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 (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
234 // There can't be stores to the value we care about inside an
236 if (invariantTag) continue;
238 // If alias analysis can tell that this store is guaranteed to not modify
239 // the query pointer, ignore it. Use getModRefInfo to handle cases where
240 // the query pointer points to constant memory etc.
241 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef)
244 // Ok, this store might clobber the query pointer. Check to see if it is
245 // a must alias: in this case, we want to return this as a def.
246 Value *Pointer = SI->getPointerOperand();
247 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
249 // If we found a pointer, check if it could be the same as our pointer.
250 AliasAnalysis::AliasResult R =
251 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
253 if (R == AliasAnalysis::NoAlias)
255 if (R == AliasAnalysis::MayAlias)
256 return MemDepResult::getClobber(Inst);
257 return MemDepResult::getDef(Inst);
260 // If this is an allocation, and if we know that the accessed pointer is to
261 // the allocation, return Def. This means that there is no dependence and
262 // the access can be optimized based on that. For example, a load could
264 // Note: Only determine this to be a malloc if Inst is the malloc call, not
265 // a subsequent bitcast of the malloc call result. There can be stores to
266 // the malloced memory between the malloc call and its bitcast uses, and we
267 // need to continue scanning until the malloc call.
268 if (isa<AllocaInst>(Inst) || extractMallocCall(Inst)) {
269 Value *AccessPtr = MemPtr->getUnderlyingObject();
271 if (AccessPtr == Inst ||
272 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
273 return MemDepResult::getDef(Inst);
277 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
278 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
279 case AliasAnalysis::NoModRef:
280 // If the call has no effect on the queried pointer, just ignore it.
282 case AliasAnalysis::Mod:
283 // If we're in an invariant region, we can ignore calls that ONLY
284 // modify the pointer.
285 if (invariantTag) continue;
286 return MemDepResult::getClobber(Inst);
287 case AliasAnalysis::Ref:
288 // If the call is known to never store to the pointer, and if this is a
289 // load query, we can safely ignore it (scan past it).
293 // Otherwise, there is a potential dependence. Return a clobber.
294 return MemDepResult::getClobber(Inst);
298 // No dependence found. If this is the entry block of the function, it is a
299 // clobber, otherwise it is non-local.
300 if (BB != &BB->getParent()->getEntryBlock())
301 return MemDepResult::getNonLocal();
302 return MemDepResult::getClobber(ScanIt);
305 /// getDependency - Return the instruction on which a memory operation
307 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
308 Instruction *ScanPos = QueryInst;
310 // Check for a cached result
311 MemDepResult &LocalCache = LocalDeps[QueryInst];
313 // If the cached entry is non-dirty, just return it. Note that this depends
314 // on MemDepResult's default constructing to 'dirty'.
315 if (!LocalCache.isDirty())
318 // Otherwise, if we have a dirty entry, we know we can start the scan at that
319 // instruction, which may save us some work.
320 if (Instruction *Inst = LocalCache.getInst()) {
323 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
326 BasicBlock *QueryParent = QueryInst->getParent();
329 uint64_t MemSize = 0;
332 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
333 // No dependence found. If this is the entry block of the function, it is a
334 // clobber, otherwise it is non-local.
335 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
336 LocalCache = MemDepResult::getNonLocal();
338 LocalCache = MemDepResult::getClobber(QueryInst);
339 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
340 // If this is a volatile store, don't mess around with it. Just return the
341 // previous instruction as a clobber.
342 if (SI->isVolatile())
343 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
345 MemPtr = SI->getPointerOperand();
346 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
348 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
349 // If this is a volatile load, don't mess around with it. Just return the
350 // previous instruction as a clobber.
351 if (LI->isVolatile())
352 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
354 MemPtr = LI->getPointerOperand();
355 MemSize = AA->getTypeStoreSize(LI->getType());
357 } else if (isFreeCall(QueryInst)) {
358 MemPtr = QueryInst->getOperand(1);
359 // calls to free() erase the entire structure, not just a field.
361 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
362 CallSite QueryCS = CallSite::get(QueryInst);
363 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
364 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
367 // Non-memory instruction.
368 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
371 // If we need to do a pointer scan, make it happen.
