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/IntrinsicInst.h"
22 #include "llvm/Function.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/Support/PredIteratorCache.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");
35 STATISTIC(NumCacheNonLocalPtr,
36 "Number of fully cached non-local ptr responses");
37 STATISTIC(NumCacheDirtyNonLocalPtr,
38 "Number of cached, but dirty, non-local ptr responses");
39 STATISTIC(NumUncacheNonLocalPtr,
40 "Number of uncached non-local ptr responses");
41 STATISTIC(NumCacheCompleteNonLocalPtr,
42 "Number of block queries that were completely cached");
44 char MemoryDependenceAnalysis::ID = 0;
46 // Register this pass...
47 static RegisterPass<MemoryDependenceAnalysis> X("memdep",
48 "Memory Dependence Analysis", false, true);
50 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
51 : FunctionPass(&ID), PredCache(0) {
53 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
56 /// Clean up memory in between runs
57 void MemoryDependenceAnalysis::releaseMemory() {
60 NonLocalPointerDeps.clear();
61 ReverseLocalDeps.clear();
62 ReverseNonLocalDeps.clear();
63 ReverseNonLocalPtrDeps.clear();
69 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
71 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
73 AU.addRequiredTransitive<AliasAnalysis>();
74 AU.addRequiredTransitive<TargetData>();
77 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
78 AA = &getAnalysis<AliasAnalysis>();
79 TD = &getAnalysis<TargetData>();
81 PredCache.reset(new PredIteratorCache());
85 /// RemoveFromReverseMap - This is a helper function that removes Val from
86 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
87 template <typename KeyTy>
88 static void RemoveFromReverseMap(DenseMap<Instruction*,
89 SmallPtrSet<KeyTy, 4> > &ReverseMap,
90 Instruction *Inst, KeyTy Val) {
91 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
92 InstIt = ReverseMap.find(Inst);
93 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
94 bool Found = InstIt->second.erase(Val);
95 assert(Found && "Invalid reverse map!"); Found=Found;
96 if (InstIt->second.empty())
97 ReverseMap.erase(InstIt);
101 /// getCallSiteDependencyFrom - Private helper for finding the local
102 /// dependencies of a call site.
103 MemDepResult MemoryDependenceAnalysis::
104 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
105 BasicBlock::iterator ScanIt, BasicBlock *BB) {
106 // Walk backwards through the block, looking for dependencies
107 while (ScanIt != BB->begin()) {
108 Instruction *Inst = --ScanIt;
110 // If this inst is a memory op, get the pointer it accessed
112 uint64_t PointerSize = 0;
113 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
114 Pointer = S->getPointerOperand();
115 PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType());
116 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
117 Pointer = V->getOperand(0);
118 PointerSize = TD->getTypeStoreSize(V->getType());
119 } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) {
120 Pointer = F->getPointerOperand();
122 // FreeInsts erase the entire structure
124 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
125 // Debug intrinsics don't cause dependences.
126 if (isa<DbgInfoIntrinsic>(Inst)) continue;
127 CallSite InstCS = CallSite::get(Inst);
128 // If these two calls do not interfere, look past it.
129 switch (AA->getModRefInfo(CS, InstCS)) {
130 case AliasAnalysis::NoModRef:
131 // If the two calls don't interact (e.g. InstCS is readnone) keep
134 case AliasAnalysis::Ref:
135 // If the two calls read the same memory locations and CS is a readonly
136 // function, then we have two cases: 1) the calls may not interfere with
137 // each other at all. 2) the calls may produce the same value. In case
138 // #1 we want to ignore the values, in case #2, we want to return Inst
139 // as a Def dependence. This allows us to CSE in cases like:
142 // Y = strlen(P); // Y = X
143 if (isReadOnlyCall) {
144 if (CS.getCalledFunction() != 0 &&
145 CS.getCalledFunction() == InstCS.getCalledFunction())
146 return MemDepResult::getDef(Inst);
147 // Ignore unrelated read/read call dependences.
152 return MemDepResult::getClobber(Inst);
155 // Non-memory instruction.
159 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef)
160 return MemDepResult::getClobber(Inst);
163 // No dependence found. If this is the entry block of the function, it is a
164 // clobber, otherwise it is non-local.
