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/MemoryBuiltins.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 (isFreeCall(Inst)) {
117 Pointer = Inst->getOperand(1);
118 // calls to free() erase the entire structure
120 } else if (isFreeCall(Inst)) {
121 Pointer = Inst->getOperand(0);
122 // calls to free() erase the entire structure
124 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
125 // Debug intrinsics don't cause dependences.
126 if (isa<DbgInfoIntrinsic>(Inst)) continue;
127 CallSite InstCS = CallSite::get(Inst);
128 // If these two calls do not interfere, look past it.
129 switch (AA->getModRefInfo(CS, InstCS)) {
130 case AliasAnalysis::NoModRef:
131 // If the two calls don't interact (e.g. InstCS is readnone) keep
134 case AliasAnalysis::Ref:
135 // If the two calls read the same memory locations and CS is a readonly
136 // function, then we have two cases: 1) the calls may not interfere with
137 // each other at all. 2) the calls may produce the same value. In case
138 // #1 we want to ignore the values, in case #2, we want to return Inst
139 // as a Def dependence. This allows us to CSE in cases like:
142 // Y = strlen(P); // Y = X
143 if (isReadOnlyCall) {
144 if (CS.getCalledFunction() != 0 &&
145 CS.getCalledFunction() == InstCS.getCalledFunction())
146 return MemDepResult::getDef(Inst);
147 // Ignore unrelated read/read call dependences.
152 return MemDepResult::getClobber(Inst);
155 // Non-memory instruction.
159 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef)
160 return MemDepResult::getClobber(Inst);
163 // No dependence found. If this is the entry block of the function, it is a
164 // clobber, otherwise it is non-local.
165 if (BB != &BB->getParent()->getEntryBlock())
166 return MemDepResult::getNonLocal();
167 return MemDepResult::getClobber(ScanIt);
170 /// getPointerDependencyFrom - Return the instruction on which a memory
171 /// location depends. If isLoad is true, this routine ignore may-aliases with
172 /// read-only operations.
173 MemDepResult MemoryDependenceAnalysis::
174 getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
175 BasicBlock::iterator ScanIt, BasicBlock *BB) {
177 Value* invariantTag = 0;
179 // Walk backwards through the basic block, looking for dependencies.
180 while (ScanIt != BB->begin()) {
181 Instruction *Inst = --ScanIt;
183 // If we're in an invariant region, no dependencies can be found before
184 // we pass an invariant-begin marker.
185 if (invariantTag == Inst) {
189 // If we pass an invariant-end marker, then we've just entered an invariant
190 // region and can start ignoring dependencies.
191 } else if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) {
192 if (II->getIntrinsicID() == Intrinsic::invariant_end) {
193 uint64_t invariantSize = ~0ULL;
194 if (ConstantInt* CI = dyn_cast<ConstantInt>(II->getOperand(2)))
195 invariantSize = CI->getZExtValue();
197 AliasAnalysis::AliasResult R =
198 AA->alias(II->getOperand(3), invariantSize, MemPtr, MemSize);
199 if (R == AliasAnalysis::MustAlias) {
200 invariantTag = II->getOperand(1);
206 // If we're querying on a load and we're in an invariant region, we're done
207 // at this point. Nothing a load depends on can live in an invariant region.
208 if (isLoad && invariantTag) continue;
210 // Debug intrinsics don't cause dependences.
211 if (isa<DbgInfoIntrinsic>(Inst)) continue;
213 // Values depend on loads if the pointers are must aliased. This means that
214 // a load depends on another must aliased load from the same value.
215 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
216 Value *Pointer = LI->getPointerOperand();
217 uint64_t PointerSize = AA->getTypeStoreSize(LI->getType());
219 // If we found a pointer, check if it could be the same as our pointer.
220 AliasAnalysis::AliasResult R =
221 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
222 if (R == AliasAnalysis::NoAlias)
225 // May-alias loads don't depend on each other without a dependence.
226 if (isLoad && R == AliasAnalysis::MayAlias)
228 // Stores depend on may and must aliased loads, loads depend on must-alias
230 return MemDepResult::getDef(Inst);
233 // If we're querying on a store and we're in an invariant region, we're done
234 // at this point. The only things that stores depend on that could exist in
235 // an invariant region are loads, which we've already checked.
