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/SmallPtrSet.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 SmallPtrSet<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 SmallPtrSet<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->erase(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: assert(0 && "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: assert(false && "Unexpected intrinsic!");
272 case Intrinsic::init_trampoline:
273 return II->getArgOperand(0);
277 static uint64_t getPointerSize(Value *V, AliasAnalysis &AA) {
278 const TargetData *TD = AA.getTargetData();
280 if (CallInst *CI = dyn_cast<CallInst>(V)) {
281 assert(isMalloc(CI) && "Expected Malloc call!");
282 if (ConstantInt *C = dyn_cast<ConstantInt>(CI->getArgOperand(0)))
283 return C->getZExtValue();
284 return AliasAnalysis::UnknownSize;
288 return AliasAnalysis::UnknownSize;
290 if (AllocaInst *A = dyn_cast<AllocaInst>(V)) {
291 // Get size information for the alloca
292 if (ConstantInt *C = dyn_cast<ConstantInt>(A->getArraySize()))
293 return C->getZExtValue() * TD->getTypeAllocSize(A->getAllocatedType());
294 return AliasAnalysis::UnknownSize;
297 assert(isa<Argument>(V) && "Expected AllocaInst, malloc call or Argument!");
298 PointerType *PT = cast<PointerType>(V->getType());
299 return TD->getTypeAllocSize(PT->getElementType());
302 /// isObjectPointerWithTrustworthySize - Return true if the specified Value* is
303 /// pointing to an object with a pointer size we can trust.
304 static bool isObjectPointerWithTrustworthySize(const Value *V) {
305 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
306 return !AI->isArrayAllocation();
307 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
308 return !GV->mayBeOverridden();
309 if (const Argument *A = dyn_cast<Argument>(V))
310 return A->hasByValAttr();
325 /// isOverwrite - Return 'OverwriteComplete' if a store to the 'Later' location
326 /// completely overwrites a store to the 'Earlier' location.
327 /// 'OverwriteEnd' if the end of the 'Earlier' location is completely
328 /// overwritten by 'Later', or 'OverwriteUnknown' if nothing can be determined
329 static OverwriteResult isOverwrite(const AliasAnalysis::Location &Later,
330 const AliasAnalysis::Location &Earlier,
334 const Value *P1 = Earlier.Ptr->stripPointerCasts();
335 const Value *P2 = Later.Ptr->stripPointerCasts();
337 // If the start pointers are the same, we just have to compare sizes to see if
338 // the later store was larger than the earlier store.
340 // If we don't know the sizes of either access, then we can't do a
342 if (Later.Size == AliasAnalysis::UnknownSize ||
343 Earlier.Size == AliasAnalysis::UnknownSize) {
344 // If we have no TargetData information around, then the size of the store
345 // is inferrable from the pointee type. If they are the same type, then
346 // we know that the store is safe.
347 if (AA.getTargetData() == 0 &&
348 Later.Ptr->getType() == Earlier.Ptr->getType())
349 return OverwriteComplete;
351 return OverwriteUnknown;
354 // Make sure that the Later size is >= the Earlier size.
355 if (Later.Size >= Earlier.Size)
356 return OverwriteComplete;
359 // Otherwise, we have to have size information, and the later store has to be
360 // larger than the earlier one.
361 if (Later.Size == AliasAnalysis::UnknownSize ||
362 Earlier.Size == AliasAnalysis::UnknownSize ||
363 AA.getTargetData() == 0)
364 return OverwriteUnknown;
366 // Check to see if the later store is to the entire object (either a global,
367 // an alloca, or a byval argument). If so, then it clearly overwrites any
368 // other store to the same object.
369 const TargetData &TD = *AA.getTargetData();
371 const Value *UO1 = GetUnderlyingObject(P1, &TD),
372 *UO2 = GetUnderlyingObject(P2, &TD);
374 // If we can't resolve the same pointers to the same object, then we can't
375 // analyze them at all.
377 return OverwriteUnknown;
379 // If the "Later" store is to a recognizable object, get its size.
