[snapshot] Refactor deserializer allocations

A continuation of the work in 59e4b751, this extracts logic around
memory reservation and allocations out of the Deserializer class.

Follow-up work is planned to create a specialized allocator for
builtin deserialization.

Bug: v8:6624
Change-Id: I7081cdc557ab8fb2571aadb816399e136ea2cdbb
Reviewed-on: https://chromium-review.googlesource.com/716036
Commit-Queue: Jakob Gruber <jgruber@chromium.org>
Reviewed-by: Yang Guo <yangguo@chromium.org>
Cr-Commit-Position: refs/heads/master@{#48634}

BUILD.gn

@@ -1968,6 +1968,8 @@ v8_source_set("v8_base") {
     "src/snapshot/builtin-serializer.h",
     "src/snapshot/code-serializer.cc",
     "src/snapshot/code-serializer.h",
+    "src/snapshot/default-deserializer-allocator.cc",
+    "src/snapshot/default-deserializer-allocator.h",
     "src/snapshot/default-serializer-allocator.cc",
     "src/snapshot/default-serializer-allocator.h",
     "src/snapshot/deserializer.cc",

src/snapshot/builtin-deserializer.cc

@@ -38,7 +38,7 @@ BuiltinDeserializer::BuiltinDeserializer(Isolate* isolate,
     : Deserializer(data, false) {
   // We may have to relax this at some point to pack reloc infos and handler
   // tables into the builtin blob (instead of the partial snapshot cache).
-  DCHECK(ReservesOnlyCodeSpace());
+  DCHECK(allocator()->ReservesOnlyCodeSpace());
 
   builtin_offsets_ = data->BuiltinOffsets();
   DCHECK_EQ(Builtins::builtin_count, builtin_offsets_.length());
@@ -136,7 +136,7 @@ uint32_t BuiltinDeserializer::ExtractBuiltinSize(int builtin_id) {
 }
 
 Heap::Reservation BuiltinDeserializer::CreateReservationsForEagerBuiltins() {
-  DCHECK(ReservesOnlyCodeSpace());
+  DCHECK(allocator()->ReservesOnlyCodeSpace());
 
   Heap::Reservation result;
 
@@ -236,7 +236,8 @@ Address BuiltinDeserializer::Allocate(int space_index, int size) {
   DCHECK_EQ(ExtractBuiltinSize(current_builtin_id_), size);
   Object* obj = isolate()->builtins()->builtin(current_builtin_id_);
   DCHECK(Internals::HasHeapObjectTag(obj));
-  return HeapObject::cast(obj)->address();
+  HeapObject* heap_obj = HeapObject::cast(obj);
+  return heap_obj->address();
 }
 
 }  // namespace internal

src/snapshot/builtin-deserializer.h

@@ -14,7 +14,7 @@ namespace internal {
 class BuiltinSnapshotData;
 
 // Deserializes the builtins blob.
-class BuiltinDeserializer final : public Deserializer {
+class BuiltinDeserializer final : public Deserializer<> {
  public:
   BuiltinDeserializer(Isolate* isolate, const BuiltinSnapshotData* data);
 

