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 bool RemoveUndeadPointers(Value *Ptr, uint64_t killPointerSize,
68 BasicBlock::iterator &BBI,
69 SmallPtrSet<Value*, 64> &deadPointers);
70 void DeleteDeadInstruction(Instruction *I,
71 SmallPtrSet<Value*, 64> *deadPointers = 0);
74 // getAnalysisUsage - We require post dominance frontiers (aka Control
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 /// hasMemoryWrite - Does this instruction write some memory? This only returns
98 /// true for things that we can analyze with other helpers below.
99 static bool hasMemoryWrite(Instruction *I) {
100 if (isa<StoreInst>(I))
102 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
103 switch (II->getIntrinsicID()) {
106 case Intrinsic::memset:
107 case Intrinsic::memmove:
108 case Intrinsic::memcpy:
109 case Intrinsic::init_trampoline:
110 case Intrinsic::lifetime_end:
117 /// getLocForWrite - Return a Location stored to by the specified instruction.
118 static AliasAnalysis::Location
119 getLocForWrite(Instruction *Inst, AliasAnalysis &AA) {
120 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
121 return AA.getLocation(SI);
123 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Inst)) {
124 // memcpy/memmove/memset.
125 AliasAnalysis::Location Loc = AA.getLocationForDest(MI);
126 // If we don't have target data around, an unknown size in Location means
127 // that we should use the size of the pointee type. This isn't valid for
128 // memset/memcpy, which writes more than an i8.
129 if (Loc.Size == AliasAnalysis::UnknownSize && AA.getTargetData() == 0)
130 return AliasAnalysis::Location();
134 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
135 if (II == 0) return AliasAnalysis::Location();
137 switch (II->getIntrinsicID()) {
138 default: return AliasAnalysis::Location(); // Unhandled intrinsic.
139 case Intrinsic::init_trampoline:
140 // If we don't have target data around, an unknown size in Location means
141 // that we should use the size of the pointee type. This isn't valid for
142 // init.trampoline, which writes more than an i8.
143 if (AA.getTargetData() == 0) return AliasAnalysis::Location();
145 // FIXME: We don't know the size of the trampoline, so we can't really
147 return AliasAnalysis::Location(II->getArgOperand(0));
148 case Intrinsic::lifetime_end: {
149 uint64_t Len = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
150 return AliasAnalysis::Location(II->getArgOperand(1), Len);
155 /// isRemovable - If the value of this instruction and the memory it writes to
156 /// is unused, may we delete this instruction?
157 static bool isRemovable(Instruction *I) {
158 // Don't remove volatile stores.
159 if (StoreInst *SI = dyn_cast<StoreInst>(I))
160 return !SI->isVolatile();
162 IntrinsicInst *II = cast<IntrinsicInst>(I);
163 switch (II->getIntrinsicID()) {
164 default: assert(0 && "doesn't pass 'hasMemoryWrite' predicate");
165 case Intrinsic::lifetime_end:
166 // Never remove dead lifetime_end's, e.g. because it is followed by a
169 case Intrinsic::init_trampoline:
170 // Always safe to remove init_trampoline.
173 case Intrinsic::memset:
174 case Intrinsic::memmove:
175 case Intrinsic::memcpy:
176 // Don't remove volatile memory intrinsics.
177 return !cast<MemIntrinsic>(II)->isVolatile();
181 /// getPointerOperand - Return the pointer that is being written to.
182 static Value *getPointerOperand(Instruction *I) {
183 assert(hasMemoryWrite(I));
184 if (StoreInst *SI = dyn_cast<StoreInst>(I))
185 return SI->getPointerOperand();
186 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
187 return MI->getArgOperand(0);
189 IntrinsicInst *II = cast<IntrinsicInst>(I);
190 switch (II->getIntrinsicID()) {
191 default: assert(false && "Unexpected intrinsic!");
192 case Intrinsic::init_trampoline:
193 return II->getArgOperand(0);
194 case Intrinsic::lifetime_end:
195 return II->getArgOperand(1);
199 static uint64_t getPointerSize(Value *V, AliasAnalysis &AA) {
200 const TargetData *TD = AA.getTargetData();
202 return AliasAnalysis::UnknownSize;
204 if (AllocaInst *A = dyn_cast<AllocaInst>(V)) {
205 // Get size information for the alloca
206 if (ConstantInt *C = dyn_cast<ConstantInt>(A->getArraySize()))
207 return C->getZExtValue() * TD->getTypeAllocSize(A->getAllocatedType());
208 return AliasAnalysis::UnknownSize;
211 assert(isa<Argument>(V) && "Expected AllocaInst or Argument!");
212 const PointerType *PT = cast<PointerType>(V->getType());
213 return TD->getTypeAllocSize(PT->getElementType());
217 /// isCompleteOverwrite - Return true if a store to the 'Later' location
218 /// completely overwrites a store to the 'Earlier' location.
