xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/MemoryDependenceAnalysis.cpp (revision 4b9d605768acabc460aa6dcfe8a1f8db35b16794)
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/TargetLibraryInfo.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/IR/BasicBlock.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/InstrTypes.h"
33 #include "llvm/IR/Instruction.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/LLVMContext.h"
37 #include "llvm/IR/Metadata.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/IR/PredIteratorCache.h"
40 #include "llvm/IR/Type.h"
41 #include "llvm/IR/Use.h"
42 #include "llvm/IR/Value.h"
43 #include "llvm/InitializePasses.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/AtomicOrdering.h"
46 #include "llvm/Support/Casting.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Compiler.h"
49 #include "llvm/Support/Debug.h"
50 #include <algorithm>
51 #include <cassert>
52 #include <iterator>
53 #include <utility>
54 
55 using namespace llvm;
56 
57 #define DEBUG_TYPE "memdep"
58 
59 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
60 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
61 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
62 
63 STATISTIC(NumCacheNonLocalPtr,
64           "Number of fully cached non-local ptr responses");
65 STATISTIC(NumCacheDirtyNonLocalPtr,
66           "Number of cached, but dirty, non-local ptr responses");
67 STATISTIC(NumUncacheNonLocalPtr, "Number of uncached non-local ptr responses");
68 STATISTIC(NumCacheCompleteNonLocalPtr,
69           "Number of block queries that were completely cached");
70 
71 // Limit for the number of instructions to scan in a block.
72 
73 static cl::opt<unsigned> BlockScanLimit(
74     "memdep-block-scan-limit", cl::Hidden, cl::init(100),
75     cl::desc("The number of instructions to scan in a block in memory "
76              "dependency analysis (default = 100)"));
77 
78 static cl::opt<unsigned>
79     BlockNumberLimit("memdep-block-number-limit", cl::Hidden, cl::init(200),
80                      cl::desc("The number of blocks to scan during memory "
81                               "dependency analysis (default = 200)"));
82 
83 // Limit on the number of memdep results to process.
84 static const unsigned int NumResultsLimit = 100;
85 
86 /// This is a helper function that removes Val from 'Inst's set in ReverseMap.
87 ///
88 /// If the set becomes empty, remove Inst's entry.
89 template <typename KeyTy>
90 static void
91 RemoveFromReverseMap(DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>> &ReverseMap,
92                      Instruction *Inst, KeyTy Val) {
93   typename DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>>::iterator InstIt =
94       ReverseMap.find(Inst);
95   assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
96   bool Found = InstIt->second.erase(Val);
97   assert(Found && "Invalid reverse map!");
98   (void)Found;
99   if (InstIt->second.empty())
100     ReverseMap.erase(InstIt);
101 }
102 
103 /// If the given instruction references a specific memory location, fill in Loc
104 /// with the details, otherwise set Loc.Ptr to null.
105 ///
106 /// Returns a ModRefInfo value describing the general behavior of the
107 /// instruction.
108 static ModRefInfo GetLocation(const Instruction *Inst, MemoryLocation &Loc,
109                               const TargetLibraryInfo &TLI) {
110   if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
111     if (LI->isUnordered()) {
112       Loc = MemoryLocation::get(LI);
113       return ModRefInfo::Ref;
114     }
115     if (LI->getOrdering() == AtomicOrdering::Monotonic) {
116       Loc = MemoryLocation::get(LI);
117       return ModRefInfo::ModRef;
118     }
119     Loc = MemoryLocation();
120     return ModRefInfo::ModRef;
121   }
122 
123   if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
124     if (SI->isUnordered()) {
125       Loc = MemoryLocation::get(SI);
126       return ModRefInfo::Mod;
127     }
128     if (SI->getOrdering() == AtomicOrdering::Monotonic) {
129       Loc = MemoryLocation::get(SI);
130       return ModRefInfo::ModRef;
131     }
132     Loc = MemoryLocation();
133     return ModRefInfo::ModRef;
134   }
135 
136   if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
137     Loc = MemoryLocation::get(V);
138     return ModRefInfo::ModRef;
139   }
140 
141   if (const CallBase *CB = dyn_cast<CallBase>(Inst)) {
142     if (Value *FreedOp = getFreedOperand(CB, &TLI)) {
143       // calls to free() deallocate the entire structure
144       Loc = MemoryLocation::getAfter(FreedOp);
145       return ModRefInfo::Mod;
146     }
147   }
148 
149   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
150     switch (II->getIntrinsicID()) {
151     case Intrinsic::lifetime_start:
152     case Intrinsic::lifetime_end:
153     case Intrinsic::invariant_start:
154       Loc = MemoryLocation::getForArgument(II, 1, TLI);
155       // These intrinsics don't really modify the memory, but returning Mod
156       // will allow them to be handled conservatively.
157       return ModRefInfo::Mod;
158     case Intrinsic::invariant_end:
159       Loc = MemoryLocation::getForArgument(II, 2, TLI);
160       // These intrinsics don't really modify the memory, but returning Mod
161       // will allow them to be handled conservatively.
162       return ModRefInfo::Mod;
163     case Intrinsic::masked_load:
164       Loc = MemoryLocation::getForArgument(II, 0, TLI);
165       return ModRefInfo::Ref;
166     case Intrinsic::masked_store:
167       Loc = MemoryLocation::getForArgument(II, 1, TLI);
168       return ModRefInfo::Mod;
169     default:
170       break;
171     }
172   }
173 
174   // Otherwise, just do the coarse-grained thing that always works.
175   if (Inst->mayWriteToMemory())
176     return ModRefInfo::ModRef;
177   if (Inst->mayReadFromMemory())
178     return ModRefInfo::Ref;
179   return ModRefInfo::NoModRef;
180 }
181 
182 /// Private helper for finding the local dependencies of a call site.
183 MemDepResult MemoryDependenceResults::getCallDependencyFrom(
184     CallBase *Call, bool isReadOnlyCall, BasicBlock::iterator ScanIt,
185     BasicBlock *BB) {
186   unsigned Limit = getDefaultBlockScanLimit();
187 
188   // Walk backwards through the block, looking for dependencies.
189   while (ScanIt != BB->begin()) {
190     Instruction *Inst = &*--ScanIt;
191     // Debug intrinsics don't cause dependences and should not affect Limit
192     if (isa<DbgInfoIntrinsic>(Inst))
193       continue;
194 
195     // Limit the amount of scanning we do so we don't end up with quadratic
196     // running time on extreme testcases.
197     --Limit;
198     if (!Limit)
199       return MemDepResult::getUnknown();
200 
201     // If this inst is a memory op, get the pointer it accessed
202     MemoryLocation Loc;
203     ModRefInfo MR = GetLocation(Inst, Loc, TLI);
204     if (Loc.Ptr) {
205       // A simple instruction.
206       if (isModOrRefSet(AA.getModRefInfo(Call, Loc)))
207         return MemDepResult::getClobber(Inst);
208       continue;
209     }
210 
211     if (auto *CallB = dyn_cast<CallBase>(Inst)) {
212       // If these two calls do not interfere, look past it.
213       if (isNoModRef(AA.getModRefInfo(Call, CallB))) {
214         // If the two calls are the same, return Inst as a Def, so that
215         // Call can be found redundant and eliminated.
216         if (isReadOnlyCall && !isModSet(MR) &&
217             Call->isIdenticalToWhenDefined(CallB))
218           return MemDepResult::getDef(Inst);
219 
220         // Otherwise if the two calls don't interact (e.g. CallB is readnone)
221         // keep scanning.
222         continue;
223       } else
224         return MemDepResult::getClobber(Inst);
225     }
226 
227     // If we could not obtain a pointer for the instruction and the instruction
228     // touches memory then assume that this is a dependency.
229     if (isModOrRefSet(MR))
230       return MemDepResult::getClobber(Inst);
231   }
232 
233   // No dependence found.  If this is the entry block of the function, it is
234   // unknown, otherwise it is non-local.
235   if (BB != &BB->getParent()->getEntryBlock())
236     return MemDepResult::getNonLocal();
237   return MemDepResult::getNonFuncLocal();
238 }
239 
240 MemDepResult MemoryDependenceResults::getPointerDependencyFrom(
241     const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
242     BasicBlock *BB, Instruction *QueryInst, unsigned *Limit,
243     BatchAAResults &BatchAA) {
244   MemDepResult InvariantGroupDependency = MemDepResult::getUnknown();
245   if (QueryInst != nullptr) {
246     if (auto *LI = dyn_cast<LoadInst>(QueryInst)) {
247       InvariantGroupDependency = getInvariantGroupPointerDependency(LI, BB);
248 
249       if (InvariantGroupDependency.isDef())
250         return InvariantGroupDependency;
251     }
252   }
253   MemDepResult SimpleDep = getSimplePointerDependencyFrom(
254       MemLoc, isLoad, ScanIt, BB, QueryInst, Limit, BatchAA);
255   if (SimpleDep.isDef())
256     return SimpleDep;
257   // Non-local invariant group dependency indicates there is non local Def
258   // (it only returns nonLocal if it finds nonLocal def), which is better than
259   // local clobber and everything else.
