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 dead store elimination that considers redundant stores
11 // within a basic-block as well as across basic blocks in a reverse CFG order.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/CaptureTracking.h"
21 #include "llvm/Analysis/GlobalsModRef.h"
22 #include "llvm/Analysis/MemoryBuiltins.h"
23 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
24 #include "llvm/Analysis/TargetLibraryInfo.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/Pass.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Transforms/Utils/Local.h"
39 #define DEBUG_TYPE "dse"
41 STATISTIC(NumRedundantStores, "Number of redundant stores deleted");
42 STATISTIC(NumFastStores, "Number of stores deleted");
43 STATISTIC(NumFastOther , "Number of other instrs removed");
44 STATISTIC(NumNonLocalStores, "Number of non-local stores deleted");
46 static cl::opt<bool> EnableNonLocalDSE("enable-nonlocal-dse", cl::init(false));
48 /// MaxNonLocalAttempts is an arbitrary threshold that provides
49 /// an early opportunitiy for bail out to control compile time.
50 static const unsigned MaxNonLocalAttempts = 100;
53 struct DSE : public FunctionPass {
55 MemoryDependenceAnalysis *MD;
57 const TargetLibraryInfo *TLI;
59 static char ID; // Pass identification, replacement for typeid
60 DSE() : FunctionPass(ID), AA(nullptr), MD(nullptr), DT(nullptr) {
61 initializeDSEPass(*PassRegistry::getPassRegistry());
64 bool runOnFunction(Function &F) override {
65 if (skipOptnoneFunction(F))
68 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
69 MD = &getAnalysis<MemoryDependenceAnalysis>();
70 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
71 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
74 for (BasicBlock &I : F)
75 // Only check non-dead blocks. Dead blocks may have strange pointer
76 // cycles that will confuse alias analysis.
77 if (DT->isReachableFromEntry(&I))
78 Changed |= runOnBasicBlock(I);
80 AA = nullptr; MD = nullptr; DT = nullptr;
84 bool runOnBasicBlock(BasicBlock &BB);
85 bool MemoryIsNotModifiedBetween(Instruction *FirstI, Instruction *SecondI);
86 bool HandleFree(CallInst *F);
87 bool handleNonLocalDependency(Instruction *Inst);
88 bool handleEndBlock(BasicBlock &BB);
89 void RemoveAccessedObjects(const MemoryLocation &LoadedLoc,
90 SmallSetVector<Value *, 16> &DeadStackObjects,
91 const DataLayout &DL);
93 void getAnalysisUsage(AnalysisUsage &AU) const override {
95 AU.addRequired<DominatorTreeWrapperPass>();
96 AU.addRequired<AAResultsWrapperPass>();
97 AU.addRequired<MemoryDependenceAnalysis>();
98 AU.addRequired<TargetLibraryInfoWrapperPass>();
99 AU.addPreserved<DominatorTreeWrapperPass>();
100 AU.addPreserved<GlobalsAAWrapperPass>();
101 AU.addPreserved<MemoryDependenceAnalysis>();
107 INITIALIZE_PASS_BEGIN(DSE, "dse", "Dead Store Elimination", false, false)
108 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
109 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
110 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
111 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
112 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
113 INITIALIZE_PASS_END(DSE, "dse", "Dead Store Elimination", false, false)
115 FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
117 //===----------------------------------------------------------------------===//
119 //===----------------------------------------------------------------------===//
121 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
122 /// and zero out all the operands of this instruction. If any of them become
123 /// dead, delete them and the computation tree that feeds them.
125 /// If ValueSet is non-null, remove any deleted instructions from it as well.
127 static void DeleteDeadInstruction(Instruction *I,
128 MemoryDependenceAnalysis &MD,
129 const TargetLibraryInfo &TLI,
130 SmallSetVector<Value*, 16> *ValueSet = nullptr) {
131 SmallVector<Instruction*, 32> NowDeadInsts;
133 NowDeadInsts.push_back(I);
136 // Before we touch this instruction, remove it from memdep!
138 Instruction *DeadInst = NowDeadInsts.pop_back_val();
141 // This instruction is dead, zap it, in stages. Start by removing it from
142 // MemDep, which needs to know the operands and needs it to be in the
144 MD.removeInstruction(DeadInst);
146 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
147 Value *Op = DeadInst->getOperand(op);
148 DeadInst->setOperand(op, nullptr);
150 // If this operand just became dead, add it to the NowDeadInsts list.
151 if (!Op->use_empty()) continue;
153 if (Instruction *OpI = dyn_cast<Instruction>(Op))
154 if (isInstructionTriviallyDead(OpI, &TLI))
155 NowDeadInsts.push_back(OpI);
158 DeadInst->eraseFromParent();
160 if (ValueSet) ValueSet->remove(DeadInst);
161 } while (!NowDeadInsts.empty());
165 /// hasMemoryWrite - Does this instruction write some memory? This only returns
166 /// true for things that we can analyze with other helpers below.
