LCOV - code coverage report
Current view: top level - lib/Analysis - MemorySSA.cpp (source / functions) Hit Total Coverage
Test: llvm-toolchain.info Lines: 720 774 93.0 %
Date: 2017-09-14 15:23:50 Functions: 82 96 85.4 %
Legend: Lines: hit not hit

          Line data    Source code
       1             : //===- MemorySSA.cpp - Memory SSA Builder ---------------------------------===//
       2             : //
       3             : //                     The LLVM Compiler Infrastructure
       4             : //
       5             : // This file is distributed under the University of Illinois Open Source
       6             : // License. See LICENSE.TXT for details.
       7             : //
       8             : //===----------------------------------------------------------------------===//
       9             : //
      10             : // This file implements the MemorySSA class.
      11             : //
      12             : //===----------------------------------------------------------------------===//
      13             : 
      14             : #include "llvm/Analysis/MemorySSA.h"
      15             : #include "llvm/ADT/DenseMap.h"
      16             : #include "llvm/ADT/DenseMapInfo.h"
      17             : #include "llvm/ADT/DenseSet.h"
      18             : #include "llvm/ADT/DepthFirstIterator.h"
      19             : #include "llvm/ADT/Hashing.h"
      20             : #include "llvm/ADT/None.h"
      21             : #include "llvm/ADT/Optional.h"
      22             : #include "llvm/ADT/STLExtras.h"
      23             : #include "llvm/ADT/SmallPtrSet.h"
      24             : #include "llvm/ADT/SmallVector.h"
      25             : #include "llvm/ADT/iterator.h"
      26             : #include "llvm/ADT/iterator_range.h"
      27             : #include "llvm/Analysis/AliasAnalysis.h"
      28             : #include "llvm/Analysis/IteratedDominanceFrontier.h"
      29             : #include "llvm/Analysis/MemoryLocation.h"
      30             : #include "llvm/IR/AssemblyAnnotationWriter.h"
      31             : #include "llvm/IR/BasicBlock.h"
      32             : #include "llvm/IR/CallSite.h"
      33             : #include "llvm/IR/Dominators.h"
      34             : #include "llvm/IR/Function.h"
      35             : #include "llvm/IR/Instruction.h"
      36             : #include "llvm/IR/Instructions.h"
      37             : #include "llvm/IR/IntrinsicInst.h"
      38             : #include "llvm/IR/Intrinsics.h"
      39             : #include "llvm/IR/LLVMContext.h"
      40             : #include "llvm/IR/PassManager.h"
      41             : #include "llvm/IR/Use.h"
      42             : #include "llvm/Pass.h"
      43             : #include "llvm/Support/AtomicOrdering.h"
      44             : #include "llvm/Support/Casting.h"
      45             : #include "llvm/Support/CommandLine.h"
      46             : #include "llvm/Support/Compiler.h"
      47             : #include "llvm/Support/Debug.h"
      48             : #include "llvm/Support/ErrorHandling.h"
      49             : #include "llvm/Support/FormattedStream.h"
      50             : #include "llvm/Support/raw_ostream.h"
      51             : #include <algorithm>
      52             : #include <cassert>
      53             : #include <iterator>
      54             : #include <memory>
      55             : #include <utility>
      56             : 
      57             : using namespace llvm;
      58             : 
      59             : #define DEBUG_TYPE "memoryssa"
      60             : 
      61       26360 : INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
      62             :                       true)
      63       26360 : INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
      64       26360 : INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
      65      344308 : INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
      66             :                     true)
      67             : 
      68        9158 : INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa",
      69             :                       "Memory SSA Printer", false, false)
      70        9158 : INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
      71       45850 : INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",
      72             :                     "Memory SSA Printer", false, false)
      73             : 
      74       72306 : static cl::opt<unsigned> MaxCheckLimit(
      75      216918 :     "memssa-check-limit", cl::Hidden, cl::init(100),
      76      216918 :     cl::desc("The maximum number of stores/phis MemorySSA"
      77       72306 :              "will consider trying to walk past (default = 100)"));
      78             : 
      79             : static cl::opt<bool>
      80      289224 :     VerifyMemorySSA("verify-memoryssa", cl::init(false), cl::Hidden,
      81      289224 :                     cl::desc("Verify MemorySSA in legacy printer pass."));
      82             : 
      83             : namespace llvm {
      84             : 
      85             : /// \brief An assembly annotator class to print Memory SSA information in
      86             : /// comments.
      87          82 : class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter {
      88             :   friend class MemorySSA;
      89             : 
      90             :   const MemorySSA *MSSA;
      91             : 
      92             : public:
      93          82 :   MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {}
      94             : 
      95         228 :   void emitBasicBlockStartAnnot(const BasicBlock *BB,
      96             :                                 formatted_raw_ostream &OS) override {
      97         228 :     if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
      98         138 :       OS << "; " << *MA << "\n";
      99         228 :   }
     100             : 
     101         724 :   void emitInstructionAnnot(const Instruction *I,
     102             :                             formatted_raw_ostream &OS) override {
     103         724 :     if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
     104         740 :       OS << "; " << *MA << "\n";
     105         724 :   }
     106             : };
     107             : 
     108             : } // end namespace llvm
     109             : 
     110             : namespace {
     111             : 
     112             : /// Our current alias analysis API differentiates heavily between calls and
     113             : /// non-calls, and functions called on one usually assert on the other.
     114             : /// This class encapsulates the distinction to simplify other code that wants
     115             : /// "Memory affecting instructions and related data" to use as a key.
     116             : /// For example, this class is used as a densemap key in the use optimizer.
     117             : class MemoryLocOrCall {
     118             : public:
     119             :   bool IsCall = false;
     120             : 
     121             :   MemoryLocOrCall() = default;
     122             :   MemoryLocOrCall(MemoryUseOrDef *MUD)
     123      419772 :       : MemoryLocOrCall(MUD->getMemoryInst()) {}
     124             :   MemoryLocOrCall(const MemoryUseOrDef *MUD)
     125           1 :       : MemoryLocOrCall(MUD->getMemoryInst()) {}
     126             : 
     127      419773 :   MemoryLocOrCall(Instruction *Inst) {
     128     1259319 :     if (ImmutableCallSite(Inst)) {
     129        2189 :       IsCall = true;
     130        4378 :       CS = ImmutableCallSite(Inst);
     131             :     } else {
     132             :       IsCall = false;
     133             :       // There is no such thing as a memorylocation for a fence inst, and it is
     134             :       // unique in that regard.
     135      835168 :       if (!isa<FenceInst>(Inst))
     136      835168 :         Loc = MemoryLocation::get(Inst);
     137             :     }
     138      419773 :   }
     139             : 
     140     1442276 :   explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {}
     141             : 
     142             :   ImmutableCallSite getCS() const {
     143             :     assert(IsCall);
     144             :     return CS;
     145             :   }
     146             : 
     147             :   MemoryLocation getLoc() const {
     148             :     assert(!IsCall);
     149     2897586 :     return Loc;
     150             :   }
     151             : 
     152     2408130 :   bool operator==(const MemoryLocOrCall &Other) const {
     153     2408130 :     if (IsCall != Other.IsCall)
     154             :       return false;
     155             : 
     156     2403563 :     if (IsCall)
     157        2487 :       return CS.getCalledValue() == Other.CS.getCalledValue();
     158     2402734 :     return Loc == Other.Loc;
     159             :   }
     160             : 
     161             : private:
     162             :   union {
     163             :     ImmutableCallSite CS;
     164             :     MemoryLocation Loc;
     165             :   };
     166             : };
     167             : 
     168             : } // end anonymous namespace
     169             : 
     170             : namespace llvm {
     171             : 
     172             : template <> struct DenseMapInfo<MemoryLocOrCall> {
     173             :   static inline MemoryLocOrCall getEmptyKey() {
     174     1767734 :     return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
     175             :   }
     176             : 
     177             :   static inline MemoryLocOrCall getTombstoneKey() {
     178     1116818 :     return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
     179             :   }
     180             : 
     181      556956 :   static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
     182      556956 :     if (MLOC.IsCall)
     183        4816 :       return hash_combine(MLOC.IsCall,
     184        7224 :                           DenseMapInfo<const Value *>::getHashValue(
     185        7224 :                               MLOC.getCS().getCalledValue()));
     186     1109096 :     return hash_combine(
     187     1663644 :         MLOC.IsCall, DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc()));
     188             :   }
     189             : 
     190             :   static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
     191     2408130 :     return LHS == RHS;
     192             :   }
     193             : };
     194             : 
     195             : enum class Reorderability { Always, IfNoAlias, Never };
     196             : 
     197             : } // end namespace llvm
     198             : 
     199             : /// This does one-way checks to see if Use could theoretically be hoisted above
     200             : /// MayClobber. This will not check the other way around.
     201             : ///
     202             : /// This assumes that, for the purposes of MemorySSA, Use comes directly after
     203             : /// MayClobber, with no potentially clobbering operations in between them.
     204             : /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
     205        1726 : static Reorderability getLoadReorderability(const LoadInst *Use,
     206             :                                             const LoadInst *MayClobber) {
     207        1726 :   bool VolatileUse = Use->isVolatile();
     208        1726 :   bool VolatileClobber = MayClobber->isVolatile();
     209             :   // Volatile operations may never be reordered with other volatile operations.
     210        1726 :   if (VolatileUse && VolatileClobber)
     211             :     return Reorderability::Never;
     212             : 
     213             :   // The lang ref allows reordering of volatile and non-volatile operations.
     214             :   // Whether an aliasing nonvolatile load and volatile load can be reordered,
     215             :   // though, is ambiguous. Because it may not be best to exploit this ambiguity,
     216             :   // we only allow volatile/non-volatile reordering if the volatile and
     217             :   // non-volatile operations don't alias.
     218        1726 :   Reorderability Result = VolatileUse || VolatileClobber
     219        1726 :                               ? Reorderability::IfNoAlias
     220             :                               : Reorderability::Always;
     221             : 
     222             :   // If a load is seq_cst, it cannot be moved above other loads. If its ordering
     223             :   // is weaker, it can be moved above other loads. We just need to be sure that
     224             :   // MayClobber isn't an acquire load, because loads can't be moved above
     225             :   // acquire loads.
     226             :   //
     227             :   // Note that this explicitly *does* allow the free reordering of monotonic (or
     228             :   // weaker) loads of the same address.
     229        1726 :   bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
     230        1726 :   bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
     231        1726 :                                                      AtomicOrdering::Acquire);
     232        1726 :   if (SeqCstUse || MayClobberIsAcquire)
     233             :     return Reorderability::Never;
     234        1699 :   return Result;
     235             : }
     236             : 
     237     1955750 : static bool instructionClobbersQuery(MemoryDef *MD,
     238             :                                      const MemoryLocation &UseLoc,
     239             :                                      const Instruction *UseInst,
     240             :                                      AliasAnalysis &AA) {
     241     1955750 :   Instruction *DefInst = MD->getMemoryInst();
     242             :   assert(DefInst && "Defining instruction not actually an instruction");
     243     1955750 :   ImmutableCallSite UseCS(UseInst);
     244             : 
     245      114873 :   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
     246             :     // These intrinsics will show up as affecting memory, but they are just
     247             :     // markers.
     248      114873 :     switch (II->getIntrinsicID()) {
     249       49657 :     case Intrinsic::lifetime_start:
     250       49657 :       if (UseCS)
     251             :         return false;
     252      198444 :       return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), UseLoc);
     253             :     case Intrinsic::lifetime_end:
     254             :     case Intrinsic::invariant_start:
     255             :     case Intrinsic::invariant_end:
     256             :     case Intrinsic::assume:
     257             :       return false;
     258             :     default:
     259             :       break;
     260             :     }
     261             :   }
     262             : 
     263     1843467 :   if (UseCS) {
     264        8920 :     ModRefInfo I = AA.getModRefInfo(DefInst, UseCS);
     265        8920 :     return I != MRI_NoModRef;
     266             :   }
     267             : 
     268        1726 :   if (auto *DefLoad = dyn_cast<LoadInst>(DefInst)) {
     269        1726 :     if (auto *UseLoad = dyn_cast<LoadInst>(UseInst)) {
     270        1726 :       switch (getLoadReorderability(UseLoad, DefLoad)) {
     271             :       case Reorderability::Always:
     272             :         return false;
     273          27 :       case Reorderability::Never:
     274          27 :         return true;
     275        1699 :       case Reorderability::IfNoAlias:
     276        3398 :         return !AA.isNoAlias(UseLoc, MemoryLocation::get(DefLoad));
     277             :       }
     278             :     }
     279             :   }
     280             : 
     281     3665642 :   return AA.getModRefInfo(DefInst, UseLoc) & MRI_Mod;
     282             : }
     283             : 
     284     1179431 : static bool instructionClobbersQuery(MemoryDef *MD, const MemoryUseOrDef *MU,
     285             :                                      const MemoryLocOrCall &UseMLOC,
     286             :                                      AliasAnalysis &AA) {
     287             :   // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
     288             :   // to exist while MemoryLocOrCall is pushed through places.