260   if (InvariantGroupDependency.isNonLocal())
261     return InvariantGroupDependency;
262 
263   assert(InvariantGroupDependency.isUnknown() &&
264          "InvariantGroupDependency should be only unknown at this point");
265   return SimpleDep;
266 }
267 
268 MemDepResult MemoryDependenceResults::getPointerDependencyFrom(
269     const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
270     BasicBlock *BB, Instruction *QueryInst, unsigned *Limit) {
271   BatchAAResults BatchAA(AA);
272   return getPointerDependencyFrom(MemLoc, isLoad, ScanIt, BB, QueryInst, Limit,
273                                   BatchAA);
274 }
275 
276 MemDepResult
277 MemoryDependenceResults::getInvariantGroupPointerDependency(LoadInst *LI,
278                                                             BasicBlock *BB) {
279 
280   if (!LI->hasMetadata(LLVMContext::MD_invariant_group))
281     return MemDepResult::getUnknown();
282 
283   // Take the ptr operand after all casts and geps 0. This way we can search
284   // cast graph down only.
285   Value *LoadOperand = LI->getPointerOperand()->stripPointerCasts();
286 
287   // It's is not safe to walk the use list of global value, because function
288   // passes aren't allowed to look outside their functions.
289   // FIXME: this could be fixed by filtering instructions from outside
290   // of current function.
291   if (isa<GlobalValue>(LoadOperand))
292     return MemDepResult::getUnknown();
293 
294   // Queue to process all pointers that are equivalent to load operand.
295   SmallVector<const Value *, 8> LoadOperandsQueue;
296   LoadOperandsQueue.push_back(LoadOperand);
297 
298   Instruction *ClosestDependency = nullptr;
299   // Order of instructions in uses list is unpredictible. In order to always
300   // get the same result, we will look for the closest dominance.
301   auto GetClosestDependency = [this](Instruction *Best, Instruction *Other) {
302     assert(Other && "Must call it with not null instruction");
303     if (Best == nullptr || DT.dominates(Best, Other))
304       return Other;
305     return Best;
306   };
307 
308   // FIXME: This loop is O(N^2) because dominates can be O(n) and in worst case
309   // we will see all the instructions. This should be fixed in MSSA.
310   while (!LoadOperandsQueue.empty()) {
311     const Value *Ptr = LoadOperandsQueue.pop_back_val();
312     assert(Ptr && !isa<GlobalValue>(Ptr) &&
313            "Null or GlobalValue should not be inserted");
314 
315     for (const Use &Us : Ptr->uses()) {
316       auto *U = dyn_cast<Instruction>(Us.getUser());
317       if (!U || U == LI || !DT.dominates(U, LI))
318         continue;
319 
320       // Bitcast or gep with zeros are using Ptr. Add to queue to check it's
321       // users.      U = bitcast Ptr
322       if (isa<BitCastInst>(U)) {
323         LoadOperandsQueue.push_back(U);
324         continue;
325       }
326       // Gep with zeros is equivalent to bitcast.
327       // FIXME: we are not sure if some bitcast should be canonicalized to gep 0
328       // or gep 0 to bitcast because of SROA, so there are 2 forms. When
329       // typeless pointers will be ready then both cases will be gone
330       // (and this BFS also won't be needed).
331       if (auto *GEP = dyn_cast<GetElementPtrInst>(U))
332         if (GEP->hasAllZeroIndices()) {
333           LoadOperandsQueue.push_back(U);
334           continue;
335         }
336 
337       // If we hit load/store with the same invariant.group metadata (and the
338       // same pointer operand) we can assume that value pointed by pointer
339       // operand didn't change.
340       if ((isa<LoadInst>(U) ||
341            (isa<StoreInst>(U) &&
342             cast<StoreInst>(U)->getPointerOperand() == Ptr)) &&
343           U->hasMetadata(LLVMContext::MD_invariant_group))
344         ClosestDependency = GetClosestDependency(ClosestDependency, U);
345     }
346   }
347 
348   if (!ClosestDependency)
349     return MemDepResult::getUnknown();
350   if (ClosestDependency->getParent() == BB)
351     return MemDepResult::getDef(ClosestDependency);
352   // Def(U) can't be returned here because it is non-local. If local
353   // dependency won't be found then return nonLocal counting that the
354   // user will call getNonLocalPointerDependency, which will return cached
355   // result.
356   NonLocalDefsCache.try_emplace(
357       LI, NonLocalDepResult(ClosestDependency->getParent(),
358                             MemDepResult::getDef(ClosestDependency), nullptr));
359   ReverseNonLocalDefsCache[ClosestDependency].insert(LI);
360   return MemDepResult::getNonLocal();
361 }
362 
363 MemDepResult MemoryDependenceResults::getSimplePointerDependencyFrom(
364     const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
365     BasicBlock *BB, Instruction *QueryInst, unsigned *Limit,
366     BatchAAResults &BatchAA) {
367   bool isInvariantLoad = false;
368 
369   unsigned DefaultLimit = getDefaultBlockScanLimit();
370   if (!Limit)
371     Limit = &DefaultLimit;
372 
373   // We must be careful with atomic accesses, as they may allow another thread
374   //   to touch this location, clobbering it. We are conservative: if the
375   //   QueryInst is not a simple (non-atomic) memory access, we automatically
376   //   return getClobber.
377   // If it is simple, we know based on the results of
378   // "Compiler testing via a theory of sound optimisations in the C11/C++11
379   //   memory model" in PLDI 2013, that a non-atomic location can only be
380   //   clobbered between a pair of a release and an acquire action, with no
381   //   access to the location in between.
382   // Here is an example for giving the general intuition behind this rule.
383   // In the following code:
384   //   store x 0;
385   //   release action; [1]
386   //   acquire action; [4]
387   //   %val = load x;
388   // It is unsafe to replace %val by 0 because another thread may be running:
389   //   acquire action; [2]
390   //   store x 42;
391   //   release action; [3]
392   // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
393   // being 42. A key property of this program however is that if either
394   // 1 or 4 were missing, there would be a race between the store of 42
395   // either the store of 0 or the load (making the whole program racy).
396   // The paper mentioned above shows that the same property is respected
397   // by every program that can detect any optimization of that kind: either
398   // it is racy (undefined) or there is a release followed by an acquire
399   // between the pair of accesses under consideration.
400 
401   // If the load is invariant, we "know" that it doesn't alias *any* write. We
402   // do want to respect mustalias results since defs are useful for value
403   // forwarding, but any mayalias write can be assumed to be noalias.
404   // Arguably, this logic should be pushed inside AliasAnalysis itself.
405   if (isLoad && QueryInst) {
406     LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
407     if (LI && LI->hasMetadata(LLVMContext::MD_invariant_load))
408       isInvariantLoad = true;
409   }
410 
411   // True for volatile instruction.
412   // For Load/Store return true if atomic ordering is stronger than AO,
413   // for other instruction just true if it can read or write to memory.
414   auto isComplexForReordering = [](Instruction * I, AtomicOrdering AO)->bool {
415     if (I->isVolatile())
416       return true;
417     if (auto *LI = dyn_cast<LoadInst>(I))
418       return isStrongerThan(LI->getOrdering(), AO);
419     if (auto *SI = dyn_cast<StoreInst>(I))
420       return isStrongerThan(SI->getOrdering(), AO);
421     return I->mayReadOrWriteMemory();
422   };
423 
424   // Walk backwards through the basic block, looking for dependencies.
425   while (ScanIt != BB->begin()) {
426     Instruction *Inst = &*--ScanIt;
427 
428     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
429       // Debug intrinsics don't (and can't) cause dependencies.
430       if (isa<DbgInfoIntrinsic>(II))
431         continue;
432 
433     // Limit the amount of scanning we do so we don't end up with quadratic
434     // running time on extreme testcases.
435     --*Limit;
436     if (!*Limit)
437       return MemDepResult::getUnknown();
438 
439     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
440       // If we reach a lifetime begin or end marker, then the query ends here
441       // because the value is undefined.
442       Intrinsic::ID ID = II->getIntrinsicID();
443       switch (ID) {
444       case Intrinsic::lifetime_start: {
445         // FIXME: This only considers queries directly on the invariant-tagged
446         // pointer, not on query pointers that are indexed off of them.  It'd
447         // be nice to handle that at some point (the right approach is to use
448         // GetPointerBaseWithConstantOffset).
449         MemoryLocation ArgLoc = MemoryLocation::getAfter(II->getArgOperand(1));
450         if (BatchAA.isMustAlias(ArgLoc, MemLoc))
451           return MemDepResult::getDef(II);
452         continue;
453       }
454       case Intrinsic::masked_load:
455       case Intrinsic::masked_store: {
456         MemoryLocation Loc;
457         /*ModRefInfo MR =*/ GetLocation(II, Loc, TLI);
458         AliasResult R = BatchAA.alias(Loc, MemLoc);
459         if (R == AliasResult::NoAlias)
460           continue;
461         if (R == AliasResult::MustAlias)
462           return MemDepResult::getDef(II);
463         if (ID == Intrinsic::masked_load)
464           continue;
465         return MemDepResult::getClobber(II);
466       }
467       }
468     }
469 
470     // Values depend on loads if the pointers are must aliased.  This means
471     // that a load depends on another must aliased load from the same value.
