LCOV - code coverage report
Current view: top level - lib/Analysis - MemorySSA.cpp (source / functions) Hit Total Coverage
Test: llvm-toolchain.info Lines: 543 583 93.1 %
Date: 2018-02-23 15:42:53 Functions: 81 95 85.3 %
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       29022 : INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
      62             :                       true)
      63       29022 : INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
      64       29022 : INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
      65      327062 : INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
      66             :                     true)
      67             : 
      68        9717 : INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa",
      69             :                       "Memory SSA Printer", false, false)
      70        9717 : INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
      71       38908 : INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",
      72             :                     "Memory SSA Printer", false, false)
      73             : 
      74       81686 : static cl::opt<unsigned> MaxCheckLimit(
      75      163372 :     "memssa-check-limit", cl::Hidden, cl::init(100),
      76       81686 :     cl::desc("The maximum number of stores/phis MemorySSA"
      77       81686 :              "will consider trying to walk past (default = 100)"));
      78             : 
      79             : static cl::opt<bool>
      80      245058 :     VerifyMemorySSA("verify-memoryssa", cl::init(false), cl::Hidden,
      81      245058 :                     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      441247 :       : MemoryLocOrCall(MUD->getMemoryInst()) {}
     124             :   MemoryLocOrCall(const MemoryUseOrDef *MUD)
     125           1 :       : MemoryLocOrCall(MUD->getMemoryInst()) {}
     126             : 
     127      441248 :   MemoryLocOrCall(Instruction *Inst) {
     128      441248 :     if (ImmutableCallSite(Inst)) {
     129        2197 :       IsCall = true;
     130        2197 :       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      439051 :       if (!isa<FenceInst>(Inst))
     136      439051 :         Loc = MemoryLocation::get(Inst);
     137             :     }
     138      441248 :   }
     139             : 
     140      901622 :   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     3172751 :     return Loc;
     150             :   }
     151             : 
     152     2524849 :   bool operator==(const MemoryLocOrCall &Other) const {
     153     2524849 :     if (IsCall != Other.IsCall)
     154             :       return false;
     155             : 
     156     2520765 :     if (IsCall)
     157         887 :       return CS.getCalledValue() == Other.CS.getCalledValue();
     158             :     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             :     return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
     175             :   }
     176             : 
     177             :   static inline MemoryLocOrCall getTombstoneKey() {
     178             :     return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
     179             :   }
     180             : 
     181      582559 :   static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
     182      582559 :     if (MLOC.IsCall)
     183        4844 :       return hash_combine(MLOC.IsCall,
     184        4844 :                           DenseMapInfo<const Value *>::getHashValue(
     185        4844 :                               MLOC.getCS().getCalledValue()));
     186     1160274 :     return hash_combine(
     187     1740411 :         MLOC.IsCall, DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc()));
     188             :   }
     189             : 
     190             :   static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
     191     2524849 :     return LHS == RHS;
     192             :   }
     193             : };
     194             : 
     195             : } // end namespace llvm
     196             : 
     197             : /// This does one-way checks to see if Use could theoretically be hoisted above
     198             : /// MayClobber. This will not check the other way around.
     199             : ///
     200             : /// This assumes that, for the purposes of MemorySSA, Use comes directly after
     201             : /// MayClobber, with no potentially clobbering operations in between them.
     202             : /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
     203             : static bool areLoadsReorderable(const LoadInst *Use,
     204             :                                 const LoadInst *MayClobber) {
     205             :   bool VolatileUse = Use->isVolatile();
     206             :   bool VolatileClobber = MayClobber->isVolatile();
     207             :   // Volatile operations may never be reordered with other volatile operations.
     208        1590 :   if (VolatileUse && VolatileClobber)
     209             :     return false;
     210             :   // Otherwise, volatile doesn't matter here. From the language reference:
     211             :   // 'optimizers may change the order of volatile operations relative to
     212             :   // non-volatile operations.'"
     213             : 
     214             :   // If a load is seq_cst, it cannot be moved above other loads. If its ordering
     215             :   // is weaker, it can be moved above other loads. We just need to be sure that
     216             :   // MayClobber isn't an acquire load, because loads can't be moved above
     217             :   // acquire loads.
     218             :   //
     219             :   // Note that this explicitly *does* allow the free reordering of monotonic (or
     220             :   // weaker) loads of the same address.
     221             :   bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
     222             :   bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
     223             :                                                      AtomicOrdering::Acquire);
     224        1590 :   return !(SeqCstUse || MayClobberIsAcquire);
     225             : }
     226             : 
     227     2123273 : static bool instructionClobbersQuery(MemoryDef *MD,
     228             :                                      const MemoryLocation &UseLoc,
     229             :                                      const Instruction *UseInst,
     230             :                                      AliasAnalysis &AA) {
     231     2123273 :   Instruction *DefInst = MD->getMemoryInst();
     232             :   assert(DefInst && "Defining instruction not actually an instruction");
     233             :   ImmutableCallSite UseCS(UseInst);
     234             : 
     235             :   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
     236             :     // These intrinsics will show up as affecting memory, but they are just
     237             :     // markers.
     238      119098 :     switch (II->getIntrinsicID()) {
     239             :     case Intrinsic::lifetime_start:
     240       52139 :       if (UseCS)
     241             :         return false;
     242      104148 :       return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), UseLoc);
     243             :     case Intrinsic::lifetime_end:
     244             :     case Intrinsic::invariant_start:
     245             :     case Intrinsic::invariant_end:
     246             :     case Intrinsic::assume:
     247             :       return false;
     248             :     default:
     249             :       break;
     250             :     }
     251             :   }
     252             : 
     253     2006686 :   if (UseCS) {
     254       10211 :     ModRefInfo I = AA.getModRefInfo(DefInst, UseCS);
     255       10211 :     return isModOrRefSet(I);
     256             :   }
     257             : 
     258             :   if (auto *DefLoad = dyn_cast<LoadInst>(DefInst))
     259             :     if (auto *UseLoad = dyn_cast<LoadInst>(UseInst))
     260        1590 :       return !areLoadsReorderable(UseLoad, DefLoad);
     261             : 
     262     3989770 :   return isModSet(AA.getModRefInfo(DefInst, UseLoc));
     263             : }
     264             : 
     265     1305515 : static bool instructionClobbersQuery(MemoryDef *MD, const MemoryUseOrDef *MU,
     266             :                                      const MemoryLocOrCall &UseMLOC,
     267             :                                      AliasAnalysis &AA) {
     268             :   // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
     269             :   // to exist while MemoryLocOrCall is pushed through places.
     270     1305515 :   if (UseMLOC.IsCall)
     271       18418 :     return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(),
     272        9209 :                                     AA);
     273     2592612 :   return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
     274     1296306 :                                   AA);
     275             : }
     276             : 
     277             : // Return true when MD may alias MU, return false otherwise.
     278           1 : bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
     279             :                                         AliasAnalysis &AA) {
     280           1 :   return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA);
     281             : }
     282             : 
     283             : namespace {
     284             : 
     285             : struct UpwardsMemoryQuery {
     286             :   // True if our original query started off as a call
     287             :   bool IsCall = false;
     288             :   // The pointer location we started the query with. This will be empty if
     289             :   // IsCall is true.
     290             :   MemoryLocation StartingLoc;
     291             :   // This is the instruction we were querying about.
     292             :   const Instruction *Inst = nullptr;
     293             :   // The MemoryAccess we actually got called with, used to test local domination
     294             :   const MemoryAccess *OriginalAccess = nullptr;
     295             : 
     296           2 :   UpwardsMemoryQuery() = default;
     297             : 
     298       96013 :   UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
     299      288039 :       : IsCall(ImmutableCallSite(Inst)), Inst(Inst), OriginalAccess(Access) {
     300       96013 :     if (!IsCall)
     301       95873 :       StartingLoc = MemoryLocation::get(Inst);
     302       96013 :   }
     303             : };
     304             : 
     305             : } // end anonymous namespace
     306             : 
     307     1296308 : static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc,
     308             :                            AliasAnalysis &AA) {
     309     1296308 :   Instruction *Inst = MD->getMemoryInst();
     310             :   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
     311       56503 :     switch (II->getIntrinsicID()) {
     312             :     case Intrinsic::lifetime_end:
     313       41186 :       return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), Loc);
     314             :     default:
     315             :       return false;
     316             :     }
     317             :   }
     318             :   return false;
     319             : }
     320             : 
     321      538773 : static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysis &AA,
     322             :                                                    const Instruction *I) {
     323             :   // If the memory can't be changed, then loads of the memory can't be
     324             :   // clobbered.
     325             :   //
     326             :   // FIXME: We should handle invariant groups, as well. It's a bit harder,
     327             :   // because we need to pay close attention to invariant group barriers.
     328     1598765 :   return isa<LoadInst>(I) && (I->getMetadata(LLVMContext::MD_invariant_load) ||
     329             :                               AA.pointsToConstantMemory(cast<LoadInst>(I)->
     330      538773 :                                                           getPointerOperand()));
     331             : }
     332             : 
     333             : /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
     334             : /// inbetween `Start` and `ClobberAt` can clobbers `Start`.
     335             : ///
     336             : /// This is meant to be as simple and self-contained as possible. Because it
     337             : /// uses no cache, etc., it can be relatively expensive.
     338             : ///
     339             : /// \param Start     The MemoryAccess that we want to walk from.
     340             : /// \param ClobberAt A clobber for Start.
     341             : /// \param StartLoc  The MemoryLocation for Start.
     342             : /// \param MSSA      The MemorySSA isntance that Start and ClobberAt belong to.
     343             : /// \param Query     The UpwardsMemoryQuery we used for our search.
     344             : /// \param AA        The AliasAnalysis we used for our search.
     345             : static void LLVM_ATTRIBUTE_UNUSED
     346             : checkClobberSanity(MemoryAccess *Start, MemoryAccess *ClobberAt,
     347             :                    const MemoryLocation &StartLoc, const MemorySSA &MSSA,
     348             :                    const UpwardsMemoryQuery &Query, AliasAnalysis &AA) {
     349             :   assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?");
     350             : 
     351             :   if (MSSA.isLiveOnEntryDef(Start)) {
     352             :     assert(MSSA.isLiveOnEntryDef(ClobberAt) &&
     353             :            "liveOnEntry must clobber itself");
     354             :     return;
     355             :   }
     356             : 
     357             :   bool FoundClobber = false;
     358             :   DenseSet<MemoryAccessPair> VisitedPhis;
     359             :   SmallVector<MemoryAccessPair, 8> Worklist;
     360             :   Worklist.emplace_back(Start, StartLoc);
     361             :   // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
     362             :   // is found, complain.
     363             :   while (!Worklist.empty()) {
     364             :     MemoryAccessPair MAP = Worklist.pop_back_val();
     365             :     // All we care about is that nothing from Start to ClobberAt clobbers Start.
     366             :     // We learn nothing from revisiting nodes.
     367             :     if (!VisitedPhis.insert(MAP).second)
     368             :       continue;
     369             : 
     370             :     for (MemoryAccess *MA : def_chain(MAP.first)) {
     371             :       if (MA == ClobberAt) {
     372             :         if (auto *MD = dyn_cast<MemoryDef>(MA)) {
     373             :           // instructionClobbersQuery isn't essentially free, so don't use `|=`,
     374             :           // since it won't let us short-circuit.
