1 //===- DeadStoreElimination.cpp - Fast Dead Store Elimination -------------===//
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 implements a trivial dead store elimination that only considers
11 // basic-block local redundant stores.
13 // FIXME: This should eventually be extended to be a post-dominator tree
14 // traversal. Doing so would be pretty trivial.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "dse"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Function.h"
22 #include "llvm/GlobalVariable.h"
23 #include "llvm/Instructions.h"
24 #include "llvm/IntrinsicInst.h"
25 #include "llvm/Pass.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/CaptureTracking.h"
28 #include "llvm/Analysis/Dominators.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/ADT/SetVector.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/ADT/STLExtras.h"
40 STATISTIC(NumFastStores, "Number of stores deleted");
41 STATISTIC(NumFastOther , "Number of other instrs removed");
44 struct DSE : public FunctionPass {
46 MemoryDependenceAnalysis *MD;
49 static char ID; // Pass identification, replacement for typeid
50 DSE() : FunctionPass(ID), AA(0), MD(0), DT(0) {
51 initializeDSEPass(*PassRegistry::getPassRegistry());
54 virtual bool runOnFunction(Function &F) {
55 AA = &getAnalysis<AliasAnalysis>();
56 MD = &getAnalysis<MemoryDependenceAnalysis>();
57 DT = &getAnalysis<DominatorTree>();
60 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
61 // Only check non-dead blocks. Dead blocks may have strange pointer
62 // cycles that will confuse alias analysis.
63 if (DT->isReachableFromEntry(I))
64 Changed |= runOnBasicBlock(*I);
66 AA = 0; MD = 0; DT = 0;
70 bool runOnBasicBlock(BasicBlock &BB);
71 bool HandleFree(CallInst *F);
72 bool handleEndBlock(BasicBlock &BB);
73 void RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc,
74 SmallSetVector<Value*, 16> &DeadStackObjects);
76 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
78 AU.addRequired<DominatorTree>();
79 AU.addRequired<AliasAnalysis>();
80 AU.addRequired<MemoryDependenceAnalysis>();
81 AU.addPreserved<AliasAnalysis>();
82 AU.addPreserved<DominatorTree>();
83 AU.addPreserved<MemoryDependenceAnalysis>();
89 INITIALIZE_PASS_BEGIN(DSE, "dse", "Dead Store Elimination", false, false)
90 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
91 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
92 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
93 INITIALIZE_PASS_END(DSE, "dse", "Dead Store Elimination", false, false)
95 FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
97 //===----------------------------------------------------------------------===//
99 //===----------------------------------------------------------------------===//
101 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
102 /// and zero out all the operands of this instruction. If any of them become
103 /// dead, delete them and the computation tree that feeds them.
105 /// If ValueSet is non-null, remove any deleted instructions from it as well.
107 static void DeleteDeadInstruction(Instruction *I,
108 MemoryDependenceAnalysis &MD,
109 SmallSetVector<Value*, 16> *ValueSet = 0) {
110 SmallVector<Instruction*, 32> NowDeadInsts;
112 NowDeadInsts.push_back(I);
115 // Before we touch this instruction, remove it from memdep!
117 Instruction *DeadInst = NowDeadInsts.pop_back_val();
120 // This instruction is dead, zap it, in stages. Start by removing it from
121 // MemDep, which needs to know the operands and needs it to be in the
123 MD.removeInstruction(DeadInst);
125 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
126 Value *Op = DeadInst->getOperand(op);
127 DeadInst->setOperand(op, 0);
129 // If this operand just became dead, add it to the NowDeadInsts list.
130 if (!Op->use_empty()) continue;
132 if (Instruction *OpI = dyn_cast<Instruction>(Op))
133 if (isInstructionTriviallyDead(OpI))
134 NowDeadInsts.push_back(OpI);
137 DeadInst->eraseFromParent();
139 if (ValueSet) ValueSet->remove(DeadInst);
140 } while (!NowDeadInsts.empty());
144 /// hasMemoryWrite - Does this instruction write some memory? This only returns
145 /// true for things that we can analyze with other helpers below.
