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/InstructionSimplify.h"
23 #include "llvm/Analysis/MemoryBuiltins.h"
24 #include "llvm/Analysis/PHITransAddr.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/Dominators.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 DominatorTreeWrapperPass *DTWP =
92 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
93 DT = DTWP ? &DTWP->getDomTree() : 0;
95 PredCache.reset(new PredIteratorCache());
99 /// RemoveFromReverseMap - This is a helper function that removes Val from
100 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
101 template <typename KeyTy>
102 static void RemoveFromReverseMap(DenseMap<Instruction*,
103 SmallPtrSet<KeyTy, 4> > &ReverseMap,
104 Instruction *Inst, KeyTy Val) {
105 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
106 InstIt = ReverseMap.find(Inst);
107 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
108 bool Found = InstIt->second.erase(Val);
109 assert(Found && "Invalid reverse map!"); (void)Found;
110 if (InstIt->second.empty())
111 ReverseMap.erase(InstIt);
114 /// GetLocation - If the given instruction references a specific memory
115 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
116 /// Return a ModRefInfo value describing the general behavior of the
119 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
120 AliasAnalysis::Location &Loc,
122 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
123 if (LI->isUnordered()) {
124 Loc = AA->getLocation(LI);
125 return AliasAnalysis::Ref;
127 if (LI->getOrdering() == Monotonic) {
128 Loc = AA->getLocation(LI);
129 return AliasAnalysis::ModRef;
131 Loc = AliasAnalysis::Location();
132 return AliasAnalysis::ModRef;
135 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
136 if (SI->isUnordered()) {
137 Loc = AA->getLocation(SI);
138 return AliasAnalysis::Mod;
140 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, AA->getTargetLibraryInfo())) {
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);
238 // If we could not obtain a pointer for the instruction and the instruction
239 // touches memory then assume that this is a dependency.
240 if (MR != AliasAnalysis::NoModRef)
241 return MemDepResult::getClobber(Inst);
244 // No dependence found. If this is the entry block of the function, it is
245 // unknown, otherwise it is non-local.
246 if (BB != &BB->getParent()->getEntryBlock())
247 return MemDepResult::getNonLocal();
248 return MemDepResult::getNonFuncLocal();
251 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
252 /// would fully overlap MemLoc if done as a wider legal integer load.
254 /// MemLocBase, MemLocOffset are lazily computed here the first time the
255 /// base/offs of memloc is needed.
257 isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc,
258 const Value *&MemLocBase,
261 const DataLayout *TD) {
262 // If we have no target data, we can't do this.
263 if (TD == 0) return false;
265 // If we haven't already computed the base/offset of MemLoc, do so now.
267 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, TD);
269 unsigned Size = MemoryDependenceAnalysis::
270 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size,
275 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
276 /// looks at a memory location for a load (specified by MemLocBase, Offs,
277 /// and Size) and compares it against a load. If the specified load could
278 /// be safely widened to a larger integer load that is 1) still efficient,
279 /// 2) safe for the target, and 3) would provide the specified memory
280 /// location value, then this function returns the size in bytes of the
281 /// load width to use. If not, this returns zero.
282 unsigned MemoryDependenceAnalysis::
283 getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs,
284 unsigned MemLocSize, const LoadInst *LI,
285 const DataLayout &TD) {
286 // We can only extend simple integer loads.
287 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
289 // Load widening is hostile to ThreadSanitizer: it may cause false positives
290 // or make the reports more cryptic (access sizes are wrong).
291 if (LI->getParent()->getParent()->getAttributes().
292 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeThread))
295 // Get the base of this load.
297 const Value *LIBase =
298 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, &TD);
300 // If the two pointers are not based on the same pointer, we can't tell that
302 if (LIBase != MemLocBase) return 0;
304 // Okay, the two values are based on the same pointer, but returned as
305 // no-alias. This happens when we have things like two byte loads at "P+1"
306 // and "P+3". Check to see if increasing the size of the "LI" load up to its
307 // alignment (or the largest native integer type) will allow us to load all
308 // the bits required by MemLoc.
310 // If MemLoc is before LI, then no widening of LI will help us out.
311 if (MemLocOffs < LIOffs) return 0;
313 // Get the alignment of the load in bytes. We assume that it is safe to load
314 // any legal integer up to this size without a problem. For example, if we're
315 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
316 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
318 unsigned LoadAlign = LI->getAlignment();
320 int64_t MemLocEnd = MemLocOffs+MemLocSize;
322 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
323 if (LIOffs+LoadAlign < MemLocEnd) return 0;
325 // This is the size of the load to try. Start with the next larger power of
327 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
328 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
331 // If this load size is bigger than our known alignment or would not fit
332 // into a native integer register, then we fail.
333 if (NewLoadByteSize > LoadAlign ||
334 !TD.fitsInLegalInteger(NewLoadByteSize*8))
337 if (LIOffs+NewLoadByteSize > MemLocEnd &&
338 LI->getParent()->getParent()->getAttributes().
339 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeAddress))
340 // We will be reading past the location accessed by the original program.
341 // While this is safe in a regular build, Address Safety analysis tools
342 // may start reporting false warnings. So, don't do widening.
