xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/MemorySSA.cpp (revision ebacd8013fe5f7fdf9f6a5b286f6680dd2891036)
1 //===- MemorySSA.cpp - Memory SSA Builder ---------------------------------===//
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 the MemorySSA class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/Analysis/MemorySSA.h"
14 #include "llvm/ADT/DenseMap.h"
15 #include "llvm/ADT/DenseMapInfo.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/DepthFirstIterator.h"
18 #include "llvm/ADT/Hashing.h"
19 #include "llvm/ADT/None.h"
20 #include "llvm/ADT/Optional.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/iterator.h"
26 #include "llvm/ADT/iterator_range.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/CFGPrinter.h"
29 #include "llvm/Analysis/IteratedDominanceFrontier.h"
30 #include "llvm/Analysis/MemoryLocation.h"
31 #include "llvm/Config/llvm-config.h"
32 #include "llvm/IR/AssemblyAnnotationWriter.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/Instruction.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/LLVMContext.h"
40 #include "llvm/IR/Operator.h"
41 #include "llvm/IR/PassManager.h"
42 #include "llvm/IR/Use.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 "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/FormattedStream.h"
52 #include "llvm/Support/GraphWriter.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include <algorithm>
55 #include <cassert>
56 #include <iterator>
57 #include <memory>
58 #include <utility>
59 
60 using namespace llvm;
61 
62 #define DEBUG_TYPE "memoryssa"
63 
64 static cl::opt<std::string>
65     DotCFGMSSA("dot-cfg-mssa",
66                cl::value_desc("file name for generated dot file"),
67                cl::desc("file name for generated dot file"), cl::init(""));
68 
69 INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
70                       true)
71 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
72 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
73 INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
74                     true)
75 
76 INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa",
77                       "Memory SSA Printer", false, false)
78 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
79 INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",
80                     "Memory SSA Printer", false, false)
81 
82 static cl::opt<unsigned> MaxCheckLimit(
83     "memssa-check-limit", cl::Hidden, cl::init(100),
84     cl::desc("The maximum number of stores/phis MemorySSA"
85              "will consider trying to walk past (default = 100)"));
86 
87 // Always verify MemorySSA if expensive checking is enabled.
88 #ifdef EXPENSIVE_CHECKS
89 bool llvm::VerifyMemorySSA = true;
90 #else
91 bool llvm::VerifyMemorySSA = false;
92 #endif
93 
94 static cl::opt<bool, true>
95     VerifyMemorySSAX("verify-memoryssa", cl::location(VerifyMemorySSA),
96                      cl::Hidden, cl::desc("Enable verification of MemorySSA."));
97 
98 const static char LiveOnEntryStr[] = "liveOnEntry";
99 
100 namespace {
101 
102 /// An assembly annotator class to print Memory SSA information in
103 /// comments.
104 class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter {
105   const MemorySSA *MSSA;
106 
107 public:
108   MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {}
109 
110   void emitBasicBlockStartAnnot(const BasicBlock *BB,
111                                 formatted_raw_ostream &OS) override {
112     if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
113       OS << "; " << *MA << "\n";
114   }
115 
116   void emitInstructionAnnot(const Instruction *I,
117                             formatted_raw_ostream &OS) override {
118     if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
119       OS << "; " << *MA << "\n";
120   }
121 };
122 
123 /// An assembly annotator class to print Memory SSA information in
124 /// comments.
125 class MemorySSAWalkerAnnotatedWriter : public AssemblyAnnotationWriter {
126   MemorySSA *MSSA;
127   MemorySSAWalker *Walker;
128 
129 public:
130   MemorySSAWalkerAnnotatedWriter(MemorySSA *M)
131       : MSSA(M), Walker(M->getWalker()) {}
132 
133   void emitBasicBlockStartAnnot(const BasicBlock *BB,
134                                 formatted_raw_ostream &OS) override {
135     if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
136       OS << "; " << *MA << "\n";
137   }
138 
139   void emitInstructionAnnot(const Instruction *I,
140                             formatted_raw_ostream &OS) override {
141     if (MemoryAccess *MA = MSSA->getMemoryAccess(I)) {
142       MemoryAccess *Clobber = Walker->getClobberingMemoryAccess(MA);
143       OS << "; " << *MA;
144       if (Clobber) {
145         OS << " - clobbered by ";
146         if (MSSA->isLiveOnEntryDef(Clobber))
147           OS << LiveOnEntryStr;
148         else
149           OS << *Clobber;
150       }
151       OS << "\n";
152     }
153   }
154 };
155 
156 } // namespace
157 
158 namespace {
159 
160 /// Our current alias analysis API differentiates heavily between calls and
161 /// non-calls, and functions called on one usually assert on the other.
162 /// This class encapsulates the distinction to simplify other code that wants
163 /// "Memory affecting instructions and related data" to use as a key.
164 /// For example, this class is used as a densemap key in the use optimizer.
165 class MemoryLocOrCall {
166 public:
167   bool IsCall = false;
168 
169   MemoryLocOrCall(MemoryUseOrDef *MUD)
170       : MemoryLocOrCall(MUD->getMemoryInst()) {}
171   MemoryLocOrCall(const MemoryUseOrDef *MUD)
172       : MemoryLocOrCall(MUD->getMemoryInst()) {}
173 
174   MemoryLocOrCall(Instruction *Inst) {
175     if (auto *C = dyn_cast<CallBase>(Inst)) {
176       IsCall = true;
177       Call = C;
178     } else {
179       IsCall = false;
180       // There is no such thing as a memorylocation for a fence inst, and it is
181       // unique in that regard.
182       if (!isa<FenceInst>(Inst))
183         Loc = MemoryLocation::get(Inst);
184     }
185   }
186 
187   explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {}
188 
189   const CallBase *getCall() const {
190     assert(IsCall);
191     return Call;
192   }
193 
194   MemoryLocation getLoc() const {
195     assert(!IsCall);
196     return Loc;
197   }
198 
199   bool operator==(const MemoryLocOrCall &Other) const {
200     if (IsCall != Other.IsCall)
201       return false;
202 
203     if (!IsCall)
204       return Loc == Other.Loc;
205 
206     if (Call->getCalledOperand() != Other.Call->getCalledOperand())
207       return false;
208 
209     return Call->arg_size() == Other.Call->arg_size() &&
210            std::equal(Call->arg_begin(), Call->arg_end(),
211                       Other.Call->arg_begin());
212   }
213 
214 private:
215   union {
216     const CallBase *Call;
217     MemoryLocation Loc;
218   };
219 };
220 
221 } // end anonymous namespace
222 
223 namespace llvm {
224 
225 template <> struct DenseMapInfo<MemoryLocOrCall> {
226   static inline MemoryLocOrCall getEmptyKey() {
227     return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
228   }
229 
230   static inline MemoryLocOrCall getTombstoneKey() {
231     return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
232   }
233 
234   static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
235     if (!MLOC.IsCall)
236       return hash_combine(
237           MLOC.IsCall,
238           DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc()));
239 
240     hash_code hash =
241         hash_combine(MLOC.IsCall, DenseMapInfo<const Value *>::getHashValue(
242                                       MLOC.getCall()->getCalledOperand()));
243 
244     for (const Value *Arg : MLOC.getCall()->args())
245       hash = hash_combine(hash, DenseMapInfo<const Value *>::getHashValue(Arg));
246     return hash;
247   }
248 
249   static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
250     return LHS == RHS;
251   }
252 };
253 
254 } // end namespace llvm
255 
256 /// This does one-way checks to see if Use could theoretically be hoisted above
257 /// MayClobber. This will not check the other way around.
258 ///
259 /// This assumes that, for the purposes of MemorySSA, Use comes directly after
260 /// MayClobber, with no potentially clobbering operations in between them.
261 /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
262 static bool areLoadsReorderable(const LoadInst *Use,
263                                 const LoadInst *MayClobber) {
264   bool VolatileUse = Use->isVolatile();
265   bool VolatileClobber = MayClobber->isVolatile();
266   // Volatile operations may never be reordered with other volatile operations.
267   if (VolatileUse && VolatileClobber)
268     return false;
269   // Otherwise, volatile doesn't matter here. From the language reference:
270   // 'optimizers may change the order of volatile operations relative to
271   // non-volatile operations.'"
272 
273   // If a load is seq_cst, it cannot be moved above other loads. If its ordering
274   // is weaker, it can be moved above other loads. We just need to be sure that
275   // MayClobber isn't an acquire load, because loads can't be moved above
276   // acquire loads.
277   //
278   // Note that this explicitly *does* allow the free reordering of monotonic (or
279   // weaker) loads of the same address.
280   bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
281   bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
282                                                      AtomicOrdering::Acquire);
283   return !(SeqCstUse || MayClobberIsAcquire);
284 }
285 
286 namespace {
287 
288 struct ClobberAlias {
289   bool IsClobber;
290   Optional<AliasResult> AR;
291 };
292 
293 } // end anonymous namespace
294 
295 // Return a pair of {IsClobber (bool), AR (AliasResult)}. It relies on AR being
296 // ignored if IsClobber = false.
297 template <typename AliasAnalysisType>
298 static ClobberAlias
299 instructionClobbersQuery(const MemoryDef *MD, const MemoryLocation &UseLoc,
300                          const Instruction *UseInst, AliasAnalysisType &AA) {
301   Instruction *DefInst = MD->getMemoryInst();
302   assert(DefInst && "Defining instruction not actually an instruction");
303   Optional<AliasResult> AR;
304 
305   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
306     // These intrinsics will show up as affecting memory, but they are just
307     // markers, mostly.
308     //
309     // FIXME: We probably don't actually want MemorySSA to model these at all
310     // (including creating MemoryAccesses for them): we just end up inventing
311     // clobbers where they don't really exist at all. Please see D43269 for
312     // context.
313     switch (II->getIntrinsicID()) {
314     case Intrinsic::invariant_start:
315     case Intrinsic::invariant_end:
316     case Intrinsic::assume:
317     case Intrinsic::experimental_noalias_scope_decl:
318     case Intrinsic::pseudoprobe:
319       return {false, AliasResult(AliasResult::NoAlias)};
320     case Intrinsic::dbg_addr:
321     case Intrinsic::dbg_declare:
322     case Intrinsic::dbg_label:
323     case Intrinsic::dbg_value:
324       llvm_unreachable("debuginfo shouldn't have associated defs!");
325     default:
326       break;
327     }
328   }
329 
330   if (auto *CB = dyn_cast_or_null<CallBase>(UseInst)) {
331     ModRefInfo I = AA.getModRefInfo(DefInst, CB);
332     AR = isMustSet(I) ? AliasResult::MustAlias : AliasResult::MayAlias;
333     return {isModOrRefSet(I), AR};
334   }
335 
336   if (auto *DefLoad = dyn_cast<LoadInst>(DefInst))
337     if (auto *UseLoad = dyn_cast_or_null<LoadInst>(UseInst))
338       return {!areLoadsReorderable(UseLoad, DefLoad),
339               AliasResult(AliasResult::MayAlias)};
340 
341   ModRefInfo I = AA.getModRefInfo(DefInst, UseLoc);
342   AR = isMustSet(I) ? AliasResult::MustAlias : AliasResult::MayAlias;
343   return {isModSet(I), AR};
344 }
345 
346 template <typename AliasAnalysisType>
347 static ClobberAlias instructionClobbersQuery(MemoryDef *MD,
348                                              const MemoryUseOrDef *MU,
349                                              const MemoryLocOrCall &UseMLOC,
350                                              AliasAnalysisType &AA) {
351   // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
352   // to exist while MemoryLocOrCall is pushed through places.
353   if (UseMLOC.IsCall)
354     return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(),
355                                     AA);
356   return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
357                                   AA);
358 }
359 
360 // Return true when MD may alias MU, return false otherwise.
361 bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
362                                         AliasAnalysis &AA) {
363   return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA).IsClobber;
364 }
365 
366 namespace {
367 
368 struct UpwardsMemoryQuery {
369   // True if our original query started off as a call
370   bool IsCall = false;
371   // The pointer location we started the query with. This will be empty if
372   // IsCall is true.
373   MemoryLocation StartingLoc;
374   // This is the instruction we were querying about.
375   const Instruction *Inst = nullptr;
376   // The MemoryAccess we actually got called with, used to test local domination
377   const MemoryAccess *OriginalAccess = nullptr;
378   Optional<AliasResult> AR = AliasResult(AliasResult::MayAlias);
379   bool SkipSelfAccess = false;
380 
381   UpwardsMemoryQuery() = default;
382 
383   UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
384       : IsCall(isa<CallBase>(Inst)), Inst(Inst), OriginalAccess(Access) {
385     if (!IsCall)
386       StartingLoc = MemoryLocation::get(Inst);
387   }
388 };
389 
390 } // end anonymous namespace
391 
392 template <typename AliasAnalysisType>
393 static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysisType &AA,
394                                                    const Instruction *I) {
395   // If the memory can't be changed, then loads of the memory can't be
396   // clobbered.
397   if (auto *LI = dyn_cast<LoadInst>(I))
398     return I->hasMetadata(LLVMContext::MD_invariant_load) ||
399            AA.pointsToConstantMemory(MemoryLocation::get(LI));
400   return false;
401 }
402 
403 /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
404 /// inbetween `Start` and `ClobberAt` can clobbers `Start`.