     289     1179431 :   if (UseMLOC.IsCall)
     290       23739 :     return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(),
     291        7913 :                                     AA);
     292     3514554 :   return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
     293     1171518 :                                   AA);
     294             : }
     295             : 
     296             : // Return true when MD may alias MU, return false otherwise.
     297           1 : bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
     298             :                                         AliasAnalysis &AA) {
     299           1 :   return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA);
     300             : }
     301             : 
     302             : namespace {
     303             : 
     304             : struct UpwardsMemoryQuery {
     305             :   // True if our original query started off as a call
     306             :   bool IsCall = false;
     307             :   // The pointer location we started the query with. This will be empty if
     308             :   // IsCall is true.
     309             :   MemoryLocation StartingLoc;
     310             :   // This is the instruction we were querying about.
     311             :   const Instruction *Inst = nullptr;
     312             :   // The MemoryAccess we actually got called with, used to test local domination
     313             :   const MemoryAccess *OriginalAccess = nullptr;
     314             : 
     315           4 :   UpwardsMemoryQuery() = default;
     316             : 
     317       92371 :   UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
     318      369484 :       : IsCall(ImmutableCallSite(Inst)), Inst(Inst), OriginalAccess(Access) {
     319       92371 :     if (!IsCall)
     320       92240 :       StartingLoc = MemoryLocation::get(Inst);
     321       92371 :   }
     322             : };
     323             : 
     324             : } // end anonymous namespace
     325             : 
     326     1171520 : static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc,
     327             :                            AliasAnalysis &AA) {
     328     1171520 :   Instruction *Inst = MD->getMemoryInst();
     329       54033 :   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
     330       54033 :     switch (II->getIntrinsicID()) {
     331       19871 :     case Intrinsic::lifetime_end:
     332       79484 :       return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), Loc);
     333             :     default:
     334             :       return false;
     335             :     }
     336             :   }
     337             :   return false;
     338             : }
     339             : 
     340      513590 : static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysis &AA,
     341             :                                                    const Instruction *I) {
     342             :   // If the memory can't be changed, then loads of the memory can't be
     343             :   // clobbered.
     344             :   //
     345             :   // FIXME: We should handle invariant groups, as well. It's a bit harder,
     346             :   // because we need to pay close attention to invariant group barriers.
     347     2037024 :   return isa<LoadInst>(I) && (I->getMetadata(LLVMContext::MD_invariant_load) ||
     348     2044900 :                               AA.pointsToConstantMemory(cast<LoadInst>(I)->
     349      513590 :                                                           getPointerOperand()));
     350             : }
     351             : 
     352             : /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
     353             : /// inbetween `Start` and `ClobberAt` can clobbers `Start`.
     354             : ///
     355             : /// This is meant to be as simple and self-contained as possible. Because it
     356             : /// uses no cache, etc., it can be relatively expensive.
     357             : ///
     358             : /// \param Start     The MemoryAccess that we want to walk from.
     359             : /// \param ClobberAt A clobber for Start.
     360             : /// \param StartLoc  The MemoryLocation for Start.
     361             : /// \param MSSA      The MemorySSA isntance that Start and ClobberAt belong to.
     362             : /// \param Query     The UpwardsMemoryQuery we used for our search.
     363             : /// \param AA        The AliasAnalysis we used for our search.
     364             : static void LLVM_ATTRIBUTE_UNUSED
     365             : checkClobberSanity(MemoryAccess *Start, MemoryAccess *ClobberAt,
     366             :                    const MemoryLocation &StartLoc, const MemorySSA &MSSA,
     367             :                    const UpwardsMemoryQuery &Query, AliasAnalysis &AA) {
     368             :   assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?");
     369             : 
     370             :   if (MSSA.isLiveOnEntryDef(Start)) {
     371             :     assert(MSSA.isLiveOnEntryDef(ClobberAt) &&
     372             :            "liveOnEntry must clobber itself");
     373             :     return;
     374             :   }
     375             : 
     376             :   bool FoundClobber = false;
     377             :   DenseSet<MemoryAccessPair> VisitedPhis;
     378             :   SmallVector<MemoryAccessPair, 8> Worklist;
     379             :   Worklist.emplace_back(Start, StartLoc);
     380             :   // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
     381             :   // is found, complain.
     382             :   while (!Worklist.empty()) {
     383             :     MemoryAccessPair MAP = Worklist.pop_back_val();
     384             :     // All we care about is that nothing from Start to ClobberAt clobbers Start.
     385             :     // We learn nothing from revisiting nodes.
     386             :     if (!VisitedPhis.insert(MAP).second)
     387             :       continue;
     388             : 
     389             :     for (MemoryAccess *MA : def_chain(MAP.first)) {
     390             :       if (MA == ClobberAt) {
     391             :         if (auto *MD = dyn_cast<MemoryDef>(MA)) {
     392             :           // instructionClobbersQuery isn't essentially free, so don't use `|=`,
     393             :           // since it won't let us short-circuit.
     394             :           //
     395             :           // Also, note that this can't be hoisted out of the `Worklist` loop,
     396             :           // since MD may only act as a clobber for 1 of N MemoryLocations.
     397             :           FoundClobber =
     398             :               FoundClobber || MSSA.isLiveOnEntryDef(MD) ||
     399             :               instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
     400             :         }
     401             :         break;
     402             :       }
     403             : 
     404             :       // We should never hit liveOnEntry, unless it's the clobber.
     405             :       assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?");
     406             : 
     407             :       if (auto *MD = dyn_cast<MemoryDef>(MA)) {
     408             :         (void)MD;
     409             :         assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) &&
     410             :                "Found clobber before reaching ClobberAt!");
     411             :         continue;
     412             :       }
     413             : 
     414             :       assert(isa<MemoryPhi>(MA));
     415             :       Worklist.append(upward_defs_begin({MA, MAP.second}), upward_defs_end());
     416             :     }
     417             :   }
     418             : 
     419             :   // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
     420             :   // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
     421             :   assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
     422             :          "ClobberAt never acted as a clobber");
     423             : }
     424             : 
     425             : namespace {
     426             : 
     427             : /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
     428             : /// in one class.
     429      102693 : class ClobberWalker {
     430             :   /// Save a few bytes by using unsigned instead of size_t.
     431             :   using ListIndex = unsigned;
     432             : 
     433             :   /// Represents a span of contiguous MemoryDefs, potentially ending in a
     434             :   /// MemoryPhi.
     435      187804 :   struct DefPath {
     436             :     MemoryLocation Loc;
     437             :     // Note that, because we always walk in reverse, Last will always dominate
     438             :     // First. Also note that First and Last are inclusive.
     439             :     MemoryAccess *First;
     440             :     MemoryAccess *Last;
     441             :     Optional<ListIndex> Previous;
     442             : 
     443             :     DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
     444             :             Optional<ListIndex> Previous)
     445      787934 :         : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
     446             : 
     447             :     DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
     448             :             Optional<ListIndex> Previous)
     449      915932 :         : DefPath(Loc, Init, Init, Previous) {}
     450             :   };
     451             : 
     452             :   const MemorySSA &MSSA;
     453             :   AliasAnalysis &AA;
     454             :   DominatorTree &DT;
     455             :   UpwardsMemoryQuery *Query;
     456             : 
     457             :   // Phi optimization bookkeeping
     458             :   SmallVector<DefPath, 32> Paths;
     459             :   DenseSet<ConstMemoryAccessPair> VisitedPhis;
     460             : 
     461             :   /// Find the nearest def or phi that `From` can legally be optimized to.
     462       80271 :   const MemoryAccess *getWalkTarget(const MemoryPhi *From) const {
     463             :     assert(From->getNumOperands() && "Phi with no operands?");
     464             : 
     465       80271 :     BasicBlock *BB = From->getBlock();
     466      160542 :     MemoryAccess *Result = MSSA.getLiveOnEntryDef();
     467      160542 :     DomTreeNode *Node = DT.getNode(BB);
     468       84195 :     while ((Node = Node->getIDom())) {
     469      166264 :       auto *Defs = MSSA.getBlockDefs(Node->getBlock());
     470       79208 :       if (Defs)
     471       79208 :         return &*Defs->rbegin();
     472             :     }
     473             :     return Result;
     474             :   }
     475             : 
     476             :   /// Result of calling walkToPhiOrClobber.
     477             :   struct UpwardsWalkResult {
     478             :     /// The "Result" of the walk. Either a clobber, the last thing we walked, or
     479             :     /// both.
     480             :     MemoryAccess *Result;
     481             :     bool IsKnownClobber;
     482             :   };
     483             : 
     484             :   /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
     485             :   /// This will update Desc.Last as it walks. It will (optionally) also stop at
     486             :   /// StopAt.
     487             :   ///
     488             :   /// This does not test for whether StopAt is a clobber
     489             :   UpwardsWalkResult
     490      239362 :   walkToPhiOrClobber(DefPath &Desc,
     491             :                      const MemoryAccess *StopAt = nullptr) const {
     492             :     assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world");
     493             : 
     494     1507801 :     for (MemoryAccess *Current : def_chain(Desc.Last)) {
     495      921478 :       Desc.Last = Current;
     496      921478 :       if (Current == StopAt)
     497       32116 :         return {Current, false};
     498             : 
     499      781763 :       if (auto *MD = dyn_cast<MemoryDef>(Current))
     500     2339845 :         if (MSSA.isLiveOnEntryDef(MD) ||
     501      776319 :             instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA))
     502       99647 :           return {MD, true};
     503             :     }
     504             : 
     505             :     assert(isa<MemoryPhi>(Desc.Last) &&
     506             :            "Ended at a non-clobber that's not a phi?");
     507      107599 :     return {Desc.Last, false};
     508             :   }
     509             : 
     510      107599 :   void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
     511             :                    ListIndex PriorNode) {
     512      215198 :     auto UpwardDefs = make_range(upward_defs_begin({Phi, Paths[PriorNode].Loc}),
     513      430396 :                                  upward_defs_end());
     514      780763 :     for (const MemoryAccessPair &P : UpwardDefs) {
     515      457966 :       PausedSearches.push_back(Paths.size());
     516      228983 :       Paths.emplace_back(P.second, P.first, PriorNode);
     517             :     }
     518      107599 :   }
     519             : 
     520             :   /// Represents a search that terminated after finding a clobber. This clobber
     521             :   /// may or may not be present in the path of defs from LastNode..SearchStart,
     522             :   /// since it may have been retrieved from cache.
     523             :   struct TerminatedPath {
     524             :     MemoryAccess *Clobber;
     525             :     ListIndex LastNode;
     526             :   };
     527             : 
     528             :   /// Get an access that keeps us from optimizing to the given phi.
     529             :   ///
     530             :   /// PausedSearches is an array of indices into the Paths array. Its incoming
     531             :   /// value is the indices of searches that stopped at the last phi optimization
     532             :   /// target. It's left in an unspecified state.
     533             :   ///
     534             :   /// If this returns None, NewPaused is a vector of searches that terminated
     535             :   /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
     536             :   Optional<TerminatedPath>
     537       80271 :   getBlockingAccess(const MemoryAccess *StopWhere,
     538             :                     SmallVectorImpl<ListIndex> &PausedSearches,
     539             :                     SmallVectorImpl<ListIndex> &NewPaused,
     540             :                     SmallVectorImpl<TerminatedPath> &Terminated) {
     541             :     assert(!PausedSearches.empty() && "No searches to continue?");
     542             : 
     543             :     // BFS vs DFS really doesn't make a difference here, so just do a DFS with
     544             :     // PausedSearches as our stack.