219 static bool isCompleteOverwrite(const AliasAnalysis::Location &Later,
220 const AliasAnalysis::Location &Earlier,
222 const Value *P1 = Later.Ptr->stripPointerCasts();
223 const Value *P2 = Earlier.Ptr->stripPointerCasts();
225 // Make sure that the start pointers are the same.
229 // If we don't know the sizes of either access, then we can't do a comparison.
230 if (Later.Size == AliasAnalysis::UnknownSize ||
231 Earlier.Size == AliasAnalysis::UnknownSize) {
232 // If we have no TargetData information around, then the size of the store
233 // is inferrable from the pointee type. If they are the same type, then we
234 // know that the store is safe.
235 if (AA.getTargetData() == 0)
236 return Later.Ptr->getType() == Earlier.Ptr->getType();
240 // Make sure that the Later size is >= the Earlier size.
241 if (Later.Size < Earlier.Size)
247 bool DSE::runOnBasicBlock(BasicBlock &BB) {
248 bool MadeChange = false;
250 // Do a top-down walk on the BB.
251 for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
252 Instruction *Inst = BBI++;
254 // Handle 'free' calls specially.
255 if (CallInst *F = isFreeCall(Inst)) {
256 MadeChange |= HandleFree(F);
260 // If we find something that writes memory, get its memory dependence.
261 if (!hasMemoryWrite(Inst))
264 MemDepResult InstDep = MD->getDependency(Inst);
266 // Ignore non-local store liveness.
267 // FIXME: cross-block DSE would be fun. :)
268 if (InstDep.isNonLocal() ||
269 // Ignore self dependence, which happens in the entry block of the
271 InstDep.getInst() == Inst)
274 // If we're storing the same value back to a pointer that we just
275 // loaded from, then the store can be removed.
276 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
277 if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
278 if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
279 SI->getOperand(0) == DepLoad && !SI->isVolatile()) {
280 // DeleteDeadInstruction can delete the current instruction. Save BBI
281 // in case we need it.
282 WeakVH NextInst(BBI);
284 DeleteDeadInstruction(SI);
286 if (NextInst == 0) // Next instruction deleted.
288 else if (BBI != BB.begin()) // Revisit this instruction if possible.
297 // Figure out what location is being stored to.
298 AliasAnalysis::Location Loc = getLocForWrite(Inst, *AA);
300 // If we didn't get a useful location, fail.
304 while (!InstDep.isNonLocal()) {
305 // Get the memory clobbered by the instruction we depend on. MemDep will
306 // skip any instructions that 'Loc' clearly doesn't interact with. If we
307 // end up depending on a may- or must-aliased load, then we can't optimize
308 // away the store and we bail out. However, if we depend on on something
309 // that overwrites the memory location we *can* potentially optimize it.
311 // Find out what memory location the dependant instruction stores.
312 Instruction *DepWrite = InstDep.getInst();
313 AliasAnalysis::Location DepLoc = getLocForWrite(DepWrite, *AA);
314 // If we didn't get a useful location, or if it isn't a size, bail out.
318 // If we find a removable write that is completely obliterated by the
319 // store to 'Loc' then we can remove it.
320 if (isRemovable(DepWrite) && isCompleteOverwrite(Loc, DepLoc, *AA)) {
321 // Delete the store and now-dead instructions that feed it.
322 DeleteDeadInstruction(DepWrite);
326 // DeleteDeadInstruction can delete the current instruction in loop
329 if (BBI != BB.begin())
334 // If this is a may-aliased store that is clobbering the store value, we
335 // can keep searching past it for another must-aliased pointer that stores
336 // to the same location. For example, in:
340 // we can remove the first store to P even though we don't know if P and Q
342 if (DepWrite == &BB.front()) break;
344 // Can't look past this instruction if it might read 'Loc'.
