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/Analysis/ValueTracking.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/Support/PredIteratorCache.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Target/TargetData.h"
36 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
37 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
38 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
40 STATISTIC(NumCacheNonLocalPtr,
41 "Number of fully cached non-local ptr responses");
42 STATISTIC(NumCacheDirtyNonLocalPtr,
43 "Number of cached, but dirty, non-local ptr responses");
44 STATISTIC(NumUncacheNonLocalPtr,
45 "Number of uncached non-local ptr responses");
46 STATISTIC(NumCacheCompleteNonLocalPtr,
47 "Number of block queries that were completely cached");
49 char MemoryDependenceAnalysis::ID = 0;
51 // Register this pass...
52 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
53 "Memory Dependence Analysis", false, true)
54 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
55 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
56 "Memory Dependence Analysis", false, true)
58 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
59 : FunctionPass(ID), PredCache(0) {
60 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
62 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
65 /// Clean up memory in between runs
66 void MemoryDependenceAnalysis::releaseMemory() {
69 NonLocalPointerDeps.clear();
70 ReverseLocalDeps.clear();
71 ReverseNonLocalDeps.clear();
72 ReverseNonLocalPtrDeps.clear();
78 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
80 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
82 AU.addRequiredTransitive<AliasAnalysis>();
85 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
86 AA = &getAnalysis<AliasAnalysis>();
87 TD = getAnalysisIfAvailable<TargetData>();
89 PredCache.reset(new PredIteratorCache());
93 /// RemoveFromReverseMap - This is a helper function that removes Val from
94 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
95 template <typename KeyTy>
96 static void RemoveFromReverseMap(DenseMap<Instruction*,
97 SmallPtrSet<KeyTy, 4> > &ReverseMap,
98 Instruction *Inst, KeyTy Val) {
99 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
100 InstIt = ReverseMap.find(Inst);
101 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
102 bool Found = InstIt->second.erase(Val);
103 assert(Found && "Invalid reverse map!"); (void)Found;
104 if (InstIt->second.empty())
105 ReverseMap.erase(InstIt);
108 /// GetLocation - If the given instruction references a specific memory
109 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
110 /// Return a ModRefInfo value describing the general behavior of the
113 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
114 AliasAnalysis::Location &Loc,
116 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
117 if (LI->isVolatile()) {
118 Loc = AliasAnalysis::Location();
119 return AliasAnalysis::ModRef;
121 Loc = AA->getLocation(LI);
122 return AliasAnalysis::Ref;
125 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
126 if (SI->isVolatile()) {
127 Loc = AliasAnalysis::Location();
128 return AliasAnalysis::ModRef;
130 Loc = AA->getLocation(SI);
131 return AliasAnalysis::Mod;
134 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
135 Loc = AA->getLocation(V);
136 return AliasAnalysis::ModRef;
139 if (const CallInst *CI = isFreeCall(Inst)) {
140 // calls to free() deallocate the entire structure
141 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
142 return AliasAnalysis::Mod;
145 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
146 switch (II->getIntrinsicID()) {
147 case Intrinsic::lifetime_start:
148 case Intrinsic::lifetime_end:
149 case Intrinsic::invariant_start:
150 Loc = AliasAnalysis::Location(II->getArgOperand(1),
151 cast<ConstantInt>(II->getArgOperand(0))
153 II->getMetadata(LLVMContext::MD_tbaa));
154 // These intrinsics don't really modify the memory, but returning Mod
155 // will allow them to be handled conservatively.
156 return AliasAnalysis::Mod;
157 case Intrinsic::invariant_end:
158 Loc = AliasAnalysis::Location(II->getArgOperand(2),
159 cast<ConstantInt>(II->getArgOperand(1))
161 II->getMetadata(LLVMContext::MD_tbaa));
162 // These intrinsics don't really modify the memory, but returning Mod
163 // will allow them to be handled conservatively.
164 return AliasAnalysis::Mod;
169 // Otherwise, just do the coarse-grained thing that always works.
170 if (Inst->mayWriteToMemory())
171 return AliasAnalysis::ModRef;
172 if (Inst->mayReadFromMemory())
173 return AliasAnalysis::Ref;
174 return AliasAnalysis::NoModRef;
177 /// getCallSiteDependencyFrom - Private helper for finding the local
178 /// dependencies of a call site.
179 MemDepResult MemoryDependenceAnalysis::
180 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
181 BasicBlock::iterator ScanIt, BasicBlock *BB) {
182 // Walk backwards through the block, looking for dependencies
183 while (ScanIt != BB->begin()) {
184 Instruction *Inst = --ScanIt;
186 // If this inst is a memory op, get the pointer it accessed
187 AliasAnalysis::Location Loc;
188 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
190 // A simple instruction.
191 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
192 return MemDepResult::getClobber(Inst);
196 if (CallSite InstCS = cast<Value>(Inst)) {
197 // Debug intrinsics don't cause dependences.
