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/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Analysis/AliasAnalysis.h"
29 #include "llvm/Analysis/Dominators.h"
30 #include "llvm/Analysis/MemoryBuiltins.h"
31 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/Transforms/Utils/Local.h"
37 STATISTIC(NumFastStores, "Number of stores deleted");
38 STATISTIC(NumFastOther , "Number of other instrs removed");
41 struct DSE : public FunctionPass {
43 MemoryDependenceAnalysis *MD;
45 static char ID; // Pass identification, replacement for typeid
46 DSE() : FunctionPass(ID), AA(0), MD(0) {
47 initializeDSEPass(*PassRegistry::getPassRegistry());
50 virtual bool runOnFunction(Function &F) {
51 AA = &getAnalysis<AliasAnalysis>();
52 MD = &getAnalysis<MemoryDependenceAnalysis>();
53 DominatorTree &DT = getAnalysis<DominatorTree>();
56 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
57 // Only check non-dead blocks. Dead blocks may have strange pointer
58 // cycles that will confuse alias analysis.
59 if (DT.isReachableFromEntry(I))
60 Changed |= runOnBasicBlock(*I);
66 bool runOnBasicBlock(BasicBlock &BB);
67 bool HandleFree(CallInst *F);
68 bool handleEndBlock(BasicBlock &BB);
69 void RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc,
70 SmallPtrSet<Value*, 16> &DeadStackObjects);
73 // getAnalysisUsage - We require post dominance frontiers (aka Control
75 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
77 AU.addRequired<DominatorTree>();
78 AU.addRequired<AliasAnalysis>();
79 AU.addRequired<MemoryDependenceAnalysis>();
80 AU.addPreserved<AliasAnalysis>();
81 AU.addPreserved<DominatorTree>();
82 AU.addPreserved<MemoryDependenceAnalysis>();
88 INITIALIZE_PASS_BEGIN(DSE, "dse", "Dead Store Elimination", false, false)
89 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
90 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
91 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
92 INITIALIZE_PASS_END(DSE, "dse", "Dead Store Elimination", false, false)
94 FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
96 //===----------------------------------------------------------------------===//
98 //===----------------------------------------------------------------------===//
100 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
101 /// and zero out all the operands of this instruction. If any of them become
102 /// dead, delete them and the computation tree that feeds them.
104 /// If ValueSet is non-null, remove any deleted instructions from it as well.
106 static void DeleteDeadInstruction(Instruction *I,
107 MemoryDependenceAnalysis &MD,
108 SmallPtrSet<Value*, 16> *ValueSet = 0) {
109 SmallVector<Instruction*, 32> NowDeadInsts;
111 NowDeadInsts.push_back(I);
114 // Before we touch this instruction, remove it from memdep!
116 Instruction *DeadInst = NowDeadInsts.pop_back_val();
119 // This instruction is dead, zap it, in stages. Start by removing it from
120 // MemDep, which needs to know the operands and needs it to be in the
122 MD.removeInstruction(DeadInst);
124 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
125 Value *Op = DeadInst->getOperand(op);
126 DeadInst->setOperand(op, 0);
128 // If this operand just became dead, add it to the NowDeadInsts list.
129 if (!Op->use_empty()) continue;
131 if (Instruction *OpI = dyn_cast<Instruction>(Op))
132 if (isInstructionTriviallyDead(OpI))
133 NowDeadInsts.push_back(OpI);
136 DeadInst->eraseFromParent();
138 if (ValueSet) ValueSet->erase(DeadInst);
139 } while (!NowDeadInsts.empty());
143 /// hasMemoryWrite - Does this instruction write some memory? This only returns
144 /// true for things that we can analyze with other helpers below.
145 static bool hasMemoryWrite(Instruction *I) {
146 if (isa<StoreInst>(I))
148 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
149 switch (II->getIntrinsicID()) {
152 case Intrinsic::memset:
153 case Intrinsic::memmove:
154 case Intrinsic::memcpy:
155 case Intrinsic::init_trampoline:
156 case Intrinsic::lifetime_end:
163 /// getLocForWrite - Return a Location stored to by the specified instruction.