167 static bool hasMemoryWrite(Instruction *I, const TargetLibraryInfo &TLI) {
168 if (isa<StoreInst>(I))
170 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
171 switch (II->getIntrinsicID()) {
174 case Intrinsic::memset:
175 case Intrinsic::memmove:
176 case Intrinsic::memcpy:
177 case Intrinsic::init_trampoline:
178 case Intrinsic::lifetime_end:
182 if (auto CS = CallSite(I)) {
183 if (Function *F = CS.getCalledFunction()) {
184 if (TLI.has(LibFunc::strcpy) &&
185 F->getName() == TLI.getName(LibFunc::strcpy)) {
188 if (TLI.has(LibFunc::strncpy) &&
189 F->getName() == TLI.getName(LibFunc::strncpy)) {
192 if (TLI.has(LibFunc::strcat) &&
193 F->getName() == TLI.getName(LibFunc::strcat)) {
196 if (TLI.has(LibFunc::strncat) &&
197 F->getName() == TLI.getName(LibFunc::strncat)) {
205 /// getLocForWrite - Return a Location stored to by the specified instruction.
206 /// If isRemovable returns true, this function and getLocForRead completely
207 /// describe the memory operations for this instruction.
208 static MemoryLocation getLocForWrite(Instruction *Inst, AliasAnalysis &AA) {
209 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
210 return MemoryLocation::get(SI);
212 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Inst)) {
213 // memcpy/memmove/memset.
214 MemoryLocation Loc = MemoryLocation::getForDest(MI);
218 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
220 return MemoryLocation();
222 switch (II->getIntrinsicID()) {
224 return MemoryLocation(); // Unhandled intrinsic.
225 case Intrinsic::init_trampoline:
226 // FIXME: We don't know the size of the trampoline, so we can't really
228 return MemoryLocation(II->getArgOperand(0));
229 case Intrinsic::lifetime_end: {
230 uint64_t Len = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
231 return MemoryLocation(II->getArgOperand(1), Len);
236 /// getLocForRead - Return the location read by the specified "hasMemoryWrite"
237 /// instruction if any.
238 static MemoryLocation getLocForRead(Instruction *Inst,
239 const TargetLibraryInfo &TLI) {
240 assert(hasMemoryWrite(Inst, TLI) && "Unknown instruction case");
242 // The only instructions that both read and write are the mem transfer
243 // instructions (memcpy/memmove).
244 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Inst))
245 return MemoryLocation::getForSource(MTI);
246 return MemoryLocation();
250 /// isRemovable - If the value of this instruction and the memory it writes to
251 /// is unused, may we delete this instruction?
252 static bool isRemovable(Instruction *I) {
253 // Don't remove volatile/atomic stores.
254 if (StoreInst *SI = dyn_cast<StoreInst>(I))
255 return SI->isUnordered();
257 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
258 switch (II->getIntrinsicID()) {
259 default: llvm_unreachable("doesn't pass 'hasMemoryWrite' predicate");
260 case Intrinsic::lifetime_end:
261 // Never remove dead lifetime_end's, e.g. because it is followed by a
264 case Intrinsic::init_trampoline:
265 // Always safe to remove init_trampoline.
268 case Intrinsic::memset:
269 case Intrinsic::memmove:
270 case Intrinsic::memcpy:
271 // Don't remove volatile memory intrinsics.
272 return !cast<MemIntrinsic>(II)->isVolatile();
276 if (auto CS = CallSite(I))
277 return CS.getInstruction()->use_empty();
283 /// isShortenable - Returns true if this instruction can be safely shortened in
285 static bool isShortenable(Instruction *I) {
286 // Don't shorten stores for now
287 if (isa<StoreInst>(I))
290 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
291 switch (II->getIntrinsicID()) {
292 default: return false;
293 case Intrinsic::memset:
294 case Intrinsic::memcpy:
295 // Do shorten memory intrinsics.
300 // Don't shorten libcalls calls for now.
305 /// getStoredPointerOperand - Return the pointer that is being written to.
