1 //===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation -------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements an analysis that determines, for a given memory
10 // operation, what preceding memory operations it depends on.  It builds on
11 // alias analysis information, and tries to provide a lazy, caching interface to
12 // a common kind of alias information query.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/AssumptionCache.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/Analysis/MemoryLocation.h"
26 #include "llvm/Analysis/PHITransAddr.h"
27 #include "llvm/Analysis/PhiValues.h"
28 #include "llvm/Analysis/TargetLibraryInfo.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/InstrTypes.h"
34 #include "llvm/IR/Instruction.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/LLVMContext.h"
38 #include "llvm/IR/Metadata.h"
39 #include "llvm/IR/Module.h"
40 #include "llvm/IR/PredIteratorCache.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/IR/Use.h"
43 #include "llvm/IR/Value.h"
44 #include "llvm/InitializePasses.h"
45 #include "llvm/Pass.h"
46 #include "llvm/Support/AtomicOrdering.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Compiler.h"
50 #include "llvm/Support/Debug.h"
51 #include <algorithm>
52 #include <cassert>
53 #include <iterator>
54 #include <utility>
55 
56 using namespace llvm;
57 
58 #define DEBUG_TYPE "memdep"
59 
60 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
61 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
62 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
63 
64 STATISTIC(NumCacheNonLocalPtr,
65           "Number of fully cached non-local ptr responses");
66 STATISTIC(NumCacheDirtyNonLocalPtr,
67           "Number of cached, but dirty, non-local ptr responses");
68 STATISTIC(NumUncacheNonLocalPtr, "Number of uncached non-local ptr responses");
69 STATISTIC(NumCacheCompleteNonLocalPtr,
70           "Number of block queries that were completely cached");
71 
72 // Limit for the number of instructions to scan in a block.
73 
74 static cl::opt<unsigned> BlockScanLimit(
75     "memdep-block-scan-limit", cl::Hidden, cl::init(100),
76     cl::desc("The number of instructions to scan in a block in memory "
77              "dependency analysis (default = 100)"));
78 
79 static cl::opt<unsigned>
80     BlockNumberLimit("memdep-block-number-limit", cl::Hidden, cl::init(1000),
81                      cl::desc("The number of blocks to scan during memory "
82                               "dependency analysis (default = 1000)"));
83 
84 // Limit on the number of memdep results to process.
85 static const unsigned int NumResultsLimit = 100;
86 
87 /// This is a helper function that removes Val from 'Inst's set in ReverseMap.
88 ///
89 /// If the set becomes empty, remove Inst's entry.
90 template <typename KeyTy>
91 static void
92 RemoveFromReverseMap(DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>> &ReverseMap,
93                      Instruction *Inst, KeyTy Val) {
94   typename DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>>::iterator InstIt =
95       ReverseMap.find(Inst);
96   assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
97   bool Found = InstIt->second.erase(Val);
98   assert(Found && "Invalid reverse map!");
99   (void)Found;
100   if (InstIt->second.empty())
101     ReverseMap.erase(InstIt);
102 }
103 
104 /// If the given instruction references a specific memory location, fill in Loc
105 /// with the details, otherwise set Loc.Ptr to null.
106 ///
107 /// Returns a ModRefInfo value describing the general behavior of the
108 /// instruction.
109 static ModRefInfo GetLocation(const Instruction *Inst, MemoryLocation &Loc,
110                               const TargetLibraryInfo &TLI) {
111   if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
112     if (LI->isUnordered()) {
113       Loc = MemoryLocation::get(LI);
114       return ModRefInfo::Ref;
115     }
116     if (LI->getOrdering() == AtomicOrdering::Monotonic) {
117       Loc = MemoryLocation::get(LI);
118       return ModRefInfo::ModRef;
119     }
120     Loc = MemoryLocation();
121     return ModRefInfo::ModRef;
122   }
123 
124   if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
125     if (SI->isUnordered()) {
126       Loc = MemoryLocation::get(SI);
127       return ModRefInfo::Mod;
128     }
129     if (SI->getOrdering() == AtomicOrdering::Monotonic) {
130       Loc = MemoryLocation::get(SI);
131       return ModRefInfo::ModRef;
132     }
133     Loc = MemoryLocation();
134     return ModRefInfo::ModRef;
135   }
136 
137   if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
138     Loc = MemoryLocation::get(V);
139     return ModRefInfo::ModRef;
140   }
141 
142   if (const CallBase *CB = dyn_cast<CallBase>(Inst)) {
143     if (Value *FreedOp = getFreedOperand(CB, &TLI)) {
144       // calls to free() deallocate the entire structure
145       Loc = MemoryLocation::getAfter(FreedOp);
146       return ModRefInfo::Mod;
147     }
148   }
149 
150   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
151     switch (II->getIntrinsicID()) {
152     case Intrinsic::lifetime_start:
153     case Intrinsic::lifetime_end:
154     case Intrinsic::invariant_start:
155       Loc = MemoryLocation::getForArgument(II, 1, TLI);
156       // These intrinsics don't really modify the memory, but returning Mod
157       // will allow them to be handled conservatively.
158       return ModRefInfo::Mod;
159     case Intrinsic::invariant_end:
160       Loc = MemoryLocation::getForArgument(II, 2, TLI);
161       // These intrinsics don't really modify the memory, but returning Mod
162       // will allow them to be handled conservatively.
163       return ModRefInfo::Mod;
164     case Intrinsic::masked_load:
165       Loc = MemoryLocation::getForArgument(II, 0, TLI);
166       return ModRefInfo::Ref;
167     case Intrinsic::masked_store:
168       Loc = MemoryLocation::getForArgument(II, 1, TLI);
169       return ModRefInfo::Mod;
170     default:
171       break;
172     }
173   }
174 
175   // Otherwise, just do the coarse-grained thing that always works.
176   if (Inst->mayWriteToMemory())
177     return ModRefInfo::ModRef;
178   if (Inst->mayReadFromMemory())
179     return ModRefInfo::Ref;
180   return ModRefInfo::NoModRef;
181 }
182 
183 /// Private helper for finding the local dependencies of a call site.
184 MemDepResult MemoryDependenceResults::getCallDependencyFrom(
185     CallBase *Call, bool isReadOnlyCall, BasicBlock::iterator ScanIt,
186     BasicBlock *BB) {
187   unsigned Limit = getDefaultBlockScanLimit();
188 
189   // Walk backwards through the block, looking for dependencies.
190   while (ScanIt != BB->begin()) {
191     Instruction *Inst = &*--ScanIt;
192     // Debug intrinsics don't cause dependences and should not affect Limit
193     if (isa<DbgInfoIntrinsic>(Inst))
194       continue;
195 
196     // Limit the amount of scanning we do so we don't end up with quadratic
197     // running time on extreme testcases.
198     --Limit;
199     if (!Limit)
200       return MemDepResult::getUnknown();
201 
202     // If this inst is a memory op, get the pointer it accessed
203     MemoryLocation Loc;
204     ModRefInfo MR = GetLocation(Inst, Loc, TLI);
205     if (Loc.Ptr) {
206       // A simple instruction.
207       if (isModOrRefSet(AA.getModRefInfo(Call, Loc)))
208         return MemDepResult::getClobber(Inst);
209       continue;
210     }
211 
212     if (auto *CallB = dyn_cast<CallBase>(Inst)) {
213       // If these two calls do not interfere, look past it.
214       if (isNoModRef(AA.getModRefInfo(Call, CallB))) {
215         // If the two calls are the same, return Inst as a Def, so that
216         // Call can be found redundant and eliminated.
217         if (isReadOnlyCall && !isModSet(MR) &&
218             Call->isIdenticalToWhenDefined(CallB))
219           return MemDepResult::getDef(Inst);
220 
221         // Otherwise if the two calls don't interact (e.g. CallB is readnone)
222         // keep scanning.
223         continue;
224       } else
225         return MemDepResult::getClobber(Inst);
226     }
227 
228     // If we could not obtain a pointer for the instruction and the instruction
229     // touches memory then assume that this is a dependency.
230     if (isModOrRefSet(MR))
231       return MemDepResult::getClobber(Inst);
232   }
233 
234   // No dependence found.  If this is the entry block of the function, it is
235   // unknown, otherwise it is non-local.
236   if (BB != &BB->getParent()->getEntryBlock())
237     return MemDepResult::getNonLocal();
238   return MemDepResult::getNonFuncLocal();
239 }
240 
241 MemDepResult MemoryDependenceResults::getPointerDependencyFrom(
242     const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
243     BasicBlock *BB, Instruction *QueryInst, unsigned *Limit,
244     BatchAAResults &BatchAA) {
245   MemDepResult InvariantGroupDependency = MemDepResult::getUnknown();
246   if (QueryInst != nullptr) {
247     if (auto *LI = dyn_cast<LoadInst>(QueryInst)) {
248       InvariantGroupDependency = getInvariantGroupPointerDependency(LI, BB);
249 
250       if (InvariantGroupDependency.isDef())
251         return InvariantGroupDependency;
252     }
253   }
254   MemDepResult SimpleDep = getSimplePointerDependencyFrom(
255       MemLoc, isLoad, ScanIt, BB, QueryInst, Limit, BatchAA);
256   if (SimpleDep.isDef())
257     return SimpleDep;
258   // Non-local invariant group dependency indicates there is non local Def
259   // (it only returns nonLocal if it finds nonLocal def), which is better than
260   // local clobber and everything else.