472     // One exception is atomic loads: a value can depend on an atomic load that
473     // it does not alias with when this atomic load indicates that another
474     // thread may be accessing the location.
475     if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
476       // While volatile access cannot be eliminated, they do not have to clobber
477       // non-aliasing locations, as normal accesses, for example, can be safely
478       // reordered with volatile accesses.
479       if (LI->isVolatile()) {
480         if (!QueryInst)
481           // Original QueryInst *may* be volatile
482           return MemDepResult::getClobber(LI);
483         if (QueryInst->isVolatile())
484           // Ordering required if QueryInst is itself volatile
485           return MemDepResult::getClobber(LI);
486         // Otherwise, volatile doesn't imply any special ordering
487       }
488 
489       // Atomic loads have complications involved.
490       // A Monotonic (or higher) load is OK if the query inst is itself not
491       // atomic.
492       // FIXME: This is overly conservative.
493       if (LI->isAtomic() && isStrongerThanUnordered(LI->getOrdering())) {
494         if (!QueryInst ||
495             isComplexForReordering(QueryInst, AtomicOrdering::NotAtomic))
496           return MemDepResult::getClobber(LI);
497         if (LI->getOrdering() != AtomicOrdering::Monotonic)
498           return MemDepResult::getClobber(LI);
499       }
500 
501       MemoryLocation LoadLoc = MemoryLocation::get(LI);
502 
503       // If we found a pointer, check if it could be the same as our pointer.
504       AliasResult R = BatchAA.alias(LoadLoc, MemLoc);
505 
506       if (R == AliasResult::NoAlias)
507         continue;
508 
509       if (isLoad) {
510         // Must aliased loads are defs of each other.
511         if (R == AliasResult::MustAlias)
512           return MemDepResult::getDef(Inst);
513 
514         // If we have a partial alias, then return this as a clobber for the
515         // client to handle.
516         if (R == AliasResult::PartialAlias && R.hasOffset()) {
517           ClobberOffsets[LI] = R.getOffset();
518           return MemDepResult::getClobber(Inst);
519         }
520 
521         // Random may-alias loads don't depend on each other without a
522         // dependence.
523         continue;
524       }
525 
526       // Stores don't alias loads from read-only memory.
527       if (!isModSet(BatchAA.getModRefInfoMask(LoadLoc)))
528         continue;
529 
530       // Stores depend on may/must aliased loads.
531       return MemDepResult::getDef(Inst);
532     }
533 
534     if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
535       // Atomic stores have complications involved.
536       // A Monotonic store is OK if the query inst is itself not atomic.
537       // FIXME: This is overly conservative.
538       if (!SI->isUnordered() && SI->isAtomic()) {
539         if (!QueryInst ||
540             isComplexForReordering(QueryInst, AtomicOrdering::Unordered))
541           return MemDepResult::getClobber(SI);
542         // Ok, if we are here the guard above guarantee us that
543         // QueryInst is a non-atomic or unordered load/store.
544         // SI is atomic with monotonic or release semantic (seq_cst for store
545         // is actually a release semantic plus total order over other seq_cst
546         // instructions, as soon as QueryInst is not seq_cst we can consider it
547         // as simple release semantic).
548         // Monotonic and Release semantic allows re-ordering before store
549         // so we are safe to go further and check the aliasing. It will prohibit
550         // re-ordering in case locations are may or must alias.
551       }
552 
553       // While volatile access cannot be eliminated, they do not have to clobber
554       // non-aliasing locations, as normal accesses can for example be reordered
555       // with volatile accesses.
556       if (SI->isVolatile())
557         if (!QueryInst || QueryInst->isVolatile())
558           return MemDepResult::getClobber(SI);
559 
560       // If alias analysis can tell that this store is guaranteed to not modify
561       // the query pointer, ignore it.  Use getModRefInfo to handle cases where
562       // the query pointer points to constant memory etc.
563       if (!isModOrRefSet(BatchAA.getModRefInfo(SI, MemLoc)))
564         continue;
565 
566       // Ok, this store might clobber the query pointer.  Check to see if it is
567       // a must alias: in this case, we want to return this as a def.
568       // FIXME: Use ModRefInfo::Must bit from getModRefInfo call above.
569       MemoryLocation StoreLoc = MemoryLocation::get(SI);
570 
571       // If we found a pointer, check if it could be the same as our pointer.
572       AliasResult R = BatchAA.alias(StoreLoc, MemLoc);
573 
574       if (R == AliasResult::NoAlias)
575         continue;
576       if (R == AliasResult::MustAlias)
577         return MemDepResult::getDef(Inst);
578       if (isInvariantLoad)
579         continue;
580       return MemDepResult::getClobber(Inst);
581     }
582 
583     // If this is an allocation, and if we know that the accessed pointer is to
584     // the allocation, return Def.  This means that there is no dependence and
585     // the access can be optimized based on that.  For example, a load could
586     // turn into undef.  Note that we can bypass the allocation itself when
587     // looking for a clobber in many cases; that's an alias property and is
588     // handled by BasicAA.
589     if (isa<AllocaInst>(Inst) || isNoAliasCall(Inst)) {
590       const Value *AccessPtr = getUnderlyingObject(MemLoc.Ptr);
591       if (AccessPtr == Inst || BatchAA.isMustAlias(Inst, AccessPtr))
592         return MemDepResult::getDef(Inst);
593     }
594 
595     // If we found a select instruction for MemLoc pointer, return it as Def
596     // dependency.
597     if (isa<SelectInst>(Inst) && MemLoc.Ptr == Inst)
598       return MemDepResult::getDef(Inst);
599 
600     if (isInvariantLoad)
601       continue;
602 
603     // A release fence requires that all stores complete before it, but does
604     // not prevent the reordering of following loads or stores 'before' the
605     // fence.  As a result, we look past it when finding a dependency for
606     // loads.  DSE uses this to find preceding stores to delete and thus we
607     // can't bypass the fence if the query instruction is a store.
608     if (FenceInst *FI = dyn_cast<FenceInst>(Inst))
609       if (isLoad && FI->getOrdering() == AtomicOrdering::Release)
610         continue;
611 
612     // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
613     ModRefInfo MR = BatchAA.getModRefInfo(Inst, MemLoc);
614     // If necessary, perform additional analysis.
615     if (isModAndRefSet(MR))
616       MR = BatchAA.callCapturesBefore(Inst, MemLoc, &DT);
617     switch (MR) {
618     case ModRefInfo::NoModRef:
619       // If the call has no effect on the queried pointer, just ignore it.
620       continue;
621     case ModRefInfo::Mod:
622       return MemDepResult::getClobber(Inst);
623     case ModRefInfo::Ref:
624       // If the call is known to never store to the pointer, and if this is a
625       // load query, we can safely ignore it (scan past it).
626       if (isLoad)
627         continue;
628       [[fallthrough]];
629     default:
630       // Otherwise, there is a potential dependence.  Return a clobber.
631       return MemDepResult::getClobber(Inst);
632     }
633   }
634 
635   // No dependence found.  If this is the entry block of the function, it is
636   // unknown, otherwise it is non-local.
637   if (BB != &BB->getParent()->getEntryBlock())
638     return MemDepResult::getNonLocal();
639   return MemDepResult::getNonFuncLocal();
640 }
641 
642 MemDepResult MemoryDependenceResults::getDependency(Instruction *QueryInst) {
643   ClobberOffsets.clear();
644   Instruction *ScanPos = QueryInst;
645 
646   // Check for a cached result
647   MemDepResult &LocalCache = LocalDeps[QueryInst];
648 
649   // If the cached entry is non-dirty, just return it.  Note that this depends
650   // on MemDepResult's default constructing to 'dirty'.
651   if (!LocalCache.isDirty())
652     return LocalCache;
653 
654   // Otherwise, if we have a dirty entry, we know we can start the scan at that
655   // instruction, which may save us some work.
656   if (Instruction *Inst = LocalCache.getInst()) {
657     ScanPos = Inst;
658 
659     RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
660   }
661 
662   BasicBlock *QueryParent = QueryInst->getParent();
663 
664   // Do the scan.
665   if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
666     // No dependence found. If this is the entry block of the function, it is
667     // unknown, otherwise it is non-local.
668     if (QueryParent != &QueryParent->getParent()->getEntryBlock())
669       LocalCache = MemDepResult::getNonLocal();
670     else
671       LocalCache = MemDepResult::getNonFuncLocal();
672   } else {
673     MemoryLocation MemLoc;
674     ModRefInfo MR = GetLocation(QueryInst, MemLoc, TLI);
675     if (MemLoc.Ptr) {
676       // If we can do a pointer scan, make it happen.
677       bool isLoad = !isModSet(MR);
678       if (auto *II = dyn_cast<IntrinsicInst>(QueryInst))
679         isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
680 
681       LocalCache =
682           getPointerDependencyFrom(MemLoc, isLoad, ScanPos->getIterator(),
683                                    QueryParent, QueryInst, nullptr);
684     } else if (auto *QueryCall = dyn_cast<CallBase>(QueryInst)) {
685       bool isReadOnly = AA.onlyReadsMemory(QueryCall);
686       LocalCache = getCallDependencyFrom(QueryCall, isReadOnly,
687                                          ScanPos->getIterator(), QueryParent);
688     } else
689       // Non-memory instruction.