164 static AliasAnalysis::Location
165 getLocForWrite(Instruction *Inst, AliasAnalysis &AA) {
166 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
167 return AA.getLocation(SI);
169 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Inst)) {
170 // memcpy/memmove/memset.
171 AliasAnalysis::Location Loc = AA.getLocationForDest(MI);
172 // If we don't have target data around, an unknown size in Location means
173 // that we should use the size of the pointee type. This isn't valid for
174 // memset/memcpy, which writes more than an i8.
175 if (Loc.Size == AliasAnalysis::UnknownSize && AA.getTargetData() == 0)
176 return AliasAnalysis::Location();
180 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
181 if (II == 0) return AliasAnalysis::Location();
183 switch (II->getIntrinsicID()) {
184 default: return AliasAnalysis::Location(); // Unhandled intrinsic.
185 case Intrinsic::init_trampoline:
186 // If we don't have target data around, an unknown size in Location means
187 // that we should use the size of the pointee type. This isn't valid for
188 // init.trampoline, which writes more than an i8.
189 if (AA.getTargetData() == 0) return AliasAnalysis::Location();
191 // FIXME: We don't know the size of the trampoline, so we can't really
193 return AliasAnalysis::Location(II->getArgOperand(0));
194 case Intrinsic::lifetime_end: {
195 uint64_t Len = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
196 return AliasAnalysis::Location(II->getArgOperand(1), Len);
201 /// isRemovable - If the value of this instruction and the memory it writes to
202 /// is unused, may we delete this instruction?
203 static bool isRemovable(Instruction *I) {
204 // Don't remove volatile stores.
205 if (StoreInst *SI = dyn_cast<StoreInst>(I))
206 return !SI->isVolatile();
208 IntrinsicInst *II = cast<IntrinsicInst>(I);
209 switch (II->getIntrinsicID()) {
210 default: assert(0 && "doesn't pass 'hasMemoryWrite' predicate");
211 case Intrinsic::lifetime_end:
212 // Never remove dead lifetime_end's, e.g. because it is followed by a
215 case Intrinsic::init_trampoline:
216 // Always safe to remove init_trampoline.
219 case Intrinsic::memset:
220 case Intrinsic::memmove:
221 case Intrinsic::memcpy:
222 // Don't remove volatile memory intrinsics.
223 return !cast<MemIntrinsic>(II)->isVolatile();
227 /// getStoredPointerOperand - Return the pointer that is being written to.
228 static Value *getStoredPointerOperand(Instruction *I) {
229 if (StoreInst *SI = dyn_cast<StoreInst>(I))
230 return SI->getPointerOperand();
231 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
232 return MI->getDest();
234 IntrinsicInst *II = cast<IntrinsicInst>(I);
235 switch (II->getIntrinsicID()) {
236 default: assert(false && "Unexpected intrinsic!");
237 case Intrinsic::init_trampoline:
238 return II->getArgOperand(0);
242 static uint64_t getPointerSize(Value *V, AliasAnalysis &AA) {
243 const TargetData *TD = AA.getTargetData();
245 return AliasAnalysis::UnknownSize;
247 if (AllocaInst *A = dyn_cast<AllocaInst>(V)) {
248 // Get size information for the alloca
249 if (ConstantInt *C = dyn_cast<ConstantInt>(A->getArraySize()))
250 return C->getZExtValue() * TD->getTypeAllocSize(A->getAllocatedType());
251 return AliasAnalysis::UnknownSize;
254 assert(isa<Argument>(V) && "Expected AllocaInst or Argument!");
255 const PointerType *PT = cast<PointerType>(V->getType());
256 return TD->getTypeAllocSize(PT->getElementType());
259 /// isObjectPointerWithTrustworthySize - Return true if the specified Value* is
260 /// pointing to an object with a pointer size we can trust.
261 static bool isObjectPointerWithTrustworthySize(const Value *V) {
262 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
263 return !AI->isArrayAllocation();
264 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
265 return !GV->isWeakForLinker();
266 if (const Argument *A = dyn_cast<Argument>(V))
267 return A->hasByValAttr();
271 /// isCompleteOverwrite - Return true if a store to the 'Later' location
272 /// completely overwrites a store to the 'Earlier' location.
