1 //===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation --*- C++ -*-===//
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 an analysis that determines, for a given memory
11 // operation, what preceding memory operations it depends on. It builds on
12 // alias analysis information, and tries to provide a lazy, caching interface to
13 // a common kind of alias information query.
15 //===----------------------------------------------------------------------===//
17 #define DEBUG_TYPE "memdep"
18 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
19 #include "llvm/Constants.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/Function.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/Support/CFG.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Target/TargetData.h"
31 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
32 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
33 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
34 char MemoryDependenceAnalysis::ID = 0;
36 // Register this pass...
37 static RegisterPass<MemoryDependenceAnalysis> X("memdep",
38 "Memory Dependence Analysis", false, true);
40 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
42 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
44 AU.addRequiredTransitive<AliasAnalysis>();
45 AU.addRequiredTransitive<TargetData>();
48 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
49 AA = &getAnalysis<AliasAnalysis>();
50 TD = &getAnalysis<TargetData>();
55 /// getCallSiteDependencyFrom - Private helper for finding the local
56 /// dependencies of a call site.
57 MemDepResult MemoryDependenceAnalysis::
58 getCallSiteDependencyFrom(CallSite CS, BasicBlock::iterator ScanIt,
60 // Walk backwards through the block, looking for dependencies
61 while (ScanIt != BB->begin()) {
62 Instruction *Inst = --ScanIt;
64 // If this inst is a memory op, get the pointer it accessed
66 uint64_t PointerSize = 0;
67 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
68 Pointer = S->getPointerOperand();
69 PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType());
70 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
71 Pointer = V->getOperand(0);
72 PointerSize = TD->getTypeStoreSize(V->getType());
73 } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) {
74 Pointer = F->getPointerOperand();
76 // FreeInsts erase the entire structure
78 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
79 CallSite InstCS = CallSite::get(Inst);
80 // If these two calls do not interfere, look past it.
81 if (AA->getModRefInfo(CS, InstCS) == AliasAnalysis::NoModRef)
84 // FIXME: If this is a ref/ref result, we should ignore it!
87 // Z = strlen(P); // Z = X
89 // If they interfere, we generally return clobber. However, if they are
90 // calls to the same read-only functions we return Def.
91 if (!AA->onlyReadsMemory(CS) || CS.getCalledFunction() == 0 ||
92 CS.getCalledFunction() != InstCS.getCalledFunction())
93 return MemDepResult::getClobber(Inst);
94 return MemDepResult::getDef(Inst);
96 // Non-memory instruction.
100 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef)
101 return MemDepResult::getClobber(Inst);
104 // No dependence found.
105 return MemDepResult::getNonLocal();
108 /// getDependencyFrom - Return the instruction on which a memory operation
110 MemDepResult MemoryDependenceAnalysis::
111 getDependencyFrom(Instruction *QueryInst, BasicBlock::iterator ScanIt,
113 // The first instruction in a block is always non-local.
114 if (ScanIt == BB->begin())
115 return MemDepResult::getNonLocal();
117 // Get the pointer value for which dependence will be determined
119 uint64_t MemSize = 0;
121 if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
122 // If this is a volatile store, don't mess around with it. Just return the
123 // previous instruction as a clobber.
124 if (SI->isVolatile())
125 return MemDepResult::getClobber(--ScanIt);
127 MemPtr = SI->getPointerOperand();
128 MemSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
129 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
130 // If this is a volatile load, don't mess around with it. Just return the
131 // previous instruction as a clobber.
132 if (LI->isVolatile())
133 return MemDepResult::getClobber(--ScanIt);
135 MemPtr = LI->getPointerOperand();
136 MemSize = TD->getTypeStoreSize(LI->getType());
137 } else if (FreeInst *FI = dyn_cast<FreeInst>(QueryInst)) {
138 MemPtr = FI->getPointerOperand();
139 // FreeInsts erase the entire structure, not just a field.
141 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
142 return getCallSiteDependencyFrom(CallSite::get(QueryInst), ScanIt, BB);
144 // Otherwise, this is a vaarg or non-memory instruction, just return a
145 // clobber dependency on the previous inst.
