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/Analysis/ValueTracking.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/Function.h"
23 #include "llvm/LLVMContext.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/Analysis/CaptureTracking.h"
26 #include "llvm/Analysis/Dominators.h"
27 #include "llvm/Analysis/InstructionSimplify.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/PHITransAddr.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/ADT/STLExtras.h"
33 #include "llvm/Support/PredIteratorCache.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Target/TargetData.h"
38 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
39 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
40 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
42 STATISTIC(NumCacheNonLocalPtr,
43 "Number of fully cached non-local ptr responses");
44 STATISTIC(NumCacheDirtyNonLocalPtr,
45 "Number of cached, but dirty, non-local ptr responses");
46 STATISTIC(NumUncacheNonLocalPtr,
47 "Number of uncached non-local ptr responses");
48 STATISTIC(NumCacheCompleteNonLocalPtr,
49 "Number of block queries that were completely cached");
51 // Limit for the number of instructions to scan in a block.
52 // FIXME: Figure out what a sane value is for this.
53 // (500 is relatively insane.)
54 static const int BlockScanLimit = 500;
56 char MemoryDependenceAnalysis::ID = 0;
58 // Register this pass...
59 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
60 "Memory Dependence Analysis", false, true)
61 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
62 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
63 "Memory Dependence Analysis", false, true)
65 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
66 : FunctionPass(ID), PredCache(0) {
67 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
69 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
72 /// Clean up memory in between runs
73 void MemoryDependenceAnalysis::releaseMemory() {
76 NonLocalPointerDeps.clear();
77 ReverseLocalDeps.clear();
78 ReverseNonLocalDeps.clear();
79 ReverseNonLocalPtrDeps.clear();
85 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
87 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
89 AU.addRequiredTransitive<AliasAnalysis>();
92 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
93 AA = &getAnalysis<AliasAnalysis>();
94 TD = getAnalysisIfAvailable<TargetData>();
95 DT = getAnalysisIfAvailable<DominatorTree>();
97 PredCache.reset(new PredIteratorCache());
101 /// RemoveFromReverseMap - This is a helper function that removes Val from
102 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
103 template <typename KeyTy>
104 static void RemoveFromReverseMap(DenseMap<Instruction*,
105 SmallPtrSet<KeyTy, 4> > &ReverseMap,
106 Instruction *Inst, KeyTy Val) {
107 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
108 InstIt = ReverseMap.find(Inst);
109 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
110 bool Found = InstIt->second.erase(Val);
111 assert(Found && "Invalid reverse map!"); (void)Found;
112 if (InstIt->second.empty())
113 ReverseMap.erase(InstIt);
116 /// GetLocation - If the given instruction references a specific memory
117 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
118 /// Return a ModRefInfo value describing the general behavior of the
121 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
122 AliasAnalysis::Location &Loc,
124 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
125 if (LI->isUnordered()) {
126 Loc = AA->getLocation(LI);
127 return AliasAnalysis::Ref;
128 } else if (LI->getOrdering() == Monotonic) {
129 Loc = AA->getLocation(LI);
130 return AliasAnalysis::ModRef;
132 Loc = AliasAnalysis::Location();
133 return AliasAnalysis::ModRef;
136 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
137 if (SI->isUnordered()) {
138 Loc = AA->getLocation(SI);
139 return AliasAnalysis::Mod;
140 } else if (SI->getOrdering() == Monotonic) {
141 Loc = AA->getLocation(SI);
142 return AliasAnalysis::ModRef;
144 Loc = AliasAnalysis::Location();
145 return AliasAnalysis::ModRef;
148 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
149 Loc = AA->getLocation(V);
150 return AliasAnalysis::ModRef;
153 if (const CallInst *CI = isFreeCall(Inst)) {
154 // calls to free() deallocate the entire structure
155 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
156 return AliasAnalysis::Mod;
159 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
160 switch (II->getIntrinsicID()) {
161 case Intrinsic::lifetime_start:
162 case Intrinsic::lifetime_end:
163 case Intrinsic::invariant_start:
164 Loc = AliasAnalysis::Location(II->getArgOperand(1),
165 cast<ConstantInt>(II->getArgOperand(0))
167 II->getMetadata(LLVMContext::MD_tbaa));
168 // These intrinsics don't really modify the memory, but returning Mod
169 // will allow them to be handled conservatively.
170 return AliasAnalysis::Mod;
171 case Intrinsic::invariant_end:
172 Loc = AliasAnalysis::Location(II->getArgOperand(2),
173 cast<ConstantInt>(II->getArgOperand(1))
175 II->getMetadata(LLVMContext::MD_tbaa));
176 // These intrinsics don't really modify the memory, but returning Mod
177 // will allow them to be handled conservatively.
178 return AliasAnalysis::Mod;
183 // Otherwise, just do the coarse-grained thing that always works.
184 if (Inst->mayWriteToMemory())
185 return AliasAnalysis::ModRef;
186 if (Inst->mayReadFromMemory())
187 return AliasAnalysis::Ref;
188 return AliasAnalysis::NoModRef;
191 /// getCallSiteDependencyFrom - Private helper for finding the local
192 /// dependencies of a call site.
193 MemDepResult MemoryDependenceAnalysis::
194 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
195 BasicBlock::iterator ScanIt, BasicBlock *BB) {
196 unsigned Limit = BlockScanLimit;
198 // Walk backwards through the block, looking for dependencies
199 while (ScanIt != BB->begin()) {
200 // Limit the amount of scanning we do so we don't end up with quadratic
201 // running time on extreme testcases.
204 return MemDepResult::getUnknown();
206 Instruction *Inst = --ScanIt;
208 // If this inst is a memory op, get the pointer it accessed
209 AliasAnalysis::Location Loc;
210 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
212 // A simple instruction.
213 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
214 return MemDepResult::getClobber(Inst);
218 if (CallSite InstCS = cast<Value>(Inst)) {
219 // Debug intrinsics don't cause dependences.