261   if (InvariantGroupDependency.isNonLocal())
262     return InvariantGroupDependency;
263 
264   assert(InvariantGroupDependency.isUnknown() &&
265          "InvariantGroupDependency should be only unknown at this point");
266   return SimpleDep;
267 }
268 
269 MemDepResult MemoryDependenceResults::getPointerDependencyFrom(
270     const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
271     BasicBlock *BB, Instruction *QueryInst, unsigned *Limit) {
272   BatchAAResults BatchAA(AA);
273   return getPointerDependencyFrom(MemLoc, isLoad, ScanIt, BB, QueryInst, Limit,
274                                   BatchAA);
275 }
276 
277 MemDepResult
278 MemoryDependenceResults::getInvariantGroupPointerDependency(LoadInst *LI,
279                                                             BasicBlock *BB) {
280 
281   if (!LI->hasMetadata(LLVMContext::MD_invariant_group))
282     return MemDepResult::getUnknown();
283 
284   // Take the ptr operand after all casts and geps 0. This way we can search
285   // cast graph down only.
286   Value *LoadOperand = LI->getPointerOperand()->stripPointerCasts();
287 
288   // It's is not safe to walk the use list of global value, because function
289   // passes aren't allowed to look outside their functions.
290   // FIXME: this could be fixed by filtering instructions from outside
291   // of current function.
292   if (isa<GlobalValue>(LoadOperand))
293     return MemDepResult::getUnknown();
294 
295   // Queue to process all pointers that are equivalent to load operand.
296   SmallVector<const Value *, 8> LoadOperandsQueue;
297   LoadOperandsQueue.push_back(LoadOperand);
298 
299   Instruction *ClosestDependency = nullptr;
300   // Order of instructions in uses list is unpredictible. In order to always
301   // get the same result, we will look for the closest dominance.
302   auto GetClosestDependency = [this](Instruction *Best, Instruction *Other) {
303     assert(Other && "Must call it with not null instruction");
304     if (Best == nullptr || DT.dominates(Best, Other))
305       return Other;
306     return Best;
307   };
308 
309   // FIXME: This loop is O(N^2) because dominates can be O(n) and in worst case
310   // we will see all the instructions. This should be fixed in MSSA.
311   while (!LoadOperandsQueue.empty()) {
312     const Value *Ptr = LoadOperandsQueue.pop_back_val();
313     assert(Ptr && !isa<GlobalValue>(Ptr) &&
314            "Null or GlobalValue should not be inserted");
315 
316     for (const Use &Us : Ptr->uses()) {
317       auto *U = dyn_cast<Instruction>(Us.getUser());
318       if (!U || U == LI || !DT.dominates(U, LI))
319         continue;
320 
321       // Bitcast or gep with zeros are using Ptr. Add to queue to check it's
322       // users.      U = bitcast Ptr
323       if (isa<BitCastInst>(U)) {
324         LoadOperandsQueue.push_back(U);
325         continue;
326       }
327       // Gep with zeros is equivalent to bitcast.
328       // FIXME: we are not sure if some bitcast should be canonicalized to gep 0
329       // or gep 0 to bitcast because of SROA, so there are 2 forms. When
330       // typeless pointers will be ready then both cases will be gone
331       // (and this BFS also won't be needed).
332       if (auto *GEP = dyn_cast<GetElementPtrInst>(U))
333         if (GEP->hasAllZeroIndices()) {
334           LoadOperandsQueue.push_back(U);
335           continue;
336         }
337 
338       // If we hit load/store with the same invariant.group metadata (and the
339       // same pointer operand) we can assume that value pointed by pointer
340       // operand didn't change.
341       if ((isa<LoadInst>(U) ||
342            (isa<StoreInst>(U) &&
343             cast<StoreInst>(U)->getPointerOperand() == Ptr)) &&
344           U->hasMetadata(LLVMContext::MD_invariant_group))
345         ClosestDependency = GetClosestDependency(ClosestDependency, U);
346     }
347   }
348 
349   if (!ClosestDependency)
350     return MemDepResult::getUnknown();
351   if (ClosestDependency->getParent() == BB)
352     return MemDepResult::getDef(ClosestDependency);
353   // Def(U) can't be returned here because it is non-local. If local
354   // dependency won't be found then return nonLocal counting that the
355   // user will call getNonLocalPointerDependency, which will return cached
356   // result.
357   NonLocalDefsCache.try_emplace(
358       LI, NonLocalDepResult(ClosestDependency->getParent(),
359                             MemDepResult::getDef(ClosestDependency), nullptr));
360   ReverseNonLocalDefsCache[ClosestDependency].insert(LI);
361   return MemDepResult::getNonLocal();
362 }
363 
364 MemDepResult MemoryDependenceResults::getSimplePointerDependencyFrom(
365     const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
366     BasicBlock *BB, Instruction *QueryInst, unsigned *Limit,
367     BatchAAResults &BatchAA) {
368   bool isInvariantLoad = false;
369 
370   unsigned DefaultLimit = getDefaultBlockScanLimit();
371   if (!Limit)
372     Limit = &DefaultLimit;
373 
374   // We must be careful with atomic accesses, as they may allow another thread
375   //   to touch this location, clobbering it. We are conservative: if the
376   //   QueryInst is not a simple (non-atomic) memory access, we automatically
377   //   return getClobber.
378   // If it is simple, we know based on the results of
379   // "Compiler testing via a theory of sound optimisations in the C11/C++11
380   //   memory model" in PLDI 2013, that a non-atomic location can only be
381   //   clobbered between a pair of a release and an acquire action, with no
382   //   access to the location in between.
383   // Here is an example for giving the general intuition behind this rule.
384   // In the following code:
385   //   store x 0;
386   //   release action; [1]
387   //   acquire action; [4]
388   //   %val = load x;
389   // It is unsafe to replace %val by 0 because another thread may be running:
390   //   acquire action; [2]
391   //   store x 42;
392   //   release action; [3]
393   // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
394   // being 42. A key property of this program however is that if either
395   // 1 or 4 were missing, there would be a race between the store of 42
396   // either the store of 0 or the load (making the whole program racy).
397   // The paper mentioned above shows that the same property is respected
398   // by every program that can detect any optimization of that kind: either
399   // it is racy (undefined) or there is a release followed by an acquire
400   // between the pair of accesses under consideration.
401 
402   // If the load is invariant, we "know" that it doesn't alias *any* write. We
403   // do want to respect mustalias results since defs are useful for value
404   // forwarding, but any mayalias write can be assumed to be noalias.
405   // Arguably, this logic should be pushed inside AliasAnalysis itself.
406   if (isLoad && QueryInst) {
407     LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
408     if (LI && LI->hasMetadata(LLVMContext::MD_invariant_load))
409       isInvariantLoad = true;
410   }
411 
412   // True for volatile instruction.
413   // For Load/Store return true if atomic ordering is stronger than AO,
414   // for other instruction just true if it can read or write to memory.
415   auto isComplexForReordering = [](Instruction * I, AtomicOrdering AO)->bool {
416     if (I->isVolatile())
417       return true;
418     if (auto *LI = dyn_cast<LoadInst>(I))
419       return isStrongerThan(LI->getOrdering(), AO);
420     if (auto *SI = dyn_cast<StoreInst>(I))
421       return isStrongerThan(SI->getOrdering(), AO);
422     return I->mayReadOrWriteMemory();
423   };
424 
425   // Walk backwards through the basic block, looking for dependencies.
426   while (ScanIt != BB->begin()) {
427     Instruction *Inst = &*--ScanIt;
428 
429     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
430       // Debug intrinsics don't (and can't) cause dependencies.
431       if (isa<DbgInfoIntrinsic>(II))
432         continue;
433 
434     // Limit the amount of scanning we do so we don't end up with quadratic
435     // running time on extreme testcases.
436     --*Limit;
437     if (!*Limit)
438       return MemDepResult::getUnknown();
439 
440     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
441       // If we reach a lifetime begin or end marker, then the query ends here
442       // because the value is undefined.
443       Intrinsic::ID ID = II->getIntrinsicID();
444       switch (ID) {
445       case Intrinsic::lifetime_start: {
446         // FIXME: This only considers queries directly on the invariant-tagged
447         // pointer, not on query pointers that are indexed off of them.  It'd
448         // be nice to handle that at some point (the right approach is to use
449         // GetPointerBaseWithConstantOffset).
450         MemoryLocation ArgLoc = MemoryLocation::getAfter(II->getArgOperand(1));
451         if (BatchAA.isMustAlias(ArgLoc, MemLoc))
452           return MemDepResult::getDef(II);
453         continue;
454       }
455       case Intrinsic::masked_load:
456       case Intrinsic::masked_store: {
457         MemoryLocation Loc;
458         /*ModRefInfo MR =*/ GetLocation(II, Loc, TLI);
459         AliasResult R = BatchAA.alias(Loc, MemLoc);
460         if (R == AliasResult::NoAlias)
461           continue;
462         if (R == AliasResult::MustAlias)
463           return MemDepResult::getDef(II);
464         if (ID == Intrinsic::masked_load)
465           continue;
466         return MemDepResult::getClobber(II);
467       }
468       }
469     }
470 
471     // Values depend on loads if the pointers are must aliased.  This means
472     // that a load depends on another must aliased load from the same value.