405 ///
406 /// This is meant to be as simple and self-contained as possible. Because it
407 /// uses no cache, etc., it can be relatively expensive.
408 ///
409 /// \param Start     The MemoryAccess that we want to walk from.
410 /// \param ClobberAt A clobber for Start.
411 /// \param StartLoc  The MemoryLocation for Start.
412 /// \param MSSA      The MemorySSA instance that Start and ClobberAt belong to.
413 /// \param Query     The UpwardsMemoryQuery we used for our search.
414 /// \param AA        The AliasAnalysis we used for our search.
415 /// \param AllowImpreciseClobber Always false, unless we do relaxed verify.
416 
417 template <typename AliasAnalysisType>
418 LLVM_ATTRIBUTE_UNUSED static void
419 checkClobberSanity(const MemoryAccess *Start, MemoryAccess *ClobberAt,
420                    const MemoryLocation &StartLoc, const MemorySSA &MSSA,
421                    const UpwardsMemoryQuery &Query, AliasAnalysisType &AA,
422                    bool AllowImpreciseClobber = false) {
423   assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?");
424 
425   if (MSSA.isLiveOnEntryDef(Start)) {
426     assert(MSSA.isLiveOnEntryDef(ClobberAt) &&
427            "liveOnEntry must clobber itself");
428     return;
429   }
430 
431   bool FoundClobber = false;
432   DenseSet<ConstMemoryAccessPair> VisitedPhis;
433   SmallVector<ConstMemoryAccessPair, 8> Worklist;
434   Worklist.emplace_back(Start, StartLoc);
435   // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
436   // is found, complain.
437   while (!Worklist.empty()) {
438     auto MAP = Worklist.pop_back_val();
439     // All we care about is that nothing from Start to ClobberAt clobbers Start.
440     // We learn nothing from revisiting nodes.
441     if (!VisitedPhis.insert(MAP).second)
442       continue;
443 
444     for (const auto *MA : def_chain(MAP.first)) {
445       if (MA == ClobberAt) {
446         if (const auto *MD = dyn_cast<MemoryDef>(MA)) {
447           // instructionClobbersQuery isn't essentially free, so don't use `|=`,
448           // since it won't let us short-circuit.
449           //
450           // Also, note that this can't be hoisted out of the `Worklist` loop,
451           // since MD may only act as a clobber for 1 of N MemoryLocations.
452           FoundClobber = FoundClobber || MSSA.isLiveOnEntryDef(MD);
453           if (!FoundClobber) {
454             ClobberAlias CA =
455                 instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
456             if (CA.IsClobber) {
457               FoundClobber = true;
458               // Not used: CA.AR;
459             }
460           }
461         }
462         break;
463       }
464 
465       // We should never hit liveOnEntry, unless it's the clobber.
466       assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?");
467 
468       if (const auto *MD = dyn_cast<MemoryDef>(MA)) {
469         // If Start is a Def, skip self.
470         if (MD == Start)
471           continue;
472 
473         assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA)
474                     .IsClobber &&
475                "Found clobber before reaching ClobberAt!");
476         continue;
477       }
478 
479       if (const auto *MU = dyn_cast<MemoryUse>(MA)) {
480         (void)MU;
481         assert (MU == Start &&
482                 "Can only find use in def chain if Start is a use");
483         continue;
484       }
485 
486       assert(isa<MemoryPhi>(MA));
487 
488       // Add reachable phi predecessors
489       for (auto ItB = upward_defs_begin(
490                     {const_cast<MemoryAccess *>(MA), MAP.second},
491                     MSSA.getDomTree()),
492                 ItE = upward_defs_end();
493            ItB != ItE; ++ItB)
494         if (MSSA.getDomTree().isReachableFromEntry(ItB.getPhiArgBlock()))
495           Worklist.emplace_back(*ItB);
496     }
497   }
498 
499   // If the verify is done following an optimization, it's possible that
500   // ClobberAt was a conservative clobbering, that we can now infer is not a
501   // true clobbering access. Don't fail the verify if that's the case.
502   // We do have accesses that claim they're optimized, but could be optimized
503   // further. Updating all these can be expensive, so allow it for now (FIXME).
504   if (AllowImpreciseClobber)
505     return;
506 
507   // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
508   // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
509   assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
510          "ClobberAt never acted as a clobber");
511 }
512 
513 namespace {
514 
515 /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
516 /// in one class.
517 template <class AliasAnalysisType> class ClobberWalker {
518   /// Save a few bytes by using unsigned instead of size_t.
519   using ListIndex = unsigned;
520 
521   /// Represents a span of contiguous MemoryDefs, potentially ending in a
522   /// MemoryPhi.
523   struct DefPath {
524     MemoryLocation Loc;
525     // Note that, because we always walk in reverse, Last will always dominate
526     // First. Also note that First and Last are inclusive.
527     MemoryAccess *First;
528     MemoryAccess *Last;
529     Optional<ListIndex> Previous;
530 
531     DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
532             Optional<ListIndex> Previous)
533         : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
534 
535     DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
536             Optional<ListIndex> Previous)
537         : DefPath(Loc, Init, Init, Previous) {}
538   };
539 
540   const MemorySSA &MSSA;
541   AliasAnalysisType &AA;
542   DominatorTree &DT;
543   UpwardsMemoryQuery *Query;
544   unsigned *UpwardWalkLimit;
545 
546   // Phi optimization bookkeeping:
547   // List of DefPath to process during the current phi optimization walk.
548   SmallVector<DefPath, 32> Paths;
549   // List of visited <Access, Location> pairs; we can skip paths already
550   // visited with the same memory location.
551   DenseSet<ConstMemoryAccessPair> VisitedPhis;
552   // Record if phi translation has been performed during the current phi
553   // optimization walk, as merging alias results after phi translation can
554   // yield incorrect results. Context in PR46156.
555   bool PerformedPhiTranslation = false;
556 
557   /// Find the nearest def or phi that `From` can legally be optimized to.
558   const MemoryAccess *getWalkTarget(const MemoryPhi *From) const {
559     assert(From->getNumOperands() && "Phi with no operands?");
560 
561     BasicBlock *BB = From->getBlock();
562     MemoryAccess *Result = MSSA.getLiveOnEntryDef();
563     DomTreeNode *Node = DT.getNode(BB);
564     while ((Node = Node->getIDom())) {
565       auto *Defs = MSSA.getBlockDefs(Node->getBlock());
566       if (Defs)
567         return &*Defs->rbegin();
568     }
569     return Result;
570   }
571 
572   /// Result of calling walkToPhiOrClobber.
573   struct UpwardsWalkResult {
574     /// The "Result" of the walk. Either a clobber, the last thing we walked, or
575     /// both. Include alias info when clobber found.
576     MemoryAccess *Result;
577     bool IsKnownClobber;
578     Optional<AliasResult> AR;
579   };
580 
581   /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
582   /// This will update Desc.Last as it walks. It will (optionally) also stop at
583   /// StopAt.
584   ///
585   /// This does not test for whether StopAt is a clobber
586   UpwardsWalkResult
587   walkToPhiOrClobber(DefPath &Desc, const MemoryAccess *StopAt = nullptr,
588                      const MemoryAccess *SkipStopAt = nullptr) const {
589     assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world");
590     assert(UpwardWalkLimit && "Need a valid walk limit");
591     bool LimitAlreadyReached = false;
592     // (*UpwardWalkLimit) may be 0 here, due to the loop in tryOptimizePhi. Set
593     // it to 1. This will not do any alias() calls. It either returns in the
594     // first iteration in the loop below, or is set back to 0 if all def chains
595     // are free of MemoryDefs.
596     if (!*UpwardWalkLimit) {
597       *UpwardWalkLimit = 1;
598       LimitAlreadyReached = true;
599     }
600 
601     for (MemoryAccess *Current : def_chain(Desc.Last)) {
602       Desc.Last = Current;
603       if (Current == StopAt || Current == SkipStopAt)
604         return {Current, false, AliasResult(AliasResult::MayAlias)};
605 
606       if (auto *MD = dyn_cast<MemoryDef>(Current)) {
607         if (MSSA.isLiveOnEntryDef(MD))
608           return {MD, true, AliasResult(AliasResult::MustAlias)};
609 
610         if (!--*UpwardWalkLimit)
611           return {Current, true, AliasResult(AliasResult::MayAlias)};
612 
613         ClobberAlias CA =
614             instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA);
615         if (CA.IsClobber)
616           return {MD, true, CA.AR};
617       }
618     }
619 
620     if (LimitAlreadyReached)
621       *UpwardWalkLimit = 0;
622 
623     assert(isa<MemoryPhi>(Desc.Last) &&
624            "Ended at a non-clobber that's not a phi?");
625     return {Desc.Last, false, AliasResult(AliasResult::MayAlias)};
626   }
627 
628   void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
629                    ListIndex PriorNode) {
630     auto UpwardDefsBegin = upward_defs_begin({Phi, Paths[PriorNode].Loc}, DT,
631                                              &PerformedPhiTranslation);
632     auto UpwardDefs = make_range(UpwardDefsBegin, upward_defs_end());
633     for (const MemoryAccessPair &P : UpwardDefs) {
634       PausedSearches.push_back(Paths.size());
635       Paths.emplace_back(P.second, P.first, PriorNode);
636     }
637   }
638 
639   /// Represents a search that terminated after finding a clobber. This clobber
640   /// may or may not be present in the path of defs from LastNode..SearchStart,
641   /// since it may have been retrieved from cache.
642   struct TerminatedPath {
643     MemoryAccess *Clobber;
644     ListIndex LastNode;
645   };
646 
647   /// Get an access that keeps us from optimizing to the given phi.
648   ///
649   /// PausedSearches is an array of indices into the Paths array. Its incoming
650   /// value is the indices of searches that stopped at the last phi optimization
651   /// target. It's left in an unspecified state.
652   ///
653   /// If this returns None, NewPaused is a vector of searches that terminated
654   /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
655   Optional<TerminatedPath>
656   getBlockingAccess(const MemoryAccess *StopWhere,
657                     SmallVectorImpl<ListIndex> &PausedSearches,
658                     SmallVectorImpl<ListIndex> &NewPaused,
659                     SmallVectorImpl<TerminatedPath> &Terminated) {
660     assert(!PausedSearches.empty() && "No searches to continue?");
661 
662     // BFS vs DFS really doesn't make a difference here, so just do a DFS with
663     // PausedSearches as our stack.
664     while (!PausedSearches.empty()) {
665       ListIndex PathIndex = PausedSearches.pop_back_val();
666       DefPath &Node = Paths[PathIndex];
667 
668       // If we've already visited this path with this MemoryLocation, we don't
669       // need to do so again.
670       //
671       // NOTE: That we just drop these paths on the ground makes caching
672       // behavior sporadic. e.g. given a diamond:
673       //  A
674       // B C
675       //  D
676       //
677       // ...If we walk D, B, A, C, we'll only cache the result of phi
678       // optimization for A, B, and D; C will be skipped because it dies here.
679       // This arguably isn't the worst thing ever, since:
680       //   - We generally query things in a top-down order, so if we got below D
681       //     without needing cache entries for {C, MemLoc}, then chances are
682       //     that those cache entries would end up ultimately unused.
683       //   - We still cache things for A, so C only needs to walk up a bit.
684       // If this behavior becomes problematic, we can fix without a ton of extra
685       // work.
686       if (!VisitedPhis.insert({Node.Last, Node.Loc}).second) {
687         if (PerformedPhiTranslation) {
688           // If visiting this path performed Phi translation, don't continue,
689           // since it may not be correct to merge results from two paths if one
690           // relies on the phi translation.
691           TerminatedPath Term{Node.Last, PathIndex};
692           return Term;
693         }
694         continue;
695       }
696 
697       const MemoryAccess *SkipStopWhere = nullptr;
698       if (Query->SkipSelfAccess && Node.Loc == Query->StartingLoc) {
699         assert(isa<MemoryDef>(Query->OriginalAccess));
700         SkipStopWhere = Query->OriginalAccess;
701       }
702 
703       UpwardsWalkResult Res = walkToPhiOrClobber(Node,
704                                                  /*StopAt=*/StopWhere,
705                                                  /*SkipStopAt=*/SkipStopWhere);
706       if (Res.IsKnownClobber) {
707         assert(Res.Result != StopWhere && Res.Result != SkipStopWhere);
708 
709         // If this wasn't a cache hit, we hit a clobber when walking. That's a
710         // failure.
711         TerminatedPath Term{Res.Result, PathIndex};
712         if (!MSSA.dominates(Res.Result, StopWhere))
713           return Term;
714 
715         // Otherwise, it's a valid thing to potentially optimize to.
716         Terminated.push_back(Term);
717         continue;
718       }
719 
720       if (Res.Result == StopWhere || Res.Result == SkipStopWhere) {
721         // We've hit our target. Save this path off for if we want to continue
722         // walking. If we are in the mode of skipping the OriginalAccess, and
723         // we've reached back to the OriginalAccess, do not save path, we've
724         // just looped back to self.