345 // If a load of this width would include all of MemLoc, then we succeed.
346 if (LIOffs+NewLoadByteSize >= MemLocEnd)
347 return NewLoadByteSize;
349 NewLoadByteSize <<= 1;
353 /// getPointerDependencyFrom - Return the instruction on which a memory
354 /// location depends. If isLoad is true, this routine ignores may-aliases with
355 /// read-only operations. If isLoad is false, this routine ignores may-aliases
356 /// with reads from read-only locations. If possible, pass the query
357 /// instruction as well; this function may take advantage of the metadata
358 /// annotated to the query instruction to refine the result.
359 MemDepResult MemoryDependenceAnalysis::
360 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
361 BasicBlock::iterator ScanIt, BasicBlock *BB,
362 Instruction *QueryInst) {
364 const Value *MemLocBase = 0;
365 int64_t MemLocOffset = 0;
366 unsigned Limit = BlockScanLimit;
367 bool isInvariantLoad = false;
368 if (isLoad && QueryInst) {
369 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
370 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != 0)
371 isInvariantLoad = true;
374 // Walk backwards through the basic block, looking for dependencies.
375 while (ScanIt != BB->begin()) {
376 Instruction *Inst = --ScanIt;
378 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
379 // Debug intrinsics don't (and can't) cause dependencies.
380 if (isa<DbgInfoIntrinsic>(II)) continue;
382 // Limit the amount of scanning we do so we don't end up with quadratic
383 // running time on extreme testcases.
386 return MemDepResult::getUnknown();
388 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
389 // If we reach a lifetime begin or end marker, then the query ends here
390 // because the value is undefined.
391 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
392 // FIXME: This only considers queries directly on the invariant-tagged
393 // pointer, not on query pointers that are indexed off of them. It'd
394 // be nice to handle that at some point (the right approach is to use
395 // GetPointerBaseWithConstantOffset).
396 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
398 return MemDepResult::getDef(II);
403 // Values depend on loads if the pointers are must aliased. This means that
404 // a load depends on another must aliased load from the same value.
405 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
406 // Atomic loads have complications involved.
407 // FIXME: This is overly conservative.
408 if (!LI->isUnordered())
409 return MemDepResult::getClobber(LI);
411 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
413 // If we found a pointer, check if it could be the same as our pointer.
414 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
417 if (R == AliasAnalysis::NoAlias) {
418 // If this is an over-aligned integer load (for example,
419 // "load i8* %P, align 4") see if it would obviously overlap with the
420 // queried location if widened to a larger load (e.g. if the queried
421 // location is 1 byte at P+1). If so, return it as a load/load
422 // clobber result, allowing the client to decide to widen the load if
424 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType()))
425 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() &&
426 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
427 MemLocOffset, LI, TD))
428 return MemDepResult::getClobber(Inst);
433 // Must aliased loads are defs of each other.
434 if (R == AliasAnalysis::MustAlias)
435 return MemDepResult::getDef(Inst);
437 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
438 // in terms of clobbering loads, but since it does this by looking
439 // at the clobbering load directly, it doesn't know about any
440 // phi translation that may have happened along the way.
442 // If we have a partial alias, then return this as a clobber for the
444 if (R == AliasAnalysis::PartialAlias)
445 return MemDepResult::getClobber(Inst);
448 // Random may-alias loads don't depend on each other without a
453 // Stores don't depend on other no-aliased accesses.
454 if (R == AliasAnalysis::NoAlias)
457 // Stores don't alias loads from read-only memory.
458 if (AA->pointsToConstantMemory(LoadLoc))
461 // Stores depend on may/must aliased loads.
462 return MemDepResult::getDef(Inst);
465 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
466 // Atomic stores have complications involved.
467 // FIXME: This is overly conservative.
468 if (!SI->isUnordered())
469 return MemDepResult::getClobber(SI);
471 // If alias analysis can tell that this store is guaranteed to not modify
472 // the query pointer, ignore it. Use getModRefInfo to handle cases where
473 // the query pointer points to constant memory etc.
474 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
477 // Ok, this store might clobber the query pointer. Check to see if it is
478 // a must alias: in this case, we want to return this as a def.
479 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
481 // If we found a pointer, check if it could be the same as our pointer.
482 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
484 if (R == AliasAnalysis::NoAlias)
486 if (R == AliasAnalysis::MustAlias)
487 return MemDepResult::getDef(Inst);
490 return MemDepResult::getClobber(Inst);
493 // If this is an allocation, and if we know that the accessed pointer is to
494 // the allocation, return Def. This means that there is no dependence and
495 // the access can be optimized based on that. For example, a load could
497 // Note: Only determine this to be a malloc if Inst is the malloc call, not
498 // a subsequent bitcast of the malloc call result. There can be stores to
499 // the malloced memory between the malloc call and its bitcast uses, and we
500 // need to continue scanning until the malloc call.
501 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
502 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
503 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD);
505 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
506 return MemDepResult::getDef(Inst);
507 // Be conservative if the accessed pointer may alias the allocation.