236 if (invariantTag) continue;
238 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
239 // If alias analysis can tell that this store is guaranteed to not modify
240 // the query pointer, ignore it. Use getModRefInfo to handle cases where
241 // the query pointer points to constant memory etc.
242 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef)
245 // Ok, this store might clobber the query pointer. Check to see if it is
246 // a must alias: in this case, we want to return this as a def.
247 Value *Pointer = SI->getPointerOperand();
248 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
250 // If we found a pointer, check if it could be the same as our pointer.
251 AliasAnalysis::AliasResult R =
252 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
254 if (R == AliasAnalysis::NoAlias)
256 if (R == AliasAnalysis::MayAlias)
257 return MemDepResult::getClobber(Inst);
258 return MemDepResult::getDef(Inst);
261 // If this is an allocation, and if we know that the accessed pointer is to
262 // the allocation, return Def. This means that there is no dependence and
263 // the access can be optimized based on that. For example, a load could
265 // Note: Only determine this to be a malloc if Inst is the malloc call, not
266 // a subsequent bitcast of the malloc call result. There can be stores to
267 // the malloced memory between the malloc call and its bitcast uses, and we
268 // need to continue scanning until the malloc call.
269 if (isa<AllocaInst>(Inst) || extractMallocCall(Inst)) {
270 Value *AccessPtr = MemPtr->getUnderlyingObject();
272 if (AccessPtr == Inst ||
273 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
274 return MemDepResult::getDef(Inst);
278 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
279 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
280 case AliasAnalysis::NoModRef:
281 // If the call has no effect on the queried pointer, just ignore it.
283 case AliasAnalysis::Ref:
284 // If the call is known to never store to the pointer, and if this is a
285 // load query, we can safely ignore it (scan past it).
290 // Otherwise, there is a potential dependence. Return a clobber.
291 return MemDepResult::getClobber(Inst);
295 // No dependence found. If this is the entry block of the function, it is a
296 // clobber, otherwise it is non-local.
297 if (BB != &BB->getParent()->getEntryBlock())
298 return MemDepResult::getNonLocal();
299 return MemDepResult::getClobber(ScanIt);
302 /// getDependency - Return the instruction on which a memory operation
304 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
305 Instruction *ScanPos = QueryInst;
307 // Check for a cached result
308 MemDepResult &LocalCache = LocalDeps[QueryInst];
310 // If the cached entry is non-dirty, just return it. Note that this depends
311 // on MemDepResult's default constructing to 'dirty'.
312 if (!LocalCache.isDirty())
315 // Otherwise, if we have a dirty entry, we know we can start the scan at that
316 // instruction, which may save us some work.
317 if (Instruction *Inst = LocalCache.getInst()) {
320 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
323 BasicBlock *QueryParent = QueryInst->getParent();
326 uint64_t MemSize = 0;
329 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
330 // No dependence found. If this is the entry block of the function, it is a
331 // clobber, otherwise it is non-local.
332 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
333 LocalCache = MemDepResult::getNonLocal();
335 LocalCache = MemDepResult::getClobber(QueryInst);
336 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
337 // If this is a volatile store, don't mess around with it. Just return the
338 // previous instruction as a clobber.
339 if (SI->isVolatile())
340 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
342 MemPtr = SI->getPointerOperand();
343 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
345 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
346 // If this is a volatile load, don't mess around with it. Just return the
347 // previous instruction as a clobber.
348 if (LI->isVolatile())
349 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
351 MemPtr = LI->getPointerOperand();
352 MemSize = AA->getTypeStoreSize(LI->getType());
354 } else if (isFreeCall(QueryInst)) {
355 MemPtr = QueryInst->getOperand(1);
356 // calls to free() erase the entire structure, not just a field.
358 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
359 CallSite QueryCS = CallSite::get(QueryInst);
360 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
361 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
364 // Non-memory instruction.
365 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
368 // If we need to do a pointer scan, make it happen.