     545      154423 :     while (!PausedSearches.empty()) {
     546      138960 :       ListIndex PathIndex = PausedSearches.pop_back_val();
     547      277920 :       DefPath &Node = Paths[PathIndex];
     548             : 
     549             :       // If we've already visited this path with this MemoryLocation, we don't
     550             :       // need to do so again.
     551             :       //
     552             :       // NOTE: That we just drop these paths on the ground makes caching
     553             :       // behavior sporadic. e.g. given a diamond:
     554             :       //  A
     555             :       // B C
     556             :       //  D
     557             :       //
     558             :       // ...If we walk D, B, A, C, we'll only cache the result of phi
     559             :       // optimization for A, B, and D; C will be skipped because it dies here.
     560             :       // This arguably isn't the worst thing ever, since:
     561             :       //   - We generally query things in a top-down order, so if we got below D
     562             :       //     without needing cache entries for {C, MemLoc}, then chances are
     563             :       //     that those cache entries would end up ultimately unused.
     564             :       //   - We still cache things for A, so C only needs to walk up a bit.
     565             :       // If this behavior becomes problematic, we can fix without a ton of extra
     566             :       // work.
     567      416880 :       if (!VisitedPhis.insert({Node.Last, Node.Loc}).second)
     568       77154 :         continue;
     569             : 
     570      116441 :       UpwardsWalkResult Res = walkToPhiOrClobber(Node, /*StopAt=*/StopWhere);
     571      116441 :       if (Res.IsKnownClobber) {
     572             :         assert(Res.Result != StopWhere);
     573             :         // If this wasn't a cache hit, we hit a clobber when walking. That's a
     574             :         // failure.
     575       64808 :         TerminatedPath Term{Res.Result, PathIndex};
     576       64808 :         if (!MSSA.dominates(Res.Result, StopWhere))
     577       64808 :           return Term;
     578             : 
     579             :         // Otherwise, it's a valid thing to potentially optimize to.
     580           0 :         Terminated.push_back(Term);
     581           0 :         continue;
     582             :       }
     583             : 
     584       83749 :       if (Res.Result == StopWhere) {
     585             :         // We've hit our target. Save this path off for if we want to continue
     586             :         // walking.
     587       32116 :         NewPaused.push_back(PathIndex);
     588       32116 :         continue;
     589             :       }
     590             : 
     591             :       assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber");
     592       39034 :       addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
     593             :     }
     594             : 
     595       15463 :     return None;
     596             :   }
     597             : 
     598             :   template <typename T, typename Walker>
     599     3682376 :   struct generic_def_path_iterator
     600             :       : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
     601             :                                     std::forward_iterator_tag, T *> {
     602       64808 :     generic_def_path_iterator() = default;
     603      129616 :     generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
     604             : 
     605      207706 :     T &operator*() const { return curNode(); }
     606             : 
     607             :     generic_def_path_iterator &operator++() {
     608       78090 :       N = curNode().Previous;
     609             :       return *this;
     610             :     }
     611             : 
     612             :     bool operator==(const generic_def_path_iterator &O) const {
     613      142898 :       if (N.hasValue() != O.N.hasValue())
     614             :         return false;
     615           0 :       return !N.hasValue() || *N == *O.N;
     616             :     }
     617             : 
     618             :   private:
     619      857388 :     T &curNode() const { return W->Paths[*N]; }
     620             : 
     621             :     Walker *W = nullptr;
     622             :     Optional<ListIndex> N = None;
     623             :   };
     624             : 
     625             :   using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
     626             :   using const_def_path_iterator =
     627             :       generic_def_path_iterator<const DefPath, const ClobberWalker>;
     628             : 
     629             :   iterator_range<def_path_iterator> def_path(ListIndex From) {
     630      388848 :     return make_range(def_path_iterator(this, From), def_path_iterator());
     631             :   }
     632             : 
     633             :   iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
     634             :     return make_range(const_def_path_iterator(this, From),
     635             :                       const_def_path_iterator());
     636             :   }
     637             : 
     638      145236 :   struct OptznResult {
     639             :     /// The path that contains our result.
     640             :     TerminatedPath PrimaryClobber;
     641             :     /// The paths that we can legally cache back from, but that aren't
     642             :     /// necessarily the result of the Phi optimization.
     643             :     SmallVector<TerminatedPath, 4> OtherClobbers;
     644             :   };
     645             : 
     646             :   ListIndex defPathIndex(const DefPath &N) const {
     647             :     // The assert looks nicer if we don't need to do &N
     648      207706 :     const DefPath *NP = &N;
     649             :     assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&
     650             :            "Out of bounds DefPath!");
     651      415412 :     return NP - &Paths.front();
     652             :   }
     653             : 
     654             :   /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
     655             :   /// that act as legal clobbers. Note that this won't return *all* clobbers.
     656             :   ///
     657             :   /// Phi optimization algorithm tl;dr:
     658             :   ///   - Find the earliest def/phi, A, we can optimize to
     659             :   ///   - Find if all paths from the starting memory access ultimately reach A
     660             :   ///     - If not, optimization isn't possible.
     661             :   ///     - Otherwise, walk from A to another clobber or phi, A'.
     662             :   ///       - If A' is a def, we're done.
     663             :   ///       - If A' is a phi, try to optimize it.
     664             :   ///
     665             :   /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
     666             :   /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
     667       72618 :   OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
     668             :                              const MemoryLocation &Loc) {
     669             :     assert(Paths.empty() && VisitedPhis.empty() &&
     670             :            "Reset the optimization state.");
     671             : 
     672       72618 :     Paths.emplace_back(Loc, Start, Phi, None);
     673             :     // Stores how many "valid" optimization nodes we had prior to calling
     674             :     // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
     675      145236 :     auto PriorPathsSize = Paths.size();
     676             : 
     677      145236 :     SmallVector<ListIndex, 16> PausedSearches;
     678      145236 :     SmallVector<ListIndex, 8> NewPaused;
     679      145236 :     SmallVector<TerminatedPath, 4> TerminatedPaths;
     680             : 
     681       72618 :     addSearches(Phi, PausedSearches, 0);
     682             : 
     683             :     // Moves the TerminatedPath with the "most dominated" Clobber to the end of
     684             :     // Paths.
     685        7810 :     auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
     686             :       assert(!Paths.empty() && "Need a path to move");
     687       15620 :       auto Dom = Paths.begin();
     688       22901 :       for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
     689        7281 :         if (!MSSA.dominates(I->Clobber, Dom->Clobber))
     690           0 :           Dom = I;
     691       15620 :       auto Last = Paths.end() - 1;
     692        7810 :       if (Last != Dom)
     693             :         std::iter_swap(Last, Dom);
     694       80428 :     };
     695             : 
     696       72618 :     MemoryPhi *Current = Phi;
     697             :     while (true) {
     698             :       assert(!MSSA.isLiveOnEntryDef(Current) &&
     699             :              "liveOnEntry wasn't treated as a clobber?");
     700             : 
     701       80271 :       const auto *Target = getWalkTarget(Current);
     702             :       // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
     703             :       // optimization for the prior phi.
     704             :       assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {
     705             :         return MSSA.dominates(P.Clobber, Target);
     706             :       }));
     707             : 
     708             :       // FIXME: This is broken, because the Blocker may be reported to be
     709             :       // liveOnEntry, and we'll happily wait for that to disappear (read: never)
     710             :       // For the moment, this is fine, since we do nothing with blocker info.
     711       80271 :       if (Optional<TerminatedPath> Blocker = getBlockingAccess(
     712       95734 :               Target, PausedSearches, NewPaused, TerminatedPaths)) {
     713             : 
     714             :         // Find the node we started at. We can't search based on N->Last, since
     715             :         // we may have gone around a loop with a different MemoryLocation.
     716      194424 :         auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
     717      285796 :           return defPathIndex(N) < PriorPathsSize;
     718      194424 :         });
     719             :         assert(Iter != def_path_iterator());
     720             : 
     721       64808 :         DefPath &CurNode = *Iter;
     722             :         assert(CurNode.Last == Current);
     723             : 
     724             :         // Two things:
     725             :         // A. We can't reliably cache all of NewPaused back. Consider a case
     726             :         //    where we have two paths in NewPaused; one of which can't optimize
     727             :         //    above this phi, whereas the other can. If we cache the second path
     728             :         //    back, we'll end up with suboptimal cache entries. We can handle
     729             :         //    cases like this a bit better when we either try to find all
     730             :         //    clobbers that block phi optimization, or when our cache starts
     731             :         //    supporting unfinished searches.
     732             :         // B. We can't reliably cache TerminatedPaths back here without doing
     733             :         //    extra checks; consider a case like:
     734             :         //       T
     735             :         //      / \
     736             :         //     D   C
     737             :         //      \ /
     738             :         //       S
     739             :         //    Where T is our target, C is a node with a clobber on it, D is a
     740             :         //    diamond (with a clobber *only* on the left or right node, N), and
     741             :         //    S is our start. Say we walk to D, through the node opposite N
     742             :         //    (read: ignoring the clobber), and see a cache entry in the top
     743             :         //    node of D. That cache entry gets put into TerminatedPaths. We then
     744             :         //    walk up to C (N is later in our worklist), find the clobber, and
     745             :         //    quit. If we append TerminatedPaths to OtherClobbers, we'll cache
     746             :         //    the bottom part of D to the cached clobber, ignoring the clobber
     747             :         //    in N. Again, this problem goes away if we start tracking all
     748             :         //    blockers for a given phi optimization.
     749      129616 :         TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
     750      129616 :         return {Result, {}};
     751             :       }
     752             : 
     753             :       // If there's nothing left to search, then all paths led to valid clobbers
     754             :       // that we got from our cache; pick the nearest to the start, and allow
     755             :       // the rest to be cached back.
     756       15463 :       if (NewPaused.empty()) {
     757           0 :         MoveDominatedPathToEnd(TerminatedPaths);
     758           0 :         TerminatedPath Result = TerminatedPaths.pop_back_val();
     759           0 :         return {Result, std::move(TerminatedPaths)};
     760             :       }
     761             : 
     762       15463 :       MemoryAccess *DefChainEnd = nullptr;
     763       23116 :       SmallVector<TerminatedPath, 4> Clobbers;
     764       46018 :       for (ListIndex Paused : NewPaused) {
     765       61110 :         UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
     766       30555 :         if (WR.IsKnownClobber)
     767       15091 :           Clobbers.push_back({WR.Result, Paused});
     768             :         else
     769             :           // Micro-opt: If we hit the end of the chain, save it.
     770             :           DefChainEnd = WR.Result;
     771             :       }
     772             : 
     773       15463 :       if (!TerminatedPaths.empty()) {
     774             :         // If we couldn't find the dominating phi/liveOnEntry in the above loop,
     775             :         // do it now.
     776           0 :         if (!DefChainEnd)
     777           0 :           for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target)))
     778           0 :             DefChainEnd = MA;
     779             : 
     780             :         // If any of the terminated paths don't dominate the phi we'll try to
     781             :         // optimize, we need to figure out what they are and quit.
     782           0 :         const BasicBlock *ChainBB = DefChainEnd->getBlock();
     783           0 :         for (const TerminatedPath &TP : TerminatedPaths) {
     784             :           // Because we know that DefChainEnd is as "high" as we can go, we
     785             :           // don't need local dominance checks; BB dominance is sufficient.
     786           0 :           if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
     787           0 :             Clobbers.push_back(TP);
     788             :         }
     789             :       }
     790             : 
     791             :       // If we have clobbers in the def chain, find the one closest to Current
     792             :       // and quit.
     793       15463 :       if (!Clobbers.empty()) {
     794        7810 :         MoveDominatedPathToEnd(Clobbers);
     795        7810 :         TerminatedPath Result = Clobbers.pop_back_val();
     796       15620 :         return {Result, std::move(Clobbers)};
     797             :       }
     798             : 
     799             :       assert(all_of(NewPaused,
     800             :                     [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }));
     801             : 
     802             :       // Because liveOnEntry is a clobber, this must be a phi.