198 if (isa<DbgInfoIntrinsic>(Inst)) continue;
199 // If these two calls do not interfere, look past it.
200 switch (AA->getModRefInfo(CS, InstCS)) {
201 case AliasAnalysis::NoModRef:
202 // If the two calls are the same, return InstCS as a Def, so that
203 // CS can be found redundant and eliminated.
204 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
205 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
206 return MemDepResult::getDef(Inst);
208 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
212 return MemDepResult::getClobber(Inst);
217 // No dependence found. If this is the entry block of the function, it is a
218 // clobber, otherwise it is non-local.
219 if (BB != &BB->getParent()->getEntryBlock())
220 return MemDepResult::getNonLocal();
221 return MemDepResult::getClobber(ScanIt);
224 /// getPointerDependencyFrom - Return the instruction on which a memory
225 /// location depends. If isLoad is true, this routine ignores may-aliases with
226 /// read-only operations. If isLoad is false, this routine ignores may-aliases
227 /// with reads from read-only locations.
228 MemDepResult MemoryDependenceAnalysis::
229 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
230 BasicBlock::iterator ScanIt, BasicBlock *BB) {
232 Value *InvariantTag = 0;
234 // Walk backwards through the basic block, looking for dependencies.
235 while (ScanIt != BB->begin()) {
236 Instruction *Inst = --ScanIt;
238 // If we're in an invariant region, no dependencies can be found before
239 // we pass an invariant-begin marker.
240 if (InvariantTag == Inst) {
245 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
246 // Debug intrinsics don't (and can't) cause dependences.
247 if (isa<DbgInfoIntrinsic>(II)) continue;
249 // If we pass an invariant-end marker, then we've just entered an
250 // invariant region and can start ignoring dependencies.
251 if (II->getIntrinsicID() == Intrinsic::invariant_end) {
252 // FIXME: This only considers queries directly on the invariant-tagged
253 // pointer, not on query pointers that are indexed off of them. It'd
254 // be nice to handle that at some point.
255 AliasAnalysis::AliasResult R =
256 AA->alias(AliasAnalysis::Location(II->getArgOperand(2)), MemLoc);
257 if (R == AliasAnalysis::MustAlias)
258 InvariantTag = II->getArgOperand(0);
263 // If we reach a lifetime begin or end marker, then the query ends here
264 // because the value is undefined.
265 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
266 // FIXME: This only considers queries directly on the invariant-tagged
267 // pointer, not on query pointers that are indexed off of them. It'd
268 // be nice to handle that at some point.
269 AliasAnalysis::AliasResult R =
270 AA->alias(AliasAnalysis::Location(II->getArgOperand(1)), MemLoc);
271 if (R == AliasAnalysis::MustAlias)
272 return MemDepResult::getDef(II);
277 // If we're querying on a load and we're in an invariant region, we're done
278 // at this point. Nothing a load depends on can live in an invariant region.
280 // FIXME: this will prevent us from returning load/load must-aliases, so GVN
281 // won't remove redundant loads.
282 if (isLoad && InvariantTag) continue;
284 // Values depend on loads if the pointers are must aliased. This means that
285 // a load depends on another must aliased load from the same value.
286 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
287 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
289 // If we found a pointer, check if it could be the same as our pointer.
290 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
291 if (R == AliasAnalysis::NoAlias)
294 // May-alias loads don't depend on each other without a dependence.
295 if (isLoad && R != AliasAnalysis::MustAlias)
298 // Stores don't alias loads from read-only memory.
299 if (!isLoad && AA->pointsToConstantMemory(LoadLoc))
302 // Stores depend on may and must aliased loads, loads depend on must-alias
304 return MemDepResult::getDef(Inst);
307 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
308 // There can't be stores to the value we care about inside an
310 if (InvariantTag) continue;
312 // If alias analysis can tell that this store is guaranteed to not modify
313 // the query pointer, ignore it. Use getModRefInfo to handle cases where
314 // the query pointer points to constant memory etc.
315 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
318 // Ok, this store might clobber the query pointer. Check to see if it is
319 // a must alias: in this case, we want to return this as a def.
320 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
322 // If we found a pointer, check if it could be the same as our pointer.
323 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
325 if (R == AliasAnalysis::NoAlias)
327 if (R == AliasAnalysis::MustAlias)
328 return MemDepResult::getDef(Inst);
329 return MemDepResult::getClobber(Inst);
332 // If this is an allocation, and if we know that the accessed pointer is to
333 // the allocation, return Def. This means that there is no dependence and
334 // the access can be optimized based on that. For example, a load could
336 // Note: Only determine this to be a malloc if Inst is the malloc call, not
337 // a subsequent bitcast of the malloc call result. There can be stores to
338 // the malloced memory between the malloc call and its bitcast uses, and we
339 // need to continue scanning until the malloc call.
