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 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/AssumptionCache.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/IR/PredIteratorCache.h"
33 #include "llvm/Support/Debug.h"
36 #define DEBUG_TYPE "memdep"
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 static const unsigned int BlockScanLimit = 100;
54 // Limit on the number of memdep results to process.
55 static const unsigned int NumResultsLimit = 100;
57 char MemoryDependenceAnalysis::ID = 0;
59 // Register this pass...
60 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
61 "Memory Dependence Analysis", false, true)
62 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
63 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
64 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
65 "Memory Dependence Analysis", false, true)
67 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
68 : FunctionPass(ID), PredCache() {
69 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
71 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
74 /// Clean up memory in between runs
75 void MemoryDependenceAnalysis::releaseMemory() {
78 NonLocalPointerDeps.clear();
79 ReverseLocalDeps.clear();
80 ReverseNonLocalDeps.clear();
81 ReverseNonLocalPtrDeps.clear();
85 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
87 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
89 AU.addRequired<AssumptionCacheTracker>();
90 AU.addRequiredTransitive<AliasAnalysis>();
93 bool MemoryDependenceAnalysis::runOnFunction(Function &F) {
94 AA = &getAnalysis<AliasAnalysis>();
95 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
96 DL = &F.getParent()->getDataLayout();
97 DominatorTreeWrapperPass *DTWP =
98 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
99 DT = DTWP ? &DTWP->getDomTree() : nullptr;
101 PredCache.reset(new PredIteratorCache());
105 /// RemoveFromReverseMap - This is a helper function that removes Val from
106 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
107 template <typename KeyTy>
108 static void RemoveFromReverseMap(DenseMap<Instruction*,
109 SmallPtrSet<KeyTy, 4> > &ReverseMap,
110 Instruction *Inst, KeyTy Val) {
111 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
112 InstIt = ReverseMap.find(Inst);
113 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
114 bool Found = InstIt->second.erase(Val);
115 assert(Found && "Invalid reverse map!"); (void)Found;
116 if (InstIt->second.empty())
117 ReverseMap.erase(InstIt);
120 /// GetLocation - If the given instruction references a specific memory
121 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
122 /// Return a ModRefInfo value describing the general behavior of the
125 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
126 AliasAnalysis::Location &Loc,
128 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
129 if (LI->isUnordered()) {
130 Loc = AA->getLocation(LI);
131 return AliasAnalysis::Ref;
133 if (LI->getOrdering() == Monotonic) {
134 Loc = AA->getLocation(LI);
135 return AliasAnalysis::ModRef;
137 Loc = AliasAnalysis::Location();
138 return AliasAnalysis::ModRef;
141 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
142 if (SI->isUnordered()) {
143 Loc = AA->getLocation(SI);
144 return AliasAnalysis::Mod;
146 if (SI->getOrdering() == Monotonic) {
147 Loc = AA->getLocation(SI);
148 return AliasAnalysis::ModRef;
150 Loc = AliasAnalysis::Location();
151 return AliasAnalysis::ModRef;
154 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
155 Loc = AA->getLocation(V);
156 return AliasAnalysis::ModRef;
159 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) {
160 // calls to free() deallocate the entire structure
161 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
162 return AliasAnalysis::Mod;
165 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
168 switch (II->getIntrinsicID()) {
169 case Intrinsic::lifetime_start:
170 case Intrinsic::lifetime_end:
171 case Intrinsic::invariant_start:
172 II->getAAMetadata(AAInfo);
173 Loc = AliasAnalysis::Location(II->getArgOperand(1),
174 cast<ConstantInt>(II->getArgOperand(0))
175 ->getZExtValue(), AAInfo);
176 // These intrinsics don't really modify the memory, but returning Mod
177 // will allow them to be handled conservatively.
178 return AliasAnalysis::Mod;
179 case Intrinsic::invariant_end:
180 II->getAAMetadata(AAInfo);
181 Loc = AliasAnalysis::Location(II->getArgOperand(2),
182 cast<ConstantInt>(II->getArgOperand(1))
183 ->getZExtValue(), AAInfo);
184 // These intrinsics don't really modify the memory, but returning Mod
185 // will allow them to be handled conservatively.
186 return AliasAnalysis::Mod;
192 // Otherwise, just do the coarse-grained thing that always works.
193 if (Inst->mayWriteToMemory())
194 return AliasAnalysis::ModRef;
195 if (Inst->mayReadFromMemory())
196 return AliasAnalysis::Ref;
197 return AliasAnalysis::NoModRef;
200 /// getCallSiteDependencyFrom - Private helper for finding the local
201 /// dependencies of a call site.
202 MemDepResult MemoryDependenceAnalysis::
203 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
204 BasicBlock::iterator ScanIt, BasicBlock *BB) {
205 unsigned Limit = BlockScanLimit;
207 // Walk backwards through the block, looking for dependencies
208 while (ScanIt != BB->begin()) {
209 // Limit the amount of scanning we do so we don't end up with quadratic
210 // running time on extreme testcases.
213 return MemDepResult::getUnknown();
215 Instruction *Inst = --ScanIt;
217 // If this inst is a memory op, get the pointer it accessed
218 AliasAnalysis::Location Loc;
219 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
221 // A simple instruction.
222 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
223 return MemDepResult::getClobber(Inst);
227 if (CallSite InstCS = cast<Value>(Inst)) {
228 // Debug intrinsics don't cause dependences.
229 if (isa<DbgInfoIntrinsic>(Inst)) continue;
230 // If these two calls do not interfere, look past it.
231 switch (AA->getModRefInfo(CS, InstCS)) {
232 case AliasAnalysis::NoModRef:
233 // If the two calls are the same, return InstCS as a Def, so that
234 // CS can be found redundant and eliminated.
235 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
236 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
237 return MemDepResult::getDef(Inst);
239 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
243 return MemDepResult::getClobber(Inst);
247 // If we could not obtain a pointer for the instruction and the instruction
248 // touches memory then assume that this is a dependency.
249 if (MR != AliasAnalysis::NoModRef)
250 return MemDepResult::getClobber(Inst);
253 // No dependence found. If this is the entry block of the function, it is
254 // unknown, otherwise it is non-local.
255 if (BB != &BB->getParent()->getEntryBlock())
256 return MemDepResult::getNonLocal();
257 return MemDepResult::getNonFuncLocal();
260 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
261 /// would fully overlap MemLoc if done as a wider legal integer load.