     803        7653 :       auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
     804             : 
     805       15306 :       PriorPathsSize = Paths.size();
     806        7653 :       PausedSearches.clear();
     807       38423 :       for (ListIndex I : NewPaused)
     808       15464 :         addSearches(DefChainPhi, PausedSearches, I);
     809        7653 :       NewPaused.clear();
     810             : 
     811        7653 :       Current = DefChainPhi;
     812             :     }
     813             :   }
     814             : 
     815             :   void verifyOptResult(const OptznResult &R) const {
     816             :     assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {
     817             :       return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);
     818             :     }));
     819             :   }
     820             : 
     821             :   void resetPhiOptznState() {
     822      145236 :     Paths.clear();
     823      145236 :     VisitedPhis.clear();
     824             :   }
     825             : 
     826             : public:
     827             :   ClobberWalker(const MemorySSA &MSSA, AliasAnalysis &AA, DominatorTree &DT)
     828      102693 :       : MSSA(MSSA), AA(AA), DT(DT) {}
     829             : 
     830             :   void reset() {}
     831             : 
     832             :   /// Finds the nearest clobber for the given query, optimizing phis if
     833             :   /// possible.
     834       92366 :   MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q) {
     835       92366 :     Query = &Q;
     836             : 
     837       92366 :     MemoryAccess *Current = Start;
     838             :     // This walker pretends uses don't exist. If we're handed one, silently grab
     839             :     // its def. (This has the nice side-effect of ensuring we never cache uses)
     840           0 :     if (auto *MU = dyn_cast<MemoryUse>(Start))
     841           0 :       Current = MU->getDefiningAccess();
     842             : 
     843      277098 :     DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
     844             :     // Fast path for the overly-common case (no crazy phi optimization
     845             :     // necessary)
     846       92366 :     UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
     847             :     MemoryAccess *Result;
     848       92366 :     if (WalkResult.IsKnownClobber) {
     849       19748 :       Result = WalkResult.Result;
     850             :     } else {
     851             :       OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
     852      217854 :                                           Current, Q.StartingLoc);
     853       72618 :       verifyOptResult(OptRes);
     854       72618 :       resetPhiOptznState();
     855       72618 :       Result = OptRes.PrimaryClobber.Clobber;
     856             :     }
     857             : 
     858             : #ifdef EXPENSIVE_CHECKS
     859             :     checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
     860             : #endif
     861      184732 :     return Result;
     862             :   }
     863             : 
     864             :   void verify(const MemorySSA *MSSA) { assert(MSSA == &this->MSSA); }
     865             : };
     866             : 
     867             : struct RenamePassData {
     868             :   DomTreeNode *DTN;
     869             :   DomTreeNode::const_iterator ChildIt;
     870             :   MemoryAccess *IncomingVal;
     871             : 
     872             :   RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
     873             :                  MemoryAccess *M)
     874      185188 :       : DTN(D), ChildIt(It), IncomingVal(M) {}
     875             : 
     876             :   void swap(RenamePassData &RHS) {
     877             :     std::swap(DTN, RHS.DTN);
     878             :     std::swap(ChildIt, RHS.ChildIt);
     879             :     std::swap(IncomingVal, RHS.IncomingVal);
     880             :   }
     881             : };
     882             : 
     883             : } // end anonymous namespace
     884             : 
     885             : namespace llvm {
     886             : 
     887             : /// \brief A MemorySSAWalker that does AA walks to disambiguate accesses. It no
     888             : /// longer does caching on its own,
     889             : /// but the name has been retained for the moment.
     890             : class MemorySSA::CachingWalker final : public MemorySSAWalker {
     891             :   ClobberWalker Walker;
     892             :   bool AutoResetWalker = true;
     893             : 
     894             :   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, UpwardsMemoryQuery &);
     895             :   void verifyRemoved(MemoryAccess *);
     896             : 
     897             : public:
     898             :   CachingWalker(MemorySSA *, AliasAnalysis *, DominatorTree *);
     899       34231 :   ~CachingWalker() override = default;
     900             : 
     901             :   using MemorySSAWalker::getClobberingMemoryAccess;
     902             : 
     903             :   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) override;
     904             :   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
     905             :                                           const MemoryLocation &) override;
     906             :   void invalidateInfo(MemoryAccess *) override;
     907             : 
     908             :   /// Whether we call resetClobberWalker() after each time we *actually* walk to
     909             :   /// answer a clobber query.
     910       68462 :   void setAutoResetWalker(bool AutoReset) { AutoResetWalker = AutoReset; }
     911             : 
     912             :   /// Drop the walker's persistent data structures.
     913       34231 :   void resetClobberWalker() { Walker.reset(); }
     914             : 
     915           0 :   void verify(const MemorySSA *MSSA) override {
     916          93 :     MemorySSAWalker::verify(MSSA);
     917          93 :     Walker.verify(MSSA);
     918           0 :   }
     919             : };
     920             : 
     921             : } // end namespace llvm
     922             : 
     923      185188 : void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal,
     924             :                                     bool RenameAllUses) {
     925             :   // Pass through values to our successors
     926      777774 :   for (const BasicBlock *S : successors(BB)) {
     927      203699 :     auto It = PerBlockAccesses.find(S);
     928             :     // Rename the phi nodes in our successor block
     929     1196769 :     if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
     930      111578 :       continue;
     931      184242 :     AccessList *Accesses = It->second.get();
     932      276363 :     auto *Phi = cast<MemoryPhi>(&Accesses->front());
     933       92121 :     if (RenameAllUses) {
     934           2 :       int PhiIndex = Phi->getBasicBlockIndex(BB);
     935             :       assert(PhiIndex != -1 && "Incomplete phi during partial rename");
     936           2 :       Phi->setIncomingValue(PhiIndex, IncomingVal);
     937             :     } else
     938       92119 :       Phi->addIncoming(IncomingVal, BB);
     939             :   }
     940      185188 : }
     941             : 
     942             : /// \brief Rename a single basic block into MemorySSA form.
     943             : /// Uses the standard SSA renaming algorithm.
     944             : /// \returns The new incoming value.
     945      185188 : MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal,
     946             :                                      bool RenameAllUses) {
     947      185188 :   auto It = PerBlockAccesses.find(BB);
     948             :   // Skip most processing if the list is empty.
     949      555564 :   if (It != PerBlockAccesses.end()) {
     950      339922 :     AccessList *Accesses = It->second.get();
     951     1793365 :     for (MemoryAccess &L : *Accesses) {
     952     1242444 :       if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
     953           5 :         if (MUD->getDefiningAccess() == nullptr || RenameAllUses)
     954             :           MUD->setDefiningAccess(IncomingVal);
     955     2484888 :         if (isa<MemoryDef>(&L))
     956      821223 :           IncomingVal = &L;
     957             :       } else {
     958             :         IncomingVal = &L;
     959             :       }
     960             :     }
     961             :   }
     962      185188 :   return IncomingVal;
     963             : }
     964             : 
     965             : /// \brief This is the standard SSA renaming algorithm.
     966             : ///
     967             : /// We walk the dominator tree in preorder, renaming accesses, and then filling
     968             : /// in phi nodes in our successors.
     969       34232 : void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
     970             :                            SmallPtrSetImpl<BasicBlock *> &Visited,
     971             :                            bool SkipVisited, bool RenameAllUses) {
     972       68464 :   SmallVector<RenamePassData, 32> WorkStack;
     973             :   // Skip everything if we already renamed this block and we are skipping.
     974             :   // Note: You can't sink this into the if, because we need it to occur
     975             :   // regardless of whether we skip blocks or not.
     976       34232 :   bool AlreadyVisited = !Visited.insert(Root->getBlock()).second;
     977       34232 :   if (SkipVisited && AlreadyVisited)
     978           0 :     return;
     979             : 
     980       34232 :   IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses);
     981       34232 :   renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses);
     982      102696 :   WorkStack.push_back({Root, Root->begin(), IncomingVal});
     983             : 
     984      370376 :   while (!WorkStack.empty()) {
     985      672288 :     DomTreeNode *Node = WorkStack.back().DTN;
     986      672288 :     DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
     987      672288 :     IncomingVal = WorkStack.back().IncomingVal;
     988             : 
     989      672288 :     if (ChildIt == Node->end()) {
     990             :       WorkStack.pop_back();
     991             :     } else {
     992      150956 :       DomTreeNode *Child = *ChildIt;
     993      452868 :       ++WorkStack.back().ChildIt;
     994      150956 :       BasicBlock *BB = Child->getBlock();
     995             :       // Note: You can't sink this into the if, because we need it to occur
     996             :       // regardless of whether we skip blocks or not.
     997      150956 :       AlreadyVisited = !Visited.insert(BB).second;
     998      150956 :       if (SkipVisited && AlreadyVisited) {
     999             :         // We already visited this during our renaming, which can happen when
    1000             :         // being asked to rename multiple blocks. Figure out the incoming val,
    1001             :         // which is the last def.
    1002             :         // Incoming value can only change if there is a block def, and in that
    1003             :         // case, it's the last block def in the list.
    1004           0 :         if (auto *BlockDefs = getWritableBlockDefs(BB))
    1005           0 :           IncomingVal = &*BlockDefs->rbegin();
    1006             :       } else
    1007      150956 :         IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses);
    1008      150956 :       renameSuccessorPhis(BB, IncomingVal, RenameAllUses);
    1009      452868 :       WorkStack.push_back({Child, Child->begin(), IncomingVal});
    1010             :     }
    1011             :   }
    1012             : }
    1013             : 
    1014             : /// \brief This handles unreachable block accesses by deleting phi nodes in
    1015             : /// unreachable blocks, and marking all other unreachable MemoryAccess's as
    1016             : /// being uses of the live on entry definition.
    1017        1685 : void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
    1018             :   assert(!DT->isReachableFromEntry(BB) &&
    1019             :          "Reachable block found while handling unreachable blocks");
    1020             : 
    1021             :   // Make sure phi nodes in our reachable successors end up with a
    1022             :   // LiveOnEntryDef for our incoming edge, even though our block is forward
    1023             :   // unreachable.  We could just disconnect these blocks from the CFG fully,
    1024             :   // but we do not right now.
    1025        5482 :   for (const BasicBlock *S : successors(BB)) {
    1026        1056 :     if (!DT->isReachableFromEntry(S))
    1027        1684 :       continue;
    1028         407 :     auto It = PerBlockAccesses.find(S);
    1029             :     // Rename the phi nodes in our successor block
    1030        2343 :     if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
    1031         386 :       continue;
    1032          42 :     AccessList *Accesses = It->second.get();
    1033          63 :     auto *Phi = cast<MemoryPhi>(&Accesses->front());
    1034          42 :     Phi->addIncoming(LiveOnEntryDef.get(), BB);
    1035             :   }
    1036             : 
    1037        1685 :   auto It = PerBlockAccesses.find(BB);
    1038        5055 :   if (It == PerBlockAccesses.end())
    1039         275 :     return;
    1040             : 
    1041        1410 :   auto &Accesses = It->second;
    1042        5640 :   for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
    1043        3705 :     auto Next = std::next(AI);
    1044             :     // If we have a phi, just remove it. We are going to replace all
    1045             :     // users with live on entry.
    1046        3705 :     if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
    1047        7410 :       UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
    1048             :     else
    1049           0 :       Accesses->erase(AI);
    1050             :     AI = Next;
    1051             :   }
    1052             : }
    1053             : 
    1054       34231 : MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
    1055             :     : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
    1056      273848 :       NextID(INVALID_MEMORYACCESS_ID) {
    1057       34231 :   buildMemorySSA();
    1058       34231 : }
    1059             : 
    1060      273848 : MemorySSA::~MemorySSA() {
    1061             :   // Drop all our references
    1062      273870 :   for (const auto &Pair : PerBlockAccesses)
    1063     1952836 :     for (MemoryAccess &MA : *Pair.second)
    1064     1268128 :       MA.dropAllReferences();
    1065       34231 : }
    1066             : 
    1067      209743 : MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
    1068     1048715 :   auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
    1069             : 
    1070      209743 :   if (Res.second)
    1071      171395 :     Res.first->second = llvm::make_unique<AccessList>();
    1072      419486 :   return Res.first->second.get();
    1073             : }
    1074             : 
    1075      205393 : MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) {
    1076     1026965 :   auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr));
    1077             : 
    1078      205393 :   if (Res.second)
    1079      167773 :     Res.first->second = llvm::make_unique<DefsList>();
    1080      410786 :   return Res.first->second.get();
    1081             : }
    1082             : 
    1083             : namespace llvm {
    1084             : 
    1085             : /// This class is a batch walker of all MemoryUse's in the program, and points
    1086             : /// their defining access at the thing that actually clobbers them.  Because it
    1087             : /// is a batch walker that touches everything, it does not operate like the
    1088             : /// other walkers.  This walker is basically performing a top-down SSA renaming
    1089             : /// pass, where the version stack is used as the cache.  This enables it to be
    1090             : /// significantly more time and memory efficient than using the regular walker,
    1091             : /// which is walking bottom-up.