370 LocalCache = getPointerDependencyFrom(MemPtr, MemSize,
371 isa<LoadInst>(QueryInst),
372 ScanPos, QueryParent);
374 // Remember the result!
375 if (Instruction *I = LocalCache.getInst())
376 ReverseLocalDeps[I].insert(QueryInst);
382 /// AssertSorted - This method is used when -debug is specified to verify that
383 /// cache arrays are properly kept sorted.
384 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
386 if (Count == -1) Count = Cache.size();
387 if (Count == 0) return;
389 for (unsigned i = 1; i != unsigned(Count); ++i)
390 assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!");
394 /// getNonLocalCallDependency - Perform a full dependency query for the
395 /// specified call, returning the set of blocks that the value is
396 /// potentially live across. The returned set of results will include a
397 /// "NonLocal" result for all blocks where the value is live across.
399 /// This method assumes the instruction returns a "NonLocal" dependency
400 /// within its own block.
402 /// This returns a reference to an internal data structure that may be
403 /// invalidated on the next non-local query or when an instruction is
404 /// removed. Clients must copy this data if they want it around longer than
406 const MemoryDependenceAnalysis::NonLocalDepInfo &
407 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
408 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
409 "getNonLocalCallDependency should only be used on calls with non-local deps!");
410 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
411 NonLocalDepInfo &Cache = CacheP.first;
413 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
414 /// the cached case, this can happen due to instructions being deleted etc. In
415 /// the uncached case, this starts out as the set of predecessors we care
417 SmallVector<BasicBlock*, 32> DirtyBlocks;
419 if (!Cache.empty()) {
420 // Okay, we have a cache entry. If we know it is not dirty, just return it
421 // with no computation.
422 if (!CacheP.second) {
427 // If we already have a partially computed set of results, scan them to
428 // determine what is dirty, seeding our initial DirtyBlocks worklist.
429 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
431 if (I->second.isDirty())
432 DirtyBlocks.push_back(I->first);
434 // Sort the cache so that we can do fast binary search lookups below.
435 std::sort(Cache.begin(), Cache.end());
437 ++NumCacheDirtyNonLocal;
438 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
439 // << Cache.size() << " cached: " << *QueryInst;
441 // Seed DirtyBlocks with each of the preds of QueryInst's block.
442 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
443 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
444 DirtyBlocks.push_back(*PI);
445 NumUncacheNonLocal++;
448 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
449 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
451 SmallPtrSet<BasicBlock*, 64> Visited;
453 unsigned NumSortedEntries = Cache.size();
454 DEBUG(AssertSorted(Cache));
456 // Iterate while we still have blocks to update.
457 while (!DirtyBlocks.empty()) {
458 BasicBlock *DirtyBB = DirtyBlocks.back();
459 DirtyBlocks.pop_back();
461 // Already processed this block?
462 if (!Visited.insert(DirtyBB))
465 // Do a binary search to see if we already have an entry for this block in
466 // the cache set. If so, find it.
467 DEBUG(AssertSorted(Cache, NumSortedEntries));
468 NonLocalDepInfo::iterator Entry =
469 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
470 std::make_pair(DirtyBB, MemDepResult()));
471 if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB)
474 MemDepResult *ExistingResult = 0;
475 if (Entry != Cache.begin()+NumSortedEntries &&
476 Entry->first == DirtyBB) {
477 // If we already have an entry, and if it isn't already dirty, the block
479 if (!Entry->second.isDirty())
482 // Otherwise, remember this slot so we can update the value.
483 ExistingResult = &Entry->second;
486 // If the dirty entry has a pointer, start scanning from it so we don't have
487 // to rescan the entire block.
488 BasicBlock::iterator ScanPos = DirtyBB->end();
489 if (ExistingResult) {
490 if (Instruction *Inst = ExistingResult->getInst()) {
492 // We're removing QueryInst's use of Inst.
493 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
494 QueryCS.getInstruction());
498 // Find out if this block has a local dependency for QueryInst.
501 if (ScanPos != DirtyBB->begin()) {
502 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
503 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
504 // No dependence found. If this is the entry block of the function, it is
505 // a clobber, otherwise it is non-local.