380 if (isObjectPointerWithTrustworthySize(UO2)) {
381 uint64_t ObjectSize =
382 TD.getTypeAllocSize(cast<PointerType>(UO2->getType())->getElementType());
383 if (ObjectSize == Later.Size)
384 return OverwriteComplete;
387 // Okay, we have stores to two completely different pointers. Try to
388 // decompose the pointer into a "base + constant_offset" form. If the base
389 // pointers are equal, then we can reason about the two stores.
392 const Value *BP1 = GetPointerBaseWithConstantOffset(P1, EarlierOff, TD);
393 const Value *BP2 = GetPointerBaseWithConstantOffset(P2, LaterOff, TD);
395 // If the base pointers still differ, we have two completely different stores.
397 return OverwriteUnknown;
399 // The later store completely overlaps the earlier store if:
401 // 1. Both start at the same offset and the later one's size is greater than
402 // or equal to the earlier one's, or
407 // 2. The earlier store has an offset greater than the later offset, but which
408 // still lies completely within the later store.
411 // |----- later ------|
413 // We have to be careful here as *Off is signed while *.Size is unsigned.
414 if (EarlierOff >= LaterOff &&
415 Later.Size > Earlier.Size &&
416 uint64_t(EarlierOff - LaterOff) + Earlier.Size <= Later.Size)
417 return OverwriteComplete;
419 // The other interesting case is if the later store overwrites the end of
425 // In this case we may want to trim the size of earlier to avoid generating
426 // writes to addresses which will definitely be overwritten later
427 if (LaterOff > EarlierOff &&
428 LaterOff < int64_t(EarlierOff + Earlier.Size) &&
429 LaterOff + Later.Size >= EarlierOff + Earlier.Size)
432 // Otherwise, they don't completely overlap.
433 return OverwriteUnknown;
436 /// isPossibleSelfRead - If 'Inst' might be a self read (i.e. a noop copy of a
437 /// memory region into an identical pointer) then it doesn't actually make its
438 /// input dead in the traditional sense. Consider this case:
443 /// In this case, the second store to A does not make the first store to A dead.
444 /// The usual situation isn't an explicit A<-A store like this (which can be
445 /// trivially removed) but a case where two pointers may alias.
447 /// This function detects when it is unsafe to remove a dependent instruction
448 /// because the DSE inducing instruction may be a self-read.
449 static bool isPossibleSelfRead(Instruction *Inst,
450 const AliasAnalysis::Location &InstStoreLoc,
451 Instruction *DepWrite, AliasAnalysis &AA) {
452 // Self reads can only happen for instructions that read memory. Get the
454 AliasAnalysis::Location InstReadLoc = getLocForRead(Inst, AA);
455 if (InstReadLoc.Ptr == 0) return false; // Not a reading instruction.
457 // If the read and written loc obviously don't alias, it isn't a read.
458 if (AA.isNoAlias(InstReadLoc, InstStoreLoc)) return false;
460 // Okay, 'Inst' may copy over itself. However, we can still remove a the
461 // DepWrite instruction if we can prove that it reads from the same location
462 // as Inst. This handles useful cases like:
465 // Here we don't know if A/B may alias, but we do know that B/B are must
466 // aliases, so removing the first memcpy is safe (assuming it writes <= #
467 // bytes as the second one.
468 AliasAnalysis::Location DepReadLoc = getLocForRead(DepWrite, AA);
470 if (DepReadLoc.Ptr && AA.isMustAlias(InstReadLoc.Ptr, DepReadLoc.Ptr))
473 // If DepWrite doesn't read memory or if we can't prove it is a must alias,
474 // then it can't be considered dead.
479 //===----------------------------------------------------------------------===//
481 //===----------------------------------------------------------------------===//
483 bool DSE::runOnBasicBlock(BasicBlock &BB) {
484 bool MadeChange = false;
486 // Do a top-down walk on the BB.
487 for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
488 Instruction *Inst = BBI++;
490 // Handle 'free' calls specially.
491 if (CallInst *F = isFreeCall(Inst)) {
492 MadeChange |= HandleFree(F);
496 // If we find something that writes memory, get its memory dependence.
497 if (!hasMemoryWrite(Inst))
500 MemDepResult InstDep = MD->getDependency(Inst);
502 // Ignore any store where we can't find a local dependence.
