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 // Limit for the number of instructions to scan in a block.
50 // FIXME: Figure out what a sane value is for this.
51 // (500 is relatively insane.)
52 static const int BlockScanLimit = 500;
54 char MemoryDependenceAnalysis::ID = 0;
56 // Register this pass...
57 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
58 "Memory Dependence Analysis", false, true)
59 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
60 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
61 "Memory Dependence Analysis", false, true)
63 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
64 : FunctionPass(ID), PredCache(0) {
65 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
67 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
70 /// Clean up memory in between runs
71 void MemoryDependenceAnalysis::releaseMemory() {
74 NonLocalPointerDeps.clear();
75 ReverseLocalDeps.clear();
76 ReverseNonLocalDeps.clear();
77 ReverseNonLocalPtrDeps.clear();
83 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
85 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
87 AU.addRequiredTransitive<AliasAnalysis>();
90 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
91 AA = &getAnalysis<AliasAnalysis>();
92 TD = getAnalysisIfAvailable<TargetData>();
93 DT = getAnalysisIfAvailable<DominatorTree>();
95 PredCache.reset(new PredIteratorCache());
99 /// RemoveFromReverseMap - This is a helper function that removes Val from
100 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
101 template <typename KeyTy>
102 static void RemoveFromReverseMap(DenseMap<Instruction*,
103 SmallPtrSet<KeyTy, 4> > &ReverseMap,
104 Instruction *Inst, KeyTy Val) {
105 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
106 InstIt = ReverseMap.find(Inst);
107 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
108 bool Found = InstIt->second.erase(Val);
109 assert(Found && "Invalid reverse map!"); (void)Found;
110 if (InstIt->second.empty())
111 ReverseMap.erase(InstIt);
114 /// GetLocation - If the given instruction references a specific memory
115 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
116 /// Return a ModRefInfo value describing the general behavior of the
119 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
120 AliasAnalysis::Location &Loc,
122 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
123 if (LI->isUnordered()) {
124 Loc = AA->getLocation(LI);
125 return AliasAnalysis::Ref;
126 } else if (LI->getOrdering() == Monotonic) {
127 Loc = AA->getLocation(LI);
128 return AliasAnalysis::ModRef;
130 Loc = AliasAnalysis::Location();
131 return AliasAnalysis::ModRef;
134 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
135 if (SI->isUnordered()) {
136 Loc = AA->getLocation(SI);
137 return AliasAnalysis::Mod;
138 } else if (SI->getOrdering() == Monotonic) {
139 Loc = AA->getLocation(SI);
140 return AliasAnalysis::ModRef;
142 Loc = AliasAnalysis::Location();
143 return AliasAnalysis::ModRef;
146 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
147 Loc = AA->getLocation(V);
148 return AliasAnalysis::ModRef;
151 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) {
152 // calls to free() deallocate the entire structure
153 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
154 return AliasAnalysis::Mod;
157 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
158 switch (II->getIntrinsicID()) {
159 case Intrinsic::lifetime_start:
160 case Intrinsic::lifetime_end:
161 case Intrinsic::invariant_start:
162 Loc = AliasAnalysis::Location(II->getArgOperand(1),
163 cast<ConstantInt>(II->getArgOperand(0))
165 II->getMetadata(LLVMContext::MD_tbaa));
166 // These intrinsics don't really modify the memory, but returning Mod
167 // will allow them to be handled conservatively.
168 return AliasAnalysis::Mod;
169 case Intrinsic::invariant_end:
170 Loc = AliasAnalysis::Location(II->getArgOperand(2),
171 cast<ConstantInt>(II->getArgOperand(1))
173 II->getMetadata(LLVMContext::MD_tbaa));
174 // These intrinsics don't really modify the memory, but returning Mod
175 // will allow them to be handled conservatively.
176 return AliasAnalysis::Mod;
181 // Otherwise, just do the coarse-grained thing that always works.
182 if (Inst->mayWriteToMemory())
183 return AliasAnalysis::ModRef;
184 if (Inst->mayReadFromMemory())
185 return AliasAnalysis::Ref;
186 return AliasAnalysis::NoModRef;
189 /// getCallSiteDependencyFrom - Private helper for finding the local
190 /// dependencies of a call site.
191 MemDepResult MemoryDependenceAnalysis::
192 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
193 BasicBlock::iterator ScanIt, BasicBlock *BB) {
194 unsigned Limit = BlockScanLimit;
196 // Walk backwards through the block, looking for dependencies
197 while (ScanIt != BB->begin()) {
198 // Limit the amount of scanning we do so we don't end up with quadratic
199 // running time on extreme testcases.
202 return MemDepResult::getUnknown();
204 Instruction *Inst = --ScanIt;
206 // If this inst is a memory op, get the pointer it accessed
207 AliasAnalysis::Location Loc;
208 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
210 // A simple instruction.
211 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
212 return MemDepResult::getClobber(Inst);
216 if (CallSite InstCS = cast<Value>(Inst)) {
217 // Debug intrinsics don't cause dependences.
218 if (isa<DbgInfoIntrinsic>(Inst)) continue;
219 // If these two calls do not interfere, look past it.
220 switch (AA->getModRefInfo(CS, InstCS)) {
221 case AliasAnalysis::NoModRef:
222 // If the two calls are the same, return InstCS as a Def, so that
223 // CS can be found redundant and eliminated.