220 if (isa<DbgInfoIntrinsic>(Inst)) continue;
221 // If these two calls do not interfere, look past it.
222 switch (AA->getModRefInfo(CS, InstCS)) {
223 case AliasAnalysis::NoModRef:
224 // If the two calls are the same, return InstCS as a Def, so that
225 // CS can be found redundant and eliminated.
226 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
227 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
228 return MemDepResult::getDef(Inst);
230 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
234 return MemDepResult::getClobber(Inst);
239 // No dependence found. If this is the entry block of the function, it is
240 // unknown, otherwise it is non-local.
241 if (BB != &BB->getParent()->getEntryBlock())
242 return MemDepResult::getNonLocal();
243 return MemDepResult::getNonFuncLocal();
246 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
247 /// would fully overlap MemLoc if done as a wider legal integer load.
249 /// MemLocBase, MemLocOffset are lazily computed here the first time the
250 /// base/offs of memloc is needed.
252 isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc,
253 const Value *&MemLocBase,
256 const TargetData *TD) {
257 // If we have no target data, we can't do this.
258 if (TD == 0) return false;
260 // If we haven't already computed the base/offset of MemLoc, do so now.
262 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, *TD);
264 unsigned Size = MemoryDependenceAnalysis::
265 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size,
270 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
271 /// looks at a memory location for a load (specified by MemLocBase, Offs,
272 /// and Size) and compares it against a load. If the specified load could
273 /// be safely widened to a larger integer load that is 1) still efficient,
274 /// 2) safe for the target, and 3) would provide the specified memory
275 /// location value, then this function returns the size in bytes of the
276 /// load width to use. If not, this returns zero.
277 unsigned MemoryDependenceAnalysis::
278 getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs,
279 unsigned MemLocSize, const LoadInst *LI,
280 const TargetData &TD) {
281 // We can only extend simple integer loads.
282 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
284 // Get the base of this load.
286 const Value *LIBase =
287 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, TD);
289 // If the two pointers are not based on the same pointer, we can't tell that
291 if (LIBase != MemLocBase) return 0;
293 // Okay, the two values are based on the same pointer, but returned as
294 // no-alias. This happens when we have things like two byte loads at "P+1"
295 // and "P+3". Check to see if increasing the size of the "LI" load up to its
296 // alignment (or the largest native integer type) will allow us to load all
297 // the bits required by MemLoc.
299 // If MemLoc is before LI, then no widening of LI will help us out.
300 if (MemLocOffs < LIOffs) return 0;
302 // Get the alignment of the load in bytes. We assume that it is safe to load
303 // any legal integer up to this size without a problem. For example, if we're
304 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
305 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
307 unsigned LoadAlign = LI->getAlignment();
309 int64_t MemLocEnd = MemLocOffs+MemLocSize;
311 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
312 if (LIOffs+LoadAlign < MemLocEnd) return 0;
314 // This is the size of the load to try. Start with the next larger power of
316 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
317 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
320 // If this load size is bigger than our known alignment or would not fit
321 // into a native integer register, then we fail.
322 if (NewLoadByteSize > LoadAlign ||
323 !TD.fitsInLegalInteger(NewLoadByteSize*8))
326 if (LIOffs+NewLoadByteSize > MemLocEnd &&
327 LI->getParent()->getParent()->hasFnAttr(Attribute::AddressSafety)) {
328 // We will be reading past the location accessed by the original program.
329 // While this is safe in a regular build, Address Safety analysis tools
330 // may start reporting false warnings. So, don't do widening.
334 // If a load of this width would include all of MemLoc, then we succeed.
335 if (LIOffs+NewLoadByteSize >= MemLocEnd)
336 return NewLoadByteSize;
338 NewLoadByteSize <<= 1;
343 /// Only find pointer captures which happen before the given instruction. Uses
344 /// the dominator tree to determine whether one instruction is before another.
345 struct CapturesBefore : public CaptureTracker {
346 CapturesBefore(const Instruction *I, DominatorTree *DT)
347 : BeforeHere(I), DT(DT), Captured(false) {}
349 void tooManyUses() { Captured = true; }
351 bool shouldExplore(Use *U) {
352 Instruction *I = cast<Instruction>(U->getUser());
353 if (BeforeHere != I && DT->dominates(BeforeHere, I))
358 bool captured(Use *U) {
359 Instruction *I = cast<Instruction>(U->getUser());
360 if (BeforeHere != I && DT->dominates(BeforeHere, I))
366 const Instruction *BeforeHere;
373 AliasAnalysis::ModRefResult
374 MemoryDependenceAnalysis::getModRefInfo(const Instruction *Inst,
375 const AliasAnalysis::Location &MemLoc) {
376 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
377 if (MR != AliasAnalysis::ModRef) return MR;
379 // FIXME: this is really just shoring-up a deficiency in alias analysis.
380 // BasicAA isn't willing to spend linear time determining whether an alloca
381 // was captured before or after this particular call, while we are. However,
382 // with a smarter AA in place, this test is just wasting compile time.
383 if (!DT) return AliasAnalysis::ModRef;
384 const Value *Object = GetUnderlyingObject(MemLoc.Ptr, TD);
385 if (!isIdentifiedObject(Object) || isa<GlobalValue>(Object))
386 return AliasAnalysis::ModRef;
387 ImmutableCallSite CS(Inst);
388 if (!CS.getInstruction()) return AliasAnalysis::ModRef;
390 CapturesBefore CB(Inst, DT);
391 llvm::PointerMayBeCaptured(Object, &CB);
393 if (isa<Constant>(Object) || CS.getInstruction() == Object || CB.Captured)
394 return AliasAnalysis::ModRef;
397 for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
398 CI != CE; ++CI, ++ArgNo) {
399 // Only look at the no-capture or byval pointer arguments. If this
400 // pointer were passed to arguments that were neither of these, then it
401 // couldn't be no-capture.
