmallVector<BitCastInst *, 2> BitCastUsers;
628 // Add all bitcast users of the function argument first.
629 for (User *U : OrigArg->users())
630 if (auto *BC = dyn_cast<BitCastInst>(U))
631 BitCastUsers.push_back(BC);
633 // Replace the bitcasts with the call return. Iterate until list is empty.
634 while (!BitCastUsers.empty()) {
635 auto *BC = BitCastUsers.pop_back_val();
636 for (User *U : BC->users())
637 if (auto *B = dyn_cast<BitCastInst>(U))
638 BitCastUsers.push_back(B);
644 // If this function has no escaping allocas or suspicious vararg usage,
645 // objc_storeStrong calls can be marked with the "tail" keyword.
646 if (TailOkForStoreStrongs)
647 for (CallInst *CI : StoreStrongCalls)
649 StoreStrongCalls.clear();
654 //===----------------------------------------------------------------------===//
656 //===----------------------------------------------------------------------===//
658 char ObjCARCContract::ID = 0;
659 INITIALIZE_PASS_BEGIN(ObjCARCContract, "objc-arc-contract",
660 "ObjC ARC contraction", false, false)
661 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
662 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
663 INITIALIZE_PASS_END(ObjCARCContract, "objc-arc-contract",
664 "ObjC ARC contraction", false, false)
666 void ObjCARCContract::getAnalysisUsage(AnalysisUsage &AU) const {
667 AU.addRequired<AAResultsWrapperPass>();
668 AU.addRequired<DominatorTreeWrapperPass>();
669 AU.setPreservesCFG();
672 Pass *llvm::createObjCARCContractPass() { return new ObjCARCContract(); }
674 bool ObjCARCContract::doInitialization(Module &M) {
675 // If nothing in the Module uses ARC, don't do anything.
676 Run = ModuleHasARC(M);
682 // Initialize RVInstMarker.
683 RVInstMarker = nullptr;
684 if (NamedMDNode *NMD =
685 M.getNamedMetadata("clang.arc.retainAutoreleasedReturnValueMarker"))
686 if (NMD->getNumOperands() == 1) {
687 const MDNode *N = NMD->getOperand(0);
688 if (N->getNumOperands() == 1)
689 if (const MDString *S = dyn_cast<MDString>(N->getOperand(0)))