    1092             : class MemorySSA::OptimizeUses {
    1093             : public:
    1094             :   OptimizeUses(MemorySSA *MSSA, MemorySSAWalker *Walker, AliasAnalysis *AA,
    1095             :                DominatorTree *DT)
    1096       34231 :       : MSSA(MSSA), Walker(Walker), AA(AA), DT(DT) {
    1097       34231 :     Walker = MSSA->getWalker();
    1098             :   }
    1099             : 
    1100             :   void optimizeUses();
    1101             : 
    1102             : private:
    1103             :   /// This represents where a given memorylocation is in the stack.
    1104             :   struct MemlocStackInfo {
    1105             :     // This essentially is keeping track of versions of the stack. Whenever
    1106             :     // the stack changes due to pushes or pops, these versions increase.
    1107             :     unsigned long StackEpoch;
    1108             :     unsigned long PopEpoch;
    1109             :     // This is the lower bound of places on the stack to check. It is equal to
    1110             :     // the place the last stack walk ended.
    1111             :     // Note: Correctness depends on this being initialized to 0, which densemap
    1112             :     // does
    1113             :     unsigned long LowerBound;
    1114             :     const BasicBlock *LowerBoundBlock;
    1115             :     // This is where the last walk for this memory location ended.
    1116             :     unsigned long LastKill;
    1117             :     bool LastKillValid;
    1118             :   };
    1119             : 
    1120             :   void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
    1121             :                            SmallVectorImpl<MemoryAccess *> &,
    1122             :                            DenseMap<MemoryLocOrCall, MemlocStackInfo> &);
    1123             : 
    1124             :   MemorySSA *MSSA;
    1125             :   MemorySSAWalker *Walker;
    1126             :   AliasAnalysis *AA;
    1127             :   DominatorTree *DT;
    1128             : };
    1129             : 
    1130             : } // end namespace llvm
    1131             : 
    1132             : /// Optimize the uses in a given block This is basically the SSA renaming
    1133             : /// algorithm, with one caveat: We are able to use a single stack for all
    1134             : /// MemoryUses.  This is because the set of *possible* reaching MemoryDefs is
    1135             : /// the same for every MemoryUse.  The *actual* clobbering MemoryDef is just
    1136             : /// going to be some position in that stack of possible ones.
    1137             : ///
    1138             : /// We track the stack positions that each MemoryLocation needs
    1139             : /// to check, and last ended at.  This is because we only want to check the
    1140             : /// things that changed since last time.  The same MemoryLocation should
    1141             : /// get clobbered by the same store (getModRefInfo does not use invariantness or
    1142             : /// things like this, and if they start, we can modify MemoryLocOrCall to
    1143             : /// include relevant data)
    1144      185184 : void MemorySSA::OptimizeUses::optimizeUsesInBlock(
    1145             :     const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
    1146             :     SmallVectorImpl<MemoryAccess *> &VersionStack,
    1147             :     DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) {
    1148             : 
    1149             :   /// If no accesses, nothing to do.
    1150      370368 :   MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
    1151      169958 :   if (Accesses == nullptr)
    1152             :     return;
    1153             : 
    1154             :   // Pop everything that doesn't dominate the current block off the stack,
    1155             :   // increment the PopEpoch to account for this.
    1156             :   while (true) {
    1157             :     assert(
    1158             :         !VersionStack.empty() &&
    1159             :         "Version stack should have liveOnEntry sentinel dominating everything");
    1160      593914 :     BasicBlock *BackBlock = VersionStack.back()->getBlock();
    1161      296957 :     if (DT->dominates(BackBlock, BB))
    1162             :       break;
    1163     1197250 :     while (VersionStack.back()->getBlock() == BackBlock)
    1164      471626 :       VersionStack.pop_back();
    1165      126999 :     ++PopEpoch;
    1166      126999 :   }
    1167             : 
    1168     1793350 :   for (MemoryAccess &MA : *Accesses) {
    1169      421219 :     auto *MU = dyn_cast<MemoryUse>(&MA);
    1170      862257 :     if (!MU) {
    1171      862257 :       VersionStack.push_back(&MA);
    1172      862257 :       ++StackEpoch;
    1173     1825077 :       continue;
    1174             :     }
    1175             : 
    1176      422666 :     if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
    1177        4341 :       MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true);
    1178        1447 :       continue;
    1179             :     }
    1180             : 
    1181      419772 :     MemoryLocOrCall UseMLOC(MU);
    1182      419772 :     auto &LocInfo = LocStackInfo[UseMLOC];
    1183             :     // If the pop epoch changed, it means we've removed stuff from top of
    1184             :     // stack due to changing blocks. We may have to reset the lower bound or
    1185             :     // last kill info.
    1186      419772 :     if (LocInfo.PopEpoch != PopEpoch) {
    1187      387390 :       LocInfo.PopEpoch = PopEpoch;
    1188      387390 :       LocInfo.StackEpoch = StackEpoch;
    1189             :       // If the lower bound was in something that no longer dominates us, we
    1190             :       // have to reset it.
    1191             :       // We can't simply track stack size, because the stack may have had
    1192             :       // pushes/pops in the meantime.
    1193             :       // XXX: This is non-optimal, but only is slower cases with heavily
    1194             :       // branching dominator trees.  To get the optimal number of queries would
    1195             :       // be to make lowerbound and lastkill a per-loc stack, and pop it until
    1196             :       // the top of that stack dominates us.  This does not seem worth it ATM.
    1197             :       // A much cheaper optimization would be to always explore the deepest
    1198             :       // branch of the dominator tree first. This will guarantee this resets on
    1199             :       // the smallest set of blocks.
    1200      452232 :       if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
    1201       64842 :           !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
    1202             :         // Reset the lower bound of things to check.
    1203             :         // TODO: Some day we should be able to reset to last kill, rather than
    1204             :         // 0.
    1205       31657 :         LocInfo.LowerBound = 0;
    1206       63314 :         LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
    1207       31657 :         LocInfo.LastKillValid = false;
    1208             :       }
    1209       32382 :     } else if (LocInfo.StackEpoch != StackEpoch) {
    1210             :       // If all that has changed is the StackEpoch, we only have to check the
    1211             :       // new things on the stack, because we've checked everything before.  In
    1212             :       // this case, the lower bound of things to check remains the same.
    1213       32133 :       LocInfo.PopEpoch = PopEpoch;
    1214       32133 :       LocInfo.StackEpoch = StackEpoch;
    1215             :     }
    1216      419772 :     if (!LocInfo.LastKillValid) {
    1217      757920 :       LocInfo.LastKill = VersionStack.size() - 1;
    1218      378960 :       LocInfo.LastKillValid = true;
    1219             :     }
    1220             : 
    1221             :     // At this point, we should have corrected last kill and LowerBound to be
    1222             :     // in bounds.
    1223             :     assert(LocInfo.LowerBound < VersionStack.size() &&
    1224             :            "Lower bound out of range");
    1225             :     assert(LocInfo.LastKill < VersionStack.size() &&
    1226             :            "Last kill info out of range");
    1227             :     // In any case, the new upper bound is the top of the stack.
    1228      839544 :     unsigned long UpperBound = VersionStack.size() - 1;
    1229             : 
    1230      938660 :     if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
    1231             :       DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("
    1232             :                    << *(MU->getMemoryInst()) << ")"
    1233             :                    << " because there are " << UpperBound - LocInfo.LowerBound
    1234             :                    << " stores to disambiguate\n");
    1235             :       // Because we did not walk, LastKill is no longer valid, as this may
    1236             :       // have been a kill.
    1237       99116 :       LocInfo.LastKillValid = false;
    1238       99116 :       continue;
    1239             :     }
    1240             :     bool FoundClobberResult = false;
    1241     2358998 :     while (UpperBound > LocInfo.LowerBound) {
    1242     3718089 :       if (isa<MemoryPhi>(VersionStack[UpperBound])) {
    1243             :         // For phis, use the walker, see where we ended up, go there
    1244       59930 :         Instruction *UseInst = MU->getMemoryInst();
    1245       59930 :         MemoryAccess *Result = Walker->getClobberingMemoryAccess(UseInst);
    1246             :         // We are guaranteed to find it or something is wrong
    1247      316861 :         while (VersionStack[UpperBound] != Result) {
    1248             :           assert(UpperBound != 0);
    1249       65667 :           --UpperBound;
    1250             :         }
    1251             :         FoundClobberResult = true;
    1252             :         break;
    1253             :       }
    1254             : 
    1255     3538299 :       MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
    1256             :       // If the lifetime of the pointer ends at this instruction, it's live on
    1257             :       // entry.
    1258     2350956 :       if (!UseMLOC.IsCall && lifetimeEndsAt(MD, UseMLOC.getLoc(), *AA)) {
    1259             :         // Reset UpperBound to liveOnEntryDef's place in the stack
    1260             :         UpperBound = 0;
    1261             :         FoundClobberResult = true;
    1262             :         break;
    1263             :       }
    1264     1179430 :       if (instructionClobbersQuery(MD, MU, UseMLOC, *AA)) {
    1265             :         FoundClobberResult = true;
    1266             :         break;
    1267             :       }
    1268     1019171 :       --UpperBound;
    1269             :     }
    1270             :     // At the end of this loop, UpperBound is either a clobber, or lower bound
    1271             :     // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
    1272      320656 :     if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
    1273      880407 :       MU->setDefiningAccess(VersionStack[UpperBound], true);
    1274             :       // We were last killed now by where we got to
    1275      293469 :       LocInfo.LastKill = UpperBound;
    1276             :     } else {
    1277             :       // Otherwise, we checked all the new ones, and now we know we can get to
    1278             :       // LastKill.
    1279       54374 :       MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true);
    1280             :     }
    1281      641312 :     LocInfo.LowerBound = VersionStack.size() - 1;
    1282      320656 :     LocInfo.LowerBoundBlock = BB;
    1283             :   }
    1284             : }
    1285             : 
    1286             : /// Optimize uses to point to their actual clobbering definitions.
    1287       34231 : void MemorySSA::OptimizeUses::optimizeUses() {
    1288       68462 :   SmallVector<MemoryAccess *, 16> VersionStack;
    1289       68462 :   DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo;
    1290       68462 :   VersionStack.push_back(MSSA->getLiveOnEntryDef());
    1291             : 
    1292       34231 :   unsigned long StackEpoch = 1;
    1293       34231 :   unsigned long PopEpoch = 1;
    1294             :   // We perform a non-recursive top-down dominator tree walk.
    1295      541523 :   for (const auto *DomNode : depth_first(DT->getRootNode()))
    1296      185184 :     optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
    1297             :                         LocStackInfo);
    1298       34231 : }
    1299             : 
    1300       34231 : void MemorySSA::placePHINodes(
    1301             :     const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks,
    1302             :     const DenseMap<const BasicBlock *, unsigned int> &BBNumbers) {
    1303             :   // Determine where our MemoryPhi's should go
    1304       68462 :   ForwardIDFCalculator IDFs(*DT);
    1305       68462 :   IDFs.setDefiningBlocks(DefiningBlocks);
    1306       68462 :   SmallVector<BasicBlock *, 32> IDFBlocks;
    1307       34231 :   IDFs.calculate(IDFBlocks);
    1308             : 
    1309      136924 :   std::sort(IDFBlocks.begin(), IDFBlocks.end(),
    1310      400256 :             [&BBNumbers](const BasicBlock *A, const BasicBlock *B) {
    1311      400256 :               return BBNumbers.lookup(A) < BBNumbers.lookup(B);
    1312      200128 :             });
    1313             : 
    1314             :   // Now place MemoryPhi nodes.
    1315      143730 :   for (auto &BB : IDFBlocks)
    1316       41037 :     createMemoryPhi(BB);
    1317       34231 : }
    1318             : 
    1319       34231 : void MemorySSA::buildMemorySSA() {
    1320             :   // We create an access to represent "live on entry", for things like
    1321             :   // arguments or users of globals, where the memory they use is defined before
    1322             :   // the beginning of the function. We do not actually insert it into the IR.
    1323             :   // We do not define a live on exit for the immediate uses, and thus our
    1324             :   // semantics do *not* imply that something with no immediate uses can simply
    1325             :   // be removed.
