1 //===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation --*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements an analysis that determines, for a given memory
11 // operation, what preceding memory operations it depends on. It builds on
12 // alias analysis information, and tries to provide a lazy, caching interface to
13 // a common kind of alias information query.
15 //===----------------------------------------------------------------------===//
17 #define DEBUG_TYPE "memdep"
18 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/IntrinsicInst.h"
21 #include "llvm/Function.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/Dominators.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/MemoryBuiltins.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/Support/PredIteratorCache.h"
29 #include "llvm/Support/Debug.h"
32 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
33 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
34 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
36 STATISTIC(NumCacheNonLocalPtr,
37 "Number of fully cached non-local ptr responses");
38 STATISTIC(NumCacheDirtyNonLocalPtr,
39 "Number of cached, but dirty, non-local ptr responses");
40 STATISTIC(NumUncacheNonLocalPtr,
41 "Number of uncached non-local ptr responses");
42 STATISTIC(NumCacheCompleteNonLocalPtr,
43 "Number of block queries that were completely cached");
45 char MemoryDependenceAnalysis::ID = 0;
47 // Register this pass...
48 static RegisterPass<MemoryDependenceAnalysis> X("memdep",
49 "Memory Dependence Analysis", false, true);
51 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
52 : FunctionPass(&ID), PredCache(0) {
54 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
57 /// Clean up memory in between runs
58 void MemoryDependenceAnalysis::releaseMemory() {
61 NonLocalPointerDeps.clear();
62 ReverseLocalDeps.clear();
63 ReverseNonLocalDeps.clear();
64 ReverseNonLocalPtrDeps.clear();
70 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
72 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
74 AU.addRequiredTransitive<AliasAnalysis>();
77 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
78 AA = &getAnalysis<AliasAnalysis>();
80 PredCache.reset(new PredIteratorCache());
84 /// RemoveFromReverseMap - This is a helper function that removes Val from
85 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
86 template <typename KeyTy>
87 static void RemoveFromReverseMap(DenseMap<Instruction*,
88 SmallPtrSet<KeyTy, 4> > &ReverseMap,
89 Instruction *Inst, KeyTy Val) {
90 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
91 InstIt = ReverseMap.find(Inst);
92 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
93 bool Found = InstIt->second.erase(Val);
94 assert(Found && "Invalid reverse map!"); Found=Found;
95 if (InstIt->second.empty())
96 ReverseMap.erase(InstIt);
100 /// getCallSiteDependencyFrom - Private helper for finding the local
101 /// dependencies of a call site.
102 MemDepResult MemoryDependenceAnalysis::
103 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
104 BasicBlock::iterator ScanIt, BasicBlock *BB) {
105 // Walk backwards through the block, looking for dependencies
106 while (ScanIt != BB->begin()) {
107 Instruction *Inst = --ScanIt;
109 // If this inst is a memory op, get the pointer it accessed
111 uint64_t PointerSize = 0;
112 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
113 Pointer = S->getPointerOperand();
114 PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType());
115 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
116 Pointer = V->getOperand(0);
117 PointerSize = AA->getTypeStoreSize(V->getType());
118 } else if (isFreeCall(Inst)) {
119 Pointer = Inst->getOperand(1);
120 // calls to free() erase the entire structure
122 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
123 // Debug intrinsics don't cause dependences.
124 if (isa<DbgInfoIntrinsic>(Inst)) continue;
125 CallSite InstCS = CallSite::get(Inst);
126 // If these two calls do not interfere, look past it.
127 switch (AA->getModRefInfo(CS, InstCS)) {
128 case AliasAnalysis::NoModRef:
129 // If the two calls don't interact (e.g. InstCS is readnone) keep
132 case AliasAnalysis::Ref:
133 // If the two calls read the same memory locations and CS is a readonly
134 // function, then we have two cases: 1) the calls may not interfere with
135 // each other at all. 2) the calls may produce the same value. In case
136 // #1 we want to ignore the values, in case #2, we want to return Inst
137 // as a Def dependence. This allows us to CSE in cases like:
140 // Y = strlen(P); // Y = X
141 if (isReadOnlyCall) {
142 if (CS.getCalledFunction() != 0 &&
143 CS.getCalledFunction() == InstCS.getCalledFunction())
144 return MemDepResult::getDef(Inst);
145 // Ignore unrelated read/read call dependences.
150 return MemDepResult::getClobber(Inst);
153 // Non-memory instruction.
157 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef)
158 return MemDepResult::getClobber(Inst);
161 // No dependence found. If this is the entry block of the function, it is a
162 // clobber, otherwise it is non-local.
163 if (BB != &BB->getParent()->getEntryBlock())
164 return MemDepResult::getNonLocal();
165 return MemDepResult::getClobber(ScanIt);
168 /// getPointerDependencyFrom - Return the instruction on which a memory
169 /// location depends. If isLoad is true, this routine ignore may-aliases with
170 /// read-only operations.
171 MemDepResult MemoryDependenceAnalysis::
172 getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
173 BasicBlock::iterator ScanIt, BasicBlock *BB) {
175 Value *InvariantTag = 0;
177 // Walk backwards through the basic block, looking for dependencies.
178 while (ScanIt != BB->begin()) {
179 Instruction *Inst = --ScanIt;
181 // If we're in an invariant region, no dependencies can be found before
182 // we pass an invariant-begin marker.
183 if (InvariantTag == Inst) {
188 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
189 // Debug intrinsics don't cause dependences.
190 if (isa<DbgInfoIntrinsic>(Inst)) continue;
192 // If we pass an invariant-end marker, then we've just entered an
193 // invariant region and can start ignoring dependencies.
194 if (II->getIntrinsicID() == Intrinsic::invariant_end) {
195 // FIXME: This only considers queries directly on the invariant-tagged
196 // pointer, not on query pointers that are indexed off of them. It'd
197 // be nice to handle that at some point.
198 AliasAnalysis::AliasResult R =
199 AA->alias(II->getOperand(3), ~0U, MemPtr, ~0U);
200 if (R == AliasAnalysis::MustAlias) {
201 InvariantTag = II->getOperand(1);
205 // If we reach a lifetime begin or end marker, then the query ends here
206 // because the value is undefined.
