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/Instructions.h"
23 #include "llvm/IntrinsicInst.h"
24 #include "llvm/Pass.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/Dominators.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
31 #include "llvm/Target/TargetData.h"
32 #include "llvm/Transforms/Utils/Local.h"
35 STATISTIC(NumFastStores, "Number of stores deleted");
36 STATISTIC(NumFastOther , "Number of other instrs removed");
39 struct DSE : public FunctionPass {
41 MemoryDependenceAnalysis *MD;
43 static char ID; // Pass identification, replacement for typeid
44 DSE() : FunctionPass(ID), AA(0), MD(0) {
45 initializeDSEPass(*PassRegistry::getPassRegistry());
48 virtual bool runOnFunction(Function &F) {
49 AA = &getAnalysis<AliasAnalysis>();
50 MD = &getAnalysis<MemoryDependenceAnalysis>();
51 DominatorTree &DT = getAnalysis<DominatorTree>();
54 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
55 // Only check non-dead blocks. Dead blocks may have strange pointer
56 // cycles that will confuse alias analysis.
57 if (DT.isReachableFromEntry(I))
58 Changed |= runOnBasicBlock(*I);
64 bool runOnBasicBlock(BasicBlock &BB);
65 bool HandleFree(CallInst *F);
66 bool handleEndBlock(BasicBlock &BB);
67 void RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc,
68 SmallPtrSet<Value*, 16> &DeadStackObjects);
71 // getAnalysisUsage - We require post dominance frontiers (aka Control
73 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
75 AU.addRequired<DominatorTree>();
76 AU.addRequired<AliasAnalysis>();
77 AU.addRequired<MemoryDependenceAnalysis>();
78 AU.addPreserved<AliasAnalysis>();
79 AU.addPreserved<DominatorTree>();
80 AU.addPreserved<MemoryDependenceAnalysis>();
86 INITIALIZE_PASS_BEGIN(DSE, "dse", "Dead Store Elimination", false, false)
87 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
88 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
89 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
90 INITIALIZE_PASS_END(DSE, "dse", "Dead Store Elimination", false, false)
92 FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
94 //===----------------------------------------------------------------------===//
96 //===----------------------------------------------------------------------===//
98 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
99 /// and zero out all the operands of this instruction. If any of them become
100 /// dead, delete them and the computation tree that feeds them.
102 /// If ValueSet is non-null, remove any deleted instructions from it as well.
104 static void DeleteDeadInstruction(Instruction *I,
105 MemoryDependenceAnalysis &MD,
106 SmallPtrSet<Value*, 16> *ValueSet = 0) {
107 SmallVector<Instruction*, 32> NowDeadInsts;
109 NowDeadInsts.push_back(I);
112 // Before we touch this instruction, remove it from memdep!
114 Instruction *DeadInst = NowDeadInsts.pop_back_val();
117 // This instruction is dead, zap it, in stages. Start by removing it from
118 // MemDep, which needs to know the operands and needs it to be in the
120 MD.removeInstruction(DeadInst);
122 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
123 Value *Op = DeadInst->getOperand(op);
124 DeadInst->setOperand(op, 0);
126 // If this operand just became dead, add it to the NowDeadInsts list.
127 if (!Op->use_empty()) continue;
129 if (Instruction *OpI = dyn_cast<Instruction>(Op))
130 if (isInstructionTriviallyDead(OpI))
131 NowDeadInsts.push_back(OpI);
134 DeadInst->eraseFromParent();
136 if (ValueSet) ValueSet->erase(DeadInst);
137 } while (!NowDeadInsts.empty());
141 /// hasMemoryWrite - Does this instruction write some memory? This only returns
142 /// true for things that we can analyze with other helpers below.
143 static bool hasMemoryWrite(Instruction *I) {
144 if (isa<StoreInst>(I))
146 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
147 switch (II->getIntrinsicID()) {
150 case Intrinsic::memset:
151 case Intrinsic::memmove:
152 case Intrinsic::memcpy:
153 case Intrinsic::init_trampoline:
154 case Intrinsic::lifetime_end:
161 /// getLocForWrite - Return a Location stored to by the specified instruction.
162 static AliasAnalysis::Location
163 getLocForWrite(Instruction *Inst, AliasAnalysis &AA) {
164 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
165 return AA.getLocation(SI);
167 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Inst)) {
168 // memcpy/memmove/memset.
169 AliasAnalysis::Location Loc = AA.getLocationForDest(MI);
170 // If we don't have target data around, an unknown size in Location means
171 // that we should use the size of the pointee type. This isn't valid for
172 // memset/memcpy, which writes more than an i8.
