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/MallocHelper.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 {
42 static char ID; // Pass identification, replacement for typeid
43 DSE() : FunctionPass(&ID) {}
45 virtual bool runOnFunction(Function &F) {
47 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
48 Changed |= runOnBasicBlock(*I);
52 bool runOnBasicBlock(BasicBlock &BB);
53 bool handleFreeWithNonTrivialDependency(Instruction *F, MemDepResult Dep);
54 bool handleEndBlock(BasicBlock &BB);
55 bool RemoveUndeadPointers(Value* Ptr, uint64_t killPointerSize,
56 BasicBlock::iterator& BBI,
57 SmallPtrSet<Value*, 64>& deadPointers);
58 void DeleteDeadInstruction(Instruction *I,
59 SmallPtrSet<Value*, 64> *deadPointers = 0);
62 // getAnalysisUsage - We require post dominance frontiers (aka Control
64 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
66 AU.addRequired<DominatorTree>();
67 AU.addRequired<AliasAnalysis>();
68 AU.addRequired<MemoryDependenceAnalysis>();
69 AU.addPreserved<DominatorTree>();
70 AU.addPreserved<AliasAnalysis>();
71 AU.addPreserved<MemoryDependenceAnalysis>();
77 static RegisterPass<DSE> X("dse", "Dead Store Elimination");
79 FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
81 bool DSE::runOnBasicBlock(BasicBlock &BB) {
82 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
83 TD = getAnalysisIfAvailable<TargetData>();
85 bool MadeChange = false;
87 // Do a top-down walk on the BB.
88 for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
89 Instruction *Inst = BBI++;
91 // If we find a store or a free, get its memory dependence.
92 if (!isa<StoreInst>(Inst) && !isa<FreeInst>(Inst) && !isFreeCall(Inst))
95 // Don't molest volatile stores or do queries that will return "clobber".
96 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
100 MemDepResult InstDep = MD.getDependency(Inst);
102 // Ignore non-local stores.
103 // FIXME: cross-block DSE would be fun. :)
104 if (InstDep.isNonLocal()) continue;
106 // Handle frees whose dependencies are non-trivial.
107 if (isa<FreeInst>(Inst) || isFreeCall(Inst)) {
108 MadeChange |= handleFreeWithNonTrivialDependency(Inst, InstDep);
112 StoreInst *SI = cast<StoreInst>(Inst);
114 // If not a definite must-alias dependency, ignore it.
115 if (!InstDep.isDef())
118 // If this is a store-store dependence, then the previous store is dead so
119 // long as this store is at least as big as it.
120 if (StoreInst *DepStore = dyn_cast<StoreInst>(InstDep.getInst()))
122 TD->getTypeStoreSize(DepStore->getOperand(0)->getType()) <=
123 TD->getTypeStoreSize(SI->getOperand(0)->getType())) {
124 // Delete the store and now-dead instructions that feed it.
125 DeleteDeadInstruction(DepStore);
129 // DeleteDeadInstruction can delete the current instruction in loop
132 if (BBI != BB.begin())
137 // If we're storing the same value back to a pointer that we just
138 // loaded from, then the store can be removed.
139 if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
140 if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
141 SI->getOperand(0) == DepLoad) {
142 // DeleteDeadInstruction can delete the current instruction. Save BBI
143 // in case we need it.
144 WeakVH NextInst(BBI);
146 DeleteDeadInstruction(SI);
148 if (NextInst == 0) // Next instruction deleted.
150 else if (BBI != BB.begin()) // Revisit this instruction if possible.
159 // If this block ends in a return, unwind, or unreachable, all allocas are
160 // dead at its end, which means stores to them are also dead.
161 if (BB.getTerminator()->getNumSuccessors() == 0)
162 MadeChange |= handleEndBlock(BB);
167 /// handleFreeWithNonTrivialDependency - Handle frees of entire structures whose
168 /// dependency is a store to a field of that structure.
169 bool DSE::handleFreeWithNonTrivialDependency(Instruction *F, MemDepResult Dep) {
170 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
172 StoreInst *Dependency = dyn_cast_or_null<StoreInst>(Dep.getInst());
173 if (!Dependency || Dependency->isVolatile())
176 Value *DepPointer = Dependency->getPointerOperand()->getUnderlyingObject();
178 // Check for aliasing.