690       LocalCache = MemDepResult::getUnknown();
691   }
692 
693   // Remember the result!
694   if (Instruction *I = LocalCache.getInst())
695     ReverseLocalDeps[I].insert(QueryInst);
696 
697   return LocalCache;
698 }
699 
700 #ifndef NDEBUG
701 /// This method is used when -debug is specified to verify that cache arrays
702 /// are properly kept sorted.
703 static void AssertSorted(MemoryDependenceResults::NonLocalDepInfo &Cache,
704                          int Count = -1) {
705   if (Count == -1)
706     Count = Cache.size();
707   assert(std::is_sorted(Cache.begin(), Cache.begin() + Count) &&
708          "Cache isn't sorted!");
709 }
710 #endif
711 
712 const MemoryDependenceResults::NonLocalDepInfo &
713 MemoryDependenceResults::getNonLocalCallDependency(CallBase *QueryCall) {
714   assert(getDependency(QueryCall).isNonLocal() &&
715          "getNonLocalCallDependency should only be used on calls with "
716          "non-local deps!");
717   PerInstNLInfo &CacheP = NonLocalDepsMap[QueryCall];
718   NonLocalDepInfo &Cache = CacheP.first;
719 
720   // This is the set of blocks that need to be recomputed.  In the cached case,
721   // this can happen due to instructions being deleted etc. In the uncached
722   // case, this starts out as the set of predecessors we care about.
723   SmallVector<BasicBlock *, 32> DirtyBlocks;
724 
725   if (!Cache.empty()) {
726     // Okay, we have a cache entry.  If we know it is not dirty, just return it
727     // with no computation.
728     if (!CacheP.second) {
729       ++NumCacheNonLocal;
730       return Cache;
731     }
732 
733     // If we already have a partially computed set of results, scan them to
734     // determine what is dirty, seeding our initial DirtyBlocks worklist.
735     for (auto &Entry : Cache)
736       if (Entry.getResult().isDirty())
737         DirtyBlocks.push_back(Entry.getBB());
738 
739     // Sort the cache so that we can do fast binary search lookups below.
740     llvm::sort(Cache);
741 
742     ++NumCacheDirtyNonLocal;
743   } else {
744     // Seed DirtyBlocks with each of the preds of QueryInst's block.
745     BasicBlock *QueryBB = QueryCall->getParent();
746     append_range(DirtyBlocks, PredCache.get(QueryBB));
747     ++NumUncacheNonLocal;
748   }
749 
750   // isReadonlyCall - If this is a read-only call, we can be more aggressive.
751   bool isReadonlyCall = AA.onlyReadsMemory(QueryCall);
752 
753   SmallPtrSet<BasicBlock *, 32> Visited;
754 
755   unsigned NumSortedEntries = Cache.size();
756   LLVM_DEBUG(AssertSorted(Cache));
757 
758   // Iterate while we still have blocks to update.
759   while (!DirtyBlocks.empty()) {
760     BasicBlock *DirtyBB = DirtyBlocks.pop_back_val();
761 
762     // Already processed this block?
763     if (!Visited.insert(DirtyBB).second)
764       continue;
765 
766     // Do a binary search to see if we already have an entry for this block in
767     // the cache set.  If so, find it.
768     LLVM_DEBUG(AssertSorted(Cache, NumSortedEntries));
769     NonLocalDepInfo::iterator Entry =
770         std::upper_bound(Cache.begin(), Cache.begin() + NumSortedEntries,
771                          NonLocalDepEntry(DirtyBB));
772     if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
773       --Entry;
774 
775     NonLocalDepEntry *ExistingResult = nullptr;
776     if (Entry != Cache.begin() + NumSortedEntries &&
777         Entry->getBB() == DirtyBB) {
778       // If we already have an entry, and if it isn't already dirty, the block
779       // is done.
780       if (!Entry->getResult().isDirty())
781         continue;
782 
783       // Otherwise, remember this slot so we can update the value.
784       ExistingResult = &*Entry;
785     }
786 
787     // If the dirty entry has a pointer, start scanning from it so we don't have
788     // to rescan the entire block.
789     BasicBlock::iterator ScanPos = DirtyBB->end();
790     if (ExistingResult) {
791       if (Instruction *Inst = ExistingResult->getResult().getInst()) {
792         ScanPos = Inst->getIterator();
793         // We're removing QueryInst's use of Inst.
794         RemoveFromReverseMap<Instruction *>(ReverseNonLocalDeps, Inst,
795                                             QueryCall);
796       }
797     }
798 
799     // Find out if this block has a local dependency for QueryInst.
800     MemDepResult Dep;
801 
802     if (ScanPos != DirtyBB->begin()) {
803       Dep = getCallDependencyFrom(QueryCall, isReadonlyCall, ScanPos, DirtyBB);
804     } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
805       // No dependence found.  If this is the entry block of the function, it is
806       // a clobber, otherwise it is unknown.
807       Dep = MemDepResult::getNonLocal();
808     } else {
809       Dep = MemDepResult::getNonFuncLocal();
810     }
811 
812     // If we had a dirty entry for the block, update it.  Otherwise, just add
813     // a new entry.
814     if (ExistingResult)
815       ExistingResult->setResult(Dep);
816     else
817       Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
818 
819     // If the block has a dependency (i.e. it isn't completely transparent to
820     // the value), remember the association!
821     if (!Dep.isNonLocal()) {
822       // Keep the ReverseNonLocalDeps map up to date so we can efficiently
823       // update this when we remove instructions.
824       if (Instruction *Inst = Dep.getInst())
825         ReverseNonLocalDeps[Inst].insert(QueryCall);
826     } else {
827 
828       // If the block *is* completely transparent to the load, we need to check
829       // the predecessors of this block.  Add them to our worklist.
830       append_range(DirtyBlocks, PredCache.get(DirtyBB));
831     }
832   }
833 
834   return Cache;
835 }
836 
837 void MemoryDependenceResults::getNonLocalPointerDependency(
838     Instruction *QueryInst, SmallVectorImpl<NonLocalDepResult> &Result) {
839   const MemoryLocation Loc = MemoryLocation::get(QueryInst);
840   bool isLoad = isa<LoadInst>(QueryInst);
841   BasicBlock *FromBB = QueryInst->getParent();
842   assert(FromBB);
843 
844   assert(Loc.Ptr->getType()->isPointerTy() &&
845          "Can't get pointer deps of a non-pointer!");
846   Result.clear();
847   {
848     // Check if there is cached Def with invariant.group.
849     auto NonLocalDefIt = NonLocalDefsCache.find(QueryInst);
850     if (NonLocalDefIt != NonLocalDefsCache.end()) {
851       Result.push_back(NonLocalDefIt->second);
852       ReverseNonLocalDefsCache[NonLocalDefIt->second.getResult().getInst()]
853           .erase(QueryInst);
854       NonLocalDefsCache.erase(NonLocalDefIt);
855       return;
856     }
857   }
858   // This routine does not expect to deal with volatile instructions.
859   // Doing so would require piping through the QueryInst all the way through.
860   // TODO: volatiles can't be elided, but they can be reordered with other
861   // non-volatile accesses.
862 
863   // We currently give up on any instruction which is ordered, but we do handle
864   // atomic instructions which are unordered.
865   // TODO: Handle ordered instructions
866   auto isOrdered = [](Instruction *Inst) {
867     if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
868       return !LI->isUnordered();
869     } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
870       return !SI->isUnordered();
871     }
872     return false;
873   };
874   if (QueryInst->isVolatile() || isOrdered(QueryInst)) {
875     Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(),
876                                        const_cast<Value *>(Loc.Ptr)));
877     return;
878   }
879   const DataLayout &DL = FromBB->getModule()->getDataLayout();
880   PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, &AC);
881 
882   // This is the set of blocks we've inspected, and the pointer we consider in
883   // each block.  Because of critical edges, we currently bail out if querying
884   // a block with multiple different pointers.  This can happen during PHI
885   // translation.
886   DenseMap<BasicBlock *, Value *> Visited;
887   if (getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
888                                    Result, Visited, true))
889     return;
890   Result.clear();
891   Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(),
892                                      const_cast<Value *>(Loc.Ptr)));
893 }
894 
895 /// Compute the memdep value for BB with Pointer/PointeeSize using either
896 /// cached information in Cache or by doing a lookup (which may use dirty cache
897 /// info if available).
898 ///
899 /// If we do a lookup, add the result to the cache.
900 MemDepResult MemoryDependenceResults::getNonLocalInfoForBlock(
901     Instruction *QueryInst, const MemoryLocation &Loc, bool isLoad,
902     BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries,
903     BatchAAResults &BatchAA) {
904 
905   bool isInvariantLoad = false;
906 
907   if (LoadInst *LI = dyn_cast_or_null<LoadInst>(QueryInst))
908     isInvariantLoad = LI->getMetadata(LLVMContext::MD_invariant_load);
909 
910   // Do a binary search to see if we already have an entry for this block in
911   // the cache set.  If so, find it.