402 if (!(*CI)->getType()->isPointerTy() ||
403 (!CS.doesNotCapture(ArgNo) && !CS.isByValArgument(ArgNo)))
406 // If this is a no-capture pointer argument, see if we can tell that it
407 // is impossible to alias the pointer we're checking. If not, we have to
408 // assume that the call could touch the pointer, even though it doesn't
410 if (!AA->isNoAlias(AliasAnalysis::Location(*CI),
411 AliasAnalysis::Location(Object))) {
412 return AliasAnalysis::ModRef;
415 return AliasAnalysis::NoModRef;
418 /// getPointerDependencyFrom - Return the instruction on which a memory
419 /// location depends. If isLoad is true, this routine ignores may-aliases with
420 /// read-only operations. If isLoad is false, this routine ignores may-aliases
421 /// with reads from read-only locations.
422 MemDepResult MemoryDependenceAnalysis::
423 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
424 BasicBlock::iterator ScanIt, BasicBlock *BB) {
426 const Value *MemLocBase = 0;
427 int64_t MemLocOffset = 0;
429 unsigned Limit = BlockScanLimit;
431 // Walk backwards through the basic block, looking for dependencies.
432 while (ScanIt != BB->begin()) {
433 // Limit the amount of scanning we do so we don't end up with quadratic
434 // running time on extreme testcases.
437 return MemDepResult::getUnknown();
439 Instruction *Inst = --ScanIt;
441 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
442 // Debug intrinsics don't (and can't) cause dependences.
443 if (isa<DbgInfoIntrinsic>(II)) continue;
445 // If we reach a lifetime begin or end marker, then the query ends here
446 // because the value is undefined.
447 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
448 // FIXME: This only considers queries directly on the invariant-tagged
449 // pointer, not on query pointers that are indexed off of them. It'd
450 // be nice to handle that at some point (the right approach is to use
451 // GetPointerBaseWithConstantOffset).
452 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
454 return MemDepResult::getDef(II);
459 // Values depend on loads if the pointers are must aliased. This means that
460 // a load depends on another must aliased load from the same value.
461 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
462 // Atomic loads have complications involved.
463 // FIXME: This is overly conservative.
464 if (!LI->isUnordered())
465 return MemDepResult::getClobber(LI);
467 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
469 // If we found a pointer, check if it could be the same as our pointer.
470 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
473 if (R == AliasAnalysis::NoAlias) {
474 // If this is an over-aligned integer load (for example,
475 // "load i8* %P, align 4") see if it would obviously overlap with the
476 // queried location if widened to a larger load (e.g. if the queried
477 // location is 1 byte at P+1). If so, return it as a load/load
478 // clobber result, allowing the client to decide to widen the load if
480 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType()))
481 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() &&
482 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
483 MemLocOffset, LI, TD))
484 return MemDepResult::getClobber(Inst);
489 // Must aliased loads are defs of each other.
490 if (R == AliasAnalysis::MustAlias)
491 return MemDepResult::getDef(Inst);
493 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
494 // in terms of clobbering loads, but since it does this by looking
495 // at the clobbering load directly, it doesn't know about any
496 // phi translation that may have happened along the way.
498 // If we have a partial alias, then return this as a clobber for the
500 if (R == AliasAnalysis::PartialAlias)
501 return MemDepResult::getClobber(Inst);
504 // Random may-alias loads don't depend on each other without a
509 // Stores don't depend on other no-aliased accesses.
510 if (R == AliasAnalysis::NoAlias)
513 // Stores don't alias loads from read-only memory.
514 if (AA->pointsToConstantMemory(LoadLoc))
517 // Stores depend on may/must aliased loads.
518 return MemDepResult::getDef(Inst);
521 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
522 // Atomic stores have complications involved.
523 // FIXME: This is overly conservative.
524 if (!SI->isUnordered())
525 return MemDepResult::getClobber(SI);
527 // If alias analysis can tell that this store is guaranteed to not modify
528 // the query pointer, ignore it. Use getModRefInfo to handle cases where
529 // the query pointer points to constant memory etc.
530 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
533 // Ok, this store might clobber the query pointer. Check to see if it is
534 // a must alias: in this case, we want to return this as a def.
535 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
537 // If we found a pointer, check if it could be the same as our pointer.
538 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
540 if (R == AliasAnalysis::NoAlias)
542 if (R == AliasAnalysis::MustAlias)
543 return MemDepResult::getDef(Inst);
544 return MemDepResult::getClobber(Inst);
547 // If this is an allocation, and if we know that the accessed pointer is to
548 // the allocation, return Def. This means that there is no dependence and
549 // the access can be optimized based on that. For example, a load could
551 // Note: Only determine this to be a malloc if Inst is the malloc call, not
552 // a subsequent bitcast of the malloc call result. There can be stores to
553 // the malloced memory between the malloc call and its bitcast uses, and we
554 // need to continue scanning until the malloc call.
555 if (isa<AllocaInst>(Inst) ||
556 (isa<CallInst>(Inst) && extractMallocCall(Inst))) {
557 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD);
559 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
560 return MemDepResult::getDef(Inst);
564 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
565 switch (getModRefInfo(Inst, MemLoc)) {
566 case AliasAnalysis::NoModRef:
567 // If the call has no effect on the queried pointer, just ignore it.
569 case AliasAnalysis::Mod:
570 return MemDepResult::getClobber(Inst);
571 case AliasAnalysis::Ref:
572 // If the call is known to never store to the pointer, and if this is a
573 // load query, we can safely ignore it (scan past it).
577 // Otherwise, there is a potential dependence. Return a clobber.
578 return MemDepResult::getClobber(Inst);
582 // No dependence found. If this is the entry block of the function, it is
583 // unknown, otherwise it is non-local.
