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/PredIteratorCache.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Target/TargetData.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>();
73 AU.addRequiredTransitive<TargetData>();
76 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
77 AA = &getAnalysis<AliasAnalysis>();
78 TD = &getAnalysis<TargetData>();
80 PredCache.reset(new PredIteratorCache());
84 /// RemoveFromReverseMap - This is a helper function that removes Val from
85 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
86 template <typename KeyTy>
87 static void RemoveFromReverseMap(DenseMap<Instruction*,
88 SmallPtrSet<KeyTy*, 4> > &ReverseMap,
89 Instruction *Inst, KeyTy *Val) {
90 typename DenseMap<Instruction*, SmallPtrSet<KeyTy*, 4> >::iterator
91 InstIt = ReverseMap.find(Inst);
92 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
93 bool Found = InstIt->second.erase(Val);
94 assert(Found && "Invalid reverse map!"); Found=Found;
95 if (InstIt->second.empty())
96 ReverseMap.erase(InstIt);
100 /// getCallSiteDependencyFrom - Private helper for finding the local
101 /// dependencies of a call site.
102 MemDepResult MemoryDependenceAnalysis::
103 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
104 BasicBlock::iterator ScanIt, BasicBlock *BB) {
105 // Walk backwards through the block, looking for dependencies
106 while (ScanIt != BB->begin()) {
107 Instruction *Inst = --ScanIt;
109 // If this inst is a memory op, get the pointer it accessed
111 uint64_t PointerSize = 0;
112 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
113 Pointer = S->getPointerOperand();
114 PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType());
115 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
116 Pointer = V->getOperand(0);
117 PointerSize = TD->getTypeStoreSize(V->getType());
118 } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) {
119 Pointer = F->getPointerOperand();
121 // FreeInsts erase the entire structure
123 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
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 // Values depend on loads if the pointers are must aliased. This means that
179 // a load depends on another must aliased load from the same value.
180 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
181 Value *Pointer = LI->getPointerOperand();
182 uint64_t PointerSize = TD->getTypeStoreSize(LI->getType());
184 // If we found a pointer, check if it could be the same as our pointer.
185 AliasAnalysis::AliasResult R =
186 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
187 if (R == AliasAnalysis::NoAlias)
190 // May-alias loads don't depend on each other without a dependence.
191 if (isLoad && R == AliasAnalysis::MayAlias)
193 // Stores depend on may and must aliased loads, loads depend on must-alias
195 return MemDepResult::getDef(Inst);
198 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
199 Value *Pointer = SI->getPointerOperand();
200 uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
202 // If we found a pointer, check if it could be the same as our pointer.
203 AliasAnalysis::AliasResult R =
204 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
206 if (R == AliasAnalysis::NoAlias)
208 if (R == AliasAnalysis::MayAlias)
209 return MemDepResult::getClobber(Inst);
210 return MemDepResult::getDef(Inst);
213 // If this is an allocation, and if we know that the accessed pointer is to
214 // the allocation, return Def. This means that there is no dependence and
215 // the access can be optimized based on that. For example, a load could
217 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
218 Value *AccessPtr = MemPtr->getUnderlyingObject();
220 if (AccessPtr == AI ||
221 AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
222 return MemDepResult::getDef(AI);
226 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
227 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
228 case AliasAnalysis::NoModRef:
229 // If the call has no effect on the queried pointer, just ignore it.
231 case AliasAnalysis::Ref:
232 // If the call is known to never store to the pointer, and if this is a
233 // load query, we can safely ignore it (scan past it).
238 // Otherwise, there is a potential dependence. Return a clobber.
239 return MemDepResult::getClobber(Inst);
243 // No dependence found. If this is the entry block of the function, it is a
244 // clobber, otherwise it is non-local.
245 if (BB != &BB->getParent()->getEntryBlock())
246 return MemDepResult::getNonLocal();
247 return MemDepResult::getClobber(ScanIt);
250 /// getDependency - Return the instruction on which a memory operation
252 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
253 Instruction *ScanPos = QueryInst;
255 // Check for a cached result
256 MemDepResult &LocalCache = LocalDeps[QueryInst];
258 // If the cached entry is non-dirty, just return it. Note that this depends
259 // on MemDepResult's default constructing to 'dirty'.
