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/Instructions.h"
20 #include "llvm/IntrinsicInst.h"
21 #include "llvm/Function.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/MallocHelper.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"
30 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
31 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
32 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
34 STATISTIC(NumCacheNonLocalPtr,
35 "Number of fully cached non-local ptr responses");
36 STATISTIC(NumCacheDirtyNonLocalPtr,
37 "Number of cached, but dirty, non-local ptr responses");
38 STATISTIC(NumUncacheNonLocalPtr,
39 "Number of uncached non-local ptr responses");
40 STATISTIC(NumCacheCompleteNonLocalPtr,
41 "Number of block queries that were completely cached");
43 char MemoryDependenceAnalysis::ID = 0;
45 // Register this pass...
46 static RegisterPass<MemoryDependenceAnalysis> X("memdep",
47 "Memory Dependence Analysis", false, true);
49 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
50 : FunctionPass(&ID), PredCache(0) {
52 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
55 /// Clean up memory in between runs
56 void MemoryDependenceAnalysis::releaseMemory() {
59 NonLocalPointerDeps.clear();
60 ReverseLocalDeps.clear();
61 ReverseNonLocalDeps.clear();
62 ReverseNonLocalPtrDeps.clear();
68 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
70 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
72 AU.addRequiredTransitive<AliasAnalysis>();
75 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
76 AA = &getAnalysis<AliasAnalysis>();
78 PredCache.reset(new PredIteratorCache());
82 /// RemoveFromReverseMap - This is a helper function that removes Val from
83 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
84 template <typename KeyTy>
85 static void RemoveFromReverseMap(DenseMap<Instruction*,
86 SmallPtrSet<KeyTy, 4> > &ReverseMap,
87 Instruction *Inst, KeyTy Val) {
88 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
89 InstIt = ReverseMap.find(Inst);
90 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
91 bool Found = InstIt->second.erase(Val);
92 assert(Found && "Invalid reverse map!"); Found=Found;
93 if (InstIt->second.empty())
94 ReverseMap.erase(InstIt);
98 /// getCallSiteDependencyFrom - Private helper for finding the local
99 /// dependencies of a call site.
100 MemDepResult MemoryDependenceAnalysis::
101 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
102 BasicBlock::iterator ScanIt, BasicBlock *BB) {
103 // Walk backwards through the block, looking for dependencies
104 while (ScanIt != BB->begin()) {
105 Instruction *Inst = --ScanIt;
107 // If this inst is a memory op, get the pointer it accessed
109 uint64_t PointerSize = 0;
110 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
111 Pointer = S->getPointerOperand();
112 PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType());
113 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
114 Pointer = V->getOperand(0);
115 PointerSize = AA->getTypeStoreSize(V->getType());
116 } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) {
117 Pointer = F->getPointerOperand();
119 // FreeInsts erase the entire structure
121 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
122 // Debug intrinsics don't cause dependences.
123 if (isa<DbgInfoIntrinsic>(Inst)) continue;
124 CallSite InstCS = CallSite::get(Inst);
125 // If these two calls do not interfere, look past it.
126 switch (AA->getModRefInfo(CS, InstCS)) {
127 case AliasAnalysis::NoModRef:
128 // If the two calls don't interact (e.g. InstCS is readnone) keep
131 case AliasAnalysis::Ref:
132 // If the two calls read the same memory locations and CS is a readonly
133 // function, then we have two cases: 1) the calls may not interfere with
134 // each other at all. 2) the calls may produce the same value. In case
135 // #1 we want to ignore the values, in case #2, we want to return Inst
136 // as a Def dependence. This allows us to CSE in cases like:
139 // Y = strlen(P); // Y = X
140 if (isReadOnlyCall) {
141 if (CS.getCalledFunction() != 0 &&
142 CS.getCalledFunction() == InstCS.getCalledFunction())
143 return MemDepResult::getDef(Inst);
144 // Ignore unrelated read/read call dependences.
149 return MemDepResult::getClobber(Inst);
152 // Non-memory instruction.
156 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef)
157 return MemDepResult::getClobber(Inst);
160 // No dependence found. If this is the entry block of the function, it is a
161 // clobber, otherwise it is non-local.