725         if (Res.Result != SkipStopWhere)
726           NewPaused.push_back(PathIndex);
727         continue;
728       }
729 
730       assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber");
731       addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
732     }
733 
734     return None;
735   }
736 
737   template <typename T, typename Walker>
738   struct generic_def_path_iterator
739       : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
740                                     std::forward_iterator_tag, T *> {
741     generic_def_path_iterator() = default;
742     generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
743 
744     T &operator*() const { return curNode(); }
745 
746     generic_def_path_iterator &operator++() {
747       N = curNode().Previous;
748       return *this;
749     }
750 
751     bool operator==(const generic_def_path_iterator &O) const {
752       if (N.has_value() != O.N.has_value())
753         return false;
754       return !N || *N == *O.N;
755     }
756 
757   private:
758     T &curNode() const { return W->Paths[*N]; }
759 
760     Walker *W = nullptr;
761     Optional<ListIndex> N = None;
762   };
763 
764   using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
765   using const_def_path_iterator =
766       generic_def_path_iterator<const DefPath, const ClobberWalker>;
767 
768   iterator_range<def_path_iterator> def_path(ListIndex From) {
769     return make_range(def_path_iterator(this, From), def_path_iterator());
770   }
771 
772   iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
773     return make_range(const_def_path_iterator(this, From),
774                       const_def_path_iterator());
775   }
776 
777   struct OptznResult {
778     /// The path that contains our result.
779     TerminatedPath PrimaryClobber;
780     /// The paths that we can legally cache back from, but that aren't
781     /// necessarily the result of the Phi optimization.
782     SmallVector<TerminatedPath, 4> OtherClobbers;
783   };
784 
785   ListIndex defPathIndex(const DefPath &N) const {
786     // The assert looks nicer if we don't need to do &N
787     const DefPath *NP = &N;
788     assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&
789            "Out of bounds DefPath!");
790     return NP - &Paths.front();
791   }
792 
793   /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
794   /// that act as legal clobbers. Note that this won't return *all* clobbers.
795   ///
796   /// Phi optimization algorithm tl;dr:
797   ///   - Find the earliest def/phi, A, we can optimize to
798   ///   - Find if all paths from the starting memory access ultimately reach A
799   ///     - If not, optimization isn't possible.
800   ///     - Otherwise, walk from A to another clobber or phi, A'.
801   ///       - If A' is a def, we're done.
802   ///       - If A' is a phi, try to optimize it.
803   ///
804   /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
805   /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
806   OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
807                              const MemoryLocation &Loc) {
808     assert(Paths.empty() && VisitedPhis.empty() && !PerformedPhiTranslation &&
809            "Reset the optimization state.");
810 
811     Paths.emplace_back(Loc, Start, Phi, None);
812     // Stores how many "valid" optimization nodes we had prior to calling
813     // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
814     auto PriorPathsSize = Paths.size();
815 
816     SmallVector<ListIndex, 16> PausedSearches;
817     SmallVector<ListIndex, 8> NewPaused;
818     SmallVector<TerminatedPath, 4> TerminatedPaths;
819 
820     addSearches(Phi, PausedSearches, 0);
821 
822     // Moves the TerminatedPath with the "most dominated" Clobber to the end of
823     // Paths.
824     auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
825       assert(!Paths.empty() && "Need a path to move");
826       auto Dom = Paths.begin();
827       for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
828         if (!MSSA.dominates(I->Clobber, Dom->Clobber))
829           Dom = I;
830       auto Last = Paths.end() - 1;
831       if (Last != Dom)
832         std::iter_swap(Last, Dom);
833     };
834 
835     MemoryPhi *Current = Phi;
836     while (true) {
837       assert(!MSSA.isLiveOnEntryDef(Current) &&
838              "liveOnEntry wasn't treated as a clobber?");
839 
840       const auto *Target = getWalkTarget(Current);
841       // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
842       // optimization for the prior phi.
843       assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {
844         return MSSA.dominates(P.Clobber, Target);
845       }));
846 
847       // FIXME: This is broken, because the Blocker may be reported to be
848       // liveOnEntry, and we'll happily wait for that to disappear (read: never)
849       // For the moment, this is fine, since we do nothing with blocker info.
850       if (Optional<TerminatedPath> Blocker = getBlockingAccess(
851               Target, PausedSearches, NewPaused, TerminatedPaths)) {
852 
853         // Find the node we started at. We can't search based on N->Last, since
854         // we may have gone around a loop with a different MemoryLocation.
855         auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
856           return defPathIndex(N) < PriorPathsSize;
857         });
858         assert(Iter != def_path_iterator());
859 
860         DefPath &CurNode = *Iter;
861         assert(CurNode.Last == Current);
862 
863         // Two things:
864         // A. We can't reliably cache all of NewPaused back. Consider a case
865         //    where we have two paths in NewPaused; one of which can't optimize
866         //    above this phi, whereas the other can. If we cache the second path
867         //    back, we'll end up with suboptimal cache entries. We can handle
868         //    cases like this a bit better when we either try to find all
869         //    clobbers that block phi optimization, or when our cache starts
870         //    supporting unfinished searches.
871         // B. We can't reliably cache TerminatedPaths back here without doing
872         //    extra checks; consider a case like:
873         //       T
874         //      / \
875         //     D   C
876         //      \ /
877         //       S
878         //    Where T is our target, C is a node with a clobber on it, D is a
879         //    diamond (with a clobber *only* on the left or right node, N), and
880         //    S is our start. Say we walk to D, through the node opposite N
881         //    (read: ignoring the clobber), and see a cache entry in the top
882         //    node of D. That cache entry gets put into TerminatedPaths. We then
883         //    walk up to C (N is later in our worklist), find the clobber, and
884         //    quit. If we append TerminatedPaths to OtherClobbers, we'll cache
885         //    the bottom part of D to the cached clobber, ignoring the clobber
886         //    in N. Again, this problem goes away if we start tracking all
887         //    blockers for a given phi optimization.
888         TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
889         return {Result, {}};
890       }
891 
892       // If there's nothing left to search, then all paths led to valid clobbers
893       // that we got from our cache; pick the nearest to the start, and allow
894       // the rest to be cached back.
895       if (NewPaused.empty()) {
896         MoveDominatedPathToEnd(TerminatedPaths);
897         TerminatedPath Result = TerminatedPaths.pop_back_val();
898         return {Result, std::move(TerminatedPaths)};
899       }
900 
901       MemoryAccess *DefChainEnd = nullptr;
902       SmallVector<TerminatedPath, 4> Clobbers;
903       for (ListIndex Paused : NewPaused) {
904         UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
905         if (WR.IsKnownClobber)
906           Clobbers.push_back({WR.Result, Paused});
907         else
908           // Micro-opt: If we hit the end of the chain, save it.
909           DefChainEnd = WR.Result;
910       }
911 
912       if (!TerminatedPaths.empty()) {
913         // If we couldn't find the dominating phi/liveOnEntry in the above loop,
914         // do it now.
915         if (!DefChainEnd)
916           for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target)))
917             DefChainEnd = MA;
918         assert(DefChainEnd && "Failed to find dominating phi/liveOnEntry");
919 
920         // If any of the terminated paths don't dominate the phi we'll try to
921         // optimize, we need to figure out what they are and quit.
922         const BasicBlock *ChainBB = DefChainEnd->getBlock();
923         for (const TerminatedPath &TP : TerminatedPaths) {
924           // Because we know that DefChainEnd is as "high" as we can go, we
925           // don't need local dominance checks; BB dominance is sufficient.
926           if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
927             Clobbers.push_back(TP);
928         }
929       }
930 
931       // If we have clobbers in the def chain, find the one closest to Current
932       // and quit.
933       if (!Clobbers.empty()) {
934         MoveDominatedPathToEnd(Clobbers);
935         TerminatedPath Result = Clobbers.pop_back_val();
936         return {Result, std::move(Clobbers)};
937       }
938 
939       assert(all_of(NewPaused,
940                     [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }));
941 
942       // Because liveOnEntry is a clobber, this must be a phi.
943       auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
944 
945       PriorPathsSize = Paths.size();
946       PausedSearches.clear();
947       for (ListIndex I : NewPaused)
948         addSearches(DefChainPhi, PausedSearches, I);
949       NewPaused.clear();
950 
951       Current = DefChainPhi;
952     }
953   }
954 
955   void verifyOptResult(const OptznResult &R) const {
956     assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {
957       return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);
958     }));
959   }
960 
961   void resetPhiOptznState() {
962     Paths.clear();
963     VisitedPhis.clear();
964     PerformedPhiTranslation = false;
965   }
966 
967 public:
968   ClobberWalker(const MemorySSA &MSSA, AliasAnalysisType &AA, DominatorTree &DT)
969       : MSSA(MSSA), AA(AA), DT(DT) {}
970 
971   AliasAnalysisType *getAA() { return &AA; }
972   /// Finds the nearest clobber for the given query, optimizing phis if
973   /// possible.
974   MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q,
975                             unsigned &UpWalkLimit) {
976     Query = &Q;
977     UpwardWalkLimit = &UpWalkLimit;
978     // Starting limit must be > 0.
979     if (!UpWalkLimit)
980       UpWalkLimit++;
981 
982     MemoryAccess *Current = Start;
983     // This walker pretends uses don't exist. If we're handed one, silently grab
984     // its def. (This has the nice side-effect of ensuring we never cache uses)
985     if (auto *MU = dyn_cast<MemoryUse>(Start))
986       Current = MU->getDefiningAccess();
987 
988     DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
989     // Fast path for the overly-common case (no crazy phi optimization
990     // necessary)
991     UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
992     MemoryAccess *Result;
993     if (WalkResult.IsKnownClobber) {
994       Result = WalkResult.Result;
995       Q.AR = WalkResult.AR;
996     } else {
997       OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
998                                           Current, Q.StartingLoc);
999       verifyOptResult(OptRes);
1000       resetPhiOptznState();
1001       Result = OptRes.PrimaryClobber.Clobber;
1002     }
1003 
1004 #ifdef EXPENSIVE_CHECKS
1005     if (!Q.SkipSelfAccess && *UpwardWalkLimit > 0)
1006       checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
1007 #endif
1008     return Result;
1009   }
1010 };
1011 
1012 struct RenamePassData {
1013   DomTreeNode *DTN;
1014   DomTreeNode::const_iterator ChildIt;
1015   MemoryAccess *IncomingVal;
1016 
1017   RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
1018                  MemoryAccess *M)
1019       : DTN(D), ChildIt(It), IncomingVal(M) {}
1020 
1021   void swap(RenamePassData &RHS) {
1022     std::swap(DTN, RHS.DTN);
1023     std::swap(ChildIt, RHS.ChildIt);
1024     std::swap(IncomingVal, RHS.IncomingVal);
1025   }
1026 };
1027 
1028 } // end anonymous namespace
1029 
1030 namespace llvm {
1031 
1032 template <class AliasAnalysisType> class MemorySSA::ClobberWalkerBase {
1033   ClobberWalker<AliasAnalysisType> Walker;
1034   MemorySSA *MSSA;
1035 
1036 public:
1037   ClobberWalkerBase(MemorySSA *M, AliasAnalysisType *A, DominatorTree *D)
1038       : Walker(*M, *A, *D), MSSA(M) {}
1039 
1040   MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *,
1041                                               const MemoryLocation &,
1042                                               unsigned &);
1043   // Third argument (bool), defines whether the clobber search should skip the
1044   // original queried access. If true, there will be a follow-up query searching
1045   // for a clobber access past "self". Note that the Optimized access is not
1046   // updated if a new clobber is found by this SkipSelf search. If this
1047   // additional query becomes heavily used we may decide to cache the result.
1048   // Walker instantiations will decide how to set the SkipSelf bool.
1049   MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, unsigned &, bool,
1050                                               bool UseInvariantGroup = true);
1051 };
1052 
1053 /// A MemorySSAWalker that does AA walks to disambiguate accesses. It no
1054 /// longer does caching on its own, but the name has been retained for the
1055 /// moment.
1056 template <class AliasAnalysisType>
1057 class MemorySSA::CachingWalker final : public MemorySSAWalker {
1058   ClobberWalkerBase<AliasAnalysisType> *Walker;
1059 
1060 public:
1061   CachingWalker(MemorySSA *M, ClobberWalkerBase<AliasAnalysisType> *W)
1062       : MemorySSAWalker(M), Walker(W) {}
1063   ~CachingWalker() override = default;
1064 
1065   using MemorySSAWalker::getClobberingMemoryAccess;
1066 
1067   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, unsigned &UWL) {
1068     return Walker->getClobberingMemoryAccessBase(MA, UWL, false);
1069   }
1070   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA,
1071                                           const MemoryLocation &Loc,
1072                                           unsigned &UWL) {
1073     return Walker->getClobberingMemoryAccessBase(MA, Loc, UWL);
1074   }
1075   // This method is not accessible outside of this file.