224 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
225 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
226 return MemDepResult::getDef(Inst);
228 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
232 return MemDepResult::getClobber(Inst);
236 // If we could not obtain a pointer for the instruction and the instruction
237 // touches memory then assume that this is a dependency.
238 if (MR != AliasAnalysis::NoModRef)
239 return MemDepResult::getClobber(Inst);
242 // No dependence found. If this is the entry block of the function, it is
243 // unknown, otherwise it is non-local.
244 if (BB != &BB->getParent()->getEntryBlock())
245 return MemDepResult::getNonLocal();
246 return MemDepResult::getNonFuncLocal();
249 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
250 /// would fully overlap MemLoc if done as a wider legal integer load.
252 /// MemLocBase, MemLocOffset are lazily computed here the first time the
253 /// base/offs of memloc is needed.
255 isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc,
256 const Value *&MemLocBase,
259 const TargetData *TD) {
260 // If we have no target data, we can't do this.
261 if (TD == 0) return false;
263 // If we haven't already computed the base/offset of MemLoc, do so now.
265 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, *TD);
267 unsigned Size = MemoryDependenceAnalysis::
268 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size,
273 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
274 /// looks at a memory location for a load (specified by MemLocBase, Offs,
275 /// and Size) and compares it against a load. If the specified load could
276 /// be safely widened to a larger integer load that is 1) still efficient,
277 /// 2) safe for the target, and 3) would provide the specified memory
278 /// location value, then this function returns the size in bytes of the
279 /// load width to use. If not, this returns zero.
280 unsigned MemoryDependenceAnalysis::
281 getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs,
282 unsigned MemLocSize, const LoadInst *LI,
283 const TargetData &TD) {
284 // We can only extend simple integer loads.
285 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
287 // Get the base of this load.
289 const Value *LIBase =
290 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, TD);
292 // If the two pointers are not based on the same pointer, we can't tell that
294 if (LIBase != MemLocBase) return 0;
296 // Okay, the two values are based on the same pointer, but returned as
297 // no-alias. This happens when we have things like two byte loads at "P+1"
298 // and "P+3". Check to see if increasing the size of the "LI" load up to its
299 // alignment (or the largest native integer type) will allow us to load all
300 // the bits required by MemLoc.
302 // If MemLoc is before LI, then no widening of LI will help us out.
303 if (MemLocOffs < LIOffs) return 0;
305 // Get the alignment of the load in bytes. We assume that it is safe to load
306 // any legal integer up to this size without a problem. For example, if we're
307 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
308 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
310 unsigned LoadAlign = LI->getAlignment();
312 int64_t MemLocEnd = MemLocOffs+MemLocSize;
314 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
315 if (LIOffs+LoadAlign < MemLocEnd) return 0;
317 // This is the size of the load to try. Start with the next larger power of
319 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
320 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
323 // If this load size is bigger than our known alignment or would not fit
324 // into a native integer register, then we fail.
325 if (NewLoadByteSize > LoadAlign ||
326 !TD.fitsInLegalInteger(NewLoadByteSize*8))
329 if (LIOffs+NewLoadByteSize > MemLocEnd &&
330 LI->getParent()->getParent()->hasFnAttr(Attribute::AddressSafety)) {
331 // We will be reading past the location accessed by the original program.
332 // While this is safe in a regular build, Address Safety analysis tools
333 // may start reporting false warnings. So, don't do widening.
337 // If a load of this width would include all of MemLoc, then we succeed.
338 if (LIOffs+NewLoadByteSize >= MemLocEnd)
339 return NewLoadByteSize;
341 NewLoadByteSize <<= 1;
345 /// getPointerDependencyFrom - Return the instruction on which a memory
346 /// location depends. If isLoad is true, this routine ignores may-aliases with
347 /// read-only operations. If isLoad is false, this routine ignores may-aliases
348 /// with reads from read-only locations.
349 MemDepResult MemoryDependenceAnalysis::
350 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
351 BasicBlock::iterator ScanIt, BasicBlock *BB) {
353 const Value *MemLocBase = 0;
354 int64_t MemLocOffset = 0;
356 unsigned Limit = BlockScanLimit;
358 // Walk backwards through the basic block, looking for dependencies.
359 while (ScanIt != BB->begin()) {
360 // Limit the amount of scanning we do so we don't end up with quadratic
361 // running time on extreme testcases.
364 return MemDepResult::getUnknown();
366 Instruction *Inst = --ScanIt;
368 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
369 // Debug intrinsics don't (and can't) cause dependences.
370 if (isa<DbgInfoIntrinsic>(II)) continue;
372 // If we reach a lifetime begin or end marker, then the query ends here
373 // because the value is undefined.
374 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
375 // FIXME: This only considers queries directly on the invariant-tagged
376 // pointer, not on query pointers that are indexed off of them. It'd
377 // be nice to handle that at some point (the right approach is to use
378 // GetPointerBaseWithConstantOffset).
379 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
381 return MemDepResult::getDef(II);
386 // Values depend on loads if the pointers are must aliased. This means that
387 // a load depends on another must aliased load from the same value.
388 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
389 // Atomic loads have complications involved.
390 // FIXME: This is overly conservative.
391 if (!LI->isUnordered())
392 return MemDepResult::getClobber(LI);
394 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
396 // If we found a pointer, check if it could be the same as our pointer.