207 } else if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
208 // FIXME: This only considers queries directly on the invariant-tagged
209 // pointer, not on query pointers that are indexed off of them. It'd
210 // be nice to handle that at some point.
211 AliasAnalysis::AliasResult R =
212 AA->alias(II->getOperand(2), ~0U, MemPtr, ~0U);
213 if (R == AliasAnalysis::MustAlias)
214 return MemDepResult::getDef(II);
218 // If we're querying on a load and we're in an invariant region, we're done
219 // at this point. Nothing a load depends on can live in an invariant region.
220 if (isLoad && InvariantTag) continue;
222 // Values depend on loads if the pointers are must aliased. This means that
223 // a load depends on another must aliased load from the same value.
224 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
225 Value *Pointer = LI->getPointerOperand();
226 uint64_t PointerSize = AA->getTypeStoreSize(LI->getType());
228 // If we found a pointer, check if it could be the same as our pointer.
229 AliasAnalysis::AliasResult R =
230 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
231 if (R == AliasAnalysis::NoAlias)
234 // May-alias loads don't depend on each other without a dependence.
235 if (isLoad && R == AliasAnalysis::MayAlias)
237 // Stores depend on may and must aliased loads, loads depend on must-alias
239 return MemDepResult::getDef(Inst);
242 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
243 // There can't be stores to the value we care about inside an
245 if (InvariantTag) continue;
247 // If alias analysis can tell that this store is guaranteed to not modify
248 // the query pointer, ignore it. Use getModRefInfo to handle cases where
249 // the query pointer points to constant memory etc.
250 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef)
253 // Ok, this store might clobber the query pointer. Check to see if it is
254 // a must alias: in this case, we want to return this as a def.
255 Value *Pointer = SI->getPointerOperand();
256 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
258 // If we found a pointer, check if it could be the same as our pointer.
259 AliasAnalysis::AliasResult R =
260 AA->alias(Pointer, PointerSize, MemPtr, MemSize);
262 if (R == AliasAnalysis::NoAlias)
264 if (R == AliasAnalysis::MayAlias)
265 return MemDepResult::getClobber(Inst);
266 return MemDepResult::getDef(Inst);
269 // If this is an allocation, and if we know that the accessed pointer is to
270 // the allocation, return Def. This means that there is no dependence and
271 // the access can be optimized based on that. For example, a load could
273 // Note: Only determine this to be a malloc if Inst is the malloc call, not
274 // a subsequent bitcast of the malloc call result. There can be stores to
275 // the malloced memory between the malloc call and its bitcast uses, and we
276 // need to continue scanning until the malloc call.
277 if (isa<AllocaInst>(Inst) || extractMallocCall(Inst)) {
278 Value *AccessPtr = MemPtr->getUnderlyingObject();
280 if (AccessPtr == Inst ||
281 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
282 return MemDepResult::getDef(Inst);
286 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
287 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
288 case AliasAnalysis::NoModRef:
289 // If the call has no effect on the queried pointer, just ignore it.
291 case AliasAnalysis::Mod:
292 // If we're in an invariant region, we can ignore calls that ONLY
293 // modify the pointer.
294 if (InvariantTag) continue;
295 return MemDepResult::getClobber(Inst);
296 case AliasAnalysis::Ref:
297 // If the call is known to never store to the pointer, and if this is a
298 // load query, we can safely ignore it (scan past it).
302 // Otherwise, there is a potential dependence. Return a clobber.
303 return MemDepResult::getClobber(Inst);
307 // No dependence found. If this is the entry block of the function, it is a
308 // clobber, otherwise it is non-local.
309 if (BB != &BB->getParent()->getEntryBlock())
310 return MemDepResult::getNonLocal();
311 return MemDepResult::getClobber(ScanIt);
314 /// getDependency - Return the instruction on which a memory operation
316 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
317 Instruction *ScanPos = QueryInst;
319 // Check for a cached result
320 MemDepResult &LocalCache = LocalDeps[QueryInst];
322 // If the cached entry is non-dirty, just return it. Note that this depends
323 // on MemDepResult's default constructing to 'dirty'.
324 if (!LocalCache.isDirty())
327 // Otherwise, if we have a dirty entry, we know we can start the scan at that
328 // instruction, which may save us some work.
329 if (Instruction *Inst = LocalCache.getInst()) {
332 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
335 BasicBlock *QueryParent = QueryInst->getParent();
338 uint64_t MemSize = 0;
341 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
342 // No dependence found. If this is the entry block of the function, it is a
343 // clobber, otherwise it is non-local.
344 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
345 LocalCache = MemDepResult::getNonLocal();
347 LocalCache = MemDepResult::getClobber(QueryInst);
348 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
349 // If this is a volatile store, don't mess around with it. Just return the
350 // previous instruction as a clobber.
351 if (SI->isVolatile())
352 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
354 MemPtr = SI->getPointerOperand();
355 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
357 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
358 // If this is a volatile load, don't mess around with it. Just return the
359 // previous instruction as a clobber.
360 if (LI->isVolatile())
361 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
363 MemPtr = LI->getPointerOperand();
364 MemSize = AA->getTypeStoreSize(LI->getType());
366 } else if (isFreeCall(QueryInst)) {
367 MemPtr = QueryInst->getOperand(1);
368 // calls to free() erase the entire structure, not just a field.
370 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
371 int IntrinsicID = 0; // Intrinsic IDs start at 1.
372 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
373 IntrinsicID = II->getIntrinsicID();
375 switch (IntrinsicID) {
376 case Intrinsic::lifetime_start:
377 case Intrinsic::lifetime_end:
378 case Intrinsic::invariant_start:
379 MemPtr = QueryInst->getOperand(2);
380 MemSize = cast<ConstantInt>(QueryInst->getOperand(1))->getZExtValue();
382 case Intrinsic::invariant_end:
383 MemPtr = QueryInst->getOperand(3);
384 MemSize = cast<ConstantInt>(QueryInst->getOperand(2))->getZExtValue();
387 CallSite QueryCS = CallSite::get(QueryInst);
388 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
389 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
394 // Non-memory instruction.