1076   MemoryAccess *getClobberingMemoryAccessWithoutInvariantGroup(MemoryAccess *MA,
1077                                                                unsigned &UWL) {
1078     return Walker->getClobberingMemoryAccessBase(MA, UWL, false, false);
1079   }
1080 
1081   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA) override {
1082     unsigned UpwardWalkLimit = MaxCheckLimit;
1083     return getClobberingMemoryAccess(MA, UpwardWalkLimit);
1084   }
1085   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA,
1086                                           const MemoryLocation &Loc) override {
1087     unsigned UpwardWalkLimit = MaxCheckLimit;
1088     return getClobberingMemoryAccess(MA, Loc, UpwardWalkLimit);
1089   }
1090 
1091   void invalidateInfo(MemoryAccess *MA) override {
1092     if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1093       MUD->resetOptimized();
1094   }
1095 };
1096 
1097 template <class AliasAnalysisType>
1098 class MemorySSA::SkipSelfWalker final : public MemorySSAWalker {
1099   ClobberWalkerBase<AliasAnalysisType> *Walker;
1100 
1101 public:
1102   SkipSelfWalker(MemorySSA *M, ClobberWalkerBase<AliasAnalysisType> *W)
1103       : MemorySSAWalker(M), Walker(W) {}
1104   ~SkipSelfWalker() override = default;
1105 
1106   using MemorySSAWalker::getClobberingMemoryAccess;
1107 
1108   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, unsigned &UWL) {
1109     return Walker->getClobberingMemoryAccessBase(MA, UWL, true);
1110   }
1111   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA,
1112                                           const MemoryLocation &Loc,
1113                                           unsigned &UWL) {
1114     return Walker->getClobberingMemoryAccessBase(MA, Loc, UWL);
1115   }
1116 
1117   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA) override {
1118     unsigned UpwardWalkLimit = MaxCheckLimit;
1119     return getClobberingMemoryAccess(MA, UpwardWalkLimit);
1120   }
1121   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA,
1122                                           const MemoryLocation &Loc) override {
1123     unsigned UpwardWalkLimit = MaxCheckLimit;
1124     return getClobberingMemoryAccess(MA, Loc, UpwardWalkLimit);
1125   }
1126 
1127   void invalidateInfo(MemoryAccess *MA) override {
1128     if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1129       MUD->resetOptimized();
1130   }
1131 };
1132 
1133 } // end namespace llvm
1134 
1135 void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal,
1136                                     bool RenameAllUses) {
1137   // Pass through values to our successors
1138   for (const BasicBlock *S : successors(BB)) {
1139     auto It = PerBlockAccesses.find(S);
1140     // Rename the phi nodes in our successor block
1141     if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1142       continue;
1143     AccessList *Accesses = It->second.get();
1144     auto *Phi = cast<MemoryPhi>(&Accesses->front());
1145     if (RenameAllUses) {
1146       bool ReplacementDone = false;
1147       for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
1148         if (Phi->getIncomingBlock(I) == BB) {
1149           Phi->setIncomingValue(I, IncomingVal);
1150           ReplacementDone = true;
1151         }
1152       (void) ReplacementDone;
1153       assert(ReplacementDone && "Incomplete phi during partial rename");
1154     } else
1155       Phi->addIncoming(IncomingVal, BB);
1156   }
1157 }
1158 
1159 /// Rename a single basic block into MemorySSA form.
1160 /// Uses the standard SSA renaming algorithm.
1161 /// \returns The new incoming value.
1162 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal,
1163                                      bool RenameAllUses) {
1164   auto It = PerBlockAccesses.find(BB);
1165   // Skip most processing if the list is empty.
1166   if (It != PerBlockAccesses.end()) {
1167     AccessList *Accesses = It->second.get();
1168     for (MemoryAccess &L : *Accesses) {
1169       if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
1170         if (MUD->getDefiningAccess() == nullptr || RenameAllUses)
1171           MUD->setDefiningAccess(IncomingVal);
1172         if (isa<MemoryDef>(&L))
1173           IncomingVal = &L;
1174       } else {
1175         IncomingVal = &L;
1176       }
1177     }
1178   }
1179   return IncomingVal;
1180 }
1181 
1182 /// This is the standard SSA renaming algorithm.
1183 ///
1184 /// We walk the dominator tree in preorder, renaming accesses, and then filling
1185 /// in phi nodes in our successors.
1186 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
1187                            SmallPtrSetImpl<BasicBlock *> &Visited,
1188                            bool SkipVisited, bool RenameAllUses) {
1189   assert(Root && "Trying to rename accesses in an unreachable block");
1190 
1191   SmallVector<RenamePassData, 32> WorkStack;
1192   // Skip everything if we already renamed this block and we are skipping.
1193   // Note: You can't sink this into the if, because we need it to occur
1194   // regardless of whether we skip blocks or not.
1195   bool AlreadyVisited = !Visited.insert(Root->getBlock()).second;
1196   if (SkipVisited && AlreadyVisited)
1197     return;
1198 
1199   IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses);
1200   renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses);
1201   WorkStack.push_back({Root, Root->begin(), IncomingVal});
1202 
1203   while (!WorkStack.empty()) {
1204     DomTreeNode *Node = WorkStack.back().DTN;
1205     DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
1206     IncomingVal = WorkStack.back().IncomingVal;
1207 
1208     if (ChildIt == Node->end()) {
1209       WorkStack.pop_back();
1210     } else {
1211       DomTreeNode *Child = *ChildIt;
1212       ++WorkStack.back().ChildIt;
1213       BasicBlock *BB = Child->getBlock();
1214       // Note: You can't sink this into the if, because we need it to occur
1215       // regardless of whether we skip blocks or not.
1216       AlreadyVisited = !Visited.insert(BB).second;
1217       if (SkipVisited && AlreadyVisited) {
1218         // We already visited this during our renaming, which can happen when
1219         // being asked to rename multiple blocks. Figure out the incoming val,
1220         // which is the last def.
1221         // Incoming value can only change if there is a block def, and in that
1222         // case, it's the last block def in the list.
1223         if (auto *BlockDefs = getWritableBlockDefs(BB))
1224           IncomingVal = &*BlockDefs->rbegin();
1225       } else
1226         IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses);
1227       renameSuccessorPhis(BB, IncomingVal, RenameAllUses);
1228       WorkStack.push_back({Child, Child->begin(), IncomingVal});
1229     }
1230   }
1231 }
1232 
1233 /// This handles unreachable block accesses by deleting phi nodes in
1234 /// unreachable blocks, and marking all other unreachable MemoryAccess's as
1235 /// being uses of the live on entry definition.
1236 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
1237   assert(!DT->isReachableFromEntry(BB) &&
1238          "Reachable block found while handling unreachable blocks");
1239 
1240   // Make sure phi nodes in our reachable successors end up with a
1241   // LiveOnEntryDef for our incoming edge, even though our block is forward
1242   // unreachable.  We could just disconnect these blocks from the CFG fully,
1243   // but we do not right now.
1244   for (const BasicBlock *S : successors(BB)) {
1245     if (!DT->isReachableFromEntry(S))
1246       continue;
1247     auto It = PerBlockAccesses.find(S);
1248     // Rename the phi nodes in our successor block
1249     if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1250       continue;
1251     AccessList *Accesses = It->second.get();
1252     auto *Phi = cast<MemoryPhi>(&Accesses->front());
1253     Phi->addIncoming(LiveOnEntryDef.get(), BB);
1254   }
1255 
1256   auto It = PerBlockAccesses.find(BB);
1257   if (It == PerBlockAccesses.end())
1258     return;
1259 
1260   auto &Accesses = It->second;
1261   for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
1262     auto Next = std::next(AI);
1263     // If we have a phi, just remove it. We are going to replace all
1264     // users with live on entry.
1265     if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
1266       UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
1267     else
1268       Accesses->erase(AI);
1269     AI = Next;
1270   }
1271 }
1272 
1273 MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
1274     : DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
1275       SkipWalker(nullptr) {
1276   // Build MemorySSA using a batch alias analysis. This reuses the internal
1277   // state that AA collects during an alias()/getModRefInfo() call. This is
1278   // safe because there are no CFG changes while building MemorySSA and can
1279   // significantly reduce the time spent by the compiler in AA, because we will
1280   // make queries about all the instructions in the Function.
1281   assert(AA && "No alias analysis?");
1282   BatchAAResults BatchAA(*AA);
1283   buildMemorySSA(BatchAA);
1284   // Intentionally leave AA to nullptr while building so we don't accidently
1285   // use non-batch AliasAnalysis.
1286   this->AA = AA;
1287   // Also create the walker here.
1288   getWalker();
1289 }
1290 
1291 MemorySSA::~MemorySSA() {
1292   // Drop all our references
1293   for (const auto &Pair : PerBlockAccesses)
1294     for (MemoryAccess &MA : *Pair.second)
1295       MA.dropAllReferences();
1296 }
1297 
1298 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
1299   auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
1300 
1301   if (Res.second)
1302     Res.first->second = std::make_unique<AccessList>();
1303   return Res.first->second.get();
1304 }
1305 
1306 MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) {
1307   auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr));
1308 
1309   if (Res.second)
1310     Res.first->second = std::make_unique<DefsList>();
1311   return Res.first->second.get();
1312 }
1313 
1314 namespace llvm {
1315 
1316 /// This class is a batch walker of all MemoryUse's in the program, and points
1317 /// their defining access at the thing that actually clobbers them.  Because it
1318 /// is a batch walker that touches everything, it does not operate like the
1319 /// other walkers.  This walker is basically performing a top-down SSA renaming
1320 /// pass, where the version stack is used as the cache.  This enables it to be
1321 /// significantly more time and memory efficient than using the regular walker,
1322 /// which is walking bottom-up.
1323 class MemorySSA::OptimizeUses {
1324 public:
1325   OptimizeUses(MemorySSA *MSSA, CachingWalker<BatchAAResults> *Walker,
1326                BatchAAResults *BAA, DominatorTree *DT)
1327       : MSSA(MSSA), Walker(Walker), AA(BAA), DT(DT) {}
1328 
1329   void optimizeUses();
1330 
1331 private:
1332   /// This represents where a given memorylocation is in the stack.
1333   struct MemlocStackInfo {
1334     // This essentially is keeping track of versions of the stack. Whenever
1335     // the stack changes due to pushes or pops, these versions increase.
1336     unsigned long StackEpoch;
1337     unsigned long PopEpoch;
1338     // This is the lower bound of places on the stack to check. It is equal to
1339     // the place the last stack walk ended.
1340     // Note: Correctness depends on this being initialized to 0, which densemap
1341     // does
1342     unsigned long LowerBound;
1343     const BasicBlock *LowerBoundBlock;
1344     // This is where the last walk for this memory location ended.
1345     unsigned long LastKill;
1346     bool LastKillValid;
1347     Optional<AliasResult> AR;
1348   };
1349 
1350   void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
1351                            SmallVectorImpl<MemoryAccess *> &,
1352                            DenseMap<MemoryLocOrCall, MemlocStackInfo> &);
1353 
1354   MemorySSA *MSSA;
1355   CachingWalker<BatchAAResults> *Walker;
1356   BatchAAResults *AA;
1357   DominatorTree *DT;
1358 };
1359 
1360 } // end namespace llvm
1361 
1362 /// Optimize the uses in a given block This is basically the SSA renaming
1363 /// algorithm, with one caveat: We are able to use a single stack for all
1364 /// MemoryUses.  This is because the set of *possible* reaching MemoryDefs is
1365 /// the same for every MemoryUse.  The *actual* clobbering MemoryDef is just
1366 /// going to be some position in that stack of possible ones.
1367 ///
1368 /// We track the stack positions that each MemoryLocation needs
1369 /// to check, and last ended at.  This is because we only want to check the
1370 /// things that changed since last time.  The same MemoryLocation should
1371 /// get clobbered by the same store (getModRefInfo does not use invariantness or
1372 /// things like this, and if they start, we can modify MemoryLocOrCall to
1373 /// include relevant data)
1374 void MemorySSA::OptimizeUses::optimizeUsesInBlock(
1375     const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
1376     SmallVectorImpl<MemoryAccess *> &VersionStack,
1377     DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) {
1378 
1379   /// If no accesses, nothing to do.
1380   MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
1381   if (Accesses == nullptr)
1382     return;
1383 
1384   // Pop everything that doesn't dominate the current block off the stack,
1385   // increment the PopEpoch to account for this.
1386   while (true) {
1387     assert(
1388         !VersionStack.empty() &&
1389         "Version stack should have liveOnEntry sentinel dominating everything");
1390     BasicBlock *BackBlock = VersionStack.back()->getBlock();
1391     if (DT->dominates(BackBlock, BB))
1392       break;
1393     while (VersionStack.back()->getBlock() == BackBlock)
1394       VersionStack.pop_back();
1395     ++PopEpoch;
1396   }
1397 
1398   for (MemoryAccess &MA : *Accesses) {
1399     auto *MU = dyn_cast<MemoryUse>(&MA);
1400     if (!MU) {
1401       VersionStack.push_back(&MA);
1402       ++StackEpoch;
1403       continue;
1404     }
1405 
1406     if (MU->isOptimized())
1407       continue;
1408 
1409     if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
1410       MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true, None);
1411       continue;
1412     }
1413 
1414     MemoryLocOrCall UseMLOC(MU);
1415     auto &LocInfo = LocStackInfo[UseMLOC];
1416     // If the pop epoch changed, it means we've removed stuff from top of
1417     // stack due to changing blocks. We may have to reset the lower bound or
1418     // last kill info.
