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 cached non-local responses");
32 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 /// getCallSiteDependency - Private helper for finding the local dependencies
50 MemoryDependenceAnalysis::DepResultTy MemoryDependenceAnalysis::
51 getCallSiteDependency(CallSite C, BasicBlock::iterator ScanIt,
53 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
54 TargetData &TD = getAnalysis<TargetData>();
56 // Walk backwards through the block, looking for dependencies
57 while (ScanIt != BB->begin()) {
58 Instruction *Inst = --ScanIt;
60 // If this inst is a memory op, get the pointer it accessed
62 uint64_t PointerSize = 0;
63 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
64 Pointer = S->getPointerOperand();
65 PointerSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
66 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
67 Pointer = V->getOperand(0);
68 PointerSize = TD.getTypeStoreSize(V->getType());
69 } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) {
70 Pointer = F->getPointerOperand();
72 // FreeInsts erase the entire structure
74 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
75 if (AA.getModRefBehavior(CallSite::get(Inst)) ==
76 AliasAnalysis::DoesNotAccessMemory)
78 return DepResultTy(Inst, Normal);
80 // Non-memory instruction.
84 if (AA.getModRefInfo(C, Pointer, PointerSize) != AliasAnalysis::NoModRef)
85 return DepResultTy(Inst, Normal);
88 // No dependence found.
89 return DepResultTy(0, NonLocal);
92 /// getDependency - Return the instruction on which a memory operation
93 /// depends. The local parameter indicates if the query should only
94 /// evaluate dependencies within the same basic block.
95 MemoryDependenceAnalysis::DepResultTy MemoryDependenceAnalysis::
96 getDependencyFromInternal(Instruction *QueryInst, BasicBlock::iterator ScanIt,
98 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
99 TargetData &TD = getAnalysis<TargetData>();
101 // Get the pointer value for which dependence will be determined
103 uint64_t MemSize = 0;
104 bool MemVolatile = false;
106 if (StoreInst* S = dyn_cast<StoreInst>(QueryInst)) {
107 MemPtr = S->getPointerOperand();
108 MemSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
109 MemVolatile = S->isVolatile();
110 } else if (LoadInst* L = dyn_cast<LoadInst>(QueryInst)) {
111 MemPtr = L->getPointerOperand();
112 MemSize = TD.getTypeStoreSize(L->getType());
113 MemVolatile = L->isVolatile();
114 } else if (VAArgInst* V = dyn_cast<VAArgInst>(QueryInst)) {
115 MemPtr = V->getOperand(0);
116 MemSize = TD.getTypeStoreSize(V->getType());
117 } else if (FreeInst* F = dyn_cast<FreeInst>(QueryInst)) {
118 MemPtr = F->getPointerOperand();
119 // FreeInsts erase the entire structure, not just a field.
121 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst))
122 return getCallSiteDependency(CallSite::get(QueryInst), ScanIt, BB);
123 else // Non-memory instructions depend on nothing.
124 return DepResultTy(0, None);
126 // Walk backwards through the basic block, looking for dependencies
127 while (ScanIt != BB->begin()) {
128 Instruction *Inst = --ScanIt;
130 // If the access is volatile and this is a volatile load/store, return a
133 ((isa<LoadInst>(Inst) && cast<LoadInst>(Inst)->isVolatile()) ||
134 (isa<StoreInst>(Inst) && cast<StoreInst>(Inst)->isVolatile())))
135 return DepResultTy(Inst, Normal);
137 // Values depend on loads if the pointers are must aliased. This means that
138 // a load depends on another must aliased load from the same value.
139 if (LoadInst *L = dyn_cast<LoadInst>(Inst)) {
140 Value *Pointer = L->getPointerOperand();
141 uint64_t PointerSize = TD.getTypeStoreSize(L->getType());
143 // If we found a pointer, check if it could be the same as our pointer
144 AliasAnalysis::AliasResult R =
145 AA.alias(Pointer, PointerSize, MemPtr, MemSize);
147 if (R == AliasAnalysis::NoAlias)
150 // May-alias loads don't depend on each other without a dependence.