395 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
398 // If we need to do a pointer scan, make it happen.
400 bool isLoad = !QueryInst->mayWriteToMemory();
401 if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) {
402 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end;
404 LocalCache = getPointerDependencyFrom(MemPtr, MemSize, isLoad, ScanPos,
408 // Remember the result!
409 if (Instruction *I = LocalCache.getInst())
410 ReverseLocalDeps[I].insert(QueryInst);
416 /// AssertSorted - This method is used when -debug is specified to verify that
417 /// cache arrays are properly kept sorted.
418 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
420 if (Count == -1) Count = Cache.size();
421 if (Count == 0) return;
423 for (unsigned i = 1; i != unsigned(Count); ++i)
424 assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!");
428 /// getNonLocalCallDependency - Perform a full dependency query for the
429 /// specified call, returning the set of blocks that the value is
430 /// potentially live across. The returned set of results will include a
431 /// "NonLocal" result for all blocks where the value is live across.
433 /// This method assumes the instruction returns a "NonLocal" dependency
434 /// within its own block.
436 /// This returns a reference to an internal data structure that may be
437 /// invalidated on the next non-local query or when an instruction is
438 /// removed. Clients must copy this data if they want it around longer than
440 const MemoryDependenceAnalysis::NonLocalDepInfo &
441 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
442 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
443 "getNonLocalCallDependency should only be used on calls with non-local deps!");
444 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
445 NonLocalDepInfo &Cache = CacheP.first;
447 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
448 /// the cached case, this can happen due to instructions being deleted etc. In
449 /// the uncached case, this starts out as the set of predecessors we care
451 SmallVector<BasicBlock*, 32> DirtyBlocks;
453 if (!Cache.empty()) {
454 // Okay, we have a cache entry. If we know it is not dirty, just return it
455 // with no computation.
456 if (!CacheP.second) {
461 // If we already have a partially computed set of results, scan them to
462 // determine what is dirty, seeding our initial DirtyBlocks worklist.
463 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
465 if (I->second.isDirty())
466 DirtyBlocks.push_back(I->first);
468 // Sort the cache so that we can do fast binary search lookups below.
469 std::sort(Cache.begin(), Cache.end());
471 ++NumCacheDirtyNonLocal;
472 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
473 // << Cache.size() << " cached: " << *QueryInst;
475 // Seed DirtyBlocks with each of the preds of QueryInst's block.
476 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
477 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
478 DirtyBlocks.push_back(*PI);
479 NumUncacheNonLocal++;
482 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
483 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
485 SmallPtrSet<BasicBlock*, 64> Visited;
487 unsigned NumSortedEntries = Cache.size();
488 DEBUG(AssertSorted(Cache));
490 // Iterate while we still have blocks to update.
491 while (!DirtyBlocks.empty()) {
492 BasicBlock *DirtyBB = DirtyBlocks.back();
493 DirtyBlocks.pop_back();
495 // Already processed this block?
496 if (!Visited.insert(DirtyBB))
499 // Do a binary search to see if we already have an entry for this block in
500 // the cache set. If so, find it.
501 DEBUG(AssertSorted(Cache, NumSortedEntries));
502 NonLocalDepInfo::iterator Entry =
503 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
504 std::make_pair(DirtyBB, MemDepResult()));
505 if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB)
508 MemDepResult *ExistingResult = 0;
509 if (Entry != Cache.begin()+NumSortedEntries &&
510 Entry->first == DirtyBB) {
511 // If we already have an entry, and if it isn't already dirty, the block
513 if (!Entry->second.isDirty())
516 // Otherwise, remember this slot so we can update the value.
517 ExistingResult = &Entry->second;
520 // If the dirty entry has a pointer, start scanning from it so we don't have
521 // to rescan the entire block.
522 BasicBlock::iterator ScanPos = DirtyBB->end();
523 if (ExistingResult) {
524 if (Instruction *Inst = ExistingResult->getInst()) {
526 // We're removing QueryInst's use of Inst.
527 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
528 QueryCS.getInstruction());
532 // Find out if this block has a local dependency for QueryInst.
535 if (ScanPos != DirtyBB->begin()) {
536 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
537 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
538 // No dependence found. If this is the entry block of the function, it is
539 // a clobber, otherwise it is non-local.
540 Dep = MemDepResult::getNonLocal();
542 Dep = MemDepResult::getClobber(ScanPos);
545 // If we had a dirty entry for the block, update it. Otherwise, just add
548 *ExistingResult = Dep;
550 Cache.push_back(std::make_pair(DirtyBB, Dep));
552 // If the block has a dependency (i.e. it isn't completely transparent to
553 // the value), remember the association!
554 if (!Dep.isNonLocal()) {
555 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
556 // update this when we remove instructions.
557 if (Instruction *Inst = Dep.getInst())
558 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
561 // If the block *is* completely transparent to the load, we need to check
562 // the predecessors of this block. Add them to our worklist.
563 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
564 DirtyBlocks.push_back(*PI);
571 /// getNonLocalPointerDependency - Perform a full dependency query for an
572 /// access to the specified (non-volatile) memory location, returning the
573 /// set of instructions that either define or clobber the value.
575 /// This method assumes the pointer has a "NonLocal" dependency within its
578 void MemoryDependenceAnalysis::
579 getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
580 SmallVectorImpl<NonLocalDepEntry> &Result) {
581 assert(isa<PointerType>(Pointer->getType()) &&
582 "Can't get pointer deps of a non-pointer!");
585 // We know that the pointer value is live into FromBB find the def/clobbers
586 // from presecessors.
587 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
588 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy);
590 // This is the set of blocks we've inspected, and the pointer we consider in
591 // each block. Because of critical edges, we currently bail out if querying
592 // a block with multiple different pointers. This can happen during PHI
594 DenseMap<BasicBlock*, Value*> Visited;
595 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB,
596 Result, Visited, true))
599 Result.push_back(std::make_pair(FromBB,
600 MemDepResult::getClobber(FromBB->begin())));
603 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
604 /// Pointer/PointeeSize using either cached information in Cache or by doing a
605 /// lookup (which may use dirty cache info if available). If we do a lookup,
606 /// add the result to the cache.
