swift-mirror/lib/SILOptimizer/Transforms/DeadObjectElimination.cpp

//===--- DeadObjectElimination.cpp - Remove unused objects  ---------------===//
//
// This source file is part of the Swift.org open source project
//
// Copyright (c) 2014 - 2017 Apple Inc. and the Swift project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See https://swift.org/LICENSE.txt for license information
// See https://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
//
// This pass eliminates store only alloc_ref objects that have destructors
// without side effects.
//
// The high level overview of the algorithm is that first it visits the
// destructor and attempts to prove that the destructor is well behaved, i.e. it
// does not have any side effects outside of the destructor itself. If the
// destructor can be proven to be well behaved, it then goes through the use
// list of the alloc_ref and attempts to prove that the alloc_ref does not
// escape or is used in a way that could cause side effects. If both of those
// conditions apply, the alloc_ref and its entire use graph is eliminated.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "dead-object-elim"
#include "swift/Basic/Assertions.h"
#include "swift/Basic/IndexTrie.h"
#include "swift/AST/ResilienceExpansion.h"
#include "swift/SIL/BasicBlockUtils.h"
#include "swift/SIL/DebugUtils.h"
#include "swift/SIL/InstructionUtils.h"
#include "swift/SIL/OwnershipUtils.h"
#include "swift/SIL/Projection.h"
#include "swift/SIL/SILArgument.h"
#include "swift/SIL/SILDeclRef.h"
#include "swift/SIL/SILFunction.h"
#include "swift/SIL/SILInstruction.h"
#include "swift/SIL/SILModule.h"
#include "swift/SIL/SILUndef.h"
#include "swift/SIL/MemAccessUtils.h"
#include "swift/SILOptimizer/Analysis/ArraySemantic.h"
#include "swift/SILOptimizer/PassManager/Passes.h"
#include "swift/SILOptimizer/PassManager/Transforms.h"
#include "swift/SILOptimizer/Utils/InstOptUtils.h"
#include "swift/SILOptimizer/Utils/SILSSAUpdater.h"
#include "swift/SILOptimizer/Utils/ValueLifetime.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/Debug.h"

using namespace swift;

STATISTIC(DeadAllocRefEliminated,
          "number of AllocRef instructions removed");

STATISTIC(DeadAllocStackEliminated,
          "number of AllocStack instructions removed");

STATISTIC(DeadKeyPathEliminated,
          "number of keypath instructions removed");

STATISTIC(DeadAllocApplyEliminated,
          "number of allocating Apply instructions removed");

using UserList = llvm::SmallSetVector<SILInstruction *, 16>;

namespace {

/// Side effects of a destructor.
enum class DestructorEffects {
  None,

  /// The destructor contains a "destroyArray" builtin which destroys the tail
  /// elements of the object - like in Array.
  DestroysTailElems,

  Unknown
};

// Analyzing the body of this class destructor is valid because the object is
// dead. This means that the object is never passed to objc_setAssociatedObject,
// so its destructor cannot be extended at runtime.
static SILFunction *getDestructor(AllocRefInstBase *ARI) {

  // We can't know the destructor for an alloc_ref_dynamic instruction in
  // general.
  auto *dynamicAllocRef = dyn_cast<AllocRefDynamicInst>(ARI);
  if (dynamicAllocRef &&
      !dynamicAllocRef->isDynamicTypeDeinitAndSizeKnownEquivalentToBaseType())
    return nullptr;

  // We only support classes.
  ClassDecl *ClsDecl = ARI->getType().getClassOrBoundGenericClass();
  if (!ClsDecl)
    return nullptr;

  // Look up the destructor of ClsDecl.
  DestructorDecl *Destructor = ClsDecl->getDestructor();
  assert(Destructor && "getDestructor() should never return a nullptr.");

  // Find the destructor name via SILDeclRef.
  // FIXME: When destructors get moved into vtables, update this to use the
  // vtable for the class.
  SILDeclRef Ref(Destructor);
  SILFunction *Fn = ARI->getModule().lookUpFunction(Ref);
  if (!Fn || Fn->empty()) {
    LLVM_DEBUG(llvm::dbgs() << "    Could not find destructor.\n");
    return nullptr;
  }

  LLVM_DEBUG(llvm::dbgs() << "    Found destructor!\n");

  // If the destructor has an objc_method calling convention, we cannot
  // analyze it since it could be swapped out from under us at runtime.
  if (Fn->getRepresentation() == SILFunctionTypeRepresentation::ObjCMethod) {
    LLVM_DEBUG(llvm::dbgs() << "        Found Objective-C destructor. Can't "
               "analyze!\n");
    return nullptr;
  }

  return Fn;
}

static bool isDestroyArray(SILInstruction *inst) {
  BuiltinInst *bi = dyn_cast<BuiltinInst>(inst);
  return bi && bi->getBuiltinInfo().ID == BuiltinValueKind::DestroyArray;
}

/// Analyze the destructor for the class of ARI to see if any instructions in it
/// could have side effects on the program outside the destructor. If it does
/// not, then we can eliminate the destructor.
/// TODO: Most default destructors with non-trivial elements will have a
/// destroy_addr of the non-trivial element in the destructor, this analysis
/// will return as having side-effects in such cases, leading to conservative
/// results. Check if we can do better here.
static DestructorEffects doesDestructorHaveSideEffects(AllocRefInstBase *ARI) {
  SILFunction *Fn = getDestructor(ARI);
  // If we can't find a constructor then assume it has side effects.
  if (!Fn)
    return DestructorEffects::Unknown;

  DestructorEffects effects = DestructorEffects::None;

  // A destructor only has one argument, self.
  assert(Fn->begin()->getNumArguments() == 1 &&
         "Destructor should have only one argument, self.");
  SILArgument *Self = Fn->begin()->getArgument(0);

  LLVM_DEBUG(llvm::dbgs() << "    Analyzing destructor.\n");

  // For each BB in the destructor...
  for (auto &BB : *Fn) {
    // For each instruction I in BB...
    for (auto &I : BB) {
      LLVM_DEBUG(llvm::dbgs() << "        Visiting: " << I);

