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/LLVMContext.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/Dominators.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Analysis/MemoryBuiltins.h"
27 #include "llvm/Analysis/PHITransAddr.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/Support/PredIteratorCache.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Target/TargetData.h"
35 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
36 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
37 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
39 STATISTIC(NumCacheNonLocalPtr,
40 "Number of fully cached non-local ptr responses");
41 STATISTIC(NumCacheDirtyNonLocalPtr,
42 "Number of cached, but dirty, non-local ptr responses");
43 STATISTIC(NumUncacheNonLocalPtr,
44 "Number of uncached non-local ptr responses");
45 STATISTIC(NumCacheCompleteNonLocalPtr,
46 "Number of block queries that were completely cached");
48 char MemoryDependenceAnalysis::ID = 0;
50 // Register this pass...
51 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
52 "Memory Dependence Analysis", false, true)
53 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
54 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
55 "Memory Dependence Analysis", false, true)
57 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
58 : FunctionPass(ID), PredCache(0) {
59 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
61 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
64 /// Clean up memory in between runs
65 void MemoryDependenceAnalysis::releaseMemory() {
68 NonLocalPointerDeps.clear();
69 ReverseLocalDeps.clear();
70 ReverseNonLocalDeps.clear();
71 ReverseNonLocalPtrDeps.clear();
77 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
79 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
81 AU.addRequiredTransitive<AliasAnalysis>();
84 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
85 AA = &getAnalysis<AliasAnalysis>();
86 TD = getAnalysisIfAvailable<TargetData>();
88 PredCache.reset(new PredIteratorCache());
92 /// RemoveFromReverseMap - This is a helper function that removes Val from
93 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
94 template <typename KeyTy>
95 static void RemoveFromReverseMap(DenseMap<Instruction*,
96 SmallPtrSet<KeyTy, 4> > &ReverseMap,
97 Instruction *Inst, KeyTy Val) {
98 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
99 InstIt = ReverseMap.find(Inst);
100 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
101 bool Found = InstIt->second.erase(Val);
102 assert(Found && "Invalid reverse map!"); Found=Found;
103 if (InstIt->second.empty())
104 ReverseMap.erase(InstIt);
107 /// GetLocation - If the given instruction references a specific memory
108 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
109 /// Return a ModRefInfo value describing the general behavior of the
112 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
113 AliasAnalysis::Location &Loc,
115 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
116 if (LI->isVolatile()) {
117 Loc = AliasAnalysis::Location();
118 return AliasAnalysis::ModRef;
120 Loc = AA->getLocation(LI);
121 return AliasAnalysis::Ref;
124 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
125 if (SI->isVolatile()) {
126 Loc = AliasAnalysis::Location();
127 return AliasAnalysis::ModRef;
129 Loc = AA->getLocation(SI);
130 return AliasAnalysis::Mod;
133 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
134 Loc = AA->getLocation(V);
135 return AliasAnalysis::ModRef;
138 if (const CallInst *CI = isFreeCall(Inst)) {
139 // calls to free() deallocate the entire structure
140 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
141 return AliasAnalysis::Mod;
144 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
145 switch (II->getIntrinsicID()) {
146 case Intrinsic::lifetime_start:
147 case Intrinsic::lifetime_end:
148 case Intrinsic::invariant_start:
149 Loc = AliasAnalysis::Location(II->getArgOperand(1),
150 cast<ConstantInt>(II->getArgOperand(0))
152 II->getMetadata(LLVMContext::MD_tbaa));
153 // These intrinsics don't really modify the memory, but returning Mod
154 // will allow them to be handled conservatively.
155 return AliasAnalysis::Mod;
156 case Intrinsic::invariant_end:
157 Loc = AliasAnalysis::Location(II->getArgOperand(2),
158 cast<ConstantInt>(II->getArgOperand(1))
160 II->getMetadata(LLVMContext::MD_tbaa));
161 // These intrinsics don't really modify the memory, but returning Mod
162 // will allow them to be handled conservatively.
163 return AliasAnalysis::Mod;
168 // Otherwise, just do the coarse-grained thing that always works.
169 if (Inst->mayWriteToMemory())
170 return AliasAnalysis::ModRef;
171 if (Inst->mayReadFromMemory())
172 return AliasAnalysis::Ref;
173 return AliasAnalysis::NoModRef;
176 /// getCallSiteDependencyFrom - Private helper for finding the local
177 /// dependencies of a call site.
178 MemDepResult MemoryDependenceAnalysis::
179 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
180 BasicBlock::iterator ScanIt, BasicBlock *BB) {
181 // Walk backwards through the block, looking for dependencies
182 while (ScanIt != BB->begin()) {
183 Instruction *Inst = --ScanIt;
185 // If this inst is a memory op, get the pointer it accessed
186 AliasAnalysis::Location Loc;
187 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
189 // A simple instruction.