162 if (BB != &BB->getParent()->getEntryBlock())
163 return MemDepResult::getNonLocal();
164 return MemDepResult::getClobber(ScanIt);
167 /// getPointerDependencyFrom - Return the instruction on which a memory
168 /// location depends. If isLoad is true, this routine ignore may-aliases with
169 /// read-only operations.
170 MemDepResult MemoryDependenceAnalysis::
171 getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
172 BasicBlock::iterator ScanIt, BasicBlock *BB) {
174 // Walk backwards through the basic block, looking for dependencies.
175 while (ScanIt != BB->begin()) {
176 Instruction *Inst = --ScanIt;
178 // Debug intrinsics don't cause dependences.
179 if (isa<DbgInfoIntrinsic>(Inst)) continue;
181 // Values depend on loads if the pointers are must aliased. This means that
182 // a load depends on another must aliased load from the same value.
183 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
184 Value *Pointer = LI->getPointerOperand();
185 uint64_t PointerSize = AA->getTypeStoreSize(LI->getType());
187 // If we found a pointer, check if it could be the same as our pointer.
188 AliasAnalysis::AliasResult R =
189 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
190 if (R == AliasAnalysis::NoAlias)
193 // May-alias loads don't depend on each other without a dependence.
194 if (isLoad && R == AliasAnalysis::MayAlias)
196 // Stores depend on may and must aliased loads, loads depend on must-alias
198 return MemDepResult::getDef(Inst);
201 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
202 // If alias analysis can tell that this store is guaranteed to not modify
203 // the query pointer, ignore it. Use getModRefInfo to handle cases where
204 // the query pointer points to constant memory etc.
205 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef)
208 // Ok, this store might clobber the query pointer. Check to see if it is
209 // a must alias: in this case, we want to return this as a def.
210 Value *Pointer = SI->getPointerOperand();
211 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
213 // If we found a pointer, check if it could be the same as our pointer.
214 AliasAnalysis::AliasResult R =
215 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
217 if (R == AliasAnalysis::NoAlias)
219 if (R == AliasAnalysis::MayAlias)
220 return MemDepResult::getClobber(Inst);
221 return MemDepResult::getDef(Inst);
224 // If this is an allocation, and if we know that the accessed pointer is to
225 // the allocation, return Def. This means that there is no dependence and
226 // the access can be optimized based on that. For example, a load could
228 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
229 Value *AccessPtr = MemPtr->getUnderlyingObject();
231 if (AccessPtr == AI ||
232 AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
233 return MemDepResult::getDef(AI);
237 if (isMalloc(Inst)) {
238 Value *AccessPtr = MemPtr->getUnderlyingObject();
240 if (AccessPtr == Inst ||
241 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
242 return MemDepResult::getDef(Inst);
246 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
247 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
248 case AliasAnalysis::NoModRef:
249 // If the call has no effect on the queried pointer, just ignore it.
251 case AliasAnalysis::Ref:
252 // If the call is known to never store to the pointer, and if this is a
253 // load query, we can safely ignore it (scan past it).
258 // Otherwise, there is a potential dependence. Return a clobber.
259 return MemDepResult::getClobber(Inst);
263 // No dependence found. If this is the entry block of the function, it is a
264 // clobber, otherwise it is non-local.
265 if (BB != &BB->getParent()->getEntryBlock())
266 return MemDepResult::getNonLocal();
267 return MemDepResult::getClobber(ScanIt);
270 /// getDependency - Return the instruction on which a memory operation
272 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
273 Instruction *ScanPos = QueryInst;
275 // Check for a cached result
276 MemDepResult &LocalCache = LocalDeps[QueryInst];
278 // If the cached entry is non-dirty, just return it. Note that this depends
279 // on MemDepResult's default constructing to 'dirty'.
280 if (!LocalCache.isDirty())
283 // Otherwise, if we have a dirty entry, we know we can start the scan at that
284 // instruction, which may save us some work.
285 if (Instruction *Inst = LocalCache.getInst()) {
288 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
291 BasicBlock *QueryParent = QueryInst->getParent();
294 uint64_t MemSize = 0;
297 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
298 // No dependence found. If this is the entry block of the function, it is a
299 // clobber, otherwise it is non-local.
300 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
301 LocalCache = MemDepResult::getNonLocal();
303 LocalCache = MemDepResult::getClobber(QueryInst);
304 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
305 // If this is a volatile store, don't mess around with it. Just return the
306 // previous instruction as a clobber.