340 if (isa<AllocaInst>(Inst) ||
341 (isa<CallInst>(Inst) && extractMallocCall(Inst))) {
342 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD);
344 if (AccessPtr == Inst ||
345 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
346 return MemDepResult::getDef(Inst);
350 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
351 switch (AA->getModRefInfo(Inst, MemLoc)) {
352 case AliasAnalysis::NoModRef:
353 // If the call has no effect on the queried pointer, just ignore it.
355 case AliasAnalysis::Mod:
356 // If we're in an invariant region, we can ignore calls that ONLY
357 // modify the pointer.
358 if (InvariantTag) continue;
359 return MemDepResult::getClobber(Inst);
360 case AliasAnalysis::Ref:
361 // If the call is known to never store to the pointer, and if this is a
362 // load query, we can safely ignore it (scan past it).
366 // Otherwise, there is a potential dependence. Return a clobber.
367 return MemDepResult::getClobber(Inst);
371 // No dependence found. If this is the entry block of the function, it is a
372 // clobber, otherwise it is non-local.
373 if (BB != &BB->getParent()->getEntryBlock())
374 return MemDepResult::getNonLocal();
375 return MemDepResult::getClobber(ScanIt);
378 /// getDependency - Return the instruction on which a memory operation
380 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
381 Instruction *ScanPos = QueryInst;
383 // Check for a cached result
384 MemDepResult &LocalCache = LocalDeps[QueryInst];
386 // If the cached entry is non-dirty, just return it. Note that this depends
387 // on MemDepResult's default constructing to 'dirty'.
388 if (!LocalCache.isDirty())
391 // Otherwise, if we have a dirty entry, we know we can start the scan at that
392 // instruction, which may save us some work.
393 if (Instruction *Inst = LocalCache.getInst()) {
396 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
399 BasicBlock *QueryParent = QueryInst->getParent();
402 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
403 // No dependence found. If this is the entry block of the function, it is a
404 // clobber, otherwise it is non-local.
405 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
406 LocalCache = MemDepResult::getNonLocal();
408 LocalCache = MemDepResult::getClobber(QueryInst);
410 AliasAnalysis::Location MemLoc;
411 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
413 // If we can do a pointer scan, make it happen.
414 bool isLoad = !(MR & AliasAnalysis::Mod);
415 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
416 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end;
418 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
420 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
421 CallSite QueryCS(QueryInst);
422 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
423 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
426 // Non-memory instruction.
427 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
430 // Remember the result!
431 if (Instruction *I = LocalCache.getInst())
432 ReverseLocalDeps[I].insert(QueryInst);
438 /// AssertSorted - This method is used when -debug is specified to verify that
439 /// cache arrays are properly kept sorted.
440 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
442 if (Count == -1) Count = Cache.size();
443 if (Count == 0) return;
445 for (unsigned i = 1; i != unsigned(Count); ++i)
446 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
450 /// getNonLocalCallDependency - Perform a full dependency query for the
451 /// specified call, returning the set of blocks that the value is
452 /// potentially live across. The returned set of results will include a
453 /// "NonLocal" result for all blocks where the value is live across.
455 /// This method assumes the instruction returns a "NonLocal" dependency
456 /// within its own block.
458 /// This returns a reference to an internal data structure that may be
459 /// invalidated on the next non-local query or when an instruction is
460 /// removed. Clients must copy this data if they want it around longer than
462 const MemoryDependenceAnalysis::NonLocalDepInfo &
463 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
464 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
465 "getNonLocalCallDependency should only be used on calls with non-local deps!");
466 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
467 NonLocalDepInfo &Cache = CacheP.first;
469 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
470 /// the cached case, this can happen due to instructions being deleted etc. In
471 /// the uncached case, this starts out as the set of predecessors we care
473 SmallVector<BasicBlock*, 32> DirtyBlocks;
475 if (!Cache.empty()) {
476 // Okay, we have a cache entry. If we know it is not dirty, just return it
477 // with no computation.
478 if (!CacheP.second) {
483 // If we already have a partially computed set of results, scan them to
484 // determine what is dirty, seeding our initial DirtyBlocks worklist.
485 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
487 if (I->getResult().isDirty())
488 DirtyBlocks.push_back(I->getBB());
490 // Sort the cache so that we can do fast binary search lookups below.
491 std::sort(Cache.begin(), Cache.end());
493 ++NumCacheDirtyNonLocal;
494 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
495 // << Cache.size() << " cached: " << *QueryInst;
497 // Seed DirtyBlocks with each of the preds of QueryInst's block.
498 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
499 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
500 DirtyBlocks.push_back(*PI);
501 ++NumUncacheNonLocal;
504 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
505 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
507 SmallPtrSet<BasicBlock*, 64> Visited;
509 unsigned NumSortedEntries = Cache.size();
510 DEBUG(AssertSorted(Cache));
512 // Iterate while we still have blocks to update.