508 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
509 return MemDepResult::getClobber(Inst);
510 // If the allocation is not aliased and does not read memory (like
511 // strdup), it is safe to ignore.
512 if (isa<AllocaInst>(Inst) ||
513 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
517 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
518 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
519 // If necessary, perform additional analysis.
520 if (MR == AliasAnalysis::ModRef)
521 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
523 case AliasAnalysis::NoModRef:
524 // If the call has no effect on the queried pointer, just ignore it.
526 case AliasAnalysis::Mod:
527 return MemDepResult::getClobber(Inst);
528 case AliasAnalysis::Ref:
529 // If the call is known to never store to the pointer, and if this is a
530 // load query, we can safely ignore it (scan past it).
534 // Otherwise, there is a potential dependence. Return a clobber.
535 return MemDepResult::getClobber(Inst);
539 // No dependence found. If this is the entry block of the function, it is
540 // unknown, otherwise it is non-local.
541 if (BB != &BB->getParent()->getEntryBlock())
542 return MemDepResult::getNonLocal();
543 return MemDepResult::getNonFuncLocal();
546 /// getDependency - Return the instruction on which a memory operation
548 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
549 Instruction *ScanPos = QueryInst;
551 // Check for a cached result
552 MemDepResult &LocalCache = LocalDeps[QueryInst];
554 // If the cached entry is non-dirty, just return it. Note that this depends
555 // on MemDepResult's default constructing to 'dirty'.
556 if (!LocalCache.isDirty())
559 // Otherwise, if we have a dirty entry, we know we can start the scan at that
560 // instruction, which may save us some work.
561 if (Instruction *Inst = LocalCache.getInst()) {
564 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
567 BasicBlock *QueryParent = QueryInst->getParent();
570 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
571 // No dependence found. If this is the entry block of the function, it is
572 // unknown, otherwise it is non-local.
573 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
574 LocalCache = MemDepResult::getNonLocal();
576 LocalCache = MemDepResult::getNonFuncLocal();
578 AliasAnalysis::Location MemLoc;
579 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
581 // If we can do a pointer scan, make it happen.
582 bool isLoad = !(MR & AliasAnalysis::Mod);
583 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
584 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
586 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
587 QueryParent, QueryInst);
588 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
589 CallSite QueryCS(QueryInst);
590 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
591 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
594 // Non-memory instruction.
595 LocalCache = MemDepResult::getUnknown();
598 // Remember the result!
599 if (Instruction *I = LocalCache.getInst())
600 ReverseLocalDeps[I].insert(QueryInst);
606 /// AssertSorted - This method is used when -debug is specified to verify that
607 /// cache arrays are properly kept sorted.
608 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
610 if (Count == -1) Count = Cache.size();
611 if (Count == 0) return;
613 for (unsigned i = 1; i != unsigned(Count); ++i)
614 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
618 /// getNonLocalCallDependency - Perform a full dependency query for the
619 /// specified call, returning the set of blocks that the value is
620 /// potentially live across. The returned set of results will include a
621 /// "NonLocal" result for all blocks where the value is live across.
623 /// This method assumes the instruction returns a "NonLocal" dependency
624 /// within its own block.
626 /// This returns a reference to an internal data structure that may be
627 /// invalidated on the next non-local query or when an instruction is
628 /// removed. Clients must copy this data if they want it around longer than
630 const MemoryDependenceAnalysis::NonLocalDepInfo &
631 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
632 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
633 "getNonLocalCallDependency should only be used on calls with non-local deps!");
634 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
635 NonLocalDepInfo &Cache = CacheP.first;
637 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
638 /// the cached case, this can happen due to instructions being deleted etc. In
639 /// the uncached case, this starts out as the set of predecessors we care
641 SmallVector<BasicBlock*, 32> DirtyBlocks;
643 if (!Cache.empty()) {
644 // Okay, we have a cache entry. If we know it is not dirty, just return it
645 // with no computation.
646 if (!CacheP.second) {
651 // If we already have a partially computed set of results, scan them to
652 // determine what is dirty, seeding our initial DirtyBlocks worklist.
653 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
655 if (I->getResult().isDirty())
656 DirtyBlocks.push_back(I->getBB());
658 // Sort the cache so that we can do fast binary search lookups below.
659 std::sort(Cache.begin(), Cache.end());
661 ++NumCacheDirtyNonLocal;
662 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
663 // << Cache.size() << " cached: " << *QueryInst;
665 // Seed DirtyBlocks with each of the preds of QueryInst's block.
666 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
667 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
668 DirtyBlocks.push_back(*PI);
669 ++NumUncacheNonLocal;
672 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
673 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
675 SmallPtrSet<BasicBlock*, 64> Visited;
677 unsigned NumSortedEntries = Cache.size();
678 DEBUG(AssertSorted(Cache));
680 // Iterate while we still have blocks to update.
681 while (!DirtyBlocks.empty()) {
682 BasicBlock *DirtyBB = DirtyBlocks.back();
683 DirtyBlocks.pop_back();
685 // Already processed this block?
686 if (!Visited.insert(DirtyBB))
689 // Do a binary search to see if we already have an entry for this block in
690 // the cache set. If so, find it.