912   NonLocalDepInfo::iterator Entry = std::upper_bound(
913       Cache->begin(), Cache->begin() + NumSortedEntries, NonLocalDepEntry(BB));
914   if (Entry != Cache->begin() && (Entry - 1)->getBB() == BB)
915     --Entry;
916 
917   NonLocalDepEntry *ExistingResult = nullptr;
918   if (Entry != Cache->begin() + NumSortedEntries && Entry->getBB() == BB)
919     ExistingResult = &*Entry;
920 
921   // Use cached result for invariant load only if there is no dependency for non
922   // invariant load. In this case invariant load can not have any dependency as
923   // well.
924   if (ExistingResult && isInvariantLoad &&
925       !ExistingResult->getResult().isNonFuncLocal())
926     ExistingResult = nullptr;
927 
928   // If we have a cached entry, and it is non-dirty, use it as the value for
929   // this dependency.
930   if (ExistingResult && !ExistingResult->getResult().isDirty()) {
931     ++NumCacheNonLocalPtr;
932     return ExistingResult->getResult();
933   }
934 
935   // Otherwise, we have to scan for the value.  If we have a dirty cache
936   // entry, start scanning from its position, otherwise we scan from the end
937   // of the block.
938   BasicBlock::iterator ScanPos = BB->end();
939   if (ExistingResult && ExistingResult->getResult().getInst()) {
940     assert(ExistingResult->getResult().getInst()->getParent() == BB &&
941            "Instruction invalidated?");
942     ++NumCacheDirtyNonLocalPtr;
943     ScanPos = ExistingResult->getResult().getInst()->getIterator();
944 
945     // Eliminating the dirty entry from 'Cache', so update the reverse info.
946     ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
947     RemoveFromReverseMap(ReverseNonLocalPtrDeps, &*ScanPos, CacheKey);
948   } else {
949     ++NumUncacheNonLocalPtr;
950   }
951 
952   // Scan the block for the dependency.
953   MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB,
954                                               QueryInst, nullptr, BatchAA);
955 
956   // Don't cache results for invariant load.
957   if (isInvariantLoad)
958     return Dep;
959 
960   // If we had a dirty entry for the block, update it.  Otherwise, just add
961   // a new entry.
962   if (ExistingResult)
963     ExistingResult->setResult(Dep);
964   else
965     Cache->push_back(NonLocalDepEntry(BB, Dep));
966 
967   // If the block has a dependency (i.e. it isn't completely transparent to
968   // the value), remember the reverse association because we just added it
969   // to Cache!
970   if (!Dep.isLocal())
971     return Dep;
972 
973   // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
974   // update MemDep when we remove instructions.
975   Instruction *Inst = Dep.getInst();
976   assert(Inst && "Didn't depend on anything?");
977   ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
978   ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
979   return Dep;
980 }
981 
982 /// Sort the NonLocalDepInfo cache, given a certain number of elements in the
983 /// array that are already properly ordered.
984 ///
985 /// This is optimized for the case when only a few entries are added.
986 static void
987 SortNonLocalDepInfoCache(MemoryDependenceResults::NonLocalDepInfo &Cache,
988                          unsigned NumSortedEntries) {
989   switch (Cache.size() - NumSortedEntries) {
990   case 0:
991     // done, no new entries.
992     break;
993   case 2: {
994     // Two new entries, insert the last one into place.
995     NonLocalDepEntry Val = Cache.back();
996     Cache.pop_back();
997     MemoryDependenceResults::NonLocalDepInfo::iterator Entry =
998         std::upper_bound(Cache.begin(), Cache.end() - 1, Val);
999     Cache.insert(Entry, Val);
1000     [[fallthrough]];
1001   }
1002   case 1:
1003     // One new entry, Just insert the new value at the appropriate position.
1004     if (Cache.size() != 1) {
1005       NonLocalDepEntry Val = Cache.back();
1006       Cache.pop_back();
1007       MemoryDependenceResults::NonLocalDepInfo::iterator Entry =
1008           llvm::upper_bound(Cache, Val);
1009       Cache.insert(Entry, Val);
1010     }
1011     break;
1012   default:
1013     // Added many values, do a full scale sort.
1014     llvm::sort(Cache);
1015     break;
1016   }
1017 }
1018 
1019 /// Perform a dependency query based on pointer/pointeesize starting at the end
1020 /// of StartBB.
1021 ///
1022 /// Add any clobber/def results to the results vector and keep track of which
1023 /// blocks are visited in 'Visited'.
1024 ///
1025 /// This has special behavior for the first block queries (when SkipFirstBlock
1026 /// is true).  In this special case, it ignores the contents of the specified
1027 /// block and starts returning dependence info for its predecessors.
1028 ///
1029 /// This function returns true on success, or false to indicate that it could
1030 /// not compute dependence information for some reason.  This should be treated
1031 /// as a clobber dependence on the first instruction in the predecessor block.
1032 bool MemoryDependenceResults::getNonLocalPointerDepFromBB(
1033     Instruction *QueryInst, const PHITransAddr &Pointer,
1034     const MemoryLocation &Loc, bool isLoad, BasicBlock *StartBB,
1035     SmallVectorImpl<NonLocalDepResult> &Result,
1036     DenseMap<BasicBlock *, Value *> &Visited, bool SkipFirstBlock,
1037     bool IsIncomplete) {
1038   // Look up the cached info for Pointer.
1039   ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1040 
1041   // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1042   // CacheKey, this value will be inserted as the associated value. Otherwise,
1043   // it'll be ignored, and we'll have to check to see if the cached size and
1044   // aa tags are consistent with the current query.
1045   NonLocalPointerInfo InitialNLPI;
1046   InitialNLPI.Size = Loc.Size;
1047   InitialNLPI.AATags = Loc.AATags;
1048 
1049   bool isInvariantLoad = false;
1050   if (LoadInst *LI = dyn_cast_or_null<LoadInst>(QueryInst))
1051     isInvariantLoad = LI->getMetadata(LLVMContext::MD_invariant_load);
1052 
1053   // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1054   // already have one.
1055   std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1056       NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1057   NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1058 
1059   // If we already have a cache entry for this CacheKey, we may need to do some
1060   // work to reconcile the cache entry and the current query.
1061   // Invariant loads don't participate in caching. Thus no need to reconcile.
1062   if (!isInvariantLoad && !Pair.second) {
1063     if (CacheInfo->Size != Loc.Size) {
1064       bool ThrowOutEverything;
1065       if (CacheInfo->Size.hasValue() && Loc.Size.hasValue()) {
1066         // FIXME: We may be able to do better in the face of results with mixed
1067         // precision. We don't appear to get them in practice, though, so just
1068         // be conservative.
1069         ThrowOutEverything =
1070             CacheInfo->Size.isPrecise() != Loc.Size.isPrecise() ||
1071             CacheInfo->Size.getValue() < Loc.Size.getValue();
1072       } else {
1073         // For our purposes, unknown size > all others.
1074         ThrowOutEverything = !Loc.Size.hasValue();
1075       }
1076 
1077       if (ThrowOutEverything) {
1078         // The query's Size is greater than the cached one. Throw out the
1079         // cached data and proceed with the query at the greater size.
1080         CacheInfo->Pair = BBSkipFirstBlockPair();
1081         CacheInfo->Size = Loc.Size;
1082         for (auto &Entry : CacheInfo->NonLocalDeps)
1083           if (Instruction *Inst = Entry.getResult().getInst())
1084             RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1085         CacheInfo->NonLocalDeps.clear();
1086         // The cache is cleared (in the above line) so we will have lost
1087         // information about blocks we have already visited. We therefore must
1088         // assume that the cache information is incomplete.
1089         IsIncomplete = true;
1090       } else {
1091         // This query's Size is less than the cached one. Conservatively restart
1092         // the query using the greater size.
1093         return getNonLocalPointerDepFromBB(
1094             QueryInst, Pointer, Loc.getWithNewSize(CacheInfo->Size), isLoad,
1095             StartBB, Result, Visited, SkipFirstBlock, IsIncomplete);
1096       }
1097     }
1098 
1099     // If the query's AATags are inconsistent with the cached one,
1100     // conservatively throw out the cached data and restart the query with
1101     // no tag if needed.
1102     if (CacheInfo->AATags != Loc.AATags) {
1103       if (CacheInfo->AATags) {
1104         CacheInfo->Pair = BBSkipFirstBlockPair();
1105         CacheInfo->AATags = AAMDNodes();
1106         for (auto &Entry : CacheInfo->NonLocalDeps)
1107           if (Instruction *Inst = Entry.getResult().getInst())
1108             RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1109         CacheInfo->NonLocalDeps.clear();
1110         // The cache is cleared (in the above line) so we will have lost
1111         // information about blocks we have already visited. We therefore must
1112         // assume that the cache information is incomplete.
1113         IsIncomplete = true;
1114       }
1115       if (Loc.AATags)
1116         return getNonLocalPointerDepFromBB(
1117             QueryInst, Pointer, Loc.getWithoutAATags(), isLoad, StartBB, Result,
1118             Visited, SkipFirstBlock, IsIncomplete);
1119     }
1120   }
1121 
1122   NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1123 
1124   // If we have valid cached information for exactly the block we are
1125   // investigating, just return it with no recomputation.