513 while (!DirtyBlocks.empty()) {
514 BasicBlock *DirtyBB = DirtyBlocks.back();
515 DirtyBlocks.pop_back();
517 // Already processed this block?
518 if (!Visited.insert(DirtyBB))
521 // Do a binary search to see if we already have an entry for this block in
522 // the cache set. If so, find it.
523 DEBUG(AssertSorted(Cache, NumSortedEntries));
524 NonLocalDepInfo::iterator Entry =
525 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
526 NonLocalDepEntry(DirtyBB));
527 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB)
530 NonLocalDepEntry *ExistingResult = 0;
531 if (Entry != Cache.begin()+NumSortedEntries &&
532 Entry->getBB() == DirtyBB) {
533 // If we already have an entry, and if it isn't already dirty, the block
535 if (!Entry->getResult().isDirty())
538 // Otherwise, remember this slot so we can update the value.
539 ExistingResult = &*Entry;
542 // If the dirty entry has a pointer, start scanning from it so we don't have
543 // to rescan the entire block.
544 BasicBlock::iterator ScanPos = DirtyBB->end();
545 if (ExistingResult) {
546 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
548 // We're removing QueryInst's use of Inst.
549 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
550 QueryCS.getInstruction());
554 // Find out if this block has a local dependency for QueryInst.
557 if (ScanPos != DirtyBB->begin()) {
558 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
559 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
560 // No dependence found. If this is the entry block of the function, it is
561 // a clobber, otherwise it is non-local.
562 Dep = MemDepResult::getNonLocal();
564 Dep = MemDepResult::getClobber(ScanPos);
567 // If we had a dirty entry for the block, update it. Otherwise, just add
570 ExistingResult->setResult(Dep);
572 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
574 // If the block has a dependency (i.e. it isn't completely transparent to
575 // the value), remember the association!
576 if (!Dep.isNonLocal()) {
577 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
578 // update this when we remove instructions.
579 if (Instruction *Inst = Dep.getInst())
580 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
583 // If the block *is* completely transparent to the load, we need to check
584 // the predecessors of this block. Add them to our worklist.
585 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
586 DirtyBlocks.push_back(*PI);
593 /// getNonLocalPointerDependency - Perform a full dependency query for an
594 /// access to the specified (non-volatile) memory location, returning the
595 /// set of instructions that either define or clobber the value.
597 /// This method assumes the pointer has a "NonLocal" dependency within its
600 void MemoryDependenceAnalysis::
601 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
603 SmallVectorImpl<NonLocalDepResult> &Result) {
604 assert(Loc.Ptr->getType()->isPointerTy() &&
605 "Can't get pointer deps of a non-pointer!");
608 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD);
610 // This is the set of blocks we've inspected, and the pointer we consider in
611 // each block. Because of critical edges, we currently bail out if querying
612 // a block with multiple different pointers. This can happen during PHI
614 DenseMap<BasicBlock*, Value*> Visited;
615 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
616 Result, Visited, true))
619 Result.push_back(NonLocalDepResult(FromBB,
620 MemDepResult::getClobber(FromBB->begin()),
621 const_cast<Value *>(Loc.Ptr)));
624 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
625 /// Pointer/PointeeSize using either cached information in Cache or by doing a
626 /// lookup (which may use dirty cache info if available). If we do a lookup,
627 /// add the result to the cache.
628 MemDepResult MemoryDependenceAnalysis::
629 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
630 bool isLoad, BasicBlock *BB,
631 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
633 // Do a binary search to see if we already have an entry for this block in
634 // the cache set. If so, find it.
635 NonLocalDepInfo::iterator Entry =
636 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
637 NonLocalDepEntry(BB));
638 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
641 NonLocalDepEntry *ExistingResult = 0;
642 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
643 ExistingResult = &*Entry;
645 // If we have a cached entry, and it is non-dirty, use it as the value for
647 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
648 ++NumCacheNonLocalPtr;
649 return ExistingResult->getResult();
652 // Otherwise, we have to scan for the value. If we have a dirty cache
653 // entry, start scanning from its position, otherwise we scan from the end
655 BasicBlock::iterator ScanPos = BB->end();
656 if (ExistingResult && ExistingResult->getResult().getInst()) {
657 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
658 "Instruction invalidated?");
659 ++NumCacheDirtyNonLocalPtr;
660 ScanPos = ExistingResult->getResult().getInst();
662 // Eliminating the dirty entry from 'Cache', so update the reverse info.
663 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
664 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
666 ++NumUncacheNonLocalPtr;
669 // Scan the block for the dependency.