165 if (BB != &BB->getParent()->getEntryBlock())
166 return MemDepResult::getNonLocal();
167 return MemDepResult::getClobber(ScanIt);
170 /// getPointerDependencyFrom - Return the instruction on which a memory
171 /// location depends. If isLoad is true, this routine ignore may-aliases with
172 /// read-only operations.
173 MemDepResult MemoryDependenceAnalysis::
174 getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
175 BasicBlock::iterator ScanIt, BasicBlock *BB) {
177 // Walk backwards through the basic block, looking for dependencies.
178 while (ScanIt != BB->begin()) {
179 Instruction *Inst = --ScanIt;
181 // Debug intrinsics don't cause dependences.
182 if (isa<DbgInfoIntrinsic>(Inst)) continue;
184 // Values depend on loads if the pointers are must aliased. This means that
185 // a load depends on another must aliased load from the same value.
186 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
187 Value *Pointer = LI->getPointerOperand();
188 uint64_t PointerSize = TD->getTypeStoreSize(LI->getType());
190 // If we found a pointer, check if it could be the same as our pointer.
191 AliasAnalysis::AliasResult R =
192 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
193 if (R == AliasAnalysis::NoAlias)
196 // May-alias loads don't depend on each other without a dependence.
197 if (isLoad && R == AliasAnalysis::MayAlias)
199 // Stores depend on may and must aliased loads, loads depend on must-alias
201 return MemDepResult::getDef(Inst);
204 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
205 // If alias analysis can tell that this store is guaranteed to not modify
206 // the query pointer, ignore it. Use getModRefInfo to handle cases where
207 // the query pointer points to constant memory etc.
208 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef)
211 // Ok, this store might clobber the query pointer. Check to see if it is
212 // a must alias: in this case, we want to return this as a def.
213 Value *Pointer = SI->getPointerOperand();
214 uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
216 // If we found a pointer, check if it could be the same as our pointer.
217 AliasAnalysis::AliasResult R =
218 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
220 if (R == AliasAnalysis::NoAlias)
222 if (R == AliasAnalysis::MayAlias)
223 return MemDepResult::getClobber(Inst);
224 return MemDepResult::getDef(Inst);
227 // If this is an allocation, and if we know that the accessed pointer is to
228 // the allocation, return Def. This means that there is no dependence and
229 // the access can be optimized based on that. For example, a load could
231 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
232 Value *AccessPtr = MemPtr->getUnderlyingObject();
234 if (AccessPtr == AI ||
235 AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
236 return MemDepResult::getDef(AI);
240 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
241 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
242 case AliasAnalysis::NoModRef:
243 // If the call has no effect on the queried pointer, just ignore it.
245 case AliasAnalysis::Ref:
246 // If the call is known to never store to the pointer, and if this is a
247 // load query, we can safely ignore it (scan past it).
252 // Otherwise, there is a potential dependence. Return a clobber.
253 return MemDepResult::getClobber(Inst);
257 // No dependence found. If this is the entry block of the function, it is a
258 // clobber, otherwise it is non-local.
259 if (BB != &BB->getParent()->getEntryBlock())
260 return MemDepResult::getNonLocal();
261 return MemDepResult::getClobber(ScanIt);
264 /// getDependency - Return the instruction on which a memory operation
266 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
267 Instruction *ScanPos = QueryInst;
269 // Check for a cached result
270 MemDepResult &LocalCache = LocalDeps[QueryInst];
272 // If the cached entry is non-dirty, just return it. Note that this depends
273 // on MemDepResult's default constructing to 'dirty'.
274 if (!LocalCache.isDirty())
277 // Otherwise, if we have a dirty entry, we know we can start the scan at that
278 // instruction, which may save us some work.
279 if (Instruction *Inst = LocalCache.getInst()) {
282 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
285 BasicBlock *QueryParent = QueryInst->getParent();
288 uint64_t MemSize = 0;
291 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
292 // No dependence found. If this is the entry block of the function, it is a
293 // clobber, otherwise it is non-local.
294 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
295 LocalCache = MemDepResult::getNonLocal();
297 LocalCache = MemDepResult::getClobber(QueryInst);
298 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
299 // If this is a volatile store, don't mess around with it. Just return the
300 // previous instruction as a clobber.
301 if (SI->isVolatile())
302 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
304 MemPtr = SI->getPointerOperand();
305 MemSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
307 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
308 // If this is a volatile load, don't mess around with it. Just return the
309 // previous instruction as a clobber.