    1326       68462 :   BasicBlock &StartingPoint = F.getEntryBlock();
    1327       34231 :   LiveOnEntryDef =
    1328       68462 :       llvm::make_unique<MemoryDef>(F.getContext(), nullptr, nullptr,
    1329      102693 :                                    &StartingPoint, NextID++);
    1330       68462 :   DenseMap<const BasicBlock *, unsigned int> BBNumbers;
    1331       34231 :   unsigned NextBBNum = 0;
    1332             : 
    1333             :   // We maintain lists of memory accesses per-block, trading memory for time. We
    1334             :   // could just look up the memory access for every possible instruction in the
    1335             :   // stream.
    1336       68462 :   SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
    1337             :   // Go through each block, figure out where defs occur, and chain together all
    1338             :   // the accesses.
    1339      289562 :   for (BasicBlock &B : F) {
    1340      373738 :     BBNumbers[&B] = NextBBNum++;
    1341      186869 :     bool InsertIntoDef = false;
    1342      186869 :     AccessList *Accesses = nullptr;
    1343      186869 :     DefsList *Defs = nullptr;
    1344     2803257 :     for (Instruction &I : B) {
    1345     2242650 :       MemoryUseOrDef *MUD = createNewAccess(&I);
    1346     2242650 :       if (!MUD)
    1347      996506 :         continue;
    1348             : 
    1349     1246144 :       if (!Accesses)
    1350      168639 :         Accesses = getOrCreateAccessList(&B);
    1351     2492288 :       Accesses->push_back(MUD);
    1352     2492288 :       if (isa<MemoryDef>(MUD)) {
    1353      824186 :         InsertIntoDef = true;
    1354      824186 :         if (!Defs)
    1355      164330 :           Defs = getOrCreateDefsList(&B);
    1356      824186 :         Defs->push_back(*MUD);
    1357             :       }
    1358             :     }
    1359      186869 :     if (InsertIntoDef)
    1360      164330 :       DefiningBlocks.insert(&B);
    1361             :   }
    1362       34231 :   placePHINodes(DefiningBlocks, BBNumbers);
    1363             : 
    1364             :   // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
    1365             :   // filled in with all blocks.
    1366       68462 :   SmallPtrSet<BasicBlock *, 16> Visited;
    1367       68462 :   renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
    1368             : 
    1369       34231 :   CachingWalker *Walker = getWalkerImpl();
    1370             : 
    1371             :   // We're doing a batch of updates; don't drop useful caches between them.
    1372       34231 :   Walker->setAutoResetWalker(false);
    1373       68462 :   OptimizeUses(this, Walker, AA, DT).optimizeUses();
    1374       34231 :   Walker->setAutoResetWalker(true);
    1375       34231 :   Walker->resetClobberWalker();
    1376             : 
    1377             :   // Mark the uses in unreachable blocks as live on entry, so that they go
    1378             :   // somewhere.
    1379      289562 :   for (auto &BB : F)
    1380      186869 :     if (!Visited.count(&BB))
    1381        1685 :       markUnreachableAsLiveOnEntry(&BB);
    1382       34231 : }
    1383             : 
    1384      121245 : MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
    1385             : 
    1386      155476 : MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() {
    1387      310952 :   if (Walker)
    1388             :     return Walker.get();
    1389             : 
    1390      136924 :   Walker = llvm::make_unique<CachingWalker>(this, AA, DT);
    1391       68462 :   return Walker.get();
    1392             : }
    1393             : 
    1394             : // This is a helper function used by the creation routines. It places NewAccess
    1395             : // into the access and defs lists for a given basic block, at the given
    1396             : // insertion point.
    1397       41104 : void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess,
    1398             :                                         const BasicBlock *BB,
    1399             :                                         InsertionPlace Point) {
    1400       41104 :   auto *Accesses = getOrCreateAccessList(BB);
    1401       41104 :   if (Point == Beginning) {
    1402             :     // If it's a phi node, it goes first, otherwise, it goes after any phi
    1403             :     // nodes.
    1404       82102 :     if (isa<MemoryPhi>(NewAccess)) {
    1405       82082 :       Accesses->push_front(NewAccess);
    1406       41041 :       auto *Defs = getOrCreateDefsList(BB);
    1407       41041 :       Defs->push_front(*NewAccess);
    1408             :     } else {
    1409             :       auto AI = find_if_not(
    1410          23 :           *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
    1411          20 :       Accesses->insert(AI, NewAccess);
    1412          20 :       if (!isa<MemoryUse>(NewAccess)) {
    1413           4 :         auto *Defs = getOrCreateDefsList(BB);
    1414             :         auto DI = find_if_not(
    1415           8 :             *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
    1416           8 :         Defs->insert(DI, *NewAccess);
    1417             :       }
    1418             :     }
    1419             :   } else {
    1420         106 :     Accesses->push_back(NewAccess);
    1421         106 :     if (!isa<MemoryUse>(NewAccess)) {
    1422          12 :       auto *Defs = getOrCreateDefsList(BB);
    1423          12 :       Defs->push_back(*NewAccess);
    1424             :     }
    1425             :   }
    1426       82208 :   BlockNumberingValid.erase(BB);
    1427       41104 : }
    1428             : 
    1429           6 : void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB,
    1430             :                                       AccessList::iterator InsertPt) {
    1431           6 :   auto *Accesses = getWritableBlockAccesses(BB);
    1432          12 :   bool WasEnd = InsertPt == Accesses->end();
    1433          12 :   Accesses->insert(AccessList::iterator(InsertPt), What);
    1434          12 :   if (!isa<MemoryUse>(What)) {
    1435           6 :     auto *Defs = getOrCreateDefsList(BB);
    1436             :     // If we got asked to insert at the end, we have an easy job, just shove it
    1437             :     // at the end. If we got asked to insert before an existing def, we also get
    1438             :     // an terator. If we got asked to insert before a use, we have to hunt for
    1439             :     // the next def.
    1440           6 :     if (WasEnd) {
    1441           3 :       Defs->push_back(*What);
    1442           3 :     } else if (isa<MemoryDef>(InsertPt)) {
    1443           6 :       Defs->insert(InsertPt->getDefsIterator(), *What);
    1444             :     } else {
    1445           5 :       while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt))
    1446             :         ++InsertPt;
    1447             :       // Either we found a def, or we are inserting at the end
    1448           2 :       if (InsertPt == Accesses->end())
    1449           1 :         Defs->push_back(*What);
    1450             :       else
    1451           0 :         Defs->insert(InsertPt->getDefsIterator(), *What);
    1452             :     }
    1453             :   }
    1454          12 :   BlockNumberingValid.erase(BB);
    1455           6 : }
    1456             : 
    1457             : // Move What before Where in the IR.  The end result is taht What will belong to
    1458             : // the right lists and have the right Block set, but will not otherwise be
    1459             : // correct. It will not have the right defining access, and if it is a def,
    1460             : // things below it will not properly be updated.
    1461           4 : void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
    1462             :                        AccessList::iterator Where) {
    1463             :   // Keep it in the lookup tables, remove from the lists
    1464           4 :   removeFromLists(What, false);
    1465           8 :   What->setBlock(BB);
    1466           4 :   insertIntoListsBefore(What, BB, Where);
    1467           4 : }
    1468             : 
    1469          52 : void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
    1470             :                        InsertionPlace Point) {
    1471          52 :   removeFromLists(What, false);
    1472         104 :   What->setBlock(BB);
    1473          52 :   insertIntoListsForBlock(What, BB, Point);
    1474          52 : }
    1475             : 
    1476       41041 : MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
    1477             :   assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB");
    1478       82082 :   MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
    1479             :   // Phi's always are placed at the front of the block.
    1480       41041 :   insertIntoListsForBlock(Phi, BB, Beginning);
    1481       82082 :   ValueToMemoryAccess[BB] = Phi;
    1482       41041 :   return Phi;
    1483             : }
    1484             : 
    1485          13 : MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I,
    1486             :                                                MemoryAccess *Definition) {
    1487             :   assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI");
    1488          13 :   MemoryUseOrDef *NewAccess = createNewAccess(I);
    1489             :   assert(
    1490             :       NewAccess != nullptr &&
    1491             :       "Tried to create a memory access for a non-memory touching instruction");
    1492          13 :   NewAccess->setDefiningAccess(Definition);
    1493          13 :   return NewAccess;
    1494             : }
    1495             : 
    1496             : // Return true if the instruction has ordering constraints.
    1497             : // Note specifically that this only considers stores and loads
    1498             : // because others are still considered ModRef by getModRefInfo.
    1499     1422701 : static inline bool isOrdered(const Instruction *I) {
    1500           0 :   if (auto *SI = dyn_cast<StoreInst>(I)) {
    1501             :     if (!SI->isUnordered())
    1502             :       return true;
    1503      424043 :   } else if (auto *LI = dyn_cast<LoadInst>(I)) {
    1504             :     if (!LI->isUnordered())
    1505             :       return true;
    1506             :   }
    1507             :   return false;
    1508             : }
    1509             : 
    1510             : /// \brief Helper function to create new memory accesses
    1511     2242663 : MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I) {
    1512             :   // The assume intrinsic has a control dependency which we model by claiming
    1513             :   // that it writes arbitrarily. Ignore that fake memory dependency here.
    1514             :   // FIXME: Replace this special casing with a more accurate modelling of
    1515             :   // assume's control dependency.
    1516      122219 :   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
    1517      122219 :     if (II->getIntrinsicID() == Intrinsic::assume)
    1518             :       return nullptr;
    1519             : 
    1520             :   // Find out what affect this instruction has on memory.
    1521     6727878 :   ModRefInfo ModRef = AA->getModRefInfo(I, None);
    1522             :   // The isOrdered check is used to ensure that volatiles end up as defs
    1523             :   // (atomics end up as ModRef right now anyway).  Until we separate the
    1524             :   // ordering chain from the memory chain, this enables people to see at least
    1525             :   // some relative ordering to volatiles.  Note that getClobberingMemoryAccess
    1526             :   // will still give an answer that bypasses other volatile loads.  TODO:
    1527             :   // Separate memory aliasing and ordering into two different chains so that we
    1528             :   // can precisely represent both "what memory will this read/write/is clobbered
    1529             :   // by" and "what instructions can I move this past".
    1530     2242626 :   bool Def = bool(ModRef & MRI_Mod) || isOrdered(I);
    1531     2242626 :   bool Use = bool(ModRef & MRI_Ref);
    1532             : 
    1533             :   // It's possible for an instruction to not modify memory at all. During
    1534             :   // construction, we ignore them.
    1535     2242626 :   if (!Def && !Use)
    1536             :     return nullptr;
    1537             : 
    1538             :   assert((Def || Use) &&
    1539             :          "Trying to create a memory access with a non-memory instruction");
    1540             : 
    1541             :   MemoryUseOrDef *MUD;
    1542     1246157 :   if (Def)
    1543     2472576 :     MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
    1544             :   else
    1545     1265895 :     MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
    1546     2492314 :   ValueToMemoryAccess[I] = MUD;
    1547     1246157 :   return MUD;
    1548             : }
    1549             : 
    1550             : /// \brief Returns true if \p Replacer dominates \p Replacee .
    1551           0 : bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
    1552             :                              const MemoryAccess *Replacee) const {
    1553           0 :   if (isa<MemoryUseOrDef>(Replacee))
    1554           0 :     return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
    1555           0 :   const auto *MP = cast<MemoryPhi>(Replacee);
    1556             :   // For a phi node, the use occurs in the predecessor block of the phi node.
    1557             :   // Since we may occur multiple times in the phi node, we have to check each
    1558             :   // operand to ensure Replacer dominates each operand where Replacee occurs.
    1559           0 :   for (const Use &Arg : MP->operands()) {
    1560           0 :     if (Arg.get() != Replacee &&
    1561           0 :         !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
    1562             :       return false;
    1563             :   }
    1564             :   return true;
    1565             : }
    1566             : 
    1567             : /// \brief Properly remove \p MA from all of MemorySSA's lookup tables.