670 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
672 // If we had a dirty entry for the block, update it. Otherwise, just add
675 ExistingResult->setResult(Dep);
677 Cache->push_back(NonLocalDepEntry(BB, Dep));
679 // If the block has a dependency (i.e. it isn't completely transparent to
680 // the value), remember the reverse association because we just added it
682 if (Dep.isNonLocal())
685 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
686 // update MemDep when we remove instructions.
687 Instruction *Inst = Dep.getInst();
688 assert(Inst && "Didn't depend on anything?");
689 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
690 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
694 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
695 /// number of elements in the array that are already properly ordered. This is
696 /// optimized for the case when only a few entries are added.
698 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
699 unsigned NumSortedEntries) {
700 switch (Cache.size() - NumSortedEntries) {
702 // done, no new entries.
705 // Two new entries, insert the last one into place.
706 NonLocalDepEntry Val = Cache.back();
708 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
709 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
710 Cache.insert(Entry, Val);
714 // One new entry, Just insert the new value at the appropriate position.
715 if (Cache.size() != 1) {
716 NonLocalDepEntry Val = Cache.back();
718 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
719 std::upper_bound(Cache.begin(), Cache.end(), Val);
720 Cache.insert(Entry, Val);
724 // Added many values, do a full scale sort.
725 std::sort(Cache.begin(), Cache.end());
730 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
731 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
732 /// results to the results vector and keep track of which blocks are visited in
735 /// This has special behavior for the first block queries (when SkipFirstBlock
736 /// is true). In this special case, it ignores the contents of the specified
737 /// block and starts returning dependence info for its predecessors.
739 /// This function returns false on success, or true to indicate that it could
740 /// not compute dependence information for some reason. This should be treated
741 /// as a clobber dependence on the first instruction in the predecessor block.
742 bool MemoryDependenceAnalysis::
743 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
744 const AliasAnalysis::Location &Loc,
745 bool isLoad, BasicBlock *StartBB,
746 SmallVectorImpl<NonLocalDepResult> &Result,
747 DenseMap<BasicBlock*, Value*> &Visited,
748 bool SkipFirstBlock) {
750 // Look up the cached info for Pointer.
751 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
753 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
754 // CacheKey, this value will be inserted as the associated value. Otherwise,
755 // it'll be ignored, and we'll have to check to see if the cached size and
756 // tbaa tag are consistent with the current query.
757 NonLocalPointerInfo InitialNLPI;
758 InitialNLPI.Size = Loc.Size;
759 InitialNLPI.TBAATag = Loc.TBAATag;
761 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
763 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
764 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
765 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
767 // If we already have a cache entry for this CacheKey, we may need to do some
768 // work to reconcile the cache entry and the current query.
770 if (CacheInfo->Size < Loc.Size) {
771 // The query's Size is greater than the cached one. Throw out the
772 // cached data and procede with the query at the greater size.
773 CacheInfo->Pair = BBSkipFirstBlockPair();
774 CacheInfo->Size = Loc.Size;
775 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
776 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
777 if (Instruction *Inst = DI->getResult().getInst())
778 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
779 CacheInfo->NonLocalDeps.clear();
780 } else if (CacheInfo->Size > Loc.Size) {
781 // This query's Size is less than the cached one. Conservatively restart
782 // the query using the greater size.
783 return getNonLocalPointerDepFromBB(Pointer,
784 Loc.getWithNewSize(CacheInfo->Size),
785 isLoad, StartBB, Result, Visited,
789 // If the query's TBAATag is inconsistent with the cached one,
790 // conservatively throw out the cached data and restart the query with
792 if (CacheInfo->TBAATag != Loc.TBAATag) {
793 if (CacheInfo->TBAATag) {
794 CacheInfo->Pair = BBSkipFirstBlockPair();
795 CacheInfo->TBAATag = 0;
796 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
797 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
798 if (Instruction *Inst = DI->getResult().getInst())
799 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
800 CacheInfo->NonLocalDeps.clear();
803 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(),
804 isLoad, StartBB, Result, Visited,
809 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
811 // If we have valid cached information for exactly the block we are
812 // investigating, just return it with no recomputation.
813 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
814 // We have a fully cached result for this query then we can just return the
815 // cached results and populate the visited set. However, we have to verify
816 // that we don't already have conflicting results for these blocks. Check
817 // to ensure that if a block in the results set is in the visited set that
818 // it was for the same pointer query.
819 if (!Visited.empty()) {
820 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
822 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
823 if (VI == Visited.end() || VI->second == Pointer.getAddr())
826 // We have a pointer mismatch in a block. Just return clobber, saying
827 // that something was clobbered in this result. We could also do a
828 // non-fully cached query, but there is little point in doing this.
833 Value *Addr = Pointer.getAddr();
834 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
836 Visited.insert(std::make_pair(I->getBB(), Addr));
837 if (!I->getResult().isNonLocal())
838 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
840 ++NumCacheCompleteNonLocalPtr;
844 // Otherwise, either this is a new block, a block with an invalid cache
845 // pointer or one that we're about to invalidate by putting more info into it
846 // than its valid cache info. If empty, the result will be valid cache info,
847 // otherwise it isn't.
