1 //===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation -------------===//
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/ADT/STLExtras.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/Dominators.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/Analysis/PHITransAddr.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/PredIteratorCache.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 static const int BlockScanLimit = 100;
52 char MemoryDependenceAnalysis::ID = 0;
54 // Register this pass...
55 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
56 "Memory Dependence Analysis", false, true)
57 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
58 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
59 "Memory Dependence Analysis", false, true)
61 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
62 : FunctionPass(ID), PredCache(0) {
63 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
65 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
68 /// Clean up memory in between runs
69 void MemoryDependenceAnalysis::releaseMemory() {
72 NonLocalPointerDeps.clear();
73 ReverseLocalDeps.clear();
74 ReverseNonLocalDeps.clear();
75 ReverseNonLocalPtrDeps.clear();
81 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
83 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
85 AU.addRequiredTransitive<AliasAnalysis>();
88 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
89 AA = &getAnalysis<AliasAnalysis>();
90 TD = getAnalysisIfAvailable<DataLayout>();
91 DT = getAnalysisIfAvailable<DominatorTree>();
93 PredCache.reset(new PredIteratorCache());
97 /// RemoveFromReverseMap - This is a helper function that removes Val from
98 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
99 template <typename KeyTy>
100 static void RemoveFromReverseMap(DenseMap<Instruction*,
101 SmallPtrSet<KeyTy, 4> > &ReverseMap,
102 Instruction *Inst, KeyTy Val) {
103 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
104 InstIt = ReverseMap.find(Inst);
105 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
106 bool Found = InstIt->second.erase(Val);
107 assert(Found && "Invalid reverse map!"); (void)Found;
108 if (InstIt->second.empty())
109 ReverseMap.erase(InstIt);
112 /// GetLocation - If the given instruction references a specific memory
113 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
114 /// Return a ModRefInfo value describing the general behavior of the
117 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
118 AliasAnalysis::Location &Loc,
120 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
121 if (LI->isUnordered()) {
122 Loc = AA->getLocation(LI);
123 return AliasAnalysis::Ref;
125 if (LI->getOrdering() == Monotonic) {
126 Loc = AA->getLocation(LI);
127 return AliasAnalysis::ModRef;
129 Loc = AliasAnalysis::Location();
130 return AliasAnalysis::ModRef;
133 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
134 if (SI->isUnordered()) {
135 Loc = AA->getLocation(SI);
136 return AliasAnalysis::Mod;
138 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 DataLayout *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 DataLayout &TD) {
284 // We can only extend simple integer loads.
285 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
287 // Load widening is hostile to ThreadSanitizer: it may cause false positives
288 // or make the reports more cryptic (access sizes are wrong).
289 if (LI->getParent()->getParent()->getAttributes().
290 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeThread))
293 // Get the base of this load.
295 const Value *LIBase =
296 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, &TD);
298 // If the two pointers are not based on the same pointer, we can't tell that
300 if (LIBase != MemLocBase) return 0;
302 // Okay, the two values are based on the same pointer, but returned as
303 // no-alias. This happens when we have things like two byte loads at "P+1"
304 // and "P+3". Check to see if increasing the size of the "LI" load up to its
305 // alignment (or the largest native integer type) will allow us to load all
306 // the bits required by MemLoc.
308 // If MemLoc is before LI, then no widening of LI will help us out.
309 if (MemLocOffs < LIOffs) return 0;
311 // Get the alignment of the load in bytes. We assume that it is safe to load
312 // any legal integer up to this size without a problem. For example, if we're
313 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
314 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
316 unsigned LoadAlign = LI->getAlignment();
318 int64_t MemLocEnd = MemLocOffs+MemLocSize;
320 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
321 if (LIOffs+LoadAlign < MemLocEnd) return 0;
323 // This is the size of the load to try. Start with the next larger power of
325 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
326 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
329 // If this load size is bigger than our known alignment or would not fit
330 // into a native integer register, then we fail.
331 if (NewLoadByteSize > LoadAlign ||
332 !TD.fitsInLegalInteger(NewLoadByteSize*8))
335 if (LIOffs+NewLoadByteSize > MemLocEnd &&
336 LI->getParent()->getParent()->getAttributes().
337 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeAddress))
338 // We will be reading past the location accessed by the original program.
339 // While this is safe in a regular build, Address Safety analysis tools
340 // may start reporting false warnings. So, don't do widening.
343 // If a load of this width would include all of MemLoc, then we succeed.
344 if (LIOffs+NewLoadByteSize >= MemLocEnd)
345 return NewLoadByteSize;
347 NewLoadByteSize <<= 1;
351 /// getPointerDependencyFrom - Return the instruction on which a memory
352 /// location depends. If isLoad is true, this routine ignores may-aliases with
353 /// read-only operations. If isLoad is false, this routine ignores may-aliases
354 /// with reads from read-only locations. If possible, pass the query
355 /// instruction as well; this function may take advantage of the metadata
356 /// annotated to the query instruction to refine the result.