306 static Value *getStoredPointerOperand(Instruction *I) {
307 if (StoreInst *SI = dyn_cast<StoreInst>(I))
308 return SI->getPointerOperand();
309 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
310 return MI->getDest();
312 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
313 switch (II->getIntrinsicID()) {
314 default: llvm_unreachable("Unexpected intrinsic!");
315 case Intrinsic::init_trampoline:
316 return II->getArgOperand(0);
321 // All the supported functions so far happen to have dest as their first
323 return CS.getArgument(0);
326 static uint64_t getPointerSize(const Value *V, const DataLayout &DL,
327 const TargetLibraryInfo &TLI) {
329 if (getObjectSize(V, Size, DL, &TLI))
331 return MemoryLocation::UnknownSize;
343 /// isOverwrite - Return 'OverwriteComplete' if a store to the 'Later' location
344 /// completely overwrites a store to the 'Earlier' location.
345 /// 'OverwriteEnd' if the end of the 'Earlier' location is completely
346 /// overwritten by 'Later', or 'OverwriteUnknown' if nothing can be determined
347 static OverwriteResult isOverwrite(const MemoryLocation &Later,
348 const MemoryLocation &Earlier,
349 const DataLayout &DL,
350 const TargetLibraryInfo &TLI,
351 int64_t &EarlierOff, int64_t &LaterOff) {
352 const Value *P1 = Earlier.Ptr->stripPointerCasts();
353 const Value *P2 = Later.Ptr->stripPointerCasts();
355 // If the start pointers are the same, we just have to compare sizes to see if
356 // the later store was larger than the earlier store.
358 // If we don't know the sizes of either access, then we can't do a
360 if (Later.Size == MemoryLocation::UnknownSize ||
361 Earlier.Size == MemoryLocation::UnknownSize)
362 return OverwriteUnknown;
364 // Make sure that the Later size is >= the Earlier size.
365 if (Later.Size >= Earlier.Size)
366 return OverwriteComplete;
369 // Otherwise, we have to have size information, and the later store has to be
370 // larger than the earlier one.
371 if (Later.Size == MemoryLocation::UnknownSize ||
372 Earlier.Size == MemoryLocation::UnknownSize)
373 return OverwriteUnknown;
375 // Check to see if the later store is to the entire object (either a global,
376 // an alloca, or a byval/inalloca argument). If so, then it clearly
377 // overwrites any other store to the same object.
378 const Value *UO1 = GetUnderlyingObject(P1, DL),
379 *UO2 = GetUnderlyingObject(P2, DL);
381 // If we can't resolve the same pointers to the same object, then we can't
382 // analyze them at all.
384 return OverwriteUnknown;
386 // If the "Later" store is to a recognizable object, get its size.
387 uint64_t ObjectSize = getPointerSize(UO2, DL, TLI);
388 if (ObjectSize != MemoryLocation::UnknownSize)
389 if (ObjectSize == Later.Size && ObjectSize >= Earlier.Size)
390 return OverwriteComplete;
392 // Okay, we have stores to two completely different pointers. Try to
393 // decompose the pointer into a "base + constant_offset" form. If the base
394 // pointers are equal, then we can reason about the two stores.
397 const Value *BP1 = GetPointerBaseWithConstantOffset(P1, EarlierOff, DL);
398 const Value *BP2 = GetPointerBaseWithConstantOffset(P2, LaterOff, DL);
400 // If the base pointers still differ, we have two completely different stores.
402 return OverwriteUnknown;
404 // The later store completely overlaps the earlier store if:
406 // 1. Both start at the same offset and the later one's size is greater than
407 // or equal to the earlier one's, or
412 // 2. The earlier store has an offset greater than the later offset, but which
413 // still lies completely within the later store.
416 // |----- later ------|
418 // We have to be careful here as *Off is signed while *.Size is unsigned.
419 if (EarlierOff >= LaterOff &&
420 Later.Size >= Earlier.Size &&
421 uint64_t(EarlierOff - LaterOff) + Earlier.Size <= Later.Size)
422 return OverwriteComplete;
424 // The other interesting case is if the later store overwrites the end of
430 // In this case we may want to trim the size of earlier to avoid generating
431 // writes to addresses which will definitely be overwritten later
432 if (LaterOff > EarlierOff &&
433 LaterOff < int64_t(EarlierOff + Earlier.Size) &&
434 int64_t(LaterOff + Later.Size) >= int64_t(EarlierOff + Earlier.Size))
437 // Otherwise, they don't completely overlap.
438 return OverwriteUnknown;
441 /// isPossibleSelfRead - If 'Inst' might be a self read (i.e. a noop copy of a
442 /// memory region into an identical pointer) then it doesn't actually make its
443 /// input dead in the traditional sense. Consider this case:
448 /// In this case, the second store to A does not make the first store to A dead.
449 /// The usual situation isn't an explicit A<-A store like this (which can be
450 /// trivially removed) but a case where two pointers may alias.
452 /// This function detects when it is unsafe to remove a dependent instruction
453 /// because the DSE inducing instruction may be a self-read.