263 /// MemLocBase, MemLocOffset are lazily computed here the first time the
264 /// base/offs of memloc is needed.
266 isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc,
267 const Value *&MemLocBase,
270 const DataLayout *DL) {
271 // If we have no target data, we can't do this.
272 if (!DL) return false;
274 // If we haven't already computed the base/offset of MemLoc, do so now.
276 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
278 unsigned Size = MemoryDependenceAnalysis::
279 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size,
284 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
285 /// looks at a memory location for a load (specified by MemLocBase, Offs,
286 /// and Size) and compares it against a load. If the specified load could
287 /// be safely widened to a larger integer load that is 1) still efficient,
288 /// 2) safe for the target, and 3) would provide the specified memory
289 /// location value, then this function returns the size in bytes of the
290 /// load width to use. If not, this returns zero.
291 unsigned MemoryDependenceAnalysis::
292 getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs,
293 unsigned MemLocSize, const LoadInst *LI,
294 const DataLayout &DL) {
295 // We can only extend simple integer loads.
296 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
298 // Load widening is hostile to ThreadSanitizer: it may cause false positives
299 // or make the reports more cryptic (access sizes are wrong).
300 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
303 // Get the base of this load.
305 const Value *LIBase =
306 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, &DL);
308 // If the two pointers are not based on the same pointer, we can't tell that
310 if (LIBase != MemLocBase) return 0;
312 // Okay, the two values are based on the same pointer, but returned as
313 // no-alias. This happens when we have things like two byte loads at "P+1"
314 // and "P+3". Check to see if increasing the size of the "LI" load up to its
315 // alignment (or the largest native integer type) will allow us to load all
316 // the bits required by MemLoc.
318 // If MemLoc is before LI, then no widening of LI will help us out.
319 if (MemLocOffs < LIOffs) return 0;
321 // Get the alignment of the load in bytes. We assume that it is safe to load
322 // any legal integer up to this size without a problem. For example, if we're
323 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
324 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
326 unsigned LoadAlign = LI->getAlignment();
328 int64_t MemLocEnd = MemLocOffs+MemLocSize;
330 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
331 if (LIOffs+LoadAlign < MemLocEnd) return 0;
333 // This is the size of the load to try. Start with the next larger power of
335 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
336 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
339 // If this load size is bigger than our known alignment or would not fit
340 // into a native integer register, then we fail.
341 if (NewLoadByteSize > LoadAlign ||
342 !DL.fitsInLegalInteger(NewLoadByteSize*8))
345 if (LIOffs + NewLoadByteSize > MemLocEnd &&
346 LI->getParent()->getParent()->hasFnAttribute(
347 Attribute::SanitizeAddress))
348 // We will be reading past the location accessed by the original program.
349 // While this is safe in a regular build, Address Safety analysis tools
350 // may start reporting false warnings. So, don't do widening.
353 // If a load of this width would include all of MemLoc, then we succeed.
354 if (LIOffs+NewLoadByteSize >= MemLocEnd)
355 return NewLoadByteSize;
357 NewLoadByteSize <<= 1;
361 static bool isVolatile(Instruction *Inst) {
362 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
363 return LI->isVolatile();
364 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
365 return SI->isVolatile();
366 else if (AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Inst))
367 return AI->isVolatile();
372 /// getPointerDependencyFrom - Return the instruction on which a memory
373 /// location depends. If isLoad is true, this routine ignores may-aliases with
374 /// read-only operations. If isLoad is false, this routine ignores may-aliases
375 /// with reads from read-only locations. If possible, pass the query
376 /// instruction as well; this function may take advantage of the metadata
377 /// annotated to the query instruction to refine the result.
378 MemDepResult MemoryDependenceAnalysis::
379 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
380 BasicBlock::iterator ScanIt, BasicBlock *BB,
381 Instruction *QueryInst) {
383 const Value *MemLocBase = nullptr;
384 int64_t MemLocOffset = 0;
385 unsigned Limit = BlockScanLimit;
386 bool isInvariantLoad = false;
388 // We must be careful with atomic accesses, as they may allow another thread
389 // to touch this location, cloberring it. We are conservative: if the
390 // QueryInst is not a simple (non-atomic) memory access, we automatically
391 // return getClobber.
392 // If it is simple, we know based on the results of
393 // "Compiler testing via a theory of sound optimisations in the C11/C++11
394 // memory model" in PLDI 2013, that a non-atomic location can only be
395 // clobbered between a pair of a release and an acquire action, with no
396 // access to the location in between.
397 // Here is an example for giving the general intuition behind this rule.
398 // In the following code:
400 // release action; [1]
401 // acquire action; [4]
403 // It is unsafe to replace %val by 0 because another thread may be running:
404 // acquire action; [2]
406 // release action; [3]
407 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
408 // being 42. A key property of this program however is that if either
409 // 1 or 4 were missing, there would be a race between the store of 42
410 // either the store of 0 or the load (making the whole progam racy).
411 // The paper mentionned above shows that the same property is respected
412 // by every program that can detect any optimisation of that kind: either
413 // it is racy (undefined) or there is a release followed by an acquire
414 // between the pair of accesses under consideration.
415 bool HasSeenAcquire = false;
417 if (isLoad && QueryInst) {
418 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
419 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
420 isInvariantLoad = true;
423 // Walk backwards through the basic block, looking for dependencies.
424 while (ScanIt != BB->begin()) {
425 Instruction *Inst = --ScanIt;
427 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
428 // Debug intrinsics don't (and can't) cause dependencies.
429 if (isa<DbgInfoIntrinsic>(II)) continue;
431 // Limit the amount of scanning we do so we don't end up with quadratic
432 // running time on extreme testcases.
435 return MemDepResult::getUnknown();
437 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
438 // If we reach a lifetime begin or end marker, then the query ends here
439 // because the value is undefined.
440 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
441 // FIXME: This only considers queries directly on the invariant-tagged
442 // pointer, not on query pointers that are indexed off of them. It'd
443 // be nice to handle that at some point (the right approach is to use
444 // GetPointerBaseWithConstantOffset).
445 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
447 return MemDepResult::getDef(II);
452 // Values depend on loads if the pointers are must aliased. This means that
453 // a load depends on another must aliased load from the same value.
454 // One exception is atomic loads: a value can depend on an atomic load that it
455 // does not alias with when this atomic load indicates that another thread may
456 // be accessing the location.