397 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
400 if (R == AliasAnalysis::NoAlias) {
401 // If this is an over-aligned integer load (for example,
402 // "load i8* %P, align 4") see if it would obviously overlap with the
403 // queried location if widened to a larger load (e.g. if the queried
404 // location is 1 byte at P+1). If so, return it as a load/load
405 // clobber result, allowing the client to decide to widen the load if
407 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType()))
408 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() &&
409 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
410 MemLocOffset, LI, TD))
411 return MemDepResult::getClobber(Inst);
416 // Must aliased loads are defs of each other.
417 if (R == AliasAnalysis::MustAlias)
418 return MemDepResult::getDef(Inst);
420 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
421 // in terms of clobbering loads, but since it does this by looking
422 // at the clobbering load directly, it doesn't know about any
423 // phi translation that may have happened along the way.
425 // If we have a partial alias, then return this as a clobber for the
427 if (R == AliasAnalysis::PartialAlias)
428 return MemDepResult::getClobber(Inst);
431 // Random may-alias loads don't depend on each other without a
436 // Stores don't depend on other no-aliased accesses.
437 if (R == AliasAnalysis::NoAlias)
440 // Stores don't alias loads from read-only memory.
441 if (AA->pointsToConstantMemory(LoadLoc))
444 // Stores depend on may/must aliased loads.
445 return MemDepResult::getDef(Inst);
448 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
449 // Atomic stores have complications involved.
450 // FIXME: This is overly conservative.
451 if (!SI->isUnordered())
452 return MemDepResult::getClobber(SI);
454 // If alias analysis can tell that this store is guaranteed to not modify
455 // the query pointer, ignore it. Use getModRefInfo to handle cases where
456 // the query pointer points to constant memory etc.
457 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
460 // Ok, this store might clobber the query pointer. Check to see if it is
461 // a must alias: in this case, we want to return this as a def.
462 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
464 // If we found a pointer, check if it could be the same as our pointer.
465 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
467 if (R == AliasAnalysis::NoAlias)
469 if (R == AliasAnalysis::MustAlias)
470 return MemDepResult::getDef(Inst);
471 return MemDepResult::getClobber(Inst);
474 // If this is an allocation, and if we know that the accessed pointer is to
475 // the allocation, return Def. This means that there is no dependence and
476 // the access can be optimized based on that. For example, a load could
478 // Note: Only determine this to be a malloc if Inst is the malloc call, not
479 // a subsequent bitcast of the malloc call result. There can be stores to
480 // the malloced memory between the malloc call and its bitcast uses, and we
481 // need to continue scanning until the malloc call.
482 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
483 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
484 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD);
486 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
487 return MemDepResult::getDef(Inst);
488 // If the allocation is not aliased and does not read memory (like
489 // strdup), it is safe to ignore.
490 if (isa<AllocaInst>(Inst) ||
491 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
495 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
496 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
497 // If necessary, perform additional analysis.
498 if (MR == AliasAnalysis::ModRef)
499 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
501 case AliasAnalysis::NoModRef:
502 // If the call has no effect on the queried pointer, just ignore it.
504 case AliasAnalysis::Mod:
505 return MemDepResult::getClobber(Inst);
506 case AliasAnalysis::Ref:
507 // If the call is known to never store to the pointer, and if this is a
508 // load query, we can safely ignore it (scan past it).
512 // Otherwise, there is a potential dependence. Return a clobber.
513 return MemDepResult::getClobber(Inst);
517 // No dependence found. If this is the entry block of the function, it is
518 // unknown, otherwise it is non-local.
519 if (BB != &BB->getParent()->getEntryBlock())
520 return MemDepResult::getNonLocal();
521 return MemDepResult::getNonFuncLocal();
524 /// getDependency - Return the instruction on which a memory operation
526 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
527 Instruction *ScanPos = QueryInst;
529 // Check for a cached result
530 MemDepResult &LocalCache = LocalDeps[QueryInst];
532 // If the cached entry is non-dirty, just return it. Note that this depends
533 // on MemDepResult's default constructing to 'dirty'.
534 if (!LocalCache.isDirty())
537 // Otherwise, if we have a dirty entry, we know we can start the scan at that
538 // instruction, which may save us some work.
539 if (Instruction *Inst = LocalCache.getInst()) {
542 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
545 BasicBlock *QueryParent = QueryInst->getParent();
548 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
549 // No dependence found. If this is the entry block of the function, it is
550 // unknown, otherwise it is non-local.
551 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
552 LocalCache = MemDepResult::getNonLocal();
554 LocalCache = MemDepResult::getNonFuncLocal();
556 AliasAnalysis::Location MemLoc;
557 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
559 // If we can do a pointer scan, make it happen.
560 bool isLoad = !(MR & AliasAnalysis::Mod);
561 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
562 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
564 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
566 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
567 CallSite QueryCS(QueryInst);
568 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
569 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
572 // Non-memory instruction.
573 LocalCache = MemDepResult::getUnknown();
576 // Remember the result!
577 if (Instruction *I = LocalCache.getInst())
578 ReverseLocalDeps[I].insert(QueryInst);
584 /// AssertSorted - This method is used when -debug is specified to verify that
585 /// cache arrays are properly kept sorted.
586 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
588 if (Count == -1) Count = Cache.size();
589 if (Count == 0) return;
591 for (unsigned i = 1; i != unsigned(Count); ++i)
592 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
596 /// getNonLocalCallDependency - Perform a full dependency query for the
597 /// specified call, returning the set of blocks that the value is
598 /// potentially live across. The returned set of results will include a
599 /// "NonLocal" result for all blocks where the value is live across.
601 /// This method assumes the instruction returns a "NonLocal" dependency
602 /// within its own block.