454 static bool isPossibleSelfRead(Instruction *Inst,
455 const MemoryLocation &InstStoreLoc,
456 Instruction *DepWrite,
457 const TargetLibraryInfo &TLI,
459 // Self reads can only happen for instructions that read memory. Get the
461 MemoryLocation InstReadLoc = getLocForRead(Inst, TLI);
462 if (!InstReadLoc.Ptr) return false; // Not a reading instruction.
464 // If the read and written loc obviously don't alias, it isn't a read.
465 if (AA.isNoAlias(InstReadLoc, InstStoreLoc)) return false;
467 // Okay, 'Inst' may copy over itself. However, we can still remove a the
468 // DepWrite instruction if we can prove that it reads from the same location
469 // as Inst. This handles useful cases like:
472 // Here we don't know if A/B may alias, but we do know that B/B are must
473 // aliases, so removing the first memcpy is safe (assuming it writes <= #
474 // bytes as the second one.
475 MemoryLocation DepReadLoc = getLocForRead(DepWrite, TLI);
477 if (DepReadLoc.Ptr && AA.isMustAlias(InstReadLoc.Ptr, DepReadLoc.Ptr))
480 // If DepWrite doesn't read memory or if we can't prove it is a must alias,
481 // then it can't be considered dead.
486 //===----------------------------------------------------------------------===//
488 //===----------------------------------------------------------------------===//
490 bool DSE::runOnBasicBlock(BasicBlock &BB) {
491 const DataLayout &DL = BB.getModule()->getDataLayout();
492 bool MadeChange = false;
493 unsigned NumNonLocalAttempts = 0;
495 // Do a top-down walk on the BB.
496 for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
497 Instruction *Inst = &*BBI++;
499 // Handle 'free' calls specially.
500 if (CallInst *F = isFreeCall(Inst, TLI)) {
501 MadeChange |= HandleFree(F);
505 // If we find something that writes memory, get its memory dependence.
506 if (!hasMemoryWrite(Inst, *TLI))
509 // If we're storing the same value back to a pointer that we just
510 // loaded from, then the store can be removed.
511 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
513 auto RemoveDeadInstAndUpdateBBI = [&](Instruction *DeadInst) {
514 // DeleteDeadInstruction can delete the current instruction. Save BBI
515 // in case we need it.
516 WeakVH NextInst(&*BBI);
518 DeleteDeadInstruction(DeadInst, *MD, *TLI);
520 if (!NextInst) // Next instruction deleted.
522 else if (BBI != BB.begin()) // Revisit this instruction if possible.
524 ++NumRedundantStores;
528 if (LoadInst *DepLoad = dyn_cast<LoadInst>(SI->getValueOperand())) {
529 if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
531 MemoryIsNotModifiedBetween(DepLoad, SI)) {
533 DEBUG(dbgs() << "DSE: Remove Store Of Load from same pointer:\n "
534 << "LOAD: " << *DepLoad << "\n STORE: " << *SI << '\n');
536 RemoveDeadInstAndUpdateBBI(SI);
541 // Remove null stores into the calloc'ed objects
542 Constant *StoredConstant = dyn_cast<Constant>(SI->getValueOperand());
544 if (StoredConstant && StoredConstant->isNullValue() &&
546 Instruction *UnderlyingPointer = dyn_cast<Instruction>(
547 GetUnderlyingObject(SI->getPointerOperand(), DL));
549 if (UnderlyingPointer && isCallocLikeFn(UnderlyingPointer, TLI) &&
550 MemoryIsNotModifiedBetween(UnderlyingPointer, SI)) {
552 << "DSE: Remove null store to the calloc'ed object:\n DEAD: "
553 << *Inst << "\n OBJECT: " << *UnderlyingPointer << '\n');
555 RemoveDeadInstAndUpdateBBI(SI);
561 MemDepResult InstDep = MD->getDependency(Inst);
563 if (InstDep.isDef() || InstDep.isClobber()) {
564 // Figure out what location is being stored to.
565 MemoryLocation Loc = getLocForWrite(Inst, *AA);
567 // If we didn't get a useful location, fail.
571 while (InstDep.isDef() || InstDep.isClobber()) {
572 // Get the memory clobbered by the instruction we depend on. MemDep
573 // will skip any instructions that 'Loc' clearly doesn't interact with.
574 // If we end up depending on a may- or must-aliased load, then we can't
575 // optimize away the store and we bail out. However, if we depend on on
576 // something that overwrites the memory location we *can* potentially
579 // Find out what memory location the dependent instruction stores.
580 Instruction *DepWrite = InstDep.getInst();
581 MemoryLocation DepLoc = getLocForWrite(DepWrite, *AA);
582 // If we didn't get a useful location, or if it isn't a size, bail out.