src/snapshot/default-deserializer-allocator.cc

+// Copyright 2017 the V8 project authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "src/snapshot/default-deserializer-allocator.h"
+
+#include "src/heap/heap-inl.h"
+#include "src/snapshot/builtin-deserializer.h"
+#include "src/snapshot/deserializer.h"
+#include "src/snapshot/startup-deserializer.h"
+
+namespace v8 {
+namespace internal {
+
+DefaultDeserializerAllocator::DefaultDeserializerAllocator(
+    Deserializer<DefaultDeserializerAllocator>* deserializer)
+    : deserializer_(deserializer) {}
+
+// We know the space requirements before deserialization and can
+// pre-allocate that reserved space. During deserialization, all we need
+// to do is to bump up the pointer for each space in the reserved
+// space. This is also used for fixing back references.
+// We may have to split up the pre-allocation into several chunks
+// because it would not fit onto a single page. We do not have to keep
+// track of when to move to the next chunk. An opcode will signal this.
+// Since multiple large objects cannot be folded into one large object
+// space allocation, we have to do an actual allocation when deserializing
+// each large object. Instead of tracking offset for back references, we
+// reference large objects by index.
+Address DefaultDeserializerAllocator::AllocateRaw(AllocationSpace space,
+                                                  int size) {
+  if (space == LO_SPACE) {
+    AlwaysAllocateScope scope(isolate());
+    LargeObjectSpace* lo_space = isolate()->heap()->lo_space();
+    // TODO(jgruber): May be cleaner to pass in executability as an argument.
+    Executability exec =
+        static_cast<Executability>(deserializer_->source()->Get());
+    AllocationResult result = lo_space->AllocateRaw(size, exec);
+    HeapObject* obj = result.ToObjectChecked();
+    deserialized_large_objects_.push_back(obj);
+    return obj->address();
+  } else if (space == MAP_SPACE) {
+    DCHECK_EQ(Map::kSize, size);
+    return allocated_maps_[next_map_index_++];
+  } else {
+    DCHECK(space < kNumberOfPreallocatedSpaces);
+    Address address = high_water_[space];
+    DCHECK_NOT_NULL(address);
+    high_water_[space] += size;
+#ifdef DEBUG
+    // Assert that the current reserved chunk is still big enough.
+    const Heap::Reservation& reservation = reservations_[space];
+    int chunk_index = current_chunk_[space];
+    DCHECK_LE(high_water_[space], reservation[chunk_index].end);
+#endif
+    if (space == CODE_SPACE) SkipList::Update(address, size);
+    return address;
+  }
+}
+
+Address DefaultDeserializerAllocator::Allocate(AllocationSpace space,
+                                               int size) {
+  Address address;
+  HeapObject* obj;
+
+  if (next_alignment_ != kWordAligned) {
+    const int reserved = size + Heap::GetMaximumFillToAlign(next_alignment_);
+    address = AllocateRaw(space, reserved);
+    obj = HeapObject::FromAddress(address);
+    // If one of the following assertions fails, then we are deserializing an
+    // aligned object when the filler maps have not been deserialized yet.
+    // We require filler maps as padding to align the object.
+    Heap* heap = isolate()->heap();
+    DCHECK(heap->free_space_map()->IsMap());
+    DCHECK(heap->one_pointer_filler_map()->IsMap());
+    DCHECK(heap->two_pointer_filler_map()->IsMap());
+    obj = heap->AlignWithFiller(obj, size, reserved, next_alignment_);
+    address = obj->address();
+    next_alignment_ = kWordAligned;
+    return address;
+  } else {
+    return AllocateRaw(space, size);
+  }
+}
+
+void DefaultDeserializerAllocator::MoveToNextChunk(AllocationSpace space) {
+  DCHECK(space < kNumberOfPreallocatedSpaces);
+  uint32_t chunk_index = current_chunk_[space];
+  const Heap::Reservation& reservation = reservations_[space];
+  // Make sure the current chunk is indeed exhausted.
+  CHECK_EQ(reservation[chunk_index].end, high_water_[space]);
+  // Move to next reserved chunk.
+  chunk_index = ++current_chunk_[space];
+  CHECK_LT(chunk_index, reservation.size());
+  high_water_[space] = reservation[chunk_index].start;
+}
+
+HeapObject* DefaultDeserializerAllocator::GetMap(uint32_t index) {
+  DCHECK_LT(index, next_map_index_);
+  return HeapObject::FromAddress(allocated_maps_[index]);
+}
+
+HeapObject* DefaultDeserializerAllocator::GetLargeObject(uint32_t index) {
+  DCHECK_LT(index, deserialized_large_objects_.size());
+  return deserialized_large_objects_[index];
+}
+
+HeapObject* DefaultDeserializerAllocator::GetObject(AllocationSpace space,
+                                                    uint32_t chunk_index,
+                                                    uint32_t chunk_offset) {
+  DCHECK_LT(space, kNumberOfPreallocatedSpaces);
+  DCHECK_LE(chunk_index, current_chunk_[space]);
+  Address address = reservations_[space][chunk_index].start + chunk_offset;
+  if (next_alignment_ != kWordAligned) {
+    int padding = Heap::GetFillToAlign(address, next_alignment_);
+    next_alignment_ = kWordAligned;
+    DCHECK(padding == 0 || HeapObject::FromAddress(address)->IsFiller());
+    address += padding;
+  }
+  return HeapObject::FromAddress(address);
+}
+
+void DefaultDeserializerAllocator::DecodeReservation(
+    Vector<const SerializedData::Reservation> res) {
+  DCHECK_EQ(0, reservations_[NEW_SPACE].size());
+  STATIC_ASSERT(NEW_SPACE == 0);
+  int current_space = NEW_SPACE;
+  for (auto& r : res) {
+    reservations_[current_space].push_back({r.chunk_size(), NULL, NULL});
+    if (r.is_last()) current_space++;
+  }
+  DCHECK_EQ(kNumberOfSpaces, current_space);
+  for (int i = 0; i < kNumberOfPreallocatedSpaces; i++) current_chunk_[i] = 0;
+}
+
+bool DefaultDeserializerAllocator::ReserveSpace() {
+#ifdef DEBUG
+  for (int i = NEW_SPACE; i < kNumberOfSpaces; ++i) {
+    DCHECK(reservations_[i].size() > 0);
+  }
+#endif  // DEBUG
+  DCHECK(allocated_maps_.empty());
+  if (!isolate()->heap()->ReserveSpace(reservations_, &allocated_maps_)) {
+    return false;
+  }
+  for (int i = 0; i < kNumberOfPreallocatedSpaces; i++) {
+    high_water_[i] = reservations_[i][0].start;
+  }
+  return true;
+}
+
+// static
+bool DefaultDeserializerAllocator::ReserveSpace(
+    StartupDeserializer* startup_deserializer,
+    BuiltinDeserializer* builtin_deserializer) {
+  const int first_space = NEW_SPACE;
+  const int last_space = SerializerDeserializer::kNumberOfSpaces;
+  Isolate* isolate = startup_deserializer->isolate();
+
+  // Create a set of merged reservations to reserve space in one go.
+  // The BuiltinDeserializer's reservations are ignored, since our actual
+  // requirements vary based on whether lazy deserialization is enabled.
+  // Instead, we manually determine the required code-space.
+
+  Heap::Reservation merged_reservations[kNumberOfSpaces];
+  for (int i = first_space; i < last_space; i++) {
+    merged_reservations[i] =
+        startup_deserializer->allocator()->reservations_[i];
+  }
+
+  Heap::Reservation builtin_reservations =
+      builtin_deserializer->CreateReservationsForEagerBuiltins();
+  DCHECK(!builtin_reservations.empty());
+
+  for (const auto& c : builtin_reservations) {
+    merged_reservations[CODE_SPACE].push_back(c);
+  }
+
+  if (!isolate->heap()->ReserveSpace(
+          merged_reservations,
+          &startup_deserializer->allocator()->allocated_maps_)) {
+    return false;
+  }
+
+  DisallowHeapAllocation no_allocation;
+
+  // Distribute the successful allocations between both deserializers.
+  // There's nothing to be done here except for code space.
+
+  {
+    const int num_builtin_reservations =
+        static_cast<int>(builtin_reservations.size());
+    for (int i = num_builtin_reservations - 1; i >= 0; i--) {
+      const auto& c = merged_reservations[CODE_SPACE].back();
+      DCHECK_EQ(c.size, builtin_reservations[i].size);
+      DCHECK_EQ(c.size, c.end - c.start);
+      builtin_reservations[i].start = c.start;
+      builtin_reservations[i].end = c.end;
+      merged_reservations[CODE_SPACE].pop_back();
+    }
+
+    builtin_deserializer->InitializeBuiltinsTable(builtin_reservations);
+  }
+
+  // Write back startup reservations.
+
+  for (int i = first_space; i < last_space; i++) {
+    startup_deserializer->allocator()->reservations_[i].swap(
+        merged_reservations[i]);
+  }
+
+  for (int i = first_space; i < kNumberOfPreallocatedSpaces; i++) {
+    startup_deserializer->allocator()->high_water_[i] =
+        startup_deserializer->allocator()->reservations_[i][0].start;
+    builtin_deserializer->allocator()->high_water_[i] = nullptr;
+  }
+
+  return true;
+}
+
+bool DefaultDeserializerAllocator::ReservesOnlyCodeSpace() const {
+  for (int space = NEW_SPACE; space < kNumberOfSpaces; space++) {
+    if (space == CODE_SPACE) continue;
+    const auto& r = reservations_[space];
+    for (const Heap::Chunk& c : r)
+      if (c.size != 0) return false;
+  }
+  return true;
+}
+
+bool DefaultDeserializerAllocator::ReservationsAreFullyUsed() const {
+  for (int space = 0; space < kNumberOfPreallocatedSpaces; space++) {
+    const uint32_t chunk_index = current_chunk_[space];
+    if (reservations_[space].size() != chunk_index + 1) {
+      return false;
+    }
+    if (reservations_[space][chunk_index].end != high_water_[space]) {
+      return false;
+    }
+  }
+  return (allocated_maps_.size() == next_map_index_);
+}
+
+void DefaultDeserializerAllocator::
+    RegisterDeserializedObjectsForBlackAllocation() {
+  isolate()->heap()->RegisterDeserializedObjectsForBlackAllocation(
+      reservations_, deserialized_large_objects_, allocated_maps_);
+}
+
+Isolate* DefaultDeserializerAllocator::isolate() const {
+  return deserializer_->isolate();
+}
+
+}  // namespace internal
+}  // namespace v8