357 MemDepResult MemoryDependenceAnalysis::
358 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
359 BasicBlock::iterator ScanIt, BasicBlock *BB,
360 Instruction *QueryInst) {
362 const Value *MemLocBase = 0;
363 int64_t MemLocOffset = 0;
364 unsigned Limit = BlockScanLimit;
365 bool isInvariantLoad = false;
366 if (isLoad && QueryInst) {
367 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
368 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != 0)
369 isInvariantLoad = true;
372 // Walk backwards through the basic block, looking for dependencies.
373 while (ScanIt != BB->begin()) {
374 // Limit the amount of scanning we do so we don't end up with quadratic
375 // running time on extreme testcases.
378 return MemDepResult::getUnknown();
380 Instruction *Inst = --ScanIt;
382 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
383 // Debug intrinsics don't (and can't) cause dependences.
384 if (isa<DbgInfoIntrinsic>(II)) continue;
386 // If we reach a lifetime begin or end marker, then the query ends here
387 // because the value is undefined.
388 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
389 // FIXME: This only considers queries directly on the invariant-tagged
390 // pointer, not on query pointers that are indexed off of them. It'd
391 // be nice to handle that at some point (the right approach is to use
392 // GetPointerBaseWithConstantOffset).
393 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
395 return MemDepResult::getDef(II);
400 // Values depend on loads if the pointers are must aliased. This means that
401 // a load depends on another must aliased load from the same value.
402 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
403 // Atomic loads have complications involved.
404 // FIXME: This is overly conservative.
405 if (!LI->isUnordered())
406 return MemDepResult::getClobber(LI);
408 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
410 // If we found a pointer, check if it could be the same as our pointer.
411 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
414 if (R == AliasAnalysis::NoAlias) {
415 // If this is an over-aligned integer load (for example,
416 // "load i8* %P, align 4") see if it would obviously overlap with the
417 // queried location if widened to a larger load (e.g. if the queried
418 // location is 1 byte at P+1). If so, return it as a load/load
419 // clobber result, allowing the client to decide to widen the load if
421 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType()))
422 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() &&
423 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
424 MemLocOffset, LI, TD))
425 return MemDepResult::getClobber(Inst);
430 // Must aliased loads are defs of each other.
431 if (R == AliasAnalysis::MustAlias)
432 return MemDepResult::getDef(Inst);
434 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
435 // in terms of clobbering loads, but since it does this by looking
436 // at the clobbering load directly, it doesn't know about any
437 // phi translation that may have happened along the way.
439 // If we have a partial alias, then return this as a clobber for the
441 if (R == AliasAnalysis::PartialAlias)
442 return MemDepResult::getClobber(Inst);
445 // Random may-alias loads don't depend on each other without a
450 // Stores don't depend on other no-aliased accesses.
451 if (R == AliasAnalysis::NoAlias)
454 // Stores don't alias loads from read-only memory.
455 if (AA->pointsToConstantMemory(LoadLoc))
458 // Stores depend on may/must aliased loads.
459 return MemDepResult::getDef(Inst);
462 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
463 // Atomic stores have complications involved.
464 // FIXME: This is overly conservative.
465 if (!SI->isUnordered())
466 return MemDepResult::getClobber(SI);
468 // If alias analysis can tell that this store is guaranteed to not modify
469 // the query pointer, ignore it. Use getModRefInfo to handle cases where
470 // the query pointer points to constant memory etc.
471 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
474 // Ok, this store might clobber the query pointer. Check to see if it is
475 // a must alias: in this case, we want to return this as a def.
476 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
478 // If we found a pointer, check if it could be the same as our pointer.
479 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
481 if (R == AliasAnalysis::NoAlias)
483 if (R == AliasAnalysis::MustAlias)
484 return MemDepResult::getDef(Inst);
487 return MemDepResult::getClobber(Inst);
490 // If this is an allocation, and if we know that the accessed pointer is to
491 // the allocation, return Def. This means that there is no dependence and
492 // the access can be optimized based on that. For example, a load could
494 // Note: Only determine this to be a malloc if Inst is the malloc call, not
495 // a subsequent bitcast of the malloc call result. There can be stores to
496 // the malloced memory between the malloc call and its bitcast uses, and we
497 // need to continue scanning until the malloc call.
498 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
499 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
500 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD);
502 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
503 return MemDepResult::getDef(Inst);
504 // Be conservative if the accessed pointer may alias the allocation.
505 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
506 return MemDepResult::getClobber(Inst);
507 // If the allocation is not aliased and does not read memory (like
508 // strdup), it is safe to ignore.
509 if (isa<AllocaInst>(Inst) ||
510 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
514 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
515 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
516 // If necessary, perform additional analysis.
517 if (MR == AliasAnalysis::ModRef)
518 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
520 case AliasAnalysis::NoModRef:
521 // If the call has no effect on the queried pointer, just ignore it.
523 case AliasAnalysis::Mod:
524 return MemDepResult::getClobber(Inst);
525 case AliasAnalysis::Ref:
526 // If the call is known to never store to the pointer, and if this is a
527 // load query, we can safely ignore it (scan past it).