607 MemDepResult MemoryDependenceAnalysis::
608 GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
609 bool isLoad, BasicBlock *BB,
610 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
612 // Do a binary search to see if we already have an entry for this block in
613 // the cache set. If so, find it.
614 NonLocalDepInfo::iterator Entry =
615 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
616 std::make_pair(BB, MemDepResult()));
617 if (Entry != Cache->begin() && prior(Entry)->first == BB)
620 MemDepResult *ExistingResult = 0;
621 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
622 ExistingResult = &Entry->second;
624 // If we have a cached entry, and it is non-dirty, use it as the value for
626 if (ExistingResult && !ExistingResult->isDirty()) {
627 ++NumCacheNonLocalPtr;
628 return *ExistingResult;
631 // Otherwise, we have to scan for the value. If we have a dirty cache
632 // entry, start scanning from its position, otherwise we scan from the end
634 BasicBlock::iterator ScanPos = BB->end();
635 if (ExistingResult && ExistingResult->getInst()) {
636 assert(ExistingResult->getInst()->getParent() == BB &&
637 "Instruction invalidated?");
638 ++NumCacheDirtyNonLocalPtr;
639 ScanPos = ExistingResult->getInst();
641 // Eliminating the dirty entry from 'Cache', so update the reverse info.
642 ValueIsLoadPair CacheKey(Pointer, isLoad);
643 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
645 ++NumUncacheNonLocalPtr;
648 // Scan the block for the dependency.
649 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad,
652 // If we had a dirty entry for the block, update it. Otherwise, just add
655 *ExistingResult = Dep;
657 Cache->push_back(std::make_pair(BB, Dep));
659 // If the block has a dependency (i.e. it isn't completely transparent to
660 // the value), remember the reverse association because we just added it
662 if (Dep.isNonLocal())
665 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
666 // update MemDep when we remove instructions.
667 Instruction *Inst = Dep.getInst();
668 assert(Inst && "Didn't depend on anything?");
669 ValueIsLoadPair CacheKey(Pointer, isLoad);
670 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
674 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
675 /// number of elements in the array that are already properly ordered. This is
676 /// optimized for the case when only a few entries are added.
678 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
679 unsigned NumSortedEntries) {
680 switch (Cache.size() - NumSortedEntries) {
682 // done, no new entries.
685 // Two new entries, insert the last one into place.
686 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
688 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
689 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
690 Cache.insert(Entry, Val);
694 // One new entry, Just insert the new value at the appropriate position.
695 if (Cache.size() != 1) {
696 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
698 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
699 std::upper_bound(Cache.begin(), Cache.end(), Val);
700 Cache.insert(Entry, Val);
704 // Added many values, do a full scale sort.
705 std::sort(Cache.begin(), Cache.end());
710 /// isPHITranslatable - Return true if the specified computation is derived from
711 /// a PHI node in the current block and if it is simple enough for us to handle.
712 static bool isPHITranslatable(Instruction *Inst) {
713 if (isa<PHINode>(Inst))
716 // We can handle bitcast of a PHI, but the PHI needs to be in the same block
718 if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) {
719 Instruction *OpI = dyn_cast<Instruction>(BC->getOperand(0));
720 if (OpI == 0 || OpI->getParent() != Inst->getParent())
722 return isPHITranslatable(OpI);
725 // We can translate a GEP if all of its operands defined in this block are phi
727 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
728 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) {
729 Instruction *OpI = dyn_cast<Instruction>(GEP->getOperand(i));
730 if (OpI == 0 || OpI->getParent() != Inst->getParent())
733 if (!isPHITranslatable(OpI))
739 if (Inst->getOpcode() == Instruction::Add &&
740 isa<ConstantInt>(Inst->getOperand(1))) {
741 Instruction *OpI = dyn_cast<Instruction>(Inst->getOperand(0));
742 if (OpI == 0 || OpI->getParent() != Inst->getParent())
744 return isPHITranslatable(OpI);
747 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
748 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
749 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
754 /// GetPHITranslatedValue - Given a computation that satisfied the
755 /// isPHITranslatable predicate, see if we can translate the computation into
756 /// the specified predecessor block. If so, return that value.
757 Value *MemoryDependenceAnalysis::
758 GetPHITranslatedValue(Value *InVal, BasicBlock *CurBB, BasicBlock *Pred,
759 const TargetData *TD) const {
760 // If the input value is not an instruction, or if it is not defined in CurBB,
761 // then we don't need to phi translate it.
762 Instruction *Inst = dyn_cast<Instruction>(InVal);
763 if (Inst == 0 || Inst->getParent() != CurBB)
766 if (PHINode *PN = dyn_cast<PHINode>(Inst))
767 return PN->getIncomingValueForBlock(Pred);
769 // Handle bitcast of PHI.
770 if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) {
771 // PHI translate the input operand.
772 Value *PHIIn = GetPHITranslatedValue(BC->getOperand(0), CurBB, Pred, TD);
773 if (PHIIn == 0) return 0;
775 // Constants are trivial to phi translate.
776 if (Constant *C = dyn_cast<Constant>(PHIIn))
777 return ConstantExpr::getBitCast(C, BC->getType());
779 // Otherwise we have to see if a bitcasted version of the incoming pointer
780 // is available. If so, we can use it, otherwise we have to fail.
781 for (Value::use_iterator UI = PHIIn->use_begin(), E = PHIIn->use_end();
783 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI))
784 if (BCI->getType() == BC->getType())
790 // Handle getelementptr with at least one PHI translatable operand.
791 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
792 SmallVector<Value*, 8> GEPOps;
793 BasicBlock *CurBB = GEP->getParent();
794 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) {
795 Value *GEPOp = GEP->getOperand(i);
796 // No PHI translation is needed of operands whose values are live in to
797 // the predecessor block.
798 if (!isa<Instruction>(GEPOp) ||
799 cast<Instruction>(GEPOp)->getParent() != CurBB) {
800 GEPOps.push_back(GEPOp);
804 // If the operand is a phi node, do phi translation.