307 if (SI->isVolatile())
308 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
310 MemPtr = SI->getPointerOperand();
311 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
313 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
314 // If this is a volatile load, don't mess around with it. Just return the
315 // previous instruction as a clobber.
316 if (LI->isVolatile())
317 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
319 MemPtr = LI->getPointerOperand();
320 MemSize = AA->getTypeStoreSize(LI->getType());
322 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
323 CallSite QueryCS = CallSite::get(QueryInst);
324 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
325 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
327 } else if (FreeInst *FI = dyn_cast<FreeInst>(QueryInst)) {
328 MemPtr = FI->getPointerOperand();
329 // FreeInsts erase the entire structure, not just a field.
332 // Non-memory instruction.
333 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
336 // If we need to do a pointer scan, make it happen.
338 LocalCache = getPointerDependencyFrom(MemPtr, MemSize,
339 isa<LoadInst>(QueryInst),
340 ScanPos, QueryParent);
342 // Remember the result!
343 if (Instruction *I = LocalCache.getInst())
344 ReverseLocalDeps[I].insert(QueryInst);
350 /// AssertSorted - This method is used when -debug is specified to verify that
351 /// cache arrays are properly kept sorted.
352 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
354 if (Count == -1) Count = Cache.size();
355 if (Count == 0) return;
357 for (unsigned i = 1; i != unsigned(Count); ++i)
358 assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!");
362 /// getNonLocalCallDependency - Perform a full dependency query for the
363 /// specified call, returning the set of blocks that the value is
364 /// potentially live across. The returned set of results will include a
365 /// "NonLocal" result for all blocks where the value is live across.
367 /// This method assumes the instruction returns a "NonLocal" dependency
368 /// within its own block.
370 /// This returns a reference to an internal data structure that may be
371 /// invalidated on the next non-local query or when an instruction is
372 /// removed. Clients must copy this data if they want it around longer than
374 const MemoryDependenceAnalysis::NonLocalDepInfo &
375 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
376 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
377 "getNonLocalCallDependency should only be used on calls with non-local deps!");
378 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
379 NonLocalDepInfo &Cache = CacheP.first;
381 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
382 /// the cached case, this can happen due to instructions being deleted etc. In
383 /// the uncached case, this starts out as the set of predecessors we care
385 SmallVector<BasicBlock*, 32> DirtyBlocks;
387 if (!Cache.empty()) {
388 // Okay, we have a cache entry. If we know it is not dirty, just return it
389 // with no computation.
390 if (!CacheP.second) {
395 // If we already have a partially computed set of results, scan them to
396 // determine what is dirty, seeding our initial DirtyBlocks worklist.
397 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
399 if (I->second.isDirty())
400 DirtyBlocks.push_back(I->first);
402 // Sort the cache so that we can do fast binary search lookups below.
403 std::sort(Cache.begin(), Cache.end());
405 ++NumCacheDirtyNonLocal;
406 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
407 // << Cache.size() << " cached: " << *QueryInst;
409 // Seed DirtyBlocks with each of the preds of QueryInst's block.
410 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
411 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
412 DirtyBlocks.push_back(*PI);
413 NumUncacheNonLocal++;
416 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
417 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
419 SmallPtrSet<BasicBlock*, 64> Visited;
421 unsigned NumSortedEntries = Cache.size();
422 DEBUG(AssertSorted(Cache));
424 // Iterate while we still have blocks to update.
425 while (!DirtyBlocks.empty()) {
426 BasicBlock *DirtyBB = DirtyBlocks.back();
427 DirtyBlocks.pop_back();
429 // Already processed this block?
430 if (!Visited.insert(DirtyBB))
433 // Do a binary search to see if we already have an entry for this block in
434 // the cache set. If so, find it.
435 DEBUG(AssertSorted(Cache, NumSortedEntries));
436 NonLocalDepInfo::iterator Entry =
437 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
438 std::make_pair(DirtyBB, MemDepResult()));
439 if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB)
442 MemDepResult *ExistingResult = 0;
443 if (Entry != Cache.begin()+NumSortedEntries &&
444 Entry->first == DirtyBB) {
445 // If we already have an entry, and if it isn't already dirty, the block
447 if (!Entry->second.isDirty())
450 // Otherwise, remember this slot so we can update the value.