     375             :           //
     376             :           // Also, note that this can't be hoisted out of the `Worklist` loop,
     377             :           // since MD may only act as a clobber for 1 of N MemoryLocations.
     378             :           FoundClobber =
     379             :               FoundClobber || MSSA.isLiveOnEntryDef(MD) ||
     380             :               instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
     381             :         }
     382             :         break;
     383             :       }
     384             : 
     385             :       // We should never hit liveOnEntry, unless it's the clobber.
     386             :       assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?");
     387             : 
     388             :       if (auto *MD = dyn_cast<MemoryDef>(MA)) {
     389             :         (void)MD;
     390             :         assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) &&
     391             :                "Found clobber before reaching ClobberAt!");
     392             :         continue;
     393             :       }
     394             : 
     395             :       assert(isa<MemoryPhi>(MA));
     396             :       Worklist.append(upward_defs_begin({MA, MAP.second}), upward_defs_end());
     397             :     }
     398             :   }
     399             : 
     400             :   // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
     401             :   // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
     402             :   assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
     403             :          "ClobberAt never acted as a clobber");
     404             : }
     405             : 
     406             : namespace {
     407             : 
     408             : /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
     409             : /// in one class.
     410       68686 : class ClobberWalker {
     411             :   /// Save a few bytes by using unsigned instead of size_t.
     412             :   using ListIndex = unsigned;
     413             : 
     414             :   /// Represents a span of contiguous MemoryDefs, potentially ending in a
     415             :   /// MemoryPhi.
     416        3104 :   struct DefPath {
     417             :     MemoryLocation Loc;
     418             :     // Note that, because we always walk in reverse, Last will always dominate
     419             :     // First. Also note that First and Last are inclusive.
     420             :     MemoryAccess *First;
     421             :     MemoryAccess *Last;
     422             :     Optional<ListIndex> Previous;
     423             : 
     424             :     DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
     425             :             Optional<ListIndex> Previous)
     426      414797 :         : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
     427             : 
     428             :     DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
     429             :             Optional<ListIndex> Previous)
     430             :         : DefPath(Loc, Init, Init, Previous) {}
     431             :   };
     432             : 
     433             :   const MemorySSA &MSSA;
     434             :   AliasAnalysis &AA;
     435             :   DominatorTree &DT;
     436             :   UpwardsMemoryQuery *Query;
     437             : 
     438             :   // Phi optimization bookkeeping
     439             :   SmallVector<DefPath, 32> Paths;
     440             :   DenseSet<ConstMemoryAccessPair> VisitedPhis;
     441             : 
     442             :   /// Find the nearest def or phi that `From` can legally be optimized to.
     443       83989 :   const MemoryAccess *getWalkTarget(const MemoryPhi *From) const {
     444             :     assert(From->getNumOperands() && "Phi with no operands?");
     445             : 
     446       83989 :     BasicBlock *BB = From->getBlock();
     447       83989 :     MemoryAccess *Result = MSSA.getLiveOnEntryDef();
     448       83989 :     DomTreeNode *Node = DT.getNode(BB);
     449       87681 :     while ((Node = Node->getIDom())) {
     450       86364 :       auto *Defs = MSSA.getBlockDefs(Node->getBlock());
     451       82672 :       if (Defs)
     452             :         return &*Defs->rbegin();
     453             :     }
     454             :     return Result;
     455             :   }
     456             : 
     457             :   /// Result of calling walkToPhiOrClobber.
     458             :   struct UpwardsWalkResult {
     459             :     /// The "Result" of the walk. Either a clobber, the last thing we walked, or
     460             :     /// both.
     461             :     MemoryAccess *Result;
     462             :     bool IsKnownClobber;
     463             :   };
     464             : 
     465             :   /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
     466             :   /// This will update Desc.Last as it walks. It will (optionally) also stop at
     467             :   /// StopAt.
     468             :   ///
     469             :   /// This does not test for whether StopAt is a clobber
     470             :   UpwardsWalkResult
     471      251815 :   walkToPhiOrClobber(DefPath &Desc,
     472             :                      const MemoryAccess *StopAt = nullptr) const {
     473             :     assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world");
     474             : 
     475      251815 :     for (MemoryAccess *Current : def_chain(Desc.Last)) {
     476      971698 :       Desc.Last = Current;
     477      971698 :       if (Current == StopAt)
     478       34436 :         return {Current, false};
     479             : 
     480             :       if (auto *MD = dyn_cast<MemoryDef>(Current))
     481     2464242 :         if (MSSA.isLiveOnEntryDef(MD) ||
     482      817758 :             instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA))
     483      103359 :           return {MD, true};
     484             :     }
     485             : 
     486             :     assert(isa<MemoryPhi>(Desc.Last) &&
     487             :            "Ended at a non-clobber that's not a phi?");
     488      114020 :     return {Desc.Last, false};
     489             :   }
     490             : 
     491      114020 :   void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
     492             :                    ListIndex PriorNode) {
     493      114020 :     auto UpwardDefs = make_range(upward_defs_begin({Phi, Paths[PriorNode].Loc}),
     494      114020 :                                  upward_defs_end());
     495      243115 :     for (const MemoryAccessPair &P : UpwardDefs) {
     496      486230 :       PausedSearches.push_back(Paths.size());
     497      243115 :       Paths.emplace_back(P.second, P.first, PriorNode);
     498             :     }
     499      114020 :   }
     500             : 
     501             :   /// Represents a search that terminated after finding a clobber. This clobber
     502             :   /// may or may not be present in the path of defs from LastNode..SearchStart,
     503             :   /// since it may have been retrieved from cache.
     504             :   struct TerminatedPath {
     505             :     MemoryAccess *Clobber;
     506             :     ListIndex LastNode;
     507             :   };
     508             : 
     509             :   /// Get an access that keeps us from optimizing to the given phi.
     510             :   ///
     511             :   /// PausedSearches is an array of indices into the Paths array. Its incoming
     512             :   /// value is the indices of searches that stopped at the last phi optimization
     513             :   /// target. It's left in an unspecified state.
     514             :   ///
     515             :   /// If this returns None, NewPaused is a vector of searches that terminated
     516             :   /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
     517             :   Optional<TerminatedPath>
     518       83989 :   getBlockingAccess(const MemoryAccess *StopWhere,
     519             :                     SmallVectorImpl<ListIndex> &PausedSearches,
     520             :                     SmallVectorImpl<ListIndex> &NewPaused,
     521             :                     SmallVectorImpl<TerminatedPath> &Terminated) {
     522             :     assert(!PausedSearches.empty() && "No searches to continue?");
     523             : 
     524             :     // BFS vs DFS really doesn't make a difference here, so just do a DFS with
     525             :     // PausedSearches as our stack.
     526      163429 :     while (!PausedSearches.empty()) {
     527      147124 :       ListIndex PathIndex = PausedSearches.pop_back_val();
     528      147124 :       DefPath &Node = Paths[PathIndex];
     529             : 
     530             :       // If we've already visited this path with this MemoryLocation, we don't
     531             :       // need to do so again.
     532             :       //
     533             :       // NOTE: That we just drop these paths on the ground makes caching
     534             :       // behavior sporadic. e.g. given a diamond:
     535             :       //  A
     536             :       // B C
     537             :       //  D
     538             :       //
     539             :       // ...If we walk D, B, A, C, we'll only cache the result of phi
     540             :       // optimization for A, B, and D; C will be skipped because it dies here.
     541             :       // This arguably isn't the worst thing ever, since:
     542             :       //   - We generally query things in a top-down order, so if we got below D
     543             :       //     without needing cache entries for {C, MemLoc}, then chances are
     544             :       //     that those cache entries would end up ultimately unused.
     545             :       //   - We still cache things for A, so C only needs to walk up a bit.
     546             :       // If this behavior becomes problematic, we can fix without a ton of extra
     547             :       // work.
     548      147124 :       if (!VisitedPhis.insert({Node.Last, Node.Loc}).second)
     549       82182 :         continue;
     550             : 
     551      123251 :       UpwardsWalkResult Res = walkToPhiOrClobber(Node, /*StopAt=*/StopWhere);
     552      123251 :       if (Res.IsKnownClobber) {
     553             :         assert(Res.Result != StopWhere);
     554             :         // If this wasn't a cache hit, we hit a clobber when walking. That's a
     555             :         // failure.
     556       67684 :         TerminatedPath Term{Res.Result, PathIndex};
     557       67684 :         if (!MSSA.dominates(Res.Result, StopWhere))
     558       67684 :           return Term;
     559             : 
     560             :         // Otherwise, it's a valid thing to potentially optimize to.
     561           0 :         Terminated.push_back(Term);
     562           0 :         continue;
     563             :       }
     564             : 
     565       90003 :       if (Res.Result == StopWhere) {
     566             :         // We've hit our target. Save this path off for if we want to continue
     567             :         // walking.
     568       34436 :         NewPaused.push_back(PathIndex);
     569       34436 :         continue;
     570             :       }
     571             : 
     572             :       assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber");
     573       21131 :       addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
     574             :     }
     575             : 
     576             :     return None;
     577             :   }
     578             : 
     579             :   template <typename T, typename Walker>
     580      758758 :   struct generic_def_path_iterator
     581             :       : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
     582             :                                     std::forward_iterator_tag, T *> {
     583             :     generic_def_path_iterator() = default;
     584             :     generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
     585             : 
     586             :     T &operator*() const { return curNode(); }
     587             : 
     588             :     generic_def_path_iterator &operator++() {
     589             :       N = curNode().Previous;
     590             :       return *this;
     591             :     }
     592             : 
     593             :     bool operator==(const generic_def_path_iterator &O) const {
     594      149602 :       if (N.hasValue() != O.N.hasValue())
     595             :         return false;
     596           0 :       return !N.hasValue() || *N == *O.N;
     597             :     }
     598             : 
     599             :   private:
     600      299204 :     T &curNode() const { return W->Paths[*N]; }
     601             : 
     602             :     Walker *W = nullptr;
     603             :     Optional<ListIndex> N = None;
     604             :   };
     605             : 
     606             :   using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
     607             :   using const_def_path_iterator =
     608             :       generic_def_path_iterator<const DefPath, const ClobberWalker>;
     609             : 
     610             :   iterator_range<def_path_iterator> def_path(ListIndex From) {
     611             :     return make_range(def_path_iterator(this, From), def_path_iterator());
     612             :   }
     613             : 
     614             :   iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
     615             :     return make_range(const_def_path_iterator(this, From),
     616             :                       const_def_path_iterator());
     617             :   }
     618             : 
     619             :   struct OptznResult {
     620             :     /// The path that contains our result.
     621             :     TerminatedPath PrimaryClobber;
     622             :     /// The paths that we can legally cache back from, but that aren't
     623             :     /// necessarily the result of the Phi optimization.
     624             :     SmallVector<TerminatedPath, 4> OtherClobbers;
     625             :   };
     626             : 
     627             :   ListIndex defPathIndex(const DefPath &N) const {
     628             :     // The assert looks nicer if we don't need to do &N
     629             :     const DefPath *NP = &N;
     630             :     assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&
     631             :            "Out of bounds DefPath!");
     632      217286 :     return NP - &Paths.front();
     633             :   }
     634             : 
     635             :   /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
     636             :   /// that act as legal clobbers. Note that this won't return *all* clobbers.
     637             :   ///
     638             :   /// Phi optimization algorithm tl;dr:
     639             :   ///   - Find the earliest def/phi, A, we can optimize to
     640             :   ///   - Find if all paths from the starting memory access ultimately reach A
     641             :   ///     - If not, optimization isn't possible.