805 Value *InOp = GetPHITranslatedValue(GEPOp, CurBB, Pred, TD);
806 if (InOp == 0) return 0;
808 GEPOps.push_back(InOp);
811 // Simplify the GEP to handle 'gep x, 0' -> x etc.
812 if (Value *V = SimplifyGEPInst(&GEPOps[0], GEPOps.size(), TD))
815 // Scan to see if we have this GEP available.
816 Value *APHIOp = GEPOps[0];
817 for (Value::use_iterator UI = APHIOp->use_begin(), E = APHIOp->use_end();
819 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI))
820 if (GEPI->getType() == GEP->getType() &&
821 GEPI->getNumOperands() == GEPOps.size() &&
822 GEPI->getParent()->getParent() == CurBB->getParent()) {
823 bool Mismatch = false;
824 for (unsigned i = 0, e = GEPOps.size(); i != e; ++i)
825 if (GEPI->getOperand(i) != GEPOps[i]) {
836 // Handle add with a constant RHS.
837 if (Inst->getOpcode() == Instruction::Add &&
838 isa<ConstantInt>(Inst->getOperand(1))) {
839 // PHI translate the LHS.
841 Constant *RHS = cast<ConstantInt>(Inst->getOperand(1));
842 Instruction *OpI = dyn_cast<Instruction>(Inst->getOperand(0));
843 bool isNSW = cast<BinaryOperator>(Inst)->hasNoSignedWrap();
844 bool isNUW = cast<BinaryOperator>(Inst)->hasNoUnsignedWrap();
846 if (OpI == 0 || OpI->getParent() != Inst->getParent())
847 LHS = Inst->getOperand(0);
849 LHS = GetPHITranslatedValue(Inst->getOperand(0), CurBB, Pred, TD);
854 // If the PHI translated LHS is an add of a constant, fold the immediates.
855 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(LHS))
856 if (BOp->getOpcode() == Instruction::Add)
857 if (ConstantInt *CI = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
858 LHS = BOp->getOperand(0);
859 RHS = ConstantExpr::getAdd(RHS, CI);
860 isNSW = isNUW = false;
863 // See if the add simplifies away.
864 if (Value *Res = SimplifyAddInst(LHS, RHS, isNSW, isNUW, TD))
867 // Otherwise, see if we have this add available somewhere.
868 for (Value::use_iterator UI = LHS->use_begin(), E = LHS->use_end();
870 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(*UI))
871 if (BO->getOperand(0) == LHS && BO->getOperand(1) == RHS &&
872 BO->getParent()->getParent() == CurBB->getParent())
882 /// GetAvailablePHITranslatePointer - Return the value computed by
883 /// PHITranslatePointer if it dominates PredBB, otherwise return null.
884 Value *MemoryDependenceAnalysis::
885 GetAvailablePHITranslatedValue(Value *V,
886 BasicBlock *CurBB, BasicBlock *PredBB,
887 const TargetData *TD,
888 const DominatorTree &DT) const {
889 // See if PHI translation succeeds.
890 V = GetPHITranslatedValue(V, CurBB, PredBB, TD);
891 if (V == 0) return 0;
893 // Make sure the value is live in the predecessor.
894 if (Instruction *Inst = dyn_cast_or_null<Instruction>(V))
895 if (!DT.dominates(Inst->getParent(), PredBB))
901 /// InsertPHITranslatedPointer - Insert a computation of the PHI translated
902 /// version of 'V' for the edge PredBB->CurBB into the end of the PredBB
903 /// block. All newly created instructions are added to the NewInsts list.
905 Value *MemoryDependenceAnalysis::
906 InsertPHITranslatedPointer(Value *InVal, BasicBlock *CurBB,
907 BasicBlock *PredBB, const TargetData *TD,
908 const DominatorTree &DT,
909 SmallVectorImpl<Instruction*> &NewInsts) const {
910 // See if we have a version of this value already available and dominating
911 // PredBB. If so, there is no need to insert a new copy.
912 if (Value *Res = GetAvailablePHITranslatedValue(InVal, CurBB, PredBB, TD, DT))
915 // If we don't have an available version of this value, it must be an
917 Instruction *Inst = cast<Instruction>(InVal);
919 // Handle bitcast of PHI translatable value.
920 if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) {
921 Value *OpVal = InsertPHITranslatedPointer(BC->getOperand(0),
922 CurBB, PredBB, TD, DT, NewInsts);
923 if (OpVal == 0) return 0;
925 // Otherwise insert a bitcast at the end of PredBB.
926 BitCastInst *New = new BitCastInst(OpVal, InVal->getType(),
927 InVal->getName()+".phi.trans.insert",
928 PredBB->getTerminator());
929 NewInsts.push_back(New);
933 // Handle getelementptr with at least one PHI operand.
934 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
935 SmallVector<Value*, 8> GEPOps;
936 BasicBlock *CurBB = GEP->getParent();
937 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) {
938 Value *OpVal = InsertPHITranslatedPointer(GEP->getOperand(i),
939 CurBB, PredBB, TD, DT, NewInsts);
940 if (OpVal == 0) return 0;
941 GEPOps.push_back(OpVal);
944 GetElementPtrInst *Result =
945 GetElementPtrInst::Create(GEPOps[0], GEPOps.begin()+1, GEPOps.end(),
946 InVal->getName()+".phi.trans.insert",
947 PredBB->getTerminator());
948 Result->setIsInBounds(GEP->isInBounds());
949 NewInsts.push_back(Result);
954 // FIXME: This code works, but it is unclear that we actually want to insert
955 // a big chain of computation in order to make a value available in a block.
956 // This needs to be evaluated carefully to consider its cost trade offs.
958 // Handle add with a constant RHS.
959 if (Inst->getOpcode() == Instruction::Add &&
960 isa<ConstantInt>(Inst->getOperand(1))) {
961 // PHI translate the LHS.