691 DEBUG(AssertSorted(Cache, NumSortedEntries));
692 NonLocalDepInfo::iterator Entry =
693 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
694 NonLocalDepEntry(DirtyBB));
695 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB)
698 NonLocalDepEntry *ExistingResult = 0;
699 if (Entry != Cache.begin()+NumSortedEntries &&
700 Entry->getBB() == DirtyBB) {
701 // If we already have an entry, and if it isn't already dirty, the block
703 if (!Entry->getResult().isDirty())
706 // Otherwise, remember this slot so we can update the value.
707 ExistingResult = &*Entry;
710 // If the dirty entry has a pointer, start scanning from it so we don't have
711 // to rescan the entire block.
712 BasicBlock::iterator ScanPos = DirtyBB->end();
713 if (ExistingResult) {
714 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
716 // We're removing QueryInst's use of Inst.
717 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
718 QueryCS.getInstruction());
722 // Find out if this block has a local dependency for QueryInst.
725 if (ScanPos != DirtyBB->begin()) {
726 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
727 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
728 // No dependence found. If this is the entry block of the function, it is
729 // a clobber, otherwise it is unknown.
730 Dep = MemDepResult::getNonLocal();
732 Dep = MemDepResult::getNonFuncLocal();
735 // If we had a dirty entry for the block, update it. Otherwise, just add
738 ExistingResult->setResult(Dep);
740 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
742 // If the block has a dependency (i.e. it isn't completely transparent to
743 // the value), remember the association!
744 if (!Dep.isNonLocal()) {
745 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
746 // update this when we remove instructions.
747 if (Instruction *Inst = Dep.getInst())
748 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
751 // If the block *is* completely transparent to the load, we need to check
752 // the predecessors of this block. Add them to our worklist.
753 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
754 DirtyBlocks.push_back(*PI);
761 /// getNonLocalPointerDependency - Perform a full dependency query for an
762 /// access to the specified (non-volatile) memory location, returning the
763 /// set of instructions that either define or clobber the value.
765 /// This method assumes the pointer has a "NonLocal" dependency within its
768 void MemoryDependenceAnalysis::
769 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
771 SmallVectorImpl<NonLocalDepResult> &Result) {
772 assert(Loc.Ptr->getType()->isPointerTy() &&
773 "Can't get pointer deps of a non-pointer!");
776 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD);
778 // This is the set of blocks we've inspected, and the pointer we consider in
779 // each block. Because of critical edges, we currently bail out if querying
780 // a block with multiple different pointers. This can happen during PHI
782 DenseMap<BasicBlock*, Value*> Visited;
783 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
784 Result, Visited, true))
787 Result.push_back(NonLocalDepResult(FromBB,
788 MemDepResult::getUnknown(),
789 const_cast<Value *>(Loc.Ptr)));
792 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
793 /// Pointer/PointeeSize using either cached information in Cache or by doing a
794 /// lookup (which may use dirty cache info if available). If we do a lookup,
795 /// add the result to the cache.
796 MemDepResult MemoryDependenceAnalysis::
797 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
798 bool isLoad, BasicBlock *BB,
799 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
801 // Do a binary search to see if we already have an entry for this block in
802 // the cache set. If so, find it.
803 NonLocalDepInfo::iterator Entry =
804 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
805 NonLocalDepEntry(BB));
806 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
809 NonLocalDepEntry *ExistingResult = 0;
810 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
811 ExistingResult = &*Entry;
813 // If we have a cached entry, and it is non-dirty, use it as the value for
815 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
816 ++NumCacheNonLocalPtr;
817 return ExistingResult->getResult();
820 // Otherwise, we have to scan for the value. If we have a dirty cache
821 // entry, start scanning from its position, otherwise we scan from the end
823 BasicBlock::iterator ScanPos = BB->end();
824 if (ExistingResult && ExistingResult->getResult().getInst()) {
825 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
826 "Instruction invalidated?");
827 ++NumCacheDirtyNonLocalPtr;
828 ScanPos = ExistingResult->getResult().getInst();
830 // Eliminating the dirty entry from 'Cache', so update the reverse info.
831 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
832 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
834 ++NumUncacheNonLocalPtr;
837 // Scan the block for the dependency.
838 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
840 // If we had a dirty entry for the block, update it. Otherwise, just add
843 ExistingResult->setResult(Dep);
845 Cache->push_back(NonLocalDepEntry(BB, Dep));
847 // If the block has a dependency (i.e. it isn't completely transparent to
848 // the value), remember the reverse association because we just added it
850 if (!Dep.isDef() && !Dep.isClobber())
853 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
854 // update MemDep when we remove instructions.
855 Instruction *Inst = Dep.getInst();
856 assert(Inst && "Didn't depend on anything?");
857 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
858 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
862 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
863 /// number of elements in the array that are already properly ordered. This is
864 /// optimized for the case when only a few entries are added.
866 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
867 unsigned NumSortedEntries) {
868 switch (Cache.size() - NumSortedEntries) {
870 // done, no new entries.
873 // Two new entries, insert the last one into place.