260 if (!LocalCache.isDirty())
263 // Otherwise, if we have a dirty entry, we know we can start the scan at that
264 // instruction, which may save us some work.
265 if (Instruction *Inst = LocalCache.getInst()) {
268 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
271 BasicBlock *QueryParent = QueryInst->getParent();
274 uint64_t MemSize = 0;
277 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
278 // No dependence found. If this is the entry block of the function, it is a
279 // clobber, otherwise it is non-local.
280 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
281 LocalCache = MemDepResult::getNonLocal();
283 LocalCache = MemDepResult::getClobber(QueryInst);
284 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
285 // If this is a volatile store, don't mess around with it. Just return the
286 // previous instruction as a clobber.
287 if (SI->isVolatile())
288 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
290 MemPtr = SI->getPointerOperand();
291 MemSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
293 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
294 // If this is a volatile load, don't mess around with it. Just return the
295 // previous instruction as a clobber.
296 if (LI->isVolatile())
297 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
299 MemPtr = LI->getPointerOperand();
300 MemSize = TD->getTypeStoreSize(LI->getType());
302 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
303 CallSite QueryCS = CallSite::get(QueryInst);
304 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
305 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
307 } else if (FreeInst *FI = dyn_cast<FreeInst>(QueryInst)) {
308 MemPtr = FI->getPointerOperand();
309 // FreeInsts erase the entire structure, not just a field.
312 // Non-memory instruction.
313 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
316 // If we need to do a pointer scan, make it happen.
318 LocalCache = getPointerDependencyFrom(MemPtr, MemSize,
319 isa<LoadInst>(QueryInst),
320 ScanPos, QueryParent);
322 // Remember the result!
323 if (Instruction *I = LocalCache.getInst())
324 ReverseLocalDeps[I].insert(QueryInst);
329 /// getNonLocalCallDependency - Perform a full dependency query for the
330 /// specified call, returning the set of blocks that the value is
331 /// potentially live across. The returned set of results will include a
332 /// "NonLocal" result for all blocks where the value is live across.
334 /// This method assumes the instruction returns a "NonLocal" dependency
335 /// within its own block.
337 /// This returns a reference to an internal data structure that may be
338 /// invalidated on the next non-local query or when an instruction is
339 /// removed. Clients must copy this data if they want it around longer than
341 const MemoryDependenceAnalysis::NonLocalDepInfo &
342 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
343 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
344 "getNonLocalCallDependency should only be used on calls with non-local deps!");
345 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
346 NonLocalDepInfo &Cache = CacheP.first;
348 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
349 /// the cached case, this can happen due to instructions being deleted etc. In
350 /// the uncached case, this starts out as the set of predecessors we care
352 SmallVector<BasicBlock*, 32> DirtyBlocks;
354 if (!Cache.empty()) {
355 // Okay, we have a cache entry. If we know it is not dirty, just return it
356 // with no computation.
357 if (!CacheP.second) {
362 // If we already have a partially computed set of results, scan them to
363 // determine what is dirty, seeding our initial DirtyBlocks worklist.
364 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
366 if (I->second.isDirty())
367 DirtyBlocks.push_back(I->first);
369 // Sort the cache so that we can do fast binary search lookups below.
370 std::sort(Cache.begin(), Cache.end());
372 ++NumCacheDirtyNonLocal;
373 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
374 // << Cache.size() << " cached: " << *QueryInst;
376 // Seed DirtyBlocks with each of the preds of QueryInst's block.
377 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
378 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
379 DirtyBlocks.push_back(*PI);
380 NumUncacheNonLocal++;
383 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
384 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
386 SmallPtrSet<BasicBlock*, 64> Visited;
388 unsigned NumSortedEntries = Cache.size();
390 // Iterate while we still have blocks to update.
391 while (!DirtyBlocks.empty()) {
392 BasicBlock *DirtyBB = DirtyBlocks.back();
393 DirtyBlocks.pop_back();
395 // Already processed this block?
396 if (!Visited.insert(DirtyBB))
399 // Do a binary search to see if we already have an entry for this block in
400 // the cache set. If so, find it.