310 if (LI->isVolatile())
311 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
313 MemPtr = LI->getPointerOperand();
314 MemSize = TD->getTypeStoreSize(LI->getType());
316 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
317 CallSite QueryCS = CallSite::get(QueryInst);
318 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
319 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
321 } else if (FreeInst *FI = dyn_cast<FreeInst>(QueryInst)) {
322 MemPtr = FI->getPointerOperand();
323 // FreeInsts erase the entire structure, not just a field.
326 // Non-memory instruction.
327 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
330 // If we need to do a pointer scan, make it happen.
332 LocalCache = getPointerDependencyFrom(MemPtr, MemSize,
333 isa<LoadInst>(QueryInst),
334 ScanPos, QueryParent);
336 // Remember the result!
337 if (Instruction *I = LocalCache.getInst())
338 ReverseLocalDeps[I].insert(QueryInst);
344 /// AssertSorted - This method is used when -debug is specified to verify that
345 /// cache arrays are properly kept sorted.
346 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
348 if (Count == -1) Count = Cache.size();
349 if (Count == 0) return;
351 for (unsigned i = 1; i != unsigned(Count); ++i)
352 assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!");
356 /// getNonLocalCallDependency - Perform a full dependency query for the
357 /// specified call, returning the set of blocks that the value is
358 /// potentially live across. The returned set of results will include a
359 /// "NonLocal" result for all blocks where the value is live across.
361 /// This method assumes the instruction returns a "NonLocal" dependency
362 /// within its own block.
364 /// This returns a reference to an internal data structure that may be
365 /// invalidated on the next non-local query or when an instruction is
366 /// removed. Clients must copy this data if they want it around longer than
368 const MemoryDependenceAnalysis::NonLocalDepInfo &
369 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
370 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
371 "getNonLocalCallDependency should only be used on calls with non-local deps!");
372 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
373 NonLocalDepInfo &Cache = CacheP.first;
375 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
376 /// the cached case, this can happen due to instructions being deleted etc. In
377 /// the uncached case, this starts out as the set of predecessors we care
379 SmallVector<BasicBlock*, 32> DirtyBlocks;
381 if (!Cache.empty()) {
382 // Okay, we have a cache entry. If we know it is not dirty, just return it
383 // with no computation.
384 if (!CacheP.second) {
389 // If we already have a partially computed set of results, scan them to
390 // determine what is dirty, seeding our initial DirtyBlocks worklist.
391 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
393 if (I->second.isDirty())
394 DirtyBlocks.push_back(I->first);
396 // Sort the cache so that we can do fast binary search lookups below.
397 std::sort(Cache.begin(), Cache.end());
399 ++NumCacheDirtyNonLocal;
400 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
401 // << Cache.size() << " cached: " << *QueryInst;
403 // Seed DirtyBlocks with each of the preds of QueryInst's block.
404 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
405 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
406 DirtyBlocks.push_back(*PI);
407 NumUncacheNonLocal++;
410 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
411 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
413 SmallPtrSet<BasicBlock*, 64> Visited;
415 unsigned NumSortedEntries = Cache.size();
416 DEBUG(AssertSorted(Cache));
418 // Iterate while we still have blocks to update.
419 while (!DirtyBlocks.empty()) {
420 BasicBlock *DirtyBB = DirtyBlocks.back();
421 DirtyBlocks.pop_back();
423 // Already processed this block?
424 if (!Visited.insert(DirtyBB))
427 // Do a binary search to see if we already have an entry for this block in
428 // the cache set. If so, find it.
429 DEBUG(AssertSorted(Cache, NumSortedEntries));
430 NonLocalDepInfo::iterator Entry =
431 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
432 std::make_pair(DirtyBB, MemDepResult()));
433 if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB)
436 MemDepResult *ExistingResult = 0;
437 if (Entry != Cache.begin()+NumSortedEntries &&
438 Entry->first == DirtyBB) {
439 // If we already have an entry, and if it isn't already dirty, the block
441 if (!Entry->second.isDirty())
444 // Otherwise, remember this slot so we can update the value.
445 ExistingResult = &Entry->second;
448 // If the dirty entry has a pointer, start scanning from it so we don't have
449 // to rescan the entire block.