1126   // Don't use cached information for invariant loads since it is valid for
1127   // non-invariant loads only.
1128   if (!IsIncomplete && !isInvariantLoad &&
1129       CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1130     // We have a fully cached result for this query then we can just return the
1131     // cached results and populate the visited set.  However, we have to verify
1132     // that we don't already have conflicting results for these blocks.  Check
1133     // to ensure that if a block in the results set is in the visited set that
1134     // it was for the same pointer query.
1135     if (!Visited.empty()) {
1136       for (auto &Entry : *Cache) {
1137         DenseMap<BasicBlock *, Value *>::iterator VI =
1138             Visited.find(Entry.getBB());
1139         if (VI == Visited.end() || VI->second == Pointer.getAddr())
1140           continue;
1141 
1142         // We have a pointer mismatch in a block.  Just return false, saying
1143         // that something was clobbered in this result.  We could also do a
1144         // non-fully cached query, but there is little point in doing this.
1145         return false;
1146       }
1147     }
1148 
1149     Value *Addr = Pointer.getAddr();
1150     for (auto &Entry : *Cache) {
1151       Visited.insert(std::make_pair(Entry.getBB(), Addr));
1152       if (Entry.getResult().isNonLocal()) {
1153         continue;
1154       }
1155 
1156       if (DT.isReachableFromEntry(Entry.getBB())) {
1157         Result.push_back(
1158             NonLocalDepResult(Entry.getBB(), Entry.getResult(), Addr));
1159       }
1160     }
1161     ++NumCacheCompleteNonLocalPtr;
1162     return true;
1163   }
1164 
1165   // Otherwise, either this is a new block, a block with an invalid cache
1166   // pointer or one that we're about to invalidate by putting more info into
1167   // it than its valid cache info.  If empty and not explicitly indicated as
1168   // incomplete, the result will be valid cache info, otherwise it isn't.
1169   //
1170   // Invariant loads don't affect cache in any way thus no need to update
1171   // CacheInfo as well.
1172   if (!isInvariantLoad) {
1173     if (!IsIncomplete && Cache->empty())
1174       CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1175     else
1176       CacheInfo->Pair = BBSkipFirstBlockPair();
1177   }
1178 
1179   SmallVector<BasicBlock *, 32> Worklist;
1180   Worklist.push_back(StartBB);
1181 
1182   // PredList used inside loop.
1183   SmallVector<std::pair<BasicBlock *, PHITransAddr>, 16> PredList;
1184 
1185   // Keep track of the entries that we know are sorted.  Previously cached
1186   // entries will all be sorted.  The entries we add we only sort on demand (we
1187   // don't insert every element into its sorted position).  We know that we
1188   // won't get any reuse from currently inserted values, because we don't
1189   // revisit blocks after we insert info for them.
1190   unsigned NumSortedEntries = Cache->size();
1191   unsigned WorklistEntries = BlockNumberLimit;
1192   bool GotWorklistLimit = false;
1193   LLVM_DEBUG(AssertSorted(*Cache));
1194 
1195   BatchAAResults BatchAA(AA);
1196   while (!Worklist.empty()) {
1197     BasicBlock *BB = Worklist.pop_back_val();
1198 
1199     // If we do process a large number of blocks it becomes very expensive and
1200     // likely it isn't worth worrying about
1201     if (Result.size() > NumResultsLimit) {
1202       // Sort it now (if needed) so that recursive invocations of
1203       // getNonLocalPointerDepFromBB and other routines that could reuse the
1204       // cache value will only see properly sorted cache arrays.
1205       if (Cache && NumSortedEntries != Cache->size()) {
1206         SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1207       }
1208       // Since we bail out, the "Cache" set won't contain all of the
1209       // results for the query.  This is ok (we can still use it to accelerate
1210       // specific block queries) but we can't do the fastpath "return all
1211       // results from the set".  Clear out the indicator for this.
1212       CacheInfo->Pair = BBSkipFirstBlockPair();
1213       return false;
1214     }
1215 
1216     // Skip the first block if we have it.
1217     if (!SkipFirstBlock) {
1218       // Analyze the dependency of *Pointer in FromBB.  See if we already have
1219       // been here.
1220       assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1221 
1222       // Get the dependency info for Pointer in BB.  If we have cached
1223       // information, we will use it, otherwise we compute it.
1224       LLVM_DEBUG(AssertSorted(*Cache, NumSortedEntries));
1225       MemDepResult Dep = getNonLocalInfoForBlock(
1226           QueryInst, Loc, isLoad, BB, Cache, NumSortedEntries, BatchAA);
1227 
1228       // If we got a Def or Clobber, add this to the list of results.
1229       if (!Dep.isNonLocal()) {
1230         if (DT.isReachableFromEntry(BB)) {
1231           Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1232           continue;
1233         }
1234       }
1235     }
1236 
1237     // If 'Pointer' is an instruction defined in this block, then we need to do
1238     // phi translation to change it into a value live in the predecessor block.
1239     // If not, we just add the predecessors to the worklist and scan them with
1240     // the same Pointer.
1241     if (!Pointer.needsPHITranslationFromBlock(BB)) {
1242       SkipFirstBlock = false;
1243       SmallVector<BasicBlock *, 16> NewBlocks;
1244       for (BasicBlock *Pred : PredCache.get(BB)) {
1245         // Verify that we haven't looked at this block yet.
1246         std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes =
1247             Visited.insert(std::make_pair(Pred, Pointer.getAddr()));
1248         if (InsertRes.second) {
1249           // First time we've looked at *PI.
1250           NewBlocks.push_back(Pred);
1251           continue;
1252         }
1253 
1254         // If we have seen this block before, but it was with a different
1255         // pointer then we have a phi translation failure and we have to treat
1256         // this as a clobber.
1257         if (InsertRes.first->second != Pointer.getAddr()) {
1258           // Make sure to clean up the Visited map before continuing on to
1259           // PredTranslationFailure.
1260           for (unsigned i = 0; i < NewBlocks.size(); i++)
1261             Visited.erase(NewBlocks[i]);
1262           goto PredTranslationFailure;
1263         }
1264       }
1265       if (NewBlocks.size() > WorklistEntries) {
1266         // Make sure to clean up the Visited map before continuing on to
1267         // PredTranslationFailure.
1268         for (unsigned i = 0; i < NewBlocks.size(); i++)
1269           Visited.erase(NewBlocks[i]);
1270         GotWorklistLimit = true;
1271         goto PredTranslationFailure;
1272       }
1273       WorklistEntries -= NewBlocks.size();
1274       Worklist.append(NewBlocks.begin(), NewBlocks.end());
1275       continue;
1276     }
1277 
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;
1282 
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();
1291     }
1292     Cache = nullptr;
1293 
1294     PredList.clear();
1295     for (BasicBlock *Pred : PredCache.get(BB)) {
1296       PredList.push_back(std::make_pair(Pred, Pointer));
1297 
1298       // Get the PHI translated pointer in this predecessor.  This can fail if
1299       // not translatable, in which case the getAddr() returns null.
1300       PHITransAddr &PredPointer = PredList.back().second;
1301       Value *PredPtrVal =
1302           PredPointer.translateValue(BB, Pred, &DT, /*MustDominate=*/false);
1303 
1304       // Check to see if we have already visited this pred block with another
1305       // pointer.  If so, we can't do this lookup.  This failure can occur
1306       // with PHI translation when a critical edge exists and the PHI node in
1307       // the successor translates to a pointer value different than the
1308       // pointer the block was first analyzed with.
1309       std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes =
1310           Visited.insert(std::make_pair(Pred, PredPtrVal));
1311 
1312       if (!InsertRes.second) {
1313         // We found the pred; take it off the list of preds to visit.
1314         PredList.pop_back();
1315 
1316         // If the predecessor was visited with PredPtr, then we already did
1317         // the analysis and can ignore it.
1318         if (InsertRes.first->second == PredPtrVal)
1319           continue;
1320 
1321         // Otherwise, the block was previously analyzed with a different
1322         // pointer.  We can't represent the result of this case, so we just
1323         // treat this as a phi translation failure.
1324 
1325         // Make sure to clean up the Visited map before continuing on to
1326         // PredTranslationFailure.
1327         for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1328           Visited.erase(PredList[i].first);
1329 
1330         goto PredTranslationFailure;
1331       }
1332     }
1333 
1334     // Actually process results here; this need to be a separate loop to avoid
1335     // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1336     // any results for.  (getNonLocalPointerDepFromBB will modify our
1337     // datastructures in ways the code after the PredTranslationFailure label
1338     // doesn't expect.)
1339     for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1340       BasicBlock *Pred = PredList[i].first;
1341       PHITransAddr &PredPointer = PredList[i].second;
1342       Value *PredPtrVal = PredPointer.getAddr();
1343 
1344       bool CanTranslate = true;
1345       // If PHI translation was unable to find an available pointer in this
1346       // predecessor, then we have to assume that the pointer is clobbered in
1347       // that predecessor.  We can still do PRE of the load, which would insert
1348       // a computation of the pointer in this predecessor.