src/snapshot/default-deserializer-allocator.h

+// Copyright 2017 the V8 project authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef V8_SNAPSHOT_DEFAULT_DESERIALIZER_ALLOCATOR_H_
+#define V8_SNAPSHOT_DEFAULT_DESERIALIZER_ALLOCATOR_H_
+
+#include "src/globals.h"
+#include "src/heap/heap.h"
+#include "src/snapshot/serializer-common.h"
+
+namespace v8 {
+namespace internal {
+
+template <class AllocatorT>
+class Deserializer;
+
+class BuiltinDeserializer;
+class StartupDeserializer;
+
+class DefaultDeserializerAllocator final {
+ public:
+  DefaultDeserializerAllocator(
+      Deserializer<DefaultDeserializerAllocator>* deserializer);
+
+  // ------- Allocation Methods -------
+  // Methods related to memory allocation during deserialization.
+
+  Address Allocate(AllocationSpace space, int size);
+
+  void MoveToNextChunk(AllocationSpace space);
+  void SetAlignment(AllocationAlignment alignment) {
+    DCHECK_EQ(kWordAligned, next_alignment_);
+    DCHECK_LE(kWordAligned, alignment);
+    DCHECK_LE(alignment, kDoubleUnaligned);
+    next_alignment_ = static_cast<AllocationAlignment>(alignment);
+  }
+
+  HeapObject* GetMap(uint32_t index);
+  HeapObject* GetLargeObject(uint32_t index);
+  HeapObject* GetObject(AllocationSpace space, uint32_t chunk_index,
+                        uint32_t chunk_offset);
+
+  // ------- Reservation Methods -------
+  // Methods related to memory reservations (prior to deserialization).
+
+  void DecodeReservation(Vector<const SerializedData::Reservation> res);
+  bool ReserveSpace();
+
+  // Atomically reserves space for the two given deserializers. Guarantees
+  // reservation for both without garbage collection in-between.
+  static bool ReserveSpace(StartupDeserializer* startup_deserializer,
+                           BuiltinDeserializer* builtin_deserializer);
+
+  bool ReservesOnlyCodeSpace() const;
+  bool ReservationsAreFullyUsed() const;
+
+  // ------- Misc Utility Methods -------
+
+  void RegisterDeserializedObjectsForBlackAllocation();
+
+  // For SortMapDescriptors();
+  const std::vector<Address>& GetAllocatedMaps() const {
+    return allocated_maps_;
+  }
+
+ private:
+  Isolate* isolate() const;
+
+  // Raw allocation without considering alignment.
+  Address AllocateRaw(AllocationSpace space, int size);
+
+ private:
+  static constexpr int kNumberOfPreallocatedSpaces =
+      SerializerDeserializer::kNumberOfPreallocatedSpaces;
+  static constexpr int kNumberOfSpaces =
+      SerializerDeserializer::kNumberOfSpaces;
+
+  // The address of the next object that will be allocated in each space.
+  // Each space has a number of chunks reserved by the GC, with each chunk
+  // fitting into a page. Deserialized objects are allocated into the
+  // current chunk of the target space by bumping up high water mark.
+  Heap::Reservation reservations_[kNumberOfSpaces];
+  uint32_t current_chunk_[kNumberOfPreallocatedSpaces];
+  Address high_water_[kNumberOfPreallocatedSpaces];
+
+  // The alignment of the next allocation.
+  AllocationAlignment next_alignment_ = kWordAligned;
+
+  // All required maps are pre-allocated during reservation. {next_map_index_}
+  // stores the index of the next map to return from allocation.
+  uint32_t next_map_index_ = 0;
+  std::vector<Address> allocated_maps_;
+
+  // Allocated large objects are kept in this map and may be fetched later as
+  // back-references.
+  std::vector<HeapObject*> deserialized_large_objects_;
+
+  // The current deserializer.
+  Deserializer<DefaultDeserializerAllocator>* const deserializer_;
+
+  DISALLOW_COPY_AND_ASSIGN(DefaultDeserializerAllocator)
+};
+
+}  // namespace internal
+}  // namespace v8
+
+#endif  // V8_SNAPSHOT_DEFAULT_DESERIALIZER_ALLOCATOR_H_

src/snapshot/deserializer.cc

@@ -4,134 +4,16 @@
 
 #include "src/snapshot/deserializer.h"
 