451 ExistingResult = &Entry->second;
454 // If the dirty entry has a pointer, start scanning from it so we don't have
455 // to rescan the entire block.
456 BasicBlock::iterator ScanPos = DirtyBB->end();
457 if (ExistingResult) {
458 if (Instruction *Inst = ExistingResult->getInst()) {
460 // We're removing QueryInst's use of Inst.
461 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
462 QueryCS.getInstruction());
466 // Find out if this block has a local dependency for QueryInst.
469 if (ScanPos != DirtyBB->begin()) {
470 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
471 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
472 // No dependence found. If this is the entry block of the function, it is
473 // a clobber, otherwise it is non-local.
474 Dep = MemDepResult::getNonLocal();
476 Dep = MemDepResult::getClobber(ScanPos);
479 // If we had a dirty entry for the block, update it. Otherwise, just add
482 *ExistingResult = Dep;
484 Cache.push_back(std::make_pair(DirtyBB, Dep));
486 // If the block has a dependency (i.e. it isn't completely transparent to
487 // the value), remember the association!
488 if (!Dep.isNonLocal()) {
489 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
490 // update this when we remove instructions.
491 if (Instruction *Inst = Dep.getInst())
492 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
495 // If the block *is* completely transparent to the load, we need to check
496 // the predecessors of this block. Add them to our worklist.
497 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
498 DirtyBlocks.push_back(*PI);
505 /// getNonLocalPointerDependency - Perform a full dependency query for an
506 /// access to the specified (non-volatile) memory location, returning the
507 /// set of instructions that either define or clobber the value.
509 /// This method assumes the pointer has a "NonLocal" dependency within its
512 void MemoryDependenceAnalysis::
513 getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
514 SmallVectorImpl<NonLocalDepEntry> &Result) {
515 assert(isa<PointerType>(Pointer->getType()) &&
516 "Can't get pointer deps of a non-pointer!");
519 // We know that the pointer value is live into FromBB find the def/clobbers
520 // from presecessors.
521 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
522 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy);
524 // This is the set of blocks we've inspected, and the pointer we consider in
525 // each block. Because of critical edges, we currently bail out if querying
526 // a block with multiple different pointers. This can happen during PHI
528 DenseMap<BasicBlock*, Value*> Visited;
529 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB,
530 Result, Visited, true))
533 Result.push_back(std::make_pair(FromBB,
534 MemDepResult::getClobber(FromBB->begin())));
537 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
538 /// Pointer/PointeeSize using either cached information in Cache or by doing a
539 /// lookup (which may use dirty cache info if available). If we do a lookup,
540 /// add the result to the cache.
541 MemDepResult MemoryDependenceAnalysis::
542 GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
543 bool isLoad, BasicBlock *BB,
544 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
546 // Do a binary search to see if we already have an entry for this block in
547 // the cache set. If so, find it.
548 NonLocalDepInfo::iterator Entry =
549 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
550 std::make_pair(BB, MemDepResult()));
551 if (Entry != Cache->begin() && prior(Entry)->first == BB)
554 MemDepResult *ExistingResult = 0;
555 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
556 ExistingResult = &Entry->second;
558 // If we have a cached entry, and it is non-dirty, use it as the value for
560 if (ExistingResult && !ExistingResult->isDirty()) {
561 ++NumCacheNonLocalPtr;
562 return *ExistingResult;
565 // Otherwise, we have to scan for the value. If we have a dirty cache
566 // entry, start scanning from its position, otherwise we scan from the end
568 BasicBlock::iterator ScanPos = BB->end();
569 if (ExistingResult && ExistingResult->getInst()) {
570 assert(ExistingResult->getInst()->getParent() == BB &&
571 "Instruction invalidated?");
572 ++NumCacheDirtyNonLocalPtr;
573 ScanPos = ExistingResult->getInst();
575 // Eliminating the dirty entry from 'Cache', so update the reverse info.
576 ValueIsLoadPair CacheKey(Pointer, isLoad);
577 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
579 ++NumUncacheNonLocalPtr;
582 // Scan the block for the dependency.