273 static bool isCompleteOverwrite(const AliasAnalysis::Location &Later,
274 const AliasAnalysis::Location &Earlier,
276 const Value *P1 = Earlier.Ptr->stripPointerCasts();
277 const Value *P2 = Later.Ptr->stripPointerCasts();
279 // If the start pointers are the same, we just have to compare sizes to see if
280 // the later store was larger than the earlier store.
282 // If we don't know the sizes of either access, then we can't do a
284 if (Later.Size == AliasAnalysis::UnknownSize ||
285 Earlier.Size == AliasAnalysis::UnknownSize) {
286 // If we have no TargetData information around, then the size of the store
287 // is inferrable from the pointee type. If they are the same type, then
288 // we know that the store is safe.
289 if (AA.getTargetData() == 0)
290 return Later.Ptr->getType() == Earlier.Ptr->getType();
294 // Make sure that the Later size is >= the Earlier size.
295 if (Later.Size < Earlier.Size)
300 // Otherwise, we have to have size information, and the later store has to be
301 // larger than the earlier one.
302 if (Later.Size == AliasAnalysis::UnknownSize ||
303 Earlier.Size == AliasAnalysis::UnknownSize ||
304 Later.Size <= Earlier.Size || AA.getTargetData() == 0)
307 // Check to see if the later store is to the entire object (either a global,
308 // an alloca, or a byval argument). If so, then it clearly overwrites any
309 // other store to the same object.
310 const TargetData &TD = *AA.getTargetData();
312 const Value *UO1 = P1->getUnderlyingObject(), *UO2 = P2->getUnderlyingObject();
314 // If we can't resolve the same pointers to the same object, then we can't
315 // analyze them at all.
319 // If the "Later" store is to a recognizable object, get its size.
320 if (isObjectPointerWithTrustworthySize(UO2)) {
321 uint64_t ObjectSize =
322 TD.getTypeAllocSize(cast<PointerType>(UO2->getType())->getElementType());
323 if (ObjectSize == Later.Size)
327 // Okay, we have stores to two completely different pointers. Try to
328 // decompose the pointer into a "base + constant_offset" form. If the base
329 // pointers are equal, then we can reason about the two stores.
330 int64_t Off1 = 0, Off2 = 0;
331 const Value *BP1 = GetPointerBaseWithConstantOffset(P1, Off1, TD);
332 const Value *BP2 = GetPointerBaseWithConstantOffset(P2, Off2, TD);
334 // If the base pointers still differ, we have two completely different stores.
338 // Otherwise, we might have a situation like:
339 // store i16 -> P + 1 Byte
341 // In this case, we see if the later store completely overlaps all bytes
342 // stored by the previous store.
343 if (Off1 < Off2 || // Earlier starts before Later.
344 Off1+Earlier.Size > Off2+Later.Size) // Earlier goes beyond Later.
346 // Otherwise, we have complete overlap.
351 //===----------------------------------------------------------------------===//
353 //===----------------------------------------------------------------------===//
355 bool DSE::runOnBasicBlock(BasicBlock &BB) {
356 bool MadeChange = false;
358 // Do a top-down walk on the BB.
359 for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
360 Instruction *Inst = BBI++;
362 // Handle 'free' calls specially.
363 if (CallInst *F = isFreeCall(Inst)) {
364 MadeChange |= HandleFree(F);
368 // If we find something that writes memory, get its memory dependence.
369 if (!hasMemoryWrite(Inst))
372 MemDepResult InstDep = MD->getDependency(Inst);
374 // Ignore non-local store liveness.
375 // FIXME: cross-block DSE would be fun. :)
376 if (InstDep.isNonLocal() ||
377 // Ignore self dependence, which happens in the entry block of the
379 InstDep.getInst() == Inst)
382 // If we're storing the same value back to a pointer that we just
383 // loaded from, then the store can be removed.
384 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
385 if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
386 if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
387 SI->getOperand(0) == DepLoad && !SI->isVolatile()) {
388 // DeleteDeadInstruction can delete the current instruction. Save BBI
389 // in case we need it.
390 WeakVH NextInst(BBI);
392 DeleteDeadInstruction(SI, *MD);
394 if (NextInst == 0) // Next instruction deleted.