    1568       19070 : void MemorySSA::removeFromLookups(MemoryAccess *MA) {
    1569             :   assert(MA->use_empty() &&
    1570             :          "Trying to remove memory access that still has uses");
    1571       19070 :   BlockNumbering.erase(MA);
    1572       38125 :   if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA))
    1573             :     MUD->setDefiningAccess(nullptr);
    1574             :   // Invalidate our walker's cache if necessary
    1575       38140 :   if (!isa<MemoryUse>(MA))
    1576         714 :     Walker->invalidateInfo(MA);
    1577             :   // The call below to erase will destroy MA, so we can't change the order we
    1578             :   // are doing things here
    1579             :   Value *MemoryInst;
    1580       38125 :   if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
    1581       19055 :     MemoryInst = MUD->getMemoryInst();
    1582             :   } else {
    1583          15 :     MemoryInst = MA->getBlock();
    1584             :   }
    1585       19070 :   auto VMA = ValueToMemoryAccess.find(MemoryInst);
    1586       19070 :   if (VMA->second == MA)
    1587       19068 :     ValueToMemoryAccess.erase(VMA);
    1588       19070 : }
    1589             : 
    1590             : /// \brief Properly remove \p MA from all of MemorySSA's lists.
    1591             : ///
    1592             : /// Because of the way the intrusive list and use lists work, it is important to
    1593             : /// do removal in the right order.
    1594             : /// ShouldDelete defaults to true, and will cause the memory access to also be
    1595             : /// deleted, not just removed.
    1596       19126 : void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) {
    1597             :   // The access list owns the reference, so we erase it from the non-owning list
    1598             :   // first.
    1599       38252 :   if (!isa<MemoryUse>(MA)) {
    1600         373 :     auto DefsIt = PerBlockDefs.find(MA->getBlock());
    1601         373 :     std::unique_ptr<DefsList> &Defs = DefsIt->second;
    1602         746 :     Defs->remove(*MA);
    1603         746 :     if (Defs->empty())
    1604          55 :       PerBlockDefs.erase(DefsIt);
    1605             :   }
    1606             : 
    1607             :   // The erase call here will delete it. If we don't want it deleted, we call
    1608             :   // remove instead.
    1609       19126 :   auto AccessIt = PerBlockAccesses.find(MA->getBlock());
    1610       19126 :   std::unique_ptr<AccessList> &Accesses = AccessIt->second;
    1611       19126 :   if (ShouldDelete)
    1612       19070 :     Accesses->erase(MA);
    1613             :   else
    1614          56 :     Accesses->remove(MA);
    1615             : 
    1616       38252 :   if (Accesses->empty())
    1617         218 :     PerBlockAccesses.erase(AccessIt);
    1618       19126 : }
    1619             : 
    1620          82 : void MemorySSA::print(raw_ostream &OS) const {
    1621         164 :   MemorySSAAnnotatedWriter Writer(this);
    1622          82 :   F.print(OS, &Writer);
    1623          82 : }
    1624             : 
    1625             : #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    1626             : LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); }
    1627             : #endif
    1628             : 
    1629          93 : void MemorySSA::verifyMemorySSA() const {
    1630          93 :   verifyDefUses(F);
    1631          93 :   verifyDomination(F);
    1632          93 :   verifyOrdering(F);
    1633         279 :   Walker->verify(this);
    1634          93 : }
    1635             : 
    1636             : /// \brief Verify that the order and existence of MemoryAccesses matches the
    1637             : /// order and existence of memory affecting instructions.
    1638          93 : void MemorySSA::verifyOrdering(Function &F) const {
    1639             :   // Walk all the blocks, comparing what the lookups think and what the access
    1640             :   // lists think, as well as the order in the blocks vs the order in the access
    1641             :   // lists.
    1642         186 :   SmallVector<MemoryAccess *, 32> ActualAccesses;
    1643         186 :   SmallVector<MemoryAccess *, 32> ActualDefs;
    1644         563 :   for (BasicBlock &B : F) {
    1645         284 :     const AccessList *AL = getBlockAccesses(&B);
    1646         284 :     const auto *DL = getBlockDefs(&B);
    1647         284 :     MemoryAccess *Phi = getMemoryAccess(&B);
    1648         284 :     if (Phi) {
    1649          81 :       ActualAccesses.push_back(Phi);
    1650          81 :       ActualDefs.push_back(Phi);
    1651             :     }
    1652             : 
    1653        1654 :     for (Instruction &I : B) {
    1654         802 :       MemoryAccess *MA = getMemoryAccess(&I);
    1655             :       assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&
    1656             :              "We have memory affecting instructions "
    1657             :              "in this block but they are not in the "
    1658             :              "access list or defs list");
    1659         802 :       if (MA) {
    1660         402 :         ActualAccesses.push_back(MA);
    1661         804 :         if (isa<MemoryDef>(MA))
    1662         234 :           ActualDefs.push_back(MA);
    1663             :       }
    1664             :     }
    1665             :     // Either we hit the assert, really have no accesses, or we have both
    1666             :     // accesses and an access list.
    1667             :     // Same with defs.
    1668         284 :     if (!AL && !DL)
    1669          72 :       continue;
    1670             :     assert(AL->size() == ActualAccesses.size() &&
    1671             :            "We don't have the same number of accesses in the block as on the "
    1672             :            "access list");
    1673             :     assert((DL || ActualDefs.size() == 0) &&
    1674             :            "Either we should have a defs list, or we should have no defs");
    1675             :     assert((!DL || DL->size() == ActualDefs.size()) &&
    1676             :            "We don't have the same number of defs in the block as on the "
    1677             :            "def list");
    1678         424 :     auto ALI = AL->begin();
    1679         212 :     auto AAI = ActualAccesses.begin();
    1680        2356 :     while (ALI != AL->end() && AAI != ActualAccesses.end()) {
    1681             :       assert(&*ALI == *AAI && "Not the same accesses in the same order");
    1682         483 :       ++ALI;
    1683         483 :       ++AAI;
    1684             :     }
    1685         212 :     ActualAccesses.clear();
    1686         212 :     if (DL) {
    1687         205 :       auto DLI = DL->begin();
    1688         205 :       auto ADI = ActualDefs.begin();
    1689        1150 :       while (DLI != DL->end() && ADI != ActualDefs.end()) {
    1690             :         assert(&*DLI == *ADI && "Not the same defs in the same order");
    1691         315 :         ++DLI;
    1692         315 :         ++ADI;
    1693             :       }
    1694             :     }
    1695         212 :     ActualDefs.clear();
    1696             :   }
    1697          93 : }
    1698             : 
    1699             : /// \brief Verify the domination properties of MemorySSA by checking that each
    1700             : /// definition dominates all of its uses.
    1701          93 : void MemorySSA::verifyDomination(Function &F) const {
    1702             : #ifndef NDEBUG
    1703             :   for (BasicBlock &B : F) {
    1704             :     // Phi nodes are attached to basic blocks
    1705             :     if (MemoryPhi *MP = getMemoryAccess(&B))
    1706             :       for (const Use &U : MP->uses())
    1707             :         assert(dominates(MP, U) && "Memory PHI does not dominate it's uses");
    1708             : 
    1709             :     for (Instruction &I : B) {
    1710             :       MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
    1711             :       if (!MD)
    1712             :         continue;
    1713             : 
    1714             :       for (const Use &U : MD->uses())
    1715             :         assert(dominates(MD, U) && "Memory Def does not dominate it's uses");
    1716             :     }
    1717             :   }
    1718             : #endif
    1719          93 : }
    1720             : 
    1721             : /// \brief Verify the def-use lists in MemorySSA, by verifying that \p Use
    1722             : /// appears in the use list of \p Def.
    1723         585 : void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
    1724             : #ifndef NDEBUG
    1725             :   // The live on entry use may cause us to get a NULL def here
    1726             :   if (!Def)
    1727             :     assert(isLiveOnEntryDef(Use) &&
    1728             :            "Null def but use not point to live on entry def");
    1729             :   else
    1730             :     assert(is_contained(Def->users(), Use) &&
    1731             :            "Did not find use in def's use list");
    1732             : #endif
    1733         585 : }
    1734             : 
    1735             : /// \brief Verify the immediate use information, by walking all the memory
    1736             : /// accesses and verifying that, for each use, it appears in the
    1737             : /// appropriate def's use list
    1738          93 : void MemorySSA::verifyDefUses(Function &F) const {
    1739         563 :   for (BasicBlock &B : F) {
    1740             :     // Phi nodes are attached to basic blocks
    1741         284 :     if (MemoryPhi *Phi = getMemoryAccess(&B)) {
    1742             :       assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(
    1743             :                                           pred_begin(&B), pred_end(&B))) &&
    1744             :              "Incomplete MemoryPhi Node");
    1745         345 :       for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
    1746         183 :         verifyUseInDefs(Phi->getIncomingValue(I), Phi);
    1747             :     }
    1748             : 
    1749        1654 :     for (Instruction &I : B) {
    1750         802 :       if (MemoryUseOrDef *MA = getMemoryAccess(&I)) {
    1751         402 :         verifyUseInDefs(MA->getDefiningAccess(), MA);
    1752             :       }
    1753             :     }
    1754             :   }
    1755          93 : }
    1756             : 
    1757      379736 : MemoryUseOrDef *MemorySSA::getMemoryAccess(const Instruction *I) const {
    1758     1139208 :   return cast_or_null<MemoryUseOrDef>(ValueToMemoryAccess.lookup(I));
    1759             : }
    1760             : 
    1761        2327 : MemoryPhi *MemorySSA::getMemoryAccess(const BasicBlock *BB) const {
    1762        9308 :   return cast_or_null<MemoryPhi>(ValueToMemoryAccess.lookup(cast<Value>(BB)));
    1763             : }
    1764             : 
    1765             : /// Perform a local numbering on blocks so that instruction ordering can be
    1766             : /// determined in constant time.
    1767             : /// TODO: We currently just number in order.  If we numbered by N, we could
    1768             : /// allow at least N-1 sequences of insertBefore or insertAfter (and at least
    1769             : /// log2(N) sequences of mixed before and after) without needing to invalidate
    1770             : /// the numbering.
    1771        3874 : void MemorySSA::renumberBlock(const BasicBlock *B) const {
    1772             :   // The pre-increment ensures the numbers really start at 1.
    1773        3874 :   unsigned long CurrentNumber = 0;
    1774        3874 :   const AccessList *AL = getBlockAccesses(B);
    1775             :   assert(AL != nullptr && "Asking to renumber an empty block");
    1776       93320 :   for (const auto &I : *AL)
    1777      163396 :     BlockNumbering[&I] = ++CurrentNumber;
    1778        3874 :   BlockNumberingValid.insert(B);
    1779        3874 : }
    1780             : 
    1781             : /// \brief Determine, for two memory accesses in the same block,
    1782             : /// whether \p Dominator dominates \p Dominatee.
    1783             : /// \returns True if \p Dominator dominates \p Dominatee.
    1784        8378 : bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
    1785             :                                  const MemoryAccess *Dominatee) const {
    1786        8378 :   const BasicBlock *DominatorBlock = Dominator->getBlock();
    1787             : 
    1788             :   assert((DominatorBlock == Dominatee->getBlock()) &&
    1789             :          "Asking for local domination when accesses are in different blocks!");
    1790             :   // A node dominates itself.
    1791        8378 :   if (Dominatee == Dominator)
    1792             :     return true;
    1793             : 
    1794             :   // When Dominatee is defined on function entry, it is not dominated by another
    1795             :   // memory access.
    1796        8378 :   if (isLiveOnEntryDef(Dominatee))
    1797             :     return false;
    1798             : 
    1799             :   // When Dominator is defined on function entry, it dominates the other memory
    1800             :   // access.
    1801        8378 :   if (isLiveOnEntryDef(Dominator))
    1802             :     return true;
    1803             : 
    1804        7449 :   if (!BlockNumberingValid.count(DominatorBlock))
    1805        3874 :     renumberBlock(DominatorBlock);
    1806             : 
    1807       14898 :   unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
    1808             :   // All numbers start with 1
    1809             :   assert(DominatorNum != 0 && "Block was not numbered properly");
    1810       14898 :   unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
    1811             :   assert(DominateeNum != 0 && "Block was not numbered properly");
    1812        7449 :   return DominatorNum < DominateeNum;
    1813             : }
    1814             : 
    1815      158778 : bool MemorySSA::dominates(const MemoryAccess *Dominator,
    1816             :                           const MemoryAccess *Dominatee) const {
    1817      158778 :   if (Dominator == Dominatee)
    1818             :     return true;
    1819             : 
    1820      139713 :   if (isLiveOnEntryDef(Dominatee))
    1821             :     return false;
    1822             : 
    1823      138788 :   if (Dominator->getBlock() != Dominatee->getBlock())
    1824      130412 :     return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
    1825        8376 :   return locallyDominates(Dominator, Dominatee);
    1826             : }
    1827             : 
    1828           0 : bool MemorySSA::dominates(const MemoryAccess *Dominator,
    1829             :                           const Use &Dominatee) const {
    1830           0 :   if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
    1831           0 :     BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
    1832             :     // The def must dominate the incoming block of the phi.