450 BasicBlock::iterator ScanPos = DirtyBB->end();
451 if (ExistingResult) {
452 if (Instruction *Inst = ExistingResult->getInst()) {
454 // We're removing QueryInst's use of Inst.
455 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
456 QueryCS.getInstruction());
460 // Find out if this block has a local dependency for QueryInst.
463 if (ScanPos != DirtyBB->begin()) {
464 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
465 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
466 // No dependence found. If this is the entry block of the function, it is
467 // a clobber, otherwise it is non-local.
468 Dep = MemDepResult::getNonLocal();
470 Dep = MemDepResult::getClobber(ScanPos);
473 // If we had a dirty entry for the block, update it. Otherwise, just add
476 *ExistingResult = Dep;
478 Cache.push_back(std::make_pair(DirtyBB, Dep));
480 // If the block has a dependency (i.e. it isn't completely transparent to
481 // the value), remember the association!
482 if (!Dep.isNonLocal()) {
483 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
484 // update this when we remove instructions.
485 if (Instruction *Inst = Dep.getInst())
486 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
489 // If the block *is* completely transparent to the load, we need to check
490 // the predecessors of this block. Add them to our worklist.
491 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
492 DirtyBlocks.push_back(*PI);
499 /// getNonLocalPointerDependency - Perform a full dependency query for an
500 /// access to the specified (non-volatile) memory location, returning the
501 /// set of instructions that either define or clobber the value.
503 /// This method assumes the pointer has a "NonLocal" dependency within its
506 void MemoryDependenceAnalysis::
507 getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
508 SmallVectorImpl<NonLocalDepEntry> &Result) {
509 assert(isa<PointerType>(Pointer->getType()) &&
510 "Can't get pointer deps of a non-pointer!");
513 // We know that the pointer value is live into FromBB find the def/clobbers
514 // from presecessors.
515 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
516 uint64_t PointeeSize = TD->getTypeStoreSize(EltTy);
518 // This is the set of blocks we've inspected, and the pointer we consider in
519 // each block. Because of critical edges, we currently bail out if querying
520 // a block with multiple different pointers. This can happen during PHI
522 DenseMap<BasicBlock*, Value*> Visited;
523 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB,
524 Result, Visited, true))
527 Result.push_back(std::make_pair(FromBB,
528 MemDepResult::getClobber(FromBB->begin())));
531 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
532 /// Pointer/PointeeSize using either cached information in Cache or by doing a
533 /// lookup (which may use dirty cache info if available). If we do a lookup,
534 /// add the result to the cache.
535 MemDepResult MemoryDependenceAnalysis::
536 GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
537 bool isLoad, BasicBlock *BB,
538 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
540 // Do a binary search to see if we already have an entry for this block in
541 // the cache set. If so, find it.
542 NonLocalDepInfo::iterator Entry =
543 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
544 std::make_pair(BB, MemDepResult()));
545 if (Entry != Cache->begin() && prior(Entry)->first == BB)
548 MemDepResult *ExistingResult = 0;
549 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
550 ExistingResult = &Entry->second;
552 // If we have a cached entry, and it is non-dirty, use it as the value for
554 if (ExistingResult && !ExistingResult->isDirty()) {
555 ++NumCacheNonLocalPtr;
556 return *ExistingResult;
559 // Otherwise, we have to scan for the value. If we have a dirty cache
560 // entry, start scanning from its position, otherwise we scan from the end
562 BasicBlock::iterator ScanPos = BB->end();
563 if (ExistingResult && ExistingResult->getInst()) {
564 assert(ExistingResult->getInst()->getParent() == BB &&
565 "Instruction invalidated?");
566 ++NumCacheDirtyNonLocalPtr;
567 ScanPos = ExistingResult->getInst();
569 // Eliminating the dirty entry from 'Cache', so update the reverse info.
570 ValueIsLoadPair CacheKey(Pointer, isLoad);
571 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
573 ++NumUncacheNonLocalPtr;
576 // Scan the block for the dependency.
577 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad,
580 // If we had a dirty entry for the block, update it. Otherwise, just add
583 *ExistingResult = Dep;
585 Cache->push_back(std::make_pair(BB, Dep));
587 // If the block has a dependency (i.e. it isn't completely transparent to
588 // the value), remember the reverse association because we just added it
590 if (Dep.isNonLocal())
593 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
594 // update MemDep when we remove instructions.
