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 {
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) && !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 (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.
158 // If this is a lifetime end marker, we can throw away the store.
159 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(InstDep.getInst())) {
160 if (II->getIntrinsicID() == Intrinsic::lifetime_end) {
161 // Delete the store and now-dead instructions that feed it.
162 // DeleteDeadInstruction can delete the current instruction. Save BBI
163 // in case we need it.
164 WeakVH NextInst(BBI);
166 DeleteDeadInstruction(SI);
168 if (NextInst == 0) // Next instruction deleted.
170 else if (BBI != BB.begin()) // Revisit this instruction if possible.
179 // If this block ends in a return, unwind, or unreachable, all allocas are
180 // dead at its end, which means stores to them are also dead.
181 if (BB.getTerminator()->getNumSuccessors() == 0)
182 MadeChange |= handleEndBlock(BB);
187 /// handleFreeWithNonTrivialDependency - Handle frees of entire structures whose
188 /// dependency is a store to a field of that structure.
189 bool DSE::handleFreeWithNonTrivialDependency(Instruction *F, MemDepResult Dep) {
190 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
192 StoreInst *Dependency = dyn_cast_or_null<StoreInst>(Dep.getInst());
193 if (!Dependency || Dependency->isVolatile())
196 Value *DepPointer = Dependency->getPointerOperand()->getUnderlyingObject();
198 // Check for aliasing.
199 if (AA.alias(F->getOperand(1), 1, DepPointer, 1) !=
200 AliasAnalysis::MustAlias)
203 // DCE instructions only used to calculate that store
204 DeleteDeadInstruction(Dependency);
209 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
210 /// function end block. Ex:
213 /// store i32 1, i32* %A
215 bool DSE::handleEndBlock(BasicBlock &BB) {
216 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
218 bool MadeChange = false;
220 // Pointers alloca'd in this function are dead in the end block
221 SmallPtrSet<Value*, 64> deadPointers;
223 // Find all of the alloca'd pointers in the entry block.
224 BasicBlock *Entry = BB.getParent()->begin();
225 for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I)
226 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
227 deadPointers.insert(AI);
229 // Treat byval arguments the same, stores to them are dead at the end of the
231 for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
232 AE = BB.getParent()->arg_end(); AI != AE; ++AI)
233 if (AI->hasByValAttr())
234 deadPointers.insert(AI);
236 // Scan the basic block backwards
237 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
240 // If we find a store whose pointer is dead.
241 if (StoreInst* S = dyn_cast<StoreInst>(BBI)) {
242 if (!S->isVolatile()) {
243 // See through pointer-to-pointer bitcasts
244 Value* pointerOperand = S->getPointerOperand()->getUnderlyingObject();
246 // Alloca'd pointers or byval arguments (which are functionally like
247 // alloca's) are valid candidates for removal.
248 if (deadPointers.count(pointerOperand)) {
249 // DCE instructions only used to calculate that store.
251 DeleteDeadInstruction(S, &deadPointers);
260 // We can also remove memcpy's to local variables at the end of a function.
261 if (MemCpyInst *M = dyn_cast<MemCpyInst>(BBI)) {
262 Value *dest = M->getDest()->getUnderlyingObject();
264 if (deadPointers.count(dest)) {
266 DeleteDeadInstruction(M, &deadPointers);
272 // Because a memcpy is also a load, we can't skip it if we didn't remove
276 Value* killPointer = 0;
277 uint64_t killPointerSize = ~0UL;
279 // If we encounter a use of the pointer, it is no longer considered dead
280 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
281 // However, if this load is unused and not volatile, we can go ahead and
282 // remove it, and not have to worry about it making our pointer undead!
283 if (L->use_empty() && !L->isVolatile()) {
285 DeleteDeadInstruction(L, &deadPointers);
291 killPointer = L->getPointerOperand();
292 } else if (VAArgInst* V = dyn_cast<VAArgInst>(BBI)) {
293 killPointer = V->getOperand(0);
294 } else if (isa<MemCpyInst>(BBI) &&
295 isa<ConstantInt>(cast<MemCpyInst>(BBI)->getLength())) {
296 killPointer = cast<MemCpyInst>(BBI)->getSource();
297 killPointerSize = cast<ConstantInt>(
298 cast<MemCpyInst>(BBI)->getLength())->getZExtValue();
299 } else if (AllocaInst* A = dyn_cast<AllocaInst>(BBI)) {
300 deadPointers.erase(A);
302 // Dead alloca's can be DCE'd when we reach them
303 if (A->use_empty()) {
305 DeleteDeadInstruction(A, &deadPointers);
311 } else if (CallSite::get(BBI).getInstruction() != 0) {
312 // If this call does not access memory, it can't
313 // be undeadifying any of our pointers.
314 CallSite CS = CallSite::get(BBI);
315 if (AA.doesNotAccessMemory(CS))
321 // Remove any pointers made undead by the call from the dead set
322 std::vector<Value*> dead;
323 for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
324 E = deadPointers.end(); I != E; ++I) {
325 // HACK: if we detect that our AA is imprecise, it's not
326 // worth it to scan the rest of the deadPointers set. Just
327 // assume that the AA will return ModRef for everything, and
328 // go ahead and bail.