586 // If we find a write that is a) removable (i.e., non-volatile), b) is
587 // completely obliterated by the store to 'Loc', and c) which we know
588 // that 'Inst' doesn't load from, then we can remove it.
589 if (isRemovable(DepWrite) &&
590 !isPossibleSelfRead(Inst, Loc, DepWrite, *TLI, *AA)) {
591 int64_t InstWriteOffset, DepWriteOffset;
592 OverwriteResult OR = isOverwrite(Loc, DepLoc, DL, *TLI,
593 DepWriteOffset, InstWriteOffset);
594 if (OR == OverwriteComplete) {
595 DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *DepWrite
596 << "\n KILLER: " << *Inst << '\n');
598 // Delete the store and now-dead instructions that feed it.
599 DeleteDeadInstruction(DepWrite, *MD, *TLI);
603 // DeleteDeadInstruction can delete the current instruction in loop
605 BBI = Inst->getIterator();
606 if (BBI != BB.begin())
609 } else if (OR == OverwriteEnd && isShortenable(DepWrite)) {
610 // TODO: base this on the target vector size so that if the earlier
611 // store was too small to get vector writes anyway then its likely a
612 // good idea to shorten it.
614 // Power of 2 vector writes are probably always a bad idea to
615 // optimize as any store/memset/memcpy is likely using vector
616 // instructions so shortening it to not vector size is likely to be
618 MemIntrinsic *DepIntrinsic = cast<MemIntrinsic>(DepWrite);
619 unsigned DepWriteAlign = DepIntrinsic->getAlignment();
620 if (llvm::isPowerOf2_64(InstWriteOffset) ||
621 ((DepWriteAlign != 0) &&
622 InstWriteOffset % DepWriteAlign == 0)) {
624 DEBUG(dbgs() << "DSE: Remove Dead Store:\n OW END: " << *DepWrite
625 << "\n KILLER (offset " << InstWriteOffset << ", "
626 << DepLoc.Size << ")" << *Inst << '\n');
628 Value *DepWriteLength = DepIntrinsic->getLength();
629 Value *TrimmedLength = ConstantInt::get(
630 DepWriteLength->getType(), InstWriteOffset - DepWriteOffset);
631 DepIntrinsic->setLength(TrimmedLength);
637 // If this is a may-aliased store that is clobbering the store value, we
638 // can keep searching past it for another must-aliased pointer that
639 // stores to the same location. For example, in
643 // we can remove the first store to P even though we don't know if P and
645 if (DepWrite == &BB.front())
648 // Can't look past this instruction if it might read 'Loc'.
649 if (AA->getModRefInfo(DepWrite, Loc) & MRI_Ref)
652 InstDep = MD->getPointerDependencyFrom(Loc, false,
653 DepWrite->getIterator(), &BB);
655 } else if (EnableNonLocalDSE && InstDep.isNonLocal()) { // DSE across BB
656 if (++NumNonLocalAttempts < MaxNonLocalAttempts)
657 MadeChange |= handleNonLocalDependency(Inst);
661 // If this block ends in a return, unwind, or unreachable, all allocas are
662 // dead at its end, which means stores to them are also dead.
663 if (BB.getTerminator()->getNumSuccessors() == 0)
664 MadeChange |= handleEndBlock(BB);
669 /// A helper for handleNonLocalDependency() function to find all blocks
670 /// that lead to the input block BB and append them to the output PredBlocks.
671 /// PredBlocks will include not only predecessors of BB that unconditionally
672 /// lead to BB but also:
673 /// - single-block loops that lead to BB, and
674 /// - if-blocks for which one edge goes to BB and the other edge goes to
675 /// a block in the input SafeBlocks.
676 /// PredBlocks will not include blocks unreachable from the entry block, nor
677 /// blocks that form cycles with BB.
678 static void findSafePreds(SmallVectorImpl<BasicBlock *> &PredBlocks,
679 SmallSetVector<BasicBlock *, 8> &SafeBlocks,
680 BasicBlock *BB, DominatorTree *DT) {
681 for (auto *Pred : predecessors(BB)) {
684 // The second check below prevents adding blocks that form a cycle with BB
685 // in order to avoid potential problems due to MemoryDependenceAnalysis,
686 // isOverwrite, etc. being not loop-aware.