-#include "src/api.h"
 #include "src/assembler-inl.h"
-#include "src/bootstrapper.h"
-#include "src/deoptimizer.h"
-#include "src/external-reference-table.h"
-#include "src/heap/heap-inl.h"
 #include "src/isolate.h"
-#include "src/macro-assembler.h"
-#include "src/objects-inl.h"
-#include "src/snapshot/builtin-deserializer.h"
+#include "src/objects/string.h"
 #include "src/snapshot/natives.h"
-#include "src/snapshot/startup-deserializer.h"
-#include "src/v8.h"
-#include "src/v8threads.h"
 
 namespace v8 {
 namespace internal {
 
-void Deserializer::DecodeReservation(
-    Vector<const SerializedData::Reservation> res) {
-  DCHECK_EQ(0, reservations_[NEW_SPACE].size());
-  STATIC_ASSERT(NEW_SPACE == 0);
-  int current_space = NEW_SPACE;
-  for (auto& r : res) {
-    reservations_[current_space].push_back({r.chunk_size(), nullptr, nullptr});
-    if (r.is_last()) current_space++;
-  }
-  DCHECK_EQ(kNumberOfSpaces, current_space);
-  for (int i = 0; i < kNumberOfPreallocatedSpaces; i++) current_chunk_[i] = 0;
-}
-
-void Deserializer::RegisterDeserializedObjectsForBlackAllocation() {
-  isolate_->heap()->RegisterDeserializedObjectsForBlackAllocation(
-      reservations_, deserialized_large_objects_, allocated_maps_);
-}
-
-bool Deserializer::ReserveSpace() {
-#ifdef DEBUG
-  for (int i = NEW_SPACE; i < kNumberOfSpaces; ++i) {
-    DCHECK_GT(reservations_[i].size(), 0);
-  }
-#endif  // DEBUG
-  DCHECK(allocated_maps_.empty());
-  if (!isolate_->heap()->ReserveSpace(reservations_, &allocated_maps_))
-    return false;
-  for (int i = 0; i < kNumberOfPreallocatedSpaces; i++) {
-    high_water_[i] = reservations_[i][0].start;
-  }
-  return true;
-}
-
-// static
-bool Deserializer::ReserveSpace(StartupDeserializer* startup_deserializer,
-                                BuiltinDeserializer* builtin_deserializer) {
-  const int first_space = NEW_SPACE;
-  const int last_space = SerializerDeserializer::kNumberOfSpaces;
-  Isolate* isolate = startup_deserializer->isolate();
-
-  // Create a set of merged reservations to reserve space in one go.
-  // The BuiltinDeserializer's reservations are ignored, since our actual
-  // requirements vary based on whether lazy deserialization is enabled.
-  // Instead, we manually determine the required code-space.
-
-  DCHECK(builtin_deserializer->ReservesOnlyCodeSpace());
-  Heap::Reservation merged_reservations[kNumberOfSpaces];
-  for (int i = first_space; i < last_space; i++) {
-    merged_reservations[i] = startup_deserializer->reservations_[i];
-  }
-
-  Heap::Reservation builtin_reservations =
-      builtin_deserializer->CreateReservationsForEagerBuiltins();
-  DCHECK(!builtin_reservations.empty());
-
-  for (const auto& c : builtin_reservations) {
-    merged_reservations[CODE_SPACE].push_back(c);
-  }
-
-  if (!isolate->heap()->ReserveSpace(merged_reservations,
-                                     &startup_deserializer->allocated_maps_)) {
-    return false;
-  }
-
-  DisallowHeapAllocation no_allocation;
-
-  // Distribute the successful allocations between both deserializers.
-  // There's nothing to be done here except for code space.
-
-  {
-    const int num_builtin_reservations =
-        static_cast<int>(builtin_reservations.size());
-    for (int i = num_builtin_reservations - 1; i >= 0; i--) {
-      const auto& c = merged_reservations[CODE_SPACE].back();
-      DCHECK_EQ(c.size, builtin_reservations[i].size);
-      DCHECK_EQ(c.size, c.end - c.start);
-      builtin_reservations[i].start = c.start;
-      builtin_reservations[i].end = c.end;
-      merged_reservations[CODE_SPACE].pop_back();
-    }
-
-    builtin_deserializer->InitializeBuiltinsTable(builtin_reservations);
-  }
-
-  // Write back startup reservations.
-
-  for (int i = first_space; i < last_space; i++) {
-    startup_deserializer->reservations_[i].swap(merged_reservations[i]);
-  }
-
-  for (int i = first_space; i < kNumberOfPreallocatedSpaces; i++) {
-    startup_deserializer->high_water_[i] =
-        startup_deserializer->reservations_[i][0].start;
-    builtin_deserializer->high_water_[i] = nullptr;
-  }
-
-  return true;
-}
-
-bool Deserializer::ReservesOnlyCodeSpace() const {
-  for (int space = NEW_SPACE; space < kNumberOfSpaces; space++) {
-    if (space == CODE_SPACE) continue;
-    const auto& r = reservations_[space];
-    for (const Heap::Chunk& c : r)
-      if (c.size != 0) return false;
-  }
-  return true;
-}
-
-void Deserializer::Initialize(Isolate* isolate) {
+template <class AllocatorT>
+void Deserializer<AllocatorT>::Initialize(Isolate* isolate) {
   DCHECK_NULL(isolate_);
   DCHECK_NOT_NULL(isolate);
   isolate_ = isolate;
@@ -150,56 +32,52 @@ void Deserializer::Initialize(Isolate* isolate) {
            SerializedData::ComputeMagicNumber(external_reference_table_));
 }
 
-void Deserializer::SortMapDescriptors() {
-  for (const auto& address : allocated_maps_) {
-    Map* map = Map::cast(HeapObject::FromAddress(address));
-    if (map->instance_descriptors()->number_of_descriptors() > 1) {
-      map->instance_descriptors()->Sort();
-    }
-  }
-}
-
-bool Deserializer::IsLazyDeserializationEnabled() const {
+template <class AllocatorT>
+bool Deserializer<AllocatorT>::IsLazyDeserializationEnabled() const {
   return FLAG_lazy_deserialization && !isolate()->serializer_enabled();
 }
 