457 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
459 // While volatile access cannot be eliminated, they do not have to clobber
460 // non-aliasing locations, as normal accesses, for example, can be safely
461 // reordered with volatile accesses.
462 if (LI->isVolatile()) {
464 // Original QueryInst *may* be volatile
465 return MemDepResult::getClobber(LI);
466 if (isVolatile(QueryInst))
467 // Ordering required if QueryInst is itself volatile
468 return MemDepResult::getClobber(LI);
469 // Otherwise, volatile doesn't imply any special ordering
472 // Atomic loads have complications involved.
473 // A Monotonic (or higher) load is OK if the query inst is itself not atomic.
474 // An Acquire (or higher) load sets the HasSeenAcquire flag, so that any
475 // release store will know to return getClobber.
476 // FIXME: This is overly conservative.
477 if (LI->isAtomic() && LI->getOrdering() > Unordered) {
479 return MemDepResult::getClobber(LI);
480 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
481 if (!QueryLI->isSimple())
482 return MemDepResult::getClobber(LI);
483 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
484 if (!QuerySI->isSimple())
485 return MemDepResult::getClobber(LI);
486 } else if (QueryInst->mayReadOrWriteMemory()) {
487 return MemDepResult::getClobber(LI);
490 if (isAtLeastAcquire(LI->getOrdering()))
491 HasSeenAcquire = true;
494 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
496 // If we found a pointer, check if it could be the same as our pointer.
497 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
500 if (R == AliasAnalysis::NoAlias) {
501 // If this is an over-aligned integer load (for example,
502 // "load i8* %P, align 4") see if it would obviously overlap with the
503 // queried location if widened to a larger load (e.g. if the queried
504 // location is 1 byte at P+1). If so, return it as a load/load
505 // clobber result, allowing the client to decide to widen the load if
507 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType()))
508 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() &&
509 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
510 MemLocOffset, LI, DL))
511 return MemDepResult::getClobber(Inst);
516 // Must aliased loads are defs of each other.
517 if (R == AliasAnalysis::MustAlias)
518 return MemDepResult::getDef(Inst);
520 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
521 // in terms of clobbering loads, but since it does this by looking
522 // at the clobbering load directly, it doesn't know about any
523 // phi translation that may have happened along the way.
525 // If we have a partial alias, then return this as a clobber for the
527 if (R == AliasAnalysis::PartialAlias)
528 return MemDepResult::getClobber(Inst);
531 // Random may-alias loads don't depend on each other without a
536 // Stores don't depend on other no-aliased accesses.
537 if (R == AliasAnalysis::NoAlias)
540 // Stores don't alias loads from read-only memory.
541 if (AA->pointsToConstantMemory(LoadLoc))
544 // Stores depend on may/must aliased loads.
545 return MemDepResult::getDef(Inst);
548 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
549 // Atomic stores have complications involved.
550 // A Monotonic store is OK if the query inst is itself not atomic.
551 // A Release (or higher) store further requires that no acquire load
553 // FIXME: This is overly conservative.
554 if (!SI->isUnordered()) {
556 return MemDepResult::getClobber(SI);
557 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
558 if (!QueryLI->isSimple())
559 return MemDepResult::getClobber(SI);
560 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
561 if (!QuerySI->isSimple())
562 return MemDepResult::getClobber(SI);
563 } else if (QueryInst->mayReadOrWriteMemory()) {
564 return MemDepResult::getClobber(SI);
567 if (HasSeenAcquire && isAtLeastRelease(SI->getOrdering()))
568 return MemDepResult::getClobber(SI);
571 // FIXME: this is overly conservative.
572 // While volatile access cannot be eliminated, they do not have to clobber
573 // non-aliasing locations, as normal accesses can for example be reordered
574 // with volatile accesses.
575 if (SI->isVolatile())
576 return MemDepResult::getClobber(SI);
578 // If alias analysis can tell that this store is guaranteed to not modify
579 // the query pointer, ignore it. Use getModRefInfo to handle cases where
580 // the query pointer points to constant memory etc.
581 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
584 // Ok, this store might clobber the query pointer. Check to see if it is
585 // a must alias: in this case, we want to return this as a def.
586 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
588 // If we found a pointer, check if it could be the same as our pointer.
589 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
591 if (R == AliasAnalysis::NoAlias)
593 if (R == AliasAnalysis::MustAlias)
594 return MemDepResult::getDef(Inst);
597 return MemDepResult::getClobber(Inst);
600 // If this is an allocation, and if we know that the accessed pointer is to
601 // the allocation, return Def. This means that there is no dependence and
602 // the access can be optimized based on that. For example, a load could
604 // Note: Only determine this to be a malloc if Inst is the malloc call, not
605 // a subsequent bitcast of the malloc call result. There can be stores to
606 // the malloced memory between the malloc call and its bitcast uses, and we
607 // need to continue scanning until the malloc call.
608 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
609 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
610 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
612 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
613 return MemDepResult::getDef(Inst);
614 // Be conservative if the accessed pointer may alias the allocation.
615 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
616 return MemDepResult::getClobber(Inst);
617 // If the allocation is not aliased and does not read memory (like
618 // strdup), it is safe to ignore.
619 if (isa<AllocaInst>(Inst) ||
620 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
624 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
625 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
626 // If necessary, perform additional analysis.
627 if (MR == AliasAnalysis::ModRef)
628 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
630 case AliasAnalysis::NoModRef:
631 // If the call has no effect on the queried pointer, just ignore it.
633 case AliasAnalysis::Mod:
634 return MemDepResult::getClobber(Inst);
635 case AliasAnalysis::Ref:
636 // If the call is known to never store to the pointer, and if this is a
637 // load query, we can safely ignore it (scan past it).
641 // Otherwise, there is a potential dependence. Return a clobber.
642 return MemDepResult::getClobber(Inst);
646 // No dependence found. If this is the entry block of the function, it is
647 // unknown, otherwise it is non-local.
648 if (BB != &BB->getParent()->getEntryBlock())
649 return MemDepResult::getNonLocal();
650 return MemDepResult::getNonFuncLocal();
653 /// getDependency - Return the instruction on which a memory operation
655 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
656 Instruction *ScanPos = QueryInst;
658 // Check for a cached result
659 MemDepResult &LocalCache = LocalDeps[QueryInst];
661 // If the cached entry is non-dirty, just return it. Note that this depends
662 // on MemDepResult's default constructing to 'dirty'.