1349       if (!PredPtrVal)
1350         CanTranslate = false;
1351 
1352       // FIXME: it is entirely possible that PHI translating will end up with
1353       // the same value.  Consider PHI translating something like:
1354       // X = phi [x, bb1], [y, bb2].  PHI translating for bb1 doesn't *need*
1355       // to recurse here, pedantically speaking.
1356 
1357       // If getNonLocalPointerDepFromBB fails here, that means the cached
1358       // result conflicted with the Visited list; we have to conservatively
1359       // assume it is unknown, but this also does not block PRE of the load.
1360       if (!CanTranslate ||
1361           !getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1362                                       Loc.getWithNewPtr(PredPtrVal), isLoad,
1363                                       Pred, Result, Visited)) {
1364         // Add the entry to the Result list.
1365         NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1366         Result.push_back(Entry);
1367 
1368         // Since we had a phi translation failure, the cache for CacheKey won't
1369         // include all of the entries that we need to immediately satisfy future
1370         // queries.  Mark this in NonLocalPointerDeps by setting the
1371         // BBSkipFirstBlockPair pointer to null.  This requires reuse of the
1372         // cached value to do more work but not miss the phi trans failure.
1373         NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1374         NLPI.Pair = BBSkipFirstBlockPair();
1375         continue;
1376       }
1377     }
1378 
1379     // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1380     CacheInfo = &NonLocalPointerDeps[CacheKey];
1381     Cache = &CacheInfo->NonLocalDeps;
1382     NumSortedEntries = Cache->size();
1383 
1384     // Since we did phi translation, the "Cache" set won't contain all of the
1385     // results for the query.  This is ok (we can still use it to accelerate
1386     // specific block queries) but we can't do the fastpath "return all
1387     // results from the set"  Clear out the indicator for this.
1388     CacheInfo->Pair = BBSkipFirstBlockPair();
1389     SkipFirstBlock = false;
1390     continue;
1391 
1392   PredTranslationFailure:
1393     // The following code is "failure"; we can't produce a sane translation
1394     // for the given block.  It assumes that we haven't modified any of
1395     // our datastructures while processing the current block.
1396 
1397     if (!Cache) {
1398       // Refresh the CacheInfo/Cache pointer if it got invalidated.
1399       CacheInfo = &NonLocalPointerDeps[CacheKey];
1400       Cache = &CacheInfo->NonLocalDeps;
1401       NumSortedEntries = Cache->size();
1402     }
1403 
1404     // Since we failed phi translation, the "Cache" set won't contain all of the
1405     // results for the query.  This is ok (we can still use it to accelerate
1406     // specific block queries) but we can't do the fastpath "return all
1407     // results from the set".  Clear out the indicator for this.
1408     CacheInfo->Pair = BBSkipFirstBlockPair();
1409 
1410     // If *nothing* works, mark the pointer as unknown.
1411     //
1412     // If this is the magic first block, return this as a clobber of the whole
1413     // incoming value.  Since we can't phi translate to one of the predecessors,
1414     // we have to bail out.
1415     if (SkipFirstBlock)
1416       return false;
1417 
1418     // Results of invariant loads are not cached thus no need to update cached
1419     // information.
1420     if (!isInvariantLoad) {
1421       for (NonLocalDepEntry &I : llvm::reverse(*Cache)) {
1422         if (I.getBB() != BB)
1423           continue;
1424 
1425         assert((GotWorklistLimit || I.getResult().isNonLocal() ||
1426                 !DT.isReachableFromEntry(BB)) &&
1427                "Should only be here with transparent block");
1428 
1429         I.setResult(MemDepResult::getUnknown());
1430 
1431 
1432         break;
1433       }
1434     }
1435     (void)GotWorklistLimit;
1436     // Go ahead and report unknown dependence.
1437     Result.push_back(
1438         NonLocalDepResult(BB, MemDepResult::getUnknown(), Pointer.getAddr()));
1439   }
1440 
1441   // Okay, we're done now.  If we added new values to the cache, re-sort it.
1442   SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1443   LLVM_DEBUG(AssertSorted(*Cache));
1444   return true;
1445 }
1446 
1447 /// If P exists in CachedNonLocalPointerInfo or NonLocalDefsCache, remove it.
1448 void MemoryDependenceResults::removeCachedNonLocalPointerDependencies(
1449     ValueIsLoadPair P) {
1450 
1451   // Most of the time this cache is empty.
1452   if (!NonLocalDefsCache.empty()) {
1453     auto it = NonLocalDefsCache.find(P.getPointer());
1454     if (it != NonLocalDefsCache.end()) {
1455       RemoveFromReverseMap(ReverseNonLocalDefsCache,
1456                            it->second.getResult().getInst(), P.getPointer());
1457       NonLocalDefsCache.erase(it);
1458     }
1459 
1460     if (auto *I = dyn_cast<Instruction>(P.getPointer())) {
1461       auto toRemoveIt = ReverseNonLocalDefsCache.find(I);
1462       if (toRemoveIt != ReverseNonLocalDefsCache.end()) {
1463         for (const auto *entry : toRemoveIt->second)
1464           NonLocalDefsCache.erase(entry);
1465         ReverseNonLocalDefsCache.erase(toRemoveIt);
1466       }
1467     }
1468   }
1469 
1470   CachedNonLocalPointerInfo::iterator It = NonLocalPointerDeps.find(P);
1471   if (It == NonLocalPointerDeps.end())
1472     return;
1473 
1474   // Remove all of the entries in the BB->val map.  This involves removing
1475   // instructions from the reverse map.
1476   NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1477 
1478   for (const NonLocalDepEntry &DE : PInfo) {
1479     Instruction *Target = DE.getResult().getInst();
1480     if (!Target)
1481       continue; // Ignore non-local dep results.
1482     assert(Target->getParent() == DE.getBB());
1483 
1484     // Eliminating the dirty entry from 'Cache', so update the reverse info.
1485     RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1486   }
1487 
1488   // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1489   NonLocalPointerDeps.erase(It);
1490 }
1491 
1492 void MemoryDependenceResults::invalidateCachedPointerInfo(Value *Ptr) {
1493   // If Ptr isn't really a pointer, just ignore it.
1494   if (!Ptr->getType()->isPointerTy())
1495     return;
1496   // Flush store info for the pointer.
1497   removeCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1498   // Flush load info for the pointer.
1499   removeCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1500 }
1501 
1502 void MemoryDependenceResults::invalidateCachedPredecessors() {
1503   PredCache.clear();
1504 }
1505 
1506 void MemoryDependenceResults::removeInstruction(Instruction *RemInst) {
1507   // Walk through the Non-local dependencies, removing this one as the value
1508   // for any cached queries.
1509   NonLocalDepMapType::iterator NLDI = NonLocalDepsMap.find(RemInst);
1510   if (NLDI != NonLocalDepsMap.end()) {
1511     NonLocalDepInfo &BlockMap = NLDI->second.first;
1512     for (auto &Entry : BlockMap)
1513       if (Instruction *Inst = Entry.getResult().getInst())
1514         RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1515     NonLocalDepsMap.erase(NLDI);
1516   }
1517 
1518   // If we have a cached local dependence query for this instruction, remove it.
1519   LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1520   if (LocalDepEntry != LocalDeps.end()) {
1521     // Remove us from DepInst's reverse set now that the local dep info is gone.
1522     if (Instruction *Inst = LocalDepEntry->second.getInst())
1523       RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1524 
1525     // Remove this local dependency info.
1526     LocalDeps.erase(LocalDepEntry);
1527   }
1528 
1529   // If we have any cached dependencies on this instruction, remove
1530   // them.
1531 
1532   // If the instruction is a pointer, remove it from both the load info and the
1533   // store info.
1534   if (RemInst->getType()->isPointerTy()) {
1535     removeCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1536     removeCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1537   } else {
1538     // Otherwise, if the instructions is in the map directly, it must be a load.
1539     // Remove it.
1540     auto toRemoveIt = NonLocalDefsCache.find(RemInst);
1541     if (toRemoveIt != NonLocalDefsCache.end()) {
1542       assert(isa<LoadInst>(RemInst) &&
1543              "only load instructions should be added directly");
1544       const Instruction *DepV = toRemoveIt->second.getResult().getInst();
1545       ReverseNonLocalDefsCache.find(DepV)->second.erase(RemInst);
1546       NonLocalDefsCache.erase(toRemoveIt);
1547     }
1548   }
1549 
1550   // Loop over all of the things that depend on the instruction we're removing.
1551   SmallVector<std::pair<Instruction *, Instruction *>, 8> ReverseDepsToAdd;
1552 
1553   // If we find RemInst as a clobber or Def in any of the maps for other values,
1554   // we need to replace its entry with a dirty version of the instruction after
1555   // it.  If RemInst is a terminator, we use a null dirty value.
1556   //
1557   // Using a dirty version of the instruction after RemInst saves having to scan
1558   // the entire block to get to this point.