-Deserializer::~Deserializer() {
+template <class AllocatorT>
+Deserializer<AllocatorT>::~Deserializer() {
 #ifdef DEBUG
   // Do not perform checks if we aborted deserialization.
   if (source_.position() == 0) return;
   // Check that we only have padding bytes remaining.
   while (source_.HasMore()) DCHECK_EQ(kNop, source_.Get());
-  for (int space = 0; space < kNumberOfPreallocatedSpaces; space++) {
-    int chunk_index = current_chunk_[space];
-    DCHECK_EQ(reservations_[space].size(), chunk_index + 1);
-    DCHECK_EQ(reservations_[space][chunk_index].end, high_water_[space]);
-  }
-  DCHECK_EQ(allocated_maps_.size(), next_map_index_);
+  // Check that we've fully used all reserved space.
+  DCHECK(allocator()->ReservationsAreFullyUsed());
 #endif  // DEBUG
 }
 
 // This is called on the roots.  It is the driver of the deserialization
 // process.  It is also called on the body of each function.
-void Deserializer::VisitRootPointers(Root root, Object** start, Object** end) {
+template <class AllocatorT>
+void Deserializer<AllocatorT>::VisitRootPointers(Root root, Object** start,
+                                                 Object** end) {
   // The space must be new space.  Any other space would cause ReadChunk to try
   // to update the remembered using nullptr as the address.
   ReadData(start, end, NEW_SPACE, nullptr);
 }
 
-void Deserializer::Synchronize(VisitorSynchronization::SyncTag tag) {
+template <class AllocatorT>
+void Deserializer<AllocatorT>::Synchronize(
+    VisitorSynchronization::SyncTag tag) {
   static const byte expected = kSynchronize;
   CHECK_EQ(expected, source_.Get());
   deserializing_builtins_ = (tag == VisitorSynchronization::kHandleScope);
 }
 
-void Deserializer::DeserializeDeferredObjects() {
+template <class AllocatorT>
+void Deserializer<AllocatorT>::DeserializeDeferredObjects() {
   for (int code = source_.Get(); code != kSynchronize; code = source_.Get()) {
     switch (code) {
       case kAlignmentPrefix:
       case kAlignmentPrefix + 1:
-      case kAlignmentPrefix + 2:
-        SetAlignment(code);
+      case kAlignmentPrefix + 2: {
+        int alignment = code - (SerializerDeserializer::kAlignmentPrefix - 1);
+        allocator()->SetAlignment(static_cast<AllocationAlignment>(alignment));
         break;
+      }
       default: {
         int space = code & kSpaceMask;
         DCHECK_LE(space, kNumberOfSpaces);
@@ -241,7 +119,9 @@ uint32_t StringTableInsertionKey::ComputeHashField(String* string) {
   return string->hash_field();
 }
 
-HeapObject* Deserializer::PostProcessNewObject(HeapObject* obj, int space) {
+template <class AllocatorT>
+HeapObject* Deserializer<AllocatorT>::PostProcessNewObject(HeapObject* obj,
+                                                           int space) {
   if (deserializing_user_code()) {
     if (obj->IsString()) {
       String* string = String::cast(obj);
@@ -334,7 +214,8 @@ HeapObject* Deserializer::PostProcessNewObject(HeapObject* obj, int space) {
   return obj;
 }
 
-int Deserializer::MaybeReplaceWithDeserializeLazy(int builtin_id) {
+template <class AllocatorT>
+int Deserializer<AllocatorT>::MaybeReplaceWithDeserializeLazy(int builtin_id) {
   DCHECK(Builtins::IsBuiltinId(builtin_id));
   return (IsLazyDeserializationEnabled() && Builtins::IsLazy(builtin_id) &&
           !deserializing_builtins_)
@@ -342,66 +223,56 @@ int Deserializer::MaybeReplaceWithDeserializeLazy(int builtin_id) {
              : builtin_id;
 }
 
-HeapObject* Deserializer::GetBackReferencedObject(int space) {
+template <class AllocatorT>
+HeapObject* Deserializer<AllocatorT>::GetBackReferencedObject(int space) {
   HeapObject* obj;
   SerializerReference back_reference =
       SerializerReference::FromBitfield(source_.GetInt());
-  if (space == LO_SPACE) {
-    uint32_t index = back_reference.large_object_index();
-    obj = deserialized_large_objects_[index];
-  } else if (space == MAP_SPACE) {
-    int index = back_reference.map_index();
-    DCHECK(index < next_map_index_);
-    obj = HeapObject::FromAddress(allocated_maps_[index]);
-  } else {
-    DCHECK_LT(space, kNumberOfPreallocatedSpaces);
-    uint32_t chunk_index = back_reference.chunk_index();
-    DCHECK_LE(chunk_index, current_chunk_[space]);
-    uint32_t chunk_offset = back_reference.chunk_offset();
-    Address address = reservations_[space][chunk_index].start + chunk_offset;
-    if (next_alignment_ != kWordAligned) {
-      int padding = Heap::GetFillToAlign(address, next_alignment_);
-      next_alignment_ = kWordAligned;
-      DCHECK(padding == 0 || HeapObject::FromAddress(address)->IsFiller());
-      address += padding;
-    }
-    obj = HeapObject::FromAddress(address);
+
+  switch (space) {
+    case LO_SPACE:
+      obj = allocator()->GetLargeObject(back_reference.large_object_index());
+      break;
+    case MAP_SPACE:
+      obj = allocator()->GetMap(back_reference.map_index());
+      break;
+    default:
+      obj = allocator()->GetObject(static_cast<AllocationSpace>(space),
+                                   back_reference.chunk_index(),
+                                   back_reference.chunk_offset());
+      break;
   }
+
   if (deserializing_user_code() && obj->IsInternalizedString()) {
     obj = String::cast(obj)->GetForwardedInternalizedString();
   }
+
   hot_objects_.Add(obj);
   return obj;
 }
 