531 // Otherwise, there is a potential dependence. Return a clobber.
532 return MemDepResult::getClobber(Inst);
536 // No dependence found. If this is the entry block of the function, it is
537 // unknown, otherwise it is non-local.
538 if (BB != &BB->getParent()->getEntryBlock())
539 return MemDepResult::getNonLocal();
540 return MemDepResult::getNonFuncLocal();
543 /// getDependency - Return the instruction on which a memory operation
545 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
546 Instruction *ScanPos = QueryInst;
548 // Check for a cached result
549 MemDepResult &LocalCache = LocalDeps[QueryInst];
551 // If the cached entry is non-dirty, just return it. Note that this depends
552 // on MemDepResult's default constructing to 'dirty'.
553 if (!LocalCache.isDirty())
556 // Otherwise, if we have a dirty entry, we know we can start the scan at that
557 // instruction, which may save us some work.
558 if (Instruction *Inst = LocalCache.getInst()) {
561 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
564 BasicBlock *QueryParent = QueryInst->getParent();
567 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
568 // No dependence found. If this is the entry block of the function, it is
569 // unknown, otherwise it is non-local.
570 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
571 LocalCache = MemDepResult::getNonLocal();
573 LocalCache = MemDepResult::getNonFuncLocal();
575 AliasAnalysis::Location MemLoc;
576 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
578 // If we can do a pointer scan, make it happen.
579 bool isLoad = !(MR & AliasAnalysis::Mod);
580 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
581 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
583 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
584 QueryParent, QueryInst);
585 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
586 CallSite QueryCS(QueryInst);
587 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
588 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
591 // Non-memory instruction.
592 LocalCache = MemDepResult::getUnknown();
595 // Remember the result!
596 if (Instruction *I = LocalCache.getInst())
597 ReverseLocalDeps[I].insert(QueryInst);
603 /// AssertSorted - This method is used when -debug is specified to verify that
604 /// cache arrays are properly kept sorted.
605 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
607 if (Count == -1) Count = Cache.size();
608 if (Count == 0) return;
610 for (unsigned i = 1; i != unsigned(Count); ++i)
611 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
615 /// getNonLocalCallDependency - Perform a full dependency query for the
616 /// specified call, returning the set of blocks that the value is
617 /// potentially live across. The returned set of results will include a
618 /// "NonLocal" result for all blocks where the value is live across.
620 /// This method assumes the instruction returns a "NonLocal" dependency
621 /// within its own block.
623 /// This returns a reference to an internal data structure that may be
624 /// invalidated on the next non-local query or when an instruction is
625 /// removed. Clients must copy this data if they want it around longer than
627 const MemoryDependenceAnalysis::NonLocalDepInfo &
628 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
629 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
630 "getNonLocalCallDependency should only be used on calls with non-local deps!");
631 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
632 NonLocalDepInfo &Cache = CacheP.first;
634 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
635 /// the cached case, this can happen due to instructions being deleted etc. In
636 /// the uncached case, this starts out as the set of predecessors we care
638 SmallVector<BasicBlock*, 32> DirtyBlocks;
640 if (!Cache.empty()) {
641 // Okay, we have a cache entry. If we know it is not dirty, just return it
642 // with no computation.
643 if (!CacheP.second) {
648 // If we already have a partially computed set of results, scan them to
649 // determine what is dirty, seeding our initial DirtyBlocks worklist.
650 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
652 if (I->getResult().isDirty())
653 DirtyBlocks.push_back(I->getBB());
655 // Sort the cache so that we can do fast binary search lookups below.
656 std::sort(Cache.begin(), Cache.end());
658 ++NumCacheDirtyNonLocal;
659 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
660 // << Cache.size() << " cached: " << *QueryInst;
662 // Seed DirtyBlocks with each of the preds of QueryInst's block.
663 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
664 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
665 DirtyBlocks.push_back(*PI);
666 ++NumUncacheNonLocal;
669 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
670 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
672 SmallPtrSet<BasicBlock*, 64> Visited;
674 unsigned NumSortedEntries = Cache.size();
675 DEBUG(AssertSorted(Cache));
677 // Iterate while we still have blocks to update.
678 while (!DirtyBlocks.empty()) {
679 BasicBlock *DirtyBB = DirtyBlocks.back();
680 DirtyBlocks.pop_back();
682 // Already processed this block?
683 if (!Visited.insert(DirtyBB))
686 // Do a binary search to see if we already have an entry for this block in
687 // the cache set. If so, find it.
688 DEBUG(AssertSorted(Cache, NumSortedEntries));
689 NonLocalDepInfo::iterator Entry =
690 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
691 NonLocalDepEntry(DirtyBB));
692 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB)
695 NonLocalDepEntry *ExistingResult = 0;
696 if (Entry != Cache.begin()+NumSortedEntries &&
697 Entry->getBB() == DirtyBB) {
698 // If we already have an entry, and if it isn't already dirty, the block
700 if (!Entry->getResult().isDirty())
703 // Otherwise, remember this slot so we can update the value.