604 /// This returns a reference to an internal data structure that may be
605 /// invalidated on the next non-local query or when an instruction is
606 /// removed. Clients must copy this data if they want it around longer than
608 const MemoryDependenceAnalysis::NonLocalDepInfo &
609 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
610 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
611 "getNonLocalCallDependency should only be used on calls with non-local deps!");
612 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
613 NonLocalDepInfo &Cache = CacheP.first;
615 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
616 /// the cached case, this can happen due to instructions being deleted etc. In
617 /// the uncached case, this starts out as the set of predecessors we care
619 SmallVector<BasicBlock*, 32> DirtyBlocks;
621 if (!Cache.empty()) {
622 // Okay, we have a cache entry. If we know it is not dirty, just return it
623 // with no computation.
624 if (!CacheP.second) {
629 // If we already have a partially computed set of results, scan them to
630 // determine what is dirty, seeding our initial DirtyBlocks worklist.
631 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
633 if (I->getResult().isDirty())
634 DirtyBlocks.push_back(I->getBB());
636 // Sort the cache so that we can do fast binary search lookups below.
637 std::sort(Cache.begin(), Cache.end());
639 ++NumCacheDirtyNonLocal;
640 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
641 // << Cache.size() << " cached: " << *QueryInst;
643 // Seed DirtyBlocks with each of the preds of QueryInst's block.
644 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
645 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
646 DirtyBlocks.push_back(*PI);
647 ++NumUncacheNonLocal;
650 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
651 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
653 SmallPtrSet<BasicBlock*, 64> Visited;
655 unsigned NumSortedEntries = Cache.size();
656 DEBUG(AssertSorted(Cache));
658 // Iterate while we still have blocks to update.
659 while (!DirtyBlocks.empty()) {
660 BasicBlock *DirtyBB = DirtyBlocks.back();
661 DirtyBlocks.pop_back();
663 // Already processed this block?
664 if (!Visited.insert(DirtyBB))
667 // Do a binary search to see if we already have an entry for this block in
668 // the cache set. If so, find it.
669 DEBUG(AssertSorted(Cache, NumSortedEntries));
670 NonLocalDepInfo::iterator Entry =
671 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
672 NonLocalDepEntry(DirtyBB));
673 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB)
676 NonLocalDepEntry *ExistingResult = 0;
677 if (Entry != Cache.begin()+NumSortedEntries &&
678 Entry->getBB() == DirtyBB) {
679 // If we already have an entry, and if it isn't already dirty, the block
681 if (!Entry->getResult().isDirty())
684 // Otherwise, remember this slot so we can update the value.
685 ExistingResult = &*Entry;
688 // If the dirty entry has a pointer, start scanning from it so we don't have
689 // to rescan the entire block.
690 BasicBlock::iterator ScanPos = DirtyBB->end();
691 if (ExistingResult) {
692 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
694 // We're removing QueryInst's use of Inst.
695 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
696 QueryCS.getInstruction());
700 // Find out if this block has a local dependency for QueryInst.
703 if (ScanPos != DirtyBB->begin()) {
704 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
705 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
706 // No dependence found. If this is the entry block of the function, it is
707 // a clobber, otherwise it is unknown.
708 Dep = MemDepResult::getNonLocal();
710 Dep = MemDepResult::getNonFuncLocal();
713 // If we had a dirty entry for the block, update it. Otherwise, just add
716 ExistingResult->setResult(Dep);
718 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
720 // If the block has a dependency (i.e. it isn't completely transparent to
721 // the value), remember the association!
722 if (!Dep.isNonLocal()) {
723 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
724 // update this when we remove instructions.
725 if (Instruction *Inst = Dep.getInst())
726 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
729 // If the block *is* completely transparent to the load, we need to check
730 // the predecessors of this block. Add them to our worklist.
731 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
732 DirtyBlocks.push_back(*PI);
739 /// getNonLocalPointerDependency - Perform a full dependency query for an
740 /// access to the specified (non-volatile) memory location, returning the
741 /// set of instructions that either define or clobber the value.
743 /// This method assumes the pointer has a "NonLocal" dependency within its
746 void MemoryDependenceAnalysis::
747 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
749 SmallVectorImpl<NonLocalDepResult> &Result) {
750 assert(Loc.Ptr->getType()->isPointerTy() &&
751 "Can't get pointer deps of a non-pointer!");
754 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD);
756 // This is the set of blocks we've inspected, and the pointer we consider in
757 // each block. Because of critical edges, we currently bail out if querying
758 // a block with multiple different pointers. This can happen during PHI
760 DenseMap<BasicBlock*, Value*> Visited;
761 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
762 Result, Visited, true))
765 Result.push_back(NonLocalDepResult(FromBB,
766 MemDepResult::getUnknown(),
767 const_cast<Value *>(Loc.Ptr)));
770 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
771 /// Pointer/PointeeSize using either cached information in Cache or by doing a
772 /// lookup (which may use dirty cache info if available). If we do a lookup,
773 /// add the result to the cache.
774 MemDepResult MemoryDependenceAnalysis::
775 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
776 bool isLoad, BasicBlock *BB,
777 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
779 // Do a binary search to see if we already have an entry for this block in
780 // the cache set. If so, find it.