595 Instruction *Inst = Dep.getInst();
596 assert(Inst && "Didn't depend on anything?");
597 ValueIsLoadPair CacheKey(Pointer, isLoad);
598 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
603 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
604 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
605 /// results to the results vector and keep track of which blocks are visited in
608 /// This has special behavior for the first block queries (when SkipFirstBlock
609 /// is true). In this special case, it ignores the contents of the specified
610 /// block and starts returning dependence info for its predecessors.
612 /// This function returns false on success, or true to indicate that it could
613 /// not compute dependence information for some reason. This should be treated
614 /// as a clobber dependence on the first instruction in the predecessor block.
615 bool MemoryDependenceAnalysis::
616 getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
617 bool isLoad, BasicBlock *StartBB,
618 SmallVectorImpl<NonLocalDepEntry> &Result,
619 DenseMap<BasicBlock*, Value*> &Visited,
620 bool SkipFirstBlock) {
622 // Look up the cached info for Pointer.
623 ValueIsLoadPair CacheKey(Pointer, isLoad);
625 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
626 &NonLocalPointerDeps[CacheKey];
627 NonLocalDepInfo *Cache = &CacheInfo->second;
629 // If we have valid cached information for exactly the block we are
630 // investigating, just return it with no recomputation.
631 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
632 // We have a fully cached result for this query then we can just return the
633 // cached results and populate the visited set. However, we have to verify
634 // that we don't already have conflicting results for these blocks. Check
635 // to ensure that if a block in the results set is in the visited set that
636 // it was for the same pointer query.
637 if (!Visited.empty()) {
638 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
640 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->first);
641 if (VI == Visited.end() || VI->second == Pointer) continue;
643 // We have a pointer mismatch in a block. Just return clobber, saying
644 // that something was clobbered in this result. We could also do a
645 // non-fully cached query, but there is little point in doing this.
650 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
652 Visited.insert(std::make_pair(I->first, Pointer));
653 if (!I->second.isNonLocal())
654 Result.push_back(*I);
656 ++NumCacheCompleteNonLocalPtr;
660 // Otherwise, either this is a new block, a block with an invalid cache
661 // pointer or one that we're about to invalidate by putting more info into it
662 // than its valid cache info. If empty, the result will be valid cache info,
663 // otherwise it isn't.
665 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
667 CacheInfo->first = BBSkipFirstBlockPair();
669 SmallVector<BasicBlock*, 32> Worklist;
670 Worklist.push_back(StartBB);
672 // Keep track of the entries that we know are sorted. Previously cached
673 // entries will all be sorted. The entries we add we only sort on demand (we
674 // don't insert every element into its sorted position). We know that we
675 // won't get any reuse from currently inserted values, because we don't
676 // revisit blocks after we insert info for them.
677 unsigned NumSortedEntries = Cache->size();
678 DEBUG(AssertSorted(*Cache));
680 while (!Worklist.empty()) {
681 BasicBlock *BB = Worklist.pop_back_val();
683 // Skip the first block if we have it.
684 if (!SkipFirstBlock) {
685 // Analyze the dependency of *Pointer in FromBB. See if we already have
687 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
689 // Get the dependency info for Pointer in BB. If we have cached
690 // information, we will use it, otherwise we compute it.
691 DEBUG(AssertSorted(*Cache, NumSortedEntries));
692 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad,
693 BB, Cache, NumSortedEntries);
695 // If we got a Def or Clobber, add this to the list of results.
696 if (!Dep.isNonLocal()) {
697 Result.push_back(NonLocalDepEntry(BB, Dep));
702 // If 'Pointer' is an instruction defined in this block, then we need to do
703 // phi translation to change it into a value live in the predecessor block.
704 // If phi translation fails, then we can't continue dependence analysis.
705 Instruction *PtrInst = dyn_cast<Instruction>(Pointer);
706 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB;
708 // If no PHI translation is needed, just add all the predecessors of this
709 // block to scan them as well.
710 if (!NeedsPHITranslation) {
711 SkipFirstBlock = false;
712 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
713 // Verify that we haven't looked at this block yet.
714 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
715 InsertRes = Visited.insert(std::make_pair(*PI, Pointer));
716 if (InsertRes.second) {
717 // First time we've looked at *PI.