473     // One exception is atomic loads: a value can depend on an atomic load that
474     // it does not alias with when this atomic load indicates that another
475     // thread may be accessing the location.
476     if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
477       // While volatile access cannot be eliminated, they do not have to clobber
478       // non-aliasing locations, as normal accesses, for example, can be safely
479       // reordered with volatile accesses.
480       if (LI->isVolatile()) {
481         if (!QueryInst)
482           // Original QueryInst *may* be volatile
483           return MemDepResult::getClobber(LI);
484         if (QueryInst->isVolatile())
485           // Ordering required if QueryInst is itself volatile
486           return MemDepResult::getClobber(LI);
487         // Otherwise, volatile doesn't imply any special ordering
488       }
489 
490       // Atomic loads have complications involved.
491       // A Monotonic (or higher) load is OK if the query inst is itself not
492       // atomic.
493       // FIXME: This is overly conservative.
494       if (LI->isAtomic() && isStrongerThanUnordered(LI->getOrdering())) {
495         if (!QueryInst ||
496             isComplexForReordering(QueryInst, AtomicOrdering::NotAtomic))
497           return MemDepResult::getClobber(LI);
498         if (LI->getOrdering() != AtomicOrdering::Monotonic)
499           return MemDepResult::getClobber(LI);
500       }
501 
502       MemoryLocation LoadLoc = MemoryLocation::get(LI);
503 
504       // If we found a pointer, check if it could be the same as our pointer.
505       AliasResult R = BatchAA.alias(LoadLoc, MemLoc);
506 
507       if (R == AliasResult::NoAlias)
508         continue;
509 
510       if (isLoad) {
511         // Must aliased loads are defs of each other.
512         if (R == AliasResult::MustAlias)
513           return MemDepResult::getDef(Inst);
514 
515         // If we have a partial alias, then return this as a clobber for the
516         // client to handle.
517         if (R == AliasResult::PartialAlias && R.hasOffset()) {
518           ClobberOffsets[LI] = R.getOffset();
519           return MemDepResult::getClobber(Inst);
520         }
521 
522         // Random may-alias loads don't depend on each other without a
523         // dependence.
524         continue;
525       }
526 
527       // Stores don't alias loads from read-only memory.
528       if (BatchAA.pointsToConstantMemory(LoadLoc))
529         continue;
530 
531       // Stores depend on may/must aliased loads.
532       return MemDepResult::getDef(Inst);
533     }
534 
535     if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
536       // Atomic stores have complications involved.
537       // A Monotonic store is OK if the query inst is itself not atomic.
538       // FIXME: This is overly conservative.
539       if (!SI->isUnordered() && SI->isAtomic()) {
540         if (!QueryInst ||
541             isComplexForReordering(QueryInst, AtomicOrdering::Unordered))
542           return MemDepResult::getClobber(SI);
543         // Ok, if we are here the guard above guarantee us that
544         // QueryInst is a non-atomic or unordered load/store.
545         // SI is atomic with monotonic or release semantic (seq_cst for store
546         // is actually a release semantic plus total order over other seq_cst
547         // instructions, as soon as QueryInst is not seq_cst we can consider it
548         // as simple release semantic).
549         // Monotonic and Release semantic allows re-ordering before store
550         // so we are safe to go further and check the aliasing. It will prohibit
551         // re-ordering in case locations are may or must alias.
552       }
553 
554       // While volatile access cannot be eliminated, they do not have to clobber
555       // non-aliasing locations, as normal accesses can for example be reordered
556       // with volatile accesses.
557       if (SI->isVolatile())
558         if (!QueryInst || QueryInst->isVolatile())
559           return MemDepResult::getClobber(SI);
560 
561       // If alias analysis can tell that this store is guaranteed to not modify
562       // the query pointer, ignore it.  Use getModRefInfo to handle cases where
563       // the query pointer points to constant memory etc.
564       if (!isModOrRefSet(BatchAA.getModRefInfo(SI, MemLoc)))
565         continue;
566 
567       // Ok, this store might clobber the query pointer.  Check to see if it is
568       // a must alias: in this case, we want to return this as a def.
569       // FIXME: Use ModRefInfo::Must bit from getModRefInfo call above.
570       MemoryLocation StoreLoc = MemoryLocation::get(SI);
571 
572       // If we found a pointer, check if it could be the same as our pointer.
573       AliasResult R = BatchAA.alias(StoreLoc, MemLoc);
574 
575       if (R == AliasResult::NoAlias)
576         continue;
577       if (R == AliasResult::MustAlias)
578         return MemDepResult::getDef(Inst);
579       if (isInvariantLoad)
580         continue;
581       return MemDepResult::getClobber(Inst);
582     }
583 
584     // If this is an allocation, and if we know that the accessed pointer is to
585     // the allocation, return Def.  This means that there is no dependence and
586     // the access can be optimized based on that.  For example, a load could
587     // turn into undef.  Note that we can bypass the allocation itself when
588     // looking for a clobber in many cases; that's an alias property and is
589     // handled by BasicAA.
590     if (isa<AllocaInst>(Inst) || isNoAliasCall(Inst)) {
591       const Value *AccessPtr = getUnderlyingObject(MemLoc.Ptr);
592       if (AccessPtr == Inst || BatchAA.isMustAlias(Inst, AccessPtr))
593         return MemDepResult::getDef(Inst);
594     }
595 
596     if (isInvariantLoad)
597       continue;
598 
599     // A release fence requires that all stores complete before it, but does
600     // not prevent the reordering of following loads or stores 'before' the
601     // fence.  As a result, we look past it when finding a dependency for
602     // loads.  DSE uses this to find preceding stores to delete and thus we
603     // can't bypass the fence if the query instruction is a store.
604     if (FenceInst *FI = dyn_cast<FenceInst>(Inst))
605       if (isLoad && FI->getOrdering() == AtomicOrdering::Release)
606         continue;
607 
608     // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
609     ModRefInfo MR = BatchAA.getModRefInfo(Inst, MemLoc);
610     // If necessary, perform additional analysis.
611     if (isModAndRefSet(MR))
612       MR = BatchAA.callCapturesBefore(Inst, MemLoc, &DT);
613     switch (clearMust(MR)) {
614     case ModRefInfo::NoModRef:
615       // If the call has no effect on the queried pointer, just ignore it.
616       continue;
617     case ModRefInfo::Mod:
618       return MemDepResult::getClobber(Inst);
619     case ModRefInfo::Ref:
620       // If the call is known to never store to the pointer, and if this is a
621       // load query, we can safely ignore it (scan past it).
622       if (isLoad)
623         continue;
624       LLVM_FALLTHROUGH;
625     default:
626       // Otherwise, there is a potential dependence.  Return a clobber.
627       return MemDepResult::getClobber(Inst);
628     }
629   }
630 
631   // No dependence found.  If this is the entry block of the function, it is
632   // unknown, otherwise it is non-local.
633   if (BB != &BB->getParent()->getEntryBlock())
634     return MemDepResult::getNonLocal();
635   return MemDepResult::getNonFuncLocal();
636 }
637 
638 MemDepResult MemoryDependenceResults::getDependency(Instruction *QueryInst) {
639   ClobberOffsets.clear();
640   Instruction *ScanPos = QueryInst;
641 
642   // Check for a cached result
643   MemDepResult &LocalCache = LocalDeps[QueryInst];
644 
645   // If the cached entry is non-dirty, just return it.  Note that this depends
646   // on MemDepResult's default constructing to 'dirty'.
647   if (!LocalCache.isDirty())
648     return LocalCache;
649 
650   // Otherwise, if we have a dirty entry, we know we can start the scan at that
651   // instruction, which may save us some work.
652   if (Instruction *Inst = LocalCache.getInst()) {
653     ScanPos = Inst;
654 
655     RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
656   }
657 
658   BasicBlock *QueryParent = QueryInst->getParent();
659 
660   // Do the scan.
661   if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
662     // No dependence found. If this is the entry block of the function, it is
663     // unknown, otherwise it is non-local.
664     if (QueryParent != &QueryParent->getParent()->getEntryBlock())
665       LocalCache = MemDepResult::getNonLocal();
666     else
667       LocalCache = MemDepResult::getNonFuncLocal();
668   } else {
669     MemoryLocation MemLoc;
670     ModRefInfo MR = GetLocation(QueryInst, MemLoc, TLI);
671     if (MemLoc.Ptr) {
672       // If we can do a pointer scan, make it happen.
673       bool isLoad = !isModSet(MR);
674       if (auto *II = dyn_cast<IntrinsicInst>(QueryInst))
675         isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
676 
677       LocalCache =
678           getPointerDependencyFrom(MemLoc, isLoad, ScanPos->getIterator(),
679                                    QueryParent, QueryInst, nullptr);
680     } else if (auto *QueryCall = dyn_cast<CallBase>(QueryInst)) {
681       bool isReadOnly = AA.onlyReadsMemory(QueryCall);
682       LocalCache = getCallDependencyFrom(QueryCall, isReadOnly,
683                                          ScanPos->getIterator(), QueryParent);
684     } else
685       // Non-memory instruction.
686       LocalCache = MemDepResult::getUnknown();
687   }
688 
689   // Remember the result!