584 if (BB != &BB->getParent()->getEntryBlock())
585 return MemDepResult::getNonLocal();
586 return MemDepResult::getNonFuncLocal();
589 /// getDependency - Return the instruction on which a memory operation
591 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
592 Instruction *ScanPos = QueryInst;
594 // Check for a cached result
595 MemDepResult &LocalCache = LocalDeps[QueryInst];
597 // If the cached entry is non-dirty, just return it. Note that this depends
598 // on MemDepResult's default constructing to 'dirty'.
599 if (!LocalCache.isDirty())
602 // Otherwise, if we have a dirty entry, we know we can start the scan at that
603 // instruction, which may save us some work.
604 if (Instruction *Inst = LocalCache.getInst()) {
607 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
610 BasicBlock *QueryParent = QueryInst->getParent();
613 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
614 // No dependence found. If this is the entry block of the function, it is
615 // unknown, otherwise it is non-local.
616 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
617 LocalCache = MemDepResult::getNonLocal();
619 LocalCache = MemDepResult::getNonFuncLocal();
621 AliasAnalysis::Location MemLoc;
622 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
624 // If we can do a pointer scan, make it happen.
625 bool isLoad = !(MR & AliasAnalysis::Mod);
626 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
627 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
629 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
631 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
632 CallSite QueryCS(QueryInst);
633 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
634 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
637 // Non-memory instruction.
638 LocalCache = MemDepResult::getUnknown();
641 // Remember the result!
642 if (Instruction *I = LocalCache.getInst())
643 ReverseLocalDeps[I].insert(QueryInst);
649 /// AssertSorted - This method is used when -debug is specified to verify that
650 /// cache arrays are properly kept sorted.
651 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
653 if (Count == -1) Count = Cache.size();
654 if (Count == 0) return;
656 for (unsigned i = 1; i != unsigned(Count); ++i)
657 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
661 /// getNonLocalCallDependency - Perform a full dependency query for the
662 /// specified call, returning the set of blocks that the value is
663 /// potentially live across. The returned set of results will include a
664 /// "NonLocal" result for all blocks where the value is live across.
666 /// This method assumes the instruction returns a "NonLocal" dependency
667 /// within its own block.
669 /// This returns a reference to an internal data structure that may be
670 /// invalidated on the next non-local query or when an instruction is
671 /// removed. Clients must copy this data if they want it around longer than
673 const MemoryDependenceAnalysis::NonLocalDepInfo &
674 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
675 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
676 "getNonLocalCallDependency should only be used on calls with non-local deps!");
677 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
678 NonLocalDepInfo &Cache = CacheP.first;
680 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
681 /// the cached case, this can happen due to instructions being deleted etc. In
682 /// the uncached case, this starts out as the set of predecessors we care
684 SmallVector<BasicBlock*, 32> DirtyBlocks;
686 if (!Cache.empty()) {
687 // Okay, we have a cache entry. If we know it is not dirty, just return it
688 // with no computation.
689 if (!CacheP.second) {
694 // If we already have a partially computed set of results, scan them to
695 // determine what is dirty, seeding our initial DirtyBlocks worklist.
696 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
698 if (I->getResult().isDirty())
699 DirtyBlocks.push_back(I->getBB());
701 // Sort the cache so that we can do fast binary search lookups below.
702 std::sort(Cache.begin(), Cache.end());
704 ++NumCacheDirtyNonLocal;
705 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
706 // << Cache.size() << " cached: " << *QueryInst;
708 // Seed DirtyBlocks with each of the preds of QueryInst's block.
709 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
710 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
711 DirtyBlocks.push_back(*PI);
712 ++NumUncacheNonLocal;
715 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
716 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
718 SmallPtrSet<BasicBlock*, 64> Visited;
720 unsigned NumSortedEntries = Cache.size();
721 DEBUG(AssertSorted(Cache));
723 // Iterate while we still have blocks to update.
724 while (!DirtyBlocks.empty()) {
725 BasicBlock *DirtyBB = DirtyBlocks.back();
726 DirtyBlocks.pop_back();
728 // Already processed this block?
729 if (!Visited.insert(DirtyBB))
732 // Do a binary search to see if we already have an entry for this block in
733 // the cache set. If so, find it.
734 DEBUG(AssertSorted(Cache, NumSortedEntries));
735 NonLocalDepInfo::iterator Entry =
736 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
737 NonLocalDepEntry(DirtyBB));
738 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB)
741 NonLocalDepEntry *ExistingResult = 0;
742 if (Entry != Cache.begin()+NumSortedEntries &&
743 Entry->getBB() == DirtyBB) {
744 // If we already have an entry, and if it isn't already dirty, the block
746 if (!Entry->getResult().isDirty())
749 // Otherwise, remember this slot so we can update the value.
750 ExistingResult = &*Entry;
753 // If the dirty entry has a pointer, start scanning from it so we don't have
754 // to rescan the entire block.
755 BasicBlock::iterator ScanPos = DirtyBB->end();
756 if (ExistingResult) {
757 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
759 // We're removing QueryInst's use of Inst.
760 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
761 QueryCS.getInstruction());
765 // Find out if this block has a local dependency for QueryInst.
768 if (ScanPos != DirtyBB->begin()) {
769 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
770 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
771 // No dependence found. If this is the entry block of the function, it is
772 // a clobber, otherwise it is unknown.
773 Dep = MemDepResult::getNonLocal();
775 Dep = MemDepResult::getNonFuncLocal();
778 // If we had a dirty entry for the block, update it. Otherwise, just add
781 ExistingResult->setResult(Dep);
783 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
785 // If the block has a dependency (i.e. it isn't completely transparent to
786 // the value), remember the association!
787 if (!Dep.isNonLocal()) {
788 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
789 // update this when we remove instructions.
790 if (Instruction *Inst = Dep.getInst())
791 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
794 // If the block *is* completely transparent to the load, we need to check
795 // the predecessors of this block. Add them to our worklist.