704 ExistingResult = &*Entry;
707 // If the dirty entry has a pointer, start scanning from it so we don't have
708 // to rescan the entire block.
709 BasicBlock::iterator ScanPos = DirtyBB->end();
710 if (ExistingResult) {
711 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
713 // We're removing QueryInst's use of Inst.
714 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
715 QueryCS.getInstruction());
719 // Find out if this block has a local dependency for QueryInst.
722 if (ScanPos != DirtyBB->begin()) {
723 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
724 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
725 // No dependence found. If this is the entry block of the function, it is
726 // a clobber, otherwise it is unknown.
727 Dep = MemDepResult::getNonLocal();
729 Dep = MemDepResult::getNonFuncLocal();
732 // If we had a dirty entry for the block, update it. Otherwise, just add
735 ExistingResult->setResult(Dep);
737 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
739 // If the block has a dependency (i.e. it isn't completely transparent to
740 // the value), remember the association!
741 if (!Dep.isNonLocal()) {
742 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
743 // update this when we remove instructions.
744 if (Instruction *Inst = Dep.getInst())
745 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
748 // If the block *is* completely transparent to the load, we need to check
749 // the predecessors of this block. Add them to our worklist.
750 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
751 DirtyBlocks.push_back(*PI);
758 /// getNonLocalPointerDependency - Perform a full dependency query for an
759 /// access to the specified (non-volatile) memory location, returning the
760 /// set of instructions that either define or clobber the value.
762 /// This method assumes the pointer has a "NonLocal" dependency within its
765 void MemoryDependenceAnalysis::
766 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
768 SmallVectorImpl<NonLocalDepResult> &Result) {
769 assert(Loc.Ptr->getType()->isPointerTy() &&
770 "Can't get pointer deps of a non-pointer!");
773 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD);
775 // This is the set of blocks we've inspected, and the pointer we consider in
776 // each block. Because of critical edges, we currently bail out if querying
777 // a block with multiple different pointers. This can happen during PHI
779 DenseMap<BasicBlock*, Value*> Visited;
780 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
781 Result, Visited, true))
784 Result.push_back(NonLocalDepResult(FromBB,
785 MemDepResult::getUnknown(),
786 const_cast<Value *>(Loc.Ptr)));
789 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
790 /// Pointer/PointeeSize using either cached information in Cache or by doing a
791 /// lookup (which may use dirty cache info if available). If we do a lookup,
792 /// add the result to the cache.
793 MemDepResult MemoryDependenceAnalysis::
794 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
795 bool isLoad, BasicBlock *BB,
796 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
798 // Do a binary search to see if we already have an entry for this block in
799 // the cache set. If so, find it.
800 NonLocalDepInfo::iterator Entry =
801 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
802 NonLocalDepEntry(BB));
803 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
806 NonLocalDepEntry *ExistingResult = 0;
807 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
808 ExistingResult = &*Entry;
810 // If we have a cached entry, and it is non-dirty, use it as the value for
812 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
813 ++NumCacheNonLocalPtr;
814 return ExistingResult->getResult();
817 // Otherwise, we have to scan for the value. If we have a dirty cache
818 // entry, start scanning from its position, otherwise we scan from the end
820 BasicBlock::iterator ScanPos = BB->end();
821 if (ExistingResult && ExistingResult->getResult().getInst()) {
822 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
823 "Instruction invalidated?");
824 ++NumCacheDirtyNonLocalPtr;
825 ScanPos = ExistingResult->getResult().getInst();
827 // Eliminating the dirty entry from 'Cache', so update the reverse info.
828 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
829 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
831 ++NumUncacheNonLocalPtr;
834 // Scan the block for the dependency.
835 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
837 // If we had a dirty entry for the block, update it. Otherwise, just add
840 ExistingResult->setResult(Dep);
842 Cache->push_back(NonLocalDepEntry(BB, Dep));
844 // If the block has a dependency (i.e. it isn't completely transparent to
845 // the value), remember the reverse association because we just added it
847 if (!Dep.isDef() && !Dep.isClobber())
850 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
851 // update MemDep when we remove instructions.
852 Instruction *Inst = Dep.getInst();
853 assert(Inst && "Didn't depend on anything?");
854 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
855 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
859 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
860 /// number of elements in the array that are already properly ordered. This is
861 /// optimized for the case when only a few entries are added.
863 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
864 unsigned NumSortedEntries) {
865 switch (Cache.size() - NumSortedEntries) {
867 // done, no new entries.
870 // Two new entries, insert the last one into place.
871 NonLocalDepEntry Val = Cache.back();
873 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
874 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
875 Cache.insert(Entry, Val);
879 // One new entry, Just insert the new value at the appropriate position.
880 if (Cache.size() != 1) {
881 NonLocalDepEntry Val = Cache.back();
883 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
884 std::upper_bound(Cache.begin(), Cache.end(), Val);
885 Cache.insert(Entry, Val);
889 // Added many values, do a full scale sort.