     642             :   ///     - Otherwise, walk from A to another clobber or phi, A'.
     643             :   ///       - If A' is a def, we're done.
     644             :   ///       - If A' is a phi, try to optimize it.
     645             :   ///
     646             :   /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
     647             :   /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
     648       75674 :   OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
     649             :                              const MemoryLocation &Loc) {
     650             :     assert(Paths.empty() && VisitedPhis.empty() &&
     651             :            "Reset the optimization state.");
     652             : 
     653       75674 :     Paths.emplace_back(Loc, Start, Phi, None);
     654             :     // Stores how many "valid" optimization nodes we had prior to calling
     655             :     // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
     656       75674 :     auto PriorPathsSize = Paths.size();
     657             : 
     658             :     SmallVector<ListIndex, 16> PausedSearches;
     659             :     SmallVector<ListIndex, 8> NewPaused;
     660             :     SmallVector<TerminatedPath, 4> TerminatedPaths;
     661             : 
     662       75674 :     addSearches(Phi, PausedSearches, 0);
     663             : 
     664             :     // Moves the TerminatedPath with the "most dominated" Clobber to the end of
     665             :     // Paths.
     666        7990 :     auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
     667             :       assert(!Paths.empty() && "Need a path to move");
     668             :       auto Dom = Paths.begin();
     669       22692 :       for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
     670        7351 :         if (!MSSA.dominates(I->Clobber, Dom->Clobber))
     671             :           Dom = I;
     672        7990 :       auto Last = Paths.end() - 1;
     673        7990 :       if (Last != Dom)
     674             :         std::iter_swap(Last, Dom);
     675       83664 :     };
     676             : 
     677       75674 :     MemoryPhi *Current = Phi;
     678             :     while (true) {
     679             :       assert(!MSSA.isLiveOnEntryDef(Current) &&
     680             :              "liveOnEntry wasn't treated as a clobber?");
     681             : 
     682       83989 :       const auto *Target = getWalkTarget(Current);
     683             :       // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
     684             :       // optimization for the prior phi.
     685             :       assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {
     686             :         return MSSA.dominates(P.Clobber, Target);
     687             :       }));
     688             : 
     689             :       // FIXME: This is broken, because the Blocker may be reported to be
     690             :       // liveOnEntry, and we'll happily wait for that to disappear (read: never)
     691             :       // For the moment, this is fine, since we do nothing with blocker info.
     692       83989 :       if (Optional<TerminatedPath> Blocker = getBlockingAccess(
     693       83989 :               Target, PausedSearches, NewPaused, TerminatedPaths)) {
     694             : 
     695             :         // Find the node we started at. We can't search based on N->Last, since
     696             :         // we may have gone around a loop with a different MemoryLocation.
     697      135368 :         auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
     698      299204 :           return defPathIndex(N) < PriorPathsSize;
     699      135368 :         });
     700             :         assert(Iter != def_path_iterator());
     701             : 
     702             :         DefPath &CurNode = *Iter;
     703             :         assert(CurNode.Last == Current);
     704             : 
     705             :         // Two things:
     706             :         // A. We can't reliably cache all of NewPaused back. Consider a case
     707             :         //    where we have two paths in NewPaused; one of which can't optimize
     708             :         //    above this phi, whereas the other can. If we cache the second path
     709             :         //    back, we'll end up with suboptimal cache entries. We can handle
     710             :         //    cases like this a bit better when we either try to find all
     711             :         //    clobbers that block phi optimization, or when our cache starts
     712             :         //    supporting unfinished searches.
     713             :         // B. We can't reliably cache TerminatedPaths back here without doing
     714             :         //    extra checks; consider a case like:
     715             :         //       T
     716             :         //      / \
     717             :         //     D   C
     718             :         //      \ /
     719             :         //       S
     720             :         //    Where T is our target, C is a node with a clobber on it, D is a
     721             :         //    diamond (with a clobber *only* on the left or right node, N), and
     722             :         //    S is our start. Say we walk to D, through the node opposite N
     723             :         //    (read: ignoring the clobber), and see a cache entry in the top
     724             :         //    node of D. That cache entry gets put into TerminatedPaths. We then
     725             :         //    walk up to C (N is later in our worklist), find the clobber, and
     726             :         //    quit. If we append TerminatedPaths to OtherClobbers, we'll cache
     727             :         //    the bottom part of D to the cached clobber, ignoring the clobber
     728             :         //    in N. Again, this problem goes away if we start tracking all
     729             :         //    blockers for a given phi optimization.
     730       67684 :         TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
     731       67684 :         return {Result, {}};
     732             :       }
     733             : 
     734             :       // If there's nothing left to search, then all paths led to valid clobbers
     735             :       // that we got from our cache; pick the nearest to the start, and allow
     736             :       // the rest to be cached back.
     737       16305 :       if (NewPaused.empty()) {
     738           0 :         MoveDominatedPathToEnd(TerminatedPaths);
     739             :         TerminatedPath Result = TerminatedPaths.pop_back_val();
     740           0 :         return {Result, std::move(TerminatedPaths)};
     741             :       }
     742             : 
     743             :       MemoryAccess *DefChainEnd = nullptr;
     744             :       SmallVector<TerminatedPath, 4> Clobbers;
     745       81417 :       for (ListIndex Paused : NewPaused) {
     746       65112 :         UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
     747       32556 :         if (WR.IsKnownClobber)
     748       15341 :           Clobbers.push_back({WR.Result, Paused});
     749             :         else
     750             :           // Micro-opt: If we hit the end of the chain, save it.
     751             :           DefChainEnd = WR.Result;
     752             :       }
     753             : 
     754       16305 :       if (!TerminatedPaths.empty()) {
     755             :         // If we couldn't find the dominating phi/liveOnEntry in the above loop,
     756             :         // do it now.
     757           0 :         if (!DefChainEnd)
     758           0 :           for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target)))
     759             :             DefChainEnd = MA;
     760             : 
     761             :         // If any of the terminated paths don't dominate the phi we'll try to
     762             :         // optimize, we need to figure out what they are and quit.
     763           0 :         const BasicBlock *ChainBB = DefChainEnd->getBlock();
     764           0 :         for (const TerminatedPath &TP : TerminatedPaths) {
     765             :           // Because we know that DefChainEnd is as "high" as we can go, we
     766             :           // don't need local dominance checks; BB dominance is sufficient.
     767           0 :           if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
     768           0 :             Clobbers.push_back(TP);
     769             :         }
     770             :       }
     771             : 
     772             :       // If we have clobbers in the def chain, find the one closest to Current
     773             :       // and quit.
     774       16305 :       if (!Clobbers.empty()) {
     775        7990 :         MoveDominatedPathToEnd(Clobbers);
     776             :         TerminatedPath Result = Clobbers.pop_back_val();
     777        7990 :         return {Result, std::move(Clobbers)};
     778             :       }
     779             : 
     780             :       assert(all_of(NewPaused,
     781             :                     [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }));
     782             : 
     783             :       // Because liveOnEntry is a clobber, this must be a phi.
     784             :       auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
     785             : 
     786        8315 :       PriorPathsSize = Paths.size();
     787             :       PausedSearches.clear();
     788       42745 :       for (ListIndex I : NewPaused)
     789       17215 :         addSearches(DefChainPhi, PausedSearches, I);
     790             :       NewPaused.clear();
     791             : 
     792             :       Current = DefChainPhi;
     793             :     }
     794             :   }
     795             : 
     796             :   void verifyOptResult(const OptznResult &R) const {
     797             :     assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {
     798             :       return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);
     799             :     }));
     800             :   }
     801             : 
     802             :   void resetPhiOptznState() {
     803             :     Paths.clear();
     804             :     VisitedPhis.clear();
     805             :   }
     806             : 
     807             : public:
     808             :   ClobberWalker(const MemorySSA &MSSA, AliasAnalysis &AA, DominatorTree &DT)
     809       34342 :       : MSSA(MSSA), AA(AA), DT(DT) {}
     810             : 
     811             :   void reset() {}
     812             : 
     813             :   /// Finds the nearest clobber for the given query, optimizing phis if
     814             :   /// possible.
     815       96008 :   MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q) {
     816       96008 :     Query = &Q;
     817             : 
     818             :     MemoryAccess *Current = Start;
     819             :     // This walker pretends uses don't exist. If we're handed one, silently grab
     820             :     // its def. (This has the nice side-effect of ensuring we never cache uses)
     821             :     if (auto *MU = dyn_cast<MemoryUse>(Start))
     822             :       Current = MU->getDefiningAccess();
     823             : 
     824             :     DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
     825             :     // Fast path for the overly-common case (no crazy phi optimization
     826             :     // necessary)
     827       96008 :     UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
     828             :     MemoryAccess *Result;
     829       96008 :     if (WalkResult.IsKnownClobber) {
     830       20334 :       Result = WalkResult.Result;
     831             :     } else {
     832             :       OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
     833       75674 :                                           Current, Q.StartingLoc);
     834             :       verifyOptResult(OptRes);
     835             :       resetPhiOptznState();
     836       75674 :       Result = OptRes.PrimaryClobber.Clobber;
     837             :     }
     838             : 
     839             : #ifdef EXPENSIVE_CHECKS
     840             :     checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
     841             : #endif
     842       96008 :     return Result;
     843             :   }
     844             : 
     845             :   void verify(const MemorySSA *MSSA) { assert(MSSA == &this->MSSA); }
     846             : };
     847             : 
     848             : struct RenamePassData {
     849             :   DomTreeNode *DTN;
     850             :   DomTreeNode::const_iterator ChildIt;
     851             :   MemoryAccess *IncomingVal;
     852             : 
     853             :   RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
     854             :                  MemoryAccess *M)
     855      190766 :       : DTN(D), ChildIt(It), IncomingVal(M) {}
     856             : 
     857             :   void swap(RenamePassData &RHS) {
     858             :     std::swap(DTN, RHS.DTN);
     859             :     std::swap(ChildIt, RHS.ChildIt);
     860             :     std::swap(IncomingVal, RHS.IncomingVal);
     861             :   }
     862             : };
     863             : 
     864             : } // end anonymous namespace
     865             : 
     866             : namespace llvm {
     867             : 
     868             : /// \brief A MemorySSAWalker that does AA walks to disambiguate accesses. It no
     869             : /// longer does caching on its own,
     870             : /// but the name has been retained for the moment.
     871             : class MemorySSA::CachingWalker final : public MemorySSAWalker {
     872             :   ClobberWalker Walker;
     873             :   bool AutoResetWalker = true;
     874             : 
     875             :   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, UpwardsMemoryQuery &);
     876             : 
     877             : public:
     878             :   CachingWalker(MemorySSA *, AliasAnalysis *, DominatorTree *);
     879       34343 :   ~CachingWalker() override = default;
     880             : 
     881             :   using MemorySSAWalker::getClobberingMemoryAccess;
     882             : 
     883             :   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) override;
     884             :   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
     885             :                                           const MemoryLocation &) override;
     886             :   void invalidateInfo(MemoryAccess *) override;
     887             : 
     888             :   /// Whether we call resetClobberWalker() after each time we *actually* walk to
     889             :   /// answer a clobber query.
     890       68686 :   void setAutoResetWalker(bool AutoReset) { AutoResetWalker = AutoReset; }
     891             : 
     892             :   /// Drop the walker's persistent data structures.