796 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
797 DirtyBlocks.push_back(*PI);
804 /// getNonLocalPointerDependency - Perform a full dependency query for an
805 /// access to the specified (non-volatile) memory location, returning the
806 /// set of instructions that either define or clobber the value.
808 /// This method assumes the pointer has a "NonLocal" dependency within its
811 void MemoryDependenceAnalysis::
812 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
814 SmallVectorImpl<NonLocalDepResult> &Result) {
815 assert(Loc.Ptr->getType()->isPointerTy() &&
816 "Can't get pointer deps of a non-pointer!");
819 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD);
821 // This is the set of blocks we've inspected, and the pointer we consider in
822 // each block. Because of critical edges, we currently bail out if querying
823 // a block with multiple different pointers. This can happen during PHI
825 DenseMap<BasicBlock*, Value*> Visited;
826 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
827 Result, Visited, true))
830 Result.push_back(NonLocalDepResult(FromBB,
831 MemDepResult::getUnknown(),
832 const_cast<Value *>(Loc.Ptr)));
835 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
836 /// Pointer/PointeeSize using either cached information in Cache or by doing a
837 /// lookup (which may use dirty cache info if available). If we do a lookup,
838 /// add the result to the cache.
839 MemDepResult MemoryDependenceAnalysis::
840 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
841 bool isLoad, BasicBlock *BB,
842 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
844 // Do a binary search to see if we already have an entry for this block in
845 // the cache set. If so, find it.
846 NonLocalDepInfo::iterator Entry =
847 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
848 NonLocalDepEntry(BB));
849 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
852 NonLocalDepEntry *ExistingResult = 0;
853 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
854 ExistingResult = &*Entry;
856 // If we have a cached entry, and it is non-dirty, use it as the value for
858 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
859 ++NumCacheNonLocalPtr;
860 return ExistingResult->getResult();
863 // Otherwise, we have to scan for the value. If we have a dirty cache
864 // entry, start scanning from its position, otherwise we scan from the end
866 BasicBlock::iterator ScanPos = BB->end();
867 if (ExistingResult && ExistingResult->getResult().getInst()) {
868 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
869 "Instruction invalidated?");
870 ++NumCacheDirtyNonLocalPtr;
871 ScanPos = ExistingResult->getResult().getInst();
873 // Eliminating the dirty entry from 'Cache', so update the reverse info.
874 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
875 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
877 ++NumUncacheNonLocalPtr;
880 // Scan the block for the dependency.
881 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
883 // If we had a dirty entry for the block, update it. Otherwise, just add
886 ExistingResult->setResult(Dep);
888 Cache->push_back(NonLocalDepEntry(BB, Dep));
890 // If the block has a dependency (i.e. it isn't completely transparent to
891 // the value), remember the reverse association because we just added it
893 if (!Dep.isDef() && !Dep.isClobber())
896 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
897 // update MemDep when we remove instructions.
898 Instruction *Inst = Dep.getInst();
899 assert(Inst && "Didn't depend on anything?");
900 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
901 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
905 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
906 /// number of elements in the array that are already properly ordered. This is
907 /// optimized for the case when only a few entries are added.
909 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
910 unsigned NumSortedEntries) {
911 switch (Cache.size() - NumSortedEntries) {
913 // done, no new entries.
916 // Two new entries, insert the last one into place.
917 NonLocalDepEntry Val = Cache.back();
919 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
920 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
921 Cache.insert(Entry, Val);
925 // One new entry, Just insert the new value at the appropriate position.
926 if (Cache.size() != 1) {
927 NonLocalDepEntry Val = Cache.back();
929 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
930 std::upper_bound(Cache.begin(), Cache.end(), Val);
931 Cache.insert(Entry, Val);
935 // Added many values, do a full scale sort.
936 std::sort(Cache.begin(), Cache.end());
941 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
942 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
943 /// results to the results vector and keep track of which blocks are visited in
946 /// This has special behavior for the first block queries (when SkipFirstBlock
947 /// is true). In this special case, it ignores the contents of the specified
948 /// block and starts returning dependence info for its predecessors.
950 /// This function returns false on success, or true to indicate that it could
951 /// not compute dependence information for some reason. This should be treated
952 /// as a clobber dependence on the first instruction in the predecessor block.
953 bool MemoryDependenceAnalysis::
954 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
955 const AliasAnalysis::Location &Loc,
956 bool isLoad, BasicBlock *StartBB,
957 SmallVectorImpl<NonLocalDepResult> &Result,
958 DenseMap<BasicBlock*, Value*> &Visited,
959 bool SkipFirstBlock) {
961 // Look up the cached info for Pointer.
962 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
964 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
965 // CacheKey, this value will be inserted as the associated value. Otherwise,
966 // it'll be ignored, and we'll have to check to see if the cached size and
967 // tbaa tag are consistent with the current query.
968 NonLocalPointerInfo InitialNLPI;
969 InitialNLPI.Size = Loc.Size;
970 InitialNLPI.TBAATag = Loc.TBAATag;
972 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
974 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
975 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
976 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
978 // If we already have a cache entry for this CacheKey, we may need to do some
979 // work to reconcile the cache entry and the current query.
981 if (CacheInfo->Size < Loc.Size) {
982 // The query's Size is greater than the cached one. Throw out the
983 // cached data and procede with the query at the greater size.
984 CacheInfo->Pair = BBSkipFirstBlockPair();
985 CacheInfo->Size = Loc.Size;
986 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
987 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
988 if (Instruction *Inst = DI->getResult().getInst())
989 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
990 CacheInfo->NonLocalDeps.clear();
991 } else if (CacheInfo->Size > Loc.Size) {
992 // This query's Size is less than the cached one. Conservatively restart
993 // the query using the greater size.