373 LocalCache = getPointerDependencyFrom(MemPtr, MemSize,
374 isa<LoadInst>(QueryInst),
375 ScanPos, QueryParent);
377 // Remember the result!
378 if (Instruction *I = LocalCache.getInst())
379 ReverseLocalDeps[I].insert(QueryInst);
385 /// AssertSorted - This method is used when -debug is specified to verify that
386 /// cache arrays are properly kept sorted.
387 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
389 if (Count == -1) Count = Cache.size();
390 if (Count == 0) return;
392 for (unsigned i = 1; i != unsigned(Count); ++i)
393 assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!");
397 /// getNonLocalCallDependency - Perform a full dependency query for the
398 /// specified call, returning the set of blocks that the value is
399 /// potentially live across. The returned set of results will include a
400 /// "NonLocal" result for all blocks where the value is live across.
402 /// This method assumes the instruction returns a "NonLocal" dependency
403 /// within its own block.
405 /// This returns a reference to an internal data structure that may be
406 /// invalidated on the next non-local query or when an instruction is
407 /// removed. Clients must copy this data if they want it around longer than
409 const MemoryDependenceAnalysis::NonLocalDepInfo &
410 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
411 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
412 "getNonLocalCallDependency should only be used on calls with non-local deps!");
413 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
414 NonLocalDepInfo &Cache = CacheP.first;
416 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
417 /// the cached case, this can happen due to instructions being deleted etc. In
418 /// the uncached case, this starts out as the set of predecessors we care
420 SmallVector<BasicBlock*, 32> DirtyBlocks;
422 if (!Cache.empty()) {
423 // Okay, we have a cache entry. If we know it is not dirty, just return it
424 // with no computation.
425 if (!CacheP.second) {
430 // If we already have a partially computed set of results, scan them to
431 // determine what is dirty, seeding our initial DirtyBlocks worklist.
432 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
434 if (I->second.isDirty())
435 DirtyBlocks.push_back(I->first);
437 // Sort the cache so that we can do fast binary search lookups below.
438 std::sort(Cache.begin(), Cache.end());
440 ++NumCacheDirtyNonLocal;
441 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
442 // << Cache.size() << " cached: " << *QueryInst;
444 // Seed DirtyBlocks with each of the preds of QueryInst's block.
445 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
446 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
447 DirtyBlocks.push_back(*PI);
448 NumUncacheNonLocal++;
451 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
452 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
454 SmallPtrSet<BasicBlock*, 64> Visited;
456 unsigned NumSortedEntries = Cache.size();
457 DEBUG(AssertSorted(Cache));
459 // Iterate while we still have blocks to update.
460 while (!DirtyBlocks.empty()) {
461 BasicBlock *DirtyBB = DirtyBlocks.back();
462 DirtyBlocks.pop_back();
464 // Already processed this block?
465 if (!Visited.insert(DirtyBB))
468 // Do a binary search to see if we already have an entry for this block in
469 // the cache set. If so, find it.
470 DEBUG(AssertSorted(Cache, NumSortedEntries));
471 NonLocalDepInfo::iterator Entry =
472 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
473 std::make_pair(DirtyBB, MemDepResult()));
474 if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB)
477 MemDepResult *ExistingResult = 0;
478 if (Entry != Cache.begin()+NumSortedEntries &&
479 Entry->first == DirtyBB) {
480 // If we already have an entry, and if it isn't already dirty, the block
482 if (!Entry->second.isDirty())
485 // Otherwise, remember this slot so we can update the value.
486 ExistingResult = &Entry->second;
489 // If the dirty entry has a pointer, start scanning from it so we don't have
490 // to rescan the entire block.
491 BasicBlock::iterator ScanPos = DirtyBB->end();
492 if (ExistingResult) {
493 if (Instruction *Inst = ExistingResult->getInst()) {
495 // We're removing QueryInst's use of Inst.
496 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
497 QueryCS.getInstruction());
501 // Find out if this block has a local dependency for QueryInst.
504 if (ScanPos != DirtyBB->begin()) {
505 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
506 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
507 // No dependence found. If this is the entry block of the function, it is
508 // a clobber, otherwise it is non-local.