583 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad,
586 // If we had a dirty entry for the block, update it. Otherwise, just add
589 *ExistingResult = Dep;
591 Cache->push_back(std::make_pair(BB, Dep));
593 // If the block has a dependency (i.e. it isn't completely transparent to
594 // the value), remember the reverse association because we just added it
596 if (Dep.isNonLocal())
599 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
600 // update MemDep when we remove instructions.
601 Instruction *Inst = Dep.getInst();
602 assert(Inst && "Didn't depend on anything?");
603 ValueIsLoadPair CacheKey(Pointer, isLoad);
604 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
608 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
609 /// number of elements in the array that are already properly ordered. This is
610 /// optimized for the case when only a few entries are added.
612 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
613 unsigned NumSortedEntries) {
614 switch (Cache.size() - NumSortedEntries) {
616 // done, no new entries.
619 // Two new entries, insert the last one into place.
620 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
622 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
623 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
624 Cache.insert(Entry, Val);
628 // One new entry, Just insert the new value at the appropriate position.
629 if (Cache.size() != 1) {
630 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
632 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
633 std::upper_bound(Cache.begin(), Cache.end(), Val);
634 Cache.insert(Entry, Val);
638 // Added many values, do a full scale sort.
639 std::sort(Cache.begin(), Cache.end());
645 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
646 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
647 /// results to the results vector and keep track of which blocks are visited in
650 /// This has special behavior for the first block queries (when SkipFirstBlock
651 /// is true). In this special case, it ignores the contents of the specified
652 /// block and starts returning dependence info for its predecessors.
654 /// This function returns false on success, or true to indicate that it could
655 /// not compute dependence information for some reason. This should be treated
656 /// as a clobber dependence on the first instruction in the predecessor block.
657 bool MemoryDependenceAnalysis::
658 getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
659 bool isLoad, BasicBlock *StartBB,
660 SmallVectorImpl<NonLocalDepEntry> &Result,
661 DenseMap<BasicBlock*, Value*> &Visited,
662 bool SkipFirstBlock) {
664 // Look up the cached info for Pointer.
665 ValueIsLoadPair CacheKey(Pointer, isLoad);
667 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
668 &NonLocalPointerDeps[CacheKey];
669 NonLocalDepInfo *Cache = &CacheInfo->second;
671 // If we have valid cached information for exactly the block we are
672 // investigating, just return it with no recomputation.
673 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
674 // We have a fully cached result for this query then we can just return the
675 // cached results and populate the visited set. However, we have to verify
676 // that we don't already have conflicting results for these blocks. Check
677 // to ensure that if a block in the results set is in the visited set that
678 // it was for the same pointer query.
679 if (!Visited.empty()) {
680 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
682 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->first);
683 if (VI == Visited.end() || VI->second == Pointer) continue;
685 // We have a pointer mismatch in a block. Just return clobber, saying
686 // that something was clobbered in this result. We could also do a
687 // non-fully cached query, but there is little point in doing this.
692 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
694 Visited.insert(std::make_pair(I->first, Pointer));
695 if (!I->second.isNonLocal())
696 Result.push_back(*I);
698 ++NumCacheCompleteNonLocalPtr;
702 // Otherwise, either this is a new block, a block with an invalid cache
703 // pointer or one that we're about to invalidate by putting more info into it
704 // than its valid cache info. If empty, the result will be valid cache info,
705 // otherwise it isn't.
707 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
709 CacheInfo->first = BBSkipFirstBlockPair();
711 SmallVector<BasicBlock*, 32> Worklist;
712 Worklist.push_back(StartBB);
714 // Keep track of the entries that we know are sorted. Previously cached
715 // entries will all be sorted. The entries we add we only sort on demand (we
716 // don't insert every element into its sorted position). We know that we
717 // won't get any reuse from currently inserted values, because we don't
718 // revisit blocks after we insert info for them.
719 unsigned NumSortedEntries = Cache->size();
720 DEBUG(AssertSorted(*Cache));
722 while (!Worklist.empty()) {
723 BasicBlock *BB = Worklist.pop_back_val();
725 // Skip the first block if we have it.
726 if (!SkipFirstBlock) {
727 // Analyze the dependency of *Pointer in FromBB. See if we already have
729 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
731 // Get the dependency info for Pointer in BB. If we have cached
732 // information, we will use it, otherwise we compute it.