179 Value* FreeVal = isa<FreeInst>(F) ? F->getOperand(0) : F->getOperand(1);
180 if (AA.alias(FreeVal, 1, DepPointer, 1) !=
181 AliasAnalysis::MustAlias)
184 // DCE instructions only used to calculate that store
185 DeleteDeadInstruction(Dependency);
190 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
191 /// function end block. Ex:
194 /// store i32 1, i32* %A
196 bool DSE::handleEndBlock(BasicBlock &BB) {
197 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
199 bool MadeChange = false;
201 // Pointers alloca'd in this function are dead in the end block
202 SmallPtrSet<Value*, 64> deadPointers;
204 // Find all of the alloca'd pointers in the entry block.
205 BasicBlock *Entry = BB.getParent()->begin();
206 for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I)
207 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
208 deadPointers.insert(AI);
210 // Treat byval arguments the same, stores to them are dead at the end of the
212 for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
213 AE = BB.getParent()->arg_end(); AI != AE; ++AI)
214 if (AI->hasByValAttr())
215 deadPointers.insert(AI);
217 // Scan the basic block backwards
218 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
221 // If we find a store whose pointer is dead.
222 if (StoreInst* S = dyn_cast<StoreInst>(BBI)) {
223 if (!S->isVolatile()) {
224 // See through pointer-to-pointer bitcasts
225 Value* pointerOperand = S->getPointerOperand()->getUnderlyingObject();
227 // Alloca'd pointers or byval arguments (which are functionally like
228 // alloca's) are valid candidates for removal.
229 if (deadPointers.count(pointerOperand)) {
230 // DCE instructions only used to calculate that store.
232 DeleteDeadInstruction(S, &deadPointers);
241 // We can also remove memcpy's to local variables at the end of a function.
242 if (MemCpyInst *M = dyn_cast<MemCpyInst>(BBI)) {
243 Value *dest = M->getDest()->getUnderlyingObject();
245 if (deadPointers.count(dest)) {
247 DeleteDeadInstruction(M, &deadPointers);
253 // Because a memcpy is also a load, we can't skip it if we didn't remove
257 Value* killPointer = 0;
258 uint64_t killPointerSize = ~0UL;
260 // If we encounter a use of the pointer, it is no longer considered dead
261 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
262 // However, if this load is unused and not volatile, we can go ahead and
263 // remove it, and not have to worry about it making our pointer undead!
264 if (L->use_empty() && !L->isVolatile()) {
266 DeleteDeadInstruction(L, &deadPointers);
272 killPointer = L->getPointerOperand();
273 } else if (VAArgInst* V = dyn_cast<VAArgInst>(BBI)) {
274 killPointer = V->getOperand(0);
275 } else if (isa<MemCpyInst>(BBI) &&
276 isa<ConstantInt>(cast<MemCpyInst>(BBI)->getLength())) {
277 killPointer = cast<MemCpyInst>(BBI)->getSource();
278 killPointerSize = cast<ConstantInt>(
279 cast<MemCpyInst>(BBI)->getLength())->getZExtValue();
280 } else if (AllocaInst* A = dyn_cast<AllocaInst>(BBI)) {
281 deadPointers.erase(A);
283 // Dead alloca's can be DCE'd when we reach them
284 if (A->use_empty()) {
286 DeleteDeadInstruction(A, &deadPointers);
292 } else if (CallSite::get(BBI).getInstruction() != 0) {
293 // If this call does not access memory, it can't
294 // be undeadifying any of our pointers.
295 CallSite CS = CallSite::get(BBI);
296 if (AA.doesNotAccessMemory(CS))
302 // Remove any pointers made undead by the call from the dead set
303 std::vector<Value*> dead;
304 for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
305 E = deadPointers.end(); I != E; ++I) {
306 // HACK: if we detect that our AA is imprecise, it's not
307 // worth it to scan the rest of the deadPointers set. Just
308 // assume that the AA will return ModRef for everything, and
309 // go ahead and bail.