173 if (Loc.Size == AliasAnalysis::UnknownSize && AA.getTargetData() == 0)
174 return AliasAnalysis::Location();
178 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
179 if (II == 0) return AliasAnalysis::Location();
181 switch (II->getIntrinsicID()) {
182 default: return AliasAnalysis::Location(); // Unhandled intrinsic.
183 case Intrinsic::init_trampoline:
184 // If we don't have target data around, an unknown size in Location means
185 // that we should use the size of the pointee type. This isn't valid for
186 // init.trampoline, which writes more than an i8.
187 if (AA.getTargetData() == 0) return AliasAnalysis::Location();
189 // FIXME: We don't know the size of the trampoline, so we can't really
191 return AliasAnalysis::Location(II->getArgOperand(0));
192 case Intrinsic::lifetime_end: {
193 uint64_t Len = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
194 return AliasAnalysis::Location(II->getArgOperand(1), Len);
199 /// isRemovable - If the value of this instruction and the memory it writes to
200 /// is unused, may we delete this instruction?
201 static bool isRemovable(Instruction *I) {
202 // Don't remove volatile stores.
203 if (StoreInst *SI = dyn_cast<StoreInst>(I))
204 return !SI->isVolatile();
206 IntrinsicInst *II = cast<IntrinsicInst>(I);
207 switch (II->getIntrinsicID()) {
208 default: assert(0 && "doesn't pass 'hasMemoryWrite' predicate");
209 case Intrinsic::lifetime_end:
210 // Never remove dead lifetime_end's, e.g. because it is followed by a
213 case Intrinsic::init_trampoline:
214 // Always safe to remove init_trampoline.
217 case Intrinsic::memset:
218 case Intrinsic::memmove:
219 case Intrinsic::memcpy:
220 // Don't remove volatile memory intrinsics.
221 return !cast<MemIntrinsic>(II)->isVolatile();
225 /// getStoredPointerOperand - Return the pointer that is being written to.
226 static Value *getStoredPointerOperand(Instruction *I) {
227 if (StoreInst *SI = dyn_cast<StoreInst>(I))
228 return SI->getPointerOperand();
229 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
230 return MI->getDest();
232 IntrinsicInst *II = cast<IntrinsicInst>(I);
233 switch (II->getIntrinsicID()) {
234 default: assert(false && "Unexpected intrinsic!");
235 case Intrinsic::init_trampoline:
236 return II->getArgOperand(0);
240 static uint64_t getPointerSize(Value *V, AliasAnalysis &AA) {
241 const TargetData *TD = AA.getTargetData();
243 return AliasAnalysis::UnknownSize;
245 if (AllocaInst *A = dyn_cast<AllocaInst>(V)) {
246 // Get size information for the alloca
247 if (ConstantInt *C = dyn_cast<ConstantInt>(A->getArraySize()))
248 return C->getZExtValue() * TD->getTypeAllocSize(A->getAllocatedType());
249 return AliasAnalysis::UnknownSize;
252 assert(isa<Argument>(V) && "Expected AllocaInst or Argument!");
253 const PointerType *PT = cast<PointerType>(V->getType());
254 return TD->getTypeAllocSize(PT->getElementType());
258 /// isCompleteOverwrite - Return true if a store to the 'Later' location
259 /// completely overwrites a store to the 'Earlier' location.
260 static bool isCompleteOverwrite(const AliasAnalysis::Location &Later,
261 const AliasAnalysis::Location &Earlier,
263 const Value *P1 = Later.Ptr->stripPointerCasts();
264 const Value *P2 = Earlier.Ptr->stripPointerCasts();
266 // Make sure that the start pointers are the same.
270 // If we don't know the sizes of either access, then we can't do a comparison.
271 if (Later.Size == AliasAnalysis::UnknownSize ||
272 Earlier.Size == AliasAnalysis::UnknownSize) {
273 // If we have no TargetData information around, then the size of the store
274 // is inferrable from the pointee type. If they are the same type, then we
275 // know that the store is safe.
276 if (AA.getTargetData() == 0)
277 return Later.Ptr->getType() == Earlier.Ptr->getType();
281 // Make sure that the Later size is >= the Earlier size.
282 if (Later.Size < Earlier.Size)
289 //===----------------------------------------------------------------------===//
291 //===----------------------------------------------------------------------===//
293 bool DSE::runOnBasicBlock(BasicBlock &BB) {
294 bool MadeChange = false;
296 // Do a top-down walk on the BB.
297 for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
298 Instruction *Inst = BBI++;
300 // Handle 'free' calls specially.
301 if (CallInst *F = isFreeCall(Inst)) {
302 MadeChange |= HandleFree(F);
306 // If we find something that writes memory, get its memory dependence.
