[snapshot][ptr-compr][cleanup] Write repeats as a prefix

... rather than as a suffix.

This avoids reading previous values which in turn will make things
simpler from the pointer compression point of view in a sense that
this is a step towards replacing UnalignedSlot with MaybeObjectSlot.

Bug: v8:8794, v8:8562
Change-Id: I9a9b4a01f73b8058074d337b7e9e9f75fa1c9de0
Reviewed-on: https://chromium-review.googlesource.com/c/1456037
Commit-Queue: Igor Sheludko <ishell@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#59402}

src/snapshot/deserializer.cc

@@ -50,12 +50,6 @@ class UnalignedSlot {
     memcpy(&result, reinterpret_cast<void*>(ptr_), sizeof(result));
     return MaybeObject(result);
   }
-  MaybeObject ReadPrevious() {
-    Address result;
-    memcpy(&result, reinterpret_cast<void*>(ptr_ - kPointerSize),
-           sizeof(result));
-    return MaybeObject(result);
-  }
   inline void Write(Address value) {
     memcpy(reinterpret_cast<void*>(ptr_), &value, sizeof(value));
   }
@@ -115,8 +109,8 @@ Deserializer::~Deserializer() {
 // process.  It is also called on the body of each function.
 void Deserializer::VisitRootPointers(Root root, const char* description,
                                      FullObjectSlot start, FullObjectSlot end) {
-  // The space must be new space.  Any other space would cause ReadChunk to try
-  // to update the remembered using nullptr as the address.
+  // We are reading to a location outside of JS heap, so pass NEW_SPACE to
+  // avoid triggering write barriers.
   // TODO(ishell): this will not work once we actually compress pointers.
   STATIC_ASSERT(kTaggedSize == kSystemPointerSize);
   ReadData(UnalignedSlot(start.address()), UnalignedSlot(end.address()),
@@ -413,6 +407,26 @@ void Deserializer::ReadObject(int space_number, UnalignedSlot write_back,
 #endif  // DEBUG
 }
 
+UnalignedSlot Deserializer::ReadRepeatedObject(UnalignedSlot current,
+                                               int repeat_count) {
+  CHECK_LE(2, repeat_count);
+  MaybeObject object;
+  // We are reading to a location outside of JS heap, so pass NEW_SPACE to
+  // avoid triggering write barriers.
+  bool filled = ReadData(UnalignedSlot(&object), UnalignedSlot(&object + 1),
+                         NEW_SPACE, kNullAddress);
+  CHECK(filled);
+  DCHECK(HAS_HEAP_OBJECT_TAG(object.ptr()));
+  DCHECK(!Heap::InYoungGeneration(object));
+  for (int i = 0; i < repeat_count; i++) {
+    // Repeated values are not subject to the write barrier so we don't need
+    // to trigger it.
+    UnalignedCopy(current, object);
+    current.Advance();
+  }
+  return current;
+}
+
 static void NoExternalReferencesCallback() {
   // The following check will trigger if a function or object template
   // with references to native functions have been deserialized from
@@ -646,13 +660,8 @@ bool Deserializer::ReadData(UnalignedSlot current, UnalignedSlot limit,
       }
 
       case kVariableRepeat: {
-        int repeats = source_.GetInt();
-        MaybeObject object = current.ReadPrevious();
-        DCHECK(!Heap::InYoungGeneration(object));
-        for (int i = 0; i < repeats; i++) {
-          UnalignedCopy(current, object);
-          current.Advance();
-        }
+        int repeats = DecodeVariableRepeatCount(source_.GetInt());
+        current = ReadRepeatedObject(current, repeats);
         break;
       }
 
@@ -769,13 +778,8 @@ bool Deserializer::ReadData(UnalignedSlot current, UnalignedSlot limit,
 