849 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
851 CacheInfo->Pair = BBSkipFirstBlockPair();
853 SmallVector<BasicBlock*, 32> Worklist;
854 Worklist.push_back(StartBB);
856 // Keep track of the entries that we know are sorted. Previously cached
857 // entries will all be sorted. The entries we add we only sort on demand (we
858 // don't insert every element into its sorted position). We know that we
859 // won't get any reuse from currently inserted values, because we don't
860 // revisit blocks after we insert info for them.
861 unsigned NumSortedEntries = Cache->size();
862 DEBUG(AssertSorted(*Cache));
864 while (!Worklist.empty()) {
865 BasicBlock *BB = Worklist.pop_back_val();
867 // Skip the first block if we have it.
868 if (!SkipFirstBlock) {
869 // Analyze the dependency of *Pointer in FromBB. See if we already have
871 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
873 // Get the dependency info for Pointer in BB. If we have cached
874 // information, we will use it, otherwise we compute it.
875 DEBUG(AssertSorted(*Cache, NumSortedEntries));
876 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
879 // If we got a Def or Clobber, add this to the list of results.
880 if (!Dep.isNonLocal()) {
881 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
886 // If 'Pointer' is an instruction defined in this block, then we need to do
887 // phi translation to change it into a value live in the predecessor block.
888 // If not, we just add the predecessors to the worklist and scan them with
890 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
891 SkipFirstBlock = false;
892 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
893 // Verify that we haven't looked at this block yet.
894 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
895 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
896 if (InsertRes.second) {
897 // First time we've looked at *PI.
898 Worklist.push_back(*PI);
902 // If we have seen this block before, but it was with a different
903 // pointer then we have a phi translation failure and we have to treat
904 // this as a clobber.
905 if (InsertRes.first->second != Pointer.getAddr())
906 goto PredTranslationFailure;
911 // We do need to do phi translation, if we know ahead of time we can't phi
912 // translate this value, don't even try.
913 if (!Pointer.IsPotentiallyPHITranslatable())
914 goto PredTranslationFailure;
916 // We may have added values to the cache list before this PHI translation.
917 // If so, we haven't done anything to ensure that the cache remains sorted.
918 // Sort it now (if needed) so that recursive invocations of
919 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
920 // value will only see properly sorted cache arrays.
921 if (Cache && NumSortedEntries != Cache->size()) {
922 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
923 NumSortedEntries = Cache->size();
927 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
928 BasicBlock *Pred = *PI;
930 // Get the PHI translated pointer in this predecessor. This can fail if
931 // not translatable, in which case the getAddr() returns null.
932 PHITransAddr PredPointer(Pointer);
933 PredPointer.PHITranslateValue(BB, Pred, 0);
935 Value *PredPtrVal = PredPointer.getAddr();
937 // Check to see if we have already visited this pred block with another
938 // pointer. If so, we can't do this lookup. This failure can occur
939 // with PHI translation when a critical edge exists and the PHI node in
940 // the successor translates to a pointer value different than the
941 // pointer the block was first analyzed with.
942 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
943 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
945 if (!InsertRes.second) {
946 // If the predecessor was visited with PredPtr, then we already did
947 // the analysis and can ignore it.
948 if (InsertRes.first->second == PredPtrVal)
951 // Otherwise, the block was previously analyzed with a different
952 // pointer. We can't represent the result of this case, so we just
953 // treat this as a phi translation failure.
954 goto PredTranslationFailure;
957 // If PHI translation was unable to find an available pointer in this
958 // predecessor, then we have to assume that the pointer is clobbered in
959 // that predecessor. We can still do PRE of the load, which would insert
960 // a computation of the pointer in this predecessor.
961 if (PredPtrVal == 0) {
962 // Add the entry to the Result list.
963 NonLocalDepResult Entry(Pred,
964 MemDepResult::getClobber(Pred->getTerminator()),
966 Result.push_back(Entry);
968 // Since we had a phi translation failure, the cache for CacheKey won't
969 // include all of the entries that we need to immediately satisfy future
970 // queries. Mark this in NonLocalPointerDeps by setting the
971 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
972 // cached value to do more work but not miss the phi trans failure.
973 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
974 NLPI.Pair = BBSkipFirstBlockPair();
978 // FIXME: it is entirely possible that PHI translating will end up with
979 // the same value. Consider PHI translating something like:
980 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
981 // to recurse here, pedantically speaking.
983 // If we have a problem phi translating, fall through to the code below
984 // to handle the failure condition.
