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/MemoryDependenceAnalysis.h"
30 #include "llvm/Target/TargetData.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Support/Compiler.h"
35 STATISTIC(NumFastStores, "Number of stores deleted");
36 STATISTIC(NumFastOther , "Number of other instrs removed");
39 struct VISIBILITY_HIDDEN DSE : public FunctionPass {
40 static char ID; // Pass identification, replacement for typeid
41 DSE() : FunctionPass(&ID) {}
43 virtual bool runOnFunction(Function &F) {
45 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
46 Changed |= runOnBasicBlock(*I);
50 bool runOnBasicBlock(BasicBlock &BB);
51 bool handleFreeWithNonTrivialDependency(FreeInst *F, Instruction *Dep);
52 bool handleEndBlock(BasicBlock &BB);
53 bool RemoveUndeadPointers(Value* pointer, uint64_t killPointerSize,
54 BasicBlock::iterator& BBI,
55 SmallPtrSet<Value*, 64>& deadPointers);
56 void DeleteDeadInstruction(Instruction *I,
57 SmallPtrSet<Value*, 64> *deadPointers = 0);
60 // getAnalysisUsage - We require post dominance frontiers (aka Control
62 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
64 AU.addRequired<DominatorTree>();
65 AU.addRequired<TargetData>();
66 AU.addRequired<AliasAnalysis>();
67 AU.addRequired<MemoryDependenceAnalysis>();
68 AU.addPreserved<DominatorTree>();
69 AU.addPreserved<AliasAnalysis>();
70 AU.addPreserved<MemoryDependenceAnalysis>();
76 static RegisterPass<DSE> X("dse", "Dead Store Elimination");
78 FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
80 bool DSE::runOnBasicBlock(BasicBlock &BB) {
81 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
82 TargetData &TD = getAnalysis<TargetData>();
84 // Record the last-seen store to this pointer
85 DenseMap<Value*, StoreInst*> lastStore;
87 bool MadeChange = false;
89 // Do a top-down walk on the BB
90 for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
91 Instruction *Inst = BBI++;
93 // If we find a store or a free...
94 if (!isa<StoreInst>(Inst) && !isa<FreeInst>(Inst))
98 if (StoreInst* S = dyn_cast<StoreInst>(Inst)) {
101 pointer = S->getPointerOperand();
103 pointer = cast<FreeInst>(Inst)->getPointerOperand();
106 pointer = pointer->stripPointerCasts();
107 StoreInst *&last = lastStore[pointer];
109 // ... to a pointer that has been stored to before...
111 Instruction* dep = MD.getDependency(Inst);
112 bool deletedStore = false;
114 // ... and no other memory dependencies are between them....
115 while (dep != MemoryDependenceAnalysis::None &&
116 dep != MemoryDependenceAnalysis::NonLocal &&
117 isa<StoreInst>(dep)) {
119 TD.getTypeStoreSize(last->getOperand(0)->getType()) >
120 TD.getTypeStoreSize(Inst->getOperand(0)->getType())) {
121 dep = MD.getDependency(Inst, dep);
125 // Delete the store and now-dead instructions that feed it.
126 DeleteDeadInstruction(last);
133 // If we deleted a store, reinvestigate this instruction.
135 if (!isa<TerminatorInst>(BB.begin()))
141 // Handle frees whose dependencies are non-trivial.
142 if (FreeInst* F = dyn_cast<FreeInst>(Inst)) {
143 MadeChange |= handleFreeWithNonTrivialDependency(F, MD.getDependency(F));
145 // No known stores after the free.
148 StoreInst* S = cast<StoreInst>(Inst);
150 // If we're storing the same value back to a pointer that we just
151 // loaded from, then the store can be removed;
152 if (LoadInst* L = dyn_cast<LoadInst>(S->getOperand(0))) {
153 // FIXME: Don't do dep query if Parents don't match and other stuff!
154 Instruction* dep = MD.getDependency(S);
155 DominatorTree& DT = getAnalysis<DominatorTree>();
157 if (!S->isVolatile() && S->getParent() == L->getParent() &&
158 S->getPointerOperand() == L->getPointerOperand() &&
159 (dep == MemoryDependenceAnalysis::None ||
160 dep == MemoryDependenceAnalysis::NonLocal ||
161 DT.dominates(dep, L))) {
163 DeleteDeadInstruction(S);
164 if (!isa<TerminatorInst>(BB.begin()))
169 // Update our most-recent-store map.
172 // Update our most-recent-store map.
177 // If this block ends in a return, unwind, or unreachable, all allocas are
178 // dead at its end, which means stores to them are also dead.
179 if (BB.getTerminator()->getNumSuccessors() == 0)
180 MadeChange |= handleEndBlock(BB);
185 /// handleFreeWithNonTrivialDependency - Handle frees of entire structures whose
186 /// dependency is a store to a field of that structure.