      // If I has no side effects, we can ignore it.
      if (!I.mayHaveSideEffects()) {
        LLVM_DEBUG(llvm::dbgs() << "            SAFE! Instruction has no side "
                   "effects.\n");
        continue;
      }

      if (auto *fl = dyn_cast<FixLifetimeInst>(&I)) {
        // A fix_lifetime of self does cannot have a side effect, because in the
        // destructor, Self is deleted.
        if (stripCasts(fl->getOperand()) == Self)
          continue;
        return DestructorEffects::Unknown;
      }

      // RefCounting operations on Self are ok since we are already in the
      // destructor. RefCountingOperations on other instructions could have side
      // effects though.
      if (auto *RefInst = dyn_cast<RefCountingInst>(&I)) {
        if (stripCasts(RefInst->getOperand(0)) == Self) {
          // For now all ref counting insts have 1 operand. Put in an assert
          // just in case.
          assert(RefInst->getNumOperands() == 1 &&
                 "Make sure RefInst only has one argument.");
          LLVM_DEBUG(llvm::dbgs() << "            SAFE! Ref count operation on "
                                     "Self.\n");
          continue;
        }
        LLVM_DEBUG(llvm::dbgs() << "            UNSAFE! Ref count operation "
                                   "not on self.\n");
        return DestructorEffects::Unknown;
      }
      if (auto *destroy = dyn_cast<DestroyValueInst>(&I)) {
        if (stripCasts(destroy->getOperand()) == Self) {
          LLVM_DEBUG(llvm::dbgs() << "            SAFE! Ref count operation on "
                                     "Self.\n");
          continue;
        }
        LLVM_DEBUG(llvm::dbgs() << "            UNSAFE! Ref count operation "
                                   "not on self.\n");
        return DestructorEffects::Unknown;
      }

      // dealloc_stack can be ignored.
      if (isa<DeallocStackInst>(I)) {
        LLVM_DEBUG(llvm::dbgs() << "            SAFE! dealloc_stack can be "
                   "ignored.\n");
        continue;
      }

      if (isa<BeginBorrowInst>(I) || isa<EndBorrowInst>(I) || isa<EndLifetimeInst>(I)) {
        continue;
      }

      // dealloc_ref on self can be ignored, but dealloc_ref on anything else
      // cannot be eliminated.
      if (auto *DeallocRef = dyn_cast<DeallocRefInst>(&I)) {
        if (stripCasts(DeallocRef->getOperand()) == Self) {
          LLVM_DEBUG(llvm::dbgs() <<"            SAFE! dealloc_ref on self.\n");
          continue;
        } else {
          LLVM_DEBUG(llvm::dbgs() << "            UNSAFE! dealloc_ref on value "
                     "besides self.\n");
          return DestructorEffects::Unknown;
        }
      }

      // Storing into the object can be ignored.
      if (auto *SI = dyn_cast<StoreInst>(&I))
        if (stripAddressProjections(SI->getDest()) == Self) {
          LLVM_DEBUG(llvm::dbgs() << "            SAFE! Instruction is a store "
                     "into self.\n");
          continue;
        }

      if (isDestroyArray(&I)) {
        // Check if the "destroyArray" destroys the tail elements of the object,
        // like in Array.
        SILValue addr = I.getOperand(1);
        auto *atp = dyn_cast<AddressToPointerInst>(addr);
        if (!atp)
          return DestructorEffects::Unknown;
        auto *rta = dyn_cast<RefTailAddrInst>(atp->getOperand());
        if (!rta)
          return DestructorEffects::Unknown;
        effects = DestructorEffects::DestroysTailElems;
        if (rta->getOperand() == Self)
          continue;
      }

      LLVM_DEBUG(llvm::dbgs() << "            UNSAFE! Unknown instruction.\n");
      // Otherwise, we can't remove the deallocation completely.
      return DestructorEffects::Unknown;
    }
  }
  // We didn't find any side effects.
  return effects;
}

//===----------------------------------------------------------------------===//
//                             Use Graph Analysis
//===----------------------------------------------------------------------===//

/// Returns false if Inst is an instruction that would require us to keep the
/// alloc_ref alive.
static bool canZapInstruction(SILInstruction *Inst, bool acceptRefCountInsts,
                              bool onlyAcceptTrivialStores) {
  if (isa<DestroyValueInst>(Inst)) {
    return acceptRefCountInsts;
  }
  if (isa<CopyValueInst>(Inst) || isa<BeginBorrowInst>(Inst) ||
      isa<MoveValueInst>(Inst)) {
    return true;
  }
  if (isa<EndInitLetRefInst>(Inst) || isa<BeginDeallocRefInst>(Inst) ||
      isa<FixLifetimeInst>(Inst) || isa<EndBorrowInst>(Inst) ||
      isa<UpcastInst>(Inst) || isa<UncheckedRefCastInst>(Inst))
    return true;

  // It is ok to eliminate various retains/releases. We are either removing
  // everything or nothing.
  if (isa<RefCountingInst>(Inst) ||
      // dealloc_partial_ref invokes releases implicitly
      isa<DeallocPartialRefInst>(Inst))
    return acceptRefCountInsts;

  if (isa<InjectEnumAddrInst>(Inst))
    return true;

  if (isa<KeyPathInst>(Inst))
    return true;

  // We know that the destructor has no side effects so we can remove the
  // deallocation instruction too.
  if (isa<DeallocationInst>(Inst) || isa<AllocationInst>(Inst))
    return true;

  // Much like deallocation, destroy addr is safe.
  if (isa<DestroyAddrInst>(Inst))
    return true;

  // We have already checked that we are storing into the pointer before we
  // added it to the worklist. Here, in the case we are allowing non-trivial
  // stores, check if the store's source is lexical, if so return false.
  // Deleting a dead object with non-trivial stores, will need compensating
  // destroys at the store for it's source, which will shorten the lifetime of
  // the store's source.
  if (auto *store = dyn_cast<StoreInst>(Inst)) {
    auto storeSrc = store->getSrc();
    return storeSrc->getType().isTrivial(*store->getFunction()) ||
           (!onlyAcceptTrivialStores &&
            (!store->getFunction()->hasOwnership() || !storeSrc->isLexical()));
  }

  // Conceptually this instruction has no side-effects.
  if (isa<InitExistentialAddrInst>(Inst))
    return true;

  if (isa<BeginAccessInst>(Inst) || isa<EndAccessInst>(Inst))
    return true;