663 if (!LocalCache.isDirty())
666 // Otherwise, if we have a dirty entry, we know we can start the scan at that
667 // instruction, which may save us some work.
668 if (Instruction *Inst = LocalCache.getInst()) {
671 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
674 BasicBlock *QueryParent = QueryInst->getParent();
677 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
678 // No dependence found. If this is the entry block of the function, it is
679 // unknown, otherwise it is non-local.
680 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
681 LocalCache = MemDepResult::getNonLocal();
683 LocalCache = MemDepResult::getNonFuncLocal();
685 AliasAnalysis::Location MemLoc;
686 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
688 // If we can do a pointer scan, make it happen.
689 bool isLoad = !(MR & AliasAnalysis::Mod);
690 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
691 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
693 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
694 QueryParent, QueryInst);
695 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
696 CallSite QueryCS(QueryInst);
697 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
698 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
701 // Non-memory instruction.
702 LocalCache = MemDepResult::getUnknown();
705 // Remember the result!
706 if (Instruction *I = LocalCache.getInst())
707 ReverseLocalDeps[I].insert(QueryInst);
713 /// AssertSorted - This method is used when -debug is specified to verify that
714 /// cache arrays are properly kept sorted.
715 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
717 if (Count == -1) Count = Cache.size();
718 if (Count == 0) return;
720 for (unsigned i = 1; i != unsigned(Count); ++i)
721 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
725 /// getNonLocalCallDependency - Perform a full dependency query for the
726 /// specified call, returning the set of blocks that the value is
727 /// potentially live across. The returned set of results will include a
728 /// "NonLocal" result for all blocks where the value is live across.
730 /// This method assumes the instruction returns a "NonLocal" dependency
731 /// within its own block.
733 /// This returns a reference to an internal data structure that may be
734 /// invalidated on the next non-local query or when an instruction is
735 /// removed. Clients must copy this data if they want it around longer than
737 const MemoryDependenceAnalysis::NonLocalDepInfo &
738 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
739 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
740 "getNonLocalCallDependency should only be used on calls with non-local deps!");
741 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
742 NonLocalDepInfo &Cache = CacheP.first;
744 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
745 /// the cached case, this can happen due to instructions being deleted etc. In
746 /// the uncached case, this starts out as the set of predecessors we care
748 SmallVector<BasicBlock*, 32> DirtyBlocks;
750 if (!Cache.empty()) {
751 // Okay, we have a cache entry. If we know it is not dirty, just return it
752 // with no computation.
753 if (!CacheP.second) {
758 // If we already have a partially computed set of results, scan them to
759 // determine what is dirty, seeding our initial DirtyBlocks worklist.
760 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
762 if (I->getResult().isDirty())
763 DirtyBlocks.push_back(I->getBB());
765 // Sort the cache so that we can do fast binary search lookups below.
766 std::sort(Cache.begin(), Cache.end());
768 ++NumCacheDirtyNonLocal;
769 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
770 // << Cache.size() << " cached: " << *QueryInst;
772 // Seed DirtyBlocks with each of the preds of QueryInst's block.
773 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
774 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
775 DirtyBlocks.push_back(*PI);
776 ++NumUncacheNonLocal;
779 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
780 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
782 SmallPtrSet<BasicBlock*, 64> Visited;
784 unsigned NumSortedEntries = Cache.size();
785 DEBUG(AssertSorted(Cache));
787 // Iterate while we still have blocks to update.
788 while (!DirtyBlocks.empty()) {
789 BasicBlock *DirtyBB = DirtyBlocks.back();
790 DirtyBlocks.pop_back();
792 // Already processed this block?
793 if (!Visited.insert(DirtyBB).second)
796 // Do a binary search to see if we already have an entry for this block in
797 // the cache set. If so, find it.
798 DEBUG(AssertSorted(Cache, NumSortedEntries));
799 NonLocalDepInfo::iterator Entry =
800 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
801 NonLocalDepEntry(DirtyBB));
802 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
805 NonLocalDepEntry *ExistingResult = nullptr;
806 if (Entry != Cache.begin()+NumSortedEntries &&
807 Entry->getBB() == DirtyBB) {
808 // If we already have an entry, and if it isn't already dirty, the block
810 if (!Entry->getResult().isDirty())
813 // Otherwise, remember this slot so we can update the value.
814 ExistingResult = &*Entry;
817 // If the dirty entry has a pointer, start scanning from it so we don't have
818 // to rescan the entire block.
819 BasicBlock::iterator ScanPos = DirtyBB->end();
820 if (ExistingResult) {
821 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
823 // We're removing QueryInst's use of Inst.
824 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
825 QueryCS.getInstruction());
829 // Find out if this block has a local dependency for QueryInst.
832 if (ScanPos != DirtyBB->begin()) {
833 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
834 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
835 // No dependence found. If this is the entry block of the function, it is
836 // a clobber, otherwise it is unknown.
837 Dep = MemDepResult::getNonLocal();
839 Dep = MemDepResult::getNonFuncLocal();
842 // If we had a dirty entry for the block, update it. Otherwise, just add
845 ExistingResult->setResult(Dep);
847 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
849 // If the block has a dependency (i.e. it isn't completely transparent to
850 // the value), remember the association!
851 if (!Dep.isNonLocal()) {
852 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
853 // update this when we remove instructions.
854 if (Instruction *Inst = Dep.getInst())
855 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
858 // If the block *is* completely transparent to the load, we need to check
859 // the predecessors of this block. Add them to our worklist.
860 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
861 DirtyBlocks.push_back(*PI);
868 /// getNonLocalPointerDependency - Perform a full dependency query for an
869 /// access to the specified (non-volatile) memory location, returning the
870 /// set of instructions that either define or clobber the value.
872 /// This method assumes the pointer has a "NonLocal" dependency within its
875 void MemoryDependenceAnalysis::
876 getNonLocalPointerDependency(Instruction *QueryInst,
877 SmallVectorImpl<NonLocalDepResult> &Result) {
879 auto getLocation = [](AliasAnalysis *AA, Instruction *Inst) {
880 if (auto *I = dyn_cast<LoadInst>(Inst))
881 return AA->getLocation(I);
882 else if (auto *I = dyn_cast<StoreInst>(Inst))
883 return AA->getLocation(I);
884 else if (auto *I = dyn_cast<VAArgInst>(Inst))
885 return AA->getLocation(I);
886 else if (auto *I = dyn_cast<AtomicCmpXchgInst>(Inst))
887 return AA->getLocation(I);
888 else if (auto *I = dyn_cast<AtomicRMWInst>(Inst))
889 return AA->getLocation(I);
891 llvm_unreachable("unsupported memory instruction");
894 const AliasAnalysis::Location Loc = getLocation(AA, QueryInst);
895 bool isLoad = isa<LoadInst>(QueryInst);
896 BasicBlock *FromBB = QueryInst->getParent();
899 assert(Loc.Ptr->getType()->isPointerTy() &&
900 "Can't get pointer deps of a non-pointer!");
903 // This routine does not expect to deal with volatile instructions.