146 static bool hasMemoryWrite(Instruction *I) {
147 if (isa<StoreInst>(I))
149 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
150 switch (II->getIntrinsicID()) {
153 case Intrinsic::memset:
154 case Intrinsic::memmove:
155 case Intrinsic::memcpy:
156 case Intrinsic::init_trampoline:
157 case Intrinsic::lifetime_end:
164 /// getLocForWrite - Return a Location stored to by the specified instruction.
165 /// If isRemovable returns true, this function and getLocForRead completely
166 /// describe the memory operations for this instruction.
167 static AliasAnalysis::Location
168 getLocForWrite(Instruction *Inst, AliasAnalysis &AA) {
169 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
170 return AA.getLocation(SI);
172 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Inst)) {
173 // memcpy/memmove/memset.
174 AliasAnalysis::Location Loc = AA.getLocationForDest(MI);
175 // If we don't have target data around, an unknown size in Location means
176 // that we should use the size of the pointee type. This isn't valid for
177 // memset/memcpy, which writes more than an i8.
178 if (Loc.Size == AliasAnalysis::UnknownSize && AA.getTargetData() == 0)
179 return AliasAnalysis::Location();
183 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
184 if (II == 0) return AliasAnalysis::Location();
186 switch (II->getIntrinsicID()) {
187 default: return AliasAnalysis::Location(); // Unhandled intrinsic.
188 case Intrinsic::init_trampoline:
189 // If we don't have target data around, an unknown size in Location means
190 // that we should use the size of the pointee type. This isn't valid for
191 // init.trampoline, which writes more than an i8.
192 if (AA.getTargetData() == 0) return AliasAnalysis::Location();
194 // FIXME: We don't know the size of the trampoline, so we can't really
196 return AliasAnalysis::Location(II->getArgOperand(0));
197 case Intrinsic::lifetime_end: {
198 uint64_t Len = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
199 return AliasAnalysis::Location(II->getArgOperand(1), Len);
204 /// getLocForRead - Return the location read by the specified "hasMemoryWrite"
205 /// instruction if any.
206 static AliasAnalysis::Location
207 getLocForRead(Instruction *Inst, AliasAnalysis &AA) {
208 assert(hasMemoryWrite(Inst) && "Unknown instruction case");
210 // The only instructions that both read and write are the mem transfer
211 // instructions (memcpy/memmove).
212 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Inst))
213 return AA.getLocationForSource(MTI);
214 return AliasAnalysis::Location();
218 /// isRemovable - If the value of this instruction and the memory it writes to
219 /// is unused, may we delete this instruction?
220 static bool isRemovable(Instruction *I) {
221 // Don't remove volatile/atomic stores.
222 if (StoreInst *SI = dyn_cast<StoreInst>(I))
223 return SI->isUnordered();
225 IntrinsicInst *II = cast<IntrinsicInst>(I);
226 switch (II->getIntrinsicID()) {
227 default: llvm_unreachable("doesn't pass 'hasMemoryWrite' predicate");
228 case Intrinsic::lifetime_end:
229 // Never remove dead lifetime_end's, e.g. because it is followed by a
232 case Intrinsic::init_trampoline:
233 // Always safe to remove init_trampoline.
236 case Intrinsic::memset:
237 case Intrinsic::memmove:
238 case Intrinsic::memcpy:
239 // Don't remove volatile memory intrinsics.
240 return !cast<MemIntrinsic>(II)->isVolatile();
245 /// isShortenable - Returns true if this instruction can be safely shortened in
247 static bool isShortenable(Instruction *I) {
248 // Don't shorten stores for now
249 if (isa<StoreInst>(I))
252 IntrinsicInst *II = cast<IntrinsicInst>(I);
253 switch (II->getIntrinsicID()) {
254 default: return false;
255 case Intrinsic::memset:
256 case Intrinsic::memcpy:
257 // Do shorten memory intrinsics.