401 NonLocalDepInfo::iterator Entry =
402 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
403 std::make_pair(DirtyBB, MemDepResult()));
404 if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB)
407 MemDepResult *ExistingResult = 0;
408 if (Entry != Cache.begin()+NumSortedEntries &&
409 Entry->first == DirtyBB) {
410 // If we already have an entry, and if it isn't already dirty, the block
412 if (!Entry->second.isDirty())
415 // Otherwise, remember this slot so we can update the value.
416 ExistingResult = &Entry->second;
419 // If the dirty entry has a pointer, start scanning from it so we don't have
420 // to rescan the entire block.
421 BasicBlock::iterator ScanPos = DirtyBB->end();
422 if (ExistingResult) {
423 if (Instruction *Inst = ExistingResult->getInst()) {
425 // We're removing QueryInst's use of Inst.
426 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
427 QueryCS.getInstruction());
431 // Find out if this block has a local dependency for QueryInst.
434 if (ScanPos != DirtyBB->begin()) {
435 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
436 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
437 // No dependence found. If this is the entry block of the function, it is
438 // a clobber, otherwise it is non-local.
439 Dep = MemDepResult::getNonLocal();
441 Dep = MemDepResult::getClobber(ScanPos);
444 // If we had a dirty entry for the block, update it. Otherwise, just add
447 *ExistingResult = Dep;
449 Cache.push_back(std::make_pair(DirtyBB, Dep));
451 // If the block has a dependency (i.e. it isn't completely transparent to
452 // the value), remember the association!
453 if (!Dep.isNonLocal()) {
454 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
455 // update this when we remove instructions.
456 if (Instruction *Inst = Dep.getInst())
457 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
460 // If the block *is* completely transparent to the load, we need to check
461 // the predecessors of this block. Add them to our worklist.
462 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
463 DirtyBlocks.push_back(*PI);
470 /// getNonLocalPointerDependency - Perform a full dependency query for an
471 /// access to the specified (non-volatile) memory location, returning the
472 /// set of instructions that either define or clobber the value.
474 /// This method assumes the pointer has a "NonLocal" dependency within its
477 void MemoryDependenceAnalysis::
478 getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
479 SmallVectorImpl<NonLocalDepEntry> &Result) {
480 assert(isa<PointerType>(Pointer->getType()) &&
481 "Can't get pointer deps of a non-pointer!");
484 // We know that the pointer value is live into FromBB find the def/clobbers
485 // from presecessors.
486 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
487 uint64_t PointeeSize = TD->getTypeStoreSize(EltTy);
489 // This is the set of blocks we've inspected, and the pointer we consider in
490 // each block. Because of critical edges, we currently bail out if querying
491 // a block with multiple different pointers. This can happen during PHI
493 DenseMap<BasicBlock*, Value*> Visited;
494 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB,
495 Result, Visited, true))
498 Result.push_back(std::make_pair(FromBB,
499 MemDepResult::getClobber(FromBB->begin())));
502 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
503 /// Pointer/PointeeSize using either cached information in Cache or by doing a
504 /// lookup (which may use dirty cache info if available). If we do a lookup,
505 /// add the result to the cache.
506 MemDepResult MemoryDependenceAnalysis::
507 GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
508 bool isLoad, BasicBlock *BB,
509 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
511 // Do a binary search to see if we already have an entry for this block in
512 // the cache set. If so, find it.
513 NonLocalDepInfo::iterator Entry =
514 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
515 std::make_pair(BB, MemDepResult()));
516 if (Entry != Cache->begin() && prior(Entry)->first == BB)
519 MemDepResult *ExistingResult = 0;
520 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
521 ExistingResult = &Entry->second;
523 // If we have a cached entry, and it is non-dirty, use it as the value for
525 if (ExistingResult && !ExistingResult->isDirty()) {
526 ++NumCacheNonLocalPtr;
527 return *ExistingResult;
530 // Otherwise, we have to scan for the value. If we have a dirty cache
531 // entry, start scanning from its position, otherwise we scan from the end
533 BasicBlock::iterator ScanPos = BB->end();
534 if (ExistingResult && ExistingResult->getInst()) {
535 assert(ExistingResult->getInst()->getParent() == BB &&
536 "Instruction invalidated?");
537 ++NumCacheDirtyNonLocalPtr;
538 ScanPos = ExistingResult->getInst();
540 // Eliminating the dirty entry from 'Cache', so update the reverse info.