506 Dep = MemDepResult::getNonLocal();
508 Dep = MemDepResult::getClobber(ScanPos);
511 // If we had a dirty entry for the block, update it. Otherwise, just add
514 *ExistingResult = Dep;
516 Cache.push_back(std::make_pair(DirtyBB, Dep));
518 // If the block has a dependency (i.e. it isn't completely transparent to
519 // the value), remember the association!
520 if (!Dep.isNonLocal()) {
521 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
522 // update this when we remove instructions.
523 if (Instruction *Inst = Dep.getInst())
524 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
527 // If the block *is* completely transparent to the load, we need to check
528 // the predecessors of this block. Add them to our worklist.
529 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
530 DirtyBlocks.push_back(*PI);
537 /// getNonLocalPointerDependency - Perform a full dependency query for an
538 /// access to the specified (non-volatile) memory location, returning the
539 /// set of instructions that either define or clobber the value.
541 /// This method assumes the pointer has a "NonLocal" dependency within its
544 void MemoryDependenceAnalysis::
545 getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
546 SmallVectorImpl<NonLocalDepEntry> &Result) {
547 assert(isa<PointerType>(Pointer->getType()) &&
548 "Can't get pointer deps of a non-pointer!");
551 // We know that the pointer value is live into FromBB find the def/clobbers
552 // from presecessors.
553 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
554 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy);
556 // This is the set of blocks we've inspected, and the pointer we consider in
557 // each block. Because of critical edges, we currently bail out if querying
558 // a block with multiple different pointers. This can happen during PHI
560 DenseMap<BasicBlock*, Value*> Visited;
561 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB,
562 Result, Visited, true))
565 Result.push_back(std::make_pair(FromBB,
566 MemDepResult::getClobber(FromBB->begin())));
569 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
570 /// Pointer/PointeeSize using either cached information in Cache or by doing a
571 /// lookup (which may use dirty cache info if available). If we do a lookup,
572 /// add the result to the cache.
573 MemDepResult MemoryDependenceAnalysis::
574 GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
575 bool isLoad, BasicBlock *BB,
576 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
578 // Do a binary search to see if we already have an entry for this block in
579 // the cache set. If so, find it.
580 NonLocalDepInfo::iterator Entry =
581 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
582 std::make_pair(BB, MemDepResult()));
583 if (Entry != Cache->begin() && prior(Entry)->first == BB)
586 MemDepResult *ExistingResult = 0;
587 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
588 ExistingResult = &Entry->second;
590 // If we have a cached entry, and it is non-dirty, use it as the value for
592 if (ExistingResult && !ExistingResult->isDirty()) {
593 ++NumCacheNonLocalPtr;
594 return *ExistingResult;
597 // Otherwise, we have to scan for the value. If we have a dirty cache
598 // entry, start scanning from its position, otherwise we scan from the end
600 BasicBlock::iterator ScanPos = BB->end();
601 if (ExistingResult && ExistingResult->getInst()) {
602 assert(ExistingResult->getInst()->getParent() == BB &&
603 "Instruction invalidated?");
604 ++NumCacheDirtyNonLocalPtr;
605 ScanPos = ExistingResult->getInst();
607 // Eliminating the dirty entry from 'Cache', so update the reverse info.
608 ValueIsLoadPair CacheKey(Pointer, isLoad);
609 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
611 ++NumUncacheNonLocalPtr;
614 // Scan the block for the dependency.
615 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad,
618 // If we had a dirty entry for the block, update it. Otherwise, just add
621 *ExistingResult = Dep;
623 Cache->push_back(std::make_pair(BB, Dep));
625 // If the block has a dependency (i.e. it isn't completely transparent to
626 // the value), remember the reverse association because we just added it
628 if (Dep.isNonLocal())
631 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
632 // update MemDep when we remove instructions.
633 Instruction *Inst = Dep.getInst();
634 assert(Inst && "Didn't depend on anything?");
635 ValueIsLoadPair CacheKey(Pointer, isLoad);
636 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
640 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
641 /// number of elements in the array that are already properly ordered. This is
642 /// optimized for the case when only a few entries are added.
644 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
645 unsigned NumSortedEntries) {
646 switch (Cache.size() - NumSortedEntries) {
648 // done, no new entries.