962 Value *OpVal = InsertPHITranslatedPointer(Inst->getOperand(0),
963 CurBB, PredBB, TD, DT, NewInsts);
964 if (OpVal == 0) return 0;
966 BinaryOperator *Res = BinaryOperator::CreateAdd(OpVal, Inst->getOperand(1),
967 InVal->getName()+".phi.trans.insert",
968 PredBB->getTerminator());
969 Res->setHasNoSignedWrap(cast<BinaryOperator>(Inst)->hasNoSignedWrap());
970 Res->setHasNoUnsignedWrap(cast<BinaryOperator>(Inst)->hasNoUnsignedWrap());
971 NewInsts.push_back(Res);
979 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
980 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
981 /// results to the results vector and keep track of which blocks are visited in
984 /// This has special behavior for the first block queries (when SkipFirstBlock
985 /// is true). In this special case, it ignores the contents of the specified
986 /// block and starts returning dependence info for its predecessors.
988 /// This function returns false on success, or true to indicate that it could
989 /// not compute dependence information for some reason. This should be treated
990 /// as a clobber dependence on the first instruction in the predecessor block.
991 bool MemoryDependenceAnalysis::
992 getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
993 bool isLoad, BasicBlock *StartBB,
994 SmallVectorImpl<NonLocalDepEntry> &Result,
995 DenseMap<BasicBlock*, Value*> &Visited,
996 bool SkipFirstBlock) {
998 // Look up the cached info for Pointer.
999 ValueIsLoadPair CacheKey(Pointer, isLoad);
1001 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
1002 &NonLocalPointerDeps[CacheKey];
1003 NonLocalDepInfo *Cache = &CacheInfo->second;
1005 // If we have valid cached information for exactly the block we are
1006 // investigating, just return it with no recomputation.
1007 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1008 // We have a fully cached result for this query then we can just return the
1009 // cached results and populate the visited set. However, we have to verify
1010 // that we don't already have conflicting results for these blocks. Check
1011 // to ensure that if a block in the results set is in the visited set that
1012 // it was for the same pointer query.
1013 if (!Visited.empty()) {
1014 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1016 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->first);
1017 if (VI == Visited.end() || VI->second == Pointer) continue;
1019 // We have a pointer mismatch in a block. Just return clobber, saying
1020 // that something was clobbered in this result. We could also do a
1021 // non-fully cached query, but there is little point in doing this.
1026 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1028 Visited.insert(std::make_pair(I->first, Pointer));
1029 if (!I->second.isNonLocal())
1030 Result.push_back(*I);
1032 ++NumCacheCompleteNonLocalPtr;
1036 // Otherwise, either this is a new block, a block with an invalid cache
1037 // pointer or one that we're about to invalidate by putting more info into it
1038 // than its valid cache info. If empty, the result will be valid cache info,
1039 // otherwise it isn't.
1041 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1043 CacheInfo->first = BBSkipFirstBlockPair();
1045 SmallVector<BasicBlock*, 32> Worklist;
1046 Worklist.push_back(StartBB);
1048 // Keep track of the entries that we know are sorted. Previously cached
1049 // entries will all be sorted. The entries we add we only sort on demand (we
1050 // don't insert every element into its sorted position). We know that we
1051 // won't get any reuse from currently inserted values, because we don't
1052 // revisit blocks after we insert info for them.
1053 unsigned NumSortedEntries = Cache->size();
1054 DEBUG(AssertSorted(*Cache));
1056 while (!Worklist.empty()) {
1057 BasicBlock *BB = Worklist.pop_back_val();
1059 // Skip the first block if we have it.
1060 if (!SkipFirstBlock) {
1061 // Analyze the dependency of *Pointer in FromBB. See if we already have
1063 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1065 // Get the dependency info for Pointer in BB. If we have cached
1066 // information, we will use it, otherwise we compute it.
1067 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1068 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad,
1069 BB, Cache, NumSortedEntries);
1071 // If we got a Def or Clobber, add this to the list of results.
1072 if (!Dep.isNonLocal()) {
1073 Result.push_back(NonLocalDepEntry(BB, Dep));
1078 // If 'Pointer' is an instruction defined in this block, then we need to do
1079 // phi translation to change it into a value live in the predecessor block.
1080 // If phi translation fails, then we can't continue dependence analysis.
1081 Instruction *PtrInst = dyn_cast<Instruction>(Pointer);
1082 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB;
1084 // If no PHI translation is needed, just add all the predecessors of this
1085 // block to scan them as well.
1086 if (!NeedsPHITranslation) {
1087 SkipFirstBlock = false;
1088 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1089 // Verify that we haven't looked at this block yet.
1090 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1091 InsertRes = Visited.insert(std::make_pair(*PI, Pointer));
1092 if (InsertRes.second) {
1093 // First time we've looked at *PI.
1094 Worklist.push_back(*PI);
1098 // If we have seen this block before, but it was with a different
1099 // pointer then we have a phi translation failure and we have to treat
1100 // this as a clobber.
1101 if (InsertRes.first->second != Pointer)
1102 goto PredTranslationFailure;
1107 // If we do need to do phi translation, then there are a bunch of different
1108 // cases, because we have to find a Value* live in the predecessor block. We
1109 // know that PtrInst is defined in this block at least.
1111 // We may have added values to the cache list before this PHI translation.
1112 // If so, we haven't done anything to ensure that the cache remains sorted.
1113 // Sort it now (if needed) so that recursive invocations of
1114 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1115 // value will only see properly sorted cache arrays.
1116 if (Cache && NumSortedEntries != Cache->size()) {
1117 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1118 NumSortedEntries = Cache->size();
1121 // If this is a computation derived from a PHI node, use the suitably
1122 // translated incoming values for each pred as the phi translated version.
1123 if (!isPHITranslatable(PtrInst))
1124 goto PredTranslationFailure;
1128 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1129 BasicBlock *Pred = *PI;
1130 // Get the PHI translated pointer in this predecessor. This can fail and
1131 // return null if not translatable.
1132 Value *PredPtr = GetPHITranslatedValue(PtrInst, BB, Pred, TD);
1134 // Check to see if we have already visited this pred block with another
1135 // pointer. If so, we can't do this lookup. This failure can occur
1136 // with PHI translation when a critical edge exists and the PHI node in
1137 // the successor translates to a pointer value different than the
1138 // pointer the block was first analyzed with.