994 return getNonLocalPointerDepFromBB(Pointer,
995 Loc.getWithNewSize(CacheInfo->Size),
996 isLoad, StartBB, Result, Visited,
1000 // If the query's TBAATag is inconsistent with the cached one,
1001 // conservatively throw out the cached data and restart the query with
1002 // no tag if needed.
1003 if (CacheInfo->TBAATag != Loc.TBAATag) {
1004 if (CacheInfo->TBAATag) {
1005 CacheInfo->Pair = BBSkipFirstBlockPair();
1006 CacheInfo->TBAATag = 0;
1007 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1008 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1009 if (Instruction *Inst = DI->getResult().getInst())
1010 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1011 CacheInfo->NonLocalDeps.clear();
1014 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(),
1015 isLoad, StartBB, Result, Visited,
1020 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1022 // If we have valid cached information for exactly the block we are
1023 // investigating, just return it with no recomputation.
1024 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1025 // We have a fully cached result for this query then we can just return the
1026 // cached results and populate the visited set. However, we have to verify
1027 // that we don't already have conflicting results for these blocks. Check
1028 // to ensure that if a block in the results set is in the visited set that
1029 // it was for the same pointer query.
1030 if (!Visited.empty()) {
1031 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1033 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
1034 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1037 // We have a pointer mismatch in a block. Just return clobber, saying
1038 // that something was clobbered in this result. We could also do a
1039 // non-fully cached query, but there is little point in doing this.
1044 Value *Addr = Pointer.getAddr();
1045 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1047 Visited.insert(std::make_pair(I->getBB(), Addr));
1048 if (!I->getResult().isNonLocal())
1049 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1051 ++NumCacheCompleteNonLocalPtr;
1055 // Otherwise, either this is a new block, a block with an invalid cache
1056 // pointer or one that we're about to invalidate by putting more info into it
1057 // than its valid cache info. If empty, the result will be valid cache info,
1058 // otherwise it isn't.
1060 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1062 CacheInfo->Pair = BBSkipFirstBlockPair();
1064 SmallVector<BasicBlock*, 32> Worklist;
1065 Worklist.push_back(StartBB);
1067 // PredList used inside loop.
1068 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1070 // Keep track of the entries that we know are sorted. Previously cached
1071 // entries will all be sorted. The entries we add we only sort on demand (we
1072 // don't insert every element into its sorted position). We know that we
1073 // won't get any reuse from currently inserted values, because we don't
1074 // revisit blocks after we insert info for them.
1075 unsigned NumSortedEntries = Cache->size();
1076 DEBUG(AssertSorted(*Cache));
1078 while (!Worklist.empty()) {
1079 BasicBlock *BB = Worklist.pop_back_val();
1081 // Skip the first block if we have it.
1082 if (!SkipFirstBlock) {
1083 // Analyze the dependency of *Pointer in FromBB. See if we already have
1085 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1087 // Get the dependency info for Pointer in BB. If we have cached
1088 // information, we will use it, otherwise we compute it.
1089 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1090 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
1093 // If we got a Def or Clobber, add this to the list of results.
1094 if (!Dep.isNonLocal()) {
1095 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1100 // If 'Pointer' is an instruction defined in this block, then we need to do
1101 // phi translation to change it into a value live in the predecessor block.
1102 // If not, we just add the predecessors to the worklist and scan them with
1103 // the same Pointer.
1104 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1105 SkipFirstBlock = false;
1106 SmallVector<BasicBlock*, 16> NewBlocks;
1107 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1108 // Verify that we haven't looked at this block yet.
1109 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1110 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1111 if (InsertRes.second) {
1112 // First time we've looked at *PI.
1113 NewBlocks.push_back(*PI);
1117 // If we have seen this block before, but it was with a different
1118 // pointer then we have a phi translation failure and we have to treat
1119 // this as a clobber.
1120 if (InsertRes.first->second != Pointer.getAddr()) {
1121 // Make sure to clean up the Visited map before continuing on to
1122 // PredTranslationFailure.
1123 for (unsigned i = 0; i < NewBlocks.size(); i++)
1124 Visited.erase(NewBlocks[i]);
1125 goto PredTranslationFailure;
1128 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1132 // We do need to do phi translation, if we know ahead of time we can't phi
1133 // translate this value, don't even try.
1134 if (!Pointer.IsPotentiallyPHITranslatable())
1135 goto PredTranslationFailure;
1137 // We may have added values to the cache list before this PHI translation.
1138 // If so, we haven't done anything to ensure that the cache remains sorted.
1139 // Sort it now (if needed) so that recursive invocations of
1140 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1141 // value will only see properly sorted cache arrays.
1142 if (Cache && NumSortedEntries != Cache->size()) {
1143 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1144 NumSortedEntries = Cache->size();
1149 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1150 BasicBlock *Pred = *PI;
1151 PredList.push_back(std::make_pair(Pred, Pointer));
1153 // Get the PHI translated pointer in this predecessor. This can fail if
1154 // not translatable, in which case the getAddr() returns null.
1155 PHITransAddr &PredPointer = PredList.back().second;
1156 PredPointer.PHITranslateValue(BB, Pred, 0);
1158 Value *PredPtrVal = PredPointer.getAddr();
1160 // Check to see if we have already visited this pred block with another
1161 // pointer. If so, we can't do this lookup. This failure can occur
1162 // with PHI translation when a critical edge exists and the PHI node in
1163 // the successor translates to a pointer value different than the
1164 // pointer the block was first analyzed with.
1165 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1166 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1168 if (!InsertRes.second) {
1169 // We found the pred; take it off the list of preds to visit.
1170 PredList.pop_back();
1172 // If the predecessor was visited with PredPtr, then we already did
1173 // the analysis and can ignore it.
1174 if (InsertRes.first->second == PredPtrVal)
1177 // Otherwise, the block was previously analyzed with a different
1178 // pointer. We can't represent the result of this case, so we just
1179 // treat this as a phi translation failure.
1181 // Make sure to clean up the Visited map before continuing on to
1182 // PredTranslationFailure.