651 // Two new entries, insert the last one into place.
652 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
654 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
655 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
656 Cache.insert(Entry, Val);
660 // One new entry, Just insert the new value at the appropriate position.
661 if (Cache.size() != 1) {
662 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
664 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
665 std::upper_bound(Cache.begin(), Cache.end(), Val);
666 Cache.insert(Entry, Val);
670 // Added many values, do a full scale sort.
671 std::sort(Cache.begin(), Cache.end());
677 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
678 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
679 /// results to the results vector and keep track of which blocks are visited in
682 /// This has special behavior for the first block queries (when SkipFirstBlock
683 /// is true). In this special case, it ignores the contents of the specified
684 /// block and starts returning dependence info for its predecessors.
686 /// This function returns false on success, or true to indicate that it could
687 /// not compute dependence information for some reason. This should be treated
688 /// as a clobber dependence on the first instruction in the predecessor block.
689 bool MemoryDependenceAnalysis::
690 getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
691 bool isLoad, BasicBlock *StartBB,
692 SmallVectorImpl<NonLocalDepEntry> &Result,
693 DenseMap<BasicBlock*, Value*> &Visited,
694 bool SkipFirstBlock) {
696 // Look up the cached info for Pointer.
697 ValueIsLoadPair CacheKey(Pointer, isLoad);
699 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
700 &NonLocalPointerDeps[CacheKey];
701 NonLocalDepInfo *Cache = &CacheInfo->second;
703 // If we have valid cached information for exactly the block we are
704 // investigating, just return it with no recomputation.
705 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
706 // We have a fully cached result for this query then we can just return the
707 // cached results and populate the visited set. However, we have to verify
708 // that we don't already have conflicting results for these blocks. Check
709 // to ensure that if a block in the results set is in the visited set that
710 // it was for the same pointer query.
711 if (!Visited.empty()) {
712 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
714 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->first);
715 if (VI == Visited.end() || VI->second == Pointer) continue;
717 // We have a pointer mismatch in a block. Just return clobber, saying
718 // that something was clobbered in this result. We could also do a
719 // non-fully cached query, but there is little point in doing this.
724 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
726 Visited.insert(std::make_pair(I->first, Pointer));
727 if (!I->second.isNonLocal())
728 Result.push_back(*I);
730 ++NumCacheCompleteNonLocalPtr;
734 // Otherwise, either this is a new block, a block with an invalid cache
735 // pointer or one that we're about to invalidate by putting more info into it
736 // than its valid cache info. If empty, the result will be valid cache info,
737 // otherwise it isn't.
739 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
741 CacheInfo->first = BBSkipFirstBlockPair();
743 SmallVector<BasicBlock*, 32> Worklist;
744 Worklist.push_back(StartBB);
746 // Keep track of the entries that we know are sorted. Previously cached
747 // entries will all be sorted. The entries we add we only sort on demand (we
748 // don't insert every element into its sorted position). We know that we
749 // won't get any reuse from currently inserted values, because we don't
750 // revisit blocks after we insert info for them.
751 unsigned NumSortedEntries = Cache->size();
752 DEBUG(AssertSorted(*Cache));
754 while (!Worklist.empty()) {
755 BasicBlock *BB = Worklist.pop_back_val();
757 // Skip the first block if we have it.
758 if (!SkipFirstBlock) {
759 // Analyze the dependency of *Pointer in FromBB. See if we already have
761 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
763 // Get the dependency info for Pointer in BB. If we have cached
764 // information, we will use it, otherwise we compute it.
765 DEBUG(AssertSorted(*Cache, NumSortedEntries));
766 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad,
767 BB, Cache, NumSortedEntries);
769 // If we got a Def or Clobber, add this to the list of results.
770 if (!Dep.isNonLocal()) {
771 Result.push_back(NonLocalDepEntry(BB, Dep));
776 // If 'Pointer' is an instruction defined in this block, then we need to do
777 // phi translation to change it into a value live in the predecessor block.
778 // If phi translation fails, then we can't continue dependence analysis.