  // The value form of zero init is not a user of any operand. The address
  // form however is easily zappable because it's always a trivial store.
  if (auto bi = dyn_cast<BuiltinInst>(Inst)) {
    if (bi->getBuiltinKind() == BuiltinValueKind::ZeroInitializer ||
        bi->getBuiltinKind() == BuiltinValueKind::PrepareInitialization) {
      return true;
    }
  }

  // If Inst does not read or write to memory, have side effects, and is not a
  // terminator, we can zap it.
  if (!Inst->mayHaveSideEffects() && !Inst->mayReadFromMemory() &&
      !isa<TermInst>(Inst))
    return true;

  // Otherwise we do not know how to handle this instruction. Be conservative
  // and don't zap it.
  return false;
}

/// Returns true if all stores in \p users store to the tail elements of
/// \p allocRef, which are destroyed by the \p destroyArray builtin.
static bool onlyStoresToTailObjects(BuiltinInst *destroyArray,
                                    const UserList &users,
                                    AllocRefInstBase *allocRef) {
  // Get the number of destroyed elements.
  auto *literal = dyn_cast<IntegerLiteralInst>(destroyArray->getArguments()[2]);
  if (!literal || literal->getValue().getSignificantBits() > 32)
    return false;
  int numDestroyed = literal->getValue().getSExtValue();
  
  SILFunction *func = destroyArray->getFunction();
  SILBasicBlock *storesBlock = nullptr;

  // Check if the destroyArray destroys the tail elements of allocRef.
  auto destroyPath = AccessPath::compute(destroyArray->getArguments()[1]);
  AccessStorage storage = destroyPath.getStorage();
  if (auto *beginDealloc = dyn_cast<BeginDeallocRefInst>(storage.getRoot())) {
    destroyPath = AccessPath(
                   storage.transformReference(beginDealloc->getAllocation()),
                   destroyPath.getPathNode(),
                   destroyPath.getOffset());
  }

  if (destroyPath != AccessPath::forTailStorage(allocRef))
    return false;

  SmallVector<AccessPath, 32> pathsToCheck;

  // Check all stores to the tail elements.
  for (SILInstruction *user : users) {
    auto *store = dyn_cast<StoreInst>(user);
    if (!store)
      continue;

    assert(users.count(store->getSrc()->getDefiningInstruction()) == 0 &&
           "Storing a use of an array (that would mean the array escapes)?");

    // All stores must be in the same block. This ensure that the stores
    // dominate the destroyArray (which may be in a different block).
    if (storesBlock && store->getParent() != storesBlock)
      return false;
    storesBlock = store->getParent();

    AccessPath storePath = AccessPath::compute(store->getDest());
    if (!storePath.isValid())
      return false;

    // We don't care about trivial stores.
    if (store->getSrc()->getType().isTrivial(*func))
      continue;
    
    // Check if it's a store to the tail elements.
    if (!destroyPath.contains(storePath.withOffset(0)))
      return false;

    // Check if the store is within the range of the destroyed array. In OSSA
    // we would not need this check. Otherwise it would be a memory lifetime
    // failure.
    if (storePath.getOffset() < 0 || storePath.getOffset() >= numDestroyed)
      return false;

    pathsToCheck.push_back(storePath);
  }

  // In non-OSSA we have to check if two paths overlap, because we could end up
  // over-releasing the stored objects.
  // Group the paths by tail-element index, so that we only have to check within
  // a tail-element group.
  std::sort(pathsToCheck.begin(), pathsToCheck.end(), [](AccessPath p1, AccessPath p2) {
    return p1.getOffset() < p2.getOffset();
  });
  for (unsigned i = 0, n = pathsToCheck.size(); i < n; ++i) {
    for (unsigned j = i + 1;
      j < n && pathsToCheck[i].getOffset() == pathsToCheck[j].getOffset(); ++j) {
      if (pathsToCheck[i].mayOverlap(pathsToCheck[j]))
        return false;
      // Limit the number of checks to avoid quadratic complexity.
      if (j > i + 8)
        return false;
    }
  }
  return true;
}

/// Analyze the use graph of AllocRef for any uses that would prevent us from
/// zapping it completely.
static bool
hasUnremovableUsers(SILInstruction *allocation, UserList *Users,
                    bool acceptRefCountInsts, bool onlyAcceptTrivialStores) {
  SmallVector<SILInstruction *, 16> Worklist;
  Worklist.push_back(allocation);

  LLVM_DEBUG(llvm::dbgs() << "    Analyzing Use Graph.");

  SmallVector<RefElementAddrInst *, 8> refElementAddrs;

  BuiltinInst *destroyArray = nullptr;
  auto *allocRef = dyn_cast<AllocRefInstBase>(allocation);

  while (!Worklist.empty()) {
    SILInstruction *I = Worklist.pop_back_val();

    LLVM_DEBUG(llvm::dbgs() << "        Visiting: " << *I);

    // Insert the instruction into our InvolvedInstructions set.  If we have
    // already seen it, then don't reprocess all of the uses.
    if (Users && !Users->insert(I)) {
      LLVM_DEBUG(llvm::dbgs() << "        Already seen skipping...\n");
      continue;
    } else if (auto *rea = dyn_cast<RefElementAddrInst>(I)) {
      if (rea != allocation && !rea->getType().isTrivial(*rea->getFunction()))
        refElementAddrs.push_back(rea);
    } else if (allocRef && isDestroyArray(I)) {
      if (destroyArray)
        return true;
      destroyArray = cast<BuiltinInst>(I);
    } else if (!canZapInstruction(I, acceptRefCountInsts,
                                  onlyAcceptTrivialStores)) {
      LLVM_DEBUG(llvm::dbgs() << "        Found instruction we can't zap...\n");
      return true;
    }

    // At this point, we can remove the instruction as long as all of its users
    // can be removed as well. Scan its users and add them to the worklist for
    // recursive processing.
    for (auto result : I->getResults()) {
      for (auto *Op : result->getUses()) {
        auto *User = Op->getUser();

        // Make sure that we are only storing into our users, not storing our
        // users which would be an escape.
        if (auto *SI = dyn_cast<StoreInst>(User))
          if (Op->get() == SI->getSrc()) {
            LLVM_DEBUG(llvm::dbgs() << "        Found store of pointer. "
                                       "Failure: "
                                    << *SI);
            return true;
          }