+template <class AllocatorT>
+void Deserializer<AllocatorT>::SortMapDescriptors() {
+  for (const auto& address : allocator()->GetAllocatedMaps()) {
+    Map* map = Map::cast(HeapObject::FromAddress(address));
+    if (map->instance_descriptors()->number_of_descriptors() > 1) {
+      map->instance_descriptors()->Sort();
+    }
+  }
+}
+
 // This routine writes the new object into the pointer provided and then
 // returns true if the new object was in young space and false otherwise.
 // The reason for this strange interface is that otherwise the object is
 // written very late, which means the FreeSpace map is not set up by the
 // time we need to use it to mark the space at the end of a page free.
-void Deserializer::ReadObject(int space_number, Object** write_back) {
-  Address address;
-  HeapObject* obj;
-  int size = source_.GetInt() << kObjectAlignmentBits;
-
-  if (next_alignment_ != kWordAligned) {
-    int reserved = size + Heap::GetMaximumFillToAlign(next_alignment_);
-    address = Allocate(space_number, reserved);
-    obj = HeapObject::FromAddress(address);
-    // If one of the following assertions fails, then we are deserializing an
-    // aligned object when the filler maps have not been deserialized yet.
-    // We require filler maps as padding to align the object.
-    Heap* heap = isolate_->heap();
-    DCHECK(heap->free_space_map()->IsMap());
-    DCHECK(heap->one_pointer_filler_map()->IsMap());
-    DCHECK(heap->two_pointer_filler_map()->IsMap());
-    obj = heap->AlignWithFiller(obj, size, reserved, next_alignment_);
-    address = obj->address();
-    next_alignment_ = kWordAligned;
-  } else {
-    address = Allocate(space_number, size);
-    obj = HeapObject::FromAddress(address);
-  }
+template <class AllocatorT>
+void Deserializer<AllocatorT>::ReadObject(int space_number,
+                                          Object** write_back) {
+  const int size = source_.GetInt() << kObjectAlignmentBits;
+
+  Address address = Allocate(space_number, size);
+  HeapObject* obj = HeapObject::FromAddress(address);
 
   isolate_->heap()->OnAllocationEvent(obj, size);
   Object** current = reinterpret_cast<Object**>(address);
@@ -423,46 +294,15 @@ void Deserializer::ReadObject(int space_number, Object** write_back) {
 #endif  // DEBUG
 }
 
-// We know the space requirements before deserialization and can
-// pre-allocate that reserved space. During deserialization, all we need
-// to do is to bump up the pointer for each space in the reserved
-// space. This is also used for fixing back references.
-// We may have to split up the pre-allocation into several chunks
-// because it would not fit onto a single page. We do not have to keep
-// track of when to move to the next chunk. An opcode will signal this.
-// Since multiple large objects cannot be folded into one large object
-// space allocation, we have to do an actual allocation when deserializing
-// each large object. Instead of tracking offset for back references, we
-// reference large objects by index.
-Address Deserializer::Allocate(int space_index, int size) {
-  if (space_index == LO_SPACE) {
-    AlwaysAllocateScope scope(isolate_);
-    LargeObjectSpace* lo_space = isolate_->heap()->lo_space();
-    Executability exec = static_cast<Executability>(source_.Get());
-    AllocationResult result = lo_space->AllocateRaw(size, exec);
-    HeapObject* obj = result.ToObjectChecked();
-    deserialized_large_objects_.push_back(obj);
-    return obj->address();
-  } else if (space_index == MAP_SPACE) {
-    DCHECK_EQ(Map::kSize, size);
-    return allocated_maps_[next_map_index_++];
-  } else {
-    DCHECK_LT(space_index, kNumberOfPreallocatedSpaces);
-    Address address = high_water_[space_index];
-    DCHECK_NOT_NULL(address);
-    high_water_[space_index] += size;
-#ifdef DEBUG
-    // Assert that the current reserved chunk is still big enough.
-    const Heap::Reservation& reservation = reservations_[space_index];
-    int chunk_index = current_chunk_[space_index];
-    DCHECK_LE(high_water_[space_index], reservation[chunk_index].end);
-#endif
-    if (space_index == CODE_SPACE) SkipList::Update(address, size);
-    return address;
-  }
+template <class AllocatorT>
+Address Deserializer<AllocatorT>::Allocate(int space_index, int size) {
+  // TODO(jgruber): Remove this indirection once we have a
+  // BuiltinDeserializerAllocator.
+  return allocator()->Allocate(static_cast<AllocationSpace>(space_index), size);
 }
 
-Object* Deserializer::ReadDataSingle() {
+template <class AllocatorT>
+Object* Deserializer<AllocatorT>::ReadDataSingle() {
   Object* o;
   Object** start = &o;
   Object** end = start + 1;
@@ -474,8 +314,10 @@ Object* Deserializer::ReadDataSingle() {
   return o;
 }
 
-bool Deserializer::ReadData(Object** current, Object** limit, int source_space,
-                            Address current_object_address) {
+template <class AllocatorT>
+bool Deserializer<AllocatorT>::ReadData(Object** current, Object** limit,
+                                        int source_space,
+                                        Address current_object_address) {
   Isolate* const isolate = isolate_;
   // Write barrier support costs around 1% in startup time.  In fact there
   // are no new space objects in current boot snapshots, so it's not needed,
@@ -618,15 +460,7 @@ bool Deserializer::ReadData(Object** current, Object** limit, int source_space,
 
       case kNextChunk: {
         int space = source_.Get();
-        DCHECK_LT(space, kNumberOfPreallocatedSpaces);
-        int chunk_index = current_chunk_[space];
-        const Heap::Reservation& reservation = reservations_[space];
-        // Make sure the current chunk is indeed exhausted.
-        CHECK_EQ(reservation[chunk_index].end, high_water_[space]);
-        // Move to next reserved chunk.
-        chunk_index = ++current_chunk_[space];
-        CHECK_LT(chunk_index, reservation.size());
-        high_water_[space] = reservation[chunk_index].start;
+        allocator()->MoveToNextChunk(static_cast<AllocationSpace>(space));
         break;
       }
 
@@ -702,9 +536,11 @@ bool Deserializer::ReadData(Object** current, Object** limit, int source_space,
 
       case kAlignmentPrefix:
       case kAlignmentPrefix + 1:
-      case kAlignmentPrefix + 2:
-        SetAlignment(data);
+      case kAlignmentPrefix + 2: {
+        int alignment = data - (SerializerDeserializer::kAlignmentPrefix - 1);
+        allocator()->SetAlignment(static_cast<AllocationAlignment>(alignment));
         break;
+      }
 