262 /// getStoredPointerOperand - Return the pointer that is being written to.
263 static Value *getStoredPointerOperand(Instruction *I) {
264 if (StoreInst *SI = dyn_cast<StoreInst>(I))
265 return SI->getPointerOperand();
266 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
267 return MI->getDest();
269 IntrinsicInst *II = cast<IntrinsicInst>(I);
270 switch (II->getIntrinsicID()) {
271 default: llvm_unreachable("Unexpected intrinsic!");
272 case Intrinsic::init_trampoline:
273 return II->getArgOperand(0);
277 static uint64_t getPointerSize(const Value *V, AliasAnalysis &AA) {
279 if (getObjectSize(V, Size, AA.getTargetData()))
281 return AliasAnalysis::UnknownSize;
293 /// isOverwrite - Return 'OverwriteComplete' if a store to the 'Later' location
294 /// completely overwrites a store to the 'Earlier' location.
295 /// 'OverwriteEnd' if the end of the 'Earlier' location is completely
296 /// overwritten by 'Later', or 'OverwriteUnknown' if nothing can be determined
297 static OverwriteResult isOverwrite(const AliasAnalysis::Location &Later,
298 const AliasAnalysis::Location &Earlier,
302 const Value *P1 = Earlier.Ptr->stripPointerCasts();
303 const Value *P2 = Later.Ptr->stripPointerCasts();
305 // If the start pointers are the same, we just have to compare sizes to see if
306 // the later store was larger than the earlier store.
308 // If we don't know the sizes of either access, then we can't do a
310 if (Later.Size == AliasAnalysis::UnknownSize ||
311 Earlier.Size == AliasAnalysis::UnknownSize) {
312 // If we have no TargetData information around, then the size of the store
313 // is inferrable from the pointee type. If they are the same type, then
314 // we know that the store is safe.
315 if (AA.getTargetData() == 0 &&
316 Later.Ptr->getType() == Earlier.Ptr->getType())
317 return OverwriteComplete;
319 return OverwriteUnknown;
322 // Make sure that the Later size is >= the Earlier size.
323 if (Later.Size >= Earlier.Size)
324 return OverwriteComplete;
327 // Otherwise, we have to have size information, and the later store has to be
328 // larger than the earlier one.
329 if (Later.Size == AliasAnalysis::UnknownSize ||
330 Earlier.Size == AliasAnalysis::UnknownSize ||
331 AA.getTargetData() == 0)
332 return OverwriteUnknown;
334 // Check to see if the later store is to the entire object (either a global,
335 // an alloca, or a byval argument). If so, then it clearly overwrites any
336 // other store to the same object.
337 const TargetData &TD = *AA.getTargetData();
339 const Value *UO1 = GetUnderlyingObject(P1, &TD),
340 *UO2 = GetUnderlyingObject(P2, &TD);
342 // If we can't resolve the same pointers to the same object, then we can't
343 // analyze them at all.
345 return OverwriteUnknown;
347 // If the "Later" store is to a recognizable object, get its size.
348 uint64_t ObjectSize = getPointerSize(UO2, AA);
349 if (ObjectSize != AliasAnalysis::UnknownSize)
350 if (ObjectSize == Later.Size && ObjectSize >= Earlier.Size)
351 return OverwriteComplete;
353 // Okay, we have stores to two completely different pointers. Try to
354 // decompose the pointer into a "base + constant_offset" form. If the base
355 // pointers are equal, then we can reason about the two stores.
358 const Value *BP1 = GetPointerBaseWithConstantOffset(P1, EarlierOff, TD);
359 const Value *BP2 = GetPointerBaseWithConstantOffset(P2, LaterOff, TD);
361 // If the base pointers still differ, we have two completely different stores.
363 return OverwriteUnknown;
365 // The later store completely overlaps the earlier store if:
367 // 1. Both start at the same offset and the later one's size is greater than
368 // or equal to the earlier one's, or
373 // 2. The earlier store has an offset greater than the later offset, but which
374 // still lies completely within the later store.