890 std::sort(Cache.begin(), Cache.end());
895 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
896 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
897 /// results to the results vector and keep track of which blocks are visited in
900 /// This has special behavior for the first block queries (when SkipFirstBlock
901 /// is true). In this special case, it ignores the contents of the specified
902 /// block and starts returning dependence info for its predecessors.
904 /// This function returns false on success, or true to indicate that it could
905 /// not compute dependence information for some reason. This should be treated
906 /// as a clobber dependence on the first instruction in the predecessor block.
907 bool MemoryDependenceAnalysis::
908 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
909 const AliasAnalysis::Location &Loc,
910 bool isLoad, BasicBlock *StartBB,
911 SmallVectorImpl<NonLocalDepResult> &Result,
912 DenseMap<BasicBlock*, Value*> &Visited,
913 bool SkipFirstBlock) {
914 // Look up the cached info for Pointer.
915 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
917 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
918 // CacheKey, this value will be inserted as the associated value. Otherwise,
919 // it'll be ignored, and we'll have to check to see if the cached size and
920 // tbaa tag are consistent with the current query.
921 NonLocalPointerInfo InitialNLPI;
922 InitialNLPI.Size = Loc.Size;
923 InitialNLPI.TBAATag = Loc.TBAATag;
925 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
927 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
928 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
929 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
931 // If we already have a cache entry for this CacheKey, we may need to do some
932 // work to reconcile the cache entry and the current query.
934 if (CacheInfo->Size < Loc.Size) {
935 // The query's Size is greater than the cached one. Throw out the
936 // cached data and proceed with the query at the greater size.
937 CacheInfo->Pair = BBSkipFirstBlockPair();
938 CacheInfo->Size = Loc.Size;
939 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
940 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
941 if (Instruction *Inst = DI->getResult().getInst())
942 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
943 CacheInfo->NonLocalDeps.clear();
944 } else if (CacheInfo->Size > Loc.Size) {
945 // This query's Size is less than the cached one. Conservatively restart
946 // the query using the greater size.
947 return getNonLocalPointerDepFromBB(Pointer,
948 Loc.getWithNewSize(CacheInfo->Size),
949 isLoad, StartBB, Result, Visited,
953 // If the query's TBAATag is inconsistent with the cached one,
954 // conservatively throw out the cached data and restart the query with
956 if (CacheInfo->TBAATag != Loc.TBAATag) {
957 if (CacheInfo->TBAATag) {
958 CacheInfo->Pair = BBSkipFirstBlockPair();
959 CacheInfo->TBAATag = 0;
960 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
961 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
962 if (Instruction *Inst = DI->getResult().getInst())
963 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
964 CacheInfo->NonLocalDeps.clear();
967 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(),
968 isLoad, StartBB, Result, Visited,
973 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
975 // If we have valid cached information for exactly the block we are
976 // investigating, just return it with no recomputation.
977 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
978 // We have a fully cached result for this query then we can just return the
979 // cached results and populate the visited set. However, we have to verify
980 // that we don't already have conflicting results for these blocks. Check
981 // to ensure that if a block in the results set is in the visited set that
982 // it was for the same pointer query.
983 if (!Visited.empty()) {
984 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
986 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
987 if (VI == Visited.end() || VI->second == Pointer.getAddr())
990 // We have a pointer mismatch in a block. Just return clobber, saying
991 // that something was clobbered in this result. We could also do a
992 // non-fully cached query, but there is little point in doing this.
997 Value *Addr = Pointer.getAddr();
998 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1000 Visited.insert(std::make_pair(I->getBB(), Addr));
1001 if (I->getResult().isNonLocal()) {
1006 Result.push_back(NonLocalDepResult(I->getBB(),
1007 MemDepResult::getUnknown(),
1009 } else if (DT->isReachableFromEntry(I->getBB())) {
1010 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1013 ++NumCacheCompleteNonLocalPtr;
1017 // Otherwise, either this is a new block, a block with an invalid cache
1018 // pointer or one that we're about to invalidate by putting more info into it
1019 // than its valid cache info. If empty, the result will be valid cache info,
1020 // otherwise it isn't.
1022 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1024 CacheInfo->Pair = BBSkipFirstBlockPair();
1026 SmallVector<BasicBlock*, 32> Worklist;
1027 Worklist.push_back(StartBB);
1029 // PredList used inside loop.
1030 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1032 // Keep track of the entries that we know are sorted. Previously cached
1033 // entries will all be sorted. The entries we add we only sort on demand (we
1034 // don't insert every element into its sorted position). We know that we
1035 // won't get any reuse from currently inserted values, because we don't
1036 // revisit blocks after we insert info for them.
1037 unsigned NumSortedEntries = Cache->size();
1038 DEBUG(AssertSorted(*Cache));
1040 while (!Worklist.empty()) {
1041 BasicBlock *BB = Worklist.pop_back_val();
1043 // Skip the first block if we have it.
1044 if (!SkipFirstBlock) {
1045 // Analyze the dependency of *Pointer in FromBB. See if we already have
1047 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1049 // Get the dependency info for Pointer in BB. If we have cached
1050 // information, we will use it, otherwise we compute it.