190 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
191 return MemDepResult::getClobber(Inst);
195 if (CallSite InstCS = cast<Value>(Inst)) {
196 // Debug intrinsics don't cause dependences.
197 if (isa<DbgInfoIntrinsic>(Inst)) continue;
198 // If these two calls do not interfere, look past it.
199 switch (AA->getModRefInfo(CS, InstCS)) {
200 case AliasAnalysis::NoModRef:
201 // If the two calls are the same, return InstCS as a Def, so that
202 // CS can be found redundant and eliminated.
203 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
204 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
205 return MemDepResult::getDef(Inst);
207 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
211 return MemDepResult::getClobber(Inst);
216 // No dependence found. If this is the entry block of the function, it is a
217 // clobber, otherwise it is non-local.
218 if (BB != &BB->getParent()->getEntryBlock())
219 return MemDepResult::getNonLocal();
220 return MemDepResult::getClobber(ScanIt);
223 /// getPointerDependencyFrom - Return the instruction on which a memory
224 /// location depends. If isLoad is true, this routine ignores may-aliases with
225 /// read-only operations. If isLoad is false, this routine ignores may-aliases
226 /// with reads from read-only locations.
227 MemDepResult MemoryDependenceAnalysis::
228 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
229 BasicBlock::iterator ScanIt, BasicBlock *BB) {
231 Value *InvariantTag = 0;
233 // Walk backwards through the basic block, looking for dependencies.
234 while (ScanIt != BB->begin()) {
235 Instruction *Inst = --ScanIt;
237 // If we're in an invariant region, no dependencies can be found before
238 // we pass an invariant-begin marker.
239 if (InvariantTag == Inst) {
244 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
245 // Debug intrinsics don't (and can't) cause dependences.
246 if (isa<DbgInfoIntrinsic>(II)) continue;
248 // If we pass an invariant-end marker, then we've just entered an
249 // invariant region and can start ignoring dependencies.
250 if (II->getIntrinsicID() == Intrinsic::invariant_end) {
251 // FIXME: This only considers queries directly on the invariant-tagged
252 // pointer, not on query pointers that are indexed off of them. It'd
253 // be nice to handle that at some point.
254 AliasAnalysis::AliasResult R =
255 AA->alias(AliasAnalysis::Location(II->getArgOperand(2)), MemLoc);
256 if (R == AliasAnalysis::MustAlias)
257 InvariantTag = II->getArgOperand(0);
262 // If we reach a lifetime begin or end marker, then the query ends here
263 // because the value is undefined.
264 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
265 // FIXME: This only considers queries directly on the invariant-tagged
266 // pointer, not on query pointers that are indexed off of them. It'd
267 // be nice to handle that at some point.
268 AliasAnalysis::AliasResult R =
269 AA->alias(AliasAnalysis::Location(II->getArgOperand(1)), MemLoc);
270 if (R == AliasAnalysis::MustAlias)
271 return MemDepResult::getDef(II);
276 // If we're querying on a load and we're in an invariant region, we're done
277 // at this point. Nothing a load depends on can live in an invariant region.
279 // FIXME: this will prevent us from returning load/load must-aliases, so GVN
280 // won't remove redundant loads.
281 if (isLoad && InvariantTag) continue;
283 // Values depend on loads if the pointers are must aliased. This means that
284 // a load depends on another must aliased load from the same value.
285 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
286 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
288 // If we found a pointer, check if it could be the same as our pointer.
289 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
290 if (R == AliasAnalysis::NoAlias)
293 // May-alias loads don't depend on each other without a dependence.
294 if (isLoad && R == AliasAnalysis::MayAlias)
297 // Stores don't alias loads from read-only memory.
298 if (!isLoad && AA->pointsToConstantMemory(LoadLoc))
301 // Stores depend on may and must aliased loads, loads depend on must-alias
303 return MemDepResult::getDef(Inst);
306 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
307 // There can't be stores to the value we care about inside an
309 if (InvariantTag) continue;
311 // If alias analysis can tell that this store is guaranteed to not modify
312 // the query pointer, ignore it. Use getModRefInfo to handle cases where
313 // the query pointer points to constant memory etc.
314 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
317 // Ok, this store might clobber the query pointer. Check to see if it is
318 // a must alias: in this case, we want to return this as a def.
319 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
321 // If we found a pointer, check if it could be the same as our pointer.
322 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
324 if (R == AliasAnalysis::NoAlias)
326 if (R == AliasAnalysis::MayAlias)
327 return MemDepResult::getClobber(Inst);
328 return MemDepResult::getDef(Inst);
331 // If this is an allocation, and if we know that the accessed pointer is to
332 // the allocation, return Def. This means that there is no dependence and
333 // the access can be optimized based on that. For example, a load could
335 // Note: Only determine this to be a malloc if Inst is the malloc call, not
336 // a subsequent bitcast of the malloc call result. There can be stores to
337 // the malloced memory between the malloc call and its bitcast uses, and we
338 // need to continue scanning until the malloc call.