541 ValueIsLoadPair CacheKey(Pointer, isLoad);
542 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos,
543 CacheKey.getOpaqueValue());
545 ++NumUncacheNonLocalPtr;
548 // Scan the block for the dependency.
549 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad,
552 // If we had a dirty entry for the block, update it. Otherwise, just add
555 *ExistingResult = Dep;
557 Cache->push_back(std::make_pair(BB, Dep));
559 // If the block has a dependency (i.e. it isn't completely transparent to
560 // the value), remember the reverse association because we just added it
562 if (Dep.isNonLocal())
565 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
566 // update MemDep when we remove instructions.
567 Instruction *Inst = Dep.getInst();
568 assert(Inst && "Didn't depend on anything?");
569 ValueIsLoadPair CacheKey(Pointer, isLoad);
570 ReverseNonLocalPtrDeps[Inst].insert(CacheKey.getOpaqueValue());
575 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
576 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
577 /// results to the results vector and keep track of which blocks are visited in
580 /// This has special behavior for the first block queries (when SkipFirstBlock
581 /// is true). In this special case, it ignores the contents of the specified
582 /// block and starts returning dependence info for its predecessors.
584 /// This function returns false on success, or true to indicate that it could
585 /// not compute dependence information for some reason. This should be treated
586 /// as a clobber dependence on the first instruction in the predecessor block.
587 bool MemoryDependenceAnalysis::
588 getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
589 bool isLoad, BasicBlock *StartBB,
590 SmallVectorImpl<NonLocalDepEntry> &Result,
591 DenseMap<BasicBlock*, Value*> &Visited,
592 bool SkipFirstBlock) {
594 // Look up the cached info for Pointer.
595 ValueIsLoadPair CacheKey(Pointer, isLoad);
597 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
598 &NonLocalPointerDeps[CacheKey];
599 NonLocalDepInfo *Cache = &CacheInfo->second;
601 // If we have valid cached information for exactly the block we are
602 // investigating, just return it with no recomputation.
603 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
604 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
606 if (!I->second.isNonLocal())
607 Result.push_back(*I);
608 ++NumCacheCompleteNonLocalPtr;
612 // Otherwise, either this is a new block, a block with an invalid cache
613 // pointer or one that we're about to invalidate by putting more info into it
614 // than its valid cache info. If empty, the result will be valid cache info,
615 // otherwise it isn't.
617 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
619 CacheInfo->first = BBSkipFirstBlockPair();
621 SmallVector<BasicBlock*, 32> Worklist;
622 Worklist.push_back(StartBB);
624 // Keep track of the entries that we know are sorted. Previously cached
625 // entries will all be sorted. The entries we add we only sort on demand (we
626 // don't insert every element into its sorted position). We know that we
627 // won't get any reuse from currently inserted values, because we don't
628 // revisit blocks after we insert info for them.
629 unsigned NumSortedEntries = Cache->size();
631 while (!Worklist.empty()) {
632 BasicBlock *BB = Worklist.pop_back_val();
634 // Skip the first block if we have it.
635 if (!SkipFirstBlock) {
636 // Analyze the dependency of *Pointer in FromBB. See if we already have
638 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
640 // Get the dependency info for Pointer in BB. If we have cached
641 // information, we will use it, otherwise we compute it.
642 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad,
643 BB, Cache, NumSortedEntries);
645 // If we got a Def or Clobber, add this to the list of results.
646 if (!Dep.isNonLocal()) {
647 Result.push_back(NonLocalDepEntry(BB, Dep));
652 // If 'Pointer' is an instruction defined in this block, then we need to do
653 // phi translation to change it into a value live in the predecessor block.
654 // If phi translation fails, then we can't continue dependence analysis.
655 Instruction *PtrInst = dyn_cast<Instruction>(Pointer);
656 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB;
658 // If no PHI translation is needed, just add all the predecessors of this
659 // block to scan them as well.