509 Dep = MemDepResult::getNonLocal();
511 Dep = MemDepResult::getClobber(ScanPos);
514 // If we had a dirty entry for the block, update it. Otherwise, just add
517 *ExistingResult = Dep;
519 Cache.push_back(std::make_pair(DirtyBB, Dep));
521 // If the block has a dependency (i.e. it isn't completely transparent to
522 // the value), remember the association!
523 if (!Dep.isNonLocal()) {
524 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
525 // update this when we remove instructions.
526 if (Instruction *Inst = Dep.getInst())
527 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
530 // If the block *is* completely transparent to the load, we need to check
531 // the predecessors of this block. Add them to our worklist.
532 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
533 DirtyBlocks.push_back(*PI);
540 /// getNonLocalPointerDependency - Perform a full dependency query for an
541 /// access to the specified (non-volatile) memory location, returning the
542 /// set of instructions that either define or clobber the value.
544 /// This method assumes the pointer has a "NonLocal" dependency within its
547 void MemoryDependenceAnalysis::
548 getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
549 SmallVectorImpl<NonLocalDepEntry> &Result) {
550 assert(isa<PointerType>(Pointer->getType()) &&
551 "Can't get pointer deps of a non-pointer!");
554 // We know that the pointer value is live into FromBB find the def/clobbers
555 // from presecessors.
556 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
557 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy);
559 // This is the set of blocks we've inspected, and the pointer we consider in
560 // each block. Because of critical edges, we currently bail out if querying
561 // a block with multiple different pointers. This can happen during PHI
563 DenseMap<BasicBlock*, Value*> Visited;
564 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB,
565 Result, Visited, true))
568 Result.push_back(std::make_pair(FromBB,
569 MemDepResult::getClobber(FromBB->begin())));
572 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
573 /// Pointer/PointeeSize using either cached information in Cache or by doing a
574 /// lookup (which may use dirty cache info if available). If we do a lookup,
575 /// add the result to the cache.
576 MemDepResult MemoryDependenceAnalysis::
577 GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
578 bool isLoad, BasicBlock *BB,
579 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
581 // Do a binary search to see if we already have an entry for this block in
582 // the cache set. If so, find it.
583 NonLocalDepInfo::iterator Entry =
584 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
585 std::make_pair(BB, MemDepResult()));
586 if (Entry != Cache->begin() && prior(Entry)->first == BB)
589 MemDepResult *ExistingResult = 0;
590 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
591 ExistingResult = &Entry->second;
593 // If we have a cached entry, and it is non-dirty, use it as the value for
595 if (ExistingResult && !ExistingResult->isDirty()) {
596 ++NumCacheNonLocalPtr;
597 return *ExistingResult;
600 // Otherwise, we have to scan for the value. If we have a dirty cache
601 // entry, start scanning from its position, otherwise we scan from the end
603 BasicBlock::iterator ScanPos = BB->end();
604 if (ExistingResult && ExistingResult->getInst()) {
605 assert(ExistingResult->getInst()->getParent() == BB &&
606 "Instruction invalidated?");
607 ++NumCacheDirtyNonLocalPtr;
608 ScanPos = ExistingResult->getInst();
610 // Eliminating the dirty entry from 'Cache', so update the reverse info.
611 ValueIsLoadPair CacheKey(Pointer, isLoad);
612 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
614 ++NumUncacheNonLocalPtr;
617 // Scan the block for the dependency.
618 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad,
621 // If we had a dirty entry for the block, update it. Otherwise, just add
624 *ExistingResult = Dep;
626 Cache->push_back(std::make_pair(BB, Dep));
628 // If the block has a dependency (i.e. it isn't completely transparent to
629 // the value), remember the reverse association because we just added it
631 if (Dep.isNonLocal())
634 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
635 // update MemDep when we remove instructions.
636 Instruction *Inst = Dep.getInst();
637 assert(Inst && "Didn't depend on anything?");
638 ValueIsLoadPair CacheKey(Pointer, isLoad);
639 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
643 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
644 /// number of elements in the array that are already properly ordered. This is
645 /// optimized for the case when only a few entries are added.