339 if (isa<AllocaInst>(Inst) ||
340 (isa<CallInst>(Inst) && extractMallocCall(Inst))) {
341 const Value *AccessPtr = MemLoc.Ptr->getUnderlyingObject();
343 if (AccessPtr == Inst ||
344 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
345 return MemDepResult::getDef(Inst);
349 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
350 switch (AA->getModRefInfo(Inst, MemLoc)) {
351 case AliasAnalysis::NoModRef:
352 // If the call has no effect on the queried pointer, just ignore it.
354 case AliasAnalysis::Mod:
355 // If we're in an invariant region, we can ignore calls that ONLY
356 // modify the pointer.
357 if (InvariantTag) continue;
358 return MemDepResult::getClobber(Inst);
359 case AliasAnalysis::Ref:
360 // If the call is known to never store to the pointer, and if this is a
361 // load query, we can safely ignore it (scan past it).
365 // Otherwise, there is a potential dependence. Return a clobber.
366 return MemDepResult::getClobber(Inst);
370 // No dependence found. If this is the entry block of the function, it is a
371 // clobber, otherwise it is non-local.
372 if (BB != &BB->getParent()->getEntryBlock())
373 return MemDepResult::getNonLocal();
374 return MemDepResult::getClobber(ScanIt);
377 /// getDependency - Return the instruction on which a memory operation
379 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
380 Instruction *ScanPos = QueryInst;
382 // Check for a cached result
383 MemDepResult &LocalCache = LocalDeps[QueryInst];
385 // If the cached entry is non-dirty, just return it. Note that this depends
386 // on MemDepResult's default constructing to 'dirty'.
387 if (!LocalCache.isDirty())
390 // Otherwise, if we have a dirty entry, we know we can start the scan at that
391 // instruction, which may save us some work.
392 if (Instruction *Inst = LocalCache.getInst()) {
395 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
398 BasicBlock *QueryParent = QueryInst->getParent();
401 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
402 // No dependence found. If this is the entry block of the function, it is a
403 // clobber, otherwise it is non-local.
404 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
405 LocalCache = MemDepResult::getNonLocal();
407 LocalCache = MemDepResult::getClobber(QueryInst);
409 AliasAnalysis::Location MemLoc;
410 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
412 // If we can do a pointer scan, make it happen.
413 bool isLoad = !(MR & AliasAnalysis::Mod);
414 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
415 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end;
417 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
419 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
420 CallSite QueryCS(QueryInst);
421 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
422 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
425 // Non-memory instruction.
426 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
429 // Remember the result!
430 if (Instruction *I = LocalCache.getInst())
431 ReverseLocalDeps[I].insert(QueryInst);
437 /// AssertSorted - This method is used when -debug is specified to verify that
438 /// cache arrays are properly kept sorted.
439 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
441 if (Count == -1) Count = Cache.size();
442 if (Count == 0) return;
444 for (unsigned i = 1; i != unsigned(Count); ++i)
445 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
449 /// getNonLocalCallDependency - Perform a full dependency query for the
450 /// specified call, returning the set of blocks that the value is
451 /// potentially live across. The returned set of results will include a
452 /// "NonLocal" result for all blocks where the value is live across.
454 /// This method assumes the instruction returns a "NonLocal" dependency
455 /// within its own block.
457 /// This returns a reference to an internal data structure that may be
458 /// invalidated on the next non-local query or when an instruction is
459 /// removed. Clients must copy this data if they want it around longer than
461 const MemoryDependenceAnalysis::NonLocalDepInfo &
462 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
463 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
464 "getNonLocalCallDependency should only be used on calls with non-local deps!");
465 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
466 NonLocalDepInfo &Cache = CacheP.first;
468 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
469 /// the cached case, this can happen due to instructions being deleted etc. In
470 /// the uncached case, this starts out as the set of predecessors we care
472 SmallVector<BasicBlock*, 32> DirtyBlocks;
474 if (!Cache.empty()) {
475 // Okay, we have a cache entry. If we know it is not dirty, just return it
476 // with no computation.
477 if (!CacheP.second) {
482 // If we already have a partially computed set of results, scan them to
483 // determine what is dirty, seeding our initial DirtyBlocks worklist.
484 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
486 if (I->getResult().isDirty())
487 DirtyBlocks.push_back(I->getBB());
489 // Sort the cache so that we can do fast binary search lookups below.
490 std::sort(Cache.begin(), Cache.end());
492 ++NumCacheDirtyNonLocal;
493 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
494 // << Cache.size() << " cached: " << *QueryInst;
496 // Seed DirtyBlocks with each of the preds of QueryInst's block.