660 if (!NeedsPHITranslation) {
661 SkipFirstBlock = false;
662 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
663 // Verify that we haven't looked at this block yet.
664 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
665 InsertRes = Visited.insert(std::make_pair(*PI, Pointer));
666 if (InsertRes.second) {
667 // First time we've looked at *PI.
668 Worklist.push_back(*PI);
672 // If we have seen this block before, but it was with a different
673 // pointer then we have a phi translation failure and we have to treat
674 // this as a clobber.
675 if (InsertRes.first->second != Pointer)
676 goto PredTranslationFailure;
681 // If we do need to do phi translation, then there are a bunch of different
682 // cases, because we have to find a Value* live in the predecessor block. We
683 // know that PtrInst is defined in this block at least.
685 // If this is directly a PHI node, just use the incoming values for each
686 // pred as the phi translated version.
687 if (PHINode *PtrPHI = dyn_cast<PHINode>(PtrInst)) {
688 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI){
689 BasicBlock *Pred = *PI;
690 Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred);
692 // Check to see if we have already visited this pred block with another
693 // pointer. If so, we can't do this lookup. This failure can occur
694 // with PHI translation when a critical edge exists and the PHI node in
695 // the successor translates to a pointer value different than the
696 // pointer the block was first analyzed with.
697 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
698 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr));
700 if (!InsertRes.second) {
701 // If the predecessor was visited with PredPtr, then we already did
702 // the analysis and can ignore it.
703 if (InsertRes.first->second == PredPtr)
706 // Otherwise, the block was previously analyzed with a different
707 // pointer. We can't represent the result of this case, so we just
708 // treat this as a phi translation failure.
709 goto PredTranslationFailure;
712 // If we have a problem phi translating, fall through to the code below
713 // to handle the failure condition.
714 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred,
716 goto PredTranslationFailure;
719 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
720 CacheInfo = &NonLocalPointerDeps[CacheKey];
721 Cache = &CacheInfo->second;
723 // Since we did phi translation, the "Cache" set won't contain all of the
724 // results for the query. This is ok (we can still use it to accelerate
725 // specific block queries) but we can't do the fastpath "return all
726 // results from the set" Clear out the indicator for this.
727 CacheInfo->first = BBSkipFirstBlockPair();
728 SkipFirstBlock = false;
732 // TODO: BITCAST, GEP.
734 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
735 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
736 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
737 PredTranslationFailure:
739 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
740 CacheInfo = &NonLocalPointerDeps[CacheKey];
741 Cache = &CacheInfo->second;
743 // Since we did phi translation, the "Cache" set won't contain all of the
744 // results for the query. This is ok (we can still use it to accelerate
745 // specific block queries) but we can't do the fastpath "return all
746 // results from the set" Clear out the indicator for this.
747 CacheInfo->first = BBSkipFirstBlockPair();
749 // If *nothing* works, mark the pointer as being clobbered by the first
750 // instruction in this block.
752 // If this is the magic first block, return this as a clobber of the whole
753 // incoming value. Since we can't phi translate to one of the predecessors,
754 // we have to bail out.
758 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
759 assert(I != Cache->rend() && "Didn't find current block??");
763 assert(I->second.isNonLocal() &&
764 "Should only be here with transparent block");
765 I->second = MemDepResult::getClobber(BB->begin());
766 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey.getOpaqueValue());
767 Result.push_back(*I);
772 // Okay, we're done now. If we added new values to the cache, re-sort it.
773 switch (Cache->size()-NumSortedEntries) {
775 // done, no new entries.
778 // Two new entries, insert the last one into place.
779 NonLocalDepEntry Val = Cache->back();
781 NonLocalDepInfo::iterator Entry =
782 std::upper_bound(Cache->begin(), Cache->end()-1, Val);
783 Cache->insert(Entry, Val);
787 // One new entry, Just insert the new value at the appropriate position.
788 if (Cache->size() != 1) {
789 NonLocalDepEntry Val = Cache->back();
791 NonLocalDepInfo::iterator Entry =
792 std::upper_bound(Cache->begin(), Cache->end(), Val);
793 Cache->insert(Entry, Val);
797 // Added many values, do a full scale sort.