     893             :   void resetClobberWalker() { Walker.reset(); }
     894             : 
     895           0 :   void verify(const MemorySSA *MSSA) override {
     896             :     MemorySSAWalker::verify(MSSA);
     897             :     Walker.verify(MSSA);
     898           0 :   }
     899             : };
     900             : 
     901             : } // end namespace llvm
     902             : 
     903      190765 : void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal,
     904             :                                     bool RenameAllUses) {
     905             :   // Pass through values to our successors
     906      190765 :   for (const BasicBlock *S : successors(BB)) {
     907      211239 :     auto It = PerBlockAccesses.find(S);
     908             :     // Rename the phi nodes in our successor block
     909      413486 :     if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
     910      115927 :       continue;
     911             :     AccessList *Accesses = It->second.get();
     912             :     auto *Phi = cast<MemoryPhi>(&Accesses->front());
     913       95312 :     if (RenameAllUses) {
     914           2 :       int PhiIndex = Phi->getBasicBlockIndex(BB);
     915             :       assert(PhiIndex != -1 && "Incomplete phi during partial rename");
     916           2 :       Phi->setIncomingValue(PhiIndex, IncomingVal);
     917             :     } else
     918       95310 :       Phi->addIncoming(IncomingVal, BB);
     919             :   }
     920      190766 : }
     921             : 
     922             : /// \brief Rename a single basic block into MemorySSA form.
     923             : /// Uses the standard SSA renaming algorithm.
     924             : /// \returns The new incoming value.
     925      190765 : MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal,
     926             :                                      bool RenameAllUses) {
     927      190765 :   auto It = PerBlockAccesses.find(BB);
     928             :   // Skip most processing if the list is empty.
     929      190766 :   if (It != PerBlockAccesses.end()) {
     930             :     AccessList *Accesses = It->second.get();
     931     1516944 :     for (MemoryAccess &L : *Accesses) {
     932             :       if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
     933           5 :         if (MUD->getDefiningAccess() == nullptr || RenameAllUses)
     934             :           MUD->setDefiningAccess(IncomingVal);
     935     1297624 :         if (isa<MemoryDef>(&L))
     936             :           IncomingVal = &L;
     937             :       } else {
     938             :         IncomingVal = &L;
     939             :       }
     940             :     }
     941             :   }
     942      190766 :   return IncomingVal;
     943             : }
     944             : 
     945             : /// \brief This is the standard SSA renaming algorithm.
     946             : ///
     947             : /// We walk the dominator tree in preorder, renaming accesses, and then filling
     948             : /// in phi nodes in our successors.
     949       34343 : void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
     950             :                            SmallPtrSetImpl<BasicBlock *> &Visited,
     951             :                            bool SkipVisited, bool RenameAllUses) {
     952             :   SmallVector<RenamePassData, 32> WorkStack;
     953             :   // Skip everything if we already renamed this block and we are skipping.
     954             :   // Note: You can't sink this into the if, because we need it to occur
     955             :   // regardless of whether we skip blocks or not.
     956       34343 :   bool AlreadyVisited = !Visited.insert(Root->getBlock()).second;
     957       34344 :   if (SkipVisited && AlreadyVisited)
     958             :     return;
     959             : 
     960       34344 :   IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses);
     961       34344 :   renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses);
     962       34344 :   WorkStack.push_back({Root, Root->begin(), IncomingVal});
     963             : 
     964      381531 :   while (!WorkStack.empty()) {
     965      347188 :     DomTreeNode *Node = WorkStack.back().DTN;
     966      347188 :     DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
     967      347188 :     IncomingVal = WorkStack.back().IncomingVal;
     968             : 
     969      347188 :     if (ChildIt == Node->end()) {
     970             :       WorkStack.pop_back();
     971             :     } else {
     972      156422 :       DomTreeNode *Child = *ChildIt;
     973             :       ++WorkStack.back().ChildIt;
     974      156422 :       BasicBlock *BB = Child->getBlock();
     975             :       // Note: You can't sink this into the if, because we need it to occur
     976             :       // regardless of whether we skip blocks or not.
     977      156422 :       AlreadyVisited = !Visited.insert(BB).second;
     978      156422 :       if (SkipVisited && AlreadyVisited) {
     979             :         // We already visited this during our renaming, which can happen when
     980             :         // being asked to rename multiple blocks. Figure out the incoming val,
     981             :         // which is the last def.
     982             :         // Incoming value can only change if there is a block def, and in that
     983             :         // case, it's the last block def in the list.
     984           0 :         if (auto *BlockDefs = getWritableBlockDefs(BB))
     985             :           IncomingVal = &*BlockDefs->rbegin();
     986             :       } else
     987      156422 :         IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses);
     988      156422 :       renameSuccessorPhis(BB, IncomingVal, RenameAllUses);
     989      156422 :       WorkStack.push_back({Child, Child->begin(), IncomingVal});
     990             :     }
     991             :   }
     992             : }
     993             : 
     994             : /// \brief This handles unreachable block accesses by deleting phi nodes in
     995             : /// unreachable blocks, and marking all other unreachable MemoryAccess's as
     996             : /// being uses of the live on entry definition.
     997        1960 : void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
     998             :   assert(!DT->isReachableFromEntry(BB) &&
     999             :          "Reachable block found while handling unreachable blocks");
    1000             : 
    1001             :   // Make sure phi nodes in our reachable successors end up with a
    1002             :   // LiveOnEntryDef for our incoming edge, even though our block is forward
    1003             :   // unreachable.  We could just disconnect these blocks from the CFG fully,
    1004             :   // but we do not right now.
    1005        1960 :   for (const BasicBlock *S : successors(BB)) {
    1006        1254 :     if (!DT->isReachableFromEntry(S))
    1007        2079 :       continue;
    1008         409 :     auto It = PerBlockAccesses.find(S);
    1009             :     // Rename the phi nodes in our successor block
    1010         785 :     if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
    1011         389 :       continue;
    1012             :     AccessList *Accesses = It->second.get();
    1013             :     auto *Phi = cast<MemoryPhi>(&Accesses->front());
    1014          20 :     Phi->addIncoming(LiveOnEntryDef.get(), BB);
    1015             :   }
    1016             : 
    1017        1960 :   auto It = PerBlockAccesses.find(BB);
    1018        1960 :   if (It == PerBlockAccesses.end())
    1019         339 :     return;
    1020             : 
    1021             :   auto &Accesses = It->second;
    1022        5619 :   for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
    1023             :     auto Next = std::next(AI);
    1024             :     // If we have a phi, just remove it. We are going to replace all
    1025             :     // users with live on entry.
    1026             :     if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
    1027             :       UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
    1028             :     else
    1029           0 :       Accesses->erase(AI);
    1030             :     AI = Next;
    1031             :   }
    1032             : }
    1033             : 
    1034       34343 : MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
    1035             :     : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
    1036      171715 :       NextID(INVALID_MEMORYACCESS_ID) {
    1037       34343 :   buildMemorySSA();
    1038       34343 : }
    1039             : 
    1040      103029 : MemorySSA::~MemorySSA() {
    1041             :   // Drop all our references
    1042      247006 :   for (const auto &Pair : PerBlockAccesses)
    1043     1502841 :     for (MemoryAccess &MA : *Pair.second)
    1044     1324521 :       MA.dropAllReferences();
    1045       34343 : }
    1046             : 
    1047      218195 : MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
    1048      436390 :   auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
    1049             : 
    1050      218195 :   if (Res.second)
    1051             :     Res.first->second = llvm::make_unique<AccessList>();
    1052      218195 :   return Res.first->second.get();
    1053             : }
    1054             : 
    1055      212978 : MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) {
    1056      425956 :   auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr));
    1057             : 
    1058      212978 :   if (Res.second)
    1059             :     Res.first->second = llvm::make_unique<DefsList>();
    1060      212978 :   return Res.first->second.get();
    1061             : }
    1062             : 
    1063             : namespace llvm {
    1064             : 
    1065             : /// This class is a batch walker of all MemoryUse's in the program, and points
    1066             : /// their defining access at the thing that actually clobbers them.  Because it
    1067             : /// is a batch walker that touches everything, it does not operate like the
    1068             : /// other walkers.  This walker is basically performing a top-down SSA renaming
    1069             : /// pass, where the version stack is used as the cache.  This enables it to be
    1070             : /// significantly more time and memory efficient than using the regular walker,
    1071             : /// which is walking bottom-up.
    1072             : class MemorySSA::OptimizeUses {
    1073             : public:
    1074             :   OptimizeUses(MemorySSA *MSSA, MemorySSAWalker *Walker, AliasAnalysis *AA,
    1075             :                DominatorTree *DT)
    1076       34343 :       : MSSA(MSSA), Walker(Walker), AA(AA), DT(DT) {
    1077       34343 :     Walker = MSSA->getWalker();
    1078             :   }
    1079             : 
    1080             :   void optimizeUses();
    1081             : 
    1082             : private:
    1083             :   /// This represents where a given memorylocation is in the stack.
    1084             :   struct MemlocStackInfo {
    1085             :     // This essentially is keeping track of versions of the stack. Whenever
    1086             :     // the stack changes due to pushes or pops, these versions increase.
    1087             :     unsigned long StackEpoch;
    1088             :     unsigned long PopEpoch;
    1089             :     // This is the lower bound of places on the stack to check. It is equal to
    1090             :     // the place the last stack walk ended.
    1091             :     // Note: Correctness depends on this being initialized to 0, which densemap
    1092             :     // does
    1093             :     unsigned long LowerBound;
    1094             :     const BasicBlock *LowerBoundBlock;
    1095             :     // This is where the last walk for this memory location ended.
    1096             :     unsigned long LastKill;
    1097             :     bool LastKillValid;
    1098             :   };
    1099             : 
    1100             :   void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
    1101             :                            SmallVectorImpl<MemoryAccess *> &,
    1102             :                            DenseMap<MemoryLocOrCall, MemlocStackInfo> &);
    1103             : 
    1104             :   MemorySSA *MSSA;
    1105             :   MemorySSAWalker *Walker;
    1106             :   AliasAnalysis *AA;
    1107             :   DominatorTree *DT;
    1108             : };
    1109             : 
    1110             : } // end namespace llvm
    1111             : 
    1112             : /// Optimize the uses in a given block This is basically the SSA renaming
    1113             : /// algorithm, with one caveat: We are able to use a single stack for all
    1114             : /// MemoryUses.  This is because the set of *possible* reaching MemoryDefs is
    1115             : /// the same for every MemoryUse.  The *actual* clobbering MemoryDef is just
    1116             : /// going to be some position in that stack of possible ones.
    1117             : ///
    1118             : /// We track the stack positions that each MemoryLocation needs
    1119             : /// to check, and last ended at.  This is because we only want to check the
    1120             : /// things that changed since last time.  The same MemoryLocation should
    1121             : /// get clobbered by the same store (getModRefInfo does not use invariantness or
    1122             : /// things like this, and if they start, we can modify MemoryLocOrCall to
    1123             : /// include relevant data)
    1124      190762 : void MemorySSA::OptimizeUses::optimizeUsesInBlock(
    1125             :     const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
    1126             :     SmallVectorImpl<MemoryAccess *> &VersionStack,
    1127             :     DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) {
    1128             : 
    1129             :   /// If no accesses, nothing to do.