151 if (isa<LoadInst>(QueryInst) && R == AliasAnalysis::MayAlias)
153 return DepResultTy(Inst, Normal);
156 // If this is an allocation, and if we know that the accessed pointer is to
157 // the allocation, return None. This means that there is no dependence and
158 // the access can be optimized based on that. For example, a load could
160 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
161 Value *AccessPtr = MemPtr->getUnderlyingObject();
163 if (AccessPtr == AI ||
164 AA.alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
165 return DepResultTy(0, None);
169 // See if this instruction mod/ref's the pointer.
170 AliasAnalysis::ModRefResult MRR = AA.getModRefInfo(Inst, MemPtr, MemSize);
172 if (MRR == AliasAnalysis::NoModRef)
175 // Loads don't depend on read-only instructions.
176 if (isa<LoadInst>(QueryInst) && MRR == AliasAnalysis::Ref)
179 // Otherwise, there is a dependence.
180 return DepResultTy(Inst, Normal);
183 // If we found nothing, return the non-local flag.
184 return DepResultTy(0, NonLocal);
187 /// getDependency - Return the instruction on which a memory operation
189 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
190 Instruction *ScanPos = QueryInst;
192 // Check for a cached result
193 DepResultTy &LocalCache = LocalDeps[QueryInst];
195 // If the cached entry is non-dirty, just return it. Note that this depends
196 // on DepResultTy's default constructing to 'dirty'.
197 if (LocalCache.getInt() != Dirty)
198 return ConvToResult(LocalCache);
200 // Otherwise, if we have a dirty entry, we know we can start the scan at that
201 // instruction, which may save us some work.
202 if (Instruction *Inst = LocalCache.getPointer())
206 LocalCache = getDependencyFromInternal(QueryInst, ScanPos,
207 QueryInst->getParent());
209 // Remember the result!
210 if (Instruction *I = LocalCache.getPointer())
211 ReverseLocalDeps[I].insert(QueryInst);
213 return ConvToResult(LocalCache);
216 /// getNonLocalDependency - Perform a full dependency query for the
217 /// specified instruction, returning the set of blocks that the value is
218 /// potentially live across. The returned set of results will include a
219 /// "NonLocal" result for all blocks where the value is live across.
221 /// This method assumes the instruction returns a "nonlocal" dependency
222 /// within its own block.
224 void MemoryDependenceAnalysis::
225 getNonLocalDependency(Instruction *QueryInst,
226 SmallVectorImpl<std::pair<BasicBlock*,
227 MemDepResult> > &Result) {
228 assert(getDependency(QueryInst).isNonLocal() &&
229 "getNonLocalDependency should only be used on insts with non-local deps!");
230 DenseMap<BasicBlock*, DepResultTy>* &CacheP = NonLocalDeps[QueryInst];
231 if (CacheP == 0) CacheP = new DenseMap<BasicBlock*, DepResultTy>();
233 DenseMap<BasicBlock*, DepResultTy> &Cache = *CacheP;
235 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
236 /// the cached case, this can happen due to instructions being deleted etc. In
237 /// the uncached case, this starts out as the set of predecessors we care
239 SmallVector<BasicBlock*, 32> DirtyBlocks;
241 if (!Cache.empty()) {
242 // If we already have a partially computed set of results, scan them to
243 // determine what is dirty, seeding our initial DirtyBlocks worklist.
244 // FIXME: In the "don't need to be updated" case, this is expensive, why not
245 // have a per-"cache" flag saying it is undirty?
246 for (DenseMap<BasicBlock*, DepResultTy>::iterator I = Cache.begin(),
247 E = Cache.end(); I != E; ++I)
248 if (I->second.getInt() == Dirty)
249 DirtyBlocks.push_back(I->first);
253 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
254 // << Cache.size() << " cached: " << *QueryInst;
256 // Seed DirtyBlocks with each of the preds of QueryInst's block.