1051 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1052 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
1055 // If we got a Def or Clobber, add this to the list of results.
1056 if (!Dep.isNonLocal()) {
1058 Result.push_back(NonLocalDepResult(BB,
1059 MemDepResult::getUnknown(),
1060 Pointer.getAddr()));
1062 } else if (DT->isReachableFromEntry(BB)) {
1063 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1069 // If 'Pointer' is an instruction defined in this block, then we need to do
1070 // phi translation to change it into a value live in the predecessor block.
1071 // If not, we just add the predecessors to the worklist and scan them with
1072 // the same Pointer.
1073 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1074 SkipFirstBlock = false;
1075 SmallVector<BasicBlock*, 16> NewBlocks;
1076 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1077 // Verify that we haven't looked at this block yet.
1078 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1079 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1080 if (InsertRes.second) {
1081 // First time we've looked at *PI.
1082 NewBlocks.push_back(*PI);
1086 // If we have seen this block before, but it was with a different
1087 // pointer then we have a phi translation failure and we have to treat
1088 // this as a clobber.
1089 if (InsertRes.first->second != Pointer.getAddr()) {
1090 // Make sure to clean up the Visited map before continuing on to
1091 // PredTranslationFailure.
1092 for (unsigned i = 0; i < NewBlocks.size(); i++)
1093 Visited.erase(NewBlocks[i]);
1094 goto PredTranslationFailure;
1097 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1101 // We do need to do phi translation, if we know ahead of time we can't phi
1102 // translate this value, don't even try.
1103 if (!Pointer.IsPotentiallyPHITranslatable())
1104 goto PredTranslationFailure;
1106 // We may have added values to the cache list before this PHI translation.
1107 // If so, we haven't done anything to ensure that the cache remains sorted.
1108 // Sort it now (if needed) so that recursive invocations of
1109 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1110 // value will only see properly sorted cache arrays.
1111 if (Cache && NumSortedEntries != Cache->size()) {
1112 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1113 NumSortedEntries = Cache->size();
1118 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1119 BasicBlock *Pred = *PI;
1120 PredList.push_back(std::make_pair(Pred, Pointer));
1122 // Get the PHI translated pointer in this predecessor. This can fail if
1123 // not translatable, in which case the getAddr() returns null.
1124 PHITransAddr &PredPointer = PredList.back().second;
1125 PredPointer.PHITranslateValue(BB, Pred, 0);
1127 Value *PredPtrVal = PredPointer.getAddr();
1129 // Check to see if we have already visited this pred block with another
1130 // pointer. If so, we can't do this lookup. This failure can occur
1131 // with PHI translation when a critical edge exists and the PHI node in
1132 // the successor translates to a pointer value different than the
1133 // pointer the block was first analyzed with.
1134 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1135 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1137 if (!InsertRes.second) {
1138 // We found the pred; take it off the list of preds to visit.
1139 PredList.pop_back();
1141 // If the predecessor was visited with PredPtr, then we already did
1142 // the analysis and can ignore it.
1143 if (InsertRes.first->second == PredPtrVal)
1146 // Otherwise, the block was previously analyzed with a different
1147 // pointer. We can't represent the result of this case, so we just
1148 // treat this as a phi translation failure.
1150 // Make sure to clean up the Visited map before continuing on to
1151 // PredTranslationFailure.
1152 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1153 Visited.erase(PredList[i].first);
1155 goto PredTranslationFailure;
1159 // Actually process results here; this need to be a separate loop to avoid
1160 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1161 // any results for. (getNonLocalPointerDepFromBB will modify our
1162 // datastructures in ways the code after the PredTranslationFailure label
1164 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1165 BasicBlock *Pred = PredList[i].first;
1166 PHITransAddr &PredPointer = PredList[i].second;
1167 Value *PredPtrVal = PredPointer.getAddr();
1169 bool CanTranslate = true;
1170 // If PHI translation was unable to find an available pointer in this
1171 // predecessor, then we have to assume that the pointer is clobbered in
1172 // that predecessor. We can still do PRE of the load, which would insert
1173 // a computation of the pointer in this predecessor.
1174 if (PredPtrVal == 0)
1175 CanTranslate = false;
1177 // FIXME: it is entirely possible that PHI translating will end up with
1178 // the same value. Consider PHI translating something like:
1179 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1180 // to recurse here, pedantically speaking.
1182 // If getNonLocalPointerDepFromBB fails here, that means the cached
1183 // result conflicted with the Visited list; we have to conservatively
1184 // assume it is unknown, but this also does not block PRE of the load.
1185 if (!CanTranslate ||
1186 getNonLocalPointerDepFromBB(PredPointer,
1187 Loc.getWithNewPtr(PredPtrVal),
1190 // Add the entry to the Result list.
1191 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1192 Result.push_back(Entry);
1194 // Since we had a phi translation failure, the cache for CacheKey won't
1195 // include all of the entries that we need to immediately satisfy future
1196 // queries. Mark this in NonLocalPointerDeps by setting the
1197 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1198 // cached value to do more work but not miss the phi trans failure.