        // Otherwise, add normal instructions to the worklist for processing.
        Worklist.push_back(User);
      }
    }
  }

  if (!allocation->getFunction()->hasOwnership()) {
    // In non-ossa, if we found a destroy array builtin that destroys the tail
    // elements, ensure all stores are to the taile elems.
    if (destroyArray) {
      return !onlyStoresToTailObjects(destroyArray, *Users, allocRef);
    }
    // In non-OSSA we cannot reliably track the lifetime of non-trivial stored
    // properties. Removing the dead alloc_ref might leak a property value.
    for (RefElementAddrInst *rea : refElementAddrs) {
      // Re-run the check with not accepting non-trivial stores.
      if (hasUnremovableUsers(rea, nullptr, acceptRefCountInsts,
                              /*onlyAcceptTrivialStores*/ true))
        return true;
    }
  }

  return false;
}

//===----------------------------------------------------------------------===//
//                     NonTrivial DeadObject Elimination
//===----------------------------------------------------------------------===//

/// Determine if an object is dead. Compute its original lifetime. Find the
/// lifetime endpoints reached by each store of a refcounted object into the
/// object.
///
/// TODO: Use this to remove nontrivial dead alloc_ref/alloc_stack, not just
/// dead arrays. We just need a slightly better destructor analysis to prove
/// that it only releases elements.
class DeadObjectAnalysis {
  // Map each address projection of this object to a list of stores.
  // Do not iterate over this map's entries.
  using AddressToStoreMap =
    llvm::DenseMap<IndexTrieNode*, llvm::SmallVector<StoreInst*, 4> >;

  // The value of the object's address at the point of allocation.
  SILValue NewAddrValue;

  // Track all users that extend the lifetime of the object.
  UserList AllUsers;

  // Trie of stored locations.
  std::unique_ptr<IndexTrieNode> AddressProjectionTrie;

  // Track all stores of refcounted elements per address projection.
  AddressToStoreMap StoredLocations;

  // Are any uses behind a PointerToAddressInst?
  bool SeenPtrToAddr;

public:
  explicit DeadObjectAnalysis(SILValue V):
    NewAddrValue(V), AddressProjectionTrie(nullptr), SeenPtrToAddr(false) {}

  bool analyze();

  ArrayRef<SILInstruction*> getAllUsers() const {
    return ArrayRef<SILInstruction*>(AllUsers.begin(), AllUsers.end());
  }

  template<typename Visitor>
  void visitStoreLocations(Visitor visitor) {
    visitStoreLocations(visitor, AddressProjectionTrie.get());
  }

private:
  void addStore(StoreInst *Store, IndexTrieNode *AddressNode);
  bool recursivelyCollectInteriorUses(ValueBase *DefInst,
                                      IndexTrieNode *AddressNode,
                                      bool IsInteriorAddress);

  template<typename Visitor>
  void visitStoreLocations(Visitor visitor, IndexTrieNode *AddressNode);
};
} // end anonymous namespace

// Record a store into this object.
void DeadObjectAnalysis::
addStore(StoreInst *Store, IndexTrieNode *AddressNode) {
  if (Store->getSrc()->getType().isTrivial(*Store->getFunction()))
    return;

  // SSAUpdater cannot handle multiple defs in the same blocks. Therefore, we
  // ensure that only one store per block is present in the StoredLocations.
  auto &StoredLocs = StoredLocations[AddressNode];
  for (auto &OtherSt : StoredLocs) {
    // In case the object's address is stored in itself.
    if (OtherSt == Store)
      return;

    if (OtherSt->getParent() == Store->getParent()) {
      for (auto II = std::next(Store->getIterator()),
                IE = Store->getParent()->end();
           II != IE; ++II) {
        if (&*II == OtherSt)
          return; // Keep the other store.
      }
      // Replace OtherSt with this store.
      OtherSt = Store;
      return;
    }
  }
  StoredLocations[AddressNode].push_back(Store);
}

// Collect instructions that either initialize or release any values at the
// object defined by defInst.
//
// Populates AllUsers, AddressProjectionTrie, and StoredLocations.
//
// If a use is visited that potentially causes defInst's address to
// escape, then return false without fully populating the data structures.
//
// `InteriorAddress` is true if the current address projection already includes
// a struct/ref/tuple element address. index_addr is only expected at the top
// level. The first non-index element address encountered pushes an "zero index"
// address node to represent the implicit index_addr #0. We do not support
// nested indexed data types in native SIL.
bool DeadObjectAnalysis::
recursivelyCollectInteriorUses(ValueBase *DefInst,
                               IndexTrieNode* AddressNode,
                               bool IsInteriorAddress) {
  for (auto Op : DefInst->getUses()) {
    auto User = Op->getUser();

    // Lifetime endpoints that don't allow the address to escape.
    if (isa<RefCountingInst>(User) || isa<DebugValueInst>(User) ||
        isa<FixLifetimeInst>(User) || isa<DestroyValueInst>(User) ||
        isa<EndBorrowInst>(User)) {
      AllUsers.insert(User);
      continue;
    }
    // Initialization points.
    if (auto *Store = dyn_cast<StoreInst>(User)) {
      // Bail if this address is stored to another object.
      if (Store->getDest() != DefInst) {
        LLVM_DEBUG(llvm::dbgs() <<"        Found an escaping store: " << *User);
        return false;
      }
      IndexTrieNode *StoreAddrNode = AddressNode;
      // Push an extra zero index node for a store to noninterior address.
      if (!IsInteriorAddress)
        StoreAddrNode = AddressNode->getChild(0);

      addStore(Store, StoreAddrNode);

      AllUsers.insert(User);
      continue;
    }
    if (auto *MDI = dyn_cast<MarkDependenceInst>(User)) {
      if (!recursivelyCollectInteriorUses(MDI, AddressNode,
                                          IsInteriorAddress)) {
        return false;
      }
      continue;
    }
    if (auto *bb = dyn_cast<BeginBorrowInst>(User)) {
      if (!recursivelyCollectInteriorUses(bb, AddressNode,
                                          IsInteriorAddress)) {
        return false;
      }
      continue;
    }
    if (auto PTAI = dyn_cast<PointerToAddressInst>(User)) {
      // Only one pointer-to-address is allowed for safety.
      if (SeenPtrToAddr)
        return false;