310 if (modRef >= 16 && other == 0) {
311 deadPointers.clear();
315 // Get size information for the alloca
316 unsigned pointerSize = ~0U;
318 if (AllocaInst* A = dyn_cast<AllocaInst>(*I)) {
319 if (ConstantInt* C = dyn_cast<ConstantInt>(A->getArraySize()))
320 pointerSize = C->getZExtValue() *
321 TD->getTypeAllocSize(A->getAllocatedType());
323 const PointerType* PT = cast<PointerType>(
324 cast<Argument>(*I)->getType());
325 pointerSize = TD->getTypeAllocSize(PT->getElementType());
329 // See if the call site touches it
330 AliasAnalysis::ModRefResult A = AA.getModRefInfo(CS, *I, pointerSize);
332 if (A == AliasAnalysis::ModRef)
337 if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
341 for (std::vector<Value*>::iterator I = dead.begin(), E = dead.end();
343 deadPointers.erase(*I);
346 } else if (isInstructionTriviallyDead(BBI)) {
347 // For any non-memory-affecting non-terminators, DCE them as we reach them
348 Instruction *Inst = BBI;
350 DeleteDeadInstruction(Inst, &deadPointers);
359 killPointer = killPointer->getUnderlyingObject();
361 // Deal with undead pointers
362 MadeChange |= RemoveUndeadPointers(killPointer, killPointerSize, BBI,
369 /// RemoveUndeadPointers - check for uses of a pointer that make it
370 /// undead when scanning for dead stores to alloca's.
371 bool DSE::RemoveUndeadPointers(Value* killPointer, uint64_t killPointerSize,
372 BasicBlock::iterator &BBI,
373 SmallPtrSet<Value*, 64>& deadPointers) {
374 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
376 // If the kill pointer can be easily reduced to an alloca,
377 // don't bother doing extraneous AA queries.
378 if (deadPointers.count(killPointer)) {
379 deadPointers.erase(killPointer);
383 // A global can't be in the dead pointer set.
384 if (isa<GlobalValue>(killPointer))
387 bool MadeChange = false;
389 SmallVector<Value*, 16> undead;
391 for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
392 E = deadPointers.end(); I != E; ++I) {
393 // Get size information for the alloca.
394 unsigned pointerSize = ~0U;
396 if (AllocaInst* A = dyn_cast<AllocaInst>(*I)) {
397 if (ConstantInt* C = dyn_cast<ConstantInt>(A->getArraySize()))
398 pointerSize = C->getZExtValue() *
399 TD->getTypeAllocSize(A->getAllocatedType());
401 const PointerType* PT = cast<PointerType>(cast<Argument>(*I)->getType());
402 pointerSize = TD->getTypeAllocSize(PT->getElementType());
406 // See if this pointer could alias it
407 AliasAnalysis::AliasResult A = AA.alias(*I, pointerSize,
408 killPointer, killPointerSize);
410 // If it must-alias and a store, we can delete it
411 if (isa<StoreInst>(BBI) && A == AliasAnalysis::MustAlias) {
412 StoreInst* S = cast<StoreInst>(BBI);
416 DeleteDeadInstruction(S, &deadPointers);
422 // Otherwise, it is undead
423 } else if (A != AliasAnalysis::NoAlias)
424 undead.push_back(*I);
427 for (SmallVector<Value*, 16>::iterator I = undead.begin(), E = undead.end();
429 deadPointers.erase(*I);
434 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
435 /// and zero out all the operands of this instruction. If any of them become
436 /// dead, delete them and the computation tree that feeds them.
438 /// If ValueSet is non-null, remove any deleted instructions from it as well.
440 void DSE::DeleteDeadInstruction(Instruction *I,
441 SmallPtrSet<Value*, 64> *ValueSet) {
442 SmallVector<Instruction*, 32> NowDeadInsts;
444 NowDeadInsts.push_back(I);
447 // Before we touch this instruction, remove it from memdep!
448 MemoryDependenceAnalysis &MDA = getAnalysis<MemoryDependenceAnalysis>();
449 while (!NowDeadInsts.empty()) {
450 Instruction *DeadInst = NowDeadInsts.back();
451 NowDeadInsts.pop_back();
455 // This instruction is dead, zap it, in stages. Start by removing it from
456 // MemDep, which needs to know the operands and needs it to be in the
458 MDA.removeInstruction(DeadInst);
460 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
461 Value *Op = DeadInst->getOperand(op);
462 DeadInst->setOperand(op, 0);
464 // If this operand just became dead, add it to the NowDeadInsts list.
465 if (!Op->use_empty()) continue;
467 if (Instruction *OpI = dyn_cast<Instruction>(Op))
468 if (isInstructionTriviallyDead(OpI))
469 NowDeadInsts.push_back(OpI);
472 DeadInst->eraseFromParent();
474 if (ValueSet) ValueSet->erase(DeadInst);