779 Instruction *PtrInst = dyn_cast<Instruction>(Pointer);
780 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB;
782 // If no PHI translation is needed, just add all the predecessors of this
783 // block to scan them as well.
784 if (!NeedsPHITranslation) {
785 SkipFirstBlock = false;
786 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
787 // Verify that we haven't looked at this block yet.
788 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
789 InsertRes = Visited.insert(std::make_pair(*PI, Pointer));
790 if (InsertRes.second) {
791 // First time we've looked at *PI.
792 Worklist.push_back(*PI);
796 // If we have seen this block before, but it was with a different
797 // pointer then we have a phi translation failure and we have to treat
798 // this as a clobber.
799 if (InsertRes.first->second != Pointer)
800 goto PredTranslationFailure;
805 // If we do need to do phi translation, then there are a bunch of different
806 // cases, because we have to find a Value* live in the predecessor block. We
807 // know that PtrInst is defined in this block at least.
809 // We may have added values to the cache list before this PHI translation.
810 // If so, we haven't done anything to ensure that the cache remains sorted.
811 // Sort it now (if needed) so that recursive invocations of
812 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
813 // value will only see properly sorted cache arrays.
814 if (Cache && NumSortedEntries != Cache->size()) {
815 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
816 NumSortedEntries = Cache->size();
819 // If this is directly a PHI node, just use the incoming values for each
820 // pred as the phi translated version.
821 if (PHINode *PtrPHI = dyn_cast<PHINode>(PtrInst)) {
824 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
825 BasicBlock *Pred = *PI;
826 Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred);
828 // Check to see if we have already visited this pred block with another
829 // pointer. If so, we can't do this lookup. This failure can occur
830 // with PHI translation when a critical edge exists and the PHI node in
831 // the successor translates to a pointer value different than the
832 // pointer the block was first analyzed with.
833 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
834 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr));
836 if (!InsertRes.second) {
837 // If the predecessor was visited with PredPtr, then we already did
838 // the analysis and can ignore it.
839 if (InsertRes.first->second == PredPtr)
842 // Otherwise, the block was previously analyzed with a different
843 // pointer. We can't represent the result of this case, so we just
844 // treat this as a phi translation failure.
845 goto PredTranslationFailure;
848 // FIXME: it is entirely possible that PHI translating will end up with
849 // the same value. Consider PHI translating something like:
850 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
851 // to recurse here, pedantically speaking.
853 // If we have a problem phi translating, fall through to the code below
854 // to handle the failure condition.
855 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred,
857 goto PredTranslationFailure;
860 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
861 CacheInfo = &NonLocalPointerDeps[CacheKey];
862 Cache = &CacheInfo->second;
863 NumSortedEntries = Cache->size();
865 // Since we did phi translation, the "Cache" set won't contain all of the
866 // results for the query. This is ok (we can still use it to accelerate
867 // specific block queries) but we can't do the fastpath "return all
868 // results from the set" Clear out the indicator for this.
869 CacheInfo->first = BBSkipFirstBlockPair();
870 SkipFirstBlock = false;
874 // TODO: BITCAST, GEP.
876 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
877 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
878 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
879 PredTranslationFailure:
882 // Refresh the CacheInfo/Cache pointer if it got invalidated.
883 CacheInfo = &NonLocalPointerDeps[CacheKey];
884 Cache = &CacheInfo->second;
885 NumSortedEntries = Cache->size();
888 // Since we did phi translation, the "Cache" set won't contain all of the
889 // results for the query. This is ok (we can still use it to accelerate
890 // specific block queries) but we can't do the fastpath "return all
891 // results from the set" Clear out the indicator for this.
892 CacheInfo->first = BBSkipFirstBlockPair();
894 // If *nothing* works, mark the pointer as being clobbered by the first
895 // instruction in this block.
897 // If this is the magic first block, return this as a clobber of the whole
898 // incoming value. Since we can't phi translate to one of the predecessors,
899 // we have to bail out.
903 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
904 assert(I != Cache->rend() && "Didn't find current block??");
908 assert(I->second.isNonLocal() &&
909 "Should only be here with transparent block");
910 I->second = MemDepResult::getClobber(BB->begin());
911 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
912 Result.push_back(*I);
917 // Okay, we're done now. If we added new values to the cache, re-sort it.