      SeenPtrToAddr = true;
      if (!recursivelyCollectInteriorUses(PTAI, AddressNode, IsInteriorAddress))
        return false;

      continue;
    }
    // Recursively follow projections.
    if (auto *svi = dyn_cast<SingleValueInstruction>(User)) {
      ProjectionIndex PI(svi);
      if (PI.isValid()) {
        IndexTrieNode *ProjAddrNode = AddressNode;
        bool ProjInteriorAddr = IsInteriorAddress;
        if (Projection::isAddressProjection(svi)) {
          if (isa<IndexAddrInst>(svi)) {
            // Don't support indexing within an interior address.
            if (IsInteriorAddress)
              return false;
          }
          else if (!IsInteriorAddress) {
            // Push an extra zero index node for the first interior address.
            ProjAddrNode = AddressNode->getChild(0);
            ProjInteriorAddr = true;
          }
        }
        else if (IsInteriorAddress) {
          // Don't expect to extract values once we've taken an address.
          return false;
        }
        if (!recursivelyCollectInteriorUses(svi,
                                            ProjAddrNode->getChild(PI.Index),
                                            ProjInteriorAddr)) {
          return false;
        }
        continue;
      }
      ArraySemanticsCall AS(svi);
      if (AS.getKind() == ArrayCallKind::kArrayFinalizeIntrinsic) {
        if (!recursivelyCollectInteriorUses(svi, AddressNode, IsInteriorAddress))
          return false;
        continue;
      }
    }
    // Otherwise bail.
    LLVM_DEBUG(llvm::dbgs() << "        Found an escaping use: " << *User);
    return false;
  }
  return true;
}

// Track the lifetime, release points, and released values referenced by a
// newly allocated object.
bool DeadObjectAnalysis::analyze() {
  LLVM_DEBUG(llvm::dbgs() << "    Analyzing nontrivial dead object: "
                          << NewAddrValue);

  // Populate AllValues, AddressProjectionTrie, and StoredLocations.
  AddressProjectionTrie.reset(new IndexTrieNode());
  if (!recursivelyCollectInteriorUses(NewAddrValue,
                                      AddressProjectionTrie.get(), false)) {
    return false;
  }
  // If all stores are leaves in the AddressProjectionTrie, then we can analyze
  // the stores that reach the end of the object lifetime. Otherwise bail.
  // This iteration order is nondeterministic but has no impact.
  for (auto &AddressToStoresPair : StoredLocations) {
    IndexTrieNode *Location = AddressToStoresPair.first;
    if (!Location->isLeaf())
      return false;
  }
  return true;
}

template<typename Visitor>
void DeadObjectAnalysis::
visitStoreLocations(Visitor visitor, IndexTrieNode *AddressNode) {
  if (AddressNode->isLeaf()) {
    auto LocI = StoredLocations.find(AddressNode);
    if (LocI != StoredLocations.end())
      visitor(LocI->second);
    return;
  }
  for (auto *SubAddressNode : AddressNode->getChildren())
    visitStoreLocations(visitor, SubAddressNode);
}

// At each release point, release the reaching values that have been stored to
// this address.
//
// The caller has already determined that all Stores are to the same element
// within an otherwise dead object.
static void insertReleases(ArrayRef<StoreInst*> Stores,
                           ArrayRef<SILInstruction*> ReleasePoints,
                           SILSSAUpdater &SSAUp) {
  assert(!Stores.empty());
  SILValue StVal = Stores.front()->getSrc();

  SSAUp.initialize(StVal->getFunction(), StVal->getType(),
                   StVal->getOwnershipKind());

  for (auto *Store : Stores)
    SSAUp.addAvailableValue(Store->getParent(), Store->getSrc());

  SILLocation Loc = Stores[0]->getLoc();
  for (auto *RelPoint : ReleasePoints) {
    SILBuilder B(RelPoint);
    // This does not use the SSAUpdater::RewriteUse API because it does not do
    // the right thing for local uses. We have already ensured a single store
    // per block, and all release points occur after all stores. Therefore we
    // can simply ask SSAUpdater for the reaching store.
    SILValue RelVal = SSAUp.getValueAtEndOfBlock(RelPoint->getParent());
    B.emitDestroyValueOperation(Loc, RelVal);
  }
}

//===----------------------------------------------------------------------===//
//                            Function Processing
//===----------------------------------------------------------------------===//

/// Does this instruction perform object allocation with no other observable
/// side effect?
static bool isAllocatingApply(SILInstruction *Inst) {
  ArraySemanticsCall ArrayAlloc(Inst);
  return ArrayAlloc.getKind() == ArrayCallKind::kArrayUninitialized ||
         ArrayAlloc.getKind() == ArrayCallKind::kArrayUninitializedIntrinsic;
}

namespace {
class DeadObjectElimination : public SILFunctionTransform {

  llvm::DenseMap<SILType, DestructorEffects> DestructorAnalysisCache;

  InstructionDeleter deleter;
  DominanceInfo *domInfo = nullptr;

  void removeInstructions(ArrayRef<SILInstruction*> toRemove);
  
  /// Try to salvage the debug info for a dead instruction removed by
  /// DeadObjectElimination.
  ///
  /// Dead stores will be replaced by a debug value for the object variable,
  /// using a fragment expression. By walking from the store to the allocation,
  /// we can know which member of the object is being assigned, and create
  /// fragments for each member. Other instructions are not salvaged.
  /// Currently only supports dead stack-allocated objects.
  void salvageDebugInfo(SILInstruction *toBeRemoved);
  std::optional<SILDebugVariable> buildDIExpression(SILInstruction *current);

  bool processAllocRef(AllocRefInstBase *ARI);
  bool processAllocStack(AllocStackInst *ASI);
  bool processKeyPath(KeyPathInst *KPI);
  bool processAllocBox(AllocBoxInst *ABI){ return false;}
  bool processAllocApply(ApplyInst *AI, DeadEndBlocks &DEBlocks);

  bool insertCompensatingReleases(SILInstruction *before,
                                  const UserList &users);

  bool getDeadInstsAfterInitializerRemoved(
    ApplyInst *AI, llvm::SmallVectorImpl<SILInstruction *> &ToDestroy);
  bool removeAndReleaseArray(
    SingleValueInstruction *NewArrayValue, DeadEndBlocks &DEBlocks);

  bool processFunction(SILFunction &Fn) {
    DeadEndBlocks DEBlocks(&Fn);
    DestructorAnalysisCache.clear();