1183 for (unsigned i = 0; i < PredList.size(); i++)
1184 Visited.erase(PredList[i].first);
1186 goto PredTranslationFailure;
1190 // Actually process results here; this need to be a separate loop to avoid
1191 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1192 // any results for. (getNonLocalPointerDepFromBB will modify our
1193 // datastructures in ways the code after the PredTranslationFailure label
1195 for (unsigned i = 0; i < PredList.size(); i++) {
1196 BasicBlock *Pred = PredList[i].first;
1197 PHITransAddr &PredPointer = PredList[i].second;
1198 Value *PredPtrVal = PredPointer.getAddr();
1200 bool CanTranslate = true;
1201 // If PHI translation was unable to find an available pointer in this
1202 // predecessor, then we have to assume that the pointer is clobbered in
1203 // that predecessor. We can still do PRE of the load, which would insert
1204 // a computation of the pointer in this predecessor.
1205 if (PredPtrVal == 0)
1206 CanTranslate = false;
1208 // FIXME: it is entirely possible that PHI translating will end up with
1209 // the same value. Consider PHI translating something like:
1210 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1211 // to recurse here, pedantically speaking.
1213 // If getNonLocalPointerDepFromBB fails here, that means the cached
1214 // result conflicted with the Visited list; we have to conservatively
1215 // assume it is unknown, but this also does not block PRE of the load.
1216 if (!CanTranslate ||
1217 getNonLocalPointerDepFromBB(PredPointer,
1218 Loc.getWithNewPtr(PredPtrVal),
1221 // Add the entry to the Result list.
1222 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1223 Result.push_back(Entry);
1225 // Since we had a phi translation failure, the cache for CacheKey won't
1226 // include all of the entries that we need to immediately satisfy future
1227 // queries. Mark this in NonLocalPointerDeps by setting the
1228 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1229 // cached value to do more work but not miss the phi trans failure.
1230 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1231 NLPI.Pair = BBSkipFirstBlockPair();
1236 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1237 CacheInfo = &NonLocalPointerDeps[CacheKey];
1238 Cache = &CacheInfo->NonLocalDeps;
1239 NumSortedEntries = Cache->size();
1241 // Since we did phi translation, the "Cache" set won't contain all of the
1242 // results for the query. This is ok (we can still use it to accelerate
1243 // specific block queries) but we can't do the fastpath "return all
1244 // results from the set" Clear out the indicator for this.
1245 CacheInfo->Pair = BBSkipFirstBlockPair();
1246 SkipFirstBlock = false;
1249 PredTranslationFailure:
1250 // The following code is "failure"; we can't produce a sane translation
1251 // for the given block. It assumes that we haven't modified any of
1252 // our datastructures while processing the current block.
1255 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1256 CacheInfo = &NonLocalPointerDeps[CacheKey];
1257 Cache = &CacheInfo->NonLocalDeps;
1258 NumSortedEntries = Cache->size();
1261 // Since we failed phi translation, the "Cache" set won't contain all of the
1262 // results for the query. This is ok (we can still use it to accelerate
1263 // specific block queries) but we can't do the fastpath "return all
1264 // results from the set". Clear out the indicator for this.
1265 CacheInfo->Pair = BBSkipFirstBlockPair();
1267 // If *nothing* works, mark the pointer as unknown.
1269 // If this is the magic first block, return this as a clobber of the whole
1270 // incoming value. Since we can't phi translate to one of the predecessors,
1271 // we have to bail out.
1275 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1276 assert(I != Cache->rend() && "Didn't find current block??");
1277 if (I->getBB() != BB)
1280 assert(I->getResult().isNonLocal() &&
1281 "Should only be here with transparent block");
1282 I->setResult(MemDepResult::getUnknown());
1283 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1284 Pointer.getAddr()));
1289 // Okay, we're done now. If we added new values to the cache, re-sort it.
1290 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1291 DEBUG(AssertSorted(*Cache));
1295 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1296 /// CachedNonLocalPointerInfo, remove it.
1297 void MemoryDependenceAnalysis::
1298 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1299 CachedNonLocalPointerInfo::iterator It =
1300 NonLocalPointerDeps.find(P);
1301 if (It == NonLocalPointerDeps.end()) return;
1303 // Remove all of the entries in the BB->val map. This involves removing
1304 // instructions from the reverse map.
1305 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1307 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1308 Instruction *Target = PInfo[i].getResult().getInst();
1309 if (Target == 0) continue; // Ignore non-local dep results.
1310 assert(Target->getParent() == PInfo[i].getBB());
1312 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1313 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1316 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1317 NonLocalPointerDeps.erase(It);
1321 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1322 /// information about the specified pointer, because it may be too
1323 /// conservative in memdep. This is an optional call that can be used when
1324 /// the client detects an equivalence between the pointer and some other
1325 /// value and replaces the other value with ptr. This can make Ptr available
1326 /// in more places that cached info does not necessarily keep.
1327 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1328 // If Ptr isn't really a pointer, just ignore it.
1329 if (!Ptr->getType()->isPointerTy()) return;
1330 // Flush store info for the pointer.
1331 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1332 // Flush load info for the pointer.
1333 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1336 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1337 /// This needs to be done when the CFG changes, e.g., due to splitting
1339 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1343 /// removeInstruction - Remove an instruction from the dependence analysis,
1344 /// updating the dependence of instructions that previously depended on it.
1345 /// This method attempts to keep the cache coherent using the reverse map.
1346 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1347 // Walk through the Non-local dependencies, removing this one as the value
1348 // for any cached queries.
1349 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1350 if (NLDI != NonLocalDeps.end()) {
1351 NonLocalDepInfo &BlockMap = NLDI->second.first;
1352 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1354 if (Instruction *Inst = DI->getResult().getInst())
1355 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1356 NonLocalDeps.erase(NLDI);
1359 // If we have a cached local dependence query for this instruction, remove it.