    1130      190762 :   MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
    1131      176895 :   if (Accesses == nullptr)
    1132             :     return;
    1133             : 
    1134             :   // Pop everything that doesn't dominate the current block off the stack,
    1135             :   // increment the PopEpoch to account for this.
    1136             :   while (true) {
    1137             :     assert(
    1138             :         !VersionStack.empty() &&
    1139             :         "Version stack should have liveOnEntry sentinel dominating everything");
    1140      307997 :     BasicBlock *BackBlock = VersionStack.back()->getBlock();
    1141      307997 :     if (DT->dominates(BackBlock, BB))
    1142             :       break;
    1143      617428 :     while (VersionStack.back()->getBlock() == BackBlock)
    1144             :       VersionStack.pop_back();
    1145      131102 :     ++PopEpoch;
    1146      131102 :   }
    1147             : 
    1148     1516935 :   for (MemoryAccess &MA : *Accesses) {
    1149             :     auto *MU = dyn_cast<MemoryUse>(&MA);
    1150      897280 :     if (!MU) {
    1151      897280 :       VersionStack.push_back(&MA);
    1152      897280 :       ++StackEpoch;
    1153     1897546 :       continue;
    1154             :     }
    1155             : 
    1156      444273 :     if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
    1157        1513 :       MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true);
    1158        1513 :       continue;
    1159             :     }
    1160             : 
    1161             :     MemoryLocOrCall UseMLOC(MU);
    1162             :     auto &LocInfo = LocStackInfo[UseMLOC];
    1163             :     // If the pop epoch changed, it means we've removed stuff from top of
    1164             :     // stack due to changing blocks. We may have to reset the lower bound or
    1165             :     // last kill info.
    1166      441247 :     if (LocInfo.PopEpoch != PopEpoch) {
    1167      408085 :       LocInfo.PopEpoch = PopEpoch;
    1168      408085 :       LocInfo.StackEpoch = StackEpoch;
    1169             :       // If the lower bound was in something that no longer dominates us, we
    1170             :       // have to reset it.
    1171             :       // We can't simply track stack size, because the stack may have had
    1172             :       // pushes/pops in the meantime.
    1173             :       // XXX: This is non-optimal, but only is slower cases with heavily
    1174             :       // branching dominator trees.  To get the optimal number of queries would
    1175             :       // be to make lowerbound and lastkill a per-loc stack, and pop it until
    1176             :       // the top of that stack dominates us.  This does not seem worth it ATM.
    1177             :       // A much cheaper optimization would be to always explore the deepest
    1178             :       // branch of the dominator tree first. This will guarantee this resets on
    1179             :       // the smallest set of blocks.
    1180      474929 :       if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
    1181       66844 :           !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
    1182             :         // Reset the lower bound of things to check.
    1183             :         // TODO: Some day we should be able to reset to last kill, rather than
    1184             :         // 0.
    1185       32717 :         LocInfo.LowerBound = 0;
    1186       32717 :         LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
    1187       32717 :         LocInfo.LastKillValid = false;
    1188             :       }
    1189       33162 :     } else if (LocInfo.StackEpoch != StackEpoch) {
    1190             :       // If all that has changed is the StackEpoch, we only have to check the
    1191             :       // new things on the stack, because we've checked everything before.  In
    1192             :       // this case, the lower bound of things to check remains the same.
    1193       32896 :       LocInfo.PopEpoch = PopEpoch;
    1194       32896 :       LocInfo.StackEpoch = StackEpoch;
    1195             :     }
    1196      441247 :     if (!LocInfo.LastKillValid) {
    1197      399099 :       LocInfo.LastKill = VersionStack.size() - 1;
    1198      399099 :       LocInfo.LastKillValid = true;
    1199             :     }
    1200             : 
    1201             :     // At this point, we should have corrected last kill and LowerBound to be
    1202             :     // in bounds.
    1203             :     assert(LocInfo.LowerBound < VersionStack.size() &&
    1204             :            "Lower bound out of range");
    1205             :     assert(LocInfo.LastKill < VersionStack.size() &&
    1206             :            "Last kill info out of range");
    1207             :     // In any case, the new upper bound is the top of the stack.
    1208      441247 :     unsigned long UpperBound = VersionStack.size() - 1;
    1209             : 
    1210      983967 :     if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
    1211             :       DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("
    1212             :                    << *(MU->getMemoryInst()) << ")"
    1213             :                    << " because there are " << UpperBound - LocInfo.LowerBound
    1214             :                    << " stores to disambiguate\n");
    1215             :       // Because we did not walk, LastKill is no longer valid, as this may
    1216             :       // have been a kill.
    1217      101473 :       LocInfo.LastKillValid = false;
    1218      101473 :       continue;
    1219             :     }
    1220             :     bool FoundClobberResult = false;
    1221     2618610 :     while (UpperBound > LocInfo.LowerBound) {
    1222     2736750 :       if (isa<MemoryPhi>(VersionStack[UpperBound])) {
    1223             :         // For phis, use the walker, see where we ended up, go there
    1224       62858 :         Instruction *UseInst = MU->getMemoryInst();
    1225       62858 :         MemoryAccess *Result = Walker->getClobberingMemoryAccess(UseInst);
    1226             :         // We are guaranteed to find it or something is wrong
    1227      201264 :         while (VersionStack[UpperBound] != Result) {
    1228             :           assert(UpperBound != 0);
    1229       69203 :           --UpperBound;
    1230             :         }
    1231             :         FoundClobberResult = true;
    1232             :         break;
    1233             :       }
    1234             : 
    1235             :       MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
    1236             :       // If the lifetime of the pointer ends at this instruction, it's live on
    1237             :       // entry.
    1238     2601828 :       if (!UseMLOC.IsCall && lifetimeEndsAt(MD, UseMLOC.getLoc(), *AA)) {
    1239             :         // Reset UpperBound to liveOnEntryDef's place in the stack
    1240             :         UpperBound = 0;
    1241             :         FoundClobberResult = true;
    1242             :         break;
    1243             :       }
    1244     1305514 :       if (instructionClobbersQuery(MD, MU, UseMLOC, *AA)) {
    1245             :         FoundClobberResult = true;
    1246             :         break;
    1247             :       }
    1248     1139418 :       --UpperBound;
    1249             :     }
    1250             :     // At the end of this loop, UpperBound is either a clobber, or lower bound
    1251             :     // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
    1252      339774 :     if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
    1253      621202 :       MU->setDefiningAccess(VersionStack[UpperBound], true);
    1254             :       // We were last killed now by where we got to
    1255      310601 :       LocInfo.LastKill = UpperBound;
    1256             :     } else {
    1257             :       // Otherwise, we checked all the new ones, and now we know we can get to
    1258             :       // LastKill.
    1259       58346 :       MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true);
    1260             :     }
    1261      339774 :     LocInfo.LowerBound = VersionStack.size() - 1;
    1262      339774 :     LocInfo.LowerBoundBlock = BB;
    1263             :   }
    1264             : }
    1265             : 
    1266             : /// Optimize uses to point to their actual clobbering definitions.
    1267       34342 : void MemorySSA::OptimizeUses::optimizeUses() {
    1268             :   SmallVector<MemoryAccess *, 16> VersionStack;
    1269             :   DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo;
    1270       68686 :   VersionStack.push_back(MSSA->getLiveOnEntryDef());
    1271             : 
    1272       34342 :   unsigned long StackEpoch = 1;
    1273       34342 :   unsigned long PopEpoch = 1;
    1274             :   // We perform a non-recursive top-down dominator tree walk.
    1275      484552 :   for (const auto *DomNode : depth_first(DT->getRootNode()))
    1276      190762 :     optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
    1277             :                         LocStackInfo);
    1278       34343 : }
    1279             : 
    1280       34343 : void MemorySSA::placePHINodes(
    1281             :     const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks,
    1282             :     const DenseMap<const BasicBlock *, unsigned int> &BBNumbers) {
    1283             :   // Determine where our MemoryPhi's should go
    1284       34343 :   ForwardIDFCalculator IDFs(*DT);
    1285             :   IDFs.setDefiningBlocks(DefiningBlocks);
    1286             :   SmallVector<BasicBlock *, 32> IDFBlocks;
    1287       34343 :   IDFs.calculate(IDFBlocks);
    1288             : 
    1289             :   std::sort(IDFBlocks.begin(), IDFBlocks.end(),
    1290      416138 :             [&BBNumbers](const BasicBlock *A, const BasicBlock *B) {
    1291             :               return BBNumbers.lookup(A) < BBNumbers.lookup(B);
    1292      208069 :             });
    1293             : 
    1294             :   // Now place MemoryPhi nodes.
    1295      119185 :   for (auto &BB : IDFBlocks)
    1296       42421 :     createMemoryPhi(BB);
    1297       34343 : }
    1298             : 
    1299       34342 : void MemorySSA::buildMemorySSA() {
    1300             :   // We create an access to represent "live on entry", for things like
    1301             :   // arguments or users of globals, where the memory they use is defined before
    1302             :   // the beginning of the function. We do not actually insert it into the IR.
    1303             :   // We do not define a live on exit for the immediate uses, and thus our
    1304             :   // semantics do *not* imply that something with no immediate uses can simply
    1305             :   // be removed.
    1306       34342 :   BasicBlock &StartingPoint = F.getEntryBlock();
    1307       34343 :   LiveOnEntryDef =
    1308       68685 :       llvm::make_unique<MemoryDef>(F.getContext(), nullptr, nullptr,
    1309       68685 :                                    &StartingPoint, NextID++);
    1310             :   DenseMap<const BasicBlock *, unsigned int> BBNumbers;
    1311             :   unsigned NextBBNum = 0;
    1312             : 
    1313             :   // We maintain lists of memory accesses per-block, trading memory for time. We
    1314             :   // could just look up the memory access for every possible instruction in the
    1315             :   // stream.
    1316             :   SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
    1317             :   // Go through each block, figure out where defs occur, and chain together all
    1318             :   // the accesses.
    1319      261408 :   for (BasicBlock &B : F) {
    1320      385444 :     BBNumbers[&B] = NextBBNum++;
    1321             :     bool InsertIntoDef = false;
    1322             :     AccessList *Accesses = nullptr;
    1323             :     DefsList *Defs = nullptr;
    1324     2536595 :     for (Instruction &I : B) {
    1325     2343873 :       MemoryUseOrDef *MUD = createNewAccess(&I);
    1326     2343873 :       if (!MUD)
    1327             :         continue;
    1328             : 
    1329     1301617 :       if (!Accesses)
    1330      175707 :         Accesses = getOrCreateAccessList(&B);
    1331             :       Accesses->push_back(MUD);
    1332     1301617 :       if (isa<MemoryDef>(MUD)) {
    1333             :         InsertIntoDef = true;
    1334      858095 :         if (!Defs)
    1335      170531 :           Defs = getOrCreateDefsList(&B);
    1336             :         Defs->push_back(*MUD);
    1337             :       }
    1338             :     }
    1339      192722 :     if (InsertIntoDef)
    1340      170531 :       DefiningBlocks.insert(&B);
    1341             :   }
    1342       34343 :   placePHINodes(DefiningBlocks, BBNumbers);
    1343             : 
    1344             :   // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
    1345             :   // filled in with all blocks.
    1346             :   SmallPtrSet<BasicBlock *, 16> Visited;
    1347       34343 :   renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
    1348             : 
    1349       34343 :   CachingWalker *Walker = getWalkerImpl();
    1350             : 
    1351             :   // We're doing a batch of updates; don't drop useful caches between them.