904 // Doing so would require piping through the QueryInst all the way through.
905 // TODO: volatiles can't be elided, but they can be reordered with other
906 // non-volatile accesses.
908 // We currently give up on any instruction which is ordered, but we do handle
909 // atomic instructions which are unordered.
910 // TODO: Handle ordered instructions
911 auto isOrdered = [](Instruction *Inst) {
912 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
913 return !LI->isUnordered();
914 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
915 return !SI->isUnordered();
919 if (isVolatile(QueryInst) || isOrdered(QueryInst)) {
920 Result.push_back(NonLocalDepResult(FromBB,
921 MemDepResult::getUnknown(),
922 const_cast<Value *>(Loc.Ptr)));
927 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, AC);
929 // This is the set of blocks we've inspected, and the pointer we consider in
930 // each block. Because of critical edges, we currently bail out if querying
931 // a block with multiple different pointers. This can happen during PHI
933 DenseMap<BasicBlock*, Value*> Visited;
934 if (!getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
935 Result, Visited, true))
938 Result.push_back(NonLocalDepResult(FromBB,
939 MemDepResult::getUnknown(),
940 const_cast<Value *>(Loc.Ptr)));
943 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
944 /// Pointer/PointeeSize using either cached information in Cache or by doing a
945 /// lookup (which may use dirty cache info if available). If we do a lookup,
946 /// add the result to the cache.
947 MemDepResult MemoryDependenceAnalysis::
948 GetNonLocalInfoForBlock(Instruction *QueryInst,
949 const AliasAnalysis::Location &Loc,
950 bool isLoad, BasicBlock *BB,
951 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
953 // Do a binary search to see if we already have an entry for this block in
954 // the cache set. If so, find it.
955 NonLocalDepInfo::iterator Entry =
956 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
957 NonLocalDepEntry(BB));
958 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
961 NonLocalDepEntry *ExistingResult = nullptr;
962 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
963 ExistingResult = &*Entry;
965 // If we have a cached entry, and it is non-dirty, use it as the value for
967 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
968 ++NumCacheNonLocalPtr;
969 return ExistingResult->getResult();
972 // Otherwise, we have to scan for the value. If we have a dirty cache
973 // entry, start scanning from its position, otherwise we scan from the end
975 BasicBlock::iterator ScanPos = BB->end();
976 if (ExistingResult && ExistingResult->getResult().getInst()) {
977 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
978 "Instruction invalidated?");
979 ++NumCacheDirtyNonLocalPtr;
980 ScanPos = ExistingResult->getResult().getInst();
982 // Eliminating the dirty entry from 'Cache', so update the reverse info.
983 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
984 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
986 ++NumUncacheNonLocalPtr;
989 // Scan the block for the dependency.
990 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB,
993 // If we had a dirty entry for the block, update it. Otherwise, just add
996 ExistingResult->setResult(Dep);
998 Cache->push_back(NonLocalDepEntry(BB, Dep));
1000 // If the block has a dependency (i.e. it isn't completely transparent to
1001 // the value), remember the reverse association because we just added it
1003 if (!Dep.isDef() && !Dep.isClobber())
1006 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
1007 // update MemDep when we remove instructions.
1008 Instruction *Inst = Dep.getInst();
1009 assert(Inst && "Didn't depend on anything?");
1010 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
1011 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
1015 /// SortNonLocalDepInfoCache - Sort the NonLocalDepInfo cache, given a certain
1016 /// number of elements in the array that are already properly ordered. This is
1017 /// optimized for the case when only a few entries are added.
1019 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
1020 unsigned NumSortedEntries) {
1021 switch (Cache.size() - NumSortedEntries) {
1023 // done, no new entries.
1026 // Two new entries, insert the last one into place.
1027 NonLocalDepEntry Val = Cache.back();
1029 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1030 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
1031 Cache.insert(Entry, Val);
1035 // One new entry, Just insert the new value at the appropriate position.
1036 if (Cache.size() != 1) {
1037 NonLocalDepEntry Val = Cache.back();
1039 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1040 std::upper_bound(Cache.begin(), Cache.end(), Val);
1041 Cache.insert(Entry, Val);
1045 // Added many values, do a full scale sort.
1046 std::sort(Cache.begin(), Cache.end());
1051 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
1052 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
1053 /// results to the results vector and keep track of which blocks are visited in
1056 /// This has special behavior for the first block queries (when SkipFirstBlock
1057 /// is true). In this special case, it ignores the contents of the specified
1058 /// block and starts returning dependence info for its predecessors.
1060 /// This function returns false on success, or true to indicate that it could
1061 /// not compute dependence information for some reason. This should be treated
1062 /// as a clobber dependence on the first instruction in the predecessor block.
1063 bool MemoryDependenceAnalysis::
1064 getNonLocalPointerDepFromBB(Instruction *QueryInst,
1065 const PHITransAddr &Pointer,
1066 const AliasAnalysis::Location &Loc,
1067 bool isLoad, BasicBlock *StartBB,
1068 SmallVectorImpl<NonLocalDepResult> &Result,
1069 DenseMap<BasicBlock*, Value*> &Visited,
1070 bool SkipFirstBlock) {
1071 // Look up the cached info for Pointer.
1072 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1074 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1075 // CacheKey, this value will be inserted as the associated value. Otherwise,
1076 // it'll be ignored, and we'll have to check to see if the cached size and
1077 // aa tags are consistent with the current query.
1078 NonLocalPointerInfo InitialNLPI;
1079 InitialNLPI.Size = Loc.Size;
1080 InitialNLPI.AATags = Loc.AATags;
1082 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1083 // already have one.
1084 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1085 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1086 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1088 // If we already have a cache entry for this CacheKey, we may need to do some
1089 // work to reconcile the cache entry and the current query.
1091 if (CacheInfo->Size < Loc.Size) {
1092 // The query's Size is greater than the cached one. Throw out the
1093 // cached data and proceed with the query at the greater size.