187 bool DSE::handleFreeWithNonTrivialDependency(FreeInst* F, Instruction* dep) {
188 TargetData &TD = getAnalysis<TargetData>();
189 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
191 if (dep == MemoryDependenceAnalysis::None ||
192 dep == MemoryDependenceAnalysis::NonLocal)
195 StoreInst* dependency = dyn_cast<StoreInst>(dep);
198 else if (dependency->isVolatile())
201 Value* depPointer = dependency->getPointerOperand();
202 const Type* depType = dependency->getOperand(0)->getType();
203 unsigned depPointerSize = TD.getTypeStoreSize(depType);
205 // Check for aliasing
206 AliasAnalysis::AliasResult A = AA.alias(F->getPointerOperand(), ~0U,
207 depPointer, depPointerSize);
209 if (A != AliasAnalysis::MustAlias)
212 // DCE instructions only used to calculate that store
213 DeleteDeadInstruction(dependency);
218 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
219 /// function end block. Ex:
222 /// store i32 1, i32* %A
224 bool DSE::handleEndBlock(BasicBlock &BB) {
225 TargetData &TD = getAnalysis<TargetData>();
226 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
228 bool MadeChange = false;
230 // Pointers alloca'd in this function are dead in the end block
231 SmallPtrSet<Value*, 64> deadPointers;
233 // Find all of the alloca'd pointers in the entry block.
234 BasicBlock *Entry = BB.getParent()->begin();
235 for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I)
236 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
237 deadPointers.insert(AI);
239 // Treat byval arguments the same, stores to them are dead at the end of the
241 for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
242 AE = BB.getParent()->arg_end(); AI != AE; ++AI)
243 if (AI->hasByValAttr())
244 deadPointers.insert(AI);
246 // Scan the basic block backwards
247 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
250 // If we find a store whose pointer is dead.
251 if (StoreInst* S = dyn_cast<StoreInst>(BBI)) {
252 if (!S->isVolatile()) {
253 // See through pointer-to-pointer bitcasts
254 Value* pointerOperand = S->getPointerOperand()->getUnderlyingObject();
256 // Alloca'd pointers or byval arguments (which are functionally like
257 // alloca's) are valid candidates for removal.
258 if (deadPointers.count(pointerOperand)) {
259 // DCE instructions only used to calculate that store.
261 DeleteDeadInstruction(S, &deadPointers);
270 // We can also remove memcpy's to local variables at the end of a function.
271 if (MemCpyInst *M = dyn_cast<MemCpyInst>(BBI)) {
272 Value *dest = M->getDest()->getUnderlyingObject();
274 if (deadPointers.count(dest)) {
276 DeleteDeadInstruction(M, &deadPointers);
282 // Because a memcpy is also a load, we can't skip it if we didn't remove
286 Value* killPointer = 0;
287 uint64_t killPointerSize = ~0UL;
289 // If we encounter a use of the pointer, it is no longer considered dead
290 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
291 // However, if this load is unused and not volatile, we can go ahead and
292 // remove it, and not have to worry about it making our pointer undead!
293 if (L->use_empty() && !L->isVolatile()) {
295 DeleteDeadInstruction(L, &deadPointers);
301 killPointer = L->getPointerOperand();
302 } else if (VAArgInst* V = dyn_cast<VAArgInst>(BBI)) {
303 killPointer = V->getOperand(0);
304 } else if (isa<MemCpyInst>(BBI) &&
305 isa<ConstantInt>(cast<MemCpyInst>(BBI)->getLength())) {
306 killPointer = cast<MemCpyInst>(BBI)->getSource();
307 killPointerSize = cast<ConstantInt>(
308 cast<MemCpyInst>(BBI)->getLength())->getZExtValue();
309 } else if (AllocaInst* A = dyn_cast<AllocaInst>(BBI)) {
310 deadPointers.erase(A);
312 // Dead alloca's can be DCE'd when we reach them
313 if (A->use_empty()) {
315 DeleteDeadInstruction(A, &deadPointers);
321 } else if (CallSite::get(BBI).getInstruction() != 0) {
322 // If this call does not access memory, it can't
323 // be undeadifying any of our pointers.
324 CallSite CS = CallSite::get(BBI);
325 if (AA.doesNotAccessMemory(CS))
331 // Remove any pointers made undead by the call from the dead set
332 std::vector<Value*> dead;
333 for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
334 E = deadPointers.end(); I != E; ++I) {
335 // HACK: if we detect that our AA is imprecise, it's not
336 // worth it to scan the rest of the deadPointers set. Just
337 // assume that the AA will return ModRef for everything, and
338 // go ahead and bail.