    1352             :   Walker->setAutoResetWalker(false);
    1353       68686 :   OptimizeUses(this, Walker, AA, DT).optimizeUses();
    1354             :   Walker->setAutoResetWalker(true);
    1355             :   Walker->resetClobberWalker();
    1356             : 
    1357             :   // Mark the uses in unreachable blocks as live on entry, so that they go
    1358             :   // somewhere.
    1359      261408 :   for (auto &BB : F)
    1360      192722 :     if (!Visited.count(&BB))
    1361        1960 :       markUnreachableAsLiveOnEntry(&BB);
    1362       34343 : }
    1363             : 
    1364      123921 : MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
    1365             : 
    1366      158263 : MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() {
    1367      158263 :   if (Walker)
    1368             :     return Walker.get();
    1369             : 
    1370       34342 :   Walker = llvm::make_unique<CachingWalker>(this, AA, DT);
    1371       34343 :   return Walker.get();
    1372             : }
    1373             : 
    1374             : // This is a helper function used by the creation routines. It places NewAccess
    1375             : // into the access and defs lists for a given basic block, at the given
    1376             : // insertion point.
    1377       42488 : void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess,
    1378             :                                         const BasicBlock *BB,
    1379             :                                         InsertionPlace Point) {
    1380       42488 :   auto *Accesses = getOrCreateAccessList(BB);
    1381       42488 :   if (Point == Beginning) {
    1382             :     // If it's a phi node, it goes first, otherwise, it goes after any phi
    1383             :     // nodes.
    1384       42435 :     if (isa<MemoryPhi>(NewAccess)) {
    1385             :       Accesses->push_front(NewAccess);
    1386       42425 :       auto *Defs = getOrCreateDefsList(BB);
    1387             :       Defs->push_front(*NewAccess);
    1388             :     } else {
    1389             :       auto AI = find_if_not(
    1390             :           *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
    1391             :       Accesses->insert(AI, NewAccess);
    1392          10 :       if (!isa<MemoryUse>(NewAccess)) {
    1393           4 :         auto *Defs = getOrCreateDefsList(BB);
    1394             :         auto DI = find_if_not(
    1395             :             *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
    1396             :         Defs->insert(DI, *NewAccess);
    1397             :       }
    1398             :     }
    1399             :   } else {
    1400             :     Accesses->push_back(NewAccess);
    1401          53 :     if (!isa<MemoryUse>(NewAccess)) {
    1402          12 :       auto *Defs = getOrCreateDefsList(BB);
    1403             :       Defs->push_back(*NewAccess);
    1404             :     }
    1405             :   }
    1406             :   BlockNumberingValid.erase(BB);
    1407       42488 : }
    1408             : 
    1409           6 : void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB,
    1410             :                                       AccessList::iterator InsertPt) {
    1411             :   auto *Accesses = getWritableBlockAccesses(BB);
    1412             :   bool WasEnd = InsertPt == Accesses->end();
    1413             :   Accesses->insert(AccessList::iterator(InsertPt), What);
    1414           6 :   if (!isa<MemoryUse>(What)) {
    1415           6 :     auto *Defs = getOrCreateDefsList(BB);
    1416             :     // If we got asked to insert at the end, we have an easy job, just shove it
    1417             :     // at the end. If we got asked to insert before an existing def, we also get
    1418             :     // an terator. If we got asked to insert before a use, we have to hunt for
    1419             :     // the next def.
    1420           6 :     if (WasEnd) {
    1421             :       Defs->push_back(*What);
    1422           3 :     } else if (isa<MemoryDef>(InsertPt)) {
    1423             :       Defs->insert(InsertPt->getDefsIterator(), *What);
    1424             :     } else {
    1425           3 :       while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt))
    1426             :         ++InsertPt;
    1427             :       // Either we found a def, or we are inserting at the end
    1428           1 :       if (InsertPt == Accesses->end())
    1429             :         Defs->push_back(*What);
    1430             :       else
    1431             :         Defs->insert(InsertPt->getDefsIterator(), *What);
    1432             :     }
    1433             :   }
    1434             :   BlockNumberingValid.erase(BB);
    1435           6 : }
    1436             : 
    1437             : // Move What before Where in the IR.  The end result is taht What will belong to
    1438             : // the right lists and have the right Block set, but will not otherwise be
    1439             : // correct. It will not have the right defining access, and if it is a def,
    1440             : // things below it will not properly be updated.
    1441           4 : void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
    1442             :                        AccessList::iterator Where) {
    1443             :   // Keep it in the lookup tables, remove from the lists
    1444           4 :   removeFromLists(What, false);
    1445             :   What->setBlock(BB);
    1446           4 :   insertIntoListsBefore(What, BB, Where);
    1447           4 : }
    1448             : 
    1449          52 : void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
    1450             :                        InsertionPlace Point) {
    1451          52 :   removeFromLists(What, false);
    1452             :   What->setBlock(BB);
    1453          52 :   insertIntoListsForBlock(What, BB, Point);
    1454          52 : }
    1455             : 
    1456       42425 : MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
    1457             :   assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB");
    1458       84850 :   MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
    1459             :   // Phi's always are placed at the front of the block.
    1460       42425 :   insertIntoListsForBlock(Phi, BB, Beginning);
    1461       84850 :   ValueToMemoryAccess[BB] = Phi;
    1462       42425 :   return Phi;
    1463             : }
    1464             : 
    1465          13 : MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I,
    1466             :                                                MemoryAccess *Definition) {
    1467             :   assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI");
    1468          13 :   MemoryUseOrDef *NewAccess = createNewAccess(I);
    1469             :   assert(
    1470             :       NewAccess != nullptr &&
    1471             :       "Tried to create a memory access for a non-memory touching instruction");
    1472             :   NewAccess->setDefiningAccess(Definition);
    1473          13 :   return NewAccess;
    1474             : }
    1475             : 
    1476             : // Return true if the instruction has ordering constraints.
    1477             : // Note specifically that this only considers stores and loads
    1478             : // because others are still considered ModRef by getModRefInfo.
    1479     1490006 : static inline bool isOrdered(const Instruction *I) {
    1480             :   if (auto *SI = dyn_cast<StoreInst>(I)) {
    1481             :     if (!SI->isUnordered())
    1482             :       return true;
    1483             :   } else if (auto *LI = dyn_cast<LoadInst>(I)) {
    1484             :     if (!LI->isUnordered())
    1485             :       return true;
    1486             :   }
    1487             :   return false;
    1488             : }
    1489             : 
    1490             : /// \brief Helper function to create new memory accesses
    1491     2343886 : MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I) {
    1492             :   // The assume intrinsic has a control dependency which we model by claiming
    1493             :   // that it writes arbitrarily. Ignore that fake memory dependency here.
    1494             :   // FIXME: Replace this special casing with a more accurate modelling of
    1495             :   // assume's control dependency.
    1496             :   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
    1497      132549 :     if (II->getIntrinsicID() == Intrinsic::assume)
    1498             :       return nullptr;
    1499             : 
    1500             :   // Find out what affect this instruction has on memory.
    1501     4687698 :   ModRefInfo ModRef = AA->getModRefInfo(I, None);
    1502             :   // The isOrdered check is used to ensure that volatiles end up as defs
    1503             :   // (atomics end up as ModRef right now anyway).  Until we separate the
    1504             :   // ordering chain from the memory chain, this enables people to see at least
    1505             :   // some relative ordering to volatiles.  Note that getClobberingMemoryAccess
    1506             :   // will still give an answer that bypasses other volatile loads.  TODO:
    1507             :   // Separate memory aliasing and ordering into two different chains so that we
    1508             :   // can precisely represent both "what memory will this read/write/is clobbered
    1509             :   // by" and "what instructions can I move this past".
    1510     2343849 :   bool Def = isModSet(ModRef) || isOrdered(I);
    1511             :   bool Use = isRefSet(ModRef);
    1512             : 
    1513             :   // It's possible for an instruction to not modify memory at all. During
    1514             :   // construction, we ignore them.
    1515     2343849 :   if (!Def && !Use)
    1516             :     return nullptr;
    1517             : 
    1518             :   assert((Def || Use) &&
    1519             :          "Trying to create a memory access with a non-memory instruction");
    1520             : 
    1521             :   MemoryUseOrDef *MUD;
    1522     1301630 :   if (Def)
    1523      858101 :     MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
    1524             :   else
    1525      443529 :     MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
    1526     2603260 :   ValueToMemoryAccess[I] = MUD;
    1527     1301630 :   return MUD;
    1528             : }
    1529             : 
    1530             : /// \brief Returns true if \p Replacer dominates \p Replacee .
    1531           0 : bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
    1532             :                              const MemoryAccess *Replacee) const {
    1533             :   if (isa<MemoryUseOrDef>(Replacee))
    1534           0 :     return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
    1535             :   const auto *MP = cast<MemoryPhi>(Replacee);
    1536             :   // For a phi node, the use occurs in the predecessor block of the phi node.
    1537             :   // Since we may occur multiple times in the phi node, we have to check each
    1538             :   // operand to ensure Replacer dominates each operand where Replacee occurs.
    1539           0 :   for (const Use &Arg : MP->operands()) {
    1540           0 :     if (Arg.get() != Replacee &&
    1541           0 :         !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
    1542             :       return false;
    1543             :   }
    1544             :   return true;
    1545             : }
    1546             : 
    1547             : /// \brief Properly remove \p MA from all of MemorySSA's lookup tables.
    1548       19534 : void MemorySSA::removeFromLookups(MemoryAccess *MA) {
    1549             :   assert(MA->use_empty() &&
    1550             :          "Trying to remove memory access that still has uses");
    1551       19534 :   BlockNumbering.erase(MA);
    1552             :   if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA))
    1553             :     MUD->setDefiningAccess(nullptr);
    1554             :   // Invalidate our walker's cache if necessary
    1555       19534 :   if (!isa<MemoryUse>(MA))
    1556         343 :     Walker->invalidateInfo(MA);
    1557             :   // The call below to erase will destroy MA, so we can't change the order we
    1558             :   // are doing things here
    1559             :   Value *MemoryInst;
    1560             :   if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
    1561       19525 :     MemoryInst = MUD->getMemoryInst();
    1562             :   } else {
    1563           9 :     MemoryInst = MA->getBlock();
    1564             :   }
    1565       19534 :   auto VMA = ValueToMemoryAccess.find(MemoryInst);
    1566       19534 :   if (VMA->second == MA)
    1567             :     ValueToMemoryAccess.erase(VMA);
    1568       19534 : }
    1569             : 
    1570             : /// \brief Properly remove \p MA from all of MemorySSA's lists.
    1571             : ///
    1572             : /// Because of the way the intrusive list and use lists work, it is important to
    1573             : /// do removal in the right order.
    1574             : /// ShouldDelete defaults to true, and will cause the memory access to also be
    1575             : /// deleted, not just removed.
    1576       19590 : void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) {
    1577             :   // The access list owns the reference, so we erase it from the non-owning list
    1578             :   // first.
    1579       19590 :   if (!isa<MemoryUse>(MA)) {
    1580         359 :     auto DefsIt = PerBlockDefs.find(MA->getBlock());
    1581             :     std::unique_ptr<DefsList> &Defs = DefsIt->second;
    1582             :     Defs->remove(*MA);
    1583         359 :     if (Defs->empty())
    1584             :       PerBlockDefs.erase(DefsIt);
    1585             :   }
    1586             : 
    1587             :   // The erase call here will delete it. If we don't want it deleted, we call
    1588             :   // remove instead.