345 if (AA->getModRefInfo(DepWrite, Loc) & AliasAnalysis::Ref)
348 InstDep = MD->getPointerDependencyFrom(Loc, false, DepWrite, &BB);
352 // If this block ends in a return, unwind, or unreachable, all allocas are
353 // dead at its end, which means stores to them are also dead.
354 if (BB.getTerminator()->getNumSuccessors() == 0)
355 MadeChange |= handleEndBlock(BB);
360 /// HandleFree - Handle frees of entire structures whose dependency is a store
361 /// to a field of that structure.
362 bool DSE::HandleFree(CallInst *F) {
363 MemDepResult Dep = MD->getDependency(F);
365 if (Dep.isNonLocal()) return false;
367 Instruction *Dependency = Dep.getInst();
368 if (!hasMemoryWrite(Dependency) || !isRemovable(Dependency))
371 Value *DepPointer = getPointerOperand(Dependency)->getUnderlyingObject();
373 // Check for aliasing.
374 if (AA->alias(F->getArgOperand(0), 1, DepPointer, 1) !=
375 AliasAnalysis::MustAlias)
378 // DCE instructions only used to calculate that store
379 DeleteDeadInstruction(Dependency);
382 // Inst's old Dependency is now deleted. Compute the next dependency,
383 // which may also be dead, as in
385 // s[1] = 0; // This has just been deleted.
387 Dep = MD->getDependency(F);
388 } while (!Dep.isNonLocal());
393 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
394 /// function end block. Ex:
397 /// store i32 1, i32* %A
399 bool DSE::handleEndBlock(BasicBlock &BB) {
400 bool MadeChange = false;
402 // Pointers alloca'd in this function are dead in the end block
403 SmallPtrSet<Value*, 64> deadPointers;
405 // Find all of the alloca'd pointers in the entry block.
406 BasicBlock *Entry = BB.getParent()->begin();
407 for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I)
408 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
409 deadPointers.insert(AI);
411 // Treat byval arguments the same, stores to them are dead at the end of the
413 for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
414 AE = BB.getParent()->arg_end(); AI != AE; ++AI)
415 if (AI->hasByValAttr())
416 deadPointers.insert(AI);
418 // Scan the basic block backwards
419 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
422 // If we find a store whose pointer is dead.
423 if (hasMemoryWrite(BBI)) {
424 if (isRemovable(BBI)) {
425 // See through pointer-to-pointer bitcasts
426 Value *pointerOperand = getPointerOperand(BBI)->getUnderlyingObject();
428 // Alloca'd pointers or byval arguments (which are functionally like
429 // alloca's) are valid candidates for removal.
430 if (deadPointers.count(pointerOperand)) {
431 // DCE instructions only used to calculate that store.
432 Instruction *Dead = BBI;
434 DeleteDeadInstruction(Dead, &deadPointers);
441 // Because a memcpy or memmove is also a load, we can't skip it if we
443 if (!isa<MemTransferInst>(BBI))
447 Value *killPointer = 0;
448 uint64_t killPointerSize = AliasAnalysis::UnknownSize;
450 // If we encounter a use of the pointer, it is no longer considered dead
451 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
452 // However, if this load is unused and not volatile, we can go ahead and
453 // remove it, and not have to worry about it making our pointer undead!
454 if (L->use_empty() && !L->isVolatile()) {
456 DeleteDeadInstruction(L, &deadPointers);
462 killPointer = L->getPointerOperand();
463 } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) {
464 killPointer = V->getOperand(0);
465 } else if (isa<MemTransferInst>(BBI) &&
466 isa<ConstantInt>(cast<MemTransferInst>(BBI)->getLength())) {
467 killPointer = cast<MemTransferInst>(BBI)->getSource();
468 killPointerSize = cast<ConstantInt>(
469 cast<MemTransferInst>(BBI)->getLength())->getZExtValue();
470 } else if (AllocaInst *A = dyn_cast<AllocaInst>(BBI)) {
471 deadPointers.erase(A);
473 // Dead alloca's can be DCE'd when we reach them
474 if (A->use_empty()) {
476 DeleteDeadInstruction(A, &deadPointers);
482 } else if (CallSite CS = cast<Value>(BBI)) {
483 // If this call does not access memory, it can't
484 // be undeadifying any of our pointers.