497 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
498 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
499 DirtyBlocks.push_back(*PI);
500 ++NumUncacheNonLocal;
503 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
504 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
506 SmallPtrSet<BasicBlock*, 64> Visited;
508 unsigned NumSortedEntries = Cache.size();
509 DEBUG(AssertSorted(Cache));
511 // Iterate while we still have blocks to update.
512 while (!DirtyBlocks.empty()) {
513 BasicBlock *DirtyBB = DirtyBlocks.back();
514 DirtyBlocks.pop_back();
516 // Already processed this block?
517 if (!Visited.insert(DirtyBB))
520 // Do a binary search to see if we already have an entry for this block in
521 // the cache set. If so, find it.
522 DEBUG(AssertSorted(Cache, NumSortedEntries));
523 NonLocalDepInfo::iterator Entry =
524 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
525 NonLocalDepEntry(DirtyBB));
526 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB)
529 NonLocalDepEntry *ExistingResult = 0;
530 if (Entry != Cache.begin()+NumSortedEntries &&
531 Entry->getBB() == DirtyBB) {
532 // If we already have an entry, and if it isn't already dirty, the block
534 if (!Entry->getResult().isDirty())
537 // Otherwise, remember this slot so we can update the value.
538 ExistingResult = &*Entry;
541 // If the dirty entry has a pointer, start scanning from it so we don't have
542 // to rescan the entire block.
543 BasicBlock::iterator ScanPos = DirtyBB->end();
544 if (ExistingResult) {
545 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
547 // We're removing QueryInst's use of Inst.
548 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
549 QueryCS.getInstruction());
553 // Find out if this block has a local dependency for QueryInst.
556 if (ScanPos != DirtyBB->begin()) {
557 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
558 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
559 // No dependence found. If this is the entry block of the function, it is
560 // a clobber, otherwise it is non-local.
561 Dep = MemDepResult::getNonLocal();
563 Dep = MemDepResult::getClobber(ScanPos);
566 // If we had a dirty entry for the block, update it. Otherwise, just add
569 ExistingResult->setResult(Dep);
571 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
573 // If the block has a dependency (i.e. it isn't completely transparent to
574 // the value), remember the association!
575 if (!Dep.isNonLocal()) {
576 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
577 // update this when we remove instructions.
578 if (Instruction *Inst = Dep.getInst())
579 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
582 // If the block *is* completely transparent to the load, we need to check
583 // the predecessors of this block. Add them to our worklist.
584 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
585 DirtyBlocks.push_back(*PI);
592 /// getNonLocalPointerDependency - Perform a full dependency query for an
593 /// access to the specified (non-volatile) memory location, returning the
594 /// set of instructions that either define or clobber the value.
596 /// This method assumes the pointer has a "NonLocal" dependency within its
599 void MemoryDependenceAnalysis::
600 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
602 SmallVectorImpl<NonLocalDepResult> &Result) {
603 assert(Loc.Ptr->getType()->isPointerTy() &&
604 "Can't get pointer deps of a non-pointer!");
607 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD);
609 // This is the set of blocks we've inspected, and the pointer we consider in
610 // each block. Because of critical edges, we currently bail out if querying
611 // a block with multiple different pointers. This can happen during PHI
613 DenseMap<BasicBlock*, Value*> Visited;
614 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
615 Result, Visited, true))
618 Result.push_back(NonLocalDepResult(FromBB,
619 MemDepResult::getClobber(FromBB->begin()),
620 const_cast<Value *>(Loc.Ptr)));
623 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
624 /// Pointer/PointeeSize using either cached information in Cache or by doing a
625 /// lookup (which may use dirty cache info if available). If we do a lookup,
626 /// add the result to the cache.
627 MemDepResult MemoryDependenceAnalysis::
628 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
629 bool isLoad, BasicBlock *BB,
630 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
632 // Do a binary search to see if we already have an entry for this block in
633 // the cache set. If so, find it.
634 NonLocalDepInfo::iterator Entry =
635 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
636 NonLocalDepEntry(BB));
637 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
640 NonLocalDepEntry *ExistingResult = 0;
641 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
642 ExistingResult = &*Entry;
644 // If we have a cached entry, and it is non-dirty, use it as the value for
646 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
647 ++NumCacheNonLocalPtr;
648 return ExistingResult->getResult();
651 // Otherwise, we have to scan for the value. If we have a dirty cache
652 // entry, start scanning from its position, otherwise we scan from the end
654 BasicBlock::iterator ScanPos = BB->end();
655 if (ExistingResult && ExistingResult->getResult().getInst()) {
656 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
657 "Instruction invalidated?");
658 ++NumCacheDirtyNonLocalPtr;
659 ScanPos = ExistingResult->getResult().getInst();
661 // Eliminating the dirty entry from 'Cache', so update the reverse info.
662 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
663 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
665 ++NumUncacheNonLocalPtr;
668 // Scan the block for the dependency.