       STATIC_ASSERT(kNumberOfRootArrayConstants == Heap::kOldSpaceRoots);
       STATIC_ASSERT(kNumberOfRootArrayConstants == 32);
@@ -783,10 +619,13 @@ bool Deserializer::ReadData(Object** current, Object** limit, int source_space,
   return true;
 }
 
+template <class AllocatorT>
 template <int where, int how, int within, int space_number_if_any>
-Object** Deserializer::ReadDataCase(Isolate* isolate, Object** current,
-                                    Address current_object_address, byte data,
-                                    bool write_barrier_needed) {
+Object** Deserializer<AllocatorT>::ReadDataCase(Isolate* isolate,
+                                                Object** current,
+                                                Address current_object_address,
+                                                byte data,
+                                                bool write_barrier_needed) {
   bool emit_write_barrier = false;
   bool current_was_incremented = false;
   int space_number = space_number_if_any == kAnyOldSpace ? (data & kSpaceMask)
@@ -877,5 +716,8 @@ Object** Deserializer::ReadDataCase(Isolate* isolate, Object** current,
   return current;
 }
 
+// Explicit instantiation.
+template class Deserializer<DefaultDeserializerAllocator>;
+
 }  // namespace internal
 }  // namespace v8

src/snapshot/deserializer.h

@@ -7,14 +7,16 @@
 
 #include <vector>
 
-#include "src/heap/heap.h"
-#include "src/objects.h"
+#include "src/snapshot/default-deserializer-allocator.h"
 #include "src/snapshot/serializer-common.h"
 #include "src/snapshot/snapshot-source-sink.h"
 
 namespace v8 {
 namespace internal {
 
+class HeapObject;
+class Object;
+
 // Used for platforms with embedded constant pools to trigger deserialization
 // of objects found in code.
 #if defined(V8_TARGET_ARCH_MIPS) || defined(V8_TARGET_ARCH_MIPS64) || \
@@ -25,21 +27,12 @@ namespace internal {
 #define V8_CODE_EMBEDS_OBJECT_POINTER 0
 #endif
 
-class BuiltinDeserializer;
-class Heap;
-class StartupDeserializer;
-
 // A Deserializer reads a snapshot and reconstructs the Object graph it defines.
+template <class AllocatorT = DefaultDeserializerAllocator>
 class Deserializer : public SerializerDeserializer {
  public:
   ~Deserializer() override;
 
-  // Add an object to back an attached reference. The order to add objects must
-  // mirror the order they are added in the serializer.
-  void AddAttachedObject(Handle<HeapObject> attached_object) {
-    attached_objects_.push_back(attached_object);
-  }
-
   void SetRehashability(bool v) { can_rehash_ = v; }
 
  protected:
@@ -49,29 +42,18 @@ class Deserializer : public SerializerDeserializer {
       : isolate_(nullptr),
         source_(data->Payload()),
         magic_number_(data->GetMagicNumber()),
-        next_map_index_(0),
         external_reference_table_(nullptr),
-        deserialized_large_objects_(0),
+        allocator_(this),
         deserializing_user_code_(deserializing_user_code),
-        next_alignment_(kWordAligned),
         can_rehash_(false) {
-    DecodeReservation(data->Reservations());
-    // We start the indicies here at 1, so that we can distinguish between an
+    allocator()->DecodeReservation(data->Reservations());
+    // We start the indices here at 1, so that we can distinguish between an
     // actual index and a nullptr in a deserialized object requiring fix-up.
     off_heap_backing_stores_.push_back(nullptr);
   }
 
-  bool ReserveSpace();
-
-  // Atomically reserves space for the two given deserializers. Guarantees
-  // reservation for both without garbage collection in-between.
-  static bool ReserveSpace(StartupDeserializer* startup_deserializer,
-                           BuiltinDeserializer* builtin_deserializer);
-  bool ReservesOnlyCodeSpace() const;
-
   void Initialize(Isolate* isolate);
   void DeserializeDeferredObjects();
-  void RegisterDeserializedObjectsForBlackAllocation();
 
   virtual Address Allocate(int space_index, int size);
 
@@ -82,6 +64,12 @@ class Deserializer : public SerializerDeserializer {
   // snapshot by chunk index and offset.
   HeapObject* GetBackReferencedObject(int space);
 
+  // Add an object to back an attached reference. The order to add objects must
+  // mirror the order they are added in the serializer.
+  void AddAttachedObject(Handle<HeapObject> attached_object) {
+    attached_objects_.push_back(attached_object);
+  }
+
   // Sort descriptors of deserialized maps using new string hashes.
   void SortMapDescriptors();
 
@@ -102,6 +90,8 @@ class Deserializer : public SerializerDeserializer {
   const std::vector<TransitionArray*>& transition_arrays() const {
     return transition_arrays_;
   }
+
+  AllocatorT* allocator() { return &allocator_; }
   bool deserializing_user_code() const { return deserializing_user_code_; }
   bool can_rehash() const { return can_rehash_; }
 
@@ -112,20 +102,10 @@ class Deserializer : public SerializerDeserializer {
 
   void Synchronize(VisitorSynchronization::SyncTag tag) override;
 
-  void DecodeReservation(Vector<const SerializedData::Reservation> res);
-
   void UnalignedCopy(Object** dest, Object** src) {
     memcpy(dest, src, sizeof(*src));
   }
 
-  void SetAlignment(byte data) {
-    DCHECK_EQ(kWordAligned, next_alignment_);
-    int alignment = data - (kAlignmentPrefix - 1);
-    DCHECK_LE(kWordAligned, alignment);
-    DCHECK_LE(alignment, kDoubleUnaligned);
-    next_alignment_ = static_cast<AllocationAlignment>(alignment);
-  }
-
   // Fills in some heap data in an area from start to end (non-inclusive).  The
   // space id is used for the write barrier.  The object_address is the address
   // of the object we are writing into, or nullptr if we are not writing into an
@@ -159,19 +139,8 @@ class Deserializer : public SerializerDeserializer {
   SnapshotByteSource source_;
   uint32_t magic_number_;
 