485 if (AA->doesNotAccessMemory(CS))
491 // Remove any pointers made undead by the call from the dead set
492 std::vector<Value*> dead;
493 for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
494 E = deadPointers.end(); I != E; ++I) {
495 // HACK: if we detect that our AA is imprecise, it's not
496 // worth it to scan the rest of the deadPointers set. Just
497 // assume that the AA will return ModRef for everything, and
498 // go ahead and bail.
499 if (modRef >= 16 && other == 0) {
500 deadPointers.clear();
504 // See if the call site touches it
505 AliasAnalysis::ModRefResult A =
506 AA->getModRefInfo(CS, *I, getPointerSize(*I, *AA));
508 if (A == AliasAnalysis::ModRef)
513 if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
517 for (std::vector<Value*>::iterator I = dead.begin(), E = dead.end();
519 deadPointers.erase(*I);
522 } else if (isInstructionTriviallyDead(BBI)) {
523 // For any non-memory-affecting non-terminators, DCE them as we reach them
524 Instruction *Inst = BBI;
526 DeleteDeadInstruction(Inst, &deadPointers);
535 killPointer = killPointer->getUnderlyingObject();
537 // Deal with undead pointers
538 MadeChange |= RemoveUndeadPointers(killPointer, killPointerSize, BBI,
545 /// RemoveUndeadPointers - check for uses of a pointer that make it
546 /// undead when scanning for dead stores to alloca's.
547 bool DSE::RemoveUndeadPointers(Value *killPointer, uint64_t killPointerSize,
548 BasicBlock::iterator &BBI,
549 SmallPtrSet<Value*, 64> &deadPointers) {
550 // If the kill pointer can be easily reduced to an alloca,
551 // don't bother doing extraneous AA queries.
552 if (deadPointers.count(killPointer)) {
553 deadPointers.erase(killPointer);
557 // A global can't be in the dead pointer set.
558 if (isa<GlobalValue>(killPointer))
561 bool MadeChange = false;
563 SmallVector<Value*, 16> undead;
565 for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
566 E = deadPointers.end(); I != E; ++I) {
567 // See if this pointer could alias it
568 AliasAnalysis::AliasResult A = AA->alias(*I, getPointerSize(*I, *AA),
569 killPointer, killPointerSize);
571 // If it must-alias and a store, we can delete it
572 if (isa<StoreInst>(BBI) && A == AliasAnalysis::MustAlias) {
573 StoreInst *S = cast<StoreInst>(BBI);
577 DeleteDeadInstruction(S, &deadPointers);
583 // Otherwise, it is undead
584 } else if (A != AliasAnalysis::NoAlias)
585 undead.push_back(*I);
588 for (SmallVector<Value*, 16>::iterator I = undead.begin(), E = undead.end();
590 deadPointers.erase(*I);
595 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
596 /// and zero out all the operands of this instruction. If any of them become
597 /// dead, delete them and the computation tree that feeds them.
599 /// If ValueSet is non-null, remove any deleted instructions from it as well.
601 void DSE::DeleteDeadInstruction(Instruction *I,
602 SmallPtrSet<Value*, 64> *ValueSet) {
603 SmallVector<Instruction*, 32> NowDeadInsts;
605 NowDeadInsts.push_back(I);
608 // Before we touch this instruction, remove it from memdep!
610 Instruction *DeadInst = NowDeadInsts.pop_back_val();
614 // This instruction is dead, zap it, in stages. Start by removing it from
615 // MemDep, which needs to know the operands and needs it to be in the
617 MD->removeInstruction(DeadInst);
619 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
620 Value *Op = DeadInst->getOperand(op);
621 DeadInst->setOperand(op, 0);
623 // If this operand just became dead, add it to the NowDeadInsts list.
624 if (!Op->use_empty()) continue;
626 if (Instruction *OpI = dyn_cast<Instruction>(Op))
627 if (isInstructionTriviallyDead(OpI))
628 NowDeadInsts.push_back(OpI);
631 DeadInst->eraseFromParent();
633 if (ValueSet) ValueSet->erase(DeadInst);
634 } while (!NowDeadInsts.empty());