1199 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1200 NLPI.Pair = BBSkipFirstBlockPair();
1205 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1206 CacheInfo = &NonLocalPointerDeps[CacheKey];
1207 Cache = &CacheInfo->NonLocalDeps;
1208 NumSortedEntries = Cache->size();
1210 // Since we did phi translation, the "Cache" set won't contain all of the
1211 // results for the query. This is ok (we can still use it to accelerate
1212 // specific block queries) but we can't do the fastpath "return all
1213 // results from the set" Clear out the indicator for this.
1214 CacheInfo->Pair = BBSkipFirstBlockPair();
1215 SkipFirstBlock = false;
1218 PredTranslationFailure:
1219 // The following code is "failure"; we can't produce a sane translation
1220 // for the given block. It assumes that we haven't modified any of
1221 // our datastructures while processing the current block.
1224 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1225 CacheInfo = &NonLocalPointerDeps[CacheKey];
1226 Cache = &CacheInfo->NonLocalDeps;
1227 NumSortedEntries = Cache->size();
1230 // Since we failed phi translation, the "Cache" set won't contain all of the
1231 // results for the query. This is ok (we can still use it to accelerate
1232 // specific block queries) but we can't do the fastpath "return all
1233 // results from the set". Clear out the indicator for this.
1234 CacheInfo->Pair = BBSkipFirstBlockPair();
1236 // If *nothing* works, mark the pointer as unknown.
1238 // If this is the magic first block, return this as a clobber of the whole
1239 // incoming value. Since we can't phi translate to one of the predecessors,
1240 // we have to bail out.
1244 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1245 assert(I != Cache->rend() && "Didn't find current block??");
1246 if (I->getBB() != BB)
1249 assert(I->getResult().isNonLocal() &&
1250 "Should only be here with transparent block");
1251 I->setResult(MemDepResult::getUnknown());
1252 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1253 Pointer.getAddr()));
1258 // Okay, we're done now. If we added new values to the cache, re-sort it.
1259 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1260 DEBUG(AssertSorted(*Cache));
1264 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1265 /// CachedNonLocalPointerInfo, remove it.
1266 void MemoryDependenceAnalysis::
1267 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1268 CachedNonLocalPointerInfo::iterator It =
1269 NonLocalPointerDeps.find(P);
1270 if (It == NonLocalPointerDeps.end()) return;
1272 // Remove all of the entries in the BB->val map. This involves removing
1273 // instructions from the reverse map.
1274 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1276 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1277 Instruction *Target = PInfo[i].getResult().getInst();
1278 if (Target == 0) continue; // Ignore non-local dep results.
1279 assert(Target->getParent() == PInfo[i].getBB());
1281 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1282 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1285 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1286 NonLocalPointerDeps.erase(It);
1290 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1291 /// information about the specified pointer, because it may be too
1292 /// conservative in memdep. This is an optional call that can be used when
1293 /// the client detects an equivalence between the pointer and some other
1294 /// value and replaces the other value with ptr. This can make Ptr available
1295 /// in more places that cached info does not necessarily keep.
1296 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1297 // If Ptr isn't really a pointer, just ignore it.
1298 if (!Ptr->getType()->isPointerTy()) return;
1299 // Flush store info for the pointer.
1300 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1301 // Flush load info for the pointer.
1302 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1305 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1306 /// This needs to be done when the CFG changes, e.g., due to splitting
1308 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1312 /// removeInstruction - Remove an instruction from the dependence analysis,
1313 /// updating the dependence of instructions that previously depended on it.
1314 /// This method attempts to keep the cache coherent using the reverse map.
1315 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1316 // Walk through the Non-local dependencies, removing this one as the value
1317 // for any cached queries.
1318 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1319 if (NLDI != NonLocalDeps.end()) {
1320 NonLocalDepInfo &BlockMap = NLDI->second.first;
1321 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1323 if (Instruction *Inst = DI->getResult().getInst())
1324 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1325 NonLocalDeps.erase(NLDI);
1328 // If we have a cached local dependence query for this instruction, remove it.
1330 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1331 if (LocalDepEntry != LocalDeps.end()) {
1332 // Remove us from DepInst's reverse set now that the local dep info is gone.
1333 if (Instruction *Inst = LocalDepEntry->second.getInst())
1334 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1336 // Remove this local dependency info.
1337 LocalDeps.erase(LocalDepEntry);
1340 // If we have any cached pointer dependencies on this instruction, remove
1341 // them. If the instruction has non-pointer type, then it can't be a pointer
1344 // Remove it from both the load info and the store info. The instruction
1345 // can't be in either of these maps if it is non-pointer.
1346 if (RemInst->getType()->isPointerTy()) {
1347 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1348 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1351 // Loop over all of the things that depend on the instruction we're removing.
1353 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1355 // If we find RemInst as a clobber or Def in any of the maps for other values,
1356 // we need to replace its entry with a dirty version of the instruction after
1357 // it. If RemInst is a terminator, we use a null dirty value.