733 DEBUG(AssertSorted(*Cache, NumSortedEntries));
734 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad,
735 BB, Cache, NumSortedEntries);
737 // If we got a Def or Clobber, add this to the list of results.
738 if (!Dep.isNonLocal()) {
739 Result.push_back(NonLocalDepEntry(BB, Dep));
744 // If 'Pointer' is an instruction defined in this block, then we need to do
745 // phi translation to change it into a value live in the predecessor block.
746 // If phi translation fails, then we can't continue dependence analysis.
747 Instruction *PtrInst = dyn_cast<Instruction>(Pointer);
748 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB;
750 // If no PHI translation is needed, just add all the predecessors of this
751 // block to scan them as well.
752 if (!NeedsPHITranslation) {
753 SkipFirstBlock = false;
754 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
755 // Verify that we haven't looked at this block yet.
756 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
757 InsertRes = Visited.insert(std::make_pair(*PI, Pointer));
758 if (InsertRes.second) {
759 // First time we've looked at *PI.
760 Worklist.push_back(*PI);
764 // If we have seen this block before, but it was with a different
765 // pointer then we have a phi translation failure and we have to treat
766 // this as a clobber.
767 if (InsertRes.first->second != Pointer)
768 goto PredTranslationFailure;
773 // If we do need to do phi translation, then there are a bunch of different
774 // cases, because we have to find a Value* live in the predecessor block. We
775 // know that PtrInst is defined in this block at least.
777 // We may have added values to the cache list before this PHI translation.
778 // If so, we haven't done anything to ensure that the cache remains sorted.
779 // Sort it now (if needed) so that recursive invocations of
780 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
781 // value will only see properly sorted cache arrays.
782 if (Cache && NumSortedEntries != Cache->size()) {
783 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
784 NumSortedEntries = Cache->size();
787 // If this is directly a PHI node, just use the incoming values for each
788 // pred as the phi translated version.
789 if (PHINode *PtrPHI = dyn_cast<PHINode>(PtrInst)) {
792 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
793 BasicBlock *Pred = *PI;
794 Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred);
796 // Check to see if we have already visited this pred block with another
797 // pointer. If so, we can't do this lookup. This failure can occur
798 // with PHI translation when a critical edge exists and the PHI node in
799 // the successor translates to a pointer value different than the
800 // pointer the block was first analyzed with.
801 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
802 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr));
804 if (!InsertRes.second) {
805 // If the predecessor was visited with PredPtr, then we already did
806 // the analysis and can ignore it.
807 if (InsertRes.first->second == PredPtr)
810 // Otherwise, the block was previously analyzed with a different
811 // pointer. We can't represent the result of this case, so we just
812 // treat this as a phi translation failure.
813 goto PredTranslationFailure;
816 // FIXME: it is entirely possible that PHI translating will end up with
817 // the same value. Consider PHI translating something like:
818 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
819 // to recurse here, pedantically speaking.
821 // If we have a problem phi translating, fall through to the code below
822 // to handle the failure condition.
823 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred,
825 goto PredTranslationFailure;
828 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
829 CacheInfo = &NonLocalPointerDeps[CacheKey];
830 Cache = &CacheInfo->second;
831 NumSortedEntries = Cache->size();
833 // Since we did phi translation, the "Cache" set won't contain all of the
834 // results for the query. This is ok (we can still use it to accelerate
835 // specific block queries) but we can't do the fastpath "return all
836 // results from the set" Clear out the indicator for this.
837 CacheInfo->first = BBSkipFirstBlockPair();
838 SkipFirstBlock = false;
842 // TODO: BITCAST, GEP.
844 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
845 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
846 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
847 PredTranslationFailure:
850 // Refresh the CacheInfo/Cache pointer if it got invalidated.
851 CacheInfo = &NonLocalPointerDeps[CacheKey];
852 Cache = &CacheInfo->second;
853 NumSortedEntries = Cache->size();
856 // Since we did phi translation, the "Cache" set won't contain all of the
857 // results for the query. This is ok (we can still use it to accelerate
858 // specific block queries) but we can't do the fastpath "return all
859 // results from the set" Clear out the indicator for this.
860 CacheInfo->first = BBSkipFirstBlockPair();
862 // If *nothing* works, mark the pointer as being clobbered by the first
863 // instruction in this block.