781 NonLocalDepInfo::iterator Entry =
782 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
783 NonLocalDepEntry(BB));
784 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
787 NonLocalDepEntry *ExistingResult = 0;
788 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
789 ExistingResult = &*Entry;
791 // If we have a cached entry, and it is non-dirty, use it as the value for
793 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
794 ++NumCacheNonLocalPtr;
795 return ExistingResult->getResult();
798 // Otherwise, we have to scan for the value. If we have a dirty cache
799 // entry, start scanning from its position, otherwise we scan from the end
801 BasicBlock::iterator ScanPos = BB->end();
802 if (ExistingResult && ExistingResult->getResult().getInst()) {
803 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
804 "Instruction invalidated?");
805 ++NumCacheDirtyNonLocalPtr;
806 ScanPos = ExistingResult->getResult().getInst();
808 // Eliminating the dirty entry from 'Cache', so update the reverse info.
809 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
810 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
812 ++NumUncacheNonLocalPtr;
815 // Scan the block for the dependency.
816 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
818 // If we had a dirty entry for the block, update it. Otherwise, just add
821 ExistingResult->setResult(Dep);
823 Cache->push_back(NonLocalDepEntry(BB, Dep));
825 // If the block has a dependency (i.e. it isn't completely transparent to
826 // the value), remember the reverse association because we just added it
828 if (!Dep.isDef() && !Dep.isClobber())
831 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
832 // update MemDep when we remove instructions.
833 Instruction *Inst = Dep.getInst();
834 assert(Inst && "Didn't depend on anything?");
835 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
836 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
840 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
841 /// number of elements in the array that are already properly ordered. This is
842 /// optimized for the case when only a few entries are added.
844 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
845 unsigned NumSortedEntries) {
846 switch (Cache.size() - NumSortedEntries) {
848 // done, no new entries.
851 // Two new entries, insert the last one into place.
852 NonLocalDepEntry Val = Cache.back();
854 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
855 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
856 Cache.insert(Entry, Val);
860 // One new entry, Just insert the new value at the appropriate position.
861 if (Cache.size() != 1) {
862 NonLocalDepEntry Val = Cache.back();
864 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
865 std::upper_bound(Cache.begin(), Cache.end(), Val);
866 Cache.insert(Entry, Val);
870 // Added many values, do a full scale sort.
871 std::sort(Cache.begin(), Cache.end());
876 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
877 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
878 /// results to the results vector and keep track of which blocks are visited in
881 /// This has special behavior for the first block queries (when SkipFirstBlock
882 /// is true). In this special case, it ignores the contents of the specified
883 /// block and starts returning dependence info for its predecessors.
885 /// This function returns false on success, or true to indicate that it could
886 /// not compute dependence information for some reason. This should be treated
887 /// as a clobber dependence on the first instruction in the predecessor block.
888 bool MemoryDependenceAnalysis::
889 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
890 const AliasAnalysis::Location &Loc,
891 bool isLoad, BasicBlock *StartBB,
892 SmallVectorImpl<NonLocalDepResult> &Result,
893 DenseMap<BasicBlock*, Value*> &Visited,
894 bool SkipFirstBlock) {
896 // Look up the cached info for Pointer.
897 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
899 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
900 // CacheKey, this value will be inserted as the associated value. Otherwise,
901 // it'll be ignored, and we'll have to check to see if the cached size and
902 // tbaa tag are consistent with the current query.
903 NonLocalPointerInfo InitialNLPI;
904 InitialNLPI.Size = Loc.Size;
905 InitialNLPI.TBAATag = Loc.TBAATag;
907 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
909 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
910 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
911 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
913 // If we already have a cache entry for this CacheKey, we may need to do some
914 // work to reconcile the cache entry and the current query.
916 if (CacheInfo->Size < Loc.Size) {
917 // The query's Size is greater than the cached one. Throw out the
918 // cached data and proceed with the query at the greater size.
919 CacheInfo->Pair = BBSkipFirstBlockPair();
920 CacheInfo->Size = Loc.Size;
921 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
922 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
923 if (Instruction *Inst = DI->getResult().getInst())
924 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
925 CacheInfo->NonLocalDeps.clear();
926 } else if (CacheInfo->Size > Loc.Size) {
927 // This query's Size is less than the cached one. Conservatively restart
928 // the query using the greater size.
929 return getNonLocalPointerDepFromBB(Pointer,
930 Loc.getWithNewSize(CacheInfo->Size),
931 isLoad, StartBB, Result, Visited,
935 // If the query's TBAATag is inconsistent with the cached one,
936 // conservatively throw out the cached data and restart the query with
938 if (CacheInfo->TBAATag != Loc.TBAATag) {
939 if (CacheInfo->TBAATag) {
940 CacheInfo->Pair = BBSkipFirstBlockPair();
941 CacheInfo->TBAATag = 0;
942 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
943 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
944 if (Instruction *Inst = DI->getResult().getInst())
945 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
946 CacheInfo->NonLocalDeps.clear();
949 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(),
950 isLoad, StartBB, Result, Visited,
955 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
957 // If we have valid cached information for exactly the block we are
958 // investigating, just return it with no recomputation.
959 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
960 // We have a fully cached result for this query then we can just return the
961 // cached results and populate the visited set. However, we have to verify
962 // that we don't already have conflicting results for these blocks. Check
963 // to ensure that if a block in the results set is in the visited set that
964 // it was for the same pointer query.
965 if (!Visited.empty()) {
966 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
968 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
969 if (VI == Visited.end() || VI->second == Pointer.getAddr())
972 // We have a pointer mismatch in a block. Just return clobber, saying
973 // that something was clobbered in this result. We could also do a
974 // non-fully cached query, but there is little point in doing this.