1419     if (LocInfo.PopEpoch != PopEpoch) {
1420       LocInfo.PopEpoch = PopEpoch;
1421       LocInfo.StackEpoch = StackEpoch;
1422       // If the lower bound was in something that no longer dominates us, we
1423       // have to reset it.
1424       // We can't simply track stack size, because the stack may have had
1425       // pushes/pops in the meantime.
1426       // XXX: This is non-optimal, but only is slower cases with heavily
1427       // branching dominator trees.  To get the optimal number of queries would
1428       // be to make lowerbound and lastkill a per-loc stack, and pop it until
1429       // the top of that stack dominates us.  This does not seem worth it ATM.
1430       // A much cheaper optimization would be to always explore the deepest
1431       // branch of the dominator tree first. This will guarantee this resets on
1432       // the smallest set of blocks.
1433       if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
1434           !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
1435         // Reset the lower bound of things to check.
1436         // TODO: Some day we should be able to reset to last kill, rather than
1437         // 0.
1438         LocInfo.LowerBound = 0;
1439         LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
1440         LocInfo.LastKillValid = false;
1441       }
1442     } else if (LocInfo.StackEpoch != StackEpoch) {
1443       // If all that has changed is the StackEpoch, we only have to check the
1444       // new things on the stack, because we've checked everything before.  In
1445       // this case, the lower bound of things to check remains the same.
1446       LocInfo.PopEpoch = PopEpoch;
1447       LocInfo.StackEpoch = StackEpoch;
1448     }
1449     if (!LocInfo.LastKillValid) {
1450       LocInfo.LastKill = VersionStack.size() - 1;
1451       LocInfo.LastKillValid = true;
1452       LocInfo.AR = AliasResult::MayAlias;
1453     }
1454 
1455     // At this point, we should have corrected last kill and LowerBound to be
1456     // in bounds.
1457     assert(LocInfo.LowerBound < VersionStack.size() &&
1458            "Lower bound out of range");
1459     assert(LocInfo.LastKill < VersionStack.size() &&
1460            "Last kill info out of range");
1461     // In any case, the new upper bound is the top of the stack.
1462     unsigned long UpperBound = VersionStack.size() - 1;
1463 
1464     if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
1465       LLVM_DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("
1466                         << *(MU->getMemoryInst()) << ")"
1467                         << " because there are "
1468                         << UpperBound - LocInfo.LowerBound
1469                         << " stores to disambiguate\n");
1470       // Because we did not walk, LastKill is no longer valid, as this may
1471       // have been a kill.
1472       LocInfo.LastKillValid = false;
1473       continue;
1474     }
1475     bool FoundClobberResult = false;
1476     unsigned UpwardWalkLimit = MaxCheckLimit;
1477     while (UpperBound > LocInfo.LowerBound) {
1478       if (isa<MemoryPhi>(VersionStack[UpperBound])) {
1479         // For phis, use the walker, see where we ended up, go there.
1480         // The invariant.group handling in MemorySSA is ad-hoc and doesn't
1481         // support updates, so don't use it to optimize uses.
1482         MemoryAccess *Result =
1483             Walker->getClobberingMemoryAccessWithoutInvariantGroup(
1484                 MU, UpwardWalkLimit);
1485         // We are guaranteed to find it or something is wrong.
1486         while (VersionStack[UpperBound] != Result) {
1487           assert(UpperBound != 0);
1488           --UpperBound;
1489         }
1490         FoundClobberResult = true;
1491         break;
1492       }
1493 
1494       MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
1495       ClobberAlias CA = instructionClobbersQuery(MD, MU, UseMLOC, *AA);
1496       if (CA.IsClobber) {
1497         FoundClobberResult = true;
1498         LocInfo.AR = CA.AR;
1499         break;
1500       }
1501       --UpperBound;
1502     }
1503 
1504     // Note: Phis always have AliasResult AR set to MayAlias ATM.
1505 
1506     // At the end of this loop, UpperBound is either a clobber, or lower bound
1507     // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
1508     if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
1509       // We were last killed now by where we got to
1510       if (MSSA->isLiveOnEntryDef(VersionStack[UpperBound]))
1511         LocInfo.AR = None;
1512       MU->setDefiningAccess(VersionStack[UpperBound], true, LocInfo.AR);
1513       LocInfo.LastKill = UpperBound;
1514     } else {
1515       // Otherwise, we checked all the new ones, and now we know we can get to
1516       // LastKill.
1517       MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true, LocInfo.AR);
1518     }
1519     LocInfo.LowerBound = VersionStack.size() - 1;
1520     LocInfo.LowerBoundBlock = BB;
1521   }
1522 }
1523 
1524 /// Optimize uses to point to their actual clobbering definitions.
1525 void MemorySSA::OptimizeUses::optimizeUses() {
1526   SmallVector<MemoryAccess *, 16> VersionStack;
1527   DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo;
1528   VersionStack.push_back(MSSA->getLiveOnEntryDef());
1529 
1530   unsigned long StackEpoch = 1;
1531   unsigned long PopEpoch = 1;
1532   // We perform a non-recursive top-down dominator tree walk.
1533   for (const auto *DomNode : depth_first(DT->getRootNode()))
1534     optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
1535                         LocStackInfo);
1536 }
1537 
1538 void MemorySSA::placePHINodes(
1539     const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks) {
1540   // Determine where our MemoryPhi's should go
1541   ForwardIDFCalculator IDFs(*DT);
1542   IDFs.setDefiningBlocks(DefiningBlocks);
1543   SmallVector<BasicBlock *, 32> IDFBlocks;
1544   IDFs.calculate(IDFBlocks);
1545 
1546   // Now place MemoryPhi nodes.
1547   for (auto &BB : IDFBlocks)
1548     createMemoryPhi(BB);
1549 }
1550 
1551 void MemorySSA::buildMemorySSA(BatchAAResults &BAA) {
1552   // We create an access to represent "live on entry", for things like
1553   // arguments or users of globals, where the memory they use is defined before
1554   // the beginning of the function. We do not actually insert it into the IR.
1555   // We do not define a live on exit for the immediate uses, and thus our
1556   // semantics do *not* imply that something with no immediate uses can simply
1557   // be removed.
1558   BasicBlock &StartingPoint = F.getEntryBlock();
1559   LiveOnEntryDef.reset(new MemoryDef(F.getContext(), nullptr, nullptr,
1560                                      &StartingPoint, NextID++));
1561 
1562   // We maintain lists of memory accesses per-block, trading memory for time. We
1563   // could just look up the memory access for every possible instruction in the
1564   // stream.
1565   SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
1566   // Go through each block, figure out where defs occur, and chain together all
1567   // the accesses.
1568   for (BasicBlock &B : F) {
1569     bool InsertIntoDef = false;
1570     AccessList *Accesses = nullptr;
1571     DefsList *Defs = nullptr;
1572     for (Instruction &I : B) {
1573       MemoryUseOrDef *MUD = createNewAccess(&I, &BAA);
1574       if (!MUD)
1575         continue;
1576 
1577       if (!Accesses)
1578         Accesses = getOrCreateAccessList(&B);
1579       Accesses->push_back(MUD);
1580       if (isa<MemoryDef>(MUD)) {
1581         InsertIntoDef = true;
1582         if (!Defs)
1583           Defs = getOrCreateDefsList(&B);
1584         Defs->push_back(*MUD);
1585       }
1586     }
1587     if (InsertIntoDef)
1588       DefiningBlocks.insert(&B);
1589   }
1590   placePHINodes(DefiningBlocks);
1591 
1592   // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
1593   // filled in with all blocks.
1594   SmallPtrSet<BasicBlock *, 16> Visited;
1595   renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
1596 
1597   // Mark the uses in unreachable blocks as live on entry, so that they go
1598   // somewhere.
1599   for (auto &BB : F)
1600     if (!Visited.count(&BB))
1601       markUnreachableAsLiveOnEntry(&BB);
1602 }
1603 
1604 MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
1605 
1606 MemorySSA::CachingWalker<AliasAnalysis> *MemorySSA::getWalkerImpl() {
1607   if (Walker)
1608     return Walker.get();
1609 
1610   if (!WalkerBase)
1611     WalkerBase =
1612         std::make_unique<ClobberWalkerBase<AliasAnalysis>>(this, AA, DT);
1613 
1614   Walker =
1615       std::make_unique<CachingWalker<AliasAnalysis>>(this, WalkerBase.get());
1616   return Walker.get();
1617 }
1618 
1619 MemorySSAWalker *MemorySSA::getSkipSelfWalker() {
1620   if (SkipWalker)
1621     return SkipWalker.get();
1622 
1623   if (!WalkerBase)
1624     WalkerBase =
1625         std::make_unique<ClobberWalkerBase<AliasAnalysis>>(this, AA, DT);
1626 
1627   SkipWalker =
1628       std::make_unique<SkipSelfWalker<AliasAnalysis>>(this, WalkerBase.get());
1629   return SkipWalker.get();
1630  }
1631 
1632 
1633 // This is a helper function used by the creation routines. It places NewAccess
1634 // into the access and defs lists for a given basic block, at the given
1635 // insertion point.
1636 void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess,
1637                                         const BasicBlock *BB,
1638                                         InsertionPlace Point) {
1639   auto *Accesses = getOrCreateAccessList(BB);
1640   if (Point == Beginning) {
1641     // If it's a phi node, it goes first, otherwise, it goes after any phi
1642     // nodes.
1643     if (isa<MemoryPhi>(NewAccess)) {
1644       Accesses->push_front(NewAccess);
1645       auto *Defs = getOrCreateDefsList(BB);
1646       Defs->push_front(*NewAccess);
1647     } else {
1648       auto AI = find_if_not(
1649           *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1650       Accesses->insert(AI, NewAccess);
1651       if (!isa<MemoryUse>(NewAccess)) {
1652         auto *Defs = getOrCreateDefsList(BB);
1653         auto DI = find_if_not(
1654             *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1655         Defs->insert(DI, *NewAccess);
1656       }
1657     }
1658   } else {
1659     Accesses->push_back(NewAccess);
1660     if (!isa<MemoryUse>(NewAccess)) {
1661       auto *Defs = getOrCreateDefsList(BB);
1662       Defs->push_back(*NewAccess);
1663     }
1664   }
1665   BlockNumberingValid.erase(BB);
1666 }
1667 
1668 void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB,
1669                                       AccessList::iterator InsertPt) {
1670   auto *Accesses = getWritableBlockAccesses(BB);
1671   bool WasEnd = InsertPt == Accesses->end();
1672   Accesses->insert(AccessList::iterator(InsertPt), What);
1673   if (!isa<MemoryUse>(What)) {
1674     auto *Defs = getOrCreateDefsList(BB);
1675     // If we got asked to insert at the end, we have an easy job, just shove it
1676     // at the end. If we got asked to insert before an existing def, we also get
1677     // an iterator. If we got asked to insert before a use, we have to hunt for
1678     // the next def.
1679     if (WasEnd) {
1680       Defs->push_back(*What);
1681     } else if (isa<MemoryDef>(InsertPt)) {
1682       Defs->insert(InsertPt->getDefsIterator(), *What);
1683     } else {
1684       while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt))
1685         ++InsertPt;
1686       // Either we found a def, or we are inserting at the end
1687       if (InsertPt == Accesses->end())
1688         Defs->push_back(*What);
1689       else
1690         Defs->insert(InsertPt->getDefsIterator(), *What);
1691     }
1692   }
1693   BlockNumberingValid.erase(BB);
1694 }
1695 
1696 void MemorySSA::prepareForMoveTo(MemoryAccess *What, BasicBlock *BB) {
1697   // Keep it in the lookup tables, remove from the lists
1698   removeFromLists(What, false);
1699 
1700   // Note that moving should implicitly invalidate the optimized state of a
1701   // MemoryUse (and Phis can't be optimized). However, it doesn't do so for a
1702   // MemoryDef.
1703   if (auto *MD = dyn_cast<MemoryDef>(What))
1704     MD->resetOptimized();
1705   What->setBlock(BB);
1706 }
1707 
1708 // Move What before Where in the IR.  The end result is that What will belong to
1709 // the right lists and have the right Block set, but will not otherwise be
1710 // correct. It will not have the right defining access, and if it is a def,
1711 // things below it will not properly be updated.
1712 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1713                        AccessList::iterator Where) {
1714   prepareForMoveTo(What, BB);
1715   insertIntoListsBefore(What, BB, Where);
1716 }
1717 
1718 void MemorySSA::moveTo(MemoryAccess *What, BasicBlock *BB,
1719                        InsertionPlace Point) {
1720   if (isa<MemoryPhi>(What)) {
1721     assert(Point == Beginning &&
1722            "Can only move a Phi at the beginning of the block");
1723     // Update lookup table entry
1724     ValueToMemoryAccess.erase(What->getBlock());
1725     bool Inserted = ValueToMemoryAccess.insert({BB, What}).second;
1726     (void)Inserted;
1727     assert(Inserted && "Cannot move a Phi to a block that already has one");
1728   }
1729 
1730   prepareForMoveTo(What, BB);
1731   insertIntoListsForBlock(What, BB, Point);
1732 }
1733 
1734 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
1735   assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB");
1736   MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
1737   // Phi's always are placed at the front of the block.