865 // If this is the magic first block, return this as a clobber of the whole
866 // incoming value. Since we can't phi translate to one of the predecessors,
867 // we have to bail out.
871 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
872 assert(I != Cache->rend() && "Didn't find current block??");
876 assert(I->second.isNonLocal() &&
877 "Should only be here with transparent block");
878 I->second = MemDepResult::getClobber(BB->begin());
879 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
880 Result.push_back(*I);
885 // Okay, we're done now. If we added new values to the cache, re-sort it.
886 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
887 DEBUG(AssertSorted(*Cache));
891 /// RemoveCachedNonLocalPointerDependencies - If P exists in
892 /// CachedNonLocalPointerInfo, remove it.
893 void MemoryDependenceAnalysis::
894 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
895 CachedNonLocalPointerInfo::iterator It =
896 NonLocalPointerDeps.find(P);
897 if (It == NonLocalPointerDeps.end()) return;
899 // Remove all of the entries in the BB->val map. This involves removing
900 // instructions from the reverse map.
901 NonLocalDepInfo &PInfo = It->second.second;
903 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
904 Instruction *Target = PInfo[i].second.getInst();
905 if (Target == 0) continue; // Ignore non-local dep results.
906 assert(Target->getParent() == PInfo[i].first);
908 // Eliminating the dirty entry from 'Cache', so update the reverse info.
909 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
912 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
913 NonLocalPointerDeps.erase(It);
917 /// invalidateCachedPointerInfo - This method is used to invalidate cached
918 /// information about the specified pointer, because it may be too
919 /// conservative in memdep. This is an optional call that can be used when
920 /// the client detects an equivalence between the pointer and some other
921 /// value and replaces the other value with ptr. This can make Ptr available
922 /// in more places that cached info does not necessarily keep.
923 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
924 // If Ptr isn't really a pointer, just ignore it.
925 if (!isa<PointerType>(Ptr->getType())) return;
926 // Flush store info for the pointer.
927 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
928 // Flush load info for the pointer.
929 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
932 /// removeInstruction - Remove an instruction from the dependence analysis,
933 /// updating the dependence of instructions that previously depended on it.
934 /// This method attempts to keep the cache coherent using the reverse map.
935 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
936 // Walk through the Non-local dependencies, removing this one as the value
937 // for any cached queries.
938 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
939 if (NLDI != NonLocalDeps.end()) {
940 NonLocalDepInfo &BlockMap = NLDI->second.first;
941 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
943 if (Instruction *Inst = DI->second.getInst())
944 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
945 NonLocalDeps.erase(NLDI);
948 // If we have a cached local dependence query for this instruction, remove it.
950 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
951 if (LocalDepEntry != LocalDeps.end()) {
952 // Remove us from DepInst's reverse set now that the local dep info is gone.
953 if (Instruction *Inst = LocalDepEntry->second.getInst())
954 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
956 // Remove this local dependency info.
957 LocalDeps.erase(LocalDepEntry);
960 // If we have any cached pointer dependencies on this instruction, remove
961 // them. If the instruction has non-pointer type, then it can't be a pointer
964 // Remove it from both the load info and the store info. The instruction
965 // can't be in either of these maps if it is non-pointer.
966 if (isa<PointerType>(RemInst->getType())) {
967 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
968 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
971 // Loop over all of the things that depend on the instruction we're removing.
973 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
975 // If we find RemInst as a clobber or Def in any of the maps for other values,
976 // we need to replace its entry with a dirty version of the instruction after
977 // it. If RemInst is a terminator, we use a null dirty value.
979 // Using a dirty version of the instruction after RemInst saves having to scan
980 // the entire block to get to this point.
981 MemDepResult NewDirtyVal;
982 if (!RemInst->isTerminator())
983 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
985 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
986 if (ReverseDepIt != ReverseLocalDeps.end()) {
987 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
988 // RemInst can't be the terminator if it has local stuff depending on it.
989 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
990 "Nothing can locally depend on a terminator");
992 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
993 E = ReverseDeps.end(); I != E; ++I) {
994 Instruction *InstDependingOnRemInst = *I;
995 assert(InstDependingOnRemInst != RemInst &&
996 "Already removed our local dep info");
998 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1000 // Make sure to remember that new things depend on NewDepInst.