329 if (modRef >= 16 && other == 0) {
330 deadPointers.clear();
334 // Get size information for the alloca
335 unsigned pointerSize = ~0U;
337 if (AllocaInst* A = dyn_cast<AllocaInst>(*I)) {
338 if (ConstantInt* C = dyn_cast<ConstantInt>(A->getArraySize()))
339 pointerSize = C->getZExtValue() *
340 TD->getTypeAllocSize(A->getAllocatedType());
342 const PointerType* PT = cast<PointerType>(
343 cast<Argument>(*I)->getType());
344 pointerSize = TD->getTypeAllocSize(PT->getElementType());
348 // See if the call site touches it
349 AliasAnalysis::ModRefResult A = AA.getModRefInfo(CS, *I, pointerSize);
351 if (A == AliasAnalysis::ModRef)
356 if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
360 for (std::vector<Value*>::iterator I = dead.begin(), E = dead.end();
362 deadPointers.erase(*I);
365 } else if (isInstructionTriviallyDead(BBI)) {
366 // For any non-memory-affecting non-terminators, DCE them as we reach them
367 Instruction *Inst = BBI;
369 DeleteDeadInstruction(Inst, &deadPointers);
378 killPointer = killPointer->getUnderlyingObject();
380 // Deal with undead pointers
381 MadeChange |= RemoveUndeadPointers(killPointer, killPointerSize, BBI,
388 /// RemoveUndeadPointers - check for uses of a pointer that make it
389 /// undead when scanning for dead stores to alloca's.
390 bool DSE::RemoveUndeadPointers(Value* killPointer, uint64_t killPointerSize,
391 BasicBlock::iterator &BBI,
392 SmallPtrSet<Value*, 64>& deadPointers) {
393 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
395 // If the kill pointer can be easily reduced to an alloca,
396 // don't bother doing extraneous AA queries.
397 if (deadPointers.count(killPointer)) {
398 deadPointers.erase(killPointer);
402 // A global can't be in the dead pointer set.
403 if (isa<GlobalValue>(killPointer))
406 bool MadeChange = false;
408 SmallVector<Value*, 16> undead;
410 for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
411 E = deadPointers.end(); I != E; ++I) {
412 // Get size information for the alloca.
413 unsigned pointerSize = ~0U;
415 if (AllocaInst* A = dyn_cast<AllocaInst>(*I)) {
416 if (ConstantInt* C = dyn_cast<ConstantInt>(A->getArraySize()))
417 pointerSize = C->getZExtValue() *
418 TD->getTypeAllocSize(A->getAllocatedType());
420 const PointerType* PT = cast<PointerType>(cast<Argument>(*I)->getType());
421 pointerSize = TD->getTypeAllocSize(PT->getElementType());
425 // See if this pointer could alias it
426 AliasAnalysis::AliasResult A = AA.alias(*I, pointerSize,
427 killPointer, killPointerSize);
429 // If it must-alias and a store, we can delete it
430 if (isa<StoreInst>(BBI) && A == AliasAnalysis::MustAlias) {
431 StoreInst* S = cast<StoreInst>(BBI);
435 DeleteDeadInstruction(S, &deadPointers);
441 // Otherwise, it is undead
442 } else if (A != AliasAnalysis::NoAlias)
443 undead.push_back(*I);
446 for (SmallVector<Value*, 16>::iterator I = undead.begin(), E = undead.end();
448 deadPointers.erase(*I);
453 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
454 /// and zero out all the operands of this instruction. If any of them become
455 /// dead, delete them and the computation tree that feeds them.
457 /// If ValueSet is non-null, remove any deleted instructions from it as well.
459 void DSE::DeleteDeadInstruction(Instruction *I,
460 SmallPtrSet<Value*, 64> *ValueSet) {
461 SmallVector<Instruction*, 32> NowDeadInsts;
463 NowDeadInsts.push_back(I);
466 // Before we touch this instruction, remove it from memdep!
467 MemoryDependenceAnalysis &MDA = getAnalysis<MemoryDependenceAnalysis>();
468 while (!NowDeadInsts.empty()) {
469 Instruction *DeadInst = NowDeadInsts.back();
470 NowDeadInsts.pop_back();
474 // This instruction is dead, zap it, in stages. Start by removing it from
475 // MemDep, which needs to know the operands and needs it to be in the
477 MDA.removeInstruction(DeadInst);
479 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
480 Value *Op = DeadInst->getOperand(op);
481 DeadInst->setOperand(op, 0);
483 // If this operand just became dead, add it to the NowDeadInsts list.
484 if (!Op->use_empty()) continue;
486 if (Instruction *OpI = dyn_cast<Instruction>(Op))
487 if (isInstructionTriviallyDead(OpI))
488 NowDeadInsts.push_back(OpI);
491 DeadInst->eraseFromParent();
493 if (ValueSet) ValueSet->erase(DeadInst);