798 std::sort(Cache->begin(), Cache->end());
804 /// RemoveCachedNonLocalPointerDependencies - If P exists in
805 /// CachedNonLocalPointerInfo, remove it.
806 void MemoryDependenceAnalysis::
807 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
808 CachedNonLocalPointerInfo::iterator It =
809 NonLocalPointerDeps.find(P);
810 if (It == NonLocalPointerDeps.end()) return;
812 // Remove all of the entries in the BB->val map. This involves removing
813 // instructions from the reverse map.
814 NonLocalDepInfo &PInfo = It->second.second;
816 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
817 Instruction *Target = PInfo[i].second.getInst();
818 if (Target == 0) continue; // Ignore non-local dep results.
819 assert(Target->getParent() == PInfo[i].first);
821 // Eliminating the dirty entry from 'Cache', so update the reverse info.
822 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P.getOpaqueValue());
825 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
826 NonLocalPointerDeps.erase(It);
830 /// invalidateCachedPointerInfo - This method is used to invalidate cached
831 /// information about the specified pointer, because it may be too
832 /// conservative in memdep. This is an optional call that can be used when
833 /// the client detects an equivalence between the pointer and some other
834 /// value and replaces the other value with ptr. This can make Ptr available
835 /// in more places that cached info does not necessarily keep.
836 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
837 // If Ptr isn't really a pointer, just ignore it.
838 if (!isa<PointerType>(Ptr->getType())) return;
839 // Flush store info for the pointer.
840 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
841 // Flush load info for the pointer.
842 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
845 /// removeInstruction - Remove an instruction from the dependence analysis,
846 /// updating the dependence of instructions that previously depended on it.
847 /// This method attempts to keep the cache coherent using the reverse map.
848 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
849 // Walk through the Non-local dependencies, removing this one as the value
850 // for any cached queries.
851 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
852 if (NLDI != NonLocalDeps.end()) {
853 NonLocalDepInfo &BlockMap = NLDI->second.first;
854 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
856 if (Instruction *Inst = DI->second.getInst())
857 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
858 NonLocalDeps.erase(NLDI);
861 // If we have a cached local dependence query for this instruction, remove it.
863 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
864 if (LocalDepEntry != LocalDeps.end()) {
865 // Remove us from DepInst's reverse set now that the local dep info is gone.
866 if (Instruction *Inst = LocalDepEntry->second.getInst())
867 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
869 // Remove this local dependency info.
870 LocalDeps.erase(LocalDepEntry);
873 // If we have any cached pointer dependencies on this instruction, remove
874 // them. If the instruction has non-pointer type, then it can't be a pointer
877 // Remove it from both the load info and the store info. The instruction
878 // can't be in either of these maps if it is non-pointer.
879 if (isa<PointerType>(RemInst->getType())) {
880 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
881 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
884 // Loop over all of the things that depend on the instruction we're removing.
886 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
888 // If we find RemInst as a clobber or Def in any of the maps for other values,
889 // we need to replace its entry with a dirty version of the instruction after
890 // it. If RemInst is a terminator, we use a null dirty value.
892 // Using a dirty version of the instruction after RemInst saves having to scan
893 // the entire block to get to this point.
894 MemDepResult NewDirtyVal;
895 if (!RemInst->isTerminator())
896 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
898 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
899 if (ReverseDepIt != ReverseLocalDeps.end()) {
900 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
901 // RemInst can't be the terminator if it has local stuff depending on it.
902 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
903 "Nothing can locally depend on a terminator");
905 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
906 E = ReverseDeps.end(); I != E; ++I) {
907 Instruction *InstDependingOnRemInst = *I;
908 assert(InstDependingOnRemInst != RemInst &&
909 "Already removed our local dep info");
911 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
913 // Make sure to remember that new things depend on NewDepInst.
914 assert(NewDirtyVal.getInst() && "There is no way something else can have "
915 "a local dep on this if it is a terminator!");
916 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
917 InstDependingOnRemInst));
920 ReverseLocalDeps.erase(ReverseDepIt);
922 // Add new reverse deps after scanning the set, to avoid invalidating the
923 // 'ReverseDeps' reference.