918 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
919 DEBUG(AssertSorted(*Cache));
923 /// RemoveCachedNonLocalPointerDependencies - If P exists in
924 /// CachedNonLocalPointerInfo, remove it.
925 void MemoryDependenceAnalysis::
926 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
927 CachedNonLocalPointerInfo::iterator It =
928 NonLocalPointerDeps.find(P);
929 if (It == NonLocalPointerDeps.end()) return;
931 // Remove all of the entries in the BB->val map. This involves removing
932 // instructions from the reverse map.
933 NonLocalDepInfo &PInfo = It->second.second;
935 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
936 Instruction *Target = PInfo[i].second.getInst();
937 if (Target == 0) continue; // Ignore non-local dep results.
938 assert(Target->getParent() == PInfo[i].first);
940 // Eliminating the dirty entry from 'Cache', so update the reverse info.
941 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
944 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
945 NonLocalPointerDeps.erase(It);
949 /// invalidateCachedPointerInfo - This method is used to invalidate cached
950 /// information about the specified pointer, because it may be too
951 /// conservative in memdep. This is an optional call that can be used when
952 /// the client detects an equivalence between the pointer and some other
953 /// value and replaces the other value with ptr. This can make Ptr available
954 /// in more places that cached info does not necessarily keep.
955 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
956 // If Ptr isn't really a pointer, just ignore it.
957 if (!isa<PointerType>(Ptr->getType())) return;
958 // Flush store info for the pointer.
959 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
960 // Flush load info for the pointer.
961 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
964 /// removeInstruction - Remove an instruction from the dependence analysis,
965 /// updating the dependence of instructions that previously depended on it.
966 /// This method attempts to keep the cache coherent using the reverse map.
967 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
968 // Walk through the Non-local dependencies, removing this one as the value
969 // for any cached queries.
970 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
971 if (NLDI != NonLocalDeps.end()) {
972 NonLocalDepInfo &BlockMap = NLDI->second.first;
973 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
975 if (Instruction *Inst = DI->second.getInst())
976 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
977 NonLocalDeps.erase(NLDI);
980 // If we have a cached local dependence query for this instruction, remove it.
982 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
983 if (LocalDepEntry != LocalDeps.end()) {
984 // Remove us from DepInst's reverse set now that the local dep info is gone.
985 if (Instruction *Inst = LocalDepEntry->second.getInst())
986 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
988 // Remove this local dependency info.
989 LocalDeps.erase(LocalDepEntry);
992 // If we have any cached pointer dependencies on this instruction, remove
993 // them. If the instruction has non-pointer type, then it can't be a pointer
996 // Remove it from both the load info and the store info. The instruction
997 // can't be in either of these maps if it is non-pointer.
998 if (isa<PointerType>(RemInst->getType())) {
999 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1000 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1003 // Loop over all of the things that depend on the instruction we're removing.
1005 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1007 // If we find RemInst as a clobber or Def in any of the maps for other values,
1008 // we need to replace its entry with a dirty version of the instruction after
1009 // it. If RemInst is a terminator, we use a null dirty value.
1011 // Using a dirty version of the instruction after RemInst saves having to scan
1012 // the entire block to get to this point.
1013 MemDepResult NewDirtyVal;
1014 if (!RemInst->isTerminator())
1015 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1017 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1018 if (ReverseDepIt != ReverseLocalDeps.end()) {
1019 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1020 // RemInst can't be the terminator if it has local stuff depending on it.
1021 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1022 "Nothing can locally depend on a terminator");
1024 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1025 E = ReverseDeps.end(); I != E; ++I) {
1026 Instruction *InstDependingOnRemInst = *I;
1027 assert(InstDependingOnRemInst != RemInst &&
1028 "Already removed our local dep info");
1030 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1032 // Make sure to remember that new things depend on NewDepInst.
1033 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1034 "a local dep on this if it is a terminator!");
1035 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1036 InstDependingOnRemInst));
1039 ReverseLocalDeps.erase(ReverseDepIt);
1041 // Add new reverse deps after scanning the set, to avoid invalidating the
1042 // 'ReverseDeps' reference.