874 NonLocalDepEntry Val = Cache.back();
876 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
877 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
878 Cache.insert(Entry, Val);
882 // One new entry, Just insert the new value at the appropriate position.
883 if (Cache.size() != 1) {
884 NonLocalDepEntry Val = Cache.back();
886 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
887 std::upper_bound(Cache.begin(), Cache.end(), Val);
888 Cache.insert(Entry, Val);
892 // Added many values, do a full scale sort.
893 std::sort(Cache.begin(), Cache.end());
898 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
899 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
900 /// results to the results vector and keep track of which blocks are visited in
903 /// This has special behavior for the first block queries (when SkipFirstBlock
904 /// is true). In this special case, it ignores the contents of the specified
905 /// block and starts returning dependence info for its predecessors.
907 /// This function returns false on success, or true to indicate that it could
908 /// not compute dependence information for some reason. This should be treated
909 /// as a clobber dependence on the first instruction in the predecessor block.
910 bool MemoryDependenceAnalysis::
911 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
912 const AliasAnalysis::Location &Loc,
913 bool isLoad, BasicBlock *StartBB,
914 SmallVectorImpl<NonLocalDepResult> &Result,
915 DenseMap<BasicBlock*, Value*> &Visited,
916 bool SkipFirstBlock) {
917 // Look up the cached info for Pointer.
918 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
920 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
921 // CacheKey, this value will be inserted as the associated value. Otherwise,
922 // it'll be ignored, and we'll have to check to see if the cached size and
923 // tbaa tag are consistent with the current query.
924 NonLocalPointerInfo InitialNLPI;
925 InitialNLPI.Size = Loc.Size;
926 InitialNLPI.TBAATag = Loc.TBAATag;
928 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
930 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
931 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
932 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
934 // If we already have a cache entry for this CacheKey, we may need to do some
935 // work to reconcile the cache entry and the current query.
937 if (CacheInfo->Size < Loc.Size) {
938 // The query's Size is greater than the cached one. Throw out the
939 // cached data and proceed with the query at the greater size.
940 CacheInfo->Pair = BBSkipFirstBlockPair();
941 CacheInfo->Size = Loc.Size;
942 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
943 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
944 if (Instruction *Inst = DI->getResult().getInst())
945 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
946 CacheInfo->NonLocalDeps.clear();
947 } else if (CacheInfo->Size > Loc.Size) {
948 // This query's Size is less than the cached one. Conservatively restart
949 // the query using the greater size.
950 return getNonLocalPointerDepFromBB(Pointer,
951 Loc.getWithNewSize(CacheInfo->Size),
952 isLoad, StartBB, Result, Visited,
956 // If the query's TBAATag is inconsistent with the cached one,
957 // conservatively throw out the cached data and restart the query with
959 if (CacheInfo->TBAATag != Loc.TBAATag) {
960 if (CacheInfo->TBAATag) {
961 CacheInfo->Pair = BBSkipFirstBlockPair();
962 CacheInfo->TBAATag = 0;
963 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
964 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
965 if (Instruction *Inst = DI->getResult().getInst())
966 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
967 CacheInfo->NonLocalDeps.clear();
970 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(),
971 isLoad, StartBB, Result, Visited,
976 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
978 // If we have valid cached information for exactly the block we are
979 // investigating, just return it with no recomputation.
980 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
981 // We have a fully cached result for this query then we can just return the
982 // cached results and populate the visited set. However, we have to verify
983 // that we don't already have conflicting results for these blocks. Check
984 // to ensure that if a block in the results set is in the visited set that
985 // it was for the same pointer query.
986 if (!Visited.empty()) {
987 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
989 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
990 if (VI == Visited.end() || VI->second == Pointer.getAddr())
993 // We have a pointer mismatch in a block. Just return clobber, saying
994 // that something was clobbered in this result. We could also do a
995 // non-fully cached query, but there is little point in doing this.
1000 Value *Addr = Pointer.getAddr();
1001 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1003 Visited.insert(std::make_pair(I->getBB(), Addr));
1004 if (I->getResult().isNonLocal()) {
1009 Result.push_back(NonLocalDepResult(I->getBB(),
1010 MemDepResult::getUnknown(),
1012 } else if (DT->isReachableFromEntry(I->getBB())) {
1013 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1016 ++NumCacheCompleteNonLocalPtr;
1020 // Otherwise, either this is a new block, a block with an invalid cache
1021 // pointer or one that we're about to invalidate by putting more info into it
1022 // than its valid cache info. If empty, the result will be valid cache info,
1023 // otherwise it isn't.
1025 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1027 CacheInfo->Pair = BBSkipFirstBlockPair();
1029 SmallVector<BasicBlock*, 32> Worklist;
1030 Worklist.push_back(StartBB);
1032 // PredList used inside loop.
1033 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1035 // Keep track of the entries that we know are sorted. Previously cached
1036 // entries will all be sorted. The entries we add we only sort on demand (we
1037 // don't insert every element into its sorted position). We know that we
1038 // won't get any reuse from currently inserted values, because we don't
1039 // revisit blocks after we insert info for them.