687 if (!DT->isReachableFromEntry(Pred) || DT->dominates(BB, Pred))
690 bool PredIsSafe = true;
691 for (auto *Succ : successors(Pred)) {
692 if (Succ == BB || Succ == Pred) // shortcut, BB should be in SafeBlocks
694 if (!SafeBlocks.count(Succ)) {
700 PredBlocks.push_back(Pred);
704 static bool underlyingObjectsDoNotAlias(StoreInst *SI, LoadInst *LI,
705 const DataLayout &DL,
707 Value *AObj = GetUnderlyingObject(SI->getPointerOperand(), DL);
708 SmallVector<Value *, 4> Pointers;
709 GetUnderlyingObjects(LI->getPointerOperand(), Pointers, DL);
711 for (auto *BObj : Pointers) {
712 if (!AA.isNoAlias(AObj, DL.getTypeStoreSize(AObj->getType()), BObj,
713 DL.getTypeStoreSize(BObj->getType())))
719 /// handleNonLocalDependency - Handle a non-local dependency on
720 /// the input instruction Inst located in BB in attempt to remove
721 /// redundant stores outside BB.
722 bool DSE::handleNonLocalDependency(Instruction *Inst) {
723 auto *SI = dyn_cast<StoreInst>(Inst);
726 // Get the location being stored to.
727 // If we don't get a useful location, bail out.
728 MemoryLocation Loc = getLocForWrite(SI, *AA);
732 bool MadeChange = false;
733 BasicBlock *BB = Inst->getParent();
734 const DataLayout &DL = BB->getModule()->getDataLayout();
736 // Worklist of predecessor blocks of BB
737 SmallVector<BasicBlock *, 8> Blocks;
738 // Keep track of all predecessor blocks that are safe to search through
739 SmallSetVector<BasicBlock *, 8> SafeBlocks;
740 SafeBlocks.insert(BB);
741 findSafePreds(Blocks, SafeBlocks, BB, DT);
743 while (!Blocks.empty()) {
744 BasicBlock *PB = Blocks.pop_back_val();
746 MD->getPointerDependencyFrom(Loc, false, PB->end(), PB, SI);
747 while (Dep.isDef() || Dep.isClobber()) {
748 Instruction *Dependency = Dep.getInst();
750 // Filter out false dependency from a load to SI looking through phis.
751 if (auto *LI = dyn_cast<LoadInst>(Dependency)) {
752 if (underlyingObjectsDoNotAlias(SI, LI, DL, *AA)) {
753 Dep = MD->getPointerDependencyFrom(Loc, false,
754 Dependency->getIterator(), PB, SI);
759 // If we don't get a useful location for the dependent instruction,
760 // it doesn't write memory, it is not removable, or it might read Loc,
762 MemoryLocation DepLoc = getLocForWrite(Dependency, *AA);
763 if (!DepLoc.Ptr || !hasMemoryWrite(Dependency, *TLI) ||
764 !isRemovable(Dependency) ||
765 (AA->getModRefInfo(Dependency, Loc) & MRI_Ref))
768 // Don't remove a store within single-block loops;
769 // we need more analysis: e.g. looking for an interferring load
770 // above the store within the loop, etc.
771 bool SingleBlockLoop = false;
772 for (auto I = succ_begin(PB), E = succ_end(PB); I != E; ++I) {
773 BasicBlock *Succ = *I;
775 SingleBlockLoop = true;
782 int64_t InstWriteOffset, DepWriteOffset;
784 isOverwrite(Loc, DepLoc, DL, *TLI, DepWriteOffset, InstWriteOffset);
785 if (OR == OverwriteComplete) {
786 DEBUG(dbgs() << "DSE: Remove Non-Local Dead Store:\n DEAD: "
787 << *Dependency << "\n KILLER: " << *SI << '\n');
789 // Delete redundant store and now-dead instructions that feed it.
790 auto Next = std::next(Dependency->getIterator());
791 DeleteDeadInstruction(Dependency, *MD, *TLI);
794 Dep = MD->getPointerDependencyFrom(Loc, false, Next, PB, SI);
797 // TODO: attempt shortening of Dependency inst as in the local case
801 if (Dep.isNonLocal()) {
802 SafeBlocks.insert(PB);
803 findSafePreds(Blocks, SafeBlocks, PB, DT);
810 /// Returns true if the memory which is accessed by the second instruction is not
811 /// modified between the first and the second instruction.
812 /// Precondition: Second instruction must be dominated by the first
814 bool DSE::MemoryIsNotModifiedBetween(Instruction *FirstI,
815 Instruction *SecondI) {
816 SmallVector<BasicBlock *, 16> WorkList;
817 SmallPtrSet<BasicBlock *, 8> Visited;
818 BasicBlock::iterator FirstBBI(FirstI);
820 BasicBlock::iterator SecondBBI(SecondI);
821 BasicBlock *FirstBB = FirstI->getParent();
822 BasicBlock *SecondBB = SecondI->getParent();
823 MemoryLocation MemLoc = MemoryLocation::get(SecondI);
825 // Start checking the store-block.
826 WorkList.push_back(SecondBB);
827 bool isFirstBlock = true;
829 // Check all blocks going backward until we reach the load-block.