307 if (!hasMemoryWrite(Inst))
310 MemDepResult InstDep = MD->getDependency(Inst);
312 // Ignore non-local store liveness.
313 // FIXME: cross-block DSE would be fun. :)
314 if (InstDep.isNonLocal() ||
315 // Ignore self dependence, which happens in the entry block of the
317 InstDep.getInst() == Inst)
320 // If we're storing the same value back to a pointer that we just
321 // loaded from, then the store can be removed.
322 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
323 if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
324 if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
325 SI->getOperand(0) == DepLoad && !SI->isVolatile()) {
326 // DeleteDeadInstruction can delete the current instruction. Save BBI
327 // in case we need it.
328 WeakVH NextInst(BBI);
330 DeleteDeadInstruction(SI, *MD);
332 if (NextInst == 0) // Next instruction deleted.
334 else if (BBI != BB.begin()) // Revisit this instruction if possible.
343 // Figure out what location is being stored to.
344 AliasAnalysis::Location Loc = getLocForWrite(Inst, *AA);
346 // If we didn't get a useful location, fail.
350 while (!InstDep.isNonLocal()) {
351 // Get the memory clobbered by the instruction we depend on. MemDep will
352 // skip any instructions that 'Loc' clearly doesn't interact with. If we
353 // end up depending on a may- or must-aliased load, then we can't optimize
354 // away the store and we bail out. However, if we depend on on something
355 // that overwrites the memory location we *can* potentially optimize it.
357 // Find out what memory location the dependant instruction stores.
358 Instruction *DepWrite = InstDep.getInst();
359 AliasAnalysis::Location DepLoc = getLocForWrite(DepWrite, *AA);
360 // If we didn't get a useful location, or if it isn't a size, bail out.
364 // If we find a removable write that is completely obliterated by the
365 // store to 'Loc' then we can remove it.
366 if (isRemovable(DepWrite) && isCompleteOverwrite(Loc, DepLoc, *AA)) {
367 // Delete the store and now-dead instructions that feed it.
368 DeleteDeadInstruction(DepWrite, *MD);
372 // DeleteDeadInstruction can delete the current instruction in loop
375 if (BBI != BB.begin())
380 // If this is a may-aliased store that is clobbering the store value, we
381 // can keep searching past it for another must-aliased pointer that stores
382 // to the same location. For example, in:
386 // we can remove the first store to P even though we don't know if P and Q
388 if (DepWrite == &BB.front()) break;
390 // Can't look past this instruction if it might read 'Loc'.
391 if (AA->getModRefInfo(DepWrite, Loc) & AliasAnalysis::Ref)
394 InstDep = MD->getPointerDependencyFrom(Loc, false, DepWrite, &BB);
398 // If this block ends in a return, unwind, or unreachable, all allocas are
399 // dead at its end, which means stores to them are also dead.
400 if (BB.getTerminator()->getNumSuccessors() == 0)
401 MadeChange |= handleEndBlock(BB);
406 /// HandleFree - Handle frees of entire structures whose dependency is a store
407 /// to a field of that structure.
408 bool DSE::HandleFree(CallInst *F) {
409 MemDepResult Dep = MD->getDependency(F);
411 if (Dep.isNonLocal()) return false;
413 Instruction *Dependency = Dep.getInst();
414 if (!hasMemoryWrite(Dependency) || !isRemovable(Dependency))
418 getStoredPointerOperand(Dependency)->getUnderlyingObject();
420 // Check for aliasing.
421 if (AA->alias(F->getArgOperand(0), 1, DepPointer, 1) !=
422 AliasAnalysis::MustAlias)
425 // DCE instructions only used to calculate that store
426 DeleteDeadInstruction(Dependency, *MD);
429 // Inst's old Dependency is now deleted. Compute the next dependency,
430 // which may also be dead, as in
432 // s[1] = 0; // This has just been deleted.
434 Dep = MD->getDependency(F);
435 } while (!Dep.isNonLocal());
440 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
441 /// function end block. Ex:
444 /// store i32 1, i32* %A
446 bool DSE::handleEndBlock(BasicBlock &BB) {
447 bool MadeChange = false;
449 // Keep track of all of the stack objects that are dead at the end of the
451 SmallPtrSet<Value*, 16> DeadStackObjects;
453 // Find all of the alloca'd pointers in the entry block.