396 else if (BBI != BB.begin()) // Revisit this instruction if possible.
405 // Figure out what location is being stored to.
406 AliasAnalysis::Location Loc = getLocForWrite(Inst, *AA);
408 // If we didn't get a useful location, fail.
412 while (!InstDep.isNonLocal()) {
413 // Get the memory clobbered by the instruction we depend on. MemDep will
414 // skip any instructions that 'Loc' clearly doesn't interact with. If we
415 // end up depending on a may- or must-aliased load, then we can't optimize
416 // away the store and we bail out. However, if we depend on on something
417 // that overwrites the memory location we *can* potentially optimize it.
419 // Find out what memory location the dependant instruction stores.
420 Instruction *DepWrite = InstDep.getInst();
421 AliasAnalysis::Location DepLoc = getLocForWrite(DepWrite, *AA);
422 // If we didn't get a useful location, or if it isn't a size, bail out.
426 // If we find a removable write that is completely obliterated by the
427 // store to 'Loc' then we can remove it.
428 if (isRemovable(DepWrite) && isCompleteOverwrite(Loc, DepLoc, *AA)) {
429 // Delete the store and now-dead instructions that feed it.
430 DeleteDeadInstruction(DepWrite, *MD);
434 // DeleteDeadInstruction can delete the current instruction in loop
437 if (BBI != BB.begin())
442 // If this is a may-aliased store that is clobbering the store value, we
443 // can keep searching past it for another must-aliased pointer that stores
444 // to the same location. For example, in:
448 // we can remove the first store to P even though we don't know if P and Q
450 if (DepWrite == &BB.front()) break;
452 // Can't look past this instruction if it might read 'Loc'.
453 if (AA->getModRefInfo(DepWrite, Loc) & AliasAnalysis::Ref)
456 InstDep = MD->getPointerDependencyFrom(Loc, false, DepWrite, &BB);
460 // If this block ends in a return, unwind, or unreachable, all allocas are
461 // dead at its end, which means stores to them are also dead.
462 if (BB.getTerminator()->getNumSuccessors() == 0)
463 MadeChange |= handleEndBlock(BB);
468 /// HandleFree - Handle frees of entire structures whose dependency is a store
469 /// to a field of that structure.
470 bool DSE::HandleFree(CallInst *F) {
471 MemDepResult Dep = MD->getDependency(F);
473 if (Dep.isNonLocal()) return false;
475 Instruction *Dependency = Dep.getInst();
476 if (!hasMemoryWrite(Dependency) || !isRemovable(Dependency))
480 getStoredPointerOperand(Dependency)->getUnderlyingObject();
482 // Check for aliasing.
483 if (AA->alias(F->getArgOperand(0), 1, DepPointer, 1) !=
484 AliasAnalysis::MustAlias)
487 // DCE instructions only used to calculate that store
488 DeleteDeadInstruction(Dependency, *MD);
491 // Inst's old Dependency is now deleted. Compute the next dependency,
492 // which may also be dead, as in
494 // s[1] = 0; // This has just been deleted.
496 Dep = MD->getDependency(F);
497 } while (!Dep.isNonLocal());
502 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
503 /// function end block. Ex:
506 /// store i32 1, i32* %A
508 bool DSE::handleEndBlock(BasicBlock &BB) {
509 bool MadeChange = false;
511 // Keep track of all of the stack objects that are dead at the end of the
513 SmallPtrSet<Value*, 16> DeadStackObjects;
515 // Find all of the alloca'd pointers in the entry block.
516 BasicBlock *Entry = BB.getParent()->begin();
517 for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I)
518 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
519 DeadStackObjects.insert(AI);
521 // Treat byval arguments the same, stores to them are dead at the end of the
523 for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
524 AE = BB.getParent()->arg_end(); AI != AE; ++AI)
525 if (AI->hasByValAttr())
526 DeadStackObjects.insert(AI);
528 // Scan the basic block backwards
529 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
532 // If we find a store, check to see if it points into a dead stack value.
533 if (hasMemoryWrite(BBI) && isRemovable(BBI)) {
534 // See through pointer-to-pointer bitcasts
535 Value *Pointer = getStoredPointerOperand(BBI)->getUnderlyingObject();
537 // Stores to stack values are valid candidates for removal.