377 // |----- later ------|
379 // We have to be careful here as *Off is signed while *.Size is unsigned.
380 if (EarlierOff >= LaterOff &&
381 Later.Size > Earlier.Size &&
382 uint64_t(EarlierOff - LaterOff) + Earlier.Size <= Later.Size)
383 return OverwriteComplete;
385 // The other interesting case is if the later store overwrites the end of
391 // In this case we may want to trim the size of earlier to avoid generating
392 // writes to addresses which will definitely be overwritten later
393 if (LaterOff > EarlierOff &&
394 LaterOff < int64_t(EarlierOff + Earlier.Size) &&
395 int64_t(LaterOff + Later.Size) >= int64_t(EarlierOff + Earlier.Size))
398 // Otherwise, they don't completely overlap.
399 return OverwriteUnknown;
402 /// isPossibleSelfRead - If 'Inst' might be a self read (i.e. a noop copy of a
403 /// memory region into an identical pointer) then it doesn't actually make its
404 /// input dead in the traditional sense. Consider this case:
409 /// In this case, the second store to A does not make the first store to A dead.
410 /// The usual situation isn't an explicit A<-A store like this (which can be
411 /// trivially removed) but a case where two pointers may alias.
413 /// This function detects when it is unsafe to remove a dependent instruction
414 /// because the DSE inducing instruction may be a self-read.
415 static bool isPossibleSelfRead(Instruction *Inst,
416 const AliasAnalysis::Location &InstStoreLoc,
417 Instruction *DepWrite, AliasAnalysis &AA) {
418 // Self reads can only happen for instructions that read memory. Get the
420 AliasAnalysis::Location InstReadLoc = getLocForRead(Inst, AA);
421 if (InstReadLoc.Ptr == 0) return false; // Not a reading instruction.
423 // If the read and written loc obviously don't alias, it isn't a read.
424 if (AA.isNoAlias(InstReadLoc, InstStoreLoc)) return false;
426 // Okay, 'Inst' may copy over itself. However, we can still remove a the
427 // DepWrite instruction if we can prove that it reads from the same location
428 // as Inst. This handles useful cases like:
431 // Here we don't know if A/B may alias, but we do know that B/B are must
432 // aliases, so removing the first memcpy is safe (assuming it writes <= #
433 // bytes as the second one.
434 AliasAnalysis::Location DepReadLoc = getLocForRead(DepWrite, AA);
436 if (DepReadLoc.Ptr && AA.isMustAlias(InstReadLoc.Ptr, DepReadLoc.Ptr))
439 // If DepWrite doesn't read memory or if we can't prove it is a must alias,
440 // then it can't be considered dead.
445 //===----------------------------------------------------------------------===//
447 //===----------------------------------------------------------------------===//
449 bool DSE::runOnBasicBlock(BasicBlock &BB) {
450 bool MadeChange = false;
452 // Do a top-down walk on the BB.
453 for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
454 Instruction *Inst = BBI++;
456 // Handle 'free' calls specially.
457 if (CallInst *F = isFreeCall(Inst)) {
458 MadeChange |= HandleFree(F);
462 // If we find something that writes memory, get its memory dependence.
463 if (!hasMemoryWrite(Inst))
466 MemDepResult InstDep = MD->getDependency(Inst);
468 // Ignore any store where we can't find a local dependence.
469 // FIXME: cross-block DSE would be fun. :)
470 if (!InstDep.isDef() && !InstDep.isClobber())
473 // If we're storing the same value back to a pointer that we just
474 // loaded from, then the store can be removed.
475 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
476 if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
477 if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
478 SI->getOperand(0) == DepLoad && isRemovable(SI)) {
479 DEBUG(dbgs() << "DSE: Remove Store Of Load from same pointer:\n "
480 << "LOAD: " << *DepLoad << "\n STORE: " << *SI << '\n');
482 // DeleteDeadInstruction can delete the current instruction. Save BBI
483 // in case we need it.
484 WeakVH NextInst(BBI);
486 DeleteDeadInstruction(SI, *MD);
488 if (NextInst == 0) // Next instruction deleted.