1359 // Using a dirty version of the instruction after RemInst saves having to scan
1360 // the entire block to get to this point.
1361 MemDepResult NewDirtyVal;
1362 if (!RemInst->isTerminator())
1363 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1365 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1366 if (ReverseDepIt != ReverseLocalDeps.end()) {
1367 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1368 // RemInst can't be the terminator if it has local stuff depending on it.
1369 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1370 "Nothing can locally depend on a terminator");
1372 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1373 E = ReverseDeps.end(); I != E; ++I) {
1374 Instruction *InstDependingOnRemInst = *I;
1375 assert(InstDependingOnRemInst != RemInst &&
1376 "Already removed our local dep info");
1378 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1380 // Make sure to remember that new things depend on NewDepInst.
1381 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1382 "a local dep on this if it is a terminator!");
1383 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1384 InstDependingOnRemInst));
1387 ReverseLocalDeps.erase(ReverseDepIt);
1389 // Add new reverse deps after scanning the set, to avoid invalidating the
1390 // 'ReverseDeps' reference.
1391 while (!ReverseDepsToAdd.empty()) {
1392 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1393 .insert(ReverseDepsToAdd.back().second);
1394 ReverseDepsToAdd.pop_back();
1398 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1399 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1400 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1401 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1403 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1405 PerInstNLInfo &INLD = NonLocalDeps[*I];
1406 // The information is now dirty!
1409 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1410 DE = INLD.first.end(); DI != DE; ++DI) {
1411 if (DI->getResult().getInst() != RemInst) continue;
1413 // Convert to a dirty entry for the subsequent instruction.
1414 DI->setResult(NewDirtyVal);
1416 if (Instruction *NextI = NewDirtyVal.getInst())
1417 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1421 ReverseNonLocalDeps.erase(ReverseDepIt);
1423 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1424 while (!ReverseDepsToAdd.empty()) {
1425 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1426 .insert(ReverseDepsToAdd.back().second);
1427 ReverseDepsToAdd.pop_back();
1431 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1432 // value in the NonLocalPointerDeps info.
1433 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1434 ReverseNonLocalPtrDeps.find(RemInst);
1435 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1436 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1437 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1439 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1440 E = Set.end(); I != E; ++I) {
1441 ValueIsLoadPair P = *I;
1442 assert(P.getPointer() != RemInst &&
1443 "Already removed NonLocalPointerDeps info for RemInst");
1445 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1447 // The cache is not valid for any specific block anymore.
1448 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1450 // Update any entries for RemInst to use the instruction after it.
1451 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1453 if (DI->getResult().getInst() != RemInst) continue;
1455 // Convert to a dirty entry for the subsequent instruction.
1456 DI->setResult(NewDirtyVal);
1458 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1459 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1462 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1463 // subsequent value may invalidate the sortedness.
1464 std::sort(NLPDI.begin(), NLPDI.end());
1467 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1469 while (!ReversePtrDepsToAdd.empty()) {
1470 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1471 .insert(ReversePtrDepsToAdd.back().second);
1472 ReversePtrDepsToAdd.pop_back();
1477 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1478 AA->deleteValue(RemInst);
1479 DEBUG(verifyRemoved(RemInst));
1481 /// verifyRemoved - Verify that the specified instruction does not occur
1482 /// in our internal data structures.
1483 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1484 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1485 E = LocalDeps.end(); I != E; ++I) {
1486 assert(I->first != D && "Inst occurs in data structures");
1487 assert(I->second.getInst() != D &&
1488 "Inst occurs in data structures");
1491 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1492 E = NonLocalPointerDeps.end(); I != E; ++I) {
1493 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1494 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1495 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1497 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1500 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1501 E = NonLocalDeps.end(); I != E; ++I) {
1502 assert(I->first != D && "Inst occurs in data structures");
1503 const PerInstNLInfo &INLD = I->second;
1504 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1505 EE = INLD.first.end(); II != EE; ++II)
1506 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1509 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1510 E = ReverseLocalDeps.end(); I != E; ++I) {
1511 assert(I->first != D && "Inst occurs in data structures");
1512 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1513 EE = I->second.end(); II != EE; ++II)
1514 assert(*II != D && "Inst occurs in data structures");
1517 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1518 E = ReverseNonLocalDeps.end();
1520 assert(I->first != D && "Inst occurs in data structures");
1521 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1522 EE = I->second.end(); II != EE; ++II)
1523 assert(*II != D && "Inst occurs in data structures");
1526 for (ReverseNonLocalPtrDepTy::const_iterator
1527 I = ReverseNonLocalPtrDeps.begin(),
1528 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1529 assert(I->first != D && "Inst occurs in rev NLPD map");
1531 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1532 E = I->second.end(); II != E; ++II)
1533 assert(*II != ValueIsLoadPair(D, false) &&
1534 *II != ValueIsLoadPair(D, true) &&
1535 "Inst occurs in ReverseNonLocalPtrDeps map");