718 Worklist.push_back(*PI);
722 // If we have seen this block before, but it was with a different
723 // pointer then we have a phi translation failure and we have to treat
724 // this as a clobber.
725 if (InsertRes.first->second != Pointer)
726 goto PredTranslationFailure;
731 // If we do need to do phi translation, then there are a bunch of different
732 // cases, because we have to find a Value* live in the predecessor block. We
733 // know that PtrInst is defined in this block at least.
735 // If this is directly a PHI node, just use the incoming values for each
736 // pred as the phi translated version.
737 if (PHINode *PtrPHI = dyn_cast<PHINode>(PtrInst)) {
738 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
739 BasicBlock *Pred = *PI;
740 Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred);
742 // Check to see if we have already visited this pred block with another
743 // pointer. If so, we can't do this lookup. This failure can occur
744 // with PHI translation when a critical edge exists and the PHI node in
745 // the successor translates to a pointer value different than the
746 // pointer the block was first analyzed with.
747 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
748 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr));
750 if (!InsertRes.second) {
751 // If the predecessor was visited with PredPtr, then we already did
752 // the analysis and can ignore it.
753 if (InsertRes.first->second == PredPtr)
756 // Otherwise, the block was previously analyzed with a different
757 // pointer. We can't represent the result of this case, so we just
758 // treat this as a phi translation failure.
759 goto PredTranslationFailure;
762 // We may have added values to the cache list before this PHI
763 // translation. If so, we haven't done anything to ensure that the
764 // cache remains sorted. Sort it now (if needed) so that recursive
765 // invocations of getNonLocalPointerDepFromBB that could reuse the cache
766 // value will only see properly sorted cache arrays.
767 if (Cache && NumSortedEntries != Cache->size())
768 std::sort(Cache->begin(), Cache->end());
771 // FIXME: it is entirely possible that PHI translating will end up with
772 // the same value. Consider PHI translating something like:
773 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
774 // to recurse here, pedantically speaking.
776 // If we have a problem phi translating, fall through to the code below
777 // to handle the failure condition.
778 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred,
780 goto PredTranslationFailure;
783 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
784 CacheInfo = &NonLocalPointerDeps[CacheKey];
785 Cache = &CacheInfo->second;
786 NumSortedEntries = Cache->size();
788 // Since we did phi translation, the "Cache" set won't contain all of the
789 // results for the query. This is ok (we can still use it to accelerate
790 // specific block queries) but we can't do the fastpath "return all
791 // results from the set" Clear out the indicator for this.
792 CacheInfo->first = BBSkipFirstBlockPair();
793 SkipFirstBlock = false;
797 // TODO: BITCAST, GEP.
799 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
800 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
801 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
802 PredTranslationFailure:
805 // Refresh the CacheInfo/Cache pointer if it got invalidated.
806 CacheInfo = &NonLocalPointerDeps[CacheKey];
807 Cache = &CacheInfo->second;
808 NumSortedEntries = Cache->size();
809 } else if (NumSortedEntries != Cache->size()) {
810 std::sort(Cache->begin(), Cache->end());
811 NumSortedEntries = Cache->size();
814 // Since we did phi translation, the "Cache" set won't contain all of the
815 // results for the query. This is ok (we can still use it to accelerate
816 // specific block queries) but we can't do the fastpath "return all
817 // results from the set" Clear out the indicator for this.
818 CacheInfo->first = BBSkipFirstBlockPair();
820 // If *nothing* works, mark the pointer as being clobbered by the first
821 // instruction in this block.
823 // If this is the magic first block, return this as a clobber of the whole
824 // incoming value. Since we can't phi translate to one of the predecessors,
825 // we have to bail out.
829 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
830 assert(I != Cache->rend() && "Didn't find current block??");
834 assert(I->second.isNonLocal() &&
835 "Should only be here with transparent block");
836 I->second = MemDepResult::getClobber(BB->begin());
837 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
838 Result.push_back(*I);
843 // Okay, we're done now. If we added new values to the cache, re-sort it.
844 switch (Cache->size()-NumSortedEntries) {
846 // done, no new entries.
849 // Two new entries, insert the last one into place.
850 NonLocalDepEntry Val = Cache->back();
852 NonLocalDepInfo::iterator Entry =
853 std::upper_bound(Cache->begin(), Cache->end()-1, Val);
854 Cache->insert(Entry, Val);
858 // One new entry, Just insert the new value at the appropriate position.