1738   insertIntoListsForBlock(Phi, BB, Beginning);
1739   ValueToMemoryAccess[BB] = Phi;
1740   return Phi;
1741 }
1742 
1743 MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I,
1744                                                MemoryAccess *Definition,
1745                                                const MemoryUseOrDef *Template,
1746                                                bool CreationMustSucceed) {
1747   assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI");
1748   MemoryUseOrDef *NewAccess = createNewAccess(I, AA, Template);
1749   if (CreationMustSucceed)
1750     assert(NewAccess != nullptr && "Tried to create a memory access for a "
1751                                    "non-memory touching instruction");
1752   if (NewAccess) {
1753     assert((!Definition || !isa<MemoryUse>(Definition)) &&
1754            "A use cannot be a defining access");
1755     NewAccess->setDefiningAccess(Definition);
1756   }
1757   return NewAccess;
1758 }
1759 
1760 // Return true if the instruction has ordering constraints.
1761 // Note specifically that this only considers stores and loads
1762 // because others are still considered ModRef by getModRefInfo.
1763 static inline bool isOrdered(const Instruction *I) {
1764   if (auto *SI = dyn_cast<StoreInst>(I)) {
1765     if (!SI->isUnordered())
1766       return true;
1767   } else if (auto *LI = dyn_cast<LoadInst>(I)) {
1768     if (!LI->isUnordered())
1769       return true;
1770   }
1771   return false;
1772 }
1773 
1774 /// Helper function to create new memory accesses
1775 template <typename AliasAnalysisType>
1776 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I,
1777                                            AliasAnalysisType *AAP,
1778                                            const MemoryUseOrDef *Template) {
1779   // The assume intrinsic has a control dependency which we model by claiming
1780   // that it writes arbitrarily. Debuginfo intrinsics may be considered
1781   // clobbers when we have a nonstandard AA pipeline. Ignore these fake memory
1782   // dependencies here.
1783   // FIXME: Replace this special casing with a more accurate modelling of
1784   // assume's control dependency.
1785   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1786     switch (II->getIntrinsicID()) {
1787     default:
1788       break;
1789     case Intrinsic::assume:
1790     case Intrinsic::experimental_noalias_scope_decl:
1791     case Intrinsic::pseudoprobe:
1792       return nullptr;
1793     }
1794   }
1795 
1796   // Using a nonstandard AA pipelines might leave us with unexpected modref
1797   // results for I, so add a check to not model instructions that may not read
1798   // from or write to memory. This is necessary for correctness.
1799   if (!I->mayReadFromMemory() && !I->mayWriteToMemory())
1800     return nullptr;
1801 
1802   bool Def, Use;
1803   if (Template) {
1804     Def = isa<MemoryDef>(Template);
1805     Use = isa<MemoryUse>(Template);
1806 #if !defined(NDEBUG)
1807     ModRefInfo ModRef = AAP->getModRefInfo(I, None);
1808     bool DefCheck, UseCheck;
1809     DefCheck = isModSet(ModRef) || isOrdered(I);
1810     UseCheck = isRefSet(ModRef);
1811     assert(Def == DefCheck && (Def || Use == UseCheck) && "Invalid template");
1812 #endif
1813   } else {
1814     // Find out what affect this instruction has on memory.
1815     ModRefInfo ModRef = AAP->getModRefInfo(I, None);
1816     // The isOrdered check is used to ensure that volatiles end up as defs
1817     // (atomics end up as ModRef right now anyway).  Until we separate the
1818     // ordering chain from the memory chain, this enables people to see at least
1819     // some relative ordering to volatiles.  Note that getClobberingMemoryAccess
1820     // will still give an answer that bypasses other volatile loads.  TODO:
1821     // Separate memory aliasing and ordering into two different chains so that
1822     // we can precisely represent both "what memory will this read/write/is
1823     // clobbered by" and "what instructions can I move this past".
1824     Def = isModSet(ModRef) || isOrdered(I);
1825     Use = isRefSet(ModRef);
1826   }
1827 
1828   // It's possible for an instruction to not modify memory at all. During
1829   // construction, we ignore them.
1830   if (!Def && !Use)
1831     return nullptr;
1832 
1833   MemoryUseOrDef *MUD;
1834   if (Def)
1835     MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
1836   else
1837     MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
1838   ValueToMemoryAccess[I] = MUD;
1839   return MUD;
1840 }
1841 
1842 /// Properly remove \p MA from all of MemorySSA's lookup tables.
1843 void MemorySSA::removeFromLookups(MemoryAccess *MA) {
1844   assert(MA->use_empty() &&
1845          "Trying to remove memory access that still has uses");
1846   BlockNumbering.erase(MA);
1847   if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1848     MUD->setDefiningAccess(nullptr);
1849   // Invalidate our walker's cache if necessary
1850   if (!isa<MemoryUse>(MA))
1851     getWalker()->invalidateInfo(MA);
1852 
1853   Value *MemoryInst;
1854   if (const auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1855     MemoryInst = MUD->getMemoryInst();
1856   else
1857     MemoryInst = MA->getBlock();
1858 
1859   auto VMA = ValueToMemoryAccess.find(MemoryInst);
1860   if (VMA->second == MA)
1861     ValueToMemoryAccess.erase(VMA);
1862 }
1863 
1864 /// Properly remove \p MA from all of MemorySSA's lists.
1865 ///
1866 /// Because of the way the intrusive list and use lists work, it is important to
1867 /// do removal in the right order.
1868 /// ShouldDelete defaults to true, and will cause the memory access to also be
1869 /// deleted, not just removed.
1870 void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) {
1871   BasicBlock *BB = MA->getBlock();
1872   // The access list owns the reference, so we erase it from the non-owning list
1873   // first.
1874   if (!isa<MemoryUse>(MA)) {
1875     auto DefsIt = PerBlockDefs.find(BB);
1876     std::unique_ptr<DefsList> &Defs = DefsIt->second;
1877     Defs->remove(*MA);
1878     if (Defs->empty())
1879       PerBlockDefs.erase(DefsIt);
1880   }
1881 
1882   // The erase call here will delete it. If we don't want it deleted, we call
1883   // remove instead.
1884   auto AccessIt = PerBlockAccesses.find(BB);
1885   std::unique_ptr<AccessList> &Accesses = AccessIt->second;
1886   if (ShouldDelete)
1887     Accesses->erase(MA);
1888   else
1889     Accesses->remove(MA);
1890 
1891   if (Accesses->empty()) {
1892     PerBlockAccesses.erase(AccessIt);
1893     BlockNumberingValid.erase(BB);
1894   }
1895 }
1896 
1897 void MemorySSA::print(raw_ostream &OS) const {
1898   MemorySSAAnnotatedWriter Writer(this);
1899   F.print(OS, &Writer);
1900 }
1901 
1902 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1903 LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); }
1904 #endif
1905 
1906 void MemorySSA::verifyMemorySSA(VerificationLevel VL) const {
1907 #if !defined(NDEBUG) && defined(EXPENSIVE_CHECKS)
1908   VL = VerificationLevel::Full;
1909 #endif
1910 
1911 #ifndef NDEBUG
1912   verifyOrderingDominationAndDefUses(F, VL);
1913   verifyDominationNumbers(F);
1914   if (VL == VerificationLevel::Full)
1915     verifyPrevDefInPhis(F);
1916 #endif
1917   // Previously, the verification used to also verify that the clobberingAccess
1918   // cached by MemorySSA is the same as the clobberingAccess found at a later
1919   // query to AA. This does not hold true in general due to the current fragility
1920   // of BasicAA which has arbitrary caps on the things it analyzes before giving
1921   // up. As a result, transformations that are correct, will lead to BasicAA
1922   // returning different Alias answers before and after that transformation.
1923   // Invalidating MemorySSA is not an option, as the results in BasicAA can be so
1924   // random, in the worst case we'd need to rebuild MemorySSA from scratch after
1925   // every transformation, which defeats the purpose of using it. For such an
1926   // example, see test4 added in D51960.
1927 }
1928 
1929 void MemorySSA::verifyPrevDefInPhis(Function &F) const {
1930   for (const BasicBlock &BB : F) {
1931     if (MemoryPhi *Phi = getMemoryAccess(&BB)) {
1932       for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
1933         auto *Pred = Phi->getIncomingBlock(I);
1934         auto *IncAcc = Phi->getIncomingValue(I);
1935         // If Pred has no unreachable predecessors, get last def looking at
1936         // IDoms. If, while walkings IDoms, any of these has an unreachable
1937         // predecessor, then the incoming def can be any access.
1938         if (auto *DTNode = DT->getNode(Pred)) {
1939           while (DTNode) {
1940             if (auto *DefList = getBlockDefs(DTNode->getBlock())) {
1941               auto *LastAcc = &*(--DefList->end());
1942               assert(LastAcc == IncAcc &&
1943                      "Incorrect incoming access into phi.");
1944               (void)IncAcc;
1945               (void)LastAcc;
1946               break;
1947             }
1948             DTNode = DTNode->getIDom();
1949           }
1950         } else {
1951           // If Pred has unreachable predecessors, but has at least a Def, the
1952           // incoming access can be the last Def in Pred, or it could have been
1953           // optimized to LoE. After an update, though, the LoE may have been
1954           // replaced by another access, so IncAcc may be any access.
1955           // If Pred has unreachable predecessors and no Defs, incoming access
1956           // should be LoE; However, after an update, it may be any access.
1957         }
1958       }
1959     }
1960   }
1961 }
1962 
1963 /// Verify that all of the blocks we believe to have valid domination numbers
1964 /// actually have valid domination numbers.
1965 void MemorySSA::verifyDominationNumbers(const Function &F) const {
1966   if (BlockNumberingValid.empty())
1967     return;
1968 
1969   SmallPtrSet<const BasicBlock *, 16> ValidBlocks = BlockNumberingValid;
1970   for (const BasicBlock &BB : F) {
1971     if (!ValidBlocks.count(&BB))
1972       continue;
1973 
1974     ValidBlocks.erase(&BB);
1975 
1976     const AccessList *Accesses = getBlockAccesses(&BB);
1977     // It's correct to say an empty block has valid numbering.
1978     if (!Accesses)
1979       continue;
1980 
1981     // Block numbering starts at 1.
1982     unsigned long LastNumber = 0;
1983     for (const MemoryAccess &MA : *Accesses) {
1984       auto ThisNumberIter = BlockNumbering.find(&MA);
1985       assert(ThisNumberIter != BlockNumbering.end() &&
1986              "MemoryAccess has no domination number in a valid block!");
1987 
1988       unsigned long ThisNumber = ThisNumberIter->second;
1989       assert(ThisNumber > LastNumber &&
1990              "Domination numbers should be strictly increasing!");
1991       (void)LastNumber;
1992       LastNumber = ThisNumber;
1993     }
1994   }
1995 
1996   assert(ValidBlocks.empty() &&
1997          "All valid BasicBlocks should exist in F -- dangling pointers?");
1998 }
1999 
2000 /// Verify ordering: the order and existence of MemoryAccesses matches the
2001 /// order and existence of memory affecting instructions.
2002 /// Verify domination: each definition dominates all of its uses.
2003 /// Verify def-uses: the immediate use information - walk all the memory
2004 /// accesses and verifying that, for each use, it appears in the appropriate
2005 /// def's use list
2006 void MemorySSA::verifyOrderingDominationAndDefUses(Function &F,
2007                                                    VerificationLevel VL) const {
2008   // Walk all the blocks, comparing what the lookups think and what the access
2009   // lists think, as well as the order in the blocks vs the order in the access
2010   // lists.
2011   SmallVector<MemoryAccess *, 32> ActualAccesses;
2012   SmallVector<MemoryAccess *, 32> ActualDefs;
2013   for (BasicBlock &B : F) {
2014     const AccessList *AL = getBlockAccesses(&B);
2015     const auto *DL = getBlockDefs(&B);
2016     MemoryPhi *Phi = getMemoryAccess(&B);
2017     if (Phi) {
2018       // Verify ordering.
2019       ActualAccesses.push_back(Phi);
2020       ActualDefs.push_back(Phi);
2021       // Verify domination
2022       for (const Use &U : Phi->uses()) {
2023         assert(dominates(Phi, U) && "Memory PHI does not dominate it's uses");
2024         (void)U;
2025       }
2026       // Verify def-uses for full verify.
2027       if (VL == VerificationLevel::Full) {
2028         assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(
2029                                             pred_begin(&B), pred_end(&B))) &&
2030                "Incomplete MemoryPhi Node");
2031         for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
2032           verifyUseInDefs(Phi->getIncomingValue(I), Phi);
2033           assert(is_contained(predecessors(&B), Phi->getIncomingBlock(I)) &&
2034                  "Incoming phi block not a block predecessor");
2035         }
2036       }
2037     }
2038 
2039     for (Instruction &I : B) {
2040       MemoryUseOrDef *MA = getMemoryAccess(&I);
2041       assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&
2042              "We have memory affecting instructions "
2043              "in this block but they are not in the "
2044              "access list or defs list");
2045       if (MA) {
2046         // Verify ordering.