979 Value *Addr = Pointer.getAddr();
980 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
982 Visited.insert(std::make_pair(I->getBB(), Addr));
983 if (!I->getResult().isNonLocal())
984 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
986 ++NumCacheCompleteNonLocalPtr;
990 // Otherwise, either this is a new block, a block with an invalid cache
991 // pointer or one that we're about to invalidate by putting more info into it
992 // than its valid cache info. If empty, the result will be valid cache info,
993 // otherwise it isn't.
995 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
997 CacheInfo->Pair = BBSkipFirstBlockPair();
999 SmallVector<BasicBlock*, 32> Worklist;
1000 Worklist.push_back(StartBB);
1002 // PredList used inside loop.
1003 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1005 // Keep track of the entries that we know are sorted. Previously cached
1006 // entries will all be sorted. The entries we add we only sort on demand (we
1007 // don't insert every element into its sorted position). We know that we
1008 // won't get any reuse from currently inserted values, because we don't
1009 // revisit blocks after we insert info for them.
1010 unsigned NumSortedEntries = Cache->size();
1011 DEBUG(AssertSorted(*Cache));
1013 while (!Worklist.empty()) {
1014 BasicBlock *BB = Worklist.pop_back_val();
1016 // Skip the first block if we have it.
1017 if (!SkipFirstBlock) {
1018 // Analyze the dependency of *Pointer in FromBB. See if we already have
1020 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1022 // Get the dependency info for Pointer in BB. If we have cached
1023 // information, we will use it, otherwise we compute it.
1024 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1025 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
1028 // If we got a Def or Clobber, add this to the list of results.
1029 if (!Dep.isNonLocal()) {
1030 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1035 // If 'Pointer' is an instruction defined in this block, then we need to do
1036 // phi translation to change it into a value live in the predecessor block.
1037 // If not, we just add the predecessors to the worklist and scan them with
1038 // the same Pointer.
1039 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1040 SkipFirstBlock = false;
1041 SmallVector<BasicBlock*, 16> NewBlocks;
1042 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1043 // Verify that we haven't looked at this block yet.
1044 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1045 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1046 if (InsertRes.second) {
1047 // First time we've looked at *PI.
1048 NewBlocks.push_back(*PI);
1052 // If we have seen this block before, but it was with a different
1053 // pointer then we have a phi translation failure and we have to treat
1054 // this as a clobber.
1055 if (InsertRes.first->second != Pointer.getAddr()) {
1056 // Make sure to clean up the Visited map before continuing on to
1057 // PredTranslationFailure.
1058 for (unsigned i = 0; i < NewBlocks.size(); i++)
1059 Visited.erase(NewBlocks[i]);
1060 goto PredTranslationFailure;
1063 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1067 // We do need to do phi translation, if we know ahead of time we can't phi
1068 // translate this value, don't even try.
1069 if (!Pointer.IsPotentiallyPHITranslatable())
1070 goto PredTranslationFailure;
1072 // We may have added values to the cache list before this PHI translation.
1073 // If so, we haven't done anything to ensure that the cache remains sorted.
1074 // Sort it now (if needed) so that recursive invocations of
1075 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1076 // value will only see properly sorted cache arrays.
1077 if (Cache && NumSortedEntries != Cache->size()) {
1078 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1079 NumSortedEntries = Cache->size();
1084 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1085 BasicBlock *Pred = *PI;
1086 PredList.push_back(std::make_pair(Pred, Pointer));
1088 // Get the PHI translated pointer in this predecessor. This can fail if
1089 // not translatable, in which case the getAddr() returns null.
1090 PHITransAddr &PredPointer = PredList.back().second;
1091 PredPointer.PHITranslateValue(BB, Pred, 0);
1093 Value *PredPtrVal = PredPointer.getAddr();
1095 // Check to see if we have already visited this pred block with another
1096 // pointer. If so, we can't do this lookup. This failure can occur
1097 // with PHI translation when a critical edge exists and the PHI node in
1098 // the successor translates to a pointer value different than the
1099 // pointer the block was first analyzed with.
1100 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1101 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1103 if (!InsertRes.second) {
1104 // We found the pred; take it off the list of preds to visit.
1105 PredList.pop_back();
1107 // If the predecessor was visited with PredPtr, then we already did
1108 // the analysis and can ignore it.
1109 if (InsertRes.first->second == PredPtrVal)
1112 // Otherwise, the block was previously analyzed with a different
1113 // pointer. We can't represent the result of this case, so we just
1114 // treat this as a phi translation failure.
1116 // Make sure to clean up the Visited map before continuing on to
1117 // PredTranslationFailure.
1118 for (unsigned i = 0; i < PredList.size(); i++)
1119 Visited.erase(PredList[i].first);
1121 goto PredTranslationFailure;
1125 // Actually process results here; this need to be a separate loop to avoid
1126 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1127 // any results for. (getNonLocalPointerDepFromBB will modify our
1128 // datastructures in ways the code after the PredTranslationFailure label
1130 for (unsigned i = 0; i < PredList.size(); i++) {
1131 BasicBlock *Pred = PredList[i].first;
1132 PHITransAddr &PredPointer = PredList[i].second;
1133 Value *PredPtrVal = PredPointer.getAddr();
1135 bool CanTranslate = true;
1136 // If PHI translation was unable to find an available pointer in this
1137 // predecessor, then we have to assume that the pointer is clobbered in
1138 // that predecessor. We can still do PRE of the load, which would insert
1139 // a computation of the pointer in this predecessor.