1361 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1362 if (LocalDepEntry != LocalDeps.end()) {
1363 // Remove us from DepInst's reverse set now that the local dep info is gone.
1364 if (Instruction *Inst = LocalDepEntry->second.getInst())
1365 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1367 // Remove this local dependency info.
1368 LocalDeps.erase(LocalDepEntry);
1371 // If we have any cached pointer dependencies on this instruction, remove
1372 // them. If the instruction has non-pointer type, then it can't be a pointer
1375 // Remove it from both the load info and the store info. The instruction
1376 // can't be in either of these maps if it is non-pointer.
1377 if (RemInst->getType()->isPointerTy()) {
1378 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1379 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1382 // Loop over all of the things that depend on the instruction we're removing.
1384 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1386 // If we find RemInst as a clobber or Def in any of the maps for other values,
1387 // we need to replace its entry with a dirty version of the instruction after
1388 // it. If RemInst is a terminator, we use a null dirty value.
1390 // Using a dirty version of the instruction after RemInst saves having to scan
1391 // the entire block to get to this point.
1392 MemDepResult NewDirtyVal;
1393 if (!RemInst->isTerminator())
1394 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1396 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1397 if (ReverseDepIt != ReverseLocalDeps.end()) {
1398 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1399 // RemInst can't be the terminator if it has local stuff depending on it.
1400 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1401 "Nothing can locally depend on a terminator");
1403 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1404 E = ReverseDeps.end(); I != E; ++I) {
1405 Instruction *InstDependingOnRemInst = *I;
1406 assert(InstDependingOnRemInst != RemInst &&
1407 "Already removed our local dep info");
1409 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1411 // Make sure to remember that new things depend on NewDepInst.
1412 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1413 "a local dep on this if it is a terminator!");
1414 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1415 InstDependingOnRemInst));
1418 ReverseLocalDeps.erase(ReverseDepIt);
1420 // Add new reverse deps after scanning the set, to avoid invalidating the
1421 // 'ReverseDeps' reference.
1422 while (!ReverseDepsToAdd.empty()) {
1423 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1424 .insert(ReverseDepsToAdd.back().second);
1425 ReverseDepsToAdd.pop_back();
1429 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1430 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1431 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1432 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1434 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1436 PerInstNLInfo &INLD = NonLocalDeps[*I];
1437 // The information is now dirty!
1440 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1441 DE = INLD.first.end(); DI != DE; ++DI) {
1442 if (DI->getResult().getInst() != RemInst) continue;
1444 // Convert to a dirty entry for the subsequent instruction.
1445 DI->setResult(NewDirtyVal);
1447 if (Instruction *NextI = NewDirtyVal.getInst())
1448 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1452 ReverseNonLocalDeps.erase(ReverseDepIt);
1454 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1455 while (!ReverseDepsToAdd.empty()) {
1456 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1457 .insert(ReverseDepsToAdd.back().second);
1458 ReverseDepsToAdd.pop_back();
1462 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1463 // value in the NonLocalPointerDeps info.
1464 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1465 ReverseNonLocalPtrDeps.find(RemInst);
1466 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1467 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1468 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1470 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1471 E = Set.end(); I != E; ++I) {
1472 ValueIsLoadPair P = *I;
1473 assert(P.getPointer() != RemInst &&
1474 "Already removed NonLocalPointerDeps info for RemInst");
1476 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1478 // The cache is not valid for any specific block anymore.
1479 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1481 // Update any entries for RemInst to use the instruction after it.
1482 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1484 if (DI->getResult().getInst() != RemInst) continue;
1486 // Convert to a dirty entry for the subsequent instruction.
1487 DI->setResult(NewDirtyVal);
1489 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1490 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1493 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1494 // subsequent value may invalidate the sortedness.
1495 std::sort(NLPDI.begin(), NLPDI.end());
1498 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1500 while (!ReversePtrDepsToAdd.empty()) {
1501 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1502 .insert(ReversePtrDepsToAdd.back().second);
1503 ReversePtrDepsToAdd.pop_back();
1508 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1509 AA->deleteValue(RemInst);
1510 DEBUG(verifyRemoved(RemInst));
1512 /// verifyRemoved - Verify that the specified instruction does not occur
1513 /// in our internal data structures.
1514 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1515 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1516 E = LocalDeps.end(); I != E; ++I) {
1517 assert(I->first != D && "Inst occurs in data structures");
1518 assert(I->second.getInst() != D &&
1519 "Inst occurs in data structures");
1522 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1523 E = NonLocalPointerDeps.end(); I != E; ++I) {
1524 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1525 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1526 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1528 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1531 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1532 E = NonLocalDeps.end(); I != E; ++I) {
1533 assert(I->first != D && "Inst occurs in data structures");
1534 const PerInstNLInfo &INLD = I->second;
1535 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1536 EE = INLD.first.end(); II != EE; ++II)
1537 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1540 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1541 E = ReverseLocalDeps.end(); I != E; ++I) {
1542 assert(I->first != D && "Inst occurs in data structures");
1543 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1544 EE = I->second.end(); II != EE; ++II)
1545 assert(*II != D && "Inst occurs in data structures");
1548 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1549 E = ReverseNonLocalDeps.end();
1551 assert(I->first != D && "Inst occurs in data structures");
1552 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1553 EE = I->second.end(); II != EE; ++II)
1554 assert(*II != D && "Inst occurs in data structures");
1557 for (ReverseNonLocalPtrDepTy::const_iterator
1558 I = ReverseNonLocalPtrDeps.begin(),
1559 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1560 assert(I->first != D && "Inst occurs in rev NLPD map");
1562 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1563 E = I->second.end(); II != E; ++II)
1564 assert(*II != ValueIsLoadPair(D, false) &&
1565 *II != ValueIsLoadPair(D, true) &&
1566 "Inst occurs in ReverseNonLocalPtrDeps map");