       STATIC_ASSERT(kNumberOfFixedRepeat == 16);
       SIXTEEN_CASES(kFixedRepeat) {
-        int repeats = data - kFixedRepeatStart;
-        MaybeObject object = current.ReadPrevious();
-        DCHECK(!Heap::InYoungGeneration(object));
-        for (int i = 0; i < repeats; i++) {
-          UnalignedCopy(current, object);
-          current.Advance();
-        }
+        int repeats = DecodeFixedRepeatCount(data);
+        current = ReadRepeatedObject(current, repeats);
         break;
       }
 

src/snapshot/deserializer.h

@@ -139,6 +139,8 @@ class Deserializer : public SerializerDeserializer {
   void ReadObject(int space_number, UnalignedSlot write_back,
                   HeapObjectReferenceType reference_type);
 
+  UnalignedSlot ReadRepeatedObject(UnalignedSlot current, int repeat_count);
+
   // Special handling for serialized code like hooking up internalized strings.
   HeapObject PostProcessNewObject(HeapObject obj, int space);
 

src/snapshot/serializer-common.h

@@ -262,7 +262,6 @@ class SerializerDeserializer : public RootVisitor {
   static const int kNumberOfFixedRepeat = 0x10;
   // 0xe0..0xef
   static const int kFixedRepeat = 0xe0;
-  static const int kFixedRepeatStart = kFixedRepeat - 1;
 
   // 8 hot (recently seen or back-referenced) objects with optional skip.
   static const int kNumberOfHotObjects = 8;
@@ -279,6 +278,39 @@ class SerializerDeserializer : public RootVisitor {
   // Sentinel after a new object to indicate that double alignment is needed.
   static const int kDoubleAlignmentSentinel = 0;
 
+  // Repeat count encoding helpers.
+  static const int kFirstEncodableRepeatCount = 2;
+  static const int kLastEncodableFixedRepeatCount =
+      kFirstEncodableRepeatCount + kNumberOfFixedRepeat - 1;
+  static const int kFirstEncodableVariableRepeatCount =
+      kLastEncodableFixedRepeatCount + 1;
+
+  // Encodes repeat count into a fixed repeat bytecode.
+  static int EncodeFixedRepeat(int repeat_count) {
+    DCHECK(IsInRange(repeat_count, kFirstEncodableRepeatCount,
+                     kLastEncodableFixedRepeatCount));
+    return kFixedRepeat + repeat_count - kFirstEncodableRepeatCount;
+  }
+
+  // Decodes repeat count from a fixed repeat bytecode.
+  static int DecodeFixedRepeatCount(int bytecode) {
+    DCHECK(
+        IsInRange(bytecode, kFixedRepeat, kFixedRepeat + kNumberOfFixedRepeat));
+    return bytecode - kFixedRepeat + kFirstEncodableRepeatCount;
+  }
+
+  // Encodes repeat count into a serialized variable repeat count value.
+  static int EncodeVariableRepeatCount(int repeat_count) {
+    DCHECK_LE(kFirstEncodableVariableRepeatCount, repeat_count);
+    return repeat_count - kFirstEncodableVariableRepeatCount;
+  }
+
+  // Decodes repeat count from a serialized variable repeat count value.
+  static int DecodeVariableRepeatCount(int value) {
+    DCHECK_LE(0, value);
+    return value + kFirstEncodableVariableRepeatCount;
+  }
+
   // ---------- member variable ----------
   HotObjectsList hot_objects_;
 };

src/snapshot/serializer.cc

@@ -289,6 +289,15 @@ void Serializer::PutNextChunk(int space) {
   sink_.Put(space, "NextChunkSpace");
 }
 
+void Serializer::PutRepeat(int repeat_count) {
+  if (repeat_count <= kLastEncodableFixedRepeatCount) {
+    sink_.Put(EncodeFixedRepeat(repeat_count), "FixedRepeat");
+  } else {
+    sink_.Put(kVariableRepeat, "VariableRepeat");
+    sink_.PutInt(EncodeVariableRