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 (isFreeCall(Inst)) {
122 Pointer = Inst->getOperand(0);
123 // calls to free() erase the entire structure
125 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
126 // Debug intrinsics don't cause dependences.
127 if (isa<DbgInfoIntrinsic>(Inst)) continue;
128 CallSite InstCS = CallSite::get(Inst);
129 // If these two calls do not interfere, look past it.
130 switch (AA->getModRefInfo(CS, InstCS)) {
131 case AliasAnalysis::NoModRef:
132 // If the two calls don't interact (e.g. InstCS is readnone) keep
135 case AliasAnalysis::Ref:
136 // If the two calls read the same memory locations and CS is a readonly
137 // function, then we have two cases: 1) the calls may not interfere with
138 // each other at all. 2) the calls may produce the same value. In case
139 // #1 we want to ignore the values, in case #2, we want to return Inst
140 // as a Def dependence. This allows us to CSE in cases like:
143 // Y = strlen(P); // Y = X
144 if (isReadOnlyCall) {
145 if (CS.getCalledFunction() != 0 &&
146 CS.getCalledFunction() == InstCS.getCalledFunction())
147 return MemDepResult::getDef(Inst);
148 // Ignore unrelated read/read call dependences.
153 return MemDepResult::getClobber(Inst);
156 // Non-memory instruction.
160 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef)
161 return MemDepResult::getClobber(Inst);
164 // No dependence found. If this is the entry block of the function, it is a
165 // clobber, otherwise it is non-local.
166 if (BB != &BB->getParent()->getEntryBlock())
167 return MemDepResult::getNonLocal();
168 return MemDepResult::getClobber(ScanIt);
171 /// getPointerDependencyFrom - Return the instruction on which a memory
172 /// location depends. If isLoad is true, this routine ignore may-aliases with
173 /// read-only operations.
174 MemDepResult MemoryDependenceAnalysis::
175 getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
176 BasicBlock::iterator ScanIt, BasicBlock *BB) {
178 // Walk backwards through the basic block, looking for dependencies.
179 while (ScanIt != BB->begin()) {
180 Instruction *Inst = --ScanIt;
182 // Debug intrinsics don't cause dependences.
183 if (isa<DbgInfoIntrinsic>(Inst)) continue;
185 // Values depend on loads if the pointers are must aliased. This means that
186 // a load depends on another must aliased load from the same value.
187 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
188 Value *Pointer = LI->getPointerOperand();
189 uint64_t PointerSize = AA->getTypeStoreSize(LI->getType());
191 // If we found a pointer, check if it could be the same as our pointer.
192 AliasAnalysis::AliasResult R =
193 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
194 if (R == AliasAnalysis::NoAlias)
197 // May-alias loads don't depend on each other without a dependence.
198 if (isLoad && R == AliasAnalysis::MayAlias)
200 // Stores depend on may and must aliased loads, loads depend on must-alias
202 return MemDepResult::getDef(Inst);
205 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
206 // If alias analysis can tell that this store is guaranteed to not modify
207 // the query pointer, ignore it. Use getModRefInfo to handle cases where
208 // the query pointer points to constant memory etc.
209 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef)
212 // Ok, this store might clobber the query pointer. Check to see if it is
213 // a must alias: in this case, we want to return this as a def.
214 Value *Pointer = SI->getPointerOperand();
215 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
217 // If we found a pointer, check if it could be the same as our pointer.
218 AliasAnalysis::AliasResult R =
219 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
221 if (R == AliasAnalysis::NoAlias)
223 if (R == AliasAnalysis::MayAlias)
224 return MemDepResult::getClobber(Inst);
225 return MemDepResult::getDef(Inst);
228 // If this is an allocation, and if we know that the accessed pointer is to
229 // the allocation, return Def. This means that there is no dependence and
230 // the access can be optimized based on that. For example, a load could
232 // Note: Only determine this to be a malloc if Inst is the malloc call, not
233 // a subsequent bitcast of the malloc call result. There can be stores to
234 // the malloced memory between the malloc call and its bitcast uses, and we
235 // need to continue scanning until the malloc call.
236 if (isa<AllocaInst>(Inst) || extractMallocCall(Inst)) {
237 Value *AccessPtr = MemPtr->getUnderlyingObject();
239 if (AccessPtr == Inst ||
240 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
241 return MemDepResult::getDef(Inst);
245 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
246 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
247 case AliasAnalysis::NoModRef:
248 // If the call has no effect on the queried pointer, just ignore it.
250 case AliasAnalysis::Ref:
251 // If the call is known to never store to the pointer, and if this is a
252 // load query, we can safely ignore it (scan past it).
257 // Otherwise, there is a potential dependence. Return a clobber.
258 return MemDepResult::getClobber(Inst);
262 // No dependence found. If this is the entry block of the function, it is a
263 // clobber, otherwise it is non-local.
264 if (BB != &BB->getParent()->getEntryBlock())
265 return MemDepResult::getNonLocal();
266 return MemDepResult::getClobber(ScanIt);
269 /// getDependency - Return the instruction on which a memory operation
271 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
272 Instruction *ScanPos = QueryInst;
274 // Check for a cached result
275 MemDepResult &LocalCache = LocalDeps[QueryInst];
277 // If the cached entry is non-dirty, just return it. Note that this depends
278 // on MemDepResult's default constructing to 'dirty'.
279 if (!LocalCache.isDirty())
282 // Otherwise, if we have a dirty entry, we know we can start the scan at that
283 // instruction, which may save us some work.
284 if (Instruction *Inst = LocalCache.getInst()) {
287 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
290 BasicBlock *QueryParent = QueryInst->getParent();
293 uint64_t MemSize = 0;
296 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
297 // No dependence found. If this is the entry block of the function, it is a
298 // clobber, otherwise it is non-local.
299 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
300 LocalCache = MemDepResult::getNonLocal();
302 LocalCache = MemDepResult::getClobber(QueryInst);
303 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
304 // If this is a volatile store, don't mess around with it. Just return the
305 // previous instruction as a clobber.
306 if (SI->isVolatile())
307 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
309 MemPtr = SI->getPointerOperand();
310 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
312 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
313 // If this is a volatile load, don't mess around with it. Just return the
314 // previous instruction as a clobber.
315 if (LI->isVolatile())
316 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
318 MemPtr = LI->getPointerOperand();
319 MemSize = AA->getTypeStoreSize(LI->getType());
321 } else if (isFreeCall(QueryInst)) {
322 MemPtr = QueryInst->getOperand(0);
323 // calls to free() erase the entire structure, not just a field.
325 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
326 CallSite QueryCS = CallSite::get(QueryInst);
327 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
328 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
330 } else if (FreeInst *FI = dyn_cast<FreeInst>(QueryInst)) {
331 MemPtr = FI->getPointerOperand();
332 // FreeInsts erase the entire structure, not just a field.
335 // Non-memory instruction.
336 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
339 // If we need to do a pointer scan, make it happen.
341 LocalCache = getPointerDependencyFrom(MemPtr, MemSize,
342 isa<LoadInst>(QueryInst),
343 ScanPos, QueryParent);
345 // Remember the result!
346 if (Instruction *I = LocalCache.getInst())
347 ReverseLocalDeps[I].insert(QueryInst);
353 /// AssertSorted - This method is used when -debug is specified to verify that
354 /// cache arrays are properly kept sorted.
355 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
357 if (Count == -1) Count = Cache.size();
358 if (Count == 0) return;
360 for (unsigned i = 1; i != unsigned(Count); ++i)
361 assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!");
365 /// getNonLocalCallDependency - Perform a full dependency query for the
366 /// specified call, returning the set of blocks that the value is
367 /// potentially live across. The returned set of results will include a
368 /// "NonLocal" result for all blocks where the value is live across.
370 /// This method assumes the instruction returns a "NonLocal" dependency
371 /// within its own block.
373 /// This returns a reference to an internal data structure that may be
374 /// invalidated on the next non-local query or when an instruction is
375 /// removed. Clients must copy this data if they want it around longer than
377 const MemoryDependenceAnalysis::NonLocalDepInfo &
378 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
379 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
380 "getNonLocalCallDependency should only be used on calls with non-local deps!");
381 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
382 NonLocalDepInfo &Cache = CacheP.first;
384 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
385 /// the cached case, this can happen due to instructions being deleted etc. In
386 /// the uncached case, this starts out as the set of predecessors we care
388 SmallVector<BasicBlock*, 32> DirtyBlocks;
390 if (!Cache.empty()) {
391 // Okay, we have a cache entry. If we know it is not dirty, just return it
392 // with no computation.
393 if (!CacheP.second) {
398 // If we already have a partially computed set of results, scan them to
399 // determine what is dirty, seeding our initial DirtyBlocks worklist.
400 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
402 if (I->second.isDirty())
403 DirtyBlocks.push_back(I->first);
405 // Sort the cache so that we can do fast binary search lookups below.
406 std::sort(Cache.begin(), Cache.end());
408 ++NumCacheDirtyNonLocal;
409 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
410 // << Cache.size() << " cached: " << *QueryInst;
412 // Seed DirtyBlocks with each of the preds of QueryInst's block.
413 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
414 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
415 DirtyBlocks.push_back(*PI);
416 NumUncacheNonLocal++;
419 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
420 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
422 SmallPtrSet<BasicBlock*, 64> Visited;
424 unsigned NumSortedEntries = Cache.size();
425 DEBUG(AssertSorted(Cache));
427 // Iterate while we still have blocks to update.
428 while (!DirtyBlocks.empty()) {
429 BasicBlock *DirtyBB = DirtyBlocks.back();
430 DirtyBlocks.pop_back();
432 // Already processed this block?
433 if (!Visited.insert(DirtyBB))
436 // Do a binary search to see if we already have an entry for this block in
437 // the cache set. If so, find it.
438 DEBUG(AssertSorted(Cache, NumSortedEntries));
439 NonLocalDepInfo::iterator Entry =
440 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
441 std::make_pair(DirtyBB, MemDepResult()));
442 if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB)
445 MemDepResult *ExistingResult = 0;
446 if (Entry != Cache.begin()+NumSortedEntries &&
447 Entry->first == DirtyBB) {
448 // If we already have an entry, and if it isn't already dirty, the block
450 if (!Entry->second.isDirty())
453 // Otherwise, remember this slot so we can update the value.
454 ExistingResult = &Entry->second;
457 // If the dirty entry has a pointer, start scanning from it so we don't have
458 // to rescan the entire block.
459 BasicBlock::iterator ScanPos = DirtyBB->end();
460 if (ExistingResult) {
461 if (Instruction *Inst = ExistingResult->getInst()) {
463 // We're removing QueryInst's use of Inst.
464 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
465 QueryCS.getInstruction());
469 // Find out if this block has a local dependency for QueryInst.
472 if (ScanPos != DirtyBB->begin()) {
473 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
474 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
475 // No dependence found. If this is the entry block of the function, it is
476 // a clobber, otherwise it is non-local.
477 Dep = MemDepResult::getNonLocal();
479 Dep = MemDepResult::getClobber(ScanPos);
482 // If we had a dirty entry for the block, update it. Otherwise, just add
485 *ExistingResult = Dep;
487 Cache.push_back(std::make_pair(DirtyBB, Dep));
489 // If the block has a dependency (i.e. it isn't completely transparent to
490 // the value), remember the association!
491 if (!Dep.isNonLocal()) {
492 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
493 // update this when we remove instructions.
494 if (Instruction *Inst = Dep.getInst())
495 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
498 // If the block *is* completely transparent to the load, we need to check
499 // the predecessors of this block. Add them to our worklist.
500 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
501 DirtyBlocks.push_back(*PI);
508 /// getNonLocalPointerDependency - Perform a full dependency query for an
509 /// access to the specified (non-volatile) memory location, returning the
510 /// set of instructions that either define or clobber the value.
512 /// This method assumes the pointer has a "NonLocal" dependency within its
515 void MemoryDependenceAnalysis::
516 getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
517 SmallVectorImpl<NonLocalDepEntry> &Result) {
518 assert(isa<PointerType>(Pointer->getType()) &&
519 "Can't get pointer deps of a non-pointer!");
522 // We know that the pointer value is live into FromBB find the def/clobbers
523 // from presecessors.
524 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
525 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy);
527 // This is the set of blocks we've inspected, and the pointer we consider in
528 // each block. Because of critical edges, we currently bail out if querying
529 // a block with multiple different pointers. This can happen during PHI
531 DenseMap<BasicBlock*, Value*> Visited;
532 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB,
533 Result, Visited, true))
536 Result.push_back(std::make_pair(FromBB,
537 MemDepResult::getClobber(FromBB->begin())));
540 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
541 /// Pointer/PointeeSize using either cached information in Cache or by doing a
542 /// lookup (which may use dirty cache info if available). If we do a lookup,
543 /// add the result to the cache.
544 MemDepResult MemoryDependenceAnalysis::
545 GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
546 bool isLoad, BasicBlock *BB,
547 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
549 // Do a binary search to see if we already have an entry for this block in
550 // the cache set. If so, find it.
551 NonLocalDepInfo::iterator Entry =
552 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
553 std::make_pair(BB, MemDepResult()));
554 if (Entry != Cache->begin() && prior(Entry)->first == BB)
557 MemDepResult *ExistingResult = 0;
558 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
559 ExistingResult = &Entry->second;
561 // If we have a cached entry, and it is non-dirty, use it as the value for
563 if (ExistingResult && !ExistingResult->isDirty()) {
564 ++NumCacheNonLocalPtr;
565 return *ExistingResult;
568 // Otherwise, we have to scan for the value. If we have a dirty cache
569 // entry, start scanning from its position, otherwise we scan from the end
571 BasicBlock::iterator ScanPos = BB->end();
572 if (ExistingResult && ExistingResult->getInst()) {
573 assert(ExistingResult->getInst()->getParent() == BB &&
574 "Instruction invalidated?");
575 ++NumCacheDirtyNonLocalPtr;
576 ScanPos = ExistingResult->getInst();
578 // Eliminating the dirty entry from 'Cache', so update the reverse info.
579 ValueIsLoadPair CacheKey(Pointer, isLoad);
580 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
582 ++NumUncacheNonLocalPtr;
585 // Scan the block for the dependency.
586 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad,
589 // If we had a dirty entry for the block, update it. Otherwise, just add
592 *ExistingResult = Dep;
594 Cache->push_back(std::make_pair(BB, Dep));
596 // If the block has a dependency (i.e. it isn't completely transparent to
597 // the value), remember the reverse association because we just added it
599 if (Dep.isNonLocal())
602 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
603 // update MemDep when we remove instructions.
604 Instruction *Inst = Dep.getInst();
605 assert(Inst && "Didn't depend on anything?");
606 ValueIsLoadPair CacheKey(Pointer, isLoad);
607 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
611 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
612 /// number of elements in the array that are already properly ordered. This is
613 /// optimized for the case when only a few entries are added.
615 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
616 unsigned NumSortedEntries) {
617 switch (Cache.size() - NumSortedEntries) {
619 // done, no new entries.
622 // Two new entries, insert the last one into place.
623 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
625 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
626 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
627 Cache.insert(Entry, Val);
631 // One new entry, Just insert the new value at the appropriate position.
632 if (Cache.size() != 1) {
633 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
635 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
636 std::upper_bound(Cache.begin(), Cache.end(), Val);
637 Cache.insert(Entry, Val);
641 // Added many values, do a full scale sort.
642 std::sort(Cache.begin(), Cache.end());
648 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
649 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
650 /// results to the results vector and keep track of which blocks are visited in
653 /// This has special behavior for the first block queries (when SkipFirstBlock
654 /// is true). In this special case, it ignores the contents of the specified
655 /// block and starts returning dependence info for its predecessors.
657 /// This function returns false on success, or true to indicate that it could
658 /// not compute dependence information for some reason. This should be treated
659 /// as a clobber dependence on the first instruction in the predecessor block.
660 bool MemoryDependenceAnalysis::
661 getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
662 bool isLoad, BasicBlock *StartBB,
663 SmallVectorImpl<NonLocalDepEntry> &Result,
664 DenseMap<BasicBlock*, Value*> &Visited,
665 bool SkipFirstBlock) {
667 // Look up the cached info for Pointer.
668 ValueIsLoadPair CacheKey(Pointer, isLoad);
670 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
671 &NonLocalPointerDeps[CacheKey];
672 NonLocalDepInfo *Cache = &CacheInfo->second;
674 // If we have valid cached information for exactly the block we are
675 // investigating, just return it with no recomputation.
676 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
677 // We have a fully cached result for this query then we can just return the
678 // cached results and populate the visited set. However, we have to verify
679 // that we don't already have conflicting results for these blocks. Check
680 // to ensure that if a block in the results set is in the visited set that
681 // it was for the same pointer query.
682 if (!Visited.empty()) {
683 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
685 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->first);
686 if (VI == Visited.end() || VI->second == Pointer) continue;
688 // We have a pointer mismatch in a block. Just return clobber, saying
689 // that something was clobbered in this result. We could also do a
690 // non-fully cached query, but there is little point in doing this.
695 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
697 Visited.insert(std::make_pair(I->first, Pointer));
698 if (!I->second.isNonLocal())
699 Result.push_back(*I);
701 ++NumCacheCompleteNonLocalPtr;
705 // Otherwise, either this is a new block, a block with an invalid cache
706 // pointer or one that we're about to invalidate by putting more info into it
707 // than its valid cache info. If empty, the result will be valid cache info,
708 // otherwise it isn't.
710 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
712 CacheInfo->first = BBSkipFirstBlockPair();
714 SmallVector<BasicBlock*, 32> Worklist;
715 Worklist.push_back(StartBB);
717 // Keep track of the entries that we know are sorted. Previously cached
718 // entries will all be sorted. The entries we add we only sort on demand (we
719 // don't insert every element into its sorted position). We know that we
720 // won't get any reuse from currently inserted values, because we don't
721 // revisit blocks after we insert info for them.
722 unsigned NumSortedEntries = Cache->size();
723 DEBUG(AssertSorted(*Cache));
725 while (!Worklist.empty()) {
726 BasicBlock *BB = Worklist.pop_back_val();
728 // Skip the first block if we have it.
729 if (!SkipFirstBlock) {
730 // Analyze the dependency of *Pointer in FromBB. See if we already have
732 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
734 // Get the dependency info for Pointer in BB. If we have cached
735 // information, we will use it, otherwise we compute it.
736 DEBUG(AssertSorted(*Cache, NumSortedEntries));
737 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad,
738 BB, Cache, NumSortedEntries);
740 // If we got a Def or Clobber, add this to the list of results.
741 if (!Dep.isNonLocal()) {
742 Result.push_back(NonLocalDepEntry(BB, Dep));
747 // If 'Pointer' is an instruction defined in this block, then we need to do
748 // phi translation to change it into a value live in the predecessor block.
749 // If phi translation fails, then we can't continue dependence analysis.
750 Instruction *PtrInst = dyn_cast<Instruction>(Pointer);
751 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB;
753 // If no PHI translation is needed, just add all the predecessors of this
754 // block to scan them as well.
755 if (!NeedsPHITranslation) {
756 SkipFirstBlock = false;
757 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
758 // Verify that we haven't looked at this block yet.
759 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
760 InsertRes = Visited.insert(std::make_pair(*PI, Pointer));
761 if (InsertRes.second) {
762 // First time we've looked at *PI.
763 Worklist.push_back(*PI);
767 // If we have seen this block before, but it was with a different
768 // pointer then we have a phi translation failure and we have to treat
769 // this as a clobber.
770 if (InsertRes.first->second != Pointer)
771 goto PredTranslationFailure;
776 // If we do need to do phi translation, then there are a bunch of different
777 // cases, because we have to find a Value* live in the predecessor block. We
778 // know that PtrInst is defined in this block at least.
780 // We may have added values to the cache list before this PHI translation.
781 // If so, we haven't done anything to ensure that the cache remains sorted.
782 // Sort it now (if needed) so that recursive invocations of
783 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
784 // value will only see properly sorted cache arrays.
785 if (Cache && NumSortedEntries != Cache->size()) {
786 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
787 NumSortedEntries = Cache->size();
790 // If this is directly a PHI node, just use the incoming values for each
791 // pred as the phi translated version.
792 if (PHINode *PtrPHI = dyn_cast<PHINode>(PtrInst)) {
795 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
796 BasicBlock *Pred = *PI;
797 Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred);
799 // Check to see if we have already visited this pred block with another
800 // pointer. If so, we can't do this lookup. This failure can occur
801 // with PHI translation when a critical edge exists and the PHI node in
802 // the successor translates to a pointer value different than the
803 // pointer the block was first analyzed with.
804 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
805 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr));
807 if (!InsertRes.second) {
808 // If the predecessor was visited with PredPtr, then we already did
809 // the analysis and can ignore it.
810 if (InsertRes.first->second == PredPtr)
813 // Otherwise, the block was previously analyzed with a different
814 // pointer. We can't represent the result of this case, so we just
815 // treat this as a phi translation failure.
816 goto PredTranslationFailure;
819 // FIXME: it is entirely possible that PHI translating will end up with
820 // the same value. Consider PHI translating something like:
821 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
822 // to recurse here, pedantically speaking.
824 // If we have a problem phi translating, fall through to the code below
825 // to handle the failure condition.
826 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred,
828 goto PredTranslationFailure;
831 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
832 CacheInfo = &NonLocalPointerDeps[CacheKey];
833 Cache = &CacheInfo->second;
834 NumSortedEntries = Cache->size();
836 // Since we did phi translation, the "Cache" set won't contain all of the
837 // results for the query. This is ok (we can still use it to accelerate
838 // specific block queries) but we can't do the fastpath "return all
839 // results from the set" Clear out the indicator for this.
840 CacheInfo->first = BBSkipFirstBlockPair();
841 SkipFirstBlock = false;
845 // TODO: BITCAST, GEP.
847 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
848 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
849 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
850 PredTranslationFailure:
853 // Refresh the CacheInfo/Cache pointer if it got invalidated.
854 CacheInfo = &NonLocalPointerDeps[CacheKey];
855 Cache = &CacheInfo->second;
856 NumSortedEntries = Cache->size();
859 // Since we did phi translation, the "Cache" set won't contain all of the
860 // results for the query. This is ok (we can still use it to accelerate
861 // specific block queries) but we can't do the fastpath "return all
862 // results from the set" Clear out the indicator for this.
863 CacheInfo->first = BBSkipFirstBlockPair();
865 // If *nothing* works, mark the pointer as being clobbered by the first
866 // instruction in this block.
868 // If this is the magic first block, return this as a clobber of the whole
869 // incoming value. Since we can't phi translate to one of the predecessors,
870 // we have to bail out.
874 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
875 assert(I != Cache->rend() && "Didn't find current block??");
879 assert(I->second.isNonLocal() &&
880 "Should only be here with transparent block");
881 I->second = MemDepResult::getClobber(BB->begin());
882 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
883 Result.push_back(*I);
888 // Okay, we're done now. If we added new values to the cache, re-sort it.
889 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
890 DEBUG(AssertSorted(*Cache));
894 /// RemoveCachedNonLocalPointerDependencies - If P exists in
895 /// CachedNonLocalPointerInfo, remove it.
896 void MemoryDependenceAnalysis::
897 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
898 CachedNonLocalPointerInfo::iterator It =
899 NonLocalPointerDeps.find(P);
900 if (It == NonLocalPointerDeps.end()) return;
902 // Remove all of the entries in the BB->val map. This involves removing
903 // instructions from the reverse map.
904 NonLocalDepInfo &PInfo = It->second.second;
906 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
907 Instruction *Target = PInfo[i].second.getInst();
908 if (Target == 0) continue; // Ignore non-local dep results.
909 assert(Target->getParent() == PInfo[i].first);
911 // Eliminating the dirty entry from 'Cache', so update the reverse info.
912 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
915 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
916 NonLocalPointerDeps.erase(It);
920 /// invalidateCachedPointerInfo - This method is used to invalidate cached
921 /// information about the specified pointer, because it may be too
922 /// conservative in memdep. This is an optional call that can be used when
923 /// the client detects an equivalence between the pointer and some other
924 /// value and replaces the other value with ptr. This can make Ptr available
925 /// in more places that cached info does not necessarily keep.
926 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
927 // If Ptr isn't really a pointer, just ignore it.
928 if (!isa<PointerType>(Ptr->getType())) return;
929 // Flush store info for the pointer.
930 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
931 // Flush load info for the pointer.
932 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
935 /// removeInstruction - Remove an instruction from the dependence analysis,
936 /// updating the dependence of instructions that previously depended on it.
937 /// This method attempts to keep the cache coherent using the reverse map.
938 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
939 // Walk through the Non-local dependencies, removing this one as the value
940 // for any cached queries.
941 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
942 if (NLDI != NonLocalDeps.end()) {
943 NonLocalDepInfo &BlockMap = NLDI->second.first;
944 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
946 if (Instruction *Inst = DI->second.getInst())
947 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
948 NonLocalDeps.erase(NLDI);
951 // If we have a cached local dependence query for this instruction, remove it.
953 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
954 if (LocalDepEntry != LocalDeps.end()) {
955 // Remove us from DepInst's reverse set now that the local dep info is gone.
956 if (Instruction *Inst = LocalDepEntry->second.getInst())
957 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
959 // Remove this local dependency info.
960 LocalDeps.erase(LocalDepEntry);
963 // If we have any cached pointer dependencies on this instruction, remove
964 // them. If the instruction has non-pointer type, then it can't be a pointer
967 // Remove it from both the load info and the store info. The instruction
968 // can't be in either of these maps if it is non-pointer.
969 if (isa<PointerType>(RemInst->getType())) {
970 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
971 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
974 // Loop over all of the things that depend on the instruction we're removing.
976 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
978 // If we find RemInst as a clobber or Def in any of the maps for other values,
979 // we need to replace its entry with a dirty version of the instruction after
980 // it. If RemInst is a terminator, we use a null dirty value.
982 // Using a dirty version of the instruction after RemInst saves having to scan
983 // the entire block to get to this point.
984 MemDepResult NewDirtyVal;
985 if (!RemInst->isTerminator())
986 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
988 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
989 if (ReverseDepIt != ReverseLocalDeps.end()) {
990 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
991 // RemInst can't be the terminator if it has local stuff depending on it.
992 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
993 "Nothing can locally depend on a terminator");
995 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
996 E = ReverseDeps.end(); I != E; ++I) {
997 Instruction *InstDependingOnRemInst = *I;
998 assert(InstDependingOnRemInst != RemInst &&
999 "Already removed our local dep info");
1001 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1003 // Make sure to remember that new things depend on NewDepInst.
1004 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1005 "a local dep on this if it is a terminator!");
1006 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1007 InstDependingOnRemInst));
1010 ReverseLocalDeps.erase(ReverseDepIt);
1012 // Add new reverse deps after scanning the set, to avoid invalidating the
1013 // 'ReverseDeps' reference.
1014 while (!ReverseDepsToAdd.empty()) {
1015 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1016 .insert(ReverseDepsToAdd.back().second);
1017 ReverseDepsToAdd.pop_back();
1021 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1022 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1023 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1024 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1026 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1028 PerInstNLInfo &INLD = NonLocalDeps[*I];
1029 // The information is now dirty!
1032 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1033 DE = INLD.first.end(); DI != DE; ++DI) {
1034 if (DI->second.getInst() != RemInst) continue;
1036 // Convert to a dirty entry for the subsequent instruction.
1037 DI->second = NewDirtyVal;
1039 if (Instruction *NextI = NewDirtyVal.getInst())
1040 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1044 ReverseNonLocalDeps.erase(ReverseDepIt);
1046 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1047 while (!ReverseDepsToAdd.empty()) {
1048 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1049 .insert(ReverseDepsToAdd.back().second);
1050 ReverseDepsToAdd.pop_back();
1054 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1055 // value in the NonLocalPointerDeps info.
1056 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1057 ReverseNonLocalPtrDeps.find(RemInst);
1058 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1059 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1060 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1062 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1063 E = Set.end(); I != E; ++I) {
1064 ValueIsLoadPair P = *I;
1065 assert(P.getPointer() != RemInst &&
1066 "Already removed NonLocalPointerDeps info for RemInst");
1068 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second;
1070 // The cache is not valid for any specific block anymore.
1071 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair();
1073 // Update any entries for RemInst to use the instruction after it.
1074 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1076 if (DI->second.getInst() != RemInst) continue;
1078 // Convert to a dirty entry for the subsequent instruction.
1079 DI->second = NewDirtyVal;
1081 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1082 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1085 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1086 // subsequent value may invalidate the sortedness.
1087 std::sort(NLPDI.begin(), NLPDI.end());
1090 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1092 while (!ReversePtrDepsToAdd.empty()) {
1093 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1094 .insert(ReversePtrDepsToAdd.back().second);
1095 ReversePtrDepsToAdd.pop_back();
1100 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1101 AA->deleteValue(RemInst);
1102 DEBUG(verifyRemoved(RemInst));
1104 /// verifyRemoved - Verify that the specified instruction does not occur
1105 /// in our internal data structures.
1106 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1107 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1108 E = LocalDeps.end(); I != E; ++I) {
1109 assert(I->first != D && "Inst occurs in data structures");
1110 assert(I->second.getInst() != D &&
1111 "Inst occurs in data structures");
1114 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1115 E = NonLocalPointerDeps.end(); I != E; ++I) {
1116 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1117 const NonLocalDepInfo &Val = I->second.second;
1118 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1120 assert(II->second.getInst() != D && "Inst occurs as NLPD value");
1123 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1124 E = NonLocalDeps.end(); I != E; ++I) {
1125 assert(I->first != D && "Inst occurs in data structures");
1126 const PerInstNLInfo &INLD = I->second;
1127 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1128 EE = INLD.first.end(); II != EE; ++II)
1129 assert(II->second.getInst() != D && "Inst occurs in data structures");
1132 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1133 E = ReverseLocalDeps.end(); I != E; ++I) {
1134 assert(I->first != D && "Inst occurs in data structures");
1135 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1136 EE = I->second.end(); II != EE; ++II)
1137 assert(*II != D && "Inst occurs in data structures");
1140 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1141 E = ReverseNonLocalDeps.end();
1143 assert(I->first != D && "Inst occurs in data structures");
1144 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1145 EE = I->second.end(); II != EE; ++II)
1146 assert(*II != D && "Inst occurs in data structures");
1149 for (ReverseNonLocalPtrDepTy::const_iterator
1150 I = ReverseNonLocalPtrDeps.begin(),
1151 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1152 assert(I->first != D && "Inst occurs in rev NLPD map");
1154 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1155 E = I->second.end(); II != E; ++II)
1156 assert(*II != ValueIsLoadPair(D, false) &&
1157 *II != ValueIsLoadPair(D, true) &&
1158 "Inst occurs in ReverseNonLocalPtrDeps map");