1559   MemDepResult NewDirtyVal;
1560   if (!RemInst->isTerminator())
1561     NewDirtyVal = MemDepResult::getDirty(&*++RemInst->getIterator());
1562 
1563   ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1564   if (ReverseDepIt != ReverseLocalDeps.end()) {
1565     // RemInst can't be the terminator if it has local stuff depending on it.
1566     assert(!ReverseDepIt->second.empty() && !RemInst->isTerminator() &&
1567            "Nothing can locally depend on a terminator");
1568 
1569     for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1570       assert(InstDependingOnRemInst != RemInst &&
1571              "Already removed our local dep info");
1572 
1573       LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1574 
1575       // Make sure to remember that new things depend on NewDepInst.
1576       assert(NewDirtyVal.getInst() &&
1577              "There is no way something else can have "
1578              "a local dep on this if it is a terminator!");
1579       ReverseDepsToAdd.push_back(
1580           std::make_pair(NewDirtyVal.getInst(), InstDependingOnRemInst));
1581     }
1582 
1583     ReverseLocalDeps.erase(ReverseDepIt);
1584 
1585     // Add new reverse deps after scanning the set, to avoid invalidating the
1586     // 'ReverseDeps' reference.
1587     while (!ReverseDepsToAdd.empty()) {
1588       ReverseLocalDeps[ReverseDepsToAdd.back().first].insert(
1589           ReverseDepsToAdd.back().second);
1590       ReverseDepsToAdd.pop_back();
1591     }
1592   }
1593 
1594   ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1595   if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1596     for (Instruction *I : ReverseDepIt->second) {
1597       assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1598 
1599       PerInstNLInfo &INLD = NonLocalDepsMap[I];
1600       // The information is now dirty!
1601       INLD.second = true;
1602 
1603       for (auto &Entry : INLD.first) {
1604         if (Entry.getResult().getInst() != RemInst)
1605           continue;
1606 
1607         // Convert to a dirty entry for the subsequent instruction.
1608         Entry.setResult(NewDirtyVal);
1609 
1610         if (Instruction *NextI = NewDirtyVal.getInst())
1611           ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1612       }
1613     }
1614 
1615     ReverseNonLocalDeps.erase(ReverseDepIt);
1616 
1617     // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1618     while (!ReverseDepsToAdd.empty()) {
1619       ReverseNonLocalDeps[ReverseDepsToAdd.back().first].insert(
1620           ReverseDepsToAdd.back().second);
1621       ReverseDepsToAdd.pop_back();
1622     }
1623   }
1624 
1625   // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1626   // value in the NonLocalPointerDeps info.
1627   ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1628       ReverseNonLocalPtrDeps.find(RemInst);
1629   if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1630     SmallVector<std::pair<Instruction *, ValueIsLoadPair>, 8>
1631         ReversePtrDepsToAdd;
1632 
1633     for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1634       assert(P.getPointer() != RemInst &&
1635              "Already removed NonLocalPointerDeps info for RemInst");
1636 
1637       NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1638 
1639       // The cache is not valid for any specific block anymore.
1640       NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1641 
1642       // Update any entries for RemInst to use the instruction after it.
1643       for (auto &Entry : NLPDI) {
1644         if (Entry.getResult().getInst() != RemInst)
1645           continue;
1646 
1647         // Convert to a dirty entry for the subsequent instruction.
1648         Entry.setResult(NewDirtyVal);
1649 
1650         if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1651           ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1652       }
1653 
1654       // Re-sort the NonLocalDepInfo.  Changing the dirty entry to its
1655       // subsequent value may invalidate the sortedness.
1656       llvm::sort(NLPDI);
1657     }
1658 
1659     ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1660 
1661     while (!ReversePtrDepsToAdd.empty()) {
1662       ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first].insert(
1663           ReversePtrDepsToAdd.back().second);
1664       ReversePtrDepsToAdd.pop_back();
1665     }
1666   }
1667 
1668   assert(!NonLocalDepsMap.count(RemInst) && "RemInst got reinserted?");
1669   LLVM_DEBUG(verifyRemoved(RemInst));
1670 }
1671 
1672 /// Verify that the specified instruction does not occur in our internal data
1673 /// structures.
1674 ///
1675 /// This function verifies by asserting in debug builds.
1676 void MemoryDependenceResults::verifyRemoved(Instruction *D) const {
1677 #ifndef NDEBUG
1678   for (const auto &DepKV : LocalDeps) {
1679     assert(DepKV.first != D && "Inst occurs in data structures");
1680     assert(DepKV.second.getInst() != D && "Inst occurs in data structures");
1681   }
1682 
1683   for (const auto &DepKV : NonLocalPointerDeps) {
1684     assert(DepKV.first.getPointer() != D && "Inst occurs in NLPD map key");
1685     for (const auto &Entry : DepKV.second.NonLocalDeps)
1686       assert(Entry.getResult().getInst() != D && "Inst occurs as NLPD value");
1687   }
1688 
1689   for (const auto &DepKV : NonLocalDepsMap) {
1690     assert(DepKV.first != D && "Inst occurs in data structures");
1691     const PerInstNLInfo &INLD = DepKV.second;
1692     for (const auto &Entry : INLD.first)
1693       assert(Entry.getResult().getInst() != D &&
1694              "Inst occurs in data structures");
1695   }
1696 
1697   for (const auto &DepKV : ReverseLocalDeps) {
1698     assert(DepKV.first != D && "Inst occurs in data structures");
1699     for (Instruction *Inst : DepKV.second)
1700       assert(Inst != D && "Inst occurs in data structures");
1701   }
1702 
1703   for (const auto &DepKV : ReverseNonLocalDeps) {
1704     assert(DepKV.first != D && "Inst occurs in data structures");
1705     for (Instruction *Inst : DepKV.second)
1706       assert(Inst != D && "Inst occurs in data structures");
1707   }
1708 
1709   for (const auto &DepKV : ReverseNonLocalPtrDeps) {
1710     assert(DepKV.first != D && "Inst occurs in rev NLPD map");
1711 
1712     for (ValueIsLoadPair P : DepKV.second)
1713       assert(P != ValueIsLoadPair(D, false) && P != ValueIsLoadPair(D, true) &&
1714              "Inst occurs in ReverseNonLocalPtrDeps map");
1715   }
1716 #endif
1717 }
1718 
1719 AnalysisKey MemoryDependenceAnalysis::Key;
1720 
1721 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
1722     : DefaultBlockScanLimit(BlockScanLimit) {}
1723 
1724 MemoryDependenceResults
1725 MemoryDependenceAnalysis::run(Function &F, FunctionAnalysisManager &AM) {
1726   auto &AA = AM.getResult<AAManager>(F);
1727   auto &AC = AM.getResult<AssumptionAnalysis>(F);
1728   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1729   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1730   return MemoryDependenceResults(AA, AC, TLI, DT, DefaultBlockScanLimit);
1731 }
1732 
1733 char MemoryDependenceWrapperPass::ID = 0;
1734 
1735 INITIALIZE_PASS_BEGIN(MemoryDependenceWrapperPass, "memdep",
1736                       "Memory Dependence Analysis", false, true)
1737 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1738 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1739 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1740 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1741 INITIALIZE_PASS_END(MemoryDependenceWrapperPass, "memdep",
1742                     "Memory Dependence Analysis", false, true)
1743 
1744 MemoryDependenceWrapperPass::MemoryDependenceWrapperPass() : FunctionPass(ID) {
1745   initializeMemoryDependenceWrapperPassPass(*PassRegistry::getPassRegistry());
1746 }
1747 
1748 MemoryDependenceWrapperPass::~MemoryDependenceWrapperPass() = default;
1749 
1750 void MemoryDependenceWrapperPass::releaseMemory() {
1751   MemDep.reset();
1752 }
1753 
1754 void MemoryDependenceWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1755   AU.setPreservesAll();
1756   AU.addRequired<AssumptionCacheTracker>();
1757   AU.addRequired<DominatorTreeWrapperPass>();
1758   AU.addRequiredTransitive<AAResultsWrapperPass>();
1759   AU.addRequiredTransitive<TargetLibraryInfoWrapperPass>();
1760 }
1761 
1762 bool MemoryDependenceResults::invalidate(Function &F, const PreservedAnalyses &PA,
1763                                FunctionAnalysisManager::Invalidator &Inv) {
1764   // Check whether our analysis is preserved.
1765   auto PAC = PA.getChecker<MemoryDependenceAnalysis>();
1766   if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>())
1767     // If not, give up now.
1768     return true;
1769 
1770   // Check whether the analyses we depend on became invalid for any reason.
1771   if (Inv.invalidate<AAManager>(F, PA) ||
1772       Inv.invalidate<AssumptionAnalysis>(F, PA) ||
1773       Inv.invalidate<DominatorTreeAnalysis>(F, PA))
1774     return true;
1775 
1776   // Otherwise this analysis result remains valid.
1777   return false;
1778 }
1779 
1780 unsigned MemoryDependenceResults::getDefaultBlockScanLimit() const {
1781   return DefaultBlockScanLimit;
1782 }
1783 
1784 bool MemoryDependenceWrapperPass::runOnFunction(Function &F) {
1785   auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1786   auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1787   auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1788   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1789   MemDep.emplace(AA, AC, TLI, DT, BlockScanLimit);
1790   return false;
1791 }
1792