1094 CacheInfo->Pair = BBSkipFirstBlockPair();
1095 CacheInfo->Size = Loc.Size;
1096 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1097 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1098 if (Instruction *Inst = DI->getResult().getInst())
1099 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1100 CacheInfo->NonLocalDeps.clear();
1101 } else if (CacheInfo->Size > Loc.Size) {
1102 // This query's Size is less than the cached one. Conservatively restart
1103 // the query using the greater size.
1104 return getNonLocalPointerDepFromBB(QueryInst, Pointer,
1105 Loc.getWithNewSize(CacheInfo->Size),
1106 isLoad, StartBB, Result, Visited,
1110 // If the query's AATags are inconsistent with the cached one,
1111 // conservatively throw out the cached data and restart the query with
1112 // no tag if needed.
1113 if (CacheInfo->AATags != Loc.AATags) {
1114 if (CacheInfo->AATags) {
1115 CacheInfo->Pair = BBSkipFirstBlockPair();
1116 CacheInfo->AATags = AAMDNodes();
1117 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1118 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1119 if (Instruction *Inst = DI->getResult().getInst())
1120 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1121 CacheInfo->NonLocalDeps.clear();
1124 return getNonLocalPointerDepFromBB(QueryInst,
1125 Pointer, Loc.getWithoutAATags(),
1126 isLoad, StartBB, Result, Visited,
1131 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1133 // If we have valid cached information for exactly the block we are
1134 // investigating, just return it with no recomputation.
1135 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1136 // We have a fully cached result for this query then we can just return the
1137 // cached results and populate the visited set. However, we have to verify
1138 // that we don't already have conflicting results for these blocks. Check
1139 // to ensure that if a block in the results set is in the visited set that
1140 // it was for the same pointer query.
1141 if (!Visited.empty()) {
1142 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1144 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
1145 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1148 // We have a pointer mismatch in a block. Just return clobber, saying
1149 // that something was clobbered in this result. We could also do a
1150 // non-fully cached query, but there is little point in doing this.
1155 Value *Addr = Pointer.getAddr();
1156 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1158 Visited.insert(std::make_pair(I->getBB(), Addr));
1159 if (I->getResult().isNonLocal()) {
1164 Result.push_back(NonLocalDepResult(I->getBB(),
1165 MemDepResult::getUnknown(),
1167 } else if (DT->isReachableFromEntry(I->getBB())) {
1168 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1171 ++NumCacheCompleteNonLocalPtr;
1175 // Otherwise, either this is a new block, a block with an invalid cache
1176 // pointer or one that we're about to invalidate by putting more info into it
1177 // than its valid cache info. If empty, the result will be valid cache info,
1178 // otherwise it isn't.
1180 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1182 CacheInfo->Pair = BBSkipFirstBlockPair();
1184 SmallVector<BasicBlock*, 32> Worklist;
1185 Worklist.push_back(StartBB);
1187 // PredList used inside loop.
1188 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1190 // Keep track of the entries that we know are sorted. Previously cached
1191 // entries will all be sorted. The entries we add we only sort on demand (we
1192 // don't insert every element into its sorted position). We know that we
1193 // won't get any reuse from currently inserted values, because we don't
1194 // revisit blocks after we insert info for them.
1195 unsigned NumSortedEntries = Cache->size();
1196 DEBUG(AssertSorted(*Cache));
1198 while (!Worklist.empty()) {
1199 BasicBlock *BB = Worklist.pop_back_val();
1201 // If we do process a large number of blocks it becomes very expensive and
1202 // likely it isn't worth worrying about
1203 if (Result.size() > NumResultsLimit) {
1205 // Sort it now (if needed) so that recursive invocations of
1206 // getNonLocalPointerDepFromBB and other routines that could reuse the
1207 // cache value will only see properly sorted cache arrays.
1208 if (Cache && NumSortedEntries != Cache->size()) {
1209 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1211 // Since we bail out, the "Cache" set won't contain all of the
1212 // results for the query. This is ok (we can still use it to accelerate
1213 // specific block queries) but we can't do the fastpath "return all
1214 // results from the set". Clear out the indicator for this.
1215 CacheInfo->Pair = BBSkipFirstBlockPair();
1219 // Skip the first block if we have it.
1220 if (!SkipFirstBlock) {
1221 // Analyze the dependency of *Pointer in FromBB. See if we already have
1223 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1225 // Get the dependency info for Pointer in BB. If we have cached
1226 // information, we will use it, otherwise we compute it.
1227 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1228 MemDepResult Dep = GetNonLocalInfoForBlock(QueryInst,
1229 Loc, isLoad, BB, Cache,
1232 // If we got a Def or Clobber, add this to the list of results.
1233 if (!Dep.isNonLocal()) {
1235 Result.push_back(NonLocalDepResult(BB,
1236 MemDepResult::getUnknown(),
1237 Pointer.getAddr()));
1239 } else if (DT->isReachableFromEntry(BB)) {
1240 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1246 // If 'Pointer' is an instruction defined in this block, then we need to do
1247 // phi translation to change it into a value live in the predecessor block.
1248 // If not, we just add the predecessors to the worklist and scan them with
1249 // the same Pointer.
1250 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1251 SkipFirstBlock = false;
1252 SmallVector<BasicBlock*, 16> NewBlocks;
1253 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1254 // Verify that we haven't looked at this block yet.
1255 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1256 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1257 if (InsertRes.second) {
1258 // First time we've looked at *PI.
1259 NewBlocks.push_back(*PI);
1263 // If we have seen this block before, but it was with a different
1264 // pointer then we have a phi translation failure and we have to treat
1265 // this as a clobber.
1266 if (InsertRes.first->second != Pointer.getAddr()) {
1267 // Make sure to clean up the Visited map before continuing on to
1268 // PredTranslationFailure.
1269 for (unsigned i = 0; i < NewBlocks.size(); i++)
1270 Visited.erase(NewBlocks[i]);
1271 goto PredTranslationFailure;
1274 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1278 // We do need to do phi translation, if we know ahead of time we can't phi
1279 // translate this value, don't even try.
1280 if (!Pointer.IsPotentiallyPHITranslatable())
1281 goto PredTranslationFailure;
1283 // We may have added values to the cache list before this PHI translation.
1284 // If so, we haven't done anything to ensure that the cache remains sorted.