503 // FIXME: cross-block DSE would be fun. :)
504 if (!InstDep.isDef() && !InstDep.isClobber())
507 // If we're storing the same value back to a pointer that we just
508 // loaded from, then the store can be removed.
509 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
510 if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
511 if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
512 SI->getOperand(0) == DepLoad && isRemovable(SI)) {
513 DEBUG(dbgs() << "DSE: Remove Store Of Load from same pointer:\n "
514 << "LOAD: " << *DepLoad << "\n STORE: " << *SI << '\n');
516 // DeleteDeadInstruction can delete the current instruction. Save BBI
517 // in case we need it.
518 WeakVH NextInst(BBI);
520 DeleteDeadInstruction(SI, *MD);
522 if (NextInst == 0) // Next instruction deleted.
524 else if (BBI != BB.begin()) // Revisit this instruction if possible.
533 // Figure out what location is being stored to.
534 AliasAnalysis::Location Loc = getLocForWrite(Inst, *AA);
536 // If we didn't get a useful location, fail.
540 while (InstDep.isDef() || InstDep.isClobber()) {
541 // Get the memory clobbered by the instruction we depend on. MemDep will
542 // skip any instructions that 'Loc' clearly doesn't interact with. If we
543 // end up depending on a may- or must-aliased load, then we can't optimize
544 // away the store and we bail out. However, if we depend on on something
545 // that overwrites the memory location we *can* potentially optimize it.
547 // Find out what memory location the dependent instruction stores.
548 Instruction *DepWrite = InstDep.getInst();
549 AliasAnalysis::Location DepLoc = getLocForWrite(DepWrite, *AA);
550 // If we didn't get a useful location, or if it isn't a size, bail out.
554 // If we find a write that is a) removable (i.e., non-volatile), b) is
555 // completely obliterated by the store to 'Loc', and c) which we know that
556 // 'Inst' doesn't load from, then we can remove it.
557 if (isRemovable(DepWrite) &&
558 !isPossibleSelfRead(Inst, Loc, DepWrite, *AA)) {
559 int64_t InstWriteOffset, DepWriteOffset;
560 OverwriteResult OR = isOverwrite(Loc, DepLoc, *AA,
561 DepWriteOffset, InstWriteOffset);
562 if (OR == OverwriteComplete) {
563 DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: "
564 << *DepWrite << "\n KILLER: " << *Inst << '\n');
566 // Delete the store and now-dead instructions that feed it.
567 DeleteDeadInstruction(DepWrite, *MD);
571 // DeleteDeadInstruction can delete the current instruction in loop
574 if (BBI != BB.begin())
577 } else if (OR == OverwriteEnd && isShortenable(DepWrite)) {
578 // TODO: base this on the target vector size so that if the earlier
579 // store was too small to get vector writes anyway then its likely
580 // a good idea to shorten it
581 // Power of 2 vector writes are probably always a bad idea to optimize
582 // as any store/memset/memcpy is likely using vector instructions so
583 // shortening it to not vector size is likely to be slower
584 MemIntrinsic* DepIntrinsic = cast<MemIntrinsic>(DepWrite);
585 unsigned DepWriteAlign = DepIntrinsic->getAlignment();
586 if (llvm::isPowerOf2_64(InstWriteOffset) ||
587 ((DepWriteAlign != 0) && InstWriteOffset % DepWriteAlign == 0)) {
589 DEBUG(dbgs() << "DSE: Remove Dead Store:\n OW END: "
590 << *DepWrite << "\n KILLER (offset "
591 << InstWriteOffset << ", "
592 << DepLoc.Size << ")"
595 Value* DepWriteLength = DepIntrinsic->getLength();
596 Value* TrimmedLength = ConstantInt::get(DepWriteLength->getType(),
599 DepIntrinsic->setLength(TrimmedLength);
605 // If this is a may-aliased store that is clobbering the store value, we
606 // can keep searching past it for another must-aliased pointer that stores
607 // to the same location. For example, in:
611 // we can remove the first store to P even though we don't know if P and Q
613 if (DepWrite == &BB.front()) break;
615 // Can't look past this instruction if it might read 'Loc'.