985 if (getNonLocalPointerDepFromBB(PredPointer,
986 Loc.getWithNewPtr(PredPointer.getAddr()),
989 goto PredTranslationFailure;
992 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
993 CacheInfo = &NonLocalPointerDeps[CacheKey];
994 Cache = &CacheInfo->NonLocalDeps;
995 NumSortedEntries = Cache->size();
997 // Since we did phi translation, the "Cache" set won't contain all of the
998 // results for the query. This is ok (we can still use it to accelerate
999 // specific block queries) but we can't do the fastpath "return all
1000 // results from the set" Clear out the indicator for this.
1001 CacheInfo->Pair = BBSkipFirstBlockPair();
1002 SkipFirstBlock = false;
1005 PredTranslationFailure:
1008 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1009 CacheInfo = &NonLocalPointerDeps[CacheKey];
1010 Cache = &CacheInfo->NonLocalDeps;
1011 NumSortedEntries = Cache->size();
1014 // Since we failed phi translation, the "Cache" set won't contain all of the
1015 // results for the query. This is ok (we can still use it to accelerate
1016 // specific block queries) but we can't do the fastpath "return all
1017 // results from the set". Clear out the indicator for this.
1018 CacheInfo->Pair = BBSkipFirstBlockPair();
1020 // If *nothing* works, mark the pointer as being clobbered by the first
1021 // instruction in this block.
1023 // If this is the magic first block, return this as a clobber of the whole
1024 // incoming value. Since we can't phi translate to one of the predecessors,
1025 // we have to bail out.
1029 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1030 assert(I != Cache->rend() && "Didn't find current block??");
1031 if (I->getBB() != BB)
1034 assert(I->getResult().isNonLocal() &&
1035 "Should only be here with transparent block");
1036 I->setResult(MemDepResult::getClobber(BB->begin()));
1037 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
1038 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1039 Pointer.getAddr()));
1044 // Okay, we're done now. If we added new values to the cache, re-sort it.
1045 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1046 DEBUG(AssertSorted(*Cache));
1050 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1051 /// CachedNonLocalPointerInfo, remove it.
1052 void MemoryDependenceAnalysis::
1053 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1054 CachedNonLocalPointerInfo::iterator It =
1055 NonLocalPointerDeps.find(P);
1056 if (It == NonLocalPointerDeps.end()) return;
1058 // Remove all of the entries in the BB->val map. This involves removing
1059 // instructions from the reverse map.
1060 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1062 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1063 Instruction *Target = PInfo[i].getResult().getInst();
1064 if (Target == 0) continue; // Ignore non-local dep results.
1065 assert(Target->getParent() == PInfo[i].getBB());
1067 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1068 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1071 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1072 NonLocalPointerDeps.erase(It);
1076 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1077 /// information about the specified pointer, because it may be too
1078 /// conservative in memdep. This is an optional call that can be used when
1079 /// the client detects an equivalence between the pointer and some other
1080 /// value and replaces the other value with ptr. This can make Ptr available
1081 /// in more places that cached info does not necessarily keep.
1082 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1083 // If Ptr isn't really a pointer, just ignore it.
1084 if (!Ptr->getType()->isPointerTy()) return;
1085 // Flush store info for the pointer.
1086 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1087 // Flush load info for the pointer.
1088 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1091 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1092 /// This needs to be done when the CFG changes, e.g., due to splitting
1094 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1098 /// removeInstruction - Remove an instruction from the dependence analysis,
1099 /// updating the dependence of instructions that previously depended on it.
1100 /// This method attempts to keep the cache coherent using the reverse map.
1101 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1102 // Walk through the Non-local dependencies, removing this one as the value
1103 // for any cached queries.
1104 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1105 if (NLDI != NonLocalDeps.end()) {
1106 NonLocalDepInfo &BlockMap = NLDI->second.first;
1107 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1109 if (Instruction *Inst = DI->getResult().getInst())
1110 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1111 NonLocalDeps.erase(NLDI);
1114 // If we have a cached local dependence query for this instruction, remove it.
1116 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1117 if (LocalDepEntry != LocalDeps.end()) {
1118 // Remove us from DepInst's reverse set now that the local dep info is gone.
1119 if (Instruction *Inst = LocalDepEntry->second.getInst())
1120 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1122 // Remove this local dependency info.
1123 LocalDeps.erase(LocalDepEntry);
1126 // If we have any cached pointer dependencies on this instruction, remove
1127 // them. If the instruction has non-pointer type, then it can't be a pointer
1130 // Remove it from both the load info and the store info. The instruction
1131 // can't be in either of these maps if it is non-pointer.
1132 if (RemInst->getType()->isPointerTy()) {
1133 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1134 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1137 // Loop over all of the things that depend on the instruction we're removing.
1139 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1141 // If we find RemInst as a clobber or Def in any of the maps for other values,
1142 // we need to replace its entry with a dirty version of the instruction after
1143 // it. If RemInst is a terminator, we use a null dirty value.
1145 // Using a dirty version of the instruction after RemInst saves having to scan
1146 // the entire block to get to this point.