339 if (modRef >= 16 && other == 0) {
340 deadPointers.clear();
344 // Get size information for the alloca
345 unsigned pointerSize = ~0U;
346 if (AllocaInst* A = dyn_cast<AllocaInst>(*I)) {
347 if (ConstantInt* C = dyn_cast<ConstantInt>(A->getArraySize()))
348 pointerSize = C->getZExtValue() *
349 TD.getABITypeSize(A->getAllocatedType());
351 const PointerType* PT = cast<PointerType>(
352 cast<Argument>(*I)->getType());
353 pointerSize = TD.getABITypeSize(PT->getElementType());
356 // See if the call site touches it
357 AliasAnalysis::ModRefResult A = AA.getModRefInfo(CS, *I, pointerSize);
359 if (A == AliasAnalysis::ModRef)
364 if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
368 for (std::vector<Value*>::iterator I = dead.begin(), E = dead.end();
370 deadPointers.erase(*I);
373 } else if (isInstructionTriviallyDead(BBI)) {
374 // For any non-memory-affecting non-terminators, DCE them as we reach them
375 Instruction *Inst = BBI;
377 DeleteDeadInstruction(Inst, &deadPointers);
386 killPointer = killPointer->getUnderlyingObject();
388 // Deal with undead pointers
389 MadeChange |= RemoveUndeadPointers(killPointer, killPointerSize, BBI,
396 /// RemoveUndeadPointers - check for uses of a pointer that make it
397 /// undead when scanning for dead stores to alloca's.
398 bool DSE::RemoveUndeadPointers(Value* killPointer, uint64_t killPointerSize,
399 BasicBlock::iterator &BBI,
400 SmallPtrSet<Value*, 64>& deadPointers) {
401 TargetData &TD = getAnalysis<TargetData>();
402 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
404 // If the kill pointer can be easily reduced to an alloca,
405 // don't bother doing extraneous AA queries.
406 if (deadPointers.count(killPointer)) {
407 deadPointers.erase(killPointer);
411 // A global can't be in the dead pointer set.
412 if (isa<GlobalValue>(killPointer))
415 bool MadeChange = false;
417 SmallVector<Value*, 16> undead;
419 for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
420 E = deadPointers.end(); I != E; ++I) {
421 // Get size information for the alloca.
422 unsigned pointerSize = ~0U;
423 if (AllocaInst* A = dyn_cast<AllocaInst>(*I)) {
424 if (ConstantInt* C = dyn_cast<ConstantInt>(A->getArraySize()))
425 pointerSize = C->getZExtValue() *
426 TD.getABITypeSize(A->getAllocatedType());
428 const PointerType* PT = cast<PointerType>(cast<Argument>(*I)->getType());
429 pointerSize = TD.getABITypeSize(PT->getElementType());
432 // See if this pointer could alias it
433 AliasAnalysis::AliasResult A = AA.alias(*I, pointerSize,
434 killPointer, killPointerSize);
436 // If it must-alias and a store, we can delete it
437 if (isa<StoreInst>(BBI) && A == AliasAnalysis::MustAlias) {
438 StoreInst* S = cast<StoreInst>(BBI);
442 DeleteDeadInstruction(S, &deadPointers);
448 // Otherwise, it is undead
449 } else if (A != AliasAnalysis::NoAlias)
450 undead.push_back(*I);
453 for (SmallVector<Value*, 16>::iterator I = undead.begin(), E = undead.end();
455 deadPointers.erase(*I);
460 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
461 /// and zero out all the operands of this instruction. If any of them become
462 /// dead, delete them and the computation tree that feeds them.
464 /// If ValueSet is non-null, remove any deleted instructions from it as well.
466 void DSE::DeleteDeadInstruction(Instruction *I,
467 SmallPtrSet<Value*, 64> *ValueSet) {
468 SmallVector<Instruction*, 32> NowDeadInsts;
470 NowDeadInsts.push_back(I);
473 // Before we touch this instruction, remove it from memdep!
474 MemoryDependenceAnalysis &MDA = getAnalysis<MemoryDependenceAnalysis>();
475 while (!NowDeadInsts.empty()) {
476 Instruction *DeadInst = NowDeadInsts.back();
477 NowDeadInsts.pop_back();
481 // This instruction is dead, zap it, in stages. Start by removing it from
482 // MemDep, which needs to know the operands and needs it to be in the
484 MDA.removeInstruction(DeadInst);
486 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
487 Value *Op = DeadInst->getOperand(op);
488 DeadInst->setOperand(op, 0);
490 // If this operand just became dead, add it to the NowDeadInsts list.
491 if (!Op->use_empty()) continue;
493 if (Instruction *OpI = dyn_cast<Instruction>(Op))
494 if (isInstructionTriviallyDead(OpI))
495 NowDeadInsts.push_back(OpI);
498 DeadInst->eraseFromParent();
500 if (ValueSet) ValueSet->erase(DeadInst);