1140 if (PredPtrVal == 0)
1141 CanTranslate = false;
1143 // FIXME: it is entirely possible that PHI translating will end up with
1144 // the same value. Consider PHI translating something like:
1145 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1146 // to recurse here, pedantically speaking.
1148 // If getNonLocalPointerDepFromBB fails here, that means the cached
1149 // result conflicted with the Visited list; we have to conservatively
1150 // assume it is unknown, but this also does not block PRE of the load.
1151 if (!CanTranslate ||
1152 getNonLocalPointerDepFromBB(PredPointer,
1153 Loc.getWithNewPtr(PredPtrVal),
1156 // Add the entry to the Result list.
1157 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1158 Result.push_back(Entry);
1160 // Since we had a phi translation failure, the cache for CacheKey won't
1161 // include all of the entries that we need to immediately satisfy future
1162 // queries. Mark this in NonLocalPointerDeps by setting the
1163 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1164 // cached value to do more work but not miss the phi trans failure.
1165 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1166 NLPI.Pair = BBSkipFirstBlockPair();
1171 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1172 CacheInfo = &NonLocalPointerDeps[CacheKey];
1173 Cache = &CacheInfo->NonLocalDeps;
1174 NumSortedEntries = Cache->size();
1176 // Since we did phi translation, the "Cache" set won't contain all of the
1177 // results for the query. This is ok (we can still use it to accelerate
1178 // specific block queries) but we can't do the fastpath "return all
1179 // results from the set" Clear out the indicator for this.
1180 CacheInfo->Pair = BBSkipFirstBlockPair();
1181 SkipFirstBlock = false;
1184 PredTranslationFailure:
1185 // The following code is "failure"; we can't produce a sane translation
1186 // for the given block. It assumes that we haven't modified any of
1187 // our datastructures while processing the current block.
1190 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1191 CacheInfo = &NonLocalPointerDeps[CacheKey];
1192 Cache = &CacheInfo->NonLocalDeps;
1193 NumSortedEntries = Cache->size();
1196 // Since we failed phi translation, the "Cache" set won't contain all of the
1197 // results for the query. This is ok (we can still use it to accelerate
1198 // specific block queries) but we can't do the fastpath "return all
1199 // results from the set". Clear out the indicator for this.
1200 CacheInfo->Pair = BBSkipFirstBlockPair();
1202 // If *nothing* works, mark the pointer as unknown.
1204 // If this is the magic first block, return this as a clobber of the whole
1205 // incoming value. Since we can't phi translate to one of the predecessors,
1206 // we have to bail out.
1210 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1211 assert(I != Cache->rend() && "Didn't find current block??");
1212 if (I->getBB() != BB)
1215 assert(I->getResult().isNonLocal() &&
1216 "Should only be here with transparent block");
1217 I->setResult(MemDepResult::getUnknown());
1218 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1219 Pointer.getAddr()));
1224 // Okay, we're done now. If we added new values to the cache, re-sort it.
1225 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1226 DEBUG(AssertSorted(*Cache));
1230 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1231 /// CachedNonLocalPointerInfo, remove it.
1232 void MemoryDependenceAnalysis::
1233 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1234 CachedNonLocalPointerInfo::iterator It =
1235 NonLocalPointerDeps.find(P);
1236 if (It == NonLocalPointerDeps.end()) return;
1238 // Remove all of the entries in the BB->val map. This involves removing
1239 // instructions from the reverse map.
1240 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1242 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1243 Instruction *Target = PInfo[i].getResult().getInst();
1244 if (Target == 0) continue; // Ignore non-local dep results.
1245 assert(Target->getParent() == PInfo[i].getBB());
1247 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1248 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1251 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1252 NonLocalPointerDeps.erase(It);
1256 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1257 /// information about the specified pointer, because it may be too
1258 /// conservative in memdep. This is an optional call that can be used when
1259 /// the client detects an equivalence between the pointer and some other
1260 /// value and replaces the other value with ptr. This can make Ptr available
1261 /// in more places that cached info does not necessarily keep.
1262 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1263 // If Ptr isn't really a pointer, just ignore it.
1264 if (!Ptr->getType()->isPointerTy()) return;
1265 // Flush store info for the pointer.
1266 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1267 // Flush load info for the pointer.
1268 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1271 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1272 /// This needs to be done when the CFG changes, e.g., due to splitting
1274 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1278 /// removeInstruction - Remove an instruction from the dependence analysis,
1279 /// updating the dependence of instructions that previously depended on it.
1280 /// This method attempts to keep the cache coherent using the reverse map.
1281 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1282 // Walk through the Non-local dependencies, removing this one as the value
1283 // for any cached queries.
1284 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1285 if (NLDI != NonLocalDeps.end()) {
1286 NonLocalDepInfo &BlockMap = NLDI->second.first;
1287 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1289 if (Instruction *Inst = DI->getResult().getInst())
1290 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1291 NonLocalDeps.erase(NLDI);
1294 // If we have a cached local dependence query for this instruction, remove it.
1296 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1297 if (LocalDepEntry != LocalDeps.end()) {
1298 // Remove us from DepInst's reverse set now that the local dep info is gone.
1299 if (Instruction *Inst = LocalDepEntry->second.getInst())
1300 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1302 // Remove this local dependency info.