669 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
671 // If we had a dirty entry for the block, update it. Otherwise, just add
674 ExistingResult->setResult(Dep);
676 Cache->push_back(NonLocalDepEntry(BB, Dep));
678 // If the block has a dependency (i.e. it isn't completely transparent to
679 // the value), remember the reverse association because we just added it
681 if (Dep.isNonLocal())
684 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
685 // update MemDep when we remove instructions.
686 Instruction *Inst = Dep.getInst();
687 assert(Inst && "Didn't depend on anything?");
688 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
689 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
693 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
694 /// number of elements in the array that are already properly ordered. This is
695 /// optimized for the case when only a few entries are added.
697 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
698 unsigned NumSortedEntries) {
699 switch (Cache.size() - NumSortedEntries) {
701 // done, no new entries.
704 // Two new entries, insert the last one into place.
705 NonLocalDepEntry Val = Cache.back();
707 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
708 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
709 Cache.insert(Entry, Val);
713 // One new entry, Just insert the new value at the appropriate position.
714 if (Cache.size() != 1) {
715 NonLocalDepEntry Val = Cache.back();
717 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
718 std::upper_bound(Cache.begin(), Cache.end(), Val);
719 Cache.insert(Entry, Val);
723 // Added many values, do a full scale sort.
724 std::sort(Cache.begin(), Cache.end());
729 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
730 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
731 /// results to the results vector and keep track of which blocks are visited in
734 /// This has special behavior for the first block queries (when SkipFirstBlock
735 /// is true). In this special case, it ignores the contents of the specified
736 /// block and starts returning dependence info for its predecessors.
738 /// This function returns false on success, or true to indicate that it could
739 /// not compute dependence information for some reason. This should be treated
740 /// as a clobber dependence on the first instruction in the predecessor block.
741 bool MemoryDependenceAnalysis::
742 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
743 const AliasAnalysis::Location &Loc,
744 bool isLoad, BasicBlock *StartBB,
745 SmallVectorImpl<NonLocalDepResult> &Result,
746 DenseMap<BasicBlock*, Value*> &Visited,
747 bool SkipFirstBlock) {
749 // Look up the cached info for Pointer.
750 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
752 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
753 // CacheKey, this value will be inserted as the associated value. Otherwise,
754 // it'll be ignored, and we'll have to check to see if the cached size and
755 // tbaa tag are consistent with the current query.
756 NonLocalPointerInfo InitialNLPI;
757 InitialNLPI.Size = Loc.Size;
758 InitialNLPI.TBAATag = Loc.TBAATag;
760 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
762 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
763 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
764 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
766 // If we already have a cache entry for this CacheKey, we may need to do some
767 // work to reconcile the cache entry and the current query.
769 if (CacheInfo->Size < Loc.Size) {
770 // The query's Size is greater than the cached one. Throw out the
771 // cached data and procede with the query at the greater size.
772 CacheInfo->Pair = BBSkipFirstBlockPair();
773 CacheInfo->Size = Loc.Size;
774 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
775 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
776 if (Instruction *Inst = DI->getResult().getInst())
777 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
778 CacheInfo->NonLocalDeps.clear();
779 } else if (CacheInfo->Size > Loc.Size) {
780 // This query's Size is less than the cached one. Conservatively restart
781 // the query using the greater size.
782 return getNonLocalPointerDepFromBB(Pointer,
783 Loc.getWithNewSize(CacheInfo->Size),
784 isLoad, StartBB, Result, Visited,
788 // If the query's TBAATag is inconsistent with the cached one,
789 // conservatively throw out the cached data and restart the query with
791 if (CacheInfo->TBAATag != Loc.TBAATag) {
792 if (CacheInfo->TBAATag) {
793 CacheInfo->Pair = BBSkipFirstBlockPair();
794 CacheInfo->TBAATag = 0;
795 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
796 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
797 if (Instruction *Inst = DI->getResult().getInst())
798 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
799 CacheInfo->NonLocalDeps.clear();
802 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(),
803 isLoad, StartBB, Result, Visited,
808 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
810 // If we have valid cached information for exactly the block we are
811 // investigating, just return it with no recomputation.
812 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
813 // We have a fully cached result for this query then we can just return the
814 // cached results and populate the visited set. However, we have to verify
815 // that we don't already have conflicting results for these blocks. Check
816 // to ensure that if a block in the results set is in the visited set that
817 // it was for the same pointer query.
818 if (!Visited.empty()) {
819 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
821 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
822 if (VI == Visited.end() || VI->second == Pointer.getAddr())
825 // We have a pointer mismatch in a block. Just return clobber, saying
826 // that something was clobbered in this result. We could also do a
827 // non-fully cached query, but there is little point in doing this.
832 Value *Addr = Pointer.getAddr();
833 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
835 Visited.insert(std::make_pair(I->getBB(), Addr));
836 if (!I->getResult().isNonLocal())
837 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
839 ++NumCacheCompleteNonLocalPtr;
843 // Otherwise, either this is a new block, a block with an invalid cache
844 // pointer or one that we're about to invalidate by putting more info into it
845 // than its valid cache info. If empty, the result will be valid cache info,
846 // otherwise it isn't.