490 else if (BBI != BB.begin()) // Revisit this instruction if possible.
499 // Figure out what location is being stored to.
500 AliasAnalysis::Location Loc = getLocForWrite(Inst, *AA);
502 // If we didn't get a useful location, fail.
506 while (InstDep.isDef() || InstDep.isClobber()) {
507 // Get the memory clobbered by the instruction we depend on. MemDep will
508 // skip any instructions that 'Loc' clearly doesn't interact with. If we
509 // end up depending on a may- or must-aliased load, then we can't optimize
510 // away the store and we bail out. However, if we depend on on something
511 // that overwrites the memory location we *can* potentially optimize it.
513 // Find out what memory location the dependent instruction stores.
514 Instruction *DepWrite = InstDep.getInst();
515 AliasAnalysis::Location DepLoc = getLocForWrite(DepWrite, *AA);
516 // If we didn't get a useful location, or if it isn't a size, bail out.
520 // If we find a write that is a) removable (i.e., non-volatile), b) is
521 // completely obliterated by the store to 'Loc', and c) which we know that
522 // 'Inst' doesn't load from, then we can remove it.
523 if (isRemovable(DepWrite) &&
524 !isPossibleSelfRead(Inst, Loc, DepWrite, *AA)) {
525 int64_t InstWriteOffset, DepWriteOffset;
526 OverwriteResult OR = isOverwrite(Loc, DepLoc, *AA,
527 DepWriteOffset, InstWriteOffset);
528 if (OR == OverwriteComplete) {
529 DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: "
530 << *DepWrite << "\n KILLER: " << *Inst << '\n');
532 // Delete the store and now-dead instructions that feed it.
533 DeleteDeadInstruction(DepWrite, *MD);
537 // DeleteDeadInstruction can delete the current instruction in loop
540 if (BBI != BB.begin())
543 } else if (OR == OverwriteEnd && isShortenable(DepWrite)) {
544 // TODO: base this on the target vector size so that if the earlier
545 // store was too small to get vector writes anyway then its likely
546 // a good idea to shorten it
547 // Power of 2 vector writes are probably always a bad idea to optimize
548 // as any store/memset/memcpy is likely using vector instructions so
549 // shortening it to not vector size is likely to be slower
550 MemIntrinsic* DepIntrinsic = cast<MemIntrinsic>(DepWrite);
551 unsigned DepWriteAlign = DepIntrinsic->getAlignment();
552 if (llvm::isPowerOf2_64(InstWriteOffset) ||
553 ((DepWriteAlign != 0) && InstWriteOffset % DepWriteAlign == 0)) {
555 DEBUG(dbgs() << "DSE: Remove Dead Store:\n OW END: "
556 << *DepWrite << "\n KILLER (offset "
557 << InstWriteOffset << ", "
558 << DepLoc.Size << ")"
561 Value* DepWriteLength = DepIntrinsic->getLength();
562 Value* TrimmedLength = ConstantInt::get(DepWriteLength->getType(),
565 DepIntrinsic->setLength(TrimmedLength);
571 // If this is a may-aliased store that is clobbering the store value, we
572 // can keep searching past it for another must-aliased pointer that stores
573 // to the same location. For example, in:
577 // we can remove the first store to P even though we don't know if P and Q
579 if (DepWrite == &BB.front()) break;
581 // Can't look past this instruction if it might read 'Loc'.
582 if (AA->getModRefInfo(DepWrite, Loc) & AliasAnalysis::Ref)
585 InstDep = MD->getPointerDependencyFrom(Loc, false, DepWrite, &BB);
589 // If this block ends in a return, unwind, or unreachable, all allocas are
590 // dead at its end, which means stores to them are also dead.