647 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
648 unsigned NumSortedEntries) {
649 switch (Cache.size() - NumSortedEntries) {
651 // done, no new entries.
654 // Two new entries, insert the last one into place.
655 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
657 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
658 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
659 Cache.insert(Entry, Val);
663 // One new entry, Just insert the new value at the appropriate position.
664 if (Cache.size() != 1) {
665 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
667 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
668 std::upper_bound(Cache.begin(), Cache.end(), Val);
669 Cache.insert(Entry, Val);
673 // Added many values, do a full scale sort.
674 std::sort(Cache.begin(), Cache.end());
680 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
681 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
682 /// results to the results vector and keep track of which blocks are visited in
685 /// This has special behavior for the first block queries (when SkipFirstBlock
686 /// is true). In this special case, it ignores the contents of the specified
687 /// block and starts returning dependence info for its predecessors.
689 /// This function returns false on success, or true to indicate that it could
690 /// not compute dependence information for some reason. This should be treated
691 /// as a clobber dependence on the first instruction in the predecessor block.
692 bool MemoryDependenceAnalysis::
693 getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
694 bool isLoad, BasicBlock *StartBB,
695 SmallVectorImpl<NonLocalDepEntry> &Result,
696 DenseMap<BasicBlock*, Value*> &Visited,
697 bool SkipFirstBlock) {
699 // Look up the cached info for Pointer.
700 ValueIsLoadPair CacheKey(Pointer, isLoad);
702 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
703 &NonLocalPointerDeps[CacheKey];
704 NonLocalDepInfo *Cache = &CacheInfo->second;
706 // If we have valid cached information for exactly the block we are
707 // investigating, just return it with no recomputation.
708 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
709 // We have a fully cached result for this query then we can just return the
710 // cached results and populate the visited set. However, we have to verify
711 // that we don't already have conflicting results for these blocks. Check
712 // to ensure that if a block in the results set is in the visited set that
713 // it was for the same pointer query.
714 if (!Visited.empty()) {
715 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
717 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->first);
718 if (VI == Visited.end() || VI->second == Pointer) continue;
720 // We have a pointer mismatch in a block. Just return clobber, saying
721 // that something was clobbered in this result. We could also do a
722 // non-fully cached query, but there is little point in doing this.
727 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
729 Visited.insert(std::make_pair(I->first, Pointer));
730 if (!I->second.isNonLocal())
731 Result.push_back(*I);
733 ++NumCacheCompleteNonLocalPtr;
737 // Otherwise, either this is a new block, a block with an invalid cache
738 // pointer or one that we're about to invalidate by putting more info into it
739 // than its valid cache info. If empty, the result will be valid cache info,
740 // otherwise it isn't.
742 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
744 CacheInfo->first = BBSkipFirstBlockPair();
746 SmallVector<BasicBlock*, 32> Worklist;
747 Worklist.push_back(StartBB);
749 // Keep track of the entries that we know are sorted. Previously cached
750 // entries will all be sorted. The entries we add we only sort on demand (we
751 // don't insert every element into its sorted position). We know that we
752 // won't get any reuse from currently inserted values, because we don't
753 // revisit blocks after we insert info for them.
754 unsigned NumSortedEntries = Cache->size();
755 DEBUG(AssertSorted(*Cache));
757 while (!Worklist.empty()) {
758 BasicBlock *BB = Worklist.pop_back_val();
760 // Skip the first block if we have it.
761 if (!SkipFirstBlock) {
762 // Analyze the dependency of *Pointer in FromBB. See if we already have
764 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
766 // Get the dependency info for Pointer in BB. If we have cached
767 // information, we will use it, otherwise we compute it.
768 DEBUG(AssertSorted(*Cache, NumSortedEntries));
769 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad,
770 BB, Cache, NumSortedEntries);
772 // If we got a Def or Clobber, add this to the list of results.
773 if (!Dep.isNonLocal()) {
774 Result.push_back(NonLocalDepEntry(BB, Dep));
779 // If 'Pointer' is an instruction defined in this block, then we need to do
780 // phi translation to change it into a value live in the predecessor block.
781 // If phi translation fails, then we can't continue dependence analysis.