    1589       19590 :   auto AccessIt = PerBlockAccesses.find(MA->getBlock());
    1590             :   std::unique_ptr<AccessList> &Accesses = AccessIt->second;
    1591       19590 :   if (ShouldDelete)
    1592             :     Accesses->erase(MA);
    1593             :   else
    1594             :     Accesses->remove(MA);
    1595             : 
    1596       19590 :   if (Accesses->empty())
    1597             :     PerBlockAccesses.erase(AccessIt);
    1598       19590 : }
    1599             : 
    1600          82 : void MemorySSA::print(raw_ostream &OS) const {
    1601             :   MemorySSAAnnotatedWriter Writer(this);
    1602          82 :   F.print(OS, &Writer);
    1603          82 : }
    1604             : 
    1605             : #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    1606             : LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); }
    1607             : #endif
    1608             : 
    1609          93 : void MemorySSA::verifyMemorySSA() const {
    1610          93 :   verifyDefUses(F);
    1611          93 :   verifyDomination(F);
    1612          93 :   verifyOrdering(F);
    1613             :   Walker->verify(this);
    1614          93 : }
    1615             : 
    1616             : /// \brief Verify that the order and existence of MemoryAccesses matches the
    1617             : /// order and existence of memory affecting instructions.
    1618          93 : void MemorySSA::verifyOrdering(Function &F) const {
    1619             :   // Walk all the blocks, comparing what the lookups think and what the access
    1620             :   // lists think, as well as the order in the blocks vs the order in the access
    1621             :   // lists.
    1622             :   SmallVector<MemoryAccess *, 32> ActualAccesses;
    1623             :   SmallVector<MemoryAccess *, 32> ActualDefs;
    1624         377 :   for (BasicBlock &B : F) {
    1625             :     const AccessList *AL = getBlockAccesses(&B);
    1626             :     const auto *DL = getBlockDefs(&B);
    1627         284 :     MemoryAccess *Phi = getMemoryAccess(&B);
    1628         284 :     if (Phi) {
    1629          81 :       ActualAccesses.push_back(Phi);
    1630          81 :       ActualDefs.push_back(Phi);
    1631             :     }
    1632             : 
    1633        1086 :     for (Instruction &I : B) {
    1634         802 :       MemoryAccess *MA = getMemoryAccess(&I);
    1635             :       assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&
    1636             :              "We have memory affecting instructions "
    1637             :              "in this block but they are not in the "
    1638             :              "access list or defs list");
    1639         802 :       if (MA) {
    1640         402 :         ActualAccesses.push_back(MA);
    1641         804 :         if (isa<MemoryDef>(MA))
    1642         234 :           ActualDefs.push_back(MA);
    1643             :       }
    1644             :     }
    1645             :     // Either we hit the assert, really have no accesses, or we have both
    1646             :     // accesses and an access list.
    1647             :     // Same with defs.
    1648         284 :     if (!AL && !DL)
    1649          72 :       continue;
    1650             :     assert(AL->size() == ActualAccesses.size() &&
    1651             :            "We don't have the same number of accesses in the block as on the "
    1652             :            "access list");
    1653             :     assert((DL || ActualDefs.size() == 0) &&
    1654             :            "Either we should have a defs list, or we should have no defs");
    1655             :     assert((!DL || DL->size() == ActualDefs.size()) &&
    1656             :            "We don't have the same number of defs in the block as on the "
    1657             :            "def list");
    1658             :     auto ALI = AL->begin();
    1659             :     auto AAI = ActualAccesses.begin();
    1660        1661 :     while (ALI != AL->end() && AAI != ActualAccesses.end()) {
    1661             :       assert(&*ALI == *AAI && "Not the same accesses in the same order");
    1662             :       ++ALI;
    1663         483 :       ++AAI;
    1664             :     }
    1665             :     ActualAccesses.clear();
    1666         212 :     if (DL) {
    1667             :       auto DLI = DL->begin();
    1668             :       auto ADI = ActualDefs.begin();
    1669        1150 :       while (DLI != DL->end() && ADI != ActualDefs.end()) {
    1670             :         assert(&*DLI == *ADI && "Not the same defs in the same order");
    1671             :         ++DLI;
    1672         315 :         ++ADI;
    1673             :       }
    1674             :     }
    1675             :     ActualDefs.clear();
    1676             :   }
    1677          93 : }
    1678             : 
    1679             : /// \brief Verify the domination properties of MemorySSA by checking that each
    1680             : /// definition dominates all of its uses.
    1681          93 : void MemorySSA::verifyDomination(Function &F) const {
    1682             : #ifndef NDEBUG
    1683             :   for (BasicBlock &B : F) {
    1684             :     // Phi nodes are attached to basic blocks
    1685             :     if (MemoryPhi *MP = getMemoryAccess(&B))
    1686             :       for (const Use &U : MP->uses())
    1687             :         assert(dominates(MP, U) && "Memory PHI does not dominate it's uses");
    1688             : 
    1689             :     for (Instruction &I : B) {
    1690             :       MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
    1691             :       if (!MD)
    1692             :         continue;
    1693             : 
    1694             :       for (const Use &U : MD->uses())
    1695             :         assert(dominates(MD, U) && "Memory Def does not dominate it's uses");
    1696             :     }
    1697             :   }
    1698             : #endif
    1699          93 : }
    1700             : 
    1701             : /// \brief Verify the def-use lists in MemorySSA, by verifying that \p Use
    1702             : /// appears in the use list of \p Def.
    1703         585 : void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
    1704             : #ifndef NDEBUG
    1705             :   // The live on entry use may cause us to get a NULL def here
    1706             :   if (!Def)
    1707             :     assert(isLiveOnEntryDef(Use) &&
    1708             :            "Null def but use not point to live on entry def");
    1709             :   else
    1710             :     assert(is_contained(Def->users(), Use) &&
    1711             :            "Did not find use in def's use list");
    1712             : #endif
    1713         585 : }
    1714             : 
    1715             : /// \brief Verify the immediate use information, by walking all the memory
    1716             : /// accesses and verifying that, for each use, it appears in the
    1717             : /// appropriate def's use list
    1718          93 : void MemorySSA::verifyDefUses(Function &F) const {
    1719         377 :   for (BasicBlock &B : F) {
    1720             :     // Phi nodes are attached to basic blocks
    1721         284 :     if (MemoryPhi *Phi = getMemoryAccess(&B)) {
    1722             :       assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(
    1723             :                                           pred_begin(&B), pred_end(&B))) &&
    1724             :              "Incomplete MemoryPhi Node");
    1725         447 :       for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
    1726         183 :         verifyUseInDefs(Phi->getIncomingValue(I), Phi);
    1727             :     }
    1728             : 
    1729        1086 :     for (Instruction &I : B) {
    1730         802 :       if (MemoryUseOrDef *MA = getMemoryAccess(&I)) {
    1731         402 :         verifyUseInDefs(MA->getDefiningAccess(), MA);
    1732             :       }
    1733             :     }
    1734             :   }
    1735          93 : }
    1736             : 
    1737      392812 : MemoryUseOrDef *MemorySSA::getMemoryAccess(const Instruction *I) const {
    1738      785624 :   return cast_or_null<MemoryUseOrDef>(ValueToMemoryAccess.lookup(I));
    1739             : }
    1740             : 
    1741        2368 : MemoryPhi *MemorySSA::getMemoryAccess(const BasicBlock *BB) const {
    1742        4736 :   return cast_or_null<MemoryPhi>(ValueToMemoryAccess.lookup(cast<Value>(BB)));
    1743             : }
    1744             : 
    1745             : /// Perform a local numbering on blocks so that instruction ordering can be
    1746             : /// determined in constant time.
    1747             : /// TODO: We currently just number in order.  If we numbered by N, we could
    1748             : /// allow at least N-1 sequences of insertBefore or insertAfter (and at least
    1749             : /// log2(N) sequences of mixed before and after) without needing to invalidate
    1750             : /// the numbering.
    1751        4046 : void MemorySSA::renumberBlock(const BasicBlock *B) const {
    1752             :   // The pre-increment ensures the numbers really start at 1.
    1753             :   unsigned long CurrentNumber = 0;
    1754             :   const AccessList *AL = getBlockAccesses(B);
    1755             :   assert(AL != nullptr && "Asking to renumber an empty block");
    1756       88638 :   for (const auto &I : *AL)
    1757      169184 :     BlockNumbering[&I] = ++CurrentNumber;
    1758        4046 :   BlockNumberingValid.insert(B);
    1759        4046 : }
    1760             : 
    1761             : /// \brief Determine, for two memory accesses in the same block,
    1762             : /// whether \p Dominator dominates \p Dominatee.
    1763             : /// \returns True if \p Dominator dominates \p Dominatee.
    1764        8601 : bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
    1765             :                                  const MemoryAccess *Dominatee) const {
    1766        8601 :   const BasicBlock *DominatorBlock = Dominator->getBlock();
    1767             : 
    1768             :   assert((DominatorBlock == Dominatee->getBlock()) &&
    1769             :          "Asking for local domination when accesses are in different blocks!");
    1770             :   // A node dominates itself.
    1771        8601 :   if (Dominatee == Dominator)
    1772             :     return true;
    1773             : 
    1774             :   // When Dominatee is defined on function entry, it is not dominated by another
    1775             :   // memory access.
    1776        8601 :   if (isLiveOnEntryDef(Dominatee))
    1777             :     return false;
    1778             : 
    1779             :   // When Dominator is defined on function entry, it dominates the other memory
    1780             :   // access.
    1781        8601 :   if (isLiveOnEntryDef(Dominator))
    1782             :     return true;
    1783             : 
    1784        7667 :   if (!BlockNumberingValid.count(DominatorBlock))
    1785        4046 :     renumberBlock(DominatorBlock);
    1786             : 
    1787        7667 :   unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
    1788             :   // All numbers start with 1
    1789             :   assert(DominatorNum != 0 && "Block was not numbered properly");
    1790             :   unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
    1791             :   assert(DominateeNum != 0 && "Block was not numbered properly");
    1792        7667 :   return DominatorNum < DominateeNum;
    1793             : }
    1794             : 
    1795      164280 : bool MemorySSA::dominates(const MemoryAccess *Dominator,
    1796             :                           const MemoryAccess *Dominatee) const {
    1797      164280 :   if (Dominator == Dominatee)
    1798             :     return true;
    1799             : 
    1800      144845 :   if (isLiveOnEntryDef(Dominatee))
    1801             :     return false;
    1802             : 
    1803      143676 :   if (Dominator->getBlock() != Dominatee->getBlock())
    1804      135077 :     return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
    1805        8599 :   return locallyDominates(Dominator, Dominatee);
    1806             : }
    1807             : 
    1808           0 : bool MemorySSA::dominates(const MemoryAccess *Dominator,
    1809             :                           const Use &Dominatee) const {
    1810           0 :   if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
    1811             :     BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
    1812             :     // The def must dominate the incoming block of the phi.
    1813           0 :     if (UseBB != Dominator->getBlock())
    1814           0 :       return DT->dominates(Dominator->getBlock(), UseBB);
    1815             :     // If the UseBB and the DefBB are the same, compare locally.
    1816           0 :     return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
    1817             :   }
    1818             :   // If it's not a PHI node use, the normal dominates can already handle it.