1303 LocalDeps.erase(LocalDepEntry);
1306 // If we have any cached pointer dependencies on this instruction, remove
1307 // them. If the instruction has non-pointer type, then it can't be a pointer
1310 // Remove it from both the load info and the store info. The instruction
1311 // can't be in either of these maps if it is non-pointer.
1312 if (RemInst->getType()->isPointerTy()) {
1313 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1314 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1317 // Loop over all of the things that depend on the instruction we're removing.
1319 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1321 // If we find RemInst as a clobber or Def in any of the maps for other values,
1322 // we need to replace its entry with a dirty version of the instruction after
1323 // it. If RemInst is a terminator, we use a null dirty value.
1325 // Using a dirty version of the instruction after RemInst saves having to scan
1326 // the entire block to get to this point.
1327 MemDepResult NewDirtyVal;
1328 if (!RemInst->isTerminator())
1329 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1331 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1332 if (ReverseDepIt != ReverseLocalDeps.end()) {
1333 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1334 // RemInst can't be the terminator if it has local stuff depending on it.
1335 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1336 "Nothing can locally depend on a terminator");
1338 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1339 E = ReverseDeps.end(); I != E; ++I) {
1340 Instruction *InstDependingOnRemInst = *I;
1341 assert(InstDependingOnRemInst != RemInst &&
1342 "Already removed our local dep info");
1344 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1346 // Make sure to remember that new things depend on NewDepInst.
1347 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1348 "a local dep on this if it is a terminator!");
1349 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1350 InstDependingOnRemInst));
1353 ReverseLocalDeps.erase(ReverseDepIt);
1355 // Add new reverse deps after scanning the set, to avoid invalidating the
1356 // 'ReverseDeps' reference.
1357 while (!ReverseDepsToAdd.empty()) {
1358 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1359 .insert(ReverseDepsToAdd.back().second);
1360 ReverseDepsToAdd.pop_back();
1364 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1365 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1366 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1367 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1369 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1371 PerInstNLInfo &INLD = NonLocalDeps[*I];
1372 // The information is now dirty!
1375 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1376 DE = INLD.first.end(); DI != DE; ++DI) {
1377 if (DI->getResult().getInst() != RemInst) continue;
1379 // Convert to a dirty entry for the subsequent instruction.
1380 DI->setResult(NewDirtyVal);
1382 if (Instruction *NextI = NewDirtyVal.getInst())
1383 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1387 ReverseNonLocalDeps.erase(ReverseDepIt);
1389 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1390 while (!ReverseDepsToAdd.empty()) {
1391 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1392 .insert(ReverseDepsToAdd.back().second);
1393 ReverseDepsToAdd.pop_back();
1397 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1398 // value in the NonLocalPointerDeps info.
1399 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1400 ReverseNonLocalPtrDeps.find(RemInst);
1401 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1402 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1403 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1405 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1406 E = Set.end(); I != E; ++I) {
1407 ValueIsLoadPair P = *I;
1408 assert(P.getPointer() != RemInst &&
1409 "Already removed NonLocalPointerDeps info for RemInst");
1411 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1413 // The cache is not valid for any specific block anymore.
1414 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1416 // Update any entries for RemInst to use the instruction after it.
1417 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1419 if (DI->getResult().getInst() != RemInst) continue;
1421 // Convert to a dirty entry for the subsequent instruction.
1422 DI->setResult(NewDirtyVal);
1424 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1425 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1428 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1429 // subsequent value may invalidate the sortedness.
1430 std::sort(NLPDI.begin(), NLPDI.end());
1433 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1435 while (!ReversePtrDepsToAdd.empty()) {
1436 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1437 .insert(ReversePtrDepsToAdd.back().second);
1438 ReversePtrDepsToAdd.pop_back();
1443 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1444 AA->deleteValue(RemInst);
1445 DEBUG(verifyRemoved(RemInst));
1447 /// verifyRemoved - Verify that the specified instruction does not occur
1448 /// in our internal data structures.
1449 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1450 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1451 E = LocalDeps.end(); I != E; ++I) {
1452 assert(I->first != D && "Inst occurs in data structures");
1453 assert(I->second.getInst() != D &&
1454 "Inst occurs in data structures");
1457 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1458 E = NonLocalPointerDeps.end(); I != E; ++I) {
1459 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1460 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1461 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1463 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1466 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1467 E = NonLocalDeps.end(); I != E; ++I) {
1468 assert(I->first != D && "Inst occurs in data structures");
1469 const PerInstNLInfo &INLD = I->second;
1470 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1471 EE = INLD.first.end(); II != EE; ++II)
1472 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1475 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1476 E = ReverseLocalDeps.end(); I != E; ++I) {
1477 assert(I->first != D && "Inst occurs in data structures");
1478 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1479 EE = I->second.end(); II != EE; ++II)
1480 assert(*II != D && "Inst occurs in data structures");
1483 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1484 E = ReverseNonLocalDeps.end();
1486 assert(I->first != D && "Inst occurs in data structures");
1487 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1488 EE = I->second.end(); II != EE; ++II)
1489 assert(*II != D && "Inst occurs in data structures");
1492 for (ReverseNonLocalPtrDepTy::const_iterator
1493 I = ReverseNonLocalPtrDeps.begin(),
1494 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1495 assert(I->first != D && "Inst occurs in rev NLPD map");
1497 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1498 E = I->second.end(); II != E; ++II)
1499 assert(*II != ValueIsLoadPair(D, false) &&
1500 *II != ValueIsLoadPair(D, true) &&
1501 "Inst occurs in ReverseNonLocalPtrDeps map");