782 Instruction *PtrInst = dyn_cast<Instruction>(Pointer);
783 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB;
785 // If no PHI translation is needed, just add all the predecessors of this
786 // block to scan them as well.
787 if (!NeedsPHITranslation) {
788 SkipFirstBlock = false;
789 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
790 // Verify that we haven't looked at this block yet.
791 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
792 InsertRes = Visited.insert(std::make_pair(*PI, Pointer));
793 if (InsertRes.second) {
794 // First time we've looked at *PI.
795 Worklist.push_back(*PI);
799 // If we have seen this block before, but it was with a different
800 // pointer then we have a phi translation failure and we have to treat
801 // this as a clobber.
802 if (InsertRes.first->second != Pointer)
803 goto PredTranslationFailure;
808 // If we do need to do phi translation, then there are a bunch of different
809 // cases, because we have to find a Value* live in the predecessor block. We
810 // know that PtrInst is defined in this block at least.
812 // We may have added values to the cache list before this PHI translation.
813 // If so, we haven't done anything to ensure that the cache remains sorted.
814 // Sort it now (if needed) so that recursive invocations of
815 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
816 // value will only see properly sorted cache arrays.
817 if (Cache && NumSortedEntries != Cache->size()) {
818 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
819 NumSortedEntries = Cache->size();
822 // If this is directly a PHI node, just use the incoming values for each
823 // pred as the phi translated version.
824 if (PHINode *PtrPHI = dyn_cast<PHINode>(PtrInst)) {
827 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
828 BasicBlock *Pred = *PI;
829 Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred);
831 // Check to see if we have already visited this pred block with another
832 // pointer. If so, we can't do this lookup. This failure can occur
833 // with PHI translation when a critical edge exists and the PHI node in
834 // the successor translates to a pointer value different than the
835 // pointer the block was first analyzed with.
836 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
837 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr));
839 if (!InsertRes.second) {
840 // If the predecessor was visited with PredPtr, then we already did
841 // the analysis and can ignore it.
842 if (InsertRes.first->second == PredPtr)
845 // Otherwise, the block was previously analyzed with a different
846 // pointer. We can't represent the result of this case, so we just
847 // treat this as a phi translation failure.
848 goto PredTranslationFailure;
851 // FIXME: it is entirely possible that PHI translating will end up with
852 // the same value. Consider PHI translating something like:
853 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
854 // to recurse here, pedantically speaking.
856 // If we have a problem phi translating, fall through to the code below
857 // to handle the failure condition.
858 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred,
860 goto PredTranslationFailure;
863 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
864 CacheInfo = &NonLocalPointerDeps[CacheKey];
865 Cache = &CacheInfo->second;
866 NumSortedEntries = Cache->size();
868 // Since we did phi translation, the "Cache" set won't contain all of the
869 // results for the query. This is ok (we can still use it to accelerate
870 // specific block queries) but we can't do the fastpath "return all
871 // results from the set" Clear out the indicator for this.
872 CacheInfo->first = BBSkipFirstBlockPair();
873 SkipFirstBlock = false;
877 // TODO: BITCAST, GEP.
879 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
880 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
881 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
882 PredTranslationFailure:
885 // Refresh the CacheInfo/Cache pointer if it got invalidated.
886 CacheInfo = &NonLocalPointerDeps[CacheKey];
887 Cache = &CacheInfo->second;
888 NumSortedEntries = Cache->size();
891 // Since we did phi translation, the "Cache" set won't contain all of the
892 // results for the query. This is ok (we can still use it to accelerate
893 // specific block queries) but we can't do the fastpath "return all
894 // results from the set" Clear out the indicator for this.
895 CacheInfo->first = BBSkipFirstBlockPair();
897 // If *nothing* works, mark the pointer as being clobbered by the first
898 // instruction in this block.
900 // If this is the magic first block, return this as a clobber of the whole
901 // incoming value. Since we can't phi translate to one of the predecessors,
902 // we have to bail out.
906 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
907 assert(I != Cache->rend() && "Didn't find current block??");
911 assert(I->second.isNonLocal() &&
912 "Should only be here with transparent block");
913 I->second = MemDepResult::getClobber(BB->begin());
914 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
915 Result.push_back(*I);
920 // Okay, we're done now. If we added new values to the cache, re-sort it.