1139 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1140 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr));
1142 if (!InsertRes.second) {
1143 // If the predecessor was visited with PredPtr, then we already did
1144 // the analysis and can ignore it.
1145 if (InsertRes.first->second == PredPtr)
1148 // Otherwise, the block was previously analyzed with a different
1149 // pointer. We can't represent the result of this case, so we just
1150 // treat this as a phi translation failure.
1151 goto PredTranslationFailure;
1154 // If PHI translation was unable to find an available pointer in this
1155 // predecessor, then we have to assume that the pointer is clobbered in
1156 // that predecessor. We can still do PRE of the load, which would insert
1157 // a computation of the pointer in this predecessor.
1159 // Add the entry to the Result list.
1160 NonLocalDepEntry Entry(Pred,
1161 MemDepResult::getClobber(Pred->getTerminator()));
1162 Result.push_back(Entry);
1164 // Add it to the cache for this CacheKey so that subsequent queries get
1166 Cache = &NonLocalPointerDeps[CacheKey].second;
1167 MemoryDependenceAnalysis::NonLocalDepInfo::iterator It =
1168 std::upper_bound(Cache->begin(), Cache->end(), Entry);
1170 if (It != Cache->begin() && prior(It)->first == Pred)
1173 if (It == Cache->end() || It->first != Pred) {
1174 Cache->insert(It, Entry);
1175 // Add it to the reverse map.
1176 ReverseNonLocalPtrDeps[Pred->getTerminator()].insert(CacheKey);
1177 } else if (!It->second.isDirty()) {
1179 } else if (It->second.getInst() == Pred->getTerminator()) {
1180 // Same instruction, clear the dirty marker.
1181 It->second = Entry.second;
1182 } else if (It->second.getInst() == 0) {
1183 // Dirty, with no instruction, just add this.
1184 It->second = Entry.second;
1185 ReverseNonLocalPtrDeps[Pred->getTerminator()].insert(CacheKey);
1187 // Otherwise, dirty with a different instruction.
1188 RemoveFromReverseMap(ReverseNonLocalPtrDeps, It->second.getInst(),
1190 It->second = Entry.second;
1191 ReverseNonLocalPtrDeps[Pred->getTerminator()].insert(CacheKey);
1197 // FIXME: it is entirely possible that PHI translating will end up with
1198 // the same value. Consider PHI translating something like:
1199 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1200 // to recurse here, pedantically speaking.
1202 // If we have a problem phi translating, fall through to the code below
1203 // to handle the failure condition.
1204 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred,
1206 goto PredTranslationFailure;
1209 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1210 CacheInfo = &NonLocalPointerDeps[CacheKey];
1211 Cache = &CacheInfo->second;
1212 NumSortedEntries = Cache->size();
1214 // Since we did phi translation, the "Cache" set won't contain all of the
1215 // results for the query. This is ok (we can still use it to accelerate
1216 // specific block queries) but we can't do the fastpath "return all
1217 // results from the set" Clear out the indicator for this.
1218 CacheInfo->first = BBSkipFirstBlockPair();
1219 SkipFirstBlock = false;
1222 PredTranslationFailure:
1225 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1226 CacheInfo = &NonLocalPointerDeps[CacheKey];
1227 Cache = &CacheInfo->second;
1228 NumSortedEntries = Cache->size();
1231 // Since we did phi translation, the "Cache" set won't contain all of the
1232 // results for the query. This is ok (we can still use it to accelerate
1233 // specific block queries) but we can't do the fastpath "return all
1234 // results from the set" Clear out the indicator for this.
1235 CacheInfo->first = BBSkipFirstBlockPair();
1237 // If *nothing* works, mark the pointer as being clobbered by the first
1238 // instruction in this block.
1240 // If this is the magic first block, return this as a clobber of the whole
1241 // incoming value. Since we can't phi translate to one of the predecessors,
1242 // we have to bail out.
1246 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1247 assert(I != Cache->rend() && "Didn't find current block??");
1251 assert(I->second.isNonLocal() &&
1252 "Should only be here with transparent block");
1253 I->second = MemDepResult::getClobber(BB->begin());
1254 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
1255 Result.push_back(*I);
1260 // Okay, we're done now. If we added new values to the cache, re-sort it.
1261 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1262 DEBUG(AssertSorted(*Cache));
1266 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1267 /// CachedNonLocalPointerInfo, remove it.
1268 void MemoryDependenceAnalysis::
1269 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1270 CachedNonLocalPointerInfo::iterator It =
1271 NonLocalPointerDeps.find(P);
1272 if (It == NonLocalPointerDeps.end()) return;
1274 // Remove all of the entries in the BB->val map. This involves removing
1275 // instructions from the reverse map.
1276 NonLocalDepInfo &PInfo = It->second.second;
1278 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1279 Instruction *Target = PInfo[i].second.getInst();
1280 if (Target == 0) continue; // Ignore non-local dep results.
1281 assert(Target->getParent() == PInfo[i].first);
1283 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1284 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1287 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1288 NonLocalPointerDeps.erase(It);
1292 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1293 /// information about the specified pointer, because it may be too
1294 /// conservative in memdep. This is an optional call that can be used when
1295 /// the client detects an equivalence between the pointer and some other
1296 /// value and replaces the other value with ptr. This can make Ptr available
1297 /// in more places that cached info does not necessarily keep.
1298 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1299 // If Ptr isn't really a pointer, just ignore it.
1300 if (!isa<PointerType>(Ptr->getType())) return;
1301 // Flush store info for the pointer.
1302 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1303 // Flush load info for the pointer.
1304 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1307 /// removeInstruction - Remove an instruction from the dependence analysis,
1308 /// updating the dependence of instructions that previously depended on it.
1309 /// This method attempts to keep the cache coherent using the reverse map.
1310 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1311 // Walk through the Non-local dependencies, removing this one as the value
1312 // for any cached queries.
1313 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1314 if (NLDI != NonLocalDeps.end()) {
1315 NonLocalDepInfo &BlockMap = NLDI->second.first;
1316 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1318 if (Instruction *Inst = DI->second.getInst())
1319 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1320 NonLocalDeps.erase(NLDI);
1323 // If we have a cached local dependence query for this instruction, remove it.