848 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
850 CacheInfo->Pair = BBSkipFirstBlockPair();
852 SmallVector<BasicBlock*, 32> Worklist;
853 Worklist.push_back(StartBB);
855 // Keep track of the entries that we know are sorted. Previously cached
856 // entries will all be sorted. The entries we add we only sort on demand (we
857 // don't insert every element into its sorted position). We know that we
858 // won't get any reuse from currently inserted values, because we don't
859 // revisit blocks after we insert info for them.
860 unsigned NumSortedEntries = Cache->size();
861 DEBUG(AssertSorted(*Cache));
863 while (!Worklist.empty()) {
864 BasicBlock *BB = Worklist.pop_back_val();
866 // Skip the first block if we have it.
867 if (!SkipFirstBlock) {
868 // Analyze the dependency of *Pointer in FromBB. See if we already have
870 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
872 // Get the dependency info for Pointer in BB. If we have cached
873 // information, we will use it, otherwise we compute it.
874 DEBUG(AssertSorted(*Cache, NumSortedEntries));
875 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
878 // If we got a Def or Clobber, add this to the list of results.
879 if (!Dep.isNonLocal()) {
880 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
885 // If 'Pointer' is an instruction defined in this block, then we need to do
886 // phi translation to change it into a value live in the predecessor block.
887 // If not, we just add the predecessors to the worklist and scan them with
889 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
890 SkipFirstBlock = false;
891 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
892 // Verify that we haven't looked at this block yet.
893 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
894 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
895 if (InsertRes.second) {
896 // First time we've looked at *PI.
897 Worklist.push_back(*PI);
901 // If we have seen this block before, but it was with a different
902 // pointer then we have a phi translation failure and we have to treat
903 // this as a clobber.
904 if (InsertRes.first->second != Pointer.getAddr())
905 goto PredTranslationFailure;
910 // We do need to do phi translation, if we know ahead of time we can't phi
911 // translate this value, don't even try.
912 if (!Pointer.IsPotentiallyPHITranslatable())
913 goto PredTranslationFailure;
915 // We may have added values to the cache list before this PHI translation.
916 // If so, we haven't done anything to ensure that the cache remains sorted.
917 // Sort it now (if needed) so that recursive invocations of
918 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
919 // value will only see properly sorted cache arrays.
920 if (Cache && NumSortedEntries != Cache->size()) {
921 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
922 NumSortedEntries = Cache->size();
926 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
927 BasicBlock *Pred = *PI;
929 // Get the PHI translated pointer in this predecessor. This can fail if
930 // not translatable, in which case the getAddr() returns null.
931 PHITransAddr PredPointer(Pointer);
932 PredPointer.PHITranslateValue(BB, Pred, 0);
934 Value *PredPtrVal = PredPointer.getAddr();
936 // Check to see if we have already visited this pred block with another
937 // pointer. If so, we can't do this lookup. This failure can occur
938 // with PHI translation when a critical edge exists and the PHI node in
939 // the successor translates to a pointer value different than the
940 // pointer the block was first analyzed with.
941 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
942 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
944 if (!InsertRes.second) {
945 // If the predecessor was visited with PredPtr, then we already did
946 // the analysis and can ignore it.
947 if (InsertRes.first->second == PredPtrVal)
950 // Otherwise, the block was previously analyzed with a different
951 // pointer. We can't represent the result of this case, so we just
952 // treat this as a phi translation failure.
953 goto PredTranslationFailure;
956 // If PHI translation was unable to find an available pointer in this
957 // predecessor, then we have to assume that the pointer is clobbered in
958 // that predecessor. We can still do PRE of the load, which would insert
959 // a computation of the pointer in this predecessor.
960 if (PredPtrVal == 0) {
961 // Add the entry to the Result list.
962 NonLocalDepResult Entry(Pred,
963 MemDepResult::getClobber(Pred->getTerminator()),
965 Result.push_back(Entry);
967 // Since we had a phi translation failure, the cache for CacheKey won't
968 // include all of the entries that we need to immediately satisfy future
969 // queries. Mark this in NonLocalPointerDeps by setting the
970 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
971 // cached value to do more work but not miss the phi trans failure.
972 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
973 NLPI.Pair = BBSkipFirstBlockPair();
977 // FIXME: it is entirely possible that PHI translating will end up with
978 // the same value. Consider PHI translating something like:
979 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
980 // to recurse here, pedantically speaking.
982 // If we have a problem phi translating, fall through to the code below
983 // to handle the failure condition.