454 BasicBlock *Entry = BB.getParent()->begin();
455 for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I)
456 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
457 DeadStackObjects.insert(AI);
459 // Treat byval arguments the same, stores to them are dead at the end of the
461 for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
462 AE = BB.getParent()->arg_end(); AI != AE; ++AI)
463 if (AI->hasByValAttr())
464 DeadStackObjects.insert(AI);
466 // Scan the basic block backwards
467 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
470 // If we find a store, check to see if it points into a dead stack value.
471 if (hasMemoryWrite(BBI) && isRemovable(BBI)) {
472 // See through pointer-to-pointer bitcasts
473 Value *Pointer = getStoredPointerOperand(BBI)->getUnderlyingObject();
475 // Stores to stack values are valid candidates for removal.
476 if (DeadStackObjects.count(Pointer)) {
477 // DCE instructions only used to calculate that store.
478 Instruction *Dead = BBI++;
479 DeleteDeadInstruction(Dead, *MD, &DeadStackObjects);
486 // Remove any dead non-memory-mutating instructions.
487 if (isInstructionTriviallyDead(BBI)) {
488 Instruction *Inst = BBI++;
489 DeleteDeadInstruction(Inst, *MD, &DeadStackObjects);
495 if (AllocaInst *A = dyn_cast<AllocaInst>(BBI)) {
496 DeadStackObjects.erase(A);
500 if (CallSite CS = cast<Value>(BBI)) {
501 // If this call does not access memory, it can't be loading any of our
503 if (AA->doesNotAccessMemory(CS))
506 unsigned NumModRef = 0, NumOther = 0;
508 // If the call might load from any of our allocas, then any store above
510 SmallVector<Value*, 8> LiveAllocas;
511 for (SmallPtrSet<Value*, 16>::iterator I = DeadStackObjects.begin(),
512 E = DeadStackObjects.end(); I != E; ++I) {
513 // If we detect that our AA is imprecise, it's not worth it to scan the
514 // rest of the DeadPointers set. Just assume that the AA will return
515 // ModRef for everything, and go ahead and bail out.
516 if (NumModRef >= 16 && NumOther == 0)
519 // See if the call site touches it.
520 AliasAnalysis::ModRefResult A =
521 AA->getModRefInfo(CS, *I, getPointerSize(*I, *AA));
523 if (A == AliasAnalysis::ModRef)
528 if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
529 LiveAllocas.push_back(*I);
532 for (SmallVector<Value*, 8>::iterator I = LiveAllocas.begin(),
533 E = LiveAllocas.end(); I != E; ++I)
534 DeadStackObjects.erase(*I);
536 // If all of the allocas were clobbered by the call then we're not going
537 // to find anything else to process.
538 if (DeadStackObjects.empty())
544 AliasAnalysis::Location LoadedLoc;
546 // If we encounter a use of the pointer, it is no longer considered dead
547 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
548 LoadedLoc = AA->getLocation(L);
549 } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) {
550 LoadedLoc = AA->getLocation(V);
551 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(BBI)) {
552 LoadedLoc = AA->getLocationForSource(MTI);
554 // Not a loading instruction.
558 // Remove any allocas from the DeadPointer set that are loaded, as this
559 // makes any stores above the access live.
560 RemoveAccessedObjects(LoadedLoc, DeadStackObjects);
562 // If all of the allocas were clobbered by the access then we're not going
563 // to find anything else to process.
564 if (DeadStackObjects.empty())
571 /// RemoveAccessedObjects - Check to see if the specified location may alias any
572 /// of the stack objects in the DeadStackObjects set. If so, they become live
573 /// because the location is being loaded.
574 void DSE::RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc,
575 SmallPtrSet<Value*, 16> &DeadStackObjects) {
576 const Value *UnderlyingPointer = LoadedLoc.Ptr->getUnderlyingObject();
578 // A constant can't be in the dead pointer set.
579 if (isa<Constant>(UnderlyingPointer))
582 // If the kill pointer can be easily reduced to an alloca, don't bother doing
583 // extraneous AA queries.
584 if (isa<AllocaInst>(UnderlyingPointer) || isa<Argument>(UnderlyingPointer)) {
585 DeadStackObjects.erase(const_cast<Value*>(UnderlyingPointer));
589 SmallVector<Value*, 16> NowLive;
590 for (SmallPtrSet<Value*, 16>::iterator I = DeadStackObjects.begin(),
591 E = DeadStackObjects.end(); I != E; ++I) {
592 // See if the loaded location could alias the stack location.
593 AliasAnalysis::Location StackLoc(*I, getPointerSize(*I, *AA));
594 if (!AA->isNoAlias(StackLoc, LoadedLoc))
595 NowLive.push_back(*I);
598 for (SmallVector<Value*, 16>::iterator I = NowLive.begin(), E = NowLive.end();
600 DeadStackObjects.erase(*I);