1040 unsigned NumSortedEntries = Cache->size();
1041 DEBUG(AssertSorted(*Cache));
1043 while (!Worklist.empty()) {
1044 BasicBlock *BB = Worklist.pop_back_val();
1046 // Skip the first block if we have it.
1047 if (!SkipFirstBlock) {
1048 // Analyze the dependency of *Pointer in FromBB. See if we already have
1050 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1052 // Get the dependency info for Pointer in BB. If we have cached
1053 // information, we will use it, otherwise we compute it.
1054 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1055 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
1058 // If we got a Def or Clobber, add this to the list of results.
1059 if (!Dep.isNonLocal()) {
1061 Result.push_back(NonLocalDepResult(BB,
1062 MemDepResult::getUnknown(),
1063 Pointer.getAddr()));
1065 } else if (DT->isReachableFromEntry(BB)) {
1066 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1072 // If 'Pointer' is an instruction defined in this block, then we need to do
1073 // phi translation to change it into a value live in the predecessor block.
1074 // If not, we just add the predecessors to the worklist and scan them with
1075 // the same Pointer.
1076 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1077 SkipFirstBlock = false;
1078 SmallVector<BasicBlock*, 16> NewBlocks;
1079 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1080 // Verify that we haven't looked at this block yet.
1081 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1082 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1083 if (InsertRes.second) {
1084 // First time we've looked at *PI.
1085 NewBlocks.push_back(*PI);
1089 // If we have seen this block before, but it was with a different
1090 // pointer then we have a phi translation failure and we have to treat
1091 // this as a clobber.
1092 if (InsertRes.first->second != Pointer.getAddr()) {
1093 // Make sure to clean up the Visited map before continuing on to
1094 // PredTranslationFailure.
1095 for (unsigned i = 0; i < NewBlocks.size(); i++)
1096 Visited.erase(NewBlocks[i]);
1097 goto PredTranslationFailure;
1100 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1104 // We do need to do phi translation, if we know ahead of time we can't phi
1105 // translate this value, don't even try.
1106 if (!Pointer.IsPotentiallyPHITranslatable())
1107 goto PredTranslationFailure;
1109 // We may have added values to the cache list before this PHI translation.
1110 // If so, we haven't done anything to ensure that the cache remains sorted.
1111 // Sort it now (if needed) so that recursive invocations of
1112 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1113 // value will only see properly sorted cache arrays.
1114 if (Cache && NumSortedEntries != Cache->size()) {
1115 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1116 NumSortedEntries = Cache->size();
1121 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1122 BasicBlock *Pred = *PI;
1123 PredList.push_back(std::make_pair(Pred, Pointer));
1125 // Get the PHI translated pointer in this predecessor. This can fail if
1126 // not translatable, in which case the getAddr() returns null.
1127 PHITransAddr &PredPointer = PredList.back().second;
1128 PredPointer.PHITranslateValue(BB, Pred, 0);
1130 Value *PredPtrVal = PredPointer.getAddr();
1132 // Check to see if we have already visited this pred block with another
1133 // pointer. If so, we can't do this lookup. This failure can occur
1134 // with PHI translation when a critical edge exists and the PHI node in
1135 // the successor translates to a pointer value different than the
1136 // pointer the block was first analyzed with.
1137 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1138 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1140 if (!InsertRes.second) {
1141 // We found the pred; take it off the list of preds to visit.
1142 PredList.pop_back();
1144 // If the predecessor was visited with PredPtr, then we already did
1145 // the analysis and can ignore it.
1146 if (InsertRes.first->second == PredPtrVal)
1149 // Otherwise, the block was previously analyzed with a different
1150 // pointer. We can't represent the result of this case, so we just
1151 // treat this as a phi translation failure.
1153 // Make sure to clean up the Visited map before continuing on to
1154 // PredTranslationFailure.
1155 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1156 Visited.erase(PredList[i].first);
1158 goto PredTranslationFailure;
1162 // Actually process results here; this need to be a separate loop to avoid
1163 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1164 // any results for. (getNonLocalPointerDepFromBB will modify our
1165 // datastructures in ways the code after the PredTranslationFailure label
1167 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1168 BasicBlock *Pred = PredList[i].first;
1169 PHITransAddr &PredPointer = PredList[i].second;
1170 Value *PredPtrVal = PredPointer.getAddr();
1172 bool CanTranslate = true;
1173 // If PHI translation was unable to find an available pointer in this
1174 // predecessor, then we have to assume that the pointer is clobbered in
1175 // that predecessor. We can still do PRE of the load, which would insert
1176 // a computation of the pointer in this predecessor.
1177 if (PredPtrVal == 0)
1178 CanTranslate = false;
1180 // FIXME: it is entirely possible that PHI translating will end up with
1181 // the same value. Consider PHI translating something like:
1182 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1183 // to recurse here, pedantically speaking.
1185 // If getNonLocalPointerDepFromBB fails here, that means the cached
1186 // result conflicted with the Visited list; we have to conservatively
1187 // assume it is unknown, but this also does not block PRE of the load.
1188 if (!CanTranslate ||
1189 getNonLocalPointerDepFromBB(PredPointer,
1190 Loc.getWithNewPtr(PredPtrVal),
1193 // Add the entry to the Result list.