616 if (AA->getModRefInfo(DepWrite, Loc) & AliasAnalysis::Ref)
619 InstDep = MD->getPointerDependencyFrom(Loc, false, DepWrite, &BB);
623 // If this block ends in a return, unwind, or unreachable, all allocas are
624 // dead at its end, which means stores to them are also dead.
625 if (BB.getTerminator()->getNumSuccessors() == 0)
626 MadeChange |= handleEndBlock(BB);
631 /// Find all blocks that will unconditionally lead to the block BB and append
633 static void FindUnconditionalPreds(SmallVectorImpl<BasicBlock *> &Blocks,
634 BasicBlock *BB, DominatorTree *DT) {
635 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
636 BasicBlock *Pred = *I;
637 TerminatorInst *PredTI = Pred->getTerminator();
638 if (PredTI->getNumSuccessors() != 1)
641 if (DT->isReachableFromEntry(Pred))
642 Blocks.push_back(Pred);
646 /// HandleFree - Handle frees of entire structures whose dependency is a store
647 /// to a field of that structure.
648 bool DSE::HandleFree(CallInst *F) {
649 bool MadeChange = false;
651 AliasAnalysis::Location Loc = AliasAnalysis::Location(F->getOperand(0));
652 SmallVector<BasicBlock *, 16> Blocks;
653 Blocks.push_back(F->getParent());
655 while (!Blocks.empty()) {
656 BasicBlock *BB = Blocks.pop_back_val();
657 Instruction *InstPt = BB->getTerminator();
658 if (BB == F->getParent()) InstPt = F;
660 MemDepResult Dep = MD->getPointerDependencyFrom(Loc, false, InstPt, BB);
661 while (Dep.isDef() || Dep.isClobber()) {
662 Instruction *Dependency = Dep.getInst();
663 if (!hasMemoryWrite(Dependency) || !isRemovable(Dependency))
667 GetUnderlyingObject(getStoredPointerOperand(Dependency));
669 // Check for aliasing.
670 if (!AA->isMustAlias(F->getArgOperand(0), DepPointer))
673 Instruction *Next = llvm::next(BasicBlock::iterator(Dependency));
675 // DCE instructions only used to calculate that store
676 DeleteDeadInstruction(Dependency, *MD);
680 // Inst's old Dependency is now deleted. Compute the next dependency,
681 // which may also be dead, as in
683 // s[1] = 0; // This has just been deleted.
685 Dep = MD->getPointerDependencyFrom(Loc, false, Next, BB);
688 if (Dep.isNonLocal())
689 FindUnconditionalPreds(Blocks, BB, DT);
695 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
696 /// function end block. Ex:
699 /// store i32 1, i32* %A
701 bool DSE::handleEndBlock(BasicBlock &BB) {
702 bool MadeChange = false;
704 // Keep track of all of the stack objects that are dead at the end of the
706 SmallPtrSet<Value*, 16> DeadStackObjects;
708 // Find all of the alloca'd pointers in the entry block.
709 BasicBlock *Entry = BB.getParent()->begin();
710 for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I) {
711 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
712 DeadStackObjects.insert(AI);
714 // Okay, so these are dead heap objects, but if the pointer never escapes
715 // then it's leaked by this function anyways.
716 if (CallInst *CI = extractMallocCall(I))
717 if (!PointerMayBeCaptured(CI, true, true))
718 DeadStackObjects.insert(CI);
721 // Treat byval arguments the same, stores to them are dead at the end of the
723 for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
724 AE = BB.getParent()->arg_end(); AI != AE; ++AI)
725 if (AI->hasByValAttr())
726 DeadStackObjects.insert(AI);
728 // Scan the basic block backwards
729 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
732 // If we find a store, check to see if it points into a dead stack value.
733 if (hasMemoryWrite(BBI) && isRemovable(BBI)) {
734 // See through pointer-to-pointer bitcasts
735 Value *Pointer = GetUnderlyingObject(getStoredPointerOperand(BBI));
737 // Stores to stack values are valid candidates for removal.