257 BasicBlock *QueryBB = QueryInst->getParent();
258 DirtyBlocks.append(pred_begin(QueryBB), pred_end(QueryBB));
259 NumUncacheNonLocal++;
262 // Iterate while we still have blocks to update.
263 while (!DirtyBlocks.empty()) {
264 BasicBlock *DirtyBB = DirtyBlocks.back();
265 DirtyBlocks.pop_back();
267 // Get the entry for this block. Note that this relies on DepResultTy
268 // default initializing to Dirty.
269 DepResultTy &DirtyBBEntry = Cache[DirtyBB];
271 // If DirtyBBEntry isn't dirty, it ended up on the worklist multiple times.
272 if (DirtyBBEntry.getInt() != Dirty) continue;
274 // If the dirty entry has a pointer, start scanning from it so we don't have
275 // to rescan the entire block.
276 BasicBlock::iterator ScanPos = DirtyBB->end();
277 if (Instruction *Inst = DirtyBBEntry.getPointer()) {
280 // We're removing QueryInst's dependence on Inst.
281 SmallPtrSet<Instruction*, 4> &InstMap = ReverseNonLocalDeps[Inst];
282 InstMap.erase(QueryInst);
283 if (InstMap.empty()) ReverseNonLocalDeps.erase(Inst);
286 // Find out if this block has a local dependency for QueryInst.
287 DirtyBBEntry = getDependencyFromInternal(QueryInst, ScanPos, DirtyBB);
289 // If the block has a dependency (i.e. it isn't completely transparent to
290 // the value), remember it!
291 if (DirtyBBEntry.getInt() != NonLocal) {
292 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
293 // update this when we remove instructions.
294 if (Instruction *Inst = DirtyBBEntry.getPointer())
295 ReverseNonLocalDeps[Inst].insert(QueryInst);
299 // If the block *is* completely transparent to the load, we need to check
300 // the predecessors of this block. Add them to our worklist.
301 DirtyBlocks.append(pred_begin(DirtyBB), pred_end(DirtyBB));
305 // Copy the result into the output set.
306 for (DenseMap<BasicBlock*, DepResultTy>::iterator I = Cache.begin(),
307 E = Cache.end(); I != E; ++I)
308 Result.push_back(std::make_pair(I->first, ConvToResult(I->second)));
311 /// removeInstruction - Remove an instruction from the dependence analysis,
312 /// updating the dependence of instructions that previously depended on it.
313 /// This method attempts to keep the cache coherent using the reverse map.
314 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
315 // Walk through the Non-local dependencies, removing this one as the value
316 // for any cached queries.
317 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
318 if (NLDI != NonLocalDeps.end()) {
319 DenseMap<BasicBlock*, DepResultTy> &BlockMap = *NLDI->second;
320 for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
321 BlockMap.begin(), DE = BlockMap.end(); DI != DE; ++DI)
322 if (Instruction *Inst = DI->second.getPointer())
323 ReverseNonLocalDeps[Inst].erase(RemInst);
325 NonLocalDeps.erase(NLDI);
328 // If we have a cached local dependence query for this instruction, remove it.
330 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
331 if (LocalDepEntry != LocalDeps.end()) {
332 // Remove us from DepInst's reverse set now that the local dep info is gone.
333 if (Instruction *Inst = LocalDepEntry->second.getPointer()) {
334 SmallPtrSet<Instruction*, 4> &RLD = ReverseLocalDeps[Inst];
337 ReverseLocalDeps.erase(Inst);
340 // Remove this local dependency info.
341 LocalDeps.erase(LocalDepEntry);
344 // Loop over all of the things that depend on the instruction we're removing.