1147 MemDepResult NewDirtyVal;
1148 if (!RemInst->isTerminator())
1149 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1151 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1152 if (ReverseDepIt != ReverseLocalDeps.end()) {
1153 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1154 // RemInst can't be the terminator if it has local stuff depending on it.
1155 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1156 "Nothing can locally depend on a terminator");
1158 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1159 E = ReverseDeps.end(); I != E; ++I) {
1160 Instruction *InstDependingOnRemInst = *I;
1161 assert(InstDependingOnRemInst != RemInst &&
1162 "Already removed our local dep info");
1164 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1166 // Make sure to remember that new things depend on NewDepInst.
1167 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1168 "a local dep on this if it is a terminator!");
1169 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1170 InstDependingOnRemInst));
1173 ReverseLocalDeps.erase(ReverseDepIt);
1175 // Add new reverse deps after scanning the set, to avoid invalidating the
1176 // 'ReverseDeps' reference.
1177 while (!ReverseDepsToAdd.empty()) {
1178 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1179 .insert(ReverseDepsToAdd.back().second);
1180 ReverseDepsToAdd.pop_back();
1184 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1185 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1186 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1187 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1189 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1191 PerInstNLInfo &INLD = NonLocalDeps[*I];
1192 // The information is now dirty!
1195 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1196 DE = INLD.first.end(); DI != DE; ++DI) {
1197 if (DI->getResult().getInst() != RemInst) continue;
1199 // Convert to a dirty entry for the subsequent instruction.
1200 DI->setResult(NewDirtyVal);
1202 if (Instruction *NextI = NewDirtyVal.getInst())
1203 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1207 ReverseNonLocalDeps.erase(ReverseDepIt);
1209 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1210 while (!ReverseDepsToAdd.empty()) {
1211 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1212 .insert(ReverseDepsToAdd.back().second);
1213 ReverseDepsToAdd.pop_back();
1217 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1218 // value in the NonLocalPointerDeps info.
1219 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1220 ReverseNonLocalPtrDeps.find(RemInst);
1221 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1222 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1223 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1225 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1226 E = Set.end(); I != E; ++I) {
1227 ValueIsLoadPair P = *I;
1228 assert(P.getPointer() != RemInst &&
1229 "Already removed NonLocalPointerDeps info for RemInst");
1231 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1233 // The cache is not valid for any specific block anymore.
1234 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1236 // Update any entries for RemInst to use the instruction after it.
1237 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1239 if (DI->getResult().getInst() != RemInst) continue;
1241 // Convert to a dirty entry for the subsequent instruction.
1242 DI->setResult(NewDirtyVal);
1244 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1245 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1248 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1249 // subsequent value may invalidate the sortedness.
1250 std::sort(NLPDI.begin(), NLPDI.end());
1253 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1255 while (!ReversePtrDepsToAdd.empty()) {
1256 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1257 .insert(ReversePtrDepsToAdd.back().second);
1258 ReversePtrDepsToAdd.pop_back();
1263 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1264 AA->deleteValue(RemInst);
1265 DEBUG(verifyRemoved(RemInst));
1267 /// verifyRemoved - Verify that the specified instruction does not occur
1268 /// in our internal data structures.
1269 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1270 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1271 E = LocalDeps.end(); I != E; ++I) {
1272 assert(I->first != D && "Inst occurs in data structures");
1273 assert(I->second.getInst() != D &&
1274 "Inst occurs in data structures");
1277 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1278 E = NonLocalPointerDeps.end(); I != E; ++I) {
1279 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1280 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1281 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1283 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1286 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1287 E = NonLocalDeps.end(); I != E; ++I) {
1288 assert(I->first != D && "Inst occurs in data structures");
1289 const PerInstNLInfo &INLD = I->second;
1290 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1291 EE = INLD.first.end(); II != EE; ++II)
1292 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1295 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1296 E = ReverseLocalDeps.end(); I != E; ++I) {
1297 assert(I->first != D && "Inst occurs in data structures");
1298 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1299 EE = I->second.end(); II != EE; ++II)
1300 assert(*II != D && "Inst occurs in data structures");
1303 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1304 E = ReverseNonLocalDeps.end();
1306 assert(I->first != D && "Inst occurs in data structures");
1307 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1308 EE = I->second.end(); II != EE; ++II)
1309 assert(*II != D && "Inst occurs in data structures");
1312 for (ReverseNonLocalPtrDepTy::const_iterator
1313 I = ReverseNonLocalPtrDeps.begin(),
1314 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1315 assert(I->first != D && "Inst occurs in rev NLPD map");
1317 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1318 E = I->second.end(); II != E; ++II)
1319 assert(*II != ValueIsLoadPair(D, false) &&
1320 *II != ValueIsLoadPair(D, true) &&
1321 "Inst occurs in ReverseNonLocalPtrDeps map");