738 if (DeadStackObjects.count(Pointer)) {
739 Instruction *Dead = BBI++;
741 DEBUG(dbgs() << "DSE: Dead Store at End of Block:\n DEAD: "
742 << *Dead << "\n Object: " << *Pointer << '\n');
744 // DCE instructions only used to calculate that store.
745 DeleteDeadInstruction(Dead, *MD, &DeadStackObjects);
752 // Remove any dead non-memory-mutating instructions.
753 if (isInstructionTriviallyDead(BBI)) {
754 Instruction *Inst = BBI++;
755 DeleteDeadInstruction(Inst, *MD, &DeadStackObjects);
761 if (AllocaInst *A = dyn_cast<AllocaInst>(BBI)) {
762 DeadStackObjects.erase(A);
766 if (CallInst *CI = extractMallocCall(BBI)) {
767 DeadStackObjects.erase(CI);
771 if (CallSite CS = cast<Value>(BBI)) {
772 // If this call does not access memory, it can't be loading any of our
774 if (AA->doesNotAccessMemory(CS))
777 // If the call might load from any of our allocas, then any store above
779 SmallVector<Value*, 8> LiveAllocas;
780 for (SmallPtrSet<Value*, 16>::iterator I = DeadStackObjects.begin(),
781 E = DeadStackObjects.end(); I != E; ++I) {
782 // See if the call site touches it.
783 AliasAnalysis::ModRefResult A =
784 AA->getModRefInfo(CS, *I, getPointerSize(*I, *AA));
786 if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
787 LiveAllocas.push_back(*I);
790 for (SmallVector<Value*, 8>::iterator I = LiveAllocas.begin(),
791 E = LiveAllocas.end(); I != E; ++I)
792 DeadStackObjects.erase(*I);
794 // If all of the allocas were clobbered by the call then we're not going
795 // to find anything else to process.
796 if (DeadStackObjects.empty())
802 AliasAnalysis::Location LoadedLoc;
804 // If we encounter a use of the pointer, it is no longer considered dead
805 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
806 if (!L->isUnordered()) // Be conservative with atomic/volatile load
808 LoadedLoc = AA->getLocation(L);
809 } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) {
810 LoadedLoc = AA->getLocation(V);
811 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(BBI)) {
812 LoadedLoc = AA->getLocationForSource(MTI);
813 } else if (!BBI->mayReadFromMemory()) {
814 // Instruction doesn't read memory. Note that stores that weren't removed
815 // above will hit this case.
818 // Unknown inst; assume it clobbers everything.
822 // Remove any allocas from the DeadPointer set that are loaded, as this
823 // makes any stores above the access live.
824 RemoveAccessedObjects(LoadedLoc, DeadStackObjects);
826 // If all of the allocas were clobbered by the access then we're not going
827 // to find anything else to process.
828 if (DeadStackObjects.empty())
835 /// RemoveAccessedObjects - Check to see if the specified location may alias any
836 /// of the stack objects in the DeadStackObjects set. If so, they become live
837 /// because the location is being loaded.
838 void DSE::RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc,
839 SmallPtrSet<Value*, 16> &DeadStackObjects) {
840 const Value *UnderlyingPointer = GetUnderlyingObject(LoadedLoc.Ptr);
842 // A constant can't be in the dead pointer set.
843 if (isa<Constant>(UnderlyingPointer))
846 // If the kill pointer can be easily reduced to an alloca, don't bother doing
847 // extraneous AA queries.
848 if (isa<AllocaInst>(UnderlyingPointer) || isa<Argument>(UnderlyingPointer)) {
849 DeadStackObjects.erase(const_cast<Value*>(UnderlyingPointer));
853 SmallVector<Value*, 16> NowLive;
854 for (SmallPtrSet<Value*, 16>::iterator I = DeadStackObjects.begin(),
855 E = DeadStackObjects.end(); I != E; ++I) {
856 // See if the loaded location could alias the stack location.
857 AliasAnalysis::Location StackLoc(*I, getPointerSize(*I, *AA));
858 if (!AA->isNoAlias(StackLoc, LoadedLoc))
859 NowLive.push_back(*I);
862 for (SmallVector<Value*, 16>::iterator I = NowLive.begin(), E = NowLive.end();
864 DeadStackObjects.erase(*I);