    LLVM_DEBUG(llvm::dbgs() << "Processing " << Fn.getName() << "\n");

    bool Changed = false;

    for (auto &BB : Fn) {
      for (SILInstruction &inst : BB.deletableInstructions()) {
        if (auto *A = dyn_cast<AllocRefInstBase>(&inst))
          Changed |= processAllocRef(A);
        else if (auto *A = dyn_cast<AllocStackInst>(&inst))
          Changed |= processAllocStack(A);
        else if (auto *KPI = dyn_cast<KeyPathInst>(&inst))
          Changed |= processKeyPath(KPI);
        else if (auto *A = dyn_cast<AllocBoxInst>(&inst))
          Changed |= processAllocBox(A);
        else if (auto *A = dyn_cast<ApplyInst>(&inst))
          Changed |= processAllocApply(A, DEBlocks);
      }
      deleter.cleanupDeadInstructions();
    }
    return Changed;
  }

  void run() override {
    assert(!domInfo);

    if (processFunction(*getFunction())) {
      invalidateAnalysis(SILAnalysis::InvalidationKind::CallsAndInstructions);
    }
    domInfo = nullptr;
  }

};
} // end anonymous namespace

void
DeadObjectElimination::removeInstructions(ArrayRef<SILInstruction*> toRemove) {
  for (auto *I : toRemove) {
    I->replaceAllUsesOfAllResultsWithUndef();
    // Now we know that I should not have any uses... erase it from its parent.
    deleter.forceDelete(I);
  }
}

void DeadObjectElimination::salvageDebugInfo(SILInstruction *toBeRemoved) {
  auto *SI = dyn_cast<StoreInst>(toBeRemoved);
  if (!SI)
    return;

  auto *parent = SI->getDest()->getDefiningInstruction();
  auto varInfo = buildDIExpression(parent);
  if (!varInfo)
    return;

  // Note: The instruction should logically be in SI's scope.
  // However, LLVM does not support variables and stores in different scopes,
  // so we use the variable's scope.
  SILBuilder Builder(SI, varInfo->Scope);
  Builder.createDebugValue(SI->getLoc(), SI->getSrc(), *varInfo);
}

std::optional<SILDebugVariable>
DeadObjectElimination::buildDIExpression(SILInstruction *current) {
  if (!current)
    return {};
  if (auto dvci = dyn_cast<AllocStackInst>(current)) {
    auto var = dvci->getVarInfo();
    if (!var)
      return {};
    if (!var->Type)
      var->Type = dvci->getElementType();
    return var;
  }
  if (auto *tupleAddr = dyn_cast<TupleElementAddrInst>(current)) {
    auto *definer = tupleAddr->getOperand().getDefiningInstruction();
    auto path = buildDIExpression(definer);
    if (!path)
      return {};
    path->DIExpr.append(SILDebugInfoExpression::createTupleFragment(
      tupleAddr->getTupleType(), tupleAddr->getFieldIndex()));
    return path;
  }
  if (auto *structAddr = dyn_cast<StructElementAddrInst>(current)) {
    auto *definer = structAddr->getOperand().getDefiningInstruction();
    auto path = buildDIExpression(definer);
    if (!path)
      return {};
    path->DIExpr.append(SILDebugInfoExpression::createFragment(
      structAddr->getField()));
    return path;
  }
  return {};
}

bool DeadObjectElimination::processAllocRef(AllocRefInstBase *ARI) {
  // Ok, we have an alloc_ref. Check the cache to see if we have already
  // computed the destructor behavior for its SILType.
  DestructorEffects destructorEffects;
  SILType Type = ARI->getType();
  auto CacheSearchResult = DestructorAnalysisCache.find(Type);
  if (CacheSearchResult != DestructorAnalysisCache.end()) {
    // Ok we found a value in the cache.
    destructorEffects = CacheSearchResult->second;
  } else {
    // We did not find a value in the cache for our destructor. Analyze the
    // destructor to make sure it has no side effects. For now this only
    // supports alloc_ref of classes so any alloc_ref with a reference type
    // that is not a class this will return false for. Once we have analyzed
    // it, set Behavior to that value and insert the value into the Cache.
    //
    // TODO: We should be able to handle destructors that do nothing but release
    // members of the object.
    destructorEffects = doesDestructorHaveSideEffects(ARI);
    DestructorAnalysisCache[Type] = destructorEffects;
  }

  // Our destructor has no side effects, so if we can prove that no loads
  // escape, then we can completely remove the use graph of this alloc_ref.
  UserList UsersToRemove;
  if (hasUnremovableUsers(ARI, &UsersToRemove,
      /*acceptRefCountInsts=*/ destructorEffects != DestructorEffects::Unknown,
      /*onlyAcceptTrivialStores*/false)) {
    LLVM_DEBUG(llvm::dbgs() << "    Found a use that cannot be zapped...\n");
    return false;
  }

  if (!ARI->getFunction()->hasOwnership()) {
    // Find the instruction which releases the object's tail elements.
    SILInstruction *releaseOfTailElems = nullptr;
    for (SILInstruction *user : UsersToRemove) {
      if (isDestroyArray(user) ||
          (destructorEffects == DestructorEffects::DestroysTailElems &&
           isa<RefCountingInst>(user) && user->mayRelease())) {
        // Bail if we find multiple such instructions.
        if (releaseOfTailElems)
          return false;
        releaseOfTailElems = user;
      }
    }
    if (releaseOfTailElems) {
      if (!insertCompensatingReleases(releaseOfTailElems, UsersToRemove)) {
        return false;
      }
    }
  }

  if (ARI->getFunction()->hasOwnership()) {
    // In ossa, we are going to delete the dead element store and insert a
    // destroy_value of the store's source. This is shortening the store's
    // source lifetime. Check if there was a pointer escape of the store's
    // source, if so bail out.
    for (auto *user : UsersToRemove) {
      auto *store = dyn_cast<StoreInst>(user);
      if (!store ||
          store->getOwnershipQualifier() == StoreOwnershipQualifier::Trivial)
        continue;
      if (findPointerEscape(store->getSrc())) {
        return false;
      }
    }
    for (auto *user : UsersToRemove) {
      auto *store = dyn_cast<StoreInst>(user);
      if (!store ||
          store->getOwnershipQualifier() == StoreOwnershipQualifier::Trivial) {
        continue;
      }
      SILBuilderWithScope(store).createDestroyValue(store->getLoc(),
                                                    store->getSrc());
    }
  }