924 while (!ReverseDepsToAdd.empty()) {
925 ReverseLocalDeps[ReverseDepsToAdd.back().first]
926 .insert(ReverseDepsToAdd.back().second);
927 ReverseDepsToAdd.pop_back();
931 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
932 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
933 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
934 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
936 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
938 PerInstNLInfo &INLD = NonLocalDeps[*I];
939 // The information is now dirty!
942 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
943 DE = INLD.first.end(); DI != DE; ++DI) {
944 if (DI->second.getInst() != RemInst) continue;
946 // Convert to a dirty entry for the subsequent instruction.
947 DI->second = NewDirtyVal;
949 if (Instruction *NextI = NewDirtyVal.getInst())
950 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
954 ReverseNonLocalDeps.erase(ReverseDepIt);
956 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
957 while (!ReverseDepsToAdd.empty()) {
958 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
959 .insert(ReverseDepsToAdd.back().second);
960 ReverseDepsToAdd.pop_back();
964 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
965 // value in the NonLocalPointerDeps info.
966 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
967 ReverseNonLocalPtrDeps.find(RemInst);
968 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
969 SmallPtrSet<void*, 4> &Set = ReversePtrDepIt->second;
970 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
972 for (SmallPtrSet<void*, 4>::iterator I = Set.begin(), E = Set.end();
975 P.setFromOpaqueValue(*I);
976 assert(P.getPointer() != RemInst &&
977 "Already removed NonLocalPointerDeps info for RemInst");
979 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second;
981 // The cache is not valid for any specific block anymore.
982 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair();
984 // Update any entries for RemInst to use the instruction after it.
985 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
987 if (DI->second.getInst() != RemInst) continue;
989 // Convert to a dirty entry for the subsequent instruction.
990 DI->second = NewDirtyVal;
992 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
993 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
997 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
999 while (!ReversePtrDepsToAdd.empty()) {
1000 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1001 .insert(ReversePtrDepsToAdd.back().second.getOpaqueValue());
1002 ReversePtrDepsToAdd.pop_back();
1007 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1008 AA->deleteValue(RemInst);
1009 DEBUG(verifyRemoved(RemInst));
1011 /// verifyRemoved - Verify that the specified instruction does not occur
1012 /// in our internal data structures.
1013 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1014 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1015 E = LocalDeps.end(); I != E; ++I) {
1016 assert(I->first != D && "Inst occurs in data structures");
1017 assert(I->second.getInst() != D &&
1018 "Inst occurs in data structures");
1021 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1022 E = NonLocalPointerDeps.end(); I != E; ++I) {
1023 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1024 const NonLocalDepInfo &Val = I->second.second;
1025 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1027 assert(II->second.getInst() != D && "Inst occurs as NLPD value");
1030 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1031 E = NonLocalDeps.end(); I != E; ++I) {
1032 assert(I->first != D && "Inst occurs in data structures");
1033 const PerInstNLInfo &INLD = I->second;
1034 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1035 EE = INLD.first.end(); II != EE; ++II)
1036 assert(II->second.getInst() != D && "Inst occurs in data structures");
1039 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1040 E = ReverseLocalDeps.end(); I != E; ++I) {
1041 assert(I->first != D && "Inst occurs in data structures");
1042 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1043 EE = I->second.end(); II != EE; ++II)
1044 assert(*II != D && "Inst occurs in data structures");
1047 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1048 E = ReverseNonLocalDeps.end();
1050 assert(I->first != D && "Inst occurs in data structures");
1051 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1052 EE = I->second.end(); II != EE; ++II)
1053 assert(*II != D && "Inst occurs in data structures");
1056 for (ReverseNonLocalPtrDepTy::const_iterator
1057 I = ReverseNonLocalPtrDeps.begin(),
1058 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1059 assert(I->first != D && "Inst occurs in rev NLPD map");
1061 for (SmallPtrSet<void*, 4>::const_iterator II = I->second.begin(),
1062 E = I->second.end(); II != E; ++II)
1063 assert(*II != ValueIsLoadPair(D, false).getOpaqueValue() &&
1064 *II != ValueIsLoadPair(D, true).getOpaqueValue() &&
1065 "Inst occurs in ReverseNonLocalPtrDeps map");