1325 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1326 if (LocalDepEntry != LocalDeps.end()) {
1327 // Remove us from DepInst's reverse set now that the local dep info is gone.
1328 if (Instruction *Inst = LocalDepEntry->second.getInst())
1329 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1331 // Remove this local dependency info.
1332 LocalDeps.erase(LocalDepEntry);
1335 // If we have any cached pointer dependencies on this instruction, remove
1336 // them. If the instruction has non-pointer type, then it can't be a pointer
1339 // Remove it from both the load info and the store info. The instruction
1340 // can't be in either of these maps if it is non-pointer.
1341 if (isa<PointerType>(RemInst->getType())) {
1342 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1343 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1346 // Loop over all of the things that depend on the instruction we're removing.
1348 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1350 // If we find RemInst as a clobber or Def in any of the maps for other values,
1351 // we need to replace its entry with a dirty version of the instruction after
1352 // it. If RemInst is a terminator, we use a null dirty value.
1354 // Using a dirty version of the instruction after RemInst saves having to scan
1355 // the entire block to get to this point.
1356 MemDepResult NewDirtyVal;
1357 if (!RemInst->isTerminator())
1358 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1360 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1361 if (ReverseDepIt != ReverseLocalDeps.end()) {
1362 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1363 // RemInst can't be the terminator if it has local stuff depending on it.
1364 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1365 "Nothing can locally depend on a terminator");
1367 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1368 E = ReverseDeps.end(); I != E; ++I) {
1369 Instruction *InstDependingOnRemInst = *I;
1370 assert(InstDependingOnRemInst != RemInst &&
1371 "Already removed our local dep info");
1373 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1375 // Make sure to remember that new things depend on NewDepInst.
1376 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1377 "a local dep on this if it is a terminator!");
1378 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1379 InstDependingOnRemInst));
1382 ReverseLocalDeps.erase(ReverseDepIt);
1384 // Add new reverse deps after scanning the set, to avoid invalidating the
1385 // 'ReverseDeps' reference.
1386 while (!ReverseDepsToAdd.empty()) {
1387 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1388 .insert(ReverseDepsToAdd.back().second);
1389 ReverseDepsToAdd.pop_back();
1393 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1394 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1395 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1396 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1398 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1400 PerInstNLInfo &INLD = NonLocalDeps[*I];
1401 // The information is now dirty!
1404 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1405 DE = INLD.first.end(); DI != DE; ++DI) {
1406 if (DI->second.getInst() != RemInst) continue;
1408 // Convert to a dirty entry for the subsequent instruction.
1409 DI->second = NewDirtyVal;
1411 if (Instruction *NextI = NewDirtyVal.getInst())
1412 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1416 ReverseNonLocalDeps.erase(ReverseDepIt);
1418 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1419 while (!ReverseDepsToAdd.empty()) {
1420 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1421 .insert(ReverseDepsToAdd.back().second);
1422 ReverseDepsToAdd.pop_back();
1426 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1427 // value in the NonLocalPointerDeps info.
1428 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1429 ReverseNonLocalPtrDeps.find(RemInst);
1430 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1431 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1432 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1434 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1435 E = Set.end(); I != E; ++I) {
1436 ValueIsLoadPair P = *I;
1437 assert(P.getPointer() != RemInst &&
1438 "Already removed NonLocalPointerDeps info for RemInst");
1440 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second;
1442 // The cache is not valid for any specific block anymore.
1443 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair();
1445 // Update any entries for RemInst to use the instruction after it.
1446 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1448 if (DI->second.getInst() != RemInst) continue;
1450 // Convert to a dirty entry for the subsequent instruction.
1451 DI->second = NewDirtyVal;
1453 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1454 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1457 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1458 // subsequent value may invalidate the sortedness.
1459 std::sort(NLPDI.begin(), NLPDI.end());
1462 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1464 while (!ReversePtrDepsToAdd.empty()) {
1465 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1466 .insert(ReversePtrDepsToAdd.back().second);
1467 ReversePtrDepsToAdd.pop_back();
1472 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1473 AA->deleteValue(RemInst);
1474 DEBUG(verifyRemoved(RemInst));
1476 /// verifyRemoved - Verify that the specified instruction does not occur
1477 /// in our internal data structures.
1478 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1479 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1480 E = LocalDeps.end(); I != E; ++I) {
1481 assert(I->first != D && "Inst occurs in data structures");
1482 assert(I->second.getInst() != D &&
1483 "Inst occurs in data structures");
1486 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1487 E = NonLocalPointerDeps.end(); I != E; ++I) {
1488 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1489 const NonLocalDepInfo &Val = I->second.second;
1490 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1492 assert(II->second.getInst() != D && "Inst occurs as NLPD value");
1495 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1496 E = NonLocalDeps.end(); I != E; ++I) {
1497 assert(I->first != D && "Inst occurs in data structures");
1498 const PerInstNLInfo &INLD = I->second;
1499 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1500 EE = INLD.first.end(); II != EE; ++II)
1501 assert(II->second.getInst() != D && "Inst occurs in data structures");
1504 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1505 E = ReverseLocalDeps.end(); I != E; ++I) {
1506 assert(I->first != D && "Inst occurs in data structures");
1507 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1508 EE = I->second.end(); II != EE; ++II)
1509 assert(*II != D && "Inst occurs in data structures");
1512 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1513 E = ReverseNonLocalDeps.end();
1515 assert(I->first != D && "Inst occurs in data structures");
1516 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1517 EE = I->second.end(); II != EE; ++II)
1518 assert(*II != D && "Inst occurs in data structures");
1521 for (ReverseNonLocalPtrDepTy::const_iterator
1522 I = ReverseNonLocalPtrDeps.begin(),
1523 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1524 assert(I->first != D && "Inst occurs in rev NLPD map");
1526 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1527 E = I->second.end(); II != E; ++II)
1528 assert(*II != ValueIsLoadPair(D, false) &&
1529 *II != ValueIsLoadPair(D, true) &&
1530 "Inst occurs in ReverseNonLocalPtrDeps map");