146 return MemDepResult::getClobber(--ScanIt);
149 // Walk backwards through the basic block, looking for dependencies
150 while (ScanIt != BB->begin()) {
151 Instruction *Inst = --ScanIt;
153 // Values depend on loads if the pointers are must aliased. This means that
154 // a load depends on another must aliased load from the same value.
155 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
156 Value *Pointer = LI->getPointerOperand();
157 uint64_t PointerSize = TD->getTypeStoreSize(LI->getType());
159 // If we found a pointer, check if it could be the same as our pointer.
160 AliasAnalysis::AliasResult R =
161 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
162 if (R == AliasAnalysis::NoAlias)
165 // May-alias loads don't depend on each other without a dependence.
166 if (isa<LoadInst>(QueryInst) && R == AliasAnalysis::MayAlias)
168 return MemDepResult::getDef(Inst);
171 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
172 Value *Pointer = SI->getPointerOperand();
173 uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
175 // If we found a pointer, check if it could be the same as our pointer.
176 AliasAnalysis::AliasResult R =
177 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
179 if (R == AliasAnalysis::NoAlias)
181 if (R == AliasAnalysis::MayAlias)
182 return MemDepResult::getClobber(Inst);
183 return MemDepResult::getDef(Inst);
186 // If this is an allocation, and if we know that the accessed pointer is to
187 // the allocation, return Def. This means that there is no dependence and
188 // the access can be optimized based on that. For example, a load could
190 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
191 Value *AccessPtr = MemPtr->getUnderlyingObject();
193 if (AccessPtr == AI ||
194 AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
195 return MemDepResult::getDef(AI);
199 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
200 if (AA->getModRefInfo(Inst, MemPtr, MemSize) == AliasAnalysis::NoModRef)
203 // Otherwise, there is a dependence.
204 return MemDepResult::getClobber(Inst);
207 // If we found nothing, return the non-local flag.
208 return MemDepResult::getNonLocal();
211 /// getDependency - Return the instruction on which a memory operation
213 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
214 Instruction *ScanPos = QueryInst;
216 // Check for a cached result
217 MemDepResult &LocalCache = LocalDeps[QueryInst];
219 // If the cached entry is non-dirty, just return it. Note that this depends
220 // on MemDepResult's default constructing to 'dirty'.
221 if (!LocalCache.isDirty())
224 // Otherwise, if we have a dirty entry, we know we can start the scan at that
225 // instruction, which may save us some work.
226 if (Instruction *Inst = LocalCache.getInst()) {
229 SmallPtrSet<Instruction*, 4> &InstMap = ReverseLocalDeps[Inst];
230 InstMap.erase(QueryInst);
232 ReverseLocalDeps.erase(Inst);
236 LocalCache = getDependencyFrom(QueryInst, ScanPos, QueryInst->getParent());
238 // Remember the result!
239 if (Instruction *I = LocalCache.getInst())
240 ReverseLocalDeps[I].insert(QueryInst);
245 /// getNonLocalDependency - Perform a full dependency query for the
246 /// specified instruction, returning the set of blocks that the value is
247 /// potentially live across. The returned set of results will include a
248 /// "NonLocal" result for all blocks where the value is live across.
250 /// This method assumes the instruction returns a "nonlocal" dependency
251 /// within its own block.
253 const MemoryDependenceAnalysis::NonLocalDepInfo &
254 MemoryDependenceAnalysis::getNonLocalDependency(Instruction *QueryInst) {
255 assert(getDependency(QueryInst).isNonLocal() &&
256 "getNonLocalDependency should only be used on insts with non-local deps!");
257 PerInstNLInfo &CacheP = NonLocalDeps[QueryInst];
259 NonLocalDepInfo &Cache = CacheP.first;
261 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
262 /// the cached case, this can happen due to instructions being deleted etc. In
263 /// the uncached case, this starts out as the set of predecessors we care
265 SmallVector<BasicBlock*, 32> DirtyBlocks;
267 if (!Cache.empty()) {
268 // Okay, we have a cache entry. If we know it is not dirty, just return it
269 // with no computation.