921 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
922 DEBUG(AssertSorted(*Cache));
926 /// RemoveCachedNonLocalPointerDependencies - If P exists in
927 /// CachedNonLocalPointerInfo, remove it.
928 void MemoryDependenceAnalysis::
929 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
930 CachedNonLocalPointerInfo::iterator It =
931 NonLocalPointerDeps.find(P);
932 if (It == NonLocalPointerDeps.end()) return;
934 // Remove all of the entries in the BB->val map. This involves removing
935 // instructions from the reverse map.
936 NonLocalDepInfo &PInfo = It->second.second;
938 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
939 Instruction *Target = PInfo[i].second.getInst();
940 if (Target == 0) continue; // Ignore non-local dep results.
941 assert(Target->getParent() == PInfo[i].first);
943 // Eliminating the dirty entry from 'Cache', so update the reverse info.
944 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
947 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
948 NonLocalPointerDeps.erase(It);
952 /// invalidateCachedPointerInfo - This method is used to invalidate cached
953 /// information about the specified pointer, because it may be too
954 /// conservative in memdep. This is an optional call that can be used when
955 /// the client detects an equivalence between the pointer and some other
956 /// value and replaces the other value with ptr. This can make Ptr available
957 /// in more places that cached info does not necessarily keep.
958 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
959 // If Ptr isn't really a pointer, just ignore it.
960 if (!isa<PointerType>(Ptr->getType())) return;
961 // Flush store info for the pointer.
962 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
963 // Flush load info for the pointer.
964 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
967 /// removeInstruction - Remove an instruction from the dependence analysis,
968 /// updating the dependence of instructions that previously depended on it.
969 /// This method attempts to keep the cache coherent using the reverse map.
970 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
971 // Walk through the Non-local dependencies, removing this one as the value
972 // for any cached queries.
973 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
974 if (NLDI != NonLocalDeps.end()) {
975 NonLocalDepInfo &BlockMap = NLDI->second.first;
976 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
978 if (Instruction *Inst = DI->second.getInst())
979 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
980 NonLocalDeps.erase(NLDI);
983 // If we have a cached local dependence query for this instruction, remove it.
985 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
986 if (LocalDepEntry != LocalDeps.end()) {
987 // Remove us from DepInst's reverse set now that the local dep info is gone.
988 if (Instruction *Inst = LocalDepEntry->second.getInst())
989 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
991 // Remove this local dependency info.
992 LocalDeps.erase(LocalDepEntry);
995 // If we have any cached pointer dependencies on this instruction, remove
996 // them. If the instruction has non-pointer type, then it can't be a pointer
999 // Remove it from both the load info and the store info. The instruction
1000 // can't be in either of these maps if it is non-pointer.
1001 if (isa<PointerType>(RemInst->getType())) {
1002 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1003 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1006 // Loop over all of the things that depend on the instruction we're removing.
1008 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1010 // If we find RemInst as a clobber or Def in any of the maps for other values,
1011 // we need to replace its entry with a dirty version of the instruction after
1012 // it. If RemInst is a terminator, we use a null dirty value.
1014 // Using a dirty version of the instruction after RemInst saves having to scan
1015 // the entire block to get to this point.
1016 MemDepResult NewDirtyVal;
1017 if (!RemInst->isTerminator())
1018 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1020 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1021 if (ReverseDepIt != ReverseLocalDeps.end()) {
1022 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1023 // RemInst can't be the terminator if it has local stuff depending on it.
1024 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1025 "Nothing can locally depend on a terminator");
1027 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1028 E = ReverseDeps.end(); I != E; ++I) {
1029 Instruction *InstDependingOnRemInst = *I;
1030 assert(InstDependingOnRemInst != RemInst &&
1031 "Already removed our local dep info");
1033 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1035 // Make sure to remember that new things depend on NewDepInst.
1036 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1037 "a local dep on this if it is a terminator!");
1038 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1039 InstDependingOnRemInst));
1042 ReverseLocalDeps.erase(ReverseDepIt);
1044 // Add new reverse deps after scanning the set, to avoid invalidating the
1045 // 'ReverseDeps' reference.