-  // The address of the next object that will be allocated in each space.
-  // Each space has a number of chunks reserved by the GC, with each chunk
-  // fitting into a page. Deserialized objects are allocated into the
-  // current chunk of the target space by bumping up high water mark.
-  Heap::Reservation reservations_[kNumberOfSpaces];
-  uint32_t current_chunk_[kNumberOfPreallocatedSpaces];
-  Address high_water_[kNumberOfPreallocatedSpaces];
-  int next_map_index_;
-  std::vector<Address> allocated_maps_;
-
   ExternalReferenceTable* external_reference_table_;
 
-  std::vector<HeapObject*> deserialized_large_objects_;
   std::vector<Code*> new_code_objects_;
   std::vector<AccessorInfo*> accessor_infos_;
   std::vector<Handle<String>> new_internalized_strings_;
@@ -179,14 +148,13 @@ class Deserializer : public SerializerDeserializer {
   std::vector<TransitionArray*> transition_arrays_;
   std::vector<byte*> off_heap_backing_stores_;
 
+  AllocatorT allocator_;
   const bool deserializing_user_code_;
 
   // TODO(jgruber): This workaround will no longer be necessary once builtin
   // reference patching has been removed (through advance allocation).
   bool deserializing_builtins_ = false;
 
-  AllocationAlignment next_alignment_;
-
   // TODO(6593): generalize rehashing, and remove this flag.
   bool can_rehash_;
 
@@ -194,6 +162,9 @@ class Deserializer : public SerializerDeserializer {
   uint32_t num_api_references_;
 #endif  // DEBUG
 
+  // For source(), isolate(), and allocator().
+  friend class DefaultDeserializerAllocator;
+
   DISALLOW_COPY_AND_ASSIGN(Deserializer);
 };
 

src/snapshot/object-deserializer.cc

@@ -67,7 +67,7 @@ ObjectDeserializer::DeserializeWasmCompiledModule(
 
 MaybeHandle<HeapObject> ObjectDeserializer::Deserialize(Isolate* isolate) {
   Initialize(isolate);
-  if (!ReserveSpace()) return MaybeHandle<HeapObject>();
+  if (!allocator()->ReserveSpace()) return MaybeHandle<HeapObject>();
 
   DCHECK(deserializing_user_code());
   HandleScope scope(isolate);
@@ -79,7 +79,7 @@ MaybeHandle<HeapObject> ObjectDeserializer::Deserialize(Isolate* isolate) {
     DeserializeDeferredObjects();
     FlushICacheForNewCodeObjectsAndRecordEmbeddedObjects();
     result = Handle<HeapObject>(HeapObject::cast(root));
-    RegisterDeserializedObjectsForBlackAllocation();
+    allocator()->RegisterDeserializedObjectsForBlackAllocation();
   }
   CommitPostProcessedObjects();
   return scope.CloseAndEscape(result);

src/snapshot/object-deserializer.h

@@ -15,7 +15,7 @@ class SharedFunctionInfo;
 class WasmCompiledModule;
 
 // Deserializes the object graph rooted at a given object.
-class ObjectDeserializer final : public Deserializer {
+class ObjectDeserializer final : public Deserializer<> {
  public:
   static MaybeHandle<SharedFunctionInfo> DeserializeSharedFunctionInfo(
       Isolate* isolate, const SerializedCodeData* data, Handle<String> source);

src/snapshot/partial-deserializer.cc

@@ -30,7 +30,9 @@ MaybeHandle<Object> PartialDeserializer::Deserialize(
     Isolate* isolate, Handle<JSGlobalProxy> global_proxy,
     v8::DeserializeEmbedderFieldsCallback embedder_fields_deserializer) {
   Initialize(isolate);
-  if (!ReserveSpace()) V8::FatalProcessOutOfMemory("PartialDeserializer");
+  if (!allocator()->ReserveSpace()) {
+    V8::FatalProcessOutOfMemory("PartialDeserializer");
+  }
 
   AddAttachedObject(global_proxy);
 
@@ -44,7 +46,7 @@ MaybeHandle<Object> PartialDeserializer::Deserialize(
   DeserializeDeferredObjects();
   DeserializeEmbedderFields(embedder_fields_deserializer);
 
-  RegisterDeserializedObjectsForBlackAllocation();
+  allocator()->RegisterDeserializedObjectsForBlackAllocation();
 
   // There's no code deserialized here. If this assert fires then that's
   // changed and logging should be added to notify the profiler et al of the

src/snapshot/partial-deserializer.h

@@ -15,7 +15,7 @@ class Context;
 
 // Deserializes the context-dependent object graph rooted at a given object.
 // The PartialDeserializer is not expected to deserialize any code objects.
-class PartialDeserializer final : public Deserializer {
+class PartialDeserializer final : public Deserializer<> {
  public:
   static MaybeHandle<Context> DeserializeContext(
       Isolate* isolate, const SnapshotData* data, bool can_rehash,

src/snapshot/startup-deserializer.cc

@@ -18,7 +18,8 @@ void StartupDeserializer::DeserializeInto(Isolate* isolate) {
 
   BuiltinDeserializer builtin_deserializer(isolate, builtin_data_);
 
-  if (!Deserializer::ReserveSpace(this, &builtin_deserializer)) {
+  if (!DefaultDeserializerAllocator::ReserveSpace(this,
+                                                  &builtin_deserializer)) {
     V8::FatalProcessOutOfMemory("StartupDeserializer");
   }
 

src/snapshot/startup-deserializer.h

@@ -12,7 +12,7 @@ namespace v8 {
 namespace internal {
 
 // Initializes an isolate with context-independent data from a given snapshot.
-class StartupDeserializer final : public Deserializer {
+class StartupDeserializer final : public Deserializer<> {
  public:
   StartupDeserializer(const SnapshotData* startup_data,
                       const BuiltinSnapshotData* builtin_data)

src/v8.gyp

@@ -1345,6 +1345,8 @@
         'snapshot/builtin-serializer.h',
         'snapshot/code-serializer.cc',
         'snapshot/code-serializer.h',
+        'snapshot/default-deserializer-allocator.cc',
+        'snapshot/default-deserializer-allocator.h',
         'snapshot/default-serializer-allocator.cc',
         'snapshot/default-serializer-allocator.h',
         'snapshot/deserializer.cc',