270 if (!CacheP.second) {
275 // If we already have a partially computed set of results, scan them to
276 // determine what is dirty, seeding our initial DirtyBlocks worklist.
277 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
279 if (I->second.isDirty())
280 DirtyBlocks.push_back(I->first);
282 // Sort the cache so that we can do fast binary search lookups below.
283 std::sort(Cache.begin(), Cache.end());
285 ++NumCacheDirtyNonLocal;
286 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
287 // << Cache.size() << " cached: " << *QueryInst;
289 // Seed DirtyBlocks with each of the preds of QueryInst's block.
290 BasicBlock *QueryBB = QueryInst->getParent();
291 DirtyBlocks.append(pred_begin(QueryBB), pred_end(QueryBB));
292 NumUncacheNonLocal++;
295 // Visited checked first, vector in sorted order.
296 SmallPtrSet<BasicBlock*, 64> Visited;
298 unsigned NumSortedEntries = Cache.size();
300 // Iterate while we still have blocks to update.
301 while (!DirtyBlocks.empty()) {
302 BasicBlock *DirtyBB = DirtyBlocks.back();
303 DirtyBlocks.pop_back();
305 // Already processed this block?
306 if (!Visited.insert(DirtyBB))
309 // Do a binary search to see if we already have an entry for this block in
310 // the cache set. If so, find it.
311 NonLocalDepInfo::iterator Entry =
312 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
313 std::make_pair(DirtyBB, MemDepResult()));
314 if (Entry != Cache.begin() && (&*Entry)[-1].first == DirtyBB)
317 MemDepResult *ExistingResult = 0;
318 if (Entry != Cache.begin()+NumSortedEntries &&
319 Entry->first == DirtyBB) {
320 // If we already have an entry, and if it isn't already dirty, the block
322 if (!Entry->second.isDirty())
325 // Otherwise, remember this slot so we can update the value.
326 ExistingResult = &Entry->second;
329 // If the dirty entry has a pointer, start scanning from it so we don't have
330 // to rescan the entire block.
331 BasicBlock::iterator ScanPos = DirtyBB->end();
332 if (ExistingResult) {
333 if (Instruction *Inst = ExistingResult->getInst()) {
336 // We're removing QueryInst's use of Inst.
337 SmallPtrSet<Instruction*, 4> &InstMap = ReverseNonLocalDeps[Inst];
338 InstMap.erase(QueryInst);
339 if (InstMap.empty()) ReverseNonLocalDeps.erase(Inst);
343 // Find out if this block has a local dependency for QueryInst.
344 MemDepResult Dep = getDependencyFrom(QueryInst, ScanPos, DirtyBB);
346 // If we had a dirty entry for the block, update it. Otherwise, just add
349 *ExistingResult = Dep;
351 Cache.push_back(std::make_pair(DirtyBB, Dep));
353 // If the block has a dependency (i.e. it isn't completely transparent to
354 // the value), remember the association!
355 if (!Dep.isNonLocal()) {
356 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
357 // update this when we remove instructions.
358 if (Instruction *Inst = Dep.getInst())
359 ReverseNonLocalDeps[Inst].insert(QueryInst);
362 // If the block *is* completely transparent to the load, we need to check
363 // the predecessors of this block. Add them to our worklist.
364 DirtyBlocks.append(pred_begin(DirtyBB), pred_end(DirtyBB));
371 /// removeInstruction - Remove an instruction from the dependence analysis,
372 /// updating the dependence of instructions that previously depended on it.
373 /// This method attempts to keep the cache coherent using the reverse map.
374 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
375 // Walk through the Non-local dependencies, removing this one as the value
376 // for any cached queries.
377 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
378 if (NLDI != NonLocalDeps.end()) {
379 NonLocalDepInfo &BlockMap = NLDI->second.first;
380 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
382 if (Instruction *Inst = DI->second.getInst())
383 ReverseNonLocalDeps[Inst].erase(RemInst);
384 NonLocalDeps.erase(NLDI);
387 // If we have a cached local dependence query for this instruction, remove it.