690   if (Instruction *I = LocalCache.getInst())
691     ReverseLocalDeps[I].insert(QueryInst);
692 
693   return LocalCache;
694 }
695 
696 #ifndef NDEBUG
697 /// This method is used when -debug is specified to verify that cache arrays
698 /// are properly kept sorted.
699 static void AssertSorted(MemoryDependenceResults::NonLocalDepInfo &Cache,
700                          int Count = -1) {
701   if (Count == -1)
702     Count = Cache.size();
703   assert(std::is_sorted(Cache.begin(), Cache.begin() + Count) &&
704          "Cache isn't sorted!");
705 }
706 #endif
707 
708 const MemoryDependenceResults::NonLocalDepInfo &
709 MemoryDependenceResults::getNonLocalCallDependency(CallBase *QueryCall) {
710   assert(getDependency(QueryCall).isNonLocal() &&
711          "getNonLocalCallDependency should only be used on calls with "
712          "non-local deps!");
713   PerInstNLInfo &CacheP = NonLocalDepsMap[QueryCall];
714   NonLocalDepInfo &Cache = CacheP.first;
715 
716   // This is the set of blocks that need to be recomputed.  In the cached case,
717   // this can happen due to instructions being deleted etc. In the uncached
718   // case, this starts out as the set of predecessors we care about.
719   SmallVector<BasicBlock *, 32> DirtyBlocks;
720 
721   if (!Cache.empty()) {
722     // Okay, we have a cache entry.  If we know it is not dirty, just return it
723     // with no computation.
724     if (!CacheP.second) {
725       ++NumCacheNonLocal;
726       return Cache;
727     }
728 
729     // If we already have a partially computed set of results, scan them to
730     // determine what is dirty, seeding our initial DirtyBlocks worklist.
731     for (auto &Entry : Cache)
732       if (Entry.getResult().isDirty())
733         DirtyBlocks.push_back(Entry.getBB());
734 
735     // Sort the cache so that we can do fast binary search lookups below.
736     llvm::sort(Cache);
737 
738     ++NumCacheDirtyNonLocal;
739   } else {
740     // Seed DirtyBlocks with each of the preds of QueryInst's block.
741     BasicBlock *QueryBB = QueryCall->getParent();
742     append_range(DirtyBlocks, PredCache.get(QueryBB));
743     ++NumUncacheNonLocal;
744   }
745 
746   // isReadonlyCall - If this is a read-only call, we can be more aggressive.
747   bool isReadonlyCall = AA.onlyReadsMemory(QueryCall);
748 
749   SmallPtrSet<BasicBlock *, 32> Visited;
750 
751   unsigned NumSortedEntries = Cache.size();
752   LLVM_DEBUG(AssertSorted(Cache));
753 
754   // Iterate while we still have blocks to update.
755   while (!DirtyBlocks.empty()) {
756     BasicBlock *DirtyBB = DirtyBlocks.pop_back_val();
757 
758     // Already processed this block?
759     if (!Visited.insert(DirtyBB).second)
760       continue;
761 
762     // Do a binary search to see if we already have an entry for this block in
763     // the cache set.  If so, find it.
764     LLVM_DEBUG(AssertSorted(Cache, NumSortedEntries));
765     NonLocalDepInfo::iterator Entry =
766         std::upper_bound(Cache.begin(), Cache.begin() + NumSortedEntries,
767                          NonLocalDepEntry(DirtyBB));
768     if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
769       --Entry;
770 
771     NonLocalDepEntry *ExistingResult = nullptr;
772     if (Entry != Cache.begin() + NumSortedEntries &&
773         Entry->getBB() == DirtyBB) {
774       // If we already have an entry, and if it isn't already dirty, the block
775       // is done.
776       if (!Entry->getResult().isDirty())
777         continue;
778 
779       // Otherwise, remember this slot so we can update the value.
780       ExistingResult = &*Entry;
781     }
782 
783     // If the dirty entry has a pointer, start scanning from it so we don't have
784     // to rescan the entire block.
785     BasicBlock::iterator ScanPos = DirtyBB->end();
786     if (ExistingResult) {
787       if (Instruction *Inst = ExistingResult->getResult().getInst()) {
788         ScanPos = Inst->getIterator();
789         // We're removing QueryInst's use of Inst.
790         RemoveFromReverseMap<Instruction *>(ReverseNonLocalDeps, Inst,
791                                             QueryCall);
792       }
793     }
794 
795     // Find out if this block has a local dependency for QueryInst.
796     MemDepResult Dep;
797 
798     if (ScanPos != DirtyBB->begin()) {
799       Dep = getCallDependencyFrom(QueryCall, isReadonlyCall, ScanPos, DirtyBB);
800     } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
801       // No dependence found.  If this is the entry block of the function, it is
802       // a clobber, otherwise it is unknown.
803       Dep = MemDepResult::getNonLocal();
804     } else {
805       Dep = MemDepResult::getNonFuncLocal();
806     }
807 
808     // If we had a dirty entry for the block, update it.  Otherwise, just add
809     // a new entry.
810     if (ExistingResult)
811       ExistingResult->setResult(Dep);
812     else
813       Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
814 
815     // If the block has a dependency (i.e. it isn't completely transparent to
816     // the value), remember the association!
817     if (!Dep.isNonLocal()) {
818       // Keep the ReverseNonLocalDeps map up to date so we can efficiently
819       // update this when we remove instructions.
820       if (Instruction *Inst = Dep.getInst())
821         ReverseNonLocalDeps[Inst].insert(QueryCall);
822     } else {
823 
824       // If the block *is* completely transparent to the load, we need to check
825       // the predecessors of this block.  Add them to our worklist.
826       append_range(DirtyBlocks, PredCache.get(DirtyBB));
827     }
828   }
829 
830   return Cache;
831 }
832 
833 void MemoryDependenceResults::getNonLocalPointerDependency(
834     Instruction *QueryInst, SmallVectorImpl<NonLocalDepResult> &Result) {
835   const MemoryLocation Loc = MemoryLocation::get(QueryInst);
836   bool isLoad = isa<LoadInst>(QueryInst);
837   BasicBlock *FromBB = QueryInst->getParent();
838   assert(FromBB);
839 
840   assert(Loc.Ptr->getType()->isPointerTy() &&
841          "Can't get pointer deps of a non-pointer!");
842   Result.clear();
843   {
844     // Check if there is cached Def with invariant.group.
845     auto NonLocalDefIt = NonLocalDefsCache.find(QueryInst);
846     if (NonLocalDefIt != NonLocalDefsCache.end()) {
847       Result.push_back(NonLocalDefIt->second);
848       ReverseNonLocalDefsCache[NonLocalDefIt->second.getResult().getInst()]
849           .erase(QueryInst);
850       NonLocalDefsCache.erase(NonLocalDefIt);
851       return;
852     }
853   }
854   // This routine does not expect to deal with volatile instructions.
855   // Doing so would require piping through the QueryInst all the way through.
856   // TODO: volatiles can't be elided, but they can be reordered with other
857   // non-volatile accesses.
858 
859   // We currently give up on any instruction which is ordered, but we do handle
860   // atomic instructions which are unordered.
861   // TODO: Handle ordered instructions
862   auto isOrdered = [](Instruction *Inst) {
863     if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
864       return !LI->isUnordered();
865     } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
866       return !SI->isUnordered();
867     }
868     return false;
869   };
870   if (QueryInst->isVolatile() || isOrdered(QueryInst)) {
871     Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(),
872                                        const_cast<Value *>(Loc.Ptr)));
873     return;
874   }
875   const DataLayout &DL = FromBB->getModule()->getDataLayout();
876   PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, &AC);
877 
878   // This is the set of blocks we've inspected, and the pointer we consider in
879   // each block.  Because of critical edges, we currently bail out if querying
880   // a block with multiple different pointers.  This can happen during PHI
881   // translation.
882   DenseMap<BasicBlock *, Value *> Visited;
883   if (getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
884                                    Result, Visited, true))
885     return;
886   Result.clear();
887   Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(),
888                                      const_cast<Value *>(Loc.Ptr)));
889 }
890 
891 /// Compute the memdep value for BB with Pointer/PointeeSize using either
892 /// cached information in Cache or by doing a lookup (which may use dirty cache
893 /// info if available).
894 ///
895 /// If we do a lookup, add the result to the cache.
896 MemDepResult MemoryDependenceResults::getNonLocalInfoForBlock(
897     Instruction *QueryInst, const MemoryLocation &Loc, bool isLoad,
898     BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries,
899     BatchAAResults &BatchAA) {
900 
901   bool isInvariantLoad = false;
902 
903   if (LoadInst *LI = dyn_cast_or_null<LoadInst>(QueryInst))
904     isInvariantLoad = LI->getMetadata(LLVMContext::MD_invariant_load);
905 
906   // Do a binary search to see if we already have an entry for this block in
907   // the cache set.  If so, find it.
908   NonLocalDepInfo::iterator Entry = std::upper_bound(
909       Cache->begin(), Cache->begin() + NumSortedEntries, NonLocalDepEntry(BB));
910   if (Entry != Cache->begin() && (Entry - 1)->getBB() == BB)
911     --Entry;
912 
913   NonLocalDepEntry *ExistingResult = nullptr;
914   if (Entry != Cache->begin() + NumSortedEntries && Entry->getBB() == BB)
915     ExistingResult = &*Entry;
916 
917   // Use cached result for invariant load only if there is no dependency for non
918   // invariant load. In this case invariant load can not have any dependency as
919   // well.