  // Remove the AllocRef and all of its users.
  removeInstructions(
    ArrayRef<SILInstruction*>(UsersToRemove.begin(), UsersToRemove.end()));
  LLVM_DEBUG(llvm::dbgs() << "    Success! Eliminating alloc_ref.\n");

  ++DeadAllocRefEliminated;
  return true;
}

bool DeadObjectElimination::processAllocStack(AllocStackInst *ASI) {
  // Trivial types don't have destructors.
  bool isTrivialType = ASI->getElementType().isTrivial(*ASI->getFunction());
  // In non-ossa, only accept trivial stores if we have a non-trivial
  // alloc_stack
  bool onlyAcceptTrivialStores =
      ASI->getFunction()->hasOwnership() ? false : !isTrivialType;
  UserList UsersToRemove;
  if (hasUnremovableUsers(ASI, &UsersToRemove, /*acceptRefCountInsts=*/true,
                          onlyAcceptTrivialStores)) {
    LLVM_DEBUG(llvm::dbgs() << "    Found a use that cannot be zapped...\n");
    return false;
  }

  for (auto *I : UsersToRemove)
    salvageDebugInfo(I);

  if (ASI->getFunction()->hasOwnership()) {
    for (auto *user : UsersToRemove) {
      auto *store = dyn_cast<StoreInst>(user);
      if (!store ||
          store->getOwnershipQualifier() == StoreOwnershipQualifier::Trivial)
        continue;
      // In ossa, we are going to delete the dead store and insert a
      // destroy_value of the store's source. This is shortening the store's
      // source lifetime. Check if there was a pointer escape of the store's
      // source, if so bail out.
      if (findPointerEscape(store->getSrc())) {
        return false;
      }
    }
    for (auto *user : UsersToRemove) {
      auto *store = dyn_cast<StoreInst>(user);
      if (!store ||
          store->getOwnershipQualifier() == StoreOwnershipQualifier::Trivial)
        continue;
      SILBuilderWithScope(store).createDestroyValue(store->getLoc(),
                                                    store->getSrc());
    }
  }

  // Remove the AllocRef and all of its users.
  removeInstructions(
    ArrayRef<SILInstruction*>(UsersToRemove.begin(), UsersToRemove.end()));
  LLVM_DEBUG(llvm::dbgs() << "    Success! Eliminating alloc_stack.\n");

  ++DeadAllocStackEliminated;
  return true;
}

bool DeadObjectElimination::processKeyPath(KeyPathInst *KPI) {
  UserList UsersToRemove;
  if (hasUnremovableUsers(KPI, &UsersToRemove, /*acceptRefCountInsts=*/ true,
      /*onlyAcceptTrivialStores*/ false)) {
    LLVM_DEBUG(llvm::dbgs() << "    Found a use that cannot be zapped...\n");
    return false;
  }

  bool hasOwnership = KPI->getFunction()->hasOwnership();
  for (const Operand &Op : KPI->getPatternOperands()) {
    // In non-ossa, bail out if we have non-trivial pattern operands.
    if (!hasOwnership) {
      if (Op.get()->getType().isTrivial(*KPI->getFunction()))
        return false;
      continue;
    }
    // In ossa, bail out if we have non-trivial pattern operand values that are
    // lexical.
    if (Op.get()->isLexical()) {
      return false;
    }
  }

  if (KPI->getFunction()->hasOwnership()) {
    for (const Operand &Op : KPI->getPatternOperands()) {
      if (Op.get()->getType().isTrivial(*KPI->getFunction()))
        continue;
      // In ossa, we are going to delete the dead keypath which was consuming
      // the pattern operand and insert a destroy_value of the pattern operand
      // value. This is shortening the pattern operand value's lifetime. Check
      // if there was a pointer escape, if so bail out.
      if (findPointerEscape(Op.get())) {
        return false;
      }
    }
    for (const Operand &Op : KPI->getPatternOperands()) {
      if (Op.get()->getType().isTrivial(*KPI->getFunction()))
        continue;
      SILBuilderWithScope(KPI).createDestroyValue(KPI->getLoc(), Op.get());
    }
  }

  // Remove the keypath and all of its users.
  removeInstructions(
    ArrayRef<SILInstruction*>(UsersToRemove.begin(), UsersToRemove.end()));
  LLVM_DEBUG(llvm::dbgs() << "    Success! Eliminating keypath.\n");

  ++DeadKeyPathEliminated;
  return true;
}

/// If AI is the version of an initializer where we pass in either an apply or
/// an alloc_ref to initialize in place, validate that we are able to continue
/// optimizing and return To
bool DeadObjectElimination::getDeadInstsAfterInitializerRemoved(
    ApplyInst *AI, llvm::SmallVectorImpl<SILInstruction *> &ToDestroy) {
  assert(ToDestroy.empty() && "We assume that ToDestroy is empty, so on "
                              "failure we can clear without worrying about the "
                              "caller accumulating and thus our eliminating "
                              "passed in state.");
  SILValue Arg0 = AI->getArgument(0);

  if (Arg0->getType().isExistentialType()) {
    // This is a version of the initializer which receives a pre-allocated
    // buffer as first argument. To completely eliminate the allocation, we must
    // destroy the extra allocations as well as the initializer,
    if (auto *Result = dyn_cast<ApplyInst>(Arg0)) {
      ToDestroy.emplace_back(Result);
      return true;
    }

    return false;
  }

  if (auto *ARI = dyn_cast<AllocRefInstBase>(Arg0)) {
    if (all_of(ARI->getUses(), [&](Operand *Op) -> bool {
          auto *user = Op->getUser();
          if (user == AI)
            return true;
          if (isa<StrongReleaseInst>(user) || isa<DestroyValueInst>(user)) {
            ToDestroy.emplace_back(user);
            return true;
          }
          return false;
        })) {
      return true;
    }
  }