1001 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1002 "a local dep on this if it is a terminator!");
1003 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1004 InstDependingOnRemInst));
1007 ReverseLocalDeps.erase(ReverseDepIt);
1009 // Add new reverse deps after scanning the set, to avoid invalidating the
1010 // 'ReverseDeps' reference.
1011 while (!ReverseDepsToAdd.empty()) {
1012 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1013 .insert(ReverseDepsToAdd.back().second);
1014 ReverseDepsToAdd.pop_back();
1018 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1019 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1020 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1021 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1023 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1025 PerInstNLInfo &INLD = NonLocalDeps[*I];
1026 // The information is now dirty!
1029 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1030 DE = INLD.first.end(); DI != DE; ++DI) {
1031 if (DI->second.getInst() != RemInst) continue;
1033 // Convert to a dirty entry for the subsequent instruction.
1034 DI->second = NewDirtyVal;
1036 if (Instruction *NextI = NewDirtyVal.getInst())
1037 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1041 ReverseNonLocalDeps.erase(ReverseDepIt);
1043 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1044 while (!ReverseDepsToAdd.empty()) {
1045 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1046 .insert(ReverseDepsToAdd.back().second);
1047 ReverseDepsToAdd.pop_back();
1051 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1052 // value in the NonLocalPointerDeps info.
1053 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1054 ReverseNonLocalPtrDeps.find(RemInst);
1055 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1056 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1057 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1059 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1060 E = Set.end(); I != E; ++I) {
1061 ValueIsLoadPair P = *I;
1062 assert(P.getPointer() != RemInst &&
1063 "Already removed NonLocalPointerDeps info for RemInst");
1065 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second;
1067 // The cache is not valid for any specific block anymore.
1068 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair();
1070 // Update any entries for RemInst to use the instruction after it.
1071 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1073 if (DI->second.getInst() != RemInst) continue;
1075 // Convert to a dirty entry for the subsequent instruction.
1076 DI->second = NewDirtyVal;
1078 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1079 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1082 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1083 // subsequent value may invalidate the sortedness.
1084 std::sort(NLPDI.begin(), NLPDI.end());
1087 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1089 while (!ReversePtrDepsToAdd.empty()) {
1090 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1091 .insert(ReversePtrDepsToAdd.back().second);
1092 ReversePtrDepsToAdd.pop_back();
1097 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1098 AA->deleteValue(RemInst);
1099 DEBUG(verifyRemoved(RemInst));
1101 /// verifyRemoved - Verify that the specified instruction does not occur
1102 /// in our internal data structures.
1103 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1104 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1105 E = LocalDeps.end(); I != E; ++I) {
1106 assert(I->first != D && "Inst occurs in data structures");
1107 assert(I->second.getInst() != D &&
1108 "Inst occurs in data structures");
1111 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1112 E = NonLocalPointerDeps.end(); I != E; ++I) {
1113 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1114 const NonLocalDepInfo &Val = I->second.second;
1115 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1117 assert(II->second.getInst() != D && "Inst occurs as NLPD value");
1120 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1121 E = NonLocalDeps.end(); I != E; ++I) {
1122 assert(I->first != D && "Inst occurs in data structures");
1123 const PerInstNLInfo &INLD = I->second;
1124 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1125 EE = INLD.first.end(); II != EE; ++II)
1126 assert(II->second.getInst() != D && "Inst occurs in data structures");
1129 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1130 E = ReverseLocalDeps.end(); I != E; ++I) {
1131 assert(I->first != D && "Inst occurs in data structures");
1132 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1133 EE = I->second.end(); II != EE; ++II)
1134 assert(*II != D && "Inst occurs in data structures");
1137 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1138 E = ReverseNonLocalDeps.end();
1140 assert(I->first != D && "Inst occurs in data structures");
1141 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1142 EE = I->second.end(); II != EE; ++II)
1143 assert(*II != D && "Inst occurs in data structures");
1146 for (ReverseNonLocalPtrDepTy::const_iterator
1147 I = ReverseNonLocalPtrDeps.begin(),
1148 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1149 assert(I->first != D && "Inst occurs in rev NLPD map");
1151 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1152 E = I->second.end(); II != E; ++II)
1153 assert(*II != ValueIsLoadPair(D, false) &&
1154 *II != ValueIsLoadPair(D, true) &&
1155 "Inst occurs in ReverseNonLocalPtrDeps map");