346 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
348 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
349 if (ReverseDepIt != ReverseLocalDeps.end()) {
350 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
351 // RemInst can't be the terminator if it has stuff depending on it.
352 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
353 "Nothing can locally depend on a terminator");
355 // Anything that was locally dependent on RemInst is now going to be
356 // dependent on the instruction after RemInst. It will have the dirty flag
357 // set so it will rescan. This saves having to scan the entire block to get
359 Instruction *NewDepInst = next(BasicBlock::iterator(RemInst));
361 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
362 E = ReverseDeps.end(); I != E; ++I) {
363 Instruction *InstDependingOnRemInst = *I;
364 assert(InstDependingOnRemInst != RemInst &&
365 "Already removed our local dep info");
367 LocalDeps[InstDependingOnRemInst] = DepResultTy(NewDepInst, Dirty);
369 // Make sure to remember that new things depend on NewDepInst.
370 ReverseDepsToAdd.push_back(std::make_pair(NewDepInst,
371 InstDependingOnRemInst));
374 ReverseLocalDeps.erase(ReverseDepIt);
376 // Add new reverse deps after scanning the set, to avoid invalidating the
377 // 'ReverseDeps' reference.
378 while (!ReverseDepsToAdd.empty()) {
379 ReverseLocalDeps[ReverseDepsToAdd.back().first]
380 .insert(ReverseDepsToAdd.back().second);
381 ReverseDepsToAdd.pop_back();
385 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
386 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
387 SmallPtrSet<Instruction*, 4>& set = ReverseDepIt->second;
388 for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
390 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
392 DenseMap<BasicBlock*, DepResultTy> &INLD = *NonLocalDeps[*I];
393 assert(&INLD != 0 && "Reverse mapping out of date?");
395 for (DenseMap<BasicBlock*, DepResultTy>::iterator
396 DI = INLD.begin(), DE = INLD.end(); DI != DE; ++DI) {
397 if (DI->second.getPointer() != RemInst) continue;
399 // Convert to a dirty entry for the subsequent instruction.
400 DI->second.setInt(Dirty);
401 if (RemInst->isTerminator())
402 DI->second.setPointer(0);
404 Instruction *NextI = next(BasicBlock::iterator(RemInst));
405 DI->second.setPointer(NextI);
406 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
411 ReverseNonLocalDeps.erase(ReverseDepIt);
413 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
414 while (!ReverseDepsToAdd.empty()) {
415 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
416 .insert(ReverseDepsToAdd.back().second);
417 ReverseDepsToAdd.pop_back();
421 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
422 getAnalysis<AliasAnalysis>().deleteValue(RemInst);
423 DEBUG(verifyRemoved(RemInst));
426 /// verifyRemoved - Verify that the specified instruction does not occur
427 /// in our internal data structures.
428 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
429 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
430 E = LocalDeps.end(); I != E; ++I) {
431 assert(I->first != D && "Inst occurs in data structures");
432 assert(I->second.getPointer() != D &&
433 "Inst occurs in data structures");
436 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
437 E = NonLocalDeps.end(); I != E; ++I) {
438 assert(I->first != D && "Inst occurs in data structures");
439 DenseMap<BasicBlock*, DepResultTy> &INLD = *I->second;
440 for (DenseMap<BasicBlock*, DepResultTy>::iterator II = INLD.begin(),
441 EE = INLD.end(); II != EE; ++II)
442 assert(II->second.getPointer() != D && "Inst occurs in data structures");
445 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
446 E = ReverseLocalDeps.end(); I != E; ++I) {
447 assert(I->first != D && "Inst occurs in data structures");
448 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
449 EE = I->second.end(); II != EE; ++II)
450 assert(*II != D && "Inst occurs in data structures");
453 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
454 E = ReverseNonLocalDeps.end();
456 assert(I->first != D && "Inst occurs in data structures");
457 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
458 EE = I->second.end(); II != EE; ++II)
459 assert(*II != D && "Inst occurs in data structures");