  // We may have added elements to the array before we failed. To avoid such a
  // problem, we clear the out array here. We assert at the beginning that the
  // out array is empty, so this is safe.
  ToDestroy.clear();
  return true;
}

// Attempt to remove the array allocated at NewAddrValue and release its
// refcounted elements.
//
// This is tightly coupled with the implementation of array.uninitialized.
// The call to allocate an uninitialized array returns two values:
// (Array<E> ArrayBase, UnsafeMutable<E> ArrayElementStorage)
//
// TODO: This relies on the lowest level array.uninitialized not being
// inlined. To do better we could either run this pass before semantic inlining,
// or we could also handle calls to array.init.
bool DeadObjectElimination::removeAndReleaseArray(
    SingleValueInstruction *NewArrayValue, DeadEndBlocks &DEBlocks) {
  SILValue ArrayDef = nullptr;
  SILValue StorageAddress = nullptr;

  if (NewArrayValue->getFunction()->hasOwnership()) {
    auto *destructureTuple =
        NewArrayValue->getSingleConsumingUserOfType<DestructureTupleInst>();
    if (!destructureTuple) {
      return false;
    }
    if (destructureTuple->getNumResults() != 2) {
      return false;
    }
    ArrayDef = destructureTuple->getResult(0);
    StorageAddress = destructureTuple->getResult(1);
  } else {
    for (auto *Op : NewArrayValue->getUses()) {
      auto *TupleElt = dyn_cast<TupleExtractInst>(Op->getUser());
      if (!TupleElt)
        return false;
      if (TupleElt->getFieldIndex() == 0 && !ArrayDef) {
        ArrayDef = TupleElt;
      } else if (TupleElt->getFieldIndex() == 1 && !StorageAddress) {
        StorageAddress = TupleElt;
      } else {
        return false;
      }
    }
  }
  if (!ArrayDef)
    return false; // No Array object to delete.

  assert(!ArrayDef->getType().isTrivial(*ArrayDef->getFunction()) &&
         "Array initialization should produce the proper tuple type.");

  // Analyze the array object uses.
  DeadObjectAnalysis DeadArray(ArrayDef);
  if (!DeadArray.analyze())
    return false;

  // Require all stores to be into the array storage not the array object,
  // otherwise bail.
  bool HasStores = false;
  DeadArray.visitStoreLocations([&](ArrayRef<StoreInst*>){ HasStores = true; });
  if (HasStores)
    return false;

  // Remove references to empty arrays.
  if (!StorageAddress) {
    removeInstructions(DeadArray.getAllUsers());
    return true;
  }
  assert(StorageAddress->getType().isTrivial(*ArrayDef->getFunction()) &&
         "Array initialization should produce the proper tuple type.");

  // Analyze the array storage uses.
  DeadObjectAnalysis DeadStorage(StorageAddress);
  if (!DeadStorage.analyze())
    return false;

  // Find array object lifetime.
  ValueLifetimeAnalysis VLA(NewArrayValue, DeadArray.getAllUsers());

  // Check that all storage users are in the Array's live blocks.
  for (auto *User : DeadStorage.getAllUsers()) {
    if (!VLA.isWithinLifetime(User))
      return false;
  }
  // For each store location, insert releases.
  SILSSAUpdater SSAUp;
  ValueLifetimeAnalysis::Frontier ArrayFrontier;
  if (!VLA.computeFrontier(ArrayFrontier,
                           ValueLifetimeAnalysis::UsersMustPostDomDef,
                           &DEBlocks)) {
    // In theory the allocated object must be released on all paths in which
    // some object initialization occurs. If not (for some reason) we bail.
    return false;
  }

  DeadStorage.visitStoreLocations([&] (ArrayRef<StoreInst*> Stores) {
      insertReleases(Stores, ArrayFrontier, SSAUp);
    });

  // Delete all uses of the dead array and its storage address.
  removeInstructions(DeadArray.getAllUsers());
  removeInstructions(DeadStorage.getAllUsers());

  return true;
}

bool DeadObjectElimination::processAllocApply(ApplyInst *AI,
                                              DeadEndBlocks &DEBlocks) {
  // Currently only handle array.uninitialized
  if (!isAllocatingApply(AI))
    return false;

  llvm::SmallVector<SILInstruction *, 8> instsDeadAfterInitializerRemoved;
  if (!getDeadInstsAfterInitializerRemoved(AI,
                                           instsDeadAfterInitializerRemoved))
    return false;

  if (!removeAndReleaseArray(AI, DEBlocks))
    return false;

  LLVM_DEBUG(llvm::dbgs() << "    Success! Eliminating apply allocate(...).\n");

  auto *ARI = dyn_cast<AllocRefInst>(AI->getArgument(0));

  deleter.forceDeleteWithUsers(AI);
  for (auto *toDelete : instsDeadAfterInitializerRemoved) {
    deleter.trackIfDead(toDelete);
  }

  if (ARI) {
    deleter.forceDeleteWithUsers(ARI);
  }

  ++DeadAllocApplyEliminated;
  return true;
}

/// Inserts releases of all stores in \p users.
/// Returns false, if this is not possible.
bool DeadObjectElimination::insertCompensatingReleases(SILInstruction *before,
                                                       const UserList &users) {

  // First check if all stored values dominate the release-point.
  for (SILInstruction *user : users) {
    if (auto *store = dyn_cast<StoreInst>(user)) {
      if (!domInfo) {
        domInfo = getAnalysis<DominanceAnalysis>()->get(before->getFunction());
      }
      SILBasicBlock *srcBlock = store->getSrc()->getParentBlock();
      if (!domInfo->dominates(srcBlock, before->getParent()))
        return false;
    }
  }

  // Second, create the releases.
  for (SILInstruction *user : users) {
    if (auto *store = dyn_cast<StoreInst>(user)) {
      createDecrementBefore(store->getSrc(), before);
    }
  }
  return true;
}


//===----------------------------------------------------------------------===//
//                              Top Level Driver
//===----------------------------------------------------------------------===//

SILTransform *swift::createDeadObjectElimination() {
  return new DeadObjectElimination();
}