389 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
390 if (LocalDepEntry != LocalDeps.end()) {
391 // Remove us from DepInst's reverse set now that the local dep info is gone.
392 if (Instruction *Inst = LocalDepEntry->second.getInst()) {
393 SmallPtrSet<Instruction*, 4> &RLD = ReverseLocalDeps[Inst];
396 ReverseLocalDeps.erase(Inst);
399 // Remove this local dependency info.
400 LocalDeps.erase(LocalDepEntry);
403 // Loop over all of the things that depend on the instruction we're removing.
405 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
407 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
408 if (ReverseDepIt != ReverseLocalDeps.end()) {
409 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
410 // RemInst can't be the terminator if it has stuff depending on it.
411 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
412 "Nothing can locally depend on a terminator");
414 // Anything that was locally dependent on RemInst is now going to be
415 // dependent on the instruction after RemInst. It will have the dirty flag
416 // set so it will rescan. This saves having to scan the entire block to get
418 Instruction *NewDepInst = next(BasicBlock::iterator(RemInst));
420 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
421 E = ReverseDeps.end(); I != E; ++I) {
422 Instruction *InstDependingOnRemInst = *I;
423 assert(InstDependingOnRemInst != RemInst &&
424 "Already removed our local dep info");
426 LocalDeps[InstDependingOnRemInst] = MemDepResult::getDirty(NewDepInst);
428 // Make sure to remember that new things depend on NewDepInst.
429 ReverseDepsToAdd.push_back(std::make_pair(NewDepInst,
430 InstDependingOnRemInst));
433 ReverseLocalDeps.erase(ReverseDepIt);
435 // Add new reverse deps after scanning the set, to avoid invalidating the
436 // 'ReverseDeps' reference.
437 while (!ReverseDepsToAdd.empty()) {
438 ReverseLocalDeps[ReverseDepsToAdd.back().first]
439 .insert(ReverseDepsToAdd.back().second);
440 ReverseDepsToAdd.pop_back();
444 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
445 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
446 SmallPtrSet<Instruction*, 4>& set = ReverseDepIt->second;
447 for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
449 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
451 PerInstNLInfo &INLD = NonLocalDeps[*I];
452 // The information is now dirty!
455 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
456 DE = INLD.first.end(); DI != DE; ++DI) {
457 if (DI->second.getInst() != RemInst) continue;
459 // Convert to a dirty entry for the subsequent instruction.
460 Instruction *NextI = 0;
461 if (!RemInst->isTerminator()) {
462 NextI = next(BasicBlock::iterator(RemInst));
463 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
465 DI->second = MemDepResult::getDirty(NextI);
469 ReverseNonLocalDeps.erase(ReverseDepIt);
471 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
472 while (!ReverseDepsToAdd.empty()) {
473 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
474 .insert(ReverseDepsToAdd.back().second);
475 ReverseDepsToAdd.pop_back();
479 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
480 AA->deleteValue(RemInst);
481 DEBUG(verifyRemoved(RemInst));
484 /// verifyRemoved - Verify that the specified instruction does not occur
485 /// in our internal data structures.
486 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
487 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
488 E = LocalDeps.end(); I != E; ++I) {
489 assert(I->first != D && "Inst occurs in data structures");
490 assert(I->second.getInst() != D &&
491 "Inst occurs in data structures");
494 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
495 E = NonLocalDeps.end(); I != E; ++I) {
496 assert(I->first != D && "Inst occurs in data structures");
497 const PerInstNLInfo &INLD = I->second;
498 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
499 EE = INLD.first.end(); II != EE; ++II)
500 assert(II->second.getInst() != D && "Inst occurs in data structures");
503 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
504 E = ReverseLocalDeps.end(); I != E; ++I) {
505 assert(I->first != D && "Inst occurs in data structures");
506 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
507 EE = I->second.end(); II != EE; ++II)
508 assert(*II != D && "Inst occurs in data structures");
511 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
512 E = ReverseNonLocalDeps.end();
514 assert(I->first != D && "Inst occurs in data structures");
515 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
516 EE = I->second.end(); II != EE; ++II)
517 assert(*II != D && "Inst occurs in data structures");