2047         ActualAccesses.push_back(MA);
2048         if (MemoryAccess *MD = dyn_cast<MemoryDef>(MA)) {
2049           // Verify ordering.
2050           ActualDefs.push_back(MA);
2051           // Verify domination.
2052           for (const Use &U : MD->uses()) {
2053             assert(dominates(MD, U) &&
2054                    "Memory Def does not dominate it's uses");
2055             (void)U;
2056           }
2057         }
2058         // Verify def-uses for full verify.
2059         if (VL == VerificationLevel::Full)
2060           verifyUseInDefs(MA->getDefiningAccess(), MA);
2061       }
2062     }
2063     // Either we hit the assert, really have no accesses, or we have both
2064     // accesses and an access list. Same with defs.
2065     if (!AL && !DL)
2066       continue;
2067     // Verify ordering.
2068     assert(AL->size() == ActualAccesses.size() &&
2069            "We don't have the same number of accesses in the block as on the "
2070            "access list");
2071     assert((DL || ActualDefs.size() == 0) &&
2072            "Either we should have a defs list, or we should have no defs");
2073     assert((!DL || DL->size() == ActualDefs.size()) &&
2074            "We don't have the same number of defs in the block as on the "
2075            "def list");
2076     auto ALI = AL->begin();
2077     auto AAI = ActualAccesses.begin();
2078     while (ALI != AL->end() && AAI != ActualAccesses.end()) {
2079       assert(&*ALI == *AAI && "Not the same accesses in the same order");
2080       ++ALI;
2081       ++AAI;
2082     }
2083     ActualAccesses.clear();
2084     if (DL) {
2085       auto DLI = DL->begin();
2086       auto ADI = ActualDefs.begin();
2087       while (DLI != DL->end() && ADI != ActualDefs.end()) {
2088         assert(&*DLI == *ADI && "Not the same defs in the same order");
2089         ++DLI;
2090         ++ADI;
2091       }
2092     }
2093     ActualDefs.clear();
2094   }
2095 }
2096 
2097 /// Verify the def-use lists in MemorySSA, by verifying that \p Use
2098 /// appears in the use list of \p Def.
2099 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
2100   // The live on entry use may cause us to get a NULL def here
2101   if (!Def)
2102     assert(isLiveOnEntryDef(Use) &&
2103            "Null def but use not point to live on entry def");
2104   else
2105     assert(is_contained(Def->users(), Use) &&
2106            "Did not find use in def's use list");
2107 }
2108 
2109 /// Perform a local numbering on blocks so that instruction ordering can be
2110 /// determined in constant time.
2111 /// TODO: We currently just number in order.  If we numbered by N, we could
2112 /// allow at least N-1 sequences of insertBefore or insertAfter (and at least
2113 /// log2(N) sequences of mixed before and after) without needing to invalidate
2114 /// the numbering.
2115 void MemorySSA::renumberBlock(const BasicBlock *B) const {
2116   // The pre-increment ensures the numbers really start at 1.
2117   unsigned long CurrentNumber = 0;
2118   const AccessList *AL = getBlockAccesses(B);
2119   assert(AL != nullptr && "Asking to renumber an empty block");
2120   for (const auto &I : *AL)
2121     BlockNumbering[&I] = ++CurrentNumber;
2122   BlockNumberingValid.insert(B);
2123 }
2124 
2125 /// Determine, for two memory accesses in the same block,
2126 /// whether \p Dominator dominates \p Dominatee.
2127 /// \returns True if \p Dominator dominates \p Dominatee.
2128 bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
2129                                  const MemoryAccess *Dominatee) const {
2130   const BasicBlock *DominatorBlock = Dominator->getBlock();
2131 
2132   assert((DominatorBlock == Dominatee->getBlock()) &&
2133          "Asking for local domination when accesses are in different blocks!");
2134   // A node dominates itself.
2135   if (Dominatee == Dominator)
2136     return true;
2137 
2138   // When Dominatee is defined on function entry, it is not dominated by another
2139   // memory access.
2140   if (isLiveOnEntryDef(Dominatee))
2141     return false;
2142 
2143   // When Dominator is defined on function entry, it dominates the other memory
2144   // access.
2145   if (isLiveOnEntryDef(Dominator))
2146     return true;
2147 
2148   if (!BlockNumberingValid.count(DominatorBlock))
2149     renumberBlock(DominatorBlock);
2150 
2151   unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
2152   // All numbers start with 1
2153   assert(DominatorNum != 0 && "Block was not numbered properly");
2154   unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
2155   assert(DominateeNum != 0 && "Block was not numbered properly");
2156   return DominatorNum < DominateeNum;
2157 }
2158 
2159 bool MemorySSA::dominates(const MemoryAccess *Dominator,
2160                           const MemoryAccess *Dominatee) const {
2161   if (Dominator == Dominatee)
2162     return true;
2163 
2164   if (isLiveOnEntryDef(Dominatee))
2165     return false;
2166 
2167   if (Dominator->getBlock() != Dominatee->getBlock())
2168     return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
2169   return locallyDominates(Dominator, Dominatee);
2170 }
2171 
2172 bool MemorySSA::dominates(const MemoryAccess *Dominator,
2173                           const Use &Dominatee) const {
2174   if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
2175     BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
2176     // The def must dominate the incoming block of the phi.
2177     if (UseBB != Dominator->getBlock())
2178       return DT->dominates(Dominator->getBlock(), UseBB);
2179     // If the UseBB and the DefBB are the same, compare locally.
2180     return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
2181   }
2182   // If it's not a PHI node use, the normal dominates can already handle it.
2183   return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
2184 }
2185 
2186 void MemorySSA::ensureOptimizedUses() {
2187   if (IsOptimized)
2188     return;
2189 
2190   BatchAAResults BatchAA(*AA);
2191   ClobberWalkerBase<BatchAAResults> WalkerBase(this, &BatchAA, DT);
2192   CachingWalker<BatchAAResults> WalkerLocal(this, &WalkerBase);
2193   OptimizeUses(this, &WalkerLocal, &BatchAA, DT).optimizeUses();
2194   IsOptimized = true;
2195 }
2196 
2197 void MemoryAccess::print(raw_ostream &OS) const {
2198   switch (getValueID()) {
2199   case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS);
2200   case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS);
2201   case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS);
2202   }
2203   llvm_unreachable("invalid value id");
2204 }
2205 
2206 void MemoryDef::print(raw_ostream &OS) const {
2207   MemoryAccess *UO = getDefiningAccess();
2208 
2209   auto printID = [&OS](MemoryAccess *A) {
2210     if (A && A->getID())
2211       OS << A->getID();
2212     else
2213       OS << LiveOnEntryStr;
2214   };
2215 
2216   OS << getID() << " = MemoryDef(";
2217   printID(UO);
2218   OS << ")";
2219 
2220   if (isOptimized()) {
2221     OS << "->";
2222     printID(getOptimized());
2223 
2224     if (Optional<AliasResult> AR = getOptimizedAccessType())
2225       OS << " " << *AR;
2226   }
2227 }
2228 
2229 void MemoryPhi::print(raw_ostream &OS) const {
2230   ListSeparator LS(",");
2231   OS << getID() << " = MemoryPhi(";
2232   for (const auto &Op : operands()) {
2233     BasicBlock *BB = getIncomingBlock(Op);
2234     MemoryAccess *MA = cast<MemoryAccess>(Op);
2235 
2236     OS << LS << '{';
2237     if (BB->hasName())
2238       OS << BB->getName();
2239     else
2240       BB->printAsOperand(OS, false);
2241     OS << ',';
2242     if (unsigned ID = MA->getID())
2243       OS << ID;
2244     else
2245       OS << LiveOnEntryStr;
2246     OS << '}';
2247   }
2248   OS << ')';
2249 }
2250 
2251 void MemoryUse::print(raw_ostream &OS) const {
2252   MemoryAccess *UO = getDefiningAccess();
2253   OS << "MemoryUse(";
2254   if (UO && UO->getID())
2255     OS << UO->getID();
2256   else
2257     OS << LiveOnEntryStr;
2258   OS << ')';
2259 
2260   if (Optional<AliasResult> AR = getOptimizedAccessType())
2261     OS << " " << *AR;
2262 }
2263 
2264 void MemoryAccess::dump() const {
2265 // Cannot completely remove virtual function even in release mode.
2266 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2267   print(dbgs());
2268   dbgs() << "\n";
2269 #endif
2270 }
2271 
2272 char MemorySSAPrinterLegacyPass::ID = 0;
2273 
2274 MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) {
2275   initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry());
2276 }
2277 
2278 void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
2279   AU.setPreservesAll();
2280   AU.addRequired<MemorySSAWrapperPass>();
2281 }
2282 
2283 class DOTFuncMSSAInfo {
2284 private:
2285   const Function &F;
2286   MemorySSAAnnotatedWriter MSSAWriter;
2287 
2288 public:
2289   DOTFuncMSSAInfo(const Function &F, MemorySSA &MSSA)
2290       : F(F), MSSAWriter(&MSSA) {}
2291 
2292   const Function *getFunction() { return &F; }
2293   MemorySSAAnnotatedWriter &getWriter() { return MSSAWriter; }
2294 };
2295 
2296 namespace llvm {
2297 
2298 template <>
2299 struct GraphTraits<DOTFuncMSSAInfo *> : public GraphTraits<const BasicBlock *> {
2300   static NodeRef getEntryNode(DOTFuncMSSAInfo *CFGInfo) {
2301     return &(CFGInfo->getFunction()->getEntryBlock());
2302   }
2303 
2304   // nodes_iterator/begin/end - Allow iteration over all nodes in the graph
2305   using nodes_iterator = pointer_iterator<Function::const_iterator>;
2306 
2307   static nodes_iterator nodes_begin(DOTFuncMSSAInfo *CFGInfo) {
2308     return nodes_iterator(CFGInfo->getFunction()->begin());
2309   }
2310 
2311   static nodes_iterator nodes_end(DOTFuncMSSAInfo *CFGInfo) {
2312     return nodes_iterator(CFGInfo->getFunction()->end());
2313   }
2314 
2315   static size_t size(DOTFuncMSSAInfo *CFGInfo) {
2316     return CFGInfo->getFunction()->size();
2317   }
2318 };
2319 
2320 template <>
2321 struct DOTGraphTraits<DOTFuncMSSAInfo *> : public DefaultDOTGraphTraits {
2322 
2323   DOTGraphTraits(bool IsSimple = false) : DefaultDOTGraphTraits(IsSimple) {}
2324 
2325   static std::string getGraphName(DOTFuncMSSAInfo *CFGInfo) {
2326     return "MSSA CFG for '" + CFGInfo->getFunction()->getName().str() +
2327            "' function";
2328   }
2329 
2330   std::string getNodeLabel(const BasicBlock *Node, DOTFuncMSSAInfo *CFGInfo) {
2331     return DOTGraphTraits<DOTFuncInfo *>::getCompleteNodeLabel(
2332         Node, nullptr,
2333         [CFGInfo](raw_string_ostream &OS, const BasicBlock &BB) -> void {
2334           BB.print(OS, &CFGInfo->getWriter(), true, true);
2335         },
2336         [](std::string &S, unsigned &I, unsigned Idx) -> void {
2337           std::string Str = S.substr(I, Idx - I);
2338           StringRef SR = Str;
2339           if (SR.count(" = MemoryDef(") || SR.count(" = MemoryPhi(") ||
2340               SR.count("MemoryUse("))
2341             return;
2342           DOTGraphTraits<DOTFuncInfo *>::eraseComment(S, I, Idx);
2343         });
2344   }
2345 
2346   static std::string getEdgeSourceLabel(const BasicBlock *Node,
2347                                         const_succ_iterator I) {
2348     return DOTGraphTraits<DOTFuncInfo *>::getEdgeSourceLabel(Node, I);
2349   }
2350 
2351   /// Display the raw branch weights from PGO.