538 if (DeadStackObjects.count(Pointer)) {
539 // DCE instructions only used to calculate that store.
540 Instruction *Dead = BBI++;
541 DeleteDeadInstruction(Dead, *MD, &DeadStackObjects);
548 // Remove any dead non-memory-mutating instructions.
549 if (isInstructionTriviallyDead(BBI)) {
550 Instruction *Inst = BBI++;
551 DeleteDeadInstruction(Inst, *MD, &DeadStackObjects);
557 if (AllocaInst *A = dyn_cast<AllocaInst>(BBI)) {
558 DeadStackObjects.erase(A);
562 if (CallSite CS = cast<Value>(BBI)) {
563 // If this call does not access memory, it can't be loading any of our
565 if (AA->doesNotAccessMemory(CS))
568 unsigned NumModRef = 0, NumOther = 0;
570 // If the call might load from any of our allocas, then any store above
572 SmallVector<Value*, 8> LiveAllocas;
573 for (SmallPtrSet<Value*, 16>::iterator I = DeadStackObjects.begin(),
574 E = DeadStackObjects.end(); I != E; ++I) {
575 // If we detect that our AA is imprecise, it's not worth it to scan the
576 // rest of the DeadPointers set. Just assume that the AA will return
577 // ModRef for everything, and go ahead and bail out.
578 if (NumModRef >= 16 && NumOther == 0)
581 // See if the call site touches it.
582 AliasAnalysis::ModRefResult A =
583 AA->getModRefInfo(CS, *I, getPointerSize(*I, *AA));
585 if (A == AliasAnalysis::ModRef)
590 if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
591 LiveAllocas.push_back(*I);
594 for (SmallVector<Value*, 8>::iterator I = LiveAllocas.begin(),
595 E = LiveAllocas.end(); I != E; ++I)
596 DeadStackObjects.erase(*I);
598 // If all of the allocas were clobbered by the call then we're not going
599 // to find anything else to process.
600 if (DeadStackObjects.empty())
606 AliasAnalysis::Location LoadedLoc;
608 // If we encounter a use of the pointer, it is no longer considered dead
609 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
610 LoadedLoc = AA->getLocation(L);
611 } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) {
612 LoadedLoc = AA->getLocation(V);
613 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(BBI)) {
614 LoadedLoc = AA->getLocationForSource(MTI);
616 // Not a loading instruction.
620 // Remove any allocas from the DeadPointer set that are loaded, as this
621 // makes any stores above the access live.
622 RemoveAccessedObjects(LoadedLoc, DeadStackObjects);
624 // If all of the allocas were clobbered by the access then we're not going
625 // to find anything else to process.
626 if (DeadStackObjects.empty())
633 /// RemoveAccessedObjects - Check to see if the specified location may alias any
634 /// of the stack objects in the DeadStackObjects set. If so, they become live
635 /// because the location is being loaded.
636 void DSE::RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc,
637 SmallPtrSet<Value*, 16> &DeadStackObjects) {
638 const Value *UnderlyingPointer = LoadedLoc.Ptr->getUnderlyingObject();
640 // A constant can't be in the dead pointer set.
641 if (isa<Constant>(UnderlyingPointer))
644 // If the kill pointer can be easily reduced to an alloca, don't bother doing
645 // extraneous AA queries.
646 if (isa<AllocaInst>(UnderlyingPointer) || isa<Argument>(UnderlyingPointer)) {
647 DeadStackObjects.erase(const_cast<Value*>(UnderlyingPointer));
651 SmallVector<Value*, 16> NowLive;
652 for (SmallPtrSet<Value*, 16>::iterator I = DeadStackObjects.begin(),
653 E = DeadStackObjects.end(); I != E; ++I) {
654 // See if the loaded location could alias the stack location.
655 AliasAnalysis::Location StackLoc(*I, getPointerSize(*I, *AA));
656 if (!AA->isNoAlias(StackLoc, LoadedLoc))
657 NowLive.push_back(*I);
660 for (SmallVector<Value*, 16>::iterator I = NowLive.begin(), E = NowLive.end();
662 DeadStackObjects.erase(*I);