1194 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1195 Result.push_back(Entry);
1197 // Since we had a phi translation failure, the cache for CacheKey won't
1198 // include all of the entries that we need to immediately satisfy future
1199 // queries. Mark this in NonLocalPointerDeps by setting the
1200 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1201 // cached value to do more work but not miss the phi trans failure.
1202 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1203 NLPI.Pair = BBSkipFirstBlockPair();
1208 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1209 CacheInfo = &NonLocalPointerDeps[CacheKey];
1210 Cache = &CacheInfo->NonLocalDeps;
1211 NumSortedEntries = Cache->size();
1213 // Since we did phi translation, the "Cache" set won't contain all of the
1214 // results for the query. This is ok (we can still use it to accelerate
1215 // specific block queries) but we can't do the fastpath "return all
1216 // results from the set" Clear out the indicator for this.
1217 CacheInfo->Pair = BBSkipFirstBlockPair();
1218 SkipFirstBlock = false;
1221 PredTranslationFailure:
1222 // The following code is "failure"; we can't produce a sane translation
1223 // for the given block. It assumes that we haven't modified any of
1224 // our datastructures while processing the current block.
1227 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1228 CacheInfo = &NonLocalPointerDeps[CacheKey];
1229 Cache = &CacheInfo->NonLocalDeps;
1230 NumSortedEntries = Cache->size();
1233 // Since we failed phi translation, the "Cache" set won't contain all of the
1234 // results for the query. This is ok (we can still use it to accelerate
1235 // specific block queries) but we can't do the fastpath "return all
1236 // results from the set". Clear out the indicator for this.
1237 CacheInfo->Pair = BBSkipFirstBlockPair();
1239 // If *nothing* works, mark the pointer as unknown.
1241 // If this is the magic first block, return this as a clobber of the whole
1242 // incoming value. Since we can't phi translate to one of the predecessors,
1243 // we have to bail out.
1247 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1248 assert(I != Cache->rend() && "Didn't find current block??");
1249 if (I->getBB() != BB)
1252 assert(I->getResult().isNonLocal() &&
1253 "Should only be here with transparent block");
1254 I->setResult(MemDepResult::getUnknown());
1255 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1256 Pointer.getAddr()));
1261 // Okay, we're done now. If we added new values to the cache, re-sort it.
1262 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1263 DEBUG(AssertSorted(*Cache));
1267 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1268 /// CachedNonLocalPointerInfo, remove it.
1269 void MemoryDependenceAnalysis::
1270 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1271 CachedNonLocalPointerInfo::iterator It =
1272 NonLocalPointerDeps.find(P);
1273 if (It == NonLocalPointerDeps.end()) return;
1275 // Remove all of the entries in the BB->val map. This involves removing
1276 // instructions from the reverse map.
1277 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1279 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1280 Instruction *Target = PInfo[i].getResult().getInst();
1281 if (Target == 0) continue; // Ignore non-local dep results.
1282 assert(Target->getParent() == PInfo[i].getBB());
1284 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1285 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1288 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1289 NonLocalPointerDeps.erase(It);
1293 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1294 /// information about the specified pointer, because it may be too
1295 /// conservative in memdep. This is an optional call that can be used when
1296 /// the client detects an equivalence between the pointer and some other
1297 /// value and replaces the other value with ptr. This can make Ptr available
1298 /// in more places that cached info does not necessarily keep.
1299 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1300 // If Ptr isn't really a pointer, just ignore it.
1301 if (!Ptr->getType()->isPointerTy()) return;
1302 // Flush store info for the pointer.
1303 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1304 // Flush load info for the pointer.
1305 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1308 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1309 /// This needs to be done when the CFG changes, e.g., due to splitting
1311 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1315 /// removeInstruction - Remove an instruction from the dependence analysis,
1316 /// updating the dependence of instructions that previously depended on it.
1317 /// This method attempts to keep the cache coherent using the reverse map.
1318 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1319 // Walk through the Non-local dependencies, removing this one as the value
1320 // for any cached queries.
1321 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1322 if (NLDI != NonLocalDeps.end()) {
1323 NonLocalDepInfo &BlockMap = NLDI->second.first;
1324 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1326 if (Instruction *Inst = DI->getResult().getInst())
1327 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1328 NonLocalDeps.erase(NLDI);
1331 // If we have a cached local dependence query for this instruction, remove it.
1333 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1334 if (LocalDepEntry != LocalDeps.end()) {
1335 // Remove us from DepInst's reverse set now that the local dep info is gone.
1336 if (Instruction *Inst = LocalDepEntry->second.getInst())
1337 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1339 // Remove this local dependency info.
1340 LocalDeps.erase(LocalDepEntry);
1343 // If we have any cached pointer dependencies on this instruction, remove
1344 // them. If the instruction has non-pointer type, then it can't be a pointer
1347 // Remove it from both the load info and the store info. The instruction
1348 // can't be in either of these maps if it is non-pointer.
1349 if (RemInst->getType()->isPointerTy()) {
1350 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1351 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1354 // Loop over all of the things that depend on the instruction we're removing.