830 while (!WorkList.empty()) {
831 BasicBlock *B = WorkList.pop_back_val();
833 // Ignore instructions before LI if this is the FirstBB.
834 BasicBlock::iterator BI = (B == FirstBB ? FirstBBI : B->begin());
836 BasicBlock::iterator EI;
838 // Ignore instructions after SI if this is the first visit of SecondBB.
839 assert(B == SecondBB && "first block is not the store block");
841 isFirstBlock = false;
843 // It's not SecondBB or (in case of a loop) the second visit of SecondBB.
844 // In this case we also have to look at instructions after SI.
847 for (; BI != EI; ++BI) {
848 Instruction *I = &*BI;
849 if (I->mayWriteToMemory() && I != SecondI) {
850 auto Res = AA->getModRefInfo(I, MemLoc);
851 if (Res != MRI_NoModRef)
856 assert(B != &FirstBB->getParent()->getEntryBlock() &&
857 "Should not hit the entry block because SI must be dominated by LI");
858 for (auto *PredI : predecessors(B)) {
859 if (!Visited.insert(PredI).second)
861 WorkList.push_back(PredI);
868 /// Find all blocks that will unconditionally lead to the block BB and append
870 static void FindUnconditionalPreds(SmallVectorImpl<BasicBlock *> &Blocks,
871 BasicBlock *BB, DominatorTree *DT) {
872 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
873 BasicBlock *Pred = *I;
874 if (Pred == BB) continue;
875 TerminatorInst *PredTI = Pred->getTerminator();
876 if (PredTI->getNumSuccessors() != 1)
879 if (DT->isReachableFromEntry(Pred))
880 Blocks.push_back(Pred);
884 /// HandleFree - Handle frees of entire structures whose dependency is a store
885 /// to a field of that structure.
886 bool DSE::HandleFree(CallInst *F) {
887 bool MadeChange = false;
889 MemoryLocation Loc = MemoryLocation(F->getOperand(0));
890 SmallVector<BasicBlock *, 16> Blocks;
891 Blocks.push_back(F->getParent());
892 const DataLayout &DL = F->getModule()->getDataLayout();
894 while (!Blocks.empty()) {
895 BasicBlock *BB = Blocks.pop_back_val();
896 Instruction *InstPt = BB->getTerminator();
897 if (BB == F->getParent()) InstPt = F;
900 MD->getPointerDependencyFrom(Loc, false, InstPt->getIterator(), BB);
901 while (Dep.isDef() || Dep.isClobber()) {
902 Instruction *Dependency = Dep.getInst();
903 if (!hasMemoryWrite(Dependency, *TLI) || !isRemovable(Dependency))
907 GetUnderlyingObject(getStoredPointerOperand(Dependency), DL);
909 // Check for aliasing.
910 if (!AA->isMustAlias(F->getArgOperand(0), DepPointer))
913 auto Next = ++Dependency->getIterator();
915 // DCE instructions only used to calculate that store
916 DeleteDeadInstruction(Dependency, *MD, *TLI);
920 // Inst's old Dependency is now deleted. Compute the next dependency,
921 // which may also be dead, as in
923 // s[1] = 0; // This has just been deleted.
925 Dep = MD->getPointerDependencyFrom(Loc, false, Next, BB);
928 if (Dep.isNonLocal())
929 FindUnconditionalPreds(Blocks, BB, DT);
935 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
936 /// function end block. Ex:
939 /// store i32 1, i32* %A
941 bool DSE::handleEndBlock(BasicBlock &BB) {
942 bool MadeChange = false;
944 // Keep track of all of the stack objects that are dead at the end of the
946 SmallSetVector<Value*, 16> DeadStackObjects;
948 // Find all of the alloca'd pointers in the entry block.
949 BasicBlock &Entry = BB.getParent()->front();
950 for (Instruction &I : Entry) {
951 if (isa<AllocaInst>(&I))
952 DeadStackObjects.insert(&I);
954 // Okay, so these are dead heap objects, but if the pointer never escapes
955 // then it's leaked by this function anyways.
956 else if (isAllocLikeFn(&I, TLI) && !PointerMayBeCaptured(&I, true, true))
957 DeadStackObjects.insert(&I);
960 // Treat byval or inalloca arguments the same, stores to them are dead at the
961 // end of the function.
962 for (Argument &AI : BB.getParent()->args())
963 if (AI.hasByValOrInAllocaAttr())
964 DeadStackObjects.insert(&AI);
966 const DataLayout &DL = BB.getModule()->getDataLayout();
968 // Scan the basic block backwards
969 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
972 // If we find a store, check to see if it points into a dead stack value.