984 if (getNonLocalPointerDepFromBB(PredPointer,
985 Loc.getWithNewPtr(PredPointer.getAddr()),
988 goto PredTranslationFailure;
991 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
992 CacheInfo = &NonLocalPointerDeps[CacheKey];
993 Cache = &CacheInfo->NonLocalDeps;
994 NumSortedEntries = Cache->size();
996 // Since we did phi translation, the "Cache" set won't contain all of the
997 // results for the query. This is ok (we can still use it to accelerate
998 // specific block queries) but we can't do the fastpath "return all
999 // results from the set" Clear out the indicator for this.
1000 CacheInfo->Pair = BBSkipFirstBlockPair();
1001 SkipFirstBlock = false;
1004 PredTranslationFailure:
1007 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1008 CacheInfo = &NonLocalPointerDeps[CacheKey];
1009 Cache = &CacheInfo->NonLocalDeps;
1010 NumSortedEntries = Cache->size();
1013 // Since we failed phi translation, the "Cache" set won't contain all of the
1014 // results for the query. This is ok (we can still use it to accelerate
1015 // specific block queries) but we can't do the fastpath "return all
1016 // results from the set". Clear out the indicator for this.
1017 CacheInfo->Pair = BBSkipFirstBlockPair();
1019 // If *nothing* works, mark the pointer as being clobbered by the first
1020 // instruction in this block.
1022 // If this is the magic first block, return this as a clobber of the whole
1023 // incoming value. Since we can't phi translate to one of the predecessors,
1024 // we have to bail out.
1028 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1029 assert(I != Cache->rend() && "Didn't find current block??");
1030 if (I->getBB() != BB)
1033 assert(I->getResult().isNonLocal() &&
1034 "Should only be here with transparent block");
1035 I->setResult(MemDepResult::getClobber(BB->begin()));
1036 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
1037 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1038 Pointer.getAddr()));
1043 // Okay, we're done now. If we added new values to the cache, re-sort it.
1044 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1045 DEBUG(AssertSorted(*Cache));
1049 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1050 /// CachedNonLocalPointerInfo, remove it.
1051 void MemoryDependenceAnalysis::
1052 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1053 CachedNonLocalPointerInfo::iterator It =
1054 NonLocalPointerDeps.find(P);
1055 if (It == NonLocalPointerDeps.end()) return;
1057 // Remove all of the entries in the BB->val map. This involves removing
1058 // instructions from the reverse map.
1059 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1061 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1062 Instruction *Target = PInfo[i].getResult().getInst();
1063 if (Target == 0) continue; // Ignore non-local dep results.
1064 assert(Target->getParent() == PInfo[i].getBB());
1066 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1067 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1070 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1071 NonLocalPointerDeps.erase(It);
1075 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1076 /// information about the specified pointer, because it may be too
1077 /// conservative in memdep. This is an optional call that can be used when
1078 /// the client detects an equivalence between the pointer and some other
1079 /// value and replaces the other value with ptr. This can make Ptr available
1080 /// in more places that cached info does not necessarily keep.
1081 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1082 // If Ptr isn't really a pointer, just ignore it.
1083 if (!Ptr->getType()->isPointerTy()) return;
1084 // Flush store info for the pointer.
1085 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1086 // Flush load info for the pointer.
1087 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1090 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1091 /// This needs to be done when the CFG changes, e.g., due to splitting
1093 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1097 /// removeInstruction - Remove an instruction from the dependence analysis,
1098 /// updating the dependence of instructions that previously depended on it.
1099 /// This method attempts to keep the cache coherent using the reverse map.
1100 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1101 // Walk through the Non-local dependencies, removing this one as the value
1102 // for any cached queries.
1103 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1104 if (NLDI != NonLocalDeps.end()) {
1105 NonLocalDepInfo &BlockMap = NLDI->second.first;
1106 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1108 if (Instruction *Inst = DI->getResult().getInst())
1109 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1110 NonLocalDeps.erase(NLDI);
1113 // If we have a cached local dependence query for this instruction, remove it.
1115 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1116 if (LocalDepEntry != LocalDeps.end()) {
1117 // Remove us from DepInst's reverse set now that the local dep info is gone.
1118 if (Instruction *Inst = LocalDepEntry->second.getInst())
1119 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1121 // Remove this local dependency info.
1122 LocalDeps.erase(LocalDepEntry);
1125 // If we have any cached pointer dependencies on this instruction, remove
1126 // them. If the instruction has non-pointer type, then it can't be a pointer
1129 // Remove it from both the load info and the store info. The instruction
1130 // can't be in either of these maps if it is non-pointer.
1131 if (RemInst->getType()->isPointerTy()) {
1132 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1133 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1136 // Loop over all of the things that depend on the instruction we're removing.
1138 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1140 // If we find RemInst as a clobber or Def in any of the maps for other values,
1141 // we need to replace its entry with a dirty version of the instruction after
1142 // it. If RemInst is a terminator, we use a null dirty value.
1144 // Using a dirty version of the instruction after RemInst saves having to scan
1145 // the entire block to get to this point.