920   if (ExistingResult && isInvariantLoad &&
921       !ExistingResult->getResult().isNonFuncLocal())
922     ExistingResult = nullptr;
923 
924   // If we have a cached entry, and it is non-dirty, use it as the value for
925   // this dependency.
926   if (ExistingResult && !ExistingResult->getResult().isDirty()) {
927     ++NumCacheNonLocalPtr;
928     return ExistingResult->getResult();
929   }
930 
931   // Otherwise, we have to scan for the value.  If we have a dirty cache
932   // entry, start scanning from its position, otherwise we scan from the end
933   // of the block.
934   BasicBlock::iterator ScanPos = BB->end();
935   if (ExistingResult && ExistingResult->getResult().getInst()) {
936     assert(ExistingResult->getResult().getInst()->getParent() == BB &&
937            "Instruction invalidated?");
938     ++NumCacheDirtyNonLocalPtr;
939     ScanPos = ExistingResult->getResult().getInst()->getIterator();
940 
941     // Eliminating the dirty entry from 'Cache', so update the reverse info.
942     ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
943     RemoveFromReverseMap(ReverseNonLocalPtrDeps, &*ScanPos, CacheKey);
944   } else {
945     ++NumUncacheNonLocalPtr;
946   }
947 
948   // Scan the block for the dependency.
949   MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB,
950                                               QueryInst, nullptr, BatchAA);
951 
952   // Don't cache results for invariant load.
953   if (isInvariantLoad)
954     return Dep;
955 
956   // If we had a dirty entry for the block, update it.  Otherwise, just add
957   // a new entry.
958   if (ExistingResult)
959     ExistingResult->setResult(Dep);
960   else
961     Cache->push_back(NonLocalDepEntry(BB, Dep));
962 
963   // If the block has a dependency (i.e. it isn't completely transparent to
964   // the value), remember the reverse association because we just added it
965   // to Cache!
966   if (!Dep.isDef() && !Dep.isClobber())
967     return Dep;
968 
969   // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
970   // update MemDep when we remove instructions.
971   Instruction *Inst = Dep.getInst();
972   assert(Inst && "Didn't depend on anything?");
973   ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
974   ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
975   return Dep;
976 }
977 
978 /// Sort the NonLocalDepInfo cache, given a certain number of elements in the
979 /// array that are already properly ordered.
980 ///
981 /// This is optimized for the case when only a few entries are added.
982 static void
983 SortNonLocalDepInfoCache(MemoryDependenceResults::NonLocalDepInfo &Cache,
984                          unsigned NumSortedEntries) {
985   switch (Cache.size() - NumSortedEntries) {
986   case 0:
987     // done, no new entries.
988     break;
989   case 2: {
990     // Two new entries, insert the last one into place.
991     NonLocalDepEntry Val = Cache.back();
992     Cache.pop_back();
993     MemoryDependenceResults::NonLocalDepInfo::iterator Entry =
994         std::upper_bound(Cache.begin(), Cache.end() - 1, Val);
995     Cache.insert(Entry, Val);
996     LLVM_FALLTHROUGH;
997   }
998   case 1:
999     // One new entry, Just insert the new value at the appropriate position.
1000     if (Cache.size() != 1) {
1001       NonLocalDepEntry Val = Cache.back();
1002       Cache.pop_back();
1003       MemoryDependenceResults::NonLocalDepInfo::iterator Entry =
1004           llvm::upper_bound(Cache, Val);
1005       Cache.insert(Entry, Val);
1006     }
1007     break;
1008   default:
1009     // Added many values, do a full scale sort.
1010     llvm::sort(Cache);
1011     break;
1012   }
1013 }
1014 
1015 /// Perform a dependency query based on pointer/pointeesize starting at the end
1016 /// of StartBB.
1017 ///
1018 /// Add any clobber/def results to the results vector and keep track of which
1019 /// blocks are visited in 'Visited'.
1020 ///
1021 /// This has special behavior for the first block queries (when SkipFirstBlock
1022 /// is true).  In this special case, it ignores the contents of the specified
1023 /// block and starts returning dependence info for its predecessors.
1024 ///
1025 /// This function returns true on success, or false to indicate that it could
1026 /// not compute dependence information for some reason.  This should be treated
1027 /// as a clobber dependence on the first instruction in the predecessor block.
1028 bool MemoryDependenceResults::getNonLocalPointerDepFromBB(
1029     Instruction *QueryInst, const PHITransAddr &Pointer,
1030     const MemoryLocation &Loc, bool isLoad, BasicBlock *StartBB,
1031     SmallVectorImpl<NonLocalDepResult> &Result,
1032     DenseMap<BasicBlock *, Value *> &Visited, bool SkipFirstBlock,
1033     bool IsIncomplete) {
1034   // Look up the cached info for Pointer.
1035   ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1036 
1037   // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1038   // CacheKey, this value will be inserted as the associated value. Otherwise,
1039   // it'll be ignored, and we'll have to check to see if the cached size and
1040   // aa tags are consistent with the current query.
1041   NonLocalPointerInfo InitialNLPI;
1042   InitialNLPI.Size = Loc.Size;
1043   InitialNLPI.AATags = Loc.AATags;
1044 
1045   bool isInvariantLoad = false;
1046   if (LoadInst *LI = dyn_cast_or_null<LoadInst>(QueryInst))
1047     isInvariantLoad = LI->getMetadata(LLVMContext::MD_invariant_load);
1048 
1049   // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1050   // already have one.
1051   std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1052       NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1053   NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1054 
1055   // If we already have a cache entry for this CacheKey, we may need to do some
1056   // work to reconcile the cache entry and the current query.
1057   // Invariant loads don't participate in caching. Thus no need to reconcile.
1058   if (!isInvariantLoad && !Pair.second) {
1059     if (CacheInfo->Size != Loc.Size) {
1060       bool ThrowOutEverything;
1061       if (CacheInfo->Size.hasValue() && Loc.Size.hasValue()) {
1062         // FIXME: We may be able to do better in the face of results with mixed
1063         // precision. We don't appear to get them in practice, though, so just
1064         // be conservative.
1065         ThrowOutEverything =
1066             CacheInfo->Size.isPrecise() != Loc.Size.isPrecise() ||
1067             CacheInfo->Size.getValue() < Loc.Size.getValue();
1068       } else {
1069         // For our purposes, unknown size > all others.
1070         ThrowOutEverything = !Loc.Size.hasValue();
1071       }
1072 
1073       if (ThrowOutEverything) {
1074         // The query's Size is greater than the cached one. Throw out the
1075         // cached data and proceed with the query at the greater size.
1076         CacheInfo->Pair = BBSkipFirstBlockPair();
1077         CacheInfo->Size = Loc.Size;
1078         for (auto &Entry : CacheInfo->NonLocalDeps)
1079           if (Instruction *Inst = Entry.getResult().getInst())
1080             RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1081         CacheInfo->NonLocalDeps.clear();
1082         // The cache is cleared (in the above line) so we will have lost
1083         // information about blocks we have already visited. We therefore must
1084         // assume that the cache information is incomplete.
1085         IsIncomplete = true;
1086       } else {
1087         // This query's Size is less than the cached one. Conservatively restart
1088         // the query using the greater size.
1089         return getNonLocalPointerDepFromBB(
1090             QueryInst, Pointer, Loc.getWithNewSize(CacheInfo->Size), isLoad,
1091             StartBB, Result, Visited, SkipFirstBlock, IsIncomplete);
1092       }
1093     }
1094 
1095     // If the query's AATags are inconsistent with the cached one,
1096     // conservatively throw out the cached data and restart the query with
1097     // no tag if needed.
1098     if (CacheInfo->AATags != Loc.AATags) {
1099       if (CacheInfo->AATags) {
1100         CacheInfo->Pair = BBSkipFirstBlockPair();
1101         CacheInfo->AATags = AAMDNodes();
1102         for (auto &Entry : CacheInfo->NonLocalDeps)
1103           if (Instruction *Inst = Entry.getResult().getInst())
1104             RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1105         CacheInfo->NonLocalDeps.clear();
1106         // The cache is cleared (in the above line) so we will have lost
1107         // information about blocks we have already visited. We therefore must
1108         // assume that the cache information is incomplete.
1109         IsIncomplete = true;
1110       }
1111       if (Loc.AATags)
1112         return getNonLocalPointerDepFromBB(
1113             QueryInst, Pointer, Loc.getWithoutAATags(), isLoad, StartBB, Result,
1114             Visited, SkipFirstBlock, IsIncomplete);
1115     }
1116   }
1117 
1118   NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1119 
1120   // If we have valid cached information for exactly the block we are
1121   // investigating, just return it with no recomputation.
1122   // Don't use cached information for invariant loads since it is valid for
1123   // non-invariant loads only.
1124   if (!IsIncomplete && !isInvariantLoad &&
1125       CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1126     // We have a fully cached result for this query then we can just return the
1127     // cached results and populate the visited set.  However, we have to verify
1128     // that we don't already have conflicting results for these blocks.  Check
1129     // to ensure that if a block in the results set is in the visited set that
1130     // it was for the same pointer query.