859 if (Cache->size() != 1) {
860 NonLocalDepEntry Val = Cache->back();
862 NonLocalDepInfo::iterator Entry =
863 std::upper_bound(Cache->begin(), Cache->end(), Val);
864 Cache->insert(Entry, Val);
868 // Added many values, do a full scale sort.
869 std::sort(Cache->begin(), Cache->end());
871 DEBUG(AssertSorted(*Cache));
875 /// RemoveCachedNonLocalPointerDependencies - If P exists in
876 /// CachedNonLocalPointerInfo, remove it.
877 void MemoryDependenceAnalysis::
878 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
879 CachedNonLocalPointerInfo::iterator It =
880 NonLocalPointerDeps.find(P);
881 if (It == NonLocalPointerDeps.end()) return;
883 // Remove all of the entries in the BB->val map. This involves removing
884 // instructions from the reverse map.
885 NonLocalDepInfo &PInfo = It->second.second;
887 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
888 Instruction *Target = PInfo[i].second.getInst();
889 if (Target == 0) continue; // Ignore non-local dep results.
890 assert(Target->getParent() == PInfo[i].first);
892 // Eliminating the dirty entry from 'Cache', so update the reverse info.
893 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
896 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
897 NonLocalPointerDeps.erase(It);
901 /// invalidateCachedPointerInfo - This method is used to invalidate cached
902 /// information about the specified pointer, because it may be too
903 /// conservative in memdep. This is an optional call that can be used when
904 /// the client detects an equivalence between the pointer and some other
905 /// value and replaces the other value with ptr. This can make Ptr available
906 /// in more places that cached info does not necessarily keep.
907 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
908 // If Ptr isn't really a pointer, just ignore it.
909 if (!isa<PointerType>(Ptr->getType())) return;
910 // Flush store info for the pointer.
911 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
912 // Flush load info for the pointer.
913 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
916 /// removeInstruction - Remove an instruction from the dependence analysis,
917 /// updating the dependence of instructions that previously depended on it.
918 /// This method attempts to keep the cache coherent using the reverse map.
919 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
920 // Walk through the Non-local dependencies, removing this one as the value
921 // for any cached queries.
922 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
923 if (NLDI != NonLocalDeps.end()) {
924 NonLocalDepInfo &BlockMap = NLDI->second.first;
925 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
927 if (Instruction *Inst = DI->second.getInst())
928 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
929 NonLocalDeps.erase(NLDI);
932 // If we have a cached local dependence query for this instruction, remove it.
934 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
935 if (LocalDepEntry != LocalDeps.end()) {
936 // Remove us from DepInst's reverse set now that the local dep info is gone.
937 if (Instruction *Inst = LocalDepEntry->second.getInst())
938 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
940 // Remove this local dependency info.
941 LocalDeps.erase(LocalDepEntry);
944 // If we have any cached pointer dependencies on this instruction, remove
945 // them. If the instruction has non-pointer type, then it can't be a pointer
948 // Remove it from both the load info and the store info. The instruction
949 // can't be in either of these maps if it is non-pointer.
950 if (isa<PointerType>(RemInst->getType())) {
951 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
952 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
955 // Loop over all of the things that depend on the instruction we're removing.
957 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
959 // If we find RemInst as a clobber or Def in any of the maps for other values,
960 // we need to replace its entry with a dirty version of the instruction after
961 // it. If RemInst is a terminator, we use a null dirty value.
963 // Using a dirty version of the instruction after RemInst saves having to scan
964 // the entire block to get to this point.
965 MemDepResult NewDirtyVal;
966 if (!RemInst->isTerminator())
967 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
969 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
970 if (ReverseDepIt != ReverseLocalDeps.end()) {
971 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
972 // RemInst can't be the terminator if it has local stuff depending on it.
973 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
974 "Nothing can locally depend on a terminator");
976 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
977 E = ReverseDeps.end(); I != E; ++I) {
978 Instruction *InstDependingOnRemInst = *I;
979 assert(InstDependingOnRemInst != RemInst &&
980 "Already removed our local dep info");
982 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
984 // Make sure to remember that new things depend on NewDepInst.