1046 while (!ReverseDepsToAdd.empty()) {
1047 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1048 .insert(ReverseDepsToAdd.back().second);
1049 ReverseDepsToAdd.pop_back();
1053 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1054 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1055 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1056 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1058 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1060 PerInstNLInfo &INLD = NonLocalDeps[*I];
1061 // The information is now dirty!
1064 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1065 DE = INLD.first.end(); DI != DE; ++DI) {
1066 if (DI->second.getInst() != RemInst) continue;
1068 // Convert to a dirty entry for the subsequent instruction.
1069 DI->second = NewDirtyVal;
1071 if (Instruction *NextI = NewDirtyVal.getInst())
1072 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1076 ReverseNonLocalDeps.erase(ReverseDepIt);
1078 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1079 while (!ReverseDepsToAdd.empty()) {
1080 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1081 .insert(ReverseDepsToAdd.back().second);
1082 ReverseDepsToAdd.pop_back();
1086 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1087 // value in the NonLocalPointerDeps info.
1088 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1089 ReverseNonLocalPtrDeps.find(RemInst);
1090 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1091 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1092 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1094 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1095 E = Set.end(); I != E; ++I) {
1096 ValueIsLoadPair P = *I;
1097 assert(P.getPointer() != RemInst &&
1098 "Already removed NonLocalPointerDeps info for RemInst");
1100 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second;
1102 // The cache is not valid for any specific block anymore.
1103 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair();
1105 // Update any entries for RemInst to use the instruction after it.
1106 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1108 if (DI->second.getInst() != RemInst) continue;
1110 // Convert to a dirty entry for the subsequent instruction.
1111 DI->second = NewDirtyVal;
1113 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1114 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1117 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1118 // subsequent value may invalidate the sortedness.
1119 std::sort(NLPDI.begin(), NLPDI.end());
1122 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1124 while (!ReversePtrDepsToAdd.empty()) {
1125 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1126 .insert(ReversePtrDepsToAdd.back().second);
1127 ReversePtrDepsToAdd.pop_back();
1132 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1133 AA->deleteValue(RemInst);
1134 DEBUG(verifyRemoved(RemInst));
1136 /// verifyRemoved - Verify that the specified instruction does not occur
1137 /// in our internal data structures.
1138 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1139 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1140 E = LocalDeps.end(); I != E; ++I) {
1141 assert(I->first != D && "Inst occurs in data structures");
1142 assert(I->second.getInst() != D &&
1143 "Inst occurs in data structures");
1146 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1147 E = NonLocalPointerDeps.end(); I != E; ++I) {
1148 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1149 const NonLocalDepInfo &Val = I->second.second;
1150 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1152 assert(II->second.getInst() != D && "Inst occurs as NLPD value");
1155 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1156 E = NonLocalDeps.end(); I != E; ++I) {
1157 assert(I->first != D && "Inst occurs in data structures");
1158 const PerInstNLInfo &INLD = I->second;
1159 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1160 EE = INLD.first.end(); II != EE; ++II)
1161 assert(II->second.getInst() != D && "Inst occurs in data structures");
1164 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1165 E = ReverseLocalDeps.end(); I != E; ++I) {
1166 assert(I->first != D && "Inst occurs in data structures");
1167 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1168 EE = I->second.end(); II != EE; ++II)
1169 assert(*II != D && "Inst occurs in data structures");
1172 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1173 E = ReverseNonLocalDeps.end();
1175 assert(I->first != D && "Inst occurs in data structures");
1176 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1177 EE = I->second.end(); II != EE; ++II)
1178 assert(*II != D && "Inst occurs in data structures");
1181 for (ReverseNonLocalPtrDepTy::const_iterator
1182 I = ReverseNonLocalPtrDeps.begin(),
1183 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1184 assert(I->first != D && "Inst occurs in rev NLPD map");
1186 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1187 E = I->second.end(); II != E; ++II)
1188 assert(*II != ValueIsLoadPair(D, false) &&
1189 *II != ValueIsLoadPair(D, true) &&
1190 "Inst occurs in ReverseNonLocalPtrDeps map");