2352   std::string getEdgeAttributes(const BasicBlock *Node, const_succ_iterator I,
2353                                 DOTFuncMSSAInfo *CFGInfo) {
2354     return "";
2355   }
2356 
2357   std::string getNodeAttributes(const BasicBlock *Node,
2358                                 DOTFuncMSSAInfo *CFGInfo) {
2359     return getNodeLabel(Node, CFGInfo).find(';') != std::string::npos
2360                ? "style=filled, fillcolor=lightpink"
2361                : "";
2362   }
2363 };
2364 
2365 } // namespace llvm
2366 
2367 bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) {
2368   auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
2369   MSSA.ensureOptimizedUses();
2370   if (DotCFGMSSA != "") {
2371     DOTFuncMSSAInfo CFGInfo(F, MSSA);
2372     WriteGraph(&CFGInfo, "", false, "MSSA", DotCFGMSSA);
2373   } else
2374     MSSA.print(dbgs());
2375 
2376   if (VerifyMemorySSA)
2377     MSSA.verifyMemorySSA();
2378   return false;
2379 }
2380 
2381 AnalysisKey MemorySSAAnalysis::Key;
2382 
2383 MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F,
2384                                                  FunctionAnalysisManager &AM) {
2385   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
2386   auto &AA = AM.getResult<AAManager>(F);
2387   return MemorySSAAnalysis::Result(std::make_unique<MemorySSA>(F, &AA, &DT));
2388 }
2389 
2390 bool MemorySSAAnalysis::Result::invalidate(
2391     Function &F, const PreservedAnalyses &PA,
2392     FunctionAnalysisManager::Invalidator &Inv) {
2393   auto PAC = PA.getChecker<MemorySSAAnalysis>();
2394   return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()) ||
2395          Inv.invalidate<AAManager>(F, PA) ||
2396          Inv.invalidate<DominatorTreeAnalysis>(F, PA);
2397 }
2398 
2399 PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
2400                                             FunctionAnalysisManager &AM) {
2401   auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
2402   MSSA.ensureOptimizedUses();
2403   if (DotCFGMSSA != "") {
2404     DOTFuncMSSAInfo CFGInfo(F, MSSA);
2405     WriteGraph(&CFGInfo, "", false, "MSSA", DotCFGMSSA);
2406   } else {
2407     OS << "MemorySSA for function: " << F.getName() << "\n";
2408     MSSA.print(OS);
2409   }
2410 
2411   return PreservedAnalyses::all();
2412 }
2413 
2414 PreservedAnalyses MemorySSAWalkerPrinterPass::run(Function &F,
2415                                                   FunctionAnalysisManager &AM) {
2416   auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
2417   OS << "MemorySSA (walker) for function: " << F.getName() << "\n";
2418   MemorySSAWalkerAnnotatedWriter Writer(&MSSA);
2419   F.print(OS, &Writer);
2420 
2421   return PreservedAnalyses::all();
2422 }
2423 
2424 PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
2425                                              FunctionAnalysisManager &AM) {
2426   AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
2427 
2428   return PreservedAnalyses::all();
2429 }
2430 
2431 char MemorySSAWrapperPass::ID = 0;
2432 
2433 MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
2434   initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
2435 }
2436 
2437 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
2438 
2439 void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
2440   AU.setPreservesAll();
2441   AU.addRequiredTransitive<DominatorTreeWrapperPass>();
2442   AU.addRequiredTransitive<AAResultsWrapperPass>();
2443 }
2444 
2445 bool MemorySSAWrapperPass::runOnFunction(Function &F) {
2446   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2447   auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
2448   MSSA.reset(new MemorySSA(F, &AA, &DT));
2449   return false;
2450 }
2451 
2452 void MemorySSAWrapperPass::verifyAnalysis() const {
2453   if (VerifyMemorySSA)
2454     MSSA->verifyMemorySSA();
2455 }
2456 
2457 void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
2458   MSSA->print(OS);
2459 }
2460 
2461 MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
2462 
2463 /// Walk the use-def chains starting at \p StartingAccess and find
2464 /// the MemoryAccess that actually clobbers Loc.
2465 ///
2466 /// \returns our clobbering memory access
2467 template <typename AliasAnalysisType>
2468 MemoryAccess *
2469 MemorySSA::ClobberWalkerBase<AliasAnalysisType>::getClobberingMemoryAccessBase(
2470     MemoryAccess *StartingAccess, const MemoryLocation &Loc,
2471     unsigned &UpwardWalkLimit) {
2472   assert(!isa<MemoryUse>(StartingAccess) && "Use cannot be defining access");
2473 
2474   Instruction *I = nullptr;
2475   if (auto *StartingUseOrDef = dyn_cast<MemoryUseOrDef>(StartingAccess)) {
2476     if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
2477       return StartingUseOrDef;
2478 
2479     I = StartingUseOrDef->getMemoryInst();
2480 
2481     // Conservatively, fences are always clobbers, so don't perform the walk if
2482     // we hit a fence.
2483     if (!isa<CallBase>(I) && I->isFenceLike())
2484       return StartingUseOrDef;
2485   }
2486 
2487   UpwardsMemoryQuery Q;
2488   Q.OriginalAccess = StartingAccess;
2489   Q.StartingLoc = Loc;
2490   Q.Inst = nullptr;
2491   Q.IsCall = false;
2492 
2493   // Unlike the other function, do not walk to the def of a def, because we are
2494   // handed something we already believe is the clobbering access.
2495   // We never set SkipSelf to true in Q in this method.
2496   MemoryAccess *Clobber =
2497       Walker.findClobber(StartingAccess, Q, UpwardWalkLimit);
2498   LLVM_DEBUG({
2499     dbgs() << "Clobber starting at access " << *StartingAccess << "\n";
2500     if (I)
2501       dbgs() << "  for instruction " << *I << "\n";
2502     dbgs() << "  is " << *Clobber << "\n";
2503   });
2504   return Clobber;
2505 }
2506 
2507 static const Instruction *
2508 getInvariantGroupClobberingInstruction(Instruction &I, DominatorTree &DT) {
2509   if (!I.hasMetadata(LLVMContext::MD_invariant_group) || I.isVolatile())
2510     return nullptr;
2511 
2512   // We consider bitcasts and zero GEPs to be the same pointer value. Start by
2513   // stripping bitcasts and zero GEPs, then we will recursively look at loads
2514   // and stores through bitcasts and zero GEPs.
2515   Value *PointerOperand = getLoadStorePointerOperand(&I)->stripPointerCasts();
2516 
2517   // It's not safe to walk the use list of a global value because function
2518   // passes aren't allowed to look outside their functions.
2519   // FIXME: this could be fixed by filtering instructions from outside of
2520   // current function.
2521   if (isa<Constant>(PointerOperand))
2522     return nullptr;
2523 
2524   // Queue to process all pointers that are equivalent to load operand.
2525   SmallVector<const Value *, 8> PointerUsesQueue;
2526   PointerUsesQueue.push_back(PointerOperand);
2527 
2528   const Instruction *MostDominatingInstruction = &I;
2529 
2530   // FIXME: This loop is O(n^2) because dominates can be O(n) and in worst case
2531   // we will see all the instructions. It may not matter in practice. If it
2532   // does, we will have to support MemorySSA construction and updates.
2533   while (!PointerUsesQueue.empty()) {
2534     const Value *Ptr = PointerUsesQueue.pop_back_val();
2535     assert(Ptr && !isa<GlobalValue>(Ptr) &&
2536            "Null or GlobalValue should not be inserted");
2537 
2538     for (const User *Us : Ptr->users()) {
2539       auto *U = dyn_cast<Instruction>(Us);
2540       if (!U || U == &I || !DT.dominates(U, MostDominatingInstruction))
2541         continue;
2542 
2543       // Add bitcasts and zero GEPs to queue.
2544       if (isa<BitCastInst>(U)) {
2545         PointerUsesQueue.push_back(U);
2546         continue;
2547       }
2548       if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) {
2549         if (GEP->hasAllZeroIndices())
2550           PointerUsesQueue.push_back(U);
2551         continue;
2552       }
2553 
2554       // If we hit a load/store with an invariant.group metadata and the same
2555       // pointer operand, we can assume that value pointed to by the pointer
2556       // operand didn't change.
2557       if (U->hasMetadata(LLVMContext::MD_invariant_group) &&
2558           getLoadStorePointerOperand(U) == Ptr && !U->isVolatile()) {
2559         MostDominatingInstruction = U;
2560       }
2561     }
2562   }
2563   return MostDominatingInstruction == &I ? nullptr : MostDominatingInstruction;
2564 }
2565 
2566 template <typename AliasAnalysisType>
2567 MemoryAccess *
2568 MemorySSA::ClobberWalkerBase<AliasAnalysisType>::getClobberingMemoryAccessBase(
2569     MemoryAccess *MA, unsigned &UpwardWalkLimit, bool SkipSelf,
2570     bool UseInvariantGroup) {
2571   auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
2572   // If this is a MemoryPhi, we can't do anything.
2573   if (!StartingAccess)
2574     return MA;
2575 
2576   if (UseInvariantGroup) {
2577     if (auto *I = getInvariantGroupClobberingInstruction(
2578             *StartingAccess->getMemoryInst(), MSSA->getDomTree())) {
2579       assert(isa<LoadInst>(I) || isa<StoreInst>(I));
2580 
2581       auto *ClobberMA = MSSA->getMemoryAccess(I);
2582       assert(ClobberMA);
2583       if (isa<MemoryUse>(ClobberMA))
2584         return ClobberMA->getDefiningAccess();
2585       return ClobberMA;
2586     }
2587   }
2588 
2589   bool IsOptimized = false;
2590 
2591   // If this is an already optimized use or def, return the optimized result.
2592   // Note: Currently, we store the optimized def result in a separate field,
2593   // since we can't use the defining access.
2594   if (StartingAccess->isOptimized()) {
2595     if (!SkipSelf || !isa<MemoryDef>(StartingAccess))
2596       return StartingAccess->getOptimized();
2597     IsOptimized = true;
2598   }
2599 
2600   const Instruction *I = StartingAccess->getMemoryInst();
2601   // We can't sanely do anything with a fence, since they conservatively clobber
2602   // all memory, and have no locations to get pointers from to try to
2603   // disambiguate.
2604   if (!isa<CallBase>(I) && I->isFenceLike())
2605     return StartingAccess;
2606 
2607   UpwardsMemoryQuery Q(I, StartingAccess);
2608 
2609   if (isUseTriviallyOptimizableToLiveOnEntry(*Walker.getAA(), I)) {
2610     MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
2611     StartingAccess->setOptimized(LiveOnEntry);
2612     StartingAccess->setOptimizedAccessType(None);
2613     return LiveOnEntry;
2614   }
2615 
2616   MemoryAccess *OptimizedAccess;
2617   if (!IsOptimized) {
2618     // Start with the thing we already think clobbers this location
2619     MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
2620 
2621     // At this point, DefiningAccess may be the live on entry def.
2622     // If it is, we will not get a better result.
2623     if (MSSA->isLiveOnEntryDef(DefiningAccess)) {
2624       StartingAccess->setOptimized(DefiningAccess);
2625       StartingAccess->setOptimizedAccessType(None);
2626       return DefiningAccess;
2627     }
2628 
2629     OptimizedAccess = Walker.findClobber(DefiningAccess, Q, UpwardWalkLimit);
2630     StartingAccess->setOptimized(OptimizedAccess);
2631     if (MSSA->isLiveOnEntryDef(OptimizedAccess))
2632       StartingAccess->setOptimizedAccessType(None);
2633     else if (Q.AR && *Q.AR == AliasResult::MustAlias)
2634       StartingAccess->setOptimizedAccessType(
2635           AliasResult(AliasResult::MustAlias));
2636   } else
2637     OptimizedAccess = StartingAccess->getOptimized();
2638 
2639   LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2640   LLVM_DEBUG(dbgs() << *StartingAccess << "\n");
2641   LLVM_DEBUG(dbgs() << "Optimized Memory SSA clobber for " << *I << " is ");
2642   LLVM_DEBUG(dbgs() << *OptimizedAccess << "\n");
2643 
2644   MemoryAccess *Result;
2645   if (SkipSelf && isa<MemoryPhi>(OptimizedAccess) &&
2646       isa<MemoryDef>(StartingAccess) && UpwardWalkLimit) {
2647     assert(isa<MemoryDef>(Q.OriginalAccess));
2648     Q.SkipSelfAccess = true;
2649     Result = Walker.findClobber(OptimizedAccess, Q, UpwardWalkLimit);
2650   } else
2651     Result = OptimizedAccess;
2652 
2653   LLVM_DEBUG(dbgs() << "Result Memory SSA clobber [SkipSelf = " << SkipSelf);
2654   LLVM_DEBUG(dbgs() << "] for " << *I << " is " << *Result << "\n");
2655 
2656   return Result;
2657 }
2658 
2659 MemoryAccess *
2660 DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2661   if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
2662     return Use->getDefiningAccess();
2663   return MA;
2664 }
2665 
2666 MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
2667     MemoryAccess *StartingAccess, const MemoryLocation &) {
2668   if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
2669     return Use->getDefiningAccess();
2670   return StartingAccess;
2671 }
2672 
2673 void MemoryPhi::deleteMe(DerivedUser *Self) {
2674   delete static_cast<MemoryPhi *>(Self);
2675 }
2676 
2677 void MemoryDef::deleteMe(DerivedUser *Self) {
2678   delete static_cast<MemoryDef *>(Self);
2679 }
2680 
2681 void MemoryUse::deleteMe(DerivedUser *Self) {
2682   delete static_cast<MemoryUse *>(Self);
2683 }
2684 
2685 bool upward_defs_iterator::IsGuaranteedLoopInvariant(Value *Ptr) const {
2686   auto IsGuaranteedLoopInvariantBase = [](Value *Ptr) {
2687     Ptr = Ptr->stripPointerCasts();
2688     if (!isa<Instruction>(Ptr))
2689       return true;
2690     return isa<AllocaInst>(Ptr);
2691   };
2692 
2693   Ptr = Ptr->stripPointerCasts();
2694   if (auto *I = dyn_cast<Instruction>(Ptr)) {
2695     if (I->getParent()->isEntryBlock())
2696       return true;
2697   }
2698   if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
2699     return IsGuaranteedLoopInvariantBase(GEP->getPointerOperand()) &&
2700            GEP->hasAllConstantIndices();
2701   }
2702   return IsGuaranteedLoopInvariantBase(Ptr);
2703 }
2704