591 if (BB.getTerminator()->getNumSuccessors() == 0)
592 MadeChange |= handleEndBlock(BB);
597 /// Find all blocks that will unconditionally lead to the block BB and append
599 static void FindUnconditionalPreds(SmallVectorImpl<BasicBlock *> &Blocks,
600 BasicBlock *BB, DominatorTree *DT) {
601 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
602 BasicBlock *Pred = *I;
603 if (Pred == BB) continue;
604 TerminatorInst *PredTI = Pred->getTerminator();
605 if (PredTI->getNumSuccessors() != 1)
608 if (DT->isReachableFromEntry(Pred))
609 Blocks.push_back(Pred);
613 /// HandleFree - Handle frees of entire structures whose dependency is a store
614 /// to a field of that structure.
615 bool DSE::HandleFree(CallInst *F) {
616 bool MadeChange = false;
618 AliasAnalysis::Location Loc = AliasAnalysis::Location(F->getOperand(0));
619 SmallVector<BasicBlock *, 16> Blocks;
620 Blocks.push_back(F->getParent());
622 while (!Blocks.empty()) {
623 BasicBlock *BB = Blocks.pop_back_val();
624 Instruction *InstPt = BB->getTerminator();
625 if (BB == F->getParent()) InstPt = F;
627 MemDepResult Dep = MD->getPointerDependencyFrom(Loc, false, InstPt, BB);
628 while (Dep.isDef() || Dep.isClobber()) {
629 Instruction *Dependency = Dep.getInst();
630 if (!hasMemoryWrite(Dependency) || !isRemovable(Dependency))
634 GetUnderlyingObject(getStoredPointerOperand(Dependency));
636 // Check for aliasing.
637 if (!AA->isMustAlias(F->getArgOperand(0), DepPointer))
640 Instruction *Next = llvm::next(BasicBlock::iterator(Dependency));
642 // DCE instructions only used to calculate that store
643 DeleteDeadInstruction(Dependency, *MD);
647 // Inst's old Dependency is now deleted. Compute the next dependency,
648 // which may also be dead, as in
650 // s[1] = 0; // This has just been deleted.
652 Dep = MD->getPointerDependencyFrom(Loc, false, Next, BB);
655 if (Dep.isNonLocal())
656 FindUnconditionalPreds(Blocks, BB, DT);
662 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
663 /// function end block. Ex:
666 /// store i32 1, i32* %A
668 bool DSE::handleEndBlock(BasicBlock &BB) {
669 bool MadeChange = false;
671 // Keep track of all of the stack objects that are dead at the end of the
673 SmallSetVector<Value*, 16> DeadStackObjects;
675 // Find all of the alloca'd pointers in the entry block.
676 BasicBlock *Entry = BB.getParent()->begin();
677 for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I) {
678 if (isa<AllocaInst>(I))
679 DeadStackObjects.insert(I);
681 // Okay, so these are dead heap objects, but if the pointer never escapes
682 // then it's leaked by this function anyways.
683 else if (isAllocLikeFn(I) && !PointerMayBeCaptured(I, true, true))
684 DeadStackObjects.insert(I);
687 // Treat byval arguments the same, stores to them are dead at the end of the
689 for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
690 AE = BB.getParent()->arg_end(); AI != AE; ++AI)
691 if (AI->hasByValAttr())
692 DeadStackObjects.insert(AI);
694 // Scan the basic block backwards
695 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
698 // If we find a store, check to see if it points into a dead stack value.
699 if (hasMemoryWrite(BBI) && isRemovable(BBI)) {
700 // See through pointer-to-pointer bitcasts
701 SmallVector<Value *, 4> Pointers;
702 GetUnderlyingObjects(getStoredPointerOperand(BBI), Pointers);
704 // Stores to stack values are valid candidates for removal.
706 for (SmallVectorImpl<Value *>::iterator I = Pointers.begin(),
707 E = Pointers.end(); I != E; ++I)
708 if (!DeadStackObjects.count(*I)) {
714 Instruction *Dead = BBI++;
716 DEBUG(dbgs() << "DSE: Dead Store at End of Block:\n DEAD: "
717 << *Dead << "\n Objects: ";
718 for (SmallVectorImpl<Value *>::iterator I = Pointers.begin(),
719 E = Pointers.end(); I != E; ++I) {
721 if (llvm::next(I) != E)
726 // DCE instructions only used to calculate that store.