1131     if (!Visited.empty()) {
1132       for (auto &Entry : *Cache) {
1133         DenseMap<BasicBlock *, Value *>::iterator VI =
1134             Visited.find(Entry.getBB());
1135         if (VI == Visited.end() || VI->second == Pointer.getAddr())
1136           continue;
1137 
1138         // We have a pointer mismatch in a block.  Just return false, saying
1139         // that something was clobbered in this result.  We could also do a
1140         // non-fully cached query, but there is little point in doing this.
1141         return false;
1142       }
1143     }
1144 
1145     Value *Addr = Pointer.getAddr();
1146     for (auto &Entry : *Cache) {
1147       Visited.insert(std::make_pair(Entry.getBB(), Addr));
1148       if (Entry.getResult().isNonLocal()) {
1149         continue;
1150       }
1151 
1152       if (DT.isReachableFromEntry(Entry.getBB())) {
1153         Result.push_back(
1154             NonLocalDepResult(Entry.getBB(), Entry.getResult(), Addr));
1155       }
1156     }
1157     ++NumCacheCompleteNonLocalPtr;
1158     return true;
1159   }
1160 
1161   // Otherwise, either this is a new block, a block with an invalid cache
1162   // pointer or one that we're about to invalidate by putting more info into
1163   // it than its valid cache info.  If empty and not explicitly indicated as
1164   // incomplete, the result will be valid cache info, otherwise it isn't.
1165   //
1166   // Invariant loads don't affect cache in any way thus no need to update
1167   // CacheInfo as well.
1168   if (!isInvariantLoad) {
1169     if (!IsIncomplete && Cache->empty())
1170       CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1171     else
1172       CacheInfo->Pair = BBSkipFirstBlockPair();
1173   }
1174 
1175   SmallVector<BasicBlock *, 32> Worklist;
1176   Worklist.push_back(StartBB);
1177 
1178   // PredList used inside loop.
1179   SmallVector<std::pair<BasicBlock *, PHITransAddr>, 16> PredList;
1180 
1181   // Keep track of the entries that we know are sorted.  Previously cached
1182   // entries will all be sorted.  The entries we add we only sort on demand (we
1183   // don't insert every element into its sorted position).  We know that we
1184   // won't get any reuse from currently inserted values, because we don't
1185   // revisit blocks after we insert info for them.
1186   unsigned NumSortedEntries = Cache->size();
1187   unsigned WorklistEntries = BlockNumberLimit;
1188   bool GotWorklistLimit = false;
1189   LLVM_DEBUG(AssertSorted(*Cache));
1190 
1191   BatchAAResults BatchAA(AA);
1192   while (!Worklist.empty()) {
1193     BasicBlock *BB = Worklist.pop_back_val();
1194 
1195     // If we do process a large number of blocks it becomes very expensive and
1196     // likely it isn't worth worrying about
1197     if (Result.size() > NumResultsLimit) {
1198       // Sort it now (if needed) so that recursive invocations of
1199       // getNonLocalPointerDepFromBB and other routines that could reuse the
1200       // cache value will only see properly sorted cache arrays.
1201       if (Cache && NumSortedEntries != Cache->size()) {
1202         SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1203       }
1204       // Since we bail out, the "Cache" set won't contain all of the
1205       // results for the query.  This is ok (we can still use it to accelerate
1206       // specific block queries) but we can't do the fastpath "return all
1207       // results from the set".  Clear out the indicator for this.
1208       CacheInfo->Pair = BBSkipFirstBlockPair();
1209       return false;
1210     }
1211 
1212     // Skip the first block if we have it.
1213     if (!SkipFirstBlock) {
1214       // Analyze the dependency of *Pointer in FromBB.  See if we already have
1215       // been here.
1216       assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1217 
1218       // Get the dependency info for Pointer in BB.  If we have cached
1219       // information, we will use it, otherwise we compute it.
1220       LLVM_DEBUG(AssertSorted(*Cache, NumSortedEntries));
1221       MemDepResult Dep = getNonLocalInfoForBlock(
1222           QueryInst, Loc, isLoad, BB, Cache, NumSortedEntries, BatchAA);
1223 
1224       // If we got a Def or Clobber, add this to the list of results.
1225       if (!Dep.isNonLocal()) {
1226         if (DT.isReachableFromEntry(BB)) {
1227           Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1228           continue;
1229         }
1230       }
1231     }
1232 
1233     // If 'Pointer' is an instruction defined in this block, then we need to do
1234     // phi translation to change it into a value live in the predecessor block.
1235     // If not, we just add the predecessors to the worklist and scan them with
1236     // the same Pointer.
1237     if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1238       SkipFirstBlock = false;
1239       SmallVector<BasicBlock *, 16> NewBlocks;
1240       for (BasicBlock *Pred : PredCache.get(BB)) {
1241         // Verify that we haven't looked at this block yet.
1242         std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes =
1243             Visited.insert(std::make_pair(Pred, Pointer.getAddr()));
1244         if (InsertRes.second) {
1245           // First time we've looked at *PI.
1246           NewBlocks.push_back(Pred);
1247           continue;
1248         }
1249 
1250         // If we have seen this block before, but it was with a different
1251         // pointer then we have a phi translation failure and we have to treat
1252         // this as a clobber.
1253         if (InsertRes.first->second != Pointer.getAddr()) {
1254           // Make sure to clean up the Visited map before continuing on to
1255           // PredTranslationFailure.
1256           for (unsigned i = 0; i < NewBlocks.size(); i++)
1257             Visited.erase(NewBlocks[i]);
1258           goto PredTranslationFailure;
1259         }
1260       }
1261       if (NewBlocks.size() > WorklistEntries) {
1262         // Make sure to clean up the Visited map before continuing on to
1263         // PredTranslationFailure.
1264         for (unsigned i = 0; i < NewBlocks.size(); i++)
1265           Visited.erase(NewBlocks[i]);
1266         GotWorklistLimit = true;
1267         goto PredTranslationFailure;
1268       }
1269       WorklistEntries -= NewBlocks.size();
1270       Worklist.append(NewBlocks.begin(), NewBlocks.end());
1271       continue;
1272     }
1273 
1274     // We do need to do phi translation, if we know ahead of time we can't phi
1275     // translate this value, don't even try.
1276     if (!Pointer.IsPotentiallyPHITranslatable())
1277       goto PredTranslationFailure;
1278 
1279     // We may have added values to the cache list before this PHI translation.
1280     // If so, we haven't done anything to ensure that the cache remains sorted.
1281     // Sort it now (if needed) so that recursive invocations of
1282     // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1283     // value will only see properly sorted cache arrays.
1284     if (Cache && NumSortedEntries != Cache->size()) {
1285       SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1286       NumSortedEntries = Cache->size();
1287     }
1288     Cache = nullptr;
1289 
1290     PredList.clear();
1291     for (BasicBlock *Pred : PredCache.get(BB)) {
1292       PredList.push_back(std::make_pair(Pred, Pointer));
1293 
1294       // Get the PHI translated pointer in this predecessor.  This can fail if
1295       // not translatable, in which case the getAddr() returns null.
1296       PHITransAddr &PredPointer = PredList.back().second;
1297       PredPointer.PHITranslateValue(BB, Pred, &DT, /*MustDominate=*/false);
1298       Value *PredPtrVal = PredPointer.getAddr();
1299 
1300       // Check to see if we have already visited this pred block with another
1301       // pointer.  If so, we can't do this lookup.  This failure can occur
1302       // with PHI translation when a critical edge exists and the PHI node in
1303       // the successor translates to a pointer value different than the
1304       // pointer the block was first analyzed with.
1305       std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes =
1306           Visited.insert(std::make_pair(Pred, PredPtrVal));
1307 
1308       if (!InsertRes.second) {
1309         // We found the pred; take it off the list of preds to visit.
1310         PredList.pop_back();
1311 
1312         // If the predecessor was visited with PredPtr, then we already did
1313         // the analysis and can ignore it.
1314         if (InsertRes.first->second == PredPtrVal)
1315           continue;
1316 
1317         // Otherwise, the block was previously analyzed with a different
1318         // pointer.  We can't represent the result of this case, so we just
1319         // treat this as a phi translation failure.
1320 
1321         // Make sure to clean up the Visited map before continuing on to
1322         // PredTranslationFailure.
1323         for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1324           Visited.erase(PredList[i].first);
1325 
1326         goto PredTranslationFailure;
1327       }
1328     }
1329 
1330     // Actually process results here; this need to be a separate loop to avoid
1331     // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1332     // any results for.  (getNonLocalPointerDepFromBB will modify our
1333     // datastructures in ways the code after the PredTranslationFailure label
1334     // doesn't expect.)
1335     for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1336       BasicBlock *Pred = PredList[i].first;
1337       PHITransAddr &PredPointer = PredList[i].second;
1338       Value *PredPtrVal = PredPointer.getAddr();
1339 
1340       bool CanTranslate = true;
1341       // If PHI translation was unable to find an available pointer in this
1342       // predecessor, then we have to assume that the pointer is clobbered in
1343       // that predecessor.  We can still do PRE of the load, which would insert
1344       // a computation of the pointer in this predecessor.
1345       if (!PredPtrVal)
1346         CanTranslate = false;
1347 
1348       // FIXME: it is entirely possible that PHI translating will end up with
1349       // the same value.  Consider PHI translating something like:
1350       // X = phi [x, bb1], [y, bb2].  PHI translating for bb1 doesn't *need*
1351       // to recurse here, pedantically speaking.