973 if (hasMemoryWrite(&*BBI, *TLI) && isRemovable(&*BBI)) {
974 // See through pointer-to-pointer bitcasts
975 SmallVector<Value *, 4> Pointers;
976 GetUnderlyingObjects(getStoredPointerOperand(&*BBI), Pointers, DL);
978 // Stores to stack values are valid candidates for removal.
980 for (SmallVectorImpl<Value *>::iterator I = Pointers.begin(),
981 E = Pointers.end(); I != E; ++I)
982 if (!DeadStackObjects.count(*I)) {
988 Instruction *Dead = &*BBI++;
990 DEBUG(dbgs() << "DSE: Dead Store at End of Block:\n DEAD: "
991 << *Dead << "\n Objects: ";
992 for (SmallVectorImpl<Value *>::iterator I = Pointers.begin(),
993 E = Pointers.end(); I != E; ++I) {
995 if (std::next(I) != E)
1000 // DCE instructions only used to calculate that store.
1001 DeleteDeadInstruction(Dead, *MD, *TLI, &DeadStackObjects);
1008 // Remove any dead non-memory-mutating instructions.
1009 if (isInstructionTriviallyDead(&*BBI, TLI)) {
1010 Instruction *Inst = &*BBI++;
1011 DeleteDeadInstruction(Inst, *MD, *TLI, &DeadStackObjects);
1017 if (isa<AllocaInst>(BBI)) {
1018 // Remove allocas from the list of dead stack objects; there can't be
1019 // any references before the definition.
1020 DeadStackObjects.remove(&*BBI);
1024 if (auto CS = CallSite(&*BBI)) {
1025 // Remove allocation function calls from the list of dead stack objects;
1026 // there can't be any references before the definition.
1027 if (isAllocLikeFn(&*BBI, TLI))
1028 DeadStackObjects.remove(&*BBI);
1030 // If this call does not access memory, it can't be loading any of our
1032 if (AA->doesNotAccessMemory(CS))
1035 // If the call might load from any of our allocas, then any store above
1036 // the call is live.
1037 DeadStackObjects.remove_if([&](Value *I) {
1038 // See if the call site touches the value.
1039 ModRefInfo A = AA->getModRefInfo(CS, I, getPointerSize(I, DL, *TLI));
1041 return A == MRI_ModRef || A == MRI_Ref;
1044 // If all of the allocas were clobbered by the call then we're not going
1045 // to find anything else to process.
1046 if (DeadStackObjects.empty())
1052 MemoryLocation LoadedLoc;
1054 // If we encounter a use of the pointer, it is no longer considered dead
1055 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
1056 if (!L->isUnordered()) // Be conservative with atomic/volatile load
1058 LoadedLoc = MemoryLocation::get(L);
1059 } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) {
1060 LoadedLoc = MemoryLocation::get(V);
1061 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(BBI)) {
1062 LoadedLoc = MemoryLocation::getForSource(MTI);
1063 } else if (!BBI->mayReadFromMemory()) {
1064 // Instruction doesn't read memory. Note that stores that weren't removed
1065 // above will hit this case.
1068 // Unknown inst; assume it clobbers everything.
1072 // Remove any allocas from the DeadPointer set that are loaded, as this
1073 // makes any stores above the access live.
1074 RemoveAccessedObjects(LoadedLoc, DeadStackObjects, DL);
1076 // If all of the allocas were clobbered by the access then we're not going
1077 // to find anything else to process.
1078 if (DeadStackObjects.empty())
1085 /// RemoveAccessedObjects - Check to see if the specified location may alias any
1086 /// of the stack objects in the DeadStackObjects set. If so, they become live
1087 /// because the location is being loaded.
1088 void DSE::RemoveAccessedObjects(const MemoryLocation &LoadedLoc,
1089 SmallSetVector<Value *, 16> &DeadStackObjects,
1090 const DataLayout &DL) {
1091 const Value *UnderlyingPointer = GetUnderlyingObject(LoadedLoc.Ptr, DL);
1093 // A constant can't be in the dead pointer set.
1094 if (isa<Constant>(UnderlyingPointer))
1097 // If the kill pointer can be easily reduced to an alloca, don't bother doing
1098 // extraneous AA queries.
1099 if (isa<AllocaInst>(UnderlyingPointer) || isa<Argument>(UnderlyingPointer)) {
1100 DeadStackObjects.remove(const_cast<Value*>(UnderlyingPointer));
1104 // Remove objects that could alias LoadedLoc.
1105 DeadStackObjects.remove_if([&](Value *I) {
1106 // See if the loaded location could alias the stack location.
1107 MemoryLocation StackLoc(I, getPointerSize(I, DL, *TLI));
1108 return !AA->isNoAlias(StackLoc, LoadedLoc);