727 DeleteDeadInstruction(Dead, *MD, &DeadStackObjects);
734 // Remove any dead non-memory-mutating instructions.
735 if (isInstructionTriviallyDead(BBI)) {
736 Instruction *Inst = BBI++;
737 DeleteDeadInstruction(Inst, *MD, &DeadStackObjects);
743 if (isa<AllocaInst>(BBI) || isAllocLikeFn(BBI)) {
744 DeadStackObjects.remove(BBI);
748 if (CallSite CS = cast<Value>(BBI)) {
749 // If this call does not access memory, it can't be loading any of our
751 if (AA->doesNotAccessMemory(CS))
754 // If the call might load from any of our allocas, then any store above
756 SmallVector<Value*, 8> LiveAllocas;
757 for (SmallSetVector<Value*, 16>::iterator I = DeadStackObjects.begin(),
758 E = DeadStackObjects.end(); I != E; ++I) {
759 // See if the call site touches it.
760 AliasAnalysis::ModRefResult A =
761 AA->getModRefInfo(CS, *I, getPointerSize(*I, *AA));
763 if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
764 LiveAllocas.push_back(*I);
767 for (SmallVector<Value*, 8>::iterator I = LiveAllocas.begin(),
768 E = LiveAllocas.end(); I != E; ++I)
769 DeadStackObjects.remove(*I);
771 // If all of the allocas were clobbered by the call then we're not going
772 // to find anything else to process.
773 if (DeadStackObjects.empty())
779 AliasAnalysis::Location LoadedLoc;
781 // If we encounter a use of the pointer, it is no longer considered dead
782 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
783 if (!L->isUnordered()) // Be conservative with atomic/volatile load
785 LoadedLoc = AA->getLocation(L);
786 } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) {
787 LoadedLoc = AA->getLocation(V);
788 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(BBI)) {
789 LoadedLoc = AA->getLocationForSource(MTI);
790 } else if (!BBI->mayReadFromMemory()) {
791 // Instruction doesn't read memory. Note that stores that weren't removed
792 // above will hit this case.
795 // Unknown inst; assume it clobbers everything.
799 // Remove any allocas from the DeadPointer set that are loaded, as this
800 // makes any stores above the access live.
801 RemoveAccessedObjects(LoadedLoc, DeadStackObjects);
803 // If all of the allocas were clobbered by the access then we're not going
804 // to find anything else to process.
805 if (DeadStackObjects.empty())
812 /// RemoveAccessedObjects - Check to see if the specified location may alias any
813 /// of the stack objects in the DeadStackObjects set. If so, they become live
814 /// because the location is being loaded.
815 void DSE::RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc,
816 SmallSetVector<Value*, 16> &DeadStackObjects) {
817 const Value *UnderlyingPointer = GetUnderlyingObject(LoadedLoc.Ptr);
819 // A constant can't be in the dead pointer set.
820 if (isa<Constant>(UnderlyingPointer))
823 // If the kill pointer can be easily reduced to an alloca, don't bother doing
824 // extraneous AA queries.
825 if (isa<AllocaInst>(UnderlyingPointer) || isa<Argument>(UnderlyingPointer)) {
826 DeadStackObjects.remove(const_cast<Value*>(UnderlyingPointer));
830 SmallVector<Value*, 16> NowLive;
831 for (SmallSetVector<Value*, 16>::iterator I = DeadStackObjects.begin(),
832 E = DeadStackObjects.end(); I != E; ++I) {
833 // See if the loaded location could alias the stack location.
834 AliasAnalysis::Location StackLoc(*I, getPointerSize(*I, *AA));
835 if (!AA->isNoAlias(StackLoc, LoadedLoc))
836 NowLive.push_back(*I);
839 for (SmallVector<Value*, 16>::iterator I = NowLive.begin(), E = NowLive.end();
841 DeadStackObjects.remove(*I);