1043 while (!ReverseDepsToAdd.empty()) {
1044 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1045 .insert(ReverseDepsToAdd.back().second);
1046 ReverseDepsToAdd.pop_back();
1050 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1051 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1052 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1053 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1055 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1057 PerInstNLInfo &INLD = NonLocalDeps[*I];
1058 // The information is now dirty!
1061 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1062 DE = INLD.first.end(); DI != DE; ++DI) {
1063 if (DI->second.getInst() != RemInst) continue;
1065 // Convert to a dirty entry for the subsequent instruction.
1066 DI->second = NewDirtyVal;
1068 if (Instruction *NextI = NewDirtyVal.getInst())
1069 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1073 ReverseNonLocalDeps.erase(ReverseDepIt);
1075 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1076 while (!ReverseDepsToAdd.empty()) {
1077 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1078 .insert(ReverseDepsToAdd.back().second);
1079 ReverseDepsToAdd.pop_back();
1083 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1084 // value in the NonLocalPointerDeps info.
1085 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1086 ReverseNonLocalPtrDeps.find(RemInst);
1087 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1088 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1089 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1091 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1092 E = Set.end(); I != E; ++I) {
1093 ValueIsLoadPair P = *I;
1094 assert(P.getPointer() != RemInst &&
1095 "Already removed NonLocalPointerDeps info for RemInst");
1097 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second;
1099 // The cache is not valid for any specific block anymore.
1100 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair();
1102 // Update any entries for RemInst to use the instruction after it.
1103 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1105 if (DI->second.getInst() != RemInst) continue;
1107 // Convert to a dirty entry for the subsequent instruction.
1108 DI->second = NewDirtyVal;
1110 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1111 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1114 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1115 // subsequent value may invalidate the sortedness.
1116 std::sort(NLPDI.begin(), NLPDI.end());
1119 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1121 while (!ReversePtrDepsToAdd.empty()) {
1122 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1123 .insert(ReversePtrDepsToAdd.back().second);
1124 ReversePtrDepsToAdd.pop_back();
1129 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1130 AA->deleteValue(RemInst);
1131 DEBUG(verifyRemoved(RemInst));
1133 /// verifyRemoved - Verify that the specified instruction does not occur
1134 /// in our internal data structures.
1135 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1136 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1137 E = LocalDeps.end(); I != E; ++I) {
1138 assert(I->first != D && "Inst occurs in data structures");
1139 assert(I->second.getInst() != D &&
1140 "Inst occurs in data structures");
1143 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1144 E = NonLocalPointerDeps.end(); I != E; ++I) {
1145 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1146 const NonLocalDepInfo &Val = I->second.second;
1147 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1149 assert(II->second.getInst() != D && "Inst occurs as NLPD value");
1152 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1153 E = NonLocalDeps.end(); I != E; ++I) {
1154 assert(I->first != D && "Inst occurs in data structures");
1155 const PerInstNLInfo &INLD = I->second;
1156 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1157 EE = INLD.first.end(); II != EE; ++II)
1158 assert(II->second.getInst() != D && "Inst occurs in data structures");
1161 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1162 E = ReverseLocalDeps.end(); I != E; ++I) {
1163 assert(I->first != D && "Inst occurs in data structures");
1164 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1165 EE = I->second.end(); II != EE; ++II)
1166 assert(*II != D && "Inst occurs in data structures");
1169 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1170 E = ReverseNonLocalDeps.end();
1172 assert(I->first != D && "Inst occurs in data structures");
1173 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1174 EE = I->second.end(); II != EE; ++II)
1175 assert(*II != D && "Inst occurs in data structures");
1178 for (ReverseNonLocalPtrDepTy::const_iterator
1179 I = ReverseNonLocalPtrDeps.begin(),
1180 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1181 assert(I->first != D && "Inst occurs in rev NLPD map");
1183 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1184 E = I->second.end(); II != E; ++II)
1185 assert(*II != ValueIsLoadPair(D, false) &&
1186 *II != ValueIsLoadPair(D, true) &&
1187 "Inst occurs in ReverseNonLocalPtrDeps map");