1356 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1358 // If we find RemInst as a clobber or Def in any of the maps for other values,
1359 // we need to replace its entry with a dirty version of the instruction after
1360 // it. If RemInst is a terminator, we use a null dirty value.
1362 // Using a dirty version of the instruction after RemInst saves having to scan
1363 // the entire block to get to this point.
1364 MemDepResult NewDirtyVal;
1365 if (!RemInst->isTerminator())
1366 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1368 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1369 if (ReverseDepIt != ReverseLocalDeps.end()) {
1370 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1371 // RemInst can't be the terminator if it has local stuff depending on it.
1372 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1373 "Nothing can locally depend on a terminator");
1375 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1376 E = ReverseDeps.end(); I != E; ++I) {
1377 Instruction *InstDependingOnRemInst = *I;
1378 assert(InstDependingOnRemInst != RemInst &&
1379 "Already removed our local dep info");
1381 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1383 // Make sure to remember that new things depend on NewDepInst.
1384 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1385 "a local dep on this if it is a terminator!");
1386 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1387 InstDependingOnRemInst));
1390 ReverseLocalDeps.erase(ReverseDepIt);
1392 // Add new reverse deps after scanning the set, to avoid invalidating the
1393 // 'ReverseDeps' reference.
1394 while (!ReverseDepsToAdd.empty()) {
1395 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1396 .insert(ReverseDepsToAdd.back().second);
1397 ReverseDepsToAdd.pop_back();
1401 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1402 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1403 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1404 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1406 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1408 PerInstNLInfo &INLD = NonLocalDeps[*I];
1409 // The information is now dirty!
1412 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1413 DE = INLD.first.end(); DI != DE; ++DI) {
1414 if (DI->getResult().getInst() != RemInst) continue;
1416 // Convert to a dirty entry for the subsequent instruction.
1417 DI->setResult(NewDirtyVal);
1419 if (Instruction *NextI = NewDirtyVal.getInst())
1420 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1424 ReverseNonLocalDeps.erase(ReverseDepIt);
1426 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1427 while (!ReverseDepsToAdd.empty()) {
1428 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1429 .insert(ReverseDepsToAdd.back().second);
1430 ReverseDepsToAdd.pop_back();
1434 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1435 // value in the NonLocalPointerDeps info.
1436 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1437 ReverseNonLocalPtrDeps.find(RemInst);
1438 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1439 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1440 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1442 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1443 E = Set.end(); I != E; ++I) {
1444 ValueIsLoadPair P = *I;
1445 assert(P.getPointer() != RemInst &&
1446 "Already removed NonLocalPointerDeps info for RemInst");
1448 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1450 // The cache is not valid for any specific block anymore.
1451 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1453 // Update any entries for RemInst to use the instruction after it.
1454 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1456 if (DI->getResult().getInst() != RemInst) continue;
1458 // Convert to a dirty entry for the subsequent instruction.
1459 DI->setResult(NewDirtyVal);
1461 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1462 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1465 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1466 // subsequent value may invalidate the sortedness.
1467 std::sort(NLPDI.begin(), NLPDI.end());
1470 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1472 while (!ReversePtrDepsToAdd.empty()) {
1473 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1474 .insert(ReversePtrDepsToAdd.back().second);
1475 ReversePtrDepsToAdd.pop_back();
1480 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1481 AA->deleteValue(RemInst);
1482 DEBUG(verifyRemoved(RemInst));
1484 /// verifyRemoved - Verify that the specified instruction does not occur
1485 /// in our internal data structures.
1486 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1487 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1488 E = LocalDeps.end(); I != E; ++I) {
1489 assert(I->first != D && "Inst occurs in data structures");
1490 assert(I->second.getInst() != D &&
1491 "Inst occurs in data structures");
1494 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1495 E = NonLocalPointerDeps.end(); I != E; ++I) {
1496 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1497 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1498 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1500 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1503 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1504 E = NonLocalDeps.end(); I != E; ++I) {
1505 assert(I->first != D && "Inst occurs in data structures");
1506 const PerInstNLInfo &INLD = I->second;
1507 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1508 EE = INLD.first.end(); II != EE; ++II)
1509 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1512 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1513 E = ReverseLocalDeps.end(); I != E; ++I) {
1514 assert(I->first != D && "Inst occurs in data structures");
1515 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1516 EE = I->second.end(); II != EE; ++II)
1517 assert(*II != D && "Inst occurs in data structures");
1520 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1521 E = ReverseNonLocalDeps.end();
1523 assert(I->first != D && "Inst occurs in data structures");
1524 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1525 EE = I->second.end(); II != EE; ++II)
1526 assert(*II != D && "Inst occurs in data structures");
1529 for (ReverseNonLocalPtrDepTy::const_iterator
1530 I = ReverseNonLocalPtrDeps.begin(),
1531 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1532 assert(I->first != D && "Inst occurs in rev NLPD map");
1534 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1535 E = I->second.end(); II != E; ++II)
1536 assert(*II != ValueIsLoadPair(D, false) &&
1537 *II != ValueIsLoadPair(D, true) &&
1538 "Inst occurs in ReverseNonLocalPtrDeps map");