    1833           0 :     if (UseBB != Dominator->getBlock())
    1834           0 :       return DT->dominates(Dominator->getBlock(), UseBB);
    1835             :     // If the UseBB and the DefBB are the same, compare locally.
    1836           0 :     return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
    1837             :   }
    1838             :   // If it's not a PHI node use, the normal dominates can already handle it.
    1839           0 :   return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
    1840             : }
    1841             : 
    1842             : const static char LiveOnEntryStr[] = "liveOnEntry";
    1843             : 
    1844         439 : void MemoryAccess::print(raw_ostream &OS) const {
    1845         878 :   switch (getValueID()) {
    1846          69 :   case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS);
    1847         216 :   case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS);
    1848         154 :   case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS);
    1849             :   }
    1850           0 :   llvm_unreachable("invalid value id");
    1851             : }
    1852             : 
    1853         216 : void MemoryDef::print(raw_ostream &OS) const {
    1854         432 :   MemoryAccess *UO = getDefiningAccess();
    1855             : 
    1856         432 :   OS << getID() << " = MemoryDef(";
    1857         432 :   if (UO && UO->getID())
    1858         144 :     OS << UO->getID();
    1859             :   else
    1860          72 :     OS << LiveOnEntryStr;
    1861         216 :   OS << ')';
    1862         216 : }
    1863             : 
    1864          69 : void MemoryPhi::print(raw_ostream &OS) const {
    1865          69 :   bool First = true;
    1866         138 :   OS << getID() << " = MemoryPhi(";
    1867         297 :   for (const auto &Op : operands()) {
    1868         159 :     BasicBlock *BB = getIncomingBlock(Op);
    1869         159 :     MemoryAccess *MA = cast<MemoryAccess>(Op);
    1870         159 :     if (!First)
    1871             :       OS << ',';
    1872             :     else
    1873             :       First = false;
    1874             : 
    1875         159 :     OS << '{';
    1876         318 :     if (BB->hasName())
    1877         145 :       OS << BB->getName();
    1878             :     else
    1879          14 :       BB->printAsOperand(OS, false);
    1880         159 :     OS << ',';
    1881         159 :     if (unsigned ID = MA->getID())
    1882             :       OS << ID;
    1883             :     else
    1884          16 :       OS << LiveOnEntryStr;
    1885         159 :     OS << '}';
    1886             :   }
    1887          69 :   OS << ')';
    1888          69 : }
    1889             : 
    1890         154 : void MemoryUse::print(raw_ostream &OS) const {
    1891         308 :   MemoryAccess *UO = getDefiningAccess();
    1892         154 :   OS << "MemoryUse(";
    1893         308 :   if (UO && UO->getID())
    1894         128 :     OS << UO->getID();
    1895             :   else
    1896          26 :     OS << LiveOnEntryStr;
    1897         154 :   OS << ')';
    1898         154 : }
    1899             : 
    1900           0 : void MemoryAccess::dump() const {
    1901             : // Cannot completely remove virtual function even in release mode.
    1902             : #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    1903             :   print(dbgs());
    1904             :   dbgs() << "\n";
    1905             : #endif
    1906           0 : }
    1907             : 
    1908             : char MemorySSAPrinterLegacyPass::ID = 0;
    1909             : 
    1910          40 : MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) {
    1911          20 :   initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry());
    1912          20 : }
    1913             : 
    1914          20 : void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
    1915          40 :   AU.setPreservesAll();
    1916          20 :   AU.addRequired<MemorySSAWrapperPass>();
    1917          20 : }
    1918             : 
    1919          45 : bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) {
    1920          90 :   auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
    1921          45 :   MSSA.print(dbgs());
    1922          45 :   if (VerifyMemorySSA)
    1923          44 :     MSSA.verifyMemorySSA();
    1924          45 :   return false;
    1925             : }
    1926             : 
    1927             : AnalysisKey MemorySSAAnalysis::Key;
    1928             : 
    1929         158 : MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F,
    1930             :                                                  FunctionAnalysisManager &AM) {
    1931         158 :   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
    1932         158 :   auto &AA = AM.getResult<AAManager>(F);
    1933         474 :   return MemorySSAAnalysis::Result(llvm::make_unique<MemorySSA>(F, &AA, &DT));
    1934             : }
    1935             : 
    1936          36 : PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
    1937             :                                             FunctionAnalysisManager &AM) {
    1938          36 :   OS << "MemorySSA for function: " << F.getName() << "\n";
    1939          72 :   AM.getResult<MemorySSAAnalysis>(F).getMSSA().print(OS);
    1940             : 
    1941          36 :   return PreservedAnalyses::all();
    1942             : }
    1943             : 
    1944          33 : PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
    1945             :                                              FunctionAnalysisManager &AM) {
    1946          66 :   AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
    1947             : 
    1948          33 :   return PreservedAnalyses::all();
    1949             : }
    1950             : 
    1951             : char MemorySSAWrapperPass::ID = 0;
    1952             : 
    1953        3222 : MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
    1954        1074 :   initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
    1955        1074 : }
    1956             : 
    1957       68110 : void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
    1958             : 
    1959        1074 : void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
    1960        2148 :   AU.setPreservesAll();
    1961        1074 :   AU.addRequiredTransitive<DominatorTreeWrapperPass>();
    1962        1074 :   AU.addRequiredTransitive<AAResultsWrapperPass>();
    1963        1074 : }
    1964             : 
    1965       34055 : bool MemorySSAWrapperPass::runOnFunction(Function &F) {
    1966       68110 :   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
    1967       68110 :   auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
    1968       68110 :   MSSA.reset(new MemorySSA(F, &AA, &DT));
    1969       34055 :   return false;
    1970             : }
    1971             : 
    1972           0 : void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
    1973             : 
    1974           1 : void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
    1975           2 :   MSSA->print(OS);
    1976           1 : }
    1977             : 
    1978       34231 : MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
    1979             : 
    1980       34231 : MemorySSA::CachingWalker::CachingWalker(MemorySSA *M, AliasAnalysis *A,
    1981       34231 :                                         DominatorTree *D)
    1982       68462 :     : MemorySSAWalker(M), Walker(*M, *A, *D) {}
    1983             : 
    1984         358 : void MemorySSA::CachingWalker::invalidateInfo(MemoryAccess *MA) {
    1985         343 :   if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
    1986             :     MUD->resetOptimized();
    1987         358 : }
    1988             : 
    1989             : /// \brief Walk the use-def chains starting at \p MA and find
    1990             : /// the MemoryAccess that actually clobbers Loc.
    1991             : ///
    1992             : /// \returns our clobbering memory access
    1993             : MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
    1994             :     MemoryAccess *StartingAccess, UpwardsMemoryQuery &Q) {
    1995       92366 :   MemoryAccess *New = Walker.findClobber(StartingAccess, Q);
    1996             : #ifdef EXPENSIVE_CHECKS
    1997             :   MemoryAccess *NewNoCache = Walker.findClobber(StartingAccess, Q);
    1998             :   assert(NewNoCache == New && "Cache made us hand back a different result?");
    1999             :   (void)NewNoCache;
    2000             : #endif
    2001             :   if (AutoResetWalker)
    2002             :     resetClobberWalker();
    2003             :   return New;
    2004             : }
    2005             : 
    2006           2 : MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
    2007             :     MemoryAccess *StartingAccess, const MemoryLocation &Loc) {
    2008           4 :   if (isa<MemoryPhi>(StartingAccess))
    2009             :     return StartingAccess;
    2010             : 
    2011           4 :   auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
    2012           4 :   if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
    2013             :     return StartingUseOrDef;
    2014             : 
    2015           2 :   Instruction *I = StartingUseOrDef->getMemoryInst();
    2016             : 
    2017             :   // Conservatively, fences are always clobbers, so don't perform the walk if we
    2018             :   // hit a fence.
    2019           4 :   if (!ImmutableCallSite(I) && I->isFenceLike())
    2020             :     return StartingUseOrDef;
    2021             : 
    2022           2 :   UpwardsMemoryQuery Q;
    2023           2 :   Q.OriginalAccess = StartingUseOrDef;
    2024           2 :   Q.StartingLoc = Loc;
    2025           2 :   Q.Inst = I;
    2026             :   Q.IsCall = false;
    2027             : 
    2028             :   // Unlike the other function, do not walk to the def of a def, because we are
    2029             :   // handed something we already believe is the clobbering access.
    2030           2 :   MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
    2031           2 :                                      ? StartingUseOrDef->getDefiningAccess()
    2032           2 :                                      : StartingUseOrDef;
    2033             : 
    2034           2 :   MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q);
    2035             :   DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
    2036             :   DEBUG(dbgs() << *StartingUseOrDef << "\n");
    2037             :   DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
    2038             :   DEBUG(dbgs() << *Clobber << "\n");
    2039           2 :   return Clobber;
    2040             : }
    2041             : 
    2042             : MemoryAccess *
    2043      147169 : MemorySSA::CachingWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
    2044      147169 :   auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
    2045             :   // If this is a MemoryPhi, we can't do anything.
    2046             :   if (!StartingAccess)
    2047             :     return MA;
    2048             : 
    2049             :   // If this is an already optimized use or def, return the optimized result.
    2050             :   // Note: Currently, we do not store the optimized def result because we'd need
    2051             :   // a separate field, since we can't use it as the defining access.
    2052      147169 :   if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
    2053      147169 :     if (MUD->isOptimized())
    2054             :       return MUD->getOptimized();
    2055             : 
    2056       92371 :   const Instruction *I = StartingAccess->getMemoryInst();
    2057       92371 :   UpwardsMemoryQuery Q(I, StartingAccess);
    2058             :   // We can't sanely do anything with a fences, they conservatively
    2059             :   // clobber all memory, and have no locations to get pointers from to
    2060             :   // try to disambiguate.
    2061       92371 :   if (!Q.IsCall && I->isFenceLike())
    2062             :     return StartingAccess;
    2063             : 
    2064       92371 :   if (isUseTriviallyOptimizableToLiveOnEntry(*MSSA->AA, I)) {
    2065           2 :     MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
    2066           1 :     if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
    2067           1 :       MUD->setOptimized(LiveOnEntry);
    2068           1 :     return LiveOnEntry;
    2069             :   }
    2070             : 
    2071             :   // Start with the thing we already think clobbers this location
    2072       92370 :   MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
    2073             : 
    2074             :   // At this point, DefiningAccess may be the live on entry def.
    2075             :   // If it is, we will not get a better result.
    2076      184740 :   if (MSSA->isLiveOnEntryDef(DefiningAccess))
    2077             :     return DefiningAccess;
    2078             : 
    2079       92364 :   MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q);
    2080             :   DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
    2081             :   DEBUG(dbgs() << *DefiningAccess << "\n");
    2082             :   DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
    2083             :   DEBUG(dbgs() << *Result << "\n");
    2084       92364 :   if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
    2085       92364 :     MUD->setOptimized(Result);
    2086             : 
    2087       92364 :   return Result;
    2088             : }
    2089             : 
    2090             : MemoryAccess *
    2091           0 : DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
    2092           0 :   if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
    2093             :     return Use->getDefiningAccess();
    2094             :   return MA;
    2095             : }
    2096             : 
    2097           0 : MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
    2098             :     MemoryAccess *StartingAccess, const MemoryLocation &) {
    2099           0 :   if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
    2100             :     return Use->getDefiningAccess();
    2101             :   return StartingAccess;
    2102             : }
    2103             : 
    2104           0 : void MemoryPhi::deleteMe(DerivedUser *Self) {
    2105           0 :   delete static_cast<MemoryPhi *>(Self);
    2106           0 : }
    2107             : 
    2108           0 : void MemoryDef::deleteMe(DerivedUser *Self) {
    2109           0 :   delete static_cast<MemoryDef *>(Self);
    2110           0 : }
    2111             : 
    2112           0 : void MemoryUse::deleteMe(DerivedUser *Self) {
    2113           0 :   delete static_cast<MemoryUse *>(Self);
    2114      216918 : }

Generated by: LCOV version 1.13