1285 // Sort it now (if needed) so that recursive invocations of
1286 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1287 // value will only see properly sorted cache arrays.
1288 if (Cache && NumSortedEntries != Cache->size()) {
1289 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1290 NumSortedEntries = Cache->size();
1295 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1296 BasicBlock *Pred = *PI;
1297 PredList.push_back(std::make_pair(Pred, Pointer));
1299 // Get the PHI translated pointer in this predecessor. This can fail if
1300 // not translatable, in which case the getAddr() returns null.
1301 PHITransAddr &PredPointer = PredList.back().second;
1302 PredPointer.PHITranslateValue(BB, Pred, nullptr);
1304 Value *PredPtrVal = PredPointer.getAddr();
1306 // Check to see if we have already visited this pred block with another
1307 // pointer. If so, we can't do this lookup. This failure can occur
1308 // with PHI translation when a critical edge exists and the PHI node in
1309 // the successor translates to a pointer value different than the
1310 // pointer the block was first analyzed with.
1311 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1312 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1314 if (!InsertRes.second) {
1315 // We found the pred; take it off the list of preds to visit.
1316 PredList.pop_back();
1318 // If the predecessor was visited with PredPtr, then we already did
1319 // the analysis and can ignore it.
1320 if (InsertRes.first->second == PredPtrVal)
1323 // Otherwise, the block was previously analyzed with a different
1324 // pointer. We can't represent the result of this case, so we just
1325 // treat this as a phi translation failure.
1327 // Make sure to clean up the Visited map before continuing on to
1328 // PredTranslationFailure.
1329 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1330 Visited.erase(PredList[i].first);
1332 goto PredTranslationFailure;
1336 // Actually process results here; this need to be a separate loop to avoid
1337 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1338 // any results for. (getNonLocalPointerDepFromBB will modify our
1339 // datastructures in ways the code after the PredTranslationFailure label
1341 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1342 BasicBlock *Pred = PredList[i].first;
1343 PHITransAddr &PredPointer = PredList[i].second;
1344 Value *PredPtrVal = PredPointer.getAddr();
1346 bool CanTranslate = true;
1347 // If PHI translation was unable to find an available pointer in this
1348 // predecessor, then we have to assume that the pointer is clobbered in
1349 // that predecessor. We can still do PRE of the load, which would insert
1350 // a computation of the pointer in this predecessor.
1352 CanTranslate = false;
1354 // FIXME: it is entirely possible that PHI translating will end up with
1355 // the same value. Consider PHI translating something like:
1356 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1357 // to recurse here, pedantically speaking.
1359 // If getNonLocalPointerDepFromBB fails here, that means the cached
1360 // result conflicted with the Visited list; we have to conservatively
1361 // assume it is unknown, but this also does not block PRE of the load.
1362 if (!CanTranslate ||
1363 getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1364 Loc.getWithNewPtr(PredPtrVal),
1367 // Add the entry to the Result list.
1368 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1369 Result.push_back(Entry);
1371 // Since we had a phi translation failure, the cache for CacheKey won't
1372 // include all of the entries that we need to immediately satisfy future
1373 // queries. Mark this in NonLocalPointerDeps by setting the
1374 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1375 // cached value to do more work but not miss the phi trans failure.
1376 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1377 NLPI.Pair = BBSkipFirstBlockPair();
1382 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1383 CacheInfo = &NonLocalPointerDeps[CacheKey];
1384 Cache = &CacheInfo->NonLocalDeps;
1385 NumSortedEntries = Cache->size();
1387 // Since we did phi translation, the "Cache" set won't contain all of the
1388 // results for the query. This is ok (we can still use it to accelerate
1389 // specific block queries) but we can't do the fastpath "return all
1390 // results from the set" Clear out the indicator for this.
1391 CacheInfo->Pair = BBSkipFirstBlockPair();
1392 SkipFirstBlock = false;
1395 PredTranslationFailure:
1396 // The following code is "failure"; we can't produce a sane translation
1397 // for the given block. It assumes that we haven't modified any of
1398 // our datastructures while processing the current block.
1401 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1402 CacheInfo = &NonLocalPointerDeps[CacheKey];
1403 Cache = &CacheInfo->NonLocalDeps;
1404 NumSortedEntries = Cache->size();
1407 // Since we failed phi translation, the "Cache" set won't contain all of the
1408 // results for the query. This is ok (we can still use it to accelerate
1409 // specific block queries) but we can't do the fastpath "return all
1410 // results from the set". Clear out the indicator for this.
1411 CacheInfo->Pair = BBSkipFirstBlockPair();
1413 // If *nothing* works, mark the pointer as unknown.
1415 // If this is the magic first block, return this as a clobber of the whole
1416 // incoming value. Since we can't phi translate to one of the predecessors,
1417 // we have to bail out.
1421 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1422 assert(I != Cache->rend() && "Didn't find current block??");
1423 if (I->getBB() != BB)
1426 assert((I->getResult().isNonLocal() || !DT->isReachableFromEntry(BB)) &&
1427 "Should only be here with transparent block");
1428 I->setResult(MemDepResult::getUnknown());
1429 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1430 Pointer.getAddr()));
1435 // Okay, we're done now. If we added new values to the cache, re-sort it.
1436 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1437 DEBUG(AssertSorted(*Cache));
1441 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1442 /// CachedNonLocalPointerInfo, remove it.
1443 void MemoryDependenceAnalysis::
1444 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1445 CachedNonLocalPointerInfo::iterator It =
1446 NonLocalPointerDeps.find(P);
1447 if (It == NonLocalPointerDeps.end()) return;
1449 // Remove all of the entries in the BB->val map. This involves removing
1450 // instructions from the reverse map.
1451 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1453 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1454 Instruction *Target = PInfo[i].getResult().getInst();
1455 if (!Target) continue; // Ignore non-local dep results.
1456 assert(Target->getParent() == PInfo[i].getBB());
1458 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1459 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1462 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1463 NonLocalPointerDeps.erase(It);
1467 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1468 /// information about the specified pointer, because it may be too
1469 /// conservative in memdep. This is an optional call that can be used when
1470 /// the client detects an equivalence between the pointer and some other
1471 /// value and replaces the other value with ptr. This can make Ptr available
1472 /// in more places that cached info does not necessarily keep.
1473 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1474 // If Ptr isn't really a pointer, just ignore it.
1475 if (!Ptr->getType()->isPointerTy()) return;
1476 // Flush store info for the pointer.