1146 MemDepResult NewDirtyVal;
1147 if (!RemInst->isTerminator())
1148 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1150 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1151 if (ReverseDepIt != ReverseLocalDeps.end()) {
1152 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1153 // RemInst can't be the terminator if it has local stuff depending on it.
1154 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1155 "Nothing can locally depend on a terminator");
1157 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1158 E = ReverseDeps.end(); I != E; ++I) {
1159 Instruction *InstDependingOnRemInst = *I;
1160 assert(InstDependingOnRemInst != RemInst &&
1161 "Already removed our local dep info");
1163 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1165 // Make sure to remember that new things depend on NewDepInst.
1166 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1167 "a local dep on this if it is a terminator!");
1168 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1169 InstDependingOnRemInst));
1172 ReverseLocalDeps.erase(ReverseDepIt);
1174 // Add new reverse deps after scanning the set, to avoid invalidating the
1175 // 'ReverseDeps' reference.
1176 while (!ReverseDepsToAdd.empty()) {
1177 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1178 .insert(ReverseDepsToAdd.back().second);
1179 ReverseDepsToAdd.pop_back();
1183 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1184 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1185 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1186 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1188 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1190 PerInstNLInfo &INLD = NonLocalDeps[*I];
1191 // The information is now dirty!
1194 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1195 DE = INLD.first.end(); DI != DE; ++DI) {
1196 if (DI->getResult().getInst() != RemInst) continue;
1198 // Convert to a dirty entry for the subsequent instruction.
1199 DI->setResult(NewDirtyVal);
1201 if (Instruction *NextI = NewDirtyVal.getInst())
1202 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1206 ReverseNonLocalDeps.erase(ReverseDepIt);
1208 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1209 while (!ReverseDepsToAdd.empty()) {
1210 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1211 .insert(ReverseDepsToAdd.back().second);
1212 ReverseDepsToAdd.pop_back();
1216 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1217 // value in the NonLocalPointerDeps info.
1218 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1219 ReverseNonLocalPtrDeps.find(RemInst);
1220 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1221 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1222 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1224 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1225 E = Set.end(); I != E; ++I) {
1226 ValueIsLoadPair P = *I;
1227 assert(P.getPointer() != RemInst &&
1228 "Already removed NonLocalPointerDeps info for RemInst");
1230 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1232 // The cache is not valid for any specific block anymore.
1233 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1235 // Update any entries for RemInst to use the instruction after it.
1236 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1238 if (DI->getResult().getInst() != RemInst) continue;
1240 // Convert to a dirty entry for the subsequent instruction.
1241 DI->setResult(NewDirtyVal);
1243 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1244 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1247 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1248 // subsequent value may invalidate the sortedness.
1249 std::sort(NLPDI.begin(), NLPDI.end());
1252 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1254 while (!ReversePtrDepsToAdd.empty()) {
1255 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1256 .insert(ReversePtrDepsToAdd.back().second);
1257 ReversePtrDepsToAdd.pop_back();
1262 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1263 AA->deleteValue(RemInst);
1264 DEBUG(verifyRemoved(RemInst));
1266 /// verifyRemoved - Verify that the specified instruction does not occur
1267 /// in our internal data structures.
1268 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1269 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1270 E = LocalDeps.end(); I != E; ++I) {
1271 assert(I->first != D && "Inst occurs in data structures");
1272 assert(I->second.getInst() != D &&
1273 "Inst occurs in data structures");
1276 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1277 E = NonLocalPointerDeps.end(); I != E; ++I) {
1278 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1279 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1280 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1282 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1285 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1286 E = NonLocalDeps.end(); I != E; ++I) {
1287 assert(I->first != D && "Inst occurs in data structures");
1288 const PerInstNLInfo &INLD = I->second;
1289 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1290 EE = INLD.first.end(); II != EE; ++II)
1291 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1294 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1295 E = ReverseLocalDeps.end(); I != E; ++I) {
1296 assert(I->first != D && "Inst occurs in data structures");
1297 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1298 EE = I->second.end(); II != EE; ++II)
1299 assert(*II != D && "Inst occurs in data structures");
1302 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1303 E = ReverseNonLocalDeps.end();
1305 assert(I->first != D && "Inst occurs in data structures");
1306 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1307 EE = I->second.end(); II != EE; ++II)
1308 assert(*II != D && "Inst occurs in data structures");
1311 for (ReverseNonLocalPtrDepTy::const_iterator
1312 I = ReverseNonLocalPtrDeps.begin(),
1313 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1314 assert(I->first != D && "Inst occurs in rev NLPD map");
1316 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1317 E = I->second.end(); II != E; ++II)
1318 assert(*II != ValueIsLoadPair(D, false) &&
1319 *II != ValueIsLoadPair(D, true) &&
1320 "Inst occurs in ReverseNonLocalPtrDeps map");