1352 
1353       // If getNonLocalPointerDepFromBB fails here, that means the cached
1354       // result conflicted with the Visited list; we have to conservatively
1355       // assume it is unknown, but this also does not block PRE of the load.
1356       if (!CanTranslate ||
1357           !getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1358                                       Loc.getWithNewPtr(PredPtrVal), isLoad,
1359                                       Pred, Result, Visited)) {
1360         // Add the entry to the Result list.
1361         NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1362         Result.push_back(Entry);
1363 
1364         // Since we had a phi translation failure, the cache for CacheKey won't
1365         // include all of the entries that we need to immediately satisfy future
1366         // queries.  Mark this in NonLocalPointerDeps by setting the
1367         // BBSkipFirstBlockPair pointer to null.  This requires reuse of the
1368         // cached value to do more work but not miss the phi trans failure.
1369         NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1370         NLPI.Pair = BBSkipFirstBlockPair();
1371         continue;
1372       }
1373     }
1374 
1375     // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1376     CacheInfo = &NonLocalPointerDeps[CacheKey];
1377     Cache = &CacheInfo->NonLocalDeps;
1378     NumSortedEntries = Cache->size();
1379 
1380     // Since we did phi translation, the "Cache" set won't contain all of the
1381     // results for the query.  This is ok (we can still use it to accelerate
1382     // specific block queries) but we can't do the fastpath "return all
1383     // results from the set"  Clear out the indicator for this.
1384     CacheInfo->Pair = BBSkipFirstBlockPair();
1385     SkipFirstBlock = false;
1386     continue;
1387 
1388   PredTranslationFailure:
1389     // The following code is "failure"; we can't produce a sane translation
1390     // for the given block.  It assumes that we haven't modified any of
1391     // our datastructures while processing the current block.
1392 
1393     if (!Cache) {
1394       // Refresh the CacheInfo/Cache pointer if it got invalidated.
1395       CacheInfo = &NonLocalPointerDeps[CacheKey];
1396       Cache = &CacheInfo->NonLocalDeps;
1397       NumSortedEntries = Cache->size();
1398     }
1399 
1400     // Since we failed phi translation, the "Cache" set won't contain all of the
1401     // results for the query.  This is ok (we can still use it to accelerate
1402     // specific block queries) but we can't do the fastpath "return all
1403     // results from the set".  Clear out the indicator for this.
1404     CacheInfo->Pair = BBSkipFirstBlockPair();
1405 
1406     // If *nothing* works, mark the pointer as unknown.
1407     //
1408     // If this is the magic first block, return this as a clobber of the whole
1409     // incoming value.  Since we can't phi translate to one of the predecessors,
1410     // we have to bail out.
1411     if (SkipFirstBlock)
1412       return false;
1413 
1414     // Results of invariant loads are not cached thus no need to update cached
1415     // information.
1416     if (!isInvariantLoad) {
1417       for (NonLocalDepEntry &I : llvm::reverse(*Cache)) {
1418         if (I.getBB() != BB)
1419           continue;
1420 
1421         assert((GotWorklistLimit || I.getResult().isNonLocal() ||
1422                 !DT.isReachableFromEntry(BB)) &&
1423                "Should only be here with transparent block");
1424 
1425         I.setResult(MemDepResult::getUnknown());
1426 
1427 
1428         break;
1429       }
1430     }
1431     (void)GotWorklistLimit;
1432     // Go ahead and report unknown dependence.
1433     Result.push_back(
1434         NonLocalDepResult(BB, MemDepResult::getUnknown(), Pointer.getAddr()));
1435   }
1436 
1437   // Okay, we're done now.  If we added new values to the cache, re-sort it.
1438   SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1439   LLVM_DEBUG(AssertSorted(*Cache));
1440   return true;
1441 }
1442 
1443 /// If P exists in CachedNonLocalPointerInfo or NonLocalDefsCache, remove it.
1444 void MemoryDependenceResults::removeCachedNonLocalPointerDependencies(
1445     ValueIsLoadPair P) {
1446 
1447   // Most of the time this cache is empty.
1448   if (!NonLocalDefsCache.empty()) {
1449     auto it = NonLocalDefsCache.find(P.getPointer());
1450     if (it != NonLocalDefsCache.end()) {
1451       RemoveFromReverseMap(ReverseNonLocalDefsCache,
1452                            it->second.getResult().getInst(), P.getPointer());
1453       NonLocalDefsCache.erase(it);
1454     }
1455 
1456     if (auto *I = dyn_cast<Instruction>(P.getPointer())) {
1457       auto toRemoveIt = ReverseNonLocalDefsCache.find(I);
1458       if (toRemoveIt != ReverseNonLocalDefsCache.end()) {
1459         for (const auto *entry : toRemoveIt->second)
1460           NonLocalDefsCache.erase(entry);
1461         ReverseNonLocalDefsCache.erase(toRemoveIt);
1462       }
1463     }
1464   }
1465 
1466   CachedNonLocalPointerInfo::iterator It = NonLocalPointerDeps.find(P);
1467   if (It == NonLocalPointerDeps.end())
1468     return;
1469 
1470   // Remove all of the entries in the BB->val map.  This involves removing
1471   // instructions from the reverse map.
1472   NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1473 
1474   for (const NonLocalDepEntry &DE : PInfo) {
1475     Instruction *Target = DE.getResult().getInst();
1476     if (!Target)
1477       continue; // Ignore non-local dep results.
1478     assert(Target->getParent() == DE.getBB());
1479 
1480     // Eliminating the dirty entry from 'Cache', so update the reverse info.
1481     RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1482   }
1483 
1484   // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1485   NonLocalPointerDeps.erase(It);
1486 }
1487 
1488 void MemoryDependenceResults::invalidateCachedPointerInfo(Value *Ptr) {
1489   // If Ptr isn't really a pointer, just ignore it.
1490   if (!Ptr->getType()->isPointerTy())
1491     return;
1492   // Flush store info for the pointer.
1493   removeCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1494   // Flush load info for the pointer.
1495   removeCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1496   // Invalidate phis that use the pointer.
1497   PV.invalidateValue(Ptr);
1498 }
1499 
1500 void MemoryDependenceResults::invalidateCachedPredecessors() {
1501   PredCache.clear();
1502 }
1503 
1504 void MemoryDependenceResults::removeInstruction(Instruction *RemInst) {
1505   // Walk through the Non-local dependencies, removing this one as the value
1506   // for any cached queries.
1507   NonLocalDepMapType::iterator NLDI = NonLocalDepsMap.find(RemInst);
1508   if (NLDI != NonLocalDepsMap.end()) {
1509     NonLocalDepInfo &BlockMap = NLDI->second.first;
1510     for (auto &Entry : BlockMap)
1511       if (Instruction *Inst = Entry.getResult().getInst())
1512         RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1513     NonLocalDepsMap.erase(NLDI);
1514   }
1515 
1516   // If we have a cached local dependence query for this instruction, remove it.
1517   LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1518   if (LocalDepEntry != LocalDeps.end()) {
1519     // Remove us from DepInst's reverse set now that the local dep info is gone.
1520     if (Instruction *Inst = LocalDepEntry->second.getInst())
1521       RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1522 
1523     // Remove this local dependency info.
1524     LocalDeps.erase(LocalDepEntry);
1525   }
1526 
1527   // If we have any cached dependencies on this instruction, remove
1528   // them.
1529 
1530   // If the instruction is a pointer, remove it from both the load info and the
1531   // store info.
1532   if (RemInst->getType()->isPointerTy()) {
1533     removeCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1534     removeCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1535   } else {
1536     // Otherwise, if the instructions is in the map directly, it must be a load.
1537     // Remove it.
1538     auto toRemoveIt = NonLocalDefsCache.find(RemInst);
1539     if (toRemoveIt != NonLocalDefsCache.end()) {
1540       assert(isa<LoadInst>(RemInst) &&
1541              "only load instructions should be added directly");
1542       const Instruction *DepV = toRemoveIt->second.getResult().getInst();
1543       ReverseNonLocalDefsCache.find(DepV)->second.erase(RemInst);
1544       NonLocalDefsCache.erase(toRemoveIt);
1545     }
1546   }
1547 
1548   // Loop over all of the things that depend on the instruction we're removing.
1549   SmallVector<std::pair<Instruction *, Instruction *>, 8> ReverseDepsToAdd;
1550 
1551   // If we find RemInst as a clobber or Def in any of the maps for other values,
1552   // we need to replace its entry with a dirty version of the instruction after
1553   // it.  If RemInst is a terminator, we use a null dirty value.
1554   //
1555   // Using a dirty version of the instruction after RemInst saves having to scan
1556   // the entire block to get to this point.
1557   MemDepResult NewDirtyVal;
1558   if (!RemInst->isTerminator())
1559     NewDirtyVal = MemDepResult::getDirty(&*++RemInst->getIterator());
1560 
1561   ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1562   if (ReverseDepIt != ReverseLocalDeps.end()) {
1563     // RemInst can't be the terminator if it has local stuff depending on it.
1564     assert(!ReverseDepIt->second.empty() && !RemInst->isTerminator() &&
1565            "Nothing can locally depend on a terminator");
1566 
1567     for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1568       assert(InstDependingOnRemInst != RemInst &&
1569              "Already removed our local dep info");
1570 
1571       LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1572 
1573       // Make sure to remember that new things depend on NewDepInst.