    1819           0 :   return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
    1820             : }
    1821             : 
    1822             : const static char LiveOnEntryStr[] = "liveOnEntry";
    1823             : 
    1824         439 : void MemoryAccess::print(raw_ostream &OS) const {
    1825         878 :   switch (getValueID()) {
    1826          69 :   case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS);
    1827         216 :   case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS);
    1828         154 :   case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS);
    1829             :   }
    1830           0 :   llvm_unreachable("invalid value id");
    1831             : }
    1832             : 
    1833         216 : void MemoryDef::print(raw_ostream &OS) const {
    1834             :   MemoryAccess *UO = getDefiningAccess();
    1835             : 
    1836         432 :   OS << getID() << " = MemoryDef(";
    1837         432 :   if (UO && UO->getID())
    1838             :     OS << UO->getID();
    1839             :   else
    1840          72 :     OS << LiveOnEntryStr;
    1841             :   OS << ')';
    1842         216 : }
    1843             : 
    1844          69 : void MemoryPhi::print(raw_ostream &OS) const {
    1845             :   bool First = true;
    1846         138 :   OS << getID() << " = MemoryPhi(";
    1847         456 :   for (const auto &Op : operands()) {
    1848             :     BasicBlock *BB = getIncomingBlock(Op);
    1849             :     MemoryAccess *MA = cast<MemoryAccess>(Op);
    1850         159 :     if (!First)
    1851             :       OS << ',';
    1852             :     else
    1853             :       First = false;
    1854             : 
    1855             :     OS << '{';
    1856         159 :     if (BB->hasName())
    1857         145 :       OS << BB->getName();
    1858             :     else
    1859          14 :       BB->printAsOperand(OS, false);
    1860             :     OS << ',';
    1861         159 :     if (unsigned ID = MA->getID())
    1862             :       OS << ID;
    1863             :     else
    1864          16 :       OS << LiveOnEntryStr;
    1865             :     OS << '}';
    1866             :   }
    1867             :   OS << ')';
    1868          69 : }
    1869             : 
    1870         154 : void MemoryUse::print(raw_ostream &OS) const {
    1871             :   MemoryAccess *UO = getDefiningAccess();
    1872         154 :   OS << "MemoryUse(";
    1873         308 :   if (UO && UO->getID())
    1874             :     OS << UO->getID();
    1875             :   else
    1876          30 :     OS << LiveOnEntryStr;
    1877             :   OS << ')';
    1878         154 : }
    1879             : 
    1880           0 : void MemoryAccess::dump() const {
    1881             : // Cannot completely remove virtual function even in release mode.
    1882             : #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    1883             :   print(dbgs());
    1884             :   dbgs() << "\n";
    1885             : #endif
    1886           0 : }
    1887             : 
    1888             : char MemorySSAPrinterLegacyPass::ID = 0;
    1889             : 
    1890          40 : MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) {
    1891          20 :   initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry());
    1892          20 : }
    1893             : 
    1894          20 : void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
    1895             :   AU.setPreservesAll();
    1896             :   AU.addRequired<MemorySSAWrapperPass>();
    1897          20 : }
    1898             : 
    1899          45 : bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) {
    1900          45 :   auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
    1901          45 :   MSSA.print(dbgs());
    1902          45 :   if (VerifyMemorySSA)
    1903          44 :     MSSA.verifyMemorySSA();
    1904          45 :   return false;
    1905             : }
    1906             : 
    1907             : AnalysisKey MemorySSAAnalysis::Key;
    1908             : 
    1909         174 : MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F,
    1910             :                                                  FunctionAnalysisManager &AM) {
    1911             :   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
    1912             :   auto &AA = AM.getResult<AAManager>(F);
    1913         174 :   return MemorySSAAnalysis::Result(llvm::make_unique<MemorySSA>(F, &AA, &DT));
    1914             : }
    1915             : 
    1916          36 : PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
    1917             :                                             FunctionAnalysisManager &AM) {
    1918          36 :   OS << "MemorySSA for function: " << F.getName() << "\n";
    1919          36 :   AM.getResult<MemorySSAAnalysis>(F).getMSSA().print(OS);
    1920             : 
    1921          36 :   return PreservedAnalyses::all();
    1922             : }
    1923             : 
    1924          33 : PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
    1925             :                                              FunctionAnalysisManager &AM) {
    1926          33 :   AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
    1927             : 
    1928          33 :   return PreservedAnalyses::all();
    1929             : }
    1930             : 
    1931             : char MemorySSAWrapperPass::ID = 0;
    1932             : 
    1933        2296 : MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
    1934        1148 :   initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
    1935        1148 : }
    1936             : 
    1937       68302 : void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
    1938             : 
    1939        1148 : void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
    1940             :   AU.setPreservesAll();
    1941             :   AU.addRequiredTransitive<DominatorTreeWrapperPass>();
    1942             :   AU.addRequiredTransitive<AAResultsWrapperPass>();
    1943        1148 : }
    1944             : 
    1945       34151 : bool MemorySSAWrapperPass::runOnFunction(Function &F) {
    1946       34151 :   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
    1947       34151 :   auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
    1948       34151 :   MSSA.reset(new MemorySSA(F, &AA, &DT));
    1949       34151 :   return false;
    1950             : }
    1951             : 
    1952           0 : void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
    1953             : 
    1954           1 : void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
    1955           1 :   MSSA->print(OS);
    1956           1 : }
    1957             : 
    1958       34343 : MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
    1959             : 
    1960       34342 : MemorySSA::CachingWalker::CachingWalker(MemorySSA *M, AliasAnalysis *A,
    1961       34342 :                                         DominatorTree *D)
    1962       68685 :     : MemorySSAWalker(M), Walker(*M, *A, *D) {}
    1963             : 
    1964         344 : void MemorySSA::CachingWalker::invalidateInfo(MemoryAccess *MA) {
    1965             :   if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
    1966             :     MUD->resetOptimized();
    1967         344 : }
    1968             : 
    1969             : /// \brief Walk the use-def chains starting at \p MA and find
    1970             : /// the MemoryAccess that actually clobbers Loc.
    1971             : ///
    1972             : /// \returns our clobbering memory access
    1973             : MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
    1974             :     MemoryAccess *StartingAccess, UpwardsMemoryQuery &Q) {
    1975       96008 :   MemoryAccess *New = Walker.findClobber(StartingAccess, Q);
    1976             : #ifdef EXPENSIVE_CHECKS
    1977             :   MemoryAccess *NewNoCache = Walker.findClobber(StartingAccess, Q);
    1978             :   assert(NewNoCache == New && "Cache made us hand back a different result?");
    1979             :   (void)NewNoCache;
    1980             : #endif
    1981             :   if (AutoResetWalker)
    1982             :     resetClobberWalker();
    1983             :   return New;
    1984             : }
    1985             : 
    1986           2 : MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
    1987             :     MemoryAccess *StartingAccess, const MemoryLocation &Loc) {
    1988           2 :   if (isa<MemoryPhi>(StartingAccess))
    1989             :     return StartingAccess;
    1990             : 
    1991             :   auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
    1992           4 :   if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
    1993             :     return StartingUseOrDef;
    1994             : 
    1995           2 :   Instruction *I = StartingUseOrDef->getMemoryInst();
    1996             : 
    1997             :   // Conservatively, fences are always clobbers, so don't perform the walk if we
    1998             :   // hit a fence.
    1999           2 :   if (!ImmutableCallSite(I) && I->isFenceLike())
    2000             :     return StartingUseOrDef;
    2001             : 
    2002             :   UpwardsMemoryQuery Q;
    2003           2 :   Q.OriginalAccess = StartingUseOrDef;
    2004           2 :   Q.StartingLoc = Loc;
    2005           2 :   Q.Inst = I;
    2006             :   Q.IsCall = false;
    2007             : 
    2008             :   // Unlike the other function, do not walk to the def of a def, because we are
    2009             :   // handed something we already believe is the clobbering access.
    2010             :   MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
    2011           2 :                                      ? StartingUseOrDef->getDefiningAccess()
    2012             :                                      : StartingUseOrDef;
    2013             : 
    2014             :   MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q);
    2015             :   DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
    2016             :   DEBUG(dbgs() << *StartingUseOrDef << "\n");
    2017             :   DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
    2018             :   DEBUG(dbgs() << *Clobber << "\n");
    2019           2 :   return Clobber;
    2020             : }
    2021             : 
    2022             : MemoryAccess *
    2023      152689 : MemorySSA::CachingWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
    2024             :   auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
    2025             :   // If this is a MemoryPhi, we can't do anything.
    2026             :   if (!StartingAccess)
    2027             :     return MA;
    2028             : 
    2029             :   // If this is an already optimized use or def, return the optimized result.
    2030             :   // Note: Currently, we do not store the optimized def result because we'd need
    2031             :   // a separate field, since we can't use it as the defining access.
    2032             :   if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
    2033      152689 :     if (MUD->isOptimized())
    2034             :       return MUD->getOptimized();
    2035             : 
    2036       96013 :   const Instruction *I = StartingAccess->getMemoryInst();
    2037       96013 :   UpwardsMemoryQuery Q(I, StartingAccess);
    2038             :   // We can't sanely do anything with a fences, they conservatively
    2039             :   // clobber all memory, and have no locations to get pointers from to
    2040             :   // try to disambiguate.
    2041       96013 :   if (!Q.IsCall && I->isFenceLike())
    2042             :     return StartingAccess;
    2043             : 
    2044       96013 :   if (isUseTriviallyOptimizableToLiveOnEntry(*MSSA->AA, I)) {
    2045           1 :     MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
    2046             :     if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
    2047           1 :       MUD->setOptimized(LiveOnEntry);
    2048           1 :     return LiveOnEntry;
    2049             :   }
    2050             : 
    2051             :   // Start with the thing we already think clobbers this location
    2052             :   MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
    2053             : 
    2054             :   // At this point, DefiningAccess may be the live on entry def.
    2055             :   // If it is, we will not get a better result.
    2056      192024 :   if (MSSA->isLiveOnEntryDef(DefiningAccess))
    2057             :     return DefiningAccess;
    2058             : 
    2059             :   MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q);
    2060             :   DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
    2061             :   DEBUG(dbgs() << *DefiningAccess << "\n");
    2062             :   DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
    2063             :   DEBUG(dbgs() << *Result << "\n");
    2064             :   if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
    2065       96006 :     MUD->setOptimized(Result);
    2066             : 
    2067       96006 :   return Result;
    2068             : }
    2069             : 
    2070             : MemoryAccess *
    2071           0 : DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
    2072             :   if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
    2073             :     return Use->getDefiningAccess();
    2074             :   return MA;
    2075             : }
    2076             : 
    2077           0 : MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
    2078             :     MemoryAccess *StartingAccess, const MemoryLocation &) {
    2079             :   if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
    2080             :     return Use->getDefiningAccess();
    2081             :   return StartingAccess;
    2082             : }
    2083             : 
    2084           0 : void MemoryPhi::deleteMe(DerivedUser *Self) {
    2085           0 :   delete static_cast<MemoryPhi *>(Self);
    2086           0 : }
    2087             : 
    2088           0 : void MemoryDef::deleteMe(DerivedUser *Self) {
    2089           0 :   delete static_cast<MemoryDef *>(Self);
    2090           0 : }
    2091             : 
    2092           0 : void MemoryUse::deleteMe(DerivedUser *Self) {
    2093           0 :   delete static_cast<MemoryUse *>(Self);
    2094      245058 : }

Generated by: LCOV version 1.13