1477 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1478 // Flush load info for the pointer.
1479 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1482 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1483 /// This needs to be done when the CFG changes, e.g., due to splitting
1485 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1489 /// removeInstruction - Remove an instruction from the dependence analysis,
1490 /// updating the dependence of instructions that previously depended on it.
1491 /// This method attempts to keep the cache coherent using the reverse map.
1492 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1493 // Walk through the Non-local dependencies, removing this one as the value
1494 // for any cached queries.
1495 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1496 if (NLDI != NonLocalDeps.end()) {
1497 NonLocalDepInfo &BlockMap = NLDI->second.first;
1498 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1500 if (Instruction *Inst = DI->getResult().getInst())
1501 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1502 NonLocalDeps.erase(NLDI);
1505 // If we have a cached local dependence query for this instruction, remove it.
1507 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1508 if (LocalDepEntry != LocalDeps.end()) {
1509 // Remove us from DepInst's reverse set now that the local dep info is gone.
1510 if (Instruction *Inst = LocalDepEntry->second.getInst())
1511 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1513 // Remove this local dependency info.
1514 LocalDeps.erase(LocalDepEntry);
1517 // If we have any cached pointer dependencies on this instruction, remove
1518 // them. If the instruction has non-pointer type, then it can't be a pointer
1521 // Remove it from both the load info and the store info. The instruction
1522 // can't be in either of these maps if it is non-pointer.
1523 if (RemInst->getType()->isPointerTy()) {
1524 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1525 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1528 // Loop over all of the things that depend on the instruction we're removing.
1530 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1532 // If we find RemInst as a clobber or Def in any of the maps for other values,
1533 // we need to replace its entry with a dirty version of the instruction after
1534 // it. If RemInst is a terminator, we use a null dirty value.
1536 // Using a dirty version of the instruction after RemInst saves having to scan
1537 // the entire block to get to this point.
1538 MemDepResult NewDirtyVal;
1539 if (!RemInst->isTerminator())
1540 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1542 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1543 if (ReverseDepIt != ReverseLocalDeps.end()) {
1544 // RemInst can't be the terminator if it has local stuff depending on it.
1545 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1546 "Nothing can locally depend on a terminator");
1548 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1549 assert(InstDependingOnRemInst != RemInst &&
1550 "Already removed our local dep info");
1552 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1554 // Make sure to remember that new things depend on NewDepInst.
1555 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1556 "a local dep on this if it is a terminator!");
1557 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1558 InstDependingOnRemInst));
1561 ReverseLocalDeps.erase(ReverseDepIt);
1563 // Add new reverse deps after scanning the set, to avoid invalidating the
1564 // 'ReverseDeps' reference.
1565 while (!ReverseDepsToAdd.empty()) {
1566 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1567 .insert(ReverseDepsToAdd.back().second);
1568 ReverseDepsToAdd.pop_back();
1572 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1573 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1574 for (Instruction *I : ReverseDepIt->second) {
1575 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1577 PerInstNLInfo &INLD = NonLocalDeps[I];
1578 // The information is now dirty!
1581 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1582 DE = INLD.first.end(); DI != DE; ++DI) {
1583 if (DI->getResult().getInst() != RemInst) continue;
1585 // Convert to a dirty entry for the subsequent instruction.
1586 DI->setResult(NewDirtyVal);
1588 if (Instruction *NextI = NewDirtyVal.getInst())
1589 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1593 ReverseNonLocalDeps.erase(ReverseDepIt);
1595 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1596 while (!ReverseDepsToAdd.empty()) {
1597 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1598 .insert(ReverseDepsToAdd.back().second);
1599 ReverseDepsToAdd.pop_back();
1603 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1604 // value in the NonLocalPointerDeps info.
1605 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1606 ReverseNonLocalPtrDeps.find(RemInst);
1607 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1608 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1610 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1611 assert(P.getPointer() != RemInst &&
1612 "Already removed NonLocalPointerDeps info for RemInst");
1614 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1616 // The cache is not valid for any specific block anymore.
1617 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1619 // Update any entries for RemInst to use the instruction after it.
1620 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1622 if (DI->getResult().getInst() != RemInst) continue;
1624 // Convert to a dirty entry for the subsequent instruction.
1625 DI->setResult(NewDirtyVal);
1627 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1628 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1631 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1632 // subsequent value may invalidate the sortedness.
1633 std::sort(NLPDI.begin(), NLPDI.end());
1636 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1638 while (!ReversePtrDepsToAdd.empty()) {
1639 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1640 .insert(ReversePtrDepsToAdd.back().second);
1641 ReversePtrDepsToAdd.pop_back();
1646 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1647 AA->deleteValue(RemInst);
1648 DEBUG(verifyRemoved(RemInst));
1650 /// verifyRemoved - Verify that the specified instruction does not occur
1651 /// in our internal data structures. This function verifies by asserting in
1653 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1655 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1656 E = LocalDeps.end(); I != E; ++I) {
1657 assert(I->first != D && "Inst occurs in data structures");
1658 assert(I->second.getInst() != D &&
1659 "Inst occurs in data structures");
1662 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1663 E = NonLocalPointerDeps.end(); I != E; ++I) {
1664 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1665 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1666 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1668 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1671 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1672 E = NonLocalDeps.end(); I != E; ++I) {
1673 assert(I->first != D && "Inst occurs in data structures");
1674 const PerInstNLInfo &INLD = I->second;
1675 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1676 EE = INLD.first.end(); II != EE; ++II)
1677 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1680 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1681 E = ReverseLocalDeps.end(); I != E; ++I) {
1682 assert(I->first != D && "Inst occurs in data structures");
1683 for (Instruction *Inst : I->second)
1684 assert(Inst != D && "Inst occurs in data structures");
1687 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1688 E = ReverseNonLocalDeps.end();
1690 assert(I->first != D && "Inst occurs in data structures");
1691 for (Instruction *Inst : I->second)
1692 assert(Inst != D && "Inst occurs in data structures");
1695 for (ReverseNonLocalPtrDepTy::const_iterator
1696 I = ReverseNonLocalPtrDeps.begin(),
1697 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1698 assert(I->first != D && "Inst occurs in rev NLPD map");
1700 for (ValueIsLoadPair P : I->second)
1701 assert(P != ValueIsLoadPair(D, false) &&
1702 P != ValueIsLoadPair(D, true) &&
1703 "Inst occurs in ReverseNonLocalPtrDeps map");