1574       assert(NewDirtyVal.getInst() &&
1575              "There is no way something else can have "
1576              "a local dep on this if it is a terminator!");
1577       ReverseDepsToAdd.push_back(
1578           std::make_pair(NewDirtyVal.getInst(), InstDependingOnRemInst));
1579     }
1580 
1581     ReverseLocalDeps.erase(ReverseDepIt);
1582 
1583     // Add new reverse deps after scanning the set, to avoid invalidating the
1584     // 'ReverseDeps' reference.
1585     while (!ReverseDepsToAdd.empty()) {
1586       ReverseLocalDeps[ReverseDepsToAdd.back().first].insert(
1587           ReverseDepsToAdd.back().second);
1588       ReverseDepsToAdd.pop_back();
1589     }
1590   }
1591 
1592   ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1593   if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1594     for (Instruction *I : ReverseDepIt->second) {
1595       assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1596 
1597       PerInstNLInfo &INLD = NonLocalDepsMap[I];
1598       // The information is now dirty!
1599       INLD.second = true;
1600 
1601       for (auto &Entry : INLD.first) {
1602         if (Entry.getResult().getInst() != RemInst)
1603           continue;
1604 
1605         // Convert to a dirty entry for the subsequent instruction.
1606         Entry.setResult(NewDirtyVal);
1607 
1608         if (Instruction *NextI = NewDirtyVal.getInst())
1609           ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1610       }
1611     }
1612 
1613     ReverseNonLocalDeps.erase(ReverseDepIt);
1614 
1615     // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1616     while (!ReverseDepsToAdd.empty()) {
1617       ReverseNonLocalDeps[ReverseDepsToAdd.back().first].insert(
1618           ReverseDepsToAdd.back().second);
1619       ReverseDepsToAdd.pop_back();
1620     }
1621   }
1622 
1623   // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1624   // value in the NonLocalPointerDeps info.
1625   ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1626       ReverseNonLocalPtrDeps.find(RemInst);
1627   if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1628     SmallVector<std::pair<Instruction *, ValueIsLoadPair>, 8>
1629         ReversePtrDepsToAdd;
1630 
1631     for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1632       assert(P.getPointer() != RemInst &&
1633              "Already removed NonLocalPointerDeps info for RemInst");
1634 
1635       NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1636 
1637       // The cache is not valid for any specific block anymore.
1638       NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1639 
1640       // Update any entries for RemInst to use the instruction after it.
1641       for (auto &Entry : NLPDI) {
1642         if (Entry.getResult().getInst() != RemInst)
1643           continue;
1644 
1645         // Convert to a dirty entry for the subsequent instruction.
1646         Entry.setResult(NewDirtyVal);
1647 
1648         if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1649           ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1650       }
1651 
1652       // Re-sort the NonLocalDepInfo.  Changing the dirty entry to its
1653       // subsequent value may invalidate the sortedness.
1654       llvm::sort(NLPDI);
1655     }
1656 
1657     ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1658 
1659     while (!ReversePtrDepsToAdd.empty()) {
1660       ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first].insert(
1661           ReversePtrDepsToAdd.back().second);
1662       ReversePtrDepsToAdd.pop_back();
1663     }
1664   }
1665 
1666   // Invalidate phis that use the removed instruction.
1667   PV.invalidateValue(RemInst);
1668 
1669   assert(!NonLocalDepsMap.count(RemInst) && "RemInst got reinserted?");
1670   LLVM_DEBUG(verifyRemoved(RemInst));
1671 }
1672 
1673 /// Verify that the specified instruction does not occur in our internal data
1674 /// structures.
1675 ///
1676 /// This function verifies by asserting in debug builds.
1677 void MemoryDependenceResults::verifyRemoved(Instruction *D) const {
1678 #ifndef NDEBUG
1679   for (const auto &DepKV : LocalDeps) {
1680     assert(DepKV.first != D && "Inst occurs in data structures");
1681     assert(DepKV.second.getInst() != D && "Inst occurs in data structures");
1682   }
1683 
1684   for (const auto &DepKV : NonLocalPointerDeps) {
1685     assert(DepKV.first.getPointer() != D && "Inst occurs in NLPD map key");
1686     for (const auto &Entry : DepKV.second.NonLocalDeps)
1687       assert(Entry.getResult().getInst() != D && "Inst occurs as NLPD value");
1688   }
1689 
1690   for (const auto &DepKV : NonLocalDepsMap) {
1691     assert(DepKV.first != D && "Inst occurs in data structures");
1692     const PerInstNLInfo &INLD = DepKV.second;
1693     for (const auto &Entry : INLD.first)
1694       assert(Entry.getResult().getInst() != D &&
1695              "Inst occurs in data structures");
1696   }
1697 
1698   for (const auto &DepKV : ReverseLocalDeps) {
1699     assert(DepKV.first != D && "Inst occurs in data structures");
1700     for (Instruction *Inst : DepKV.second)
1701       assert(Inst != D && "Inst occurs in data structures");
1702   }
1703 
1704   for (const auto &DepKV : ReverseNonLocalDeps) {
1705     assert(DepKV.first != D && "Inst occurs in data structures");
1706     for (Instruction *Inst : DepKV.second)
1707       assert(Inst != D && "Inst occurs in data structures");
1708   }
1709 
1710   for (const auto &DepKV : ReverseNonLocalPtrDeps) {
1711     assert(DepKV.first != D && "Inst occurs in rev NLPD map");
1712 
1713     for (ValueIsLoadPair P : DepKV.second)
1714       assert(P != ValueIsLoadPair(D, false) && P != ValueIsLoadPair(D, true) &&
1715              "Inst occurs in ReverseNonLocalPtrDeps map");
1716   }
1717 #endif
1718 }
1719 
1720 AnalysisKey MemoryDependenceAnalysis::Key;
1721 
1722 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
1723     : DefaultBlockScanLimit(BlockScanLimit) {}
1724 
1725 MemoryDependenceResults
1726 MemoryDependenceAnalysis::run(Function &F, FunctionAnalysisManager &AM) {
1727   auto &AA = AM.getResult<AAManager>(F);
1728   auto &AC = AM.getResult<AssumptionAnalysis>(F);
1729   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1730   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1731   auto &PV = AM.getResult<PhiValuesAnalysis>(F);
1732   return MemoryDependenceResults(AA, AC, TLI, DT, PV, DefaultBlockScanLimit);
1733 }
1734 
1735 char MemoryDependenceWrapperPass::ID = 0;
1736 
1737 INITIALIZE_PASS_BEGIN(MemoryDependenceWrapperPass, "memdep",
1738                       "Memory Dependence Analysis", false, true)
1739 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1740 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1741 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1742 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1743 INITIALIZE_PASS_DEPENDENCY(PhiValuesWrapperPass)
1744 INITIALIZE_PASS_END(MemoryDependenceWrapperPass, "memdep",
1745                     "Memory Dependence Analysis", false, true)
1746 
1747 MemoryDependenceWrapperPass::MemoryDependenceWrapperPass() : FunctionPass(ID) {
1748   initializeMemoryDependenceWrapperPassPass(*PassRegistry::getPassRegistry());
1749 }
1750 
1751 MemoryDependenceWrapperPass::~MemoryDependenceWrapperPass() = default;
1752 
1753 void MemoryDependenceWrapperPass::releaseMemory() {
1754   MemDep.reset();
1755 }
1756 
1757 void MemoryDependenceWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1758   AU.setPreservesAll();
1759   AU.addRequired<AssumptionCacheTracker>();
1760   AU.addRequired<DominatorTreeWrapperPass>();
1761   AU.addRequired<PhiValuesWrapperPass>();
1762   AU.addRequiredTransitive<AAResultsWrapperPass>();
1763   AU.addRequiredTransitive<TargetLibraryInfoWrapperPass>();
1764 }
1765 
1766 bool MemoryDependenceResults::invalidate(Function &F, const PreservedAnalyses &PA,
1767                                FunctionAnalysisManager::Invalidator &Inv) {
1768   // Check whether our analysis is preserved.
1769   auto PAC = PA.getChecker<MemoryDependenceAnalysis>();
1770   if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>())
1771     // If not, give up now.
1772     return true;
1773 
1774   // Check whether the analyses we depend on became invalid for any reason.
1775   if (Inv.invalidate<AAManager>(F, PA) ||
1776       Inv.invalidate<AssumptionAnalysis>(F, PA) ||
1777       Inv.invalidate<DominatorTreeAnalysis>(F, PA) ||
1778       Inv.invalidate<PhiValuesAnalysis>(F, PA))
1779     return true;
1780 
1781   // Otherwise this analysis result remains valid.
1782   return false;
1783 }
1784 
1785 unsigned MemoryDependenceResults::getDefaultBlockScanLimit() const {
1786   return DefaultBlockScanLimit;
1787 }
1788 
1789 bool MemoryDependenceWrapperPass::runOnFunction(Function &F) {
1790   auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1791   auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1792   auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1793   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1794   auto &PV = getAnalysis<PhiValuesWrapperPass>().getResult();
1795   MemDep.emplace(AA, AC, TLI, DT, PV, BlockScanLimit);
1796   return false;
1797 }
1798