985 assert(NewDirtyVal.getInst() && "There is no way something else can have "
986 "a local dep on this if it is a terminator!");
987 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
988 InstDependingOnRemInst));
991 ReverseLocalDeps.erase(ReverseDepIt);
993 // Add new reverse deps after scanning the set, to avoid invalidating the
994 // 'ReverseDeps' reference.
995 while (!ReverseDepsToAdd.empty()) {
996 ReverseLocalDeps[ReverseDepsToAdd.back().first]
997 .insert(ReverseDepsToAdd.back().second);
998 ReverseDepsToAdd.pop_back();
1002 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1003 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1004 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1005 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1007 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1009 PerInstNLInfo &INLD = NonLocalDeps[*I];
1010 // The information is now dirty!
1013 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1014 DE = INLD.first.end(); DI != DE; ++DI) {
1015 if (DI->second.getInst() != RemInst) continue;
1017 // Convert to a dirty entry for the subsequent instruction.
1018 DI->second = NewDirtyVal;
1020 if (Instruction *NextI = NewDirtyVal.getInst())
1021 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1025 ReverseNonLocalDeps.erase(ReverseDepIt);
1027 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1028 while (!ReverseDepsToAdd.empty()) {
1029 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1030 .insert(ReverseDepsToAdd.back().second);
1031 ReverseDepsToAdd.pop_back();
1035 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1036 // value in the NonLocalPointerDeps info.
1037 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1038 ReverseNonLocalPtrDeps.find(RemInst);
1039 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1040 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1041 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1043 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1044 E = Set.end(); I != E; ++I) {
1045 ValueIsLoadPair P = *I;
1046 assert(P.getPointer() != RemInst &&
1047 "Already removed NonLocalPointerDeps info for RemInst");
1049 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second;
1051 // The cache is not valid for any specific block anymore.
1052 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair();
1054 // Update any entries for RemInst to use the instruction after it.
1055 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1057 if (DI->second.getInst() != RemInst) continue;
1059 // Convert to a dirty entry for the subsequent instruction.
1060 DI->second = NewDirtyVal;
1062 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1063 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1066 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1067 // subsequent value may invalidate the sortedness.
1068 std::sort(NLPDI.begin(), NLPDI.end());
1071 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1073 while (!ReversePtrDepsToAdd.empty()) {
1074 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1075 .insert(ReversePtrDepsToAdd.back().second);
1076 ReversePtrDepsToAdd.pop_back();
1081 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1082 AA->deleteValue(RemInst);
1083 DEBUG(verifyRemoved(RemInst));
1085 /// verifyRemoved - Verify that the specified instruction does not occur
1086 /// in our internal data structures.
1087 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1088 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1089 E = LocalDeps.end(); I != E; ++I) {
1090 assert(I->first != D && "Inst occurs in data structures");
1091 assert(I->second.getInst() != D &&
1092 "Inst occurs in data structures");
1095 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1096 E = NonLocalPointerDeps.end(); I != E; ++I) {
1097 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1098 const NonLocalDepInfo &Val = I->second.second;
1099 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1101 assert(II->second.getInst() != D && "Inst occurs as NLPD value");
1104 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1105 E = NonLocalDeps.end(); I != E; ++I) {
1106 assert(I->first != D && "Inst occurs in data structures");
1107 const PerInstNLInfo &INLD = I->second;
1108 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1109 EE = INLD.first.end(); II != EE; ++II)
1110 assert(II->second.getInst() != D && "Inst occurs in data structures");
1113 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1114 E = ReverseLocalDeps.end(); I != E; ++I) {
1115 assert(I->first != D && "Inst occurs in data structures");
1116 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1117 EE = I->second.end(); II != EE; ++II)
1118 assert(*II != D && "Inst occurs in data structures");
1121 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1122 E = ReverseNonLocalDeps.end();
1124 assert(I->first != D && "Inst occurs in data structures");
1125 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1126 EE = I->second.end(); II != EE; ++II)
1127 assert(*II != D && "Inst occurs in data structures");
1130 for (ReverseNonLocalPtrDepTy::const_iterator
1131 I = ReverseNonLocalPtrDeps.begin(),
1132 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1133 assert(I->first != D && "Inst occurs in rev NLPD map");
1135 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1136 E = I->second.end(); II != E; ++II)
1137 assert(*II != ValueIsLoadPair(D, false) &&
1138 *II != ValueIsLoadPair(D, true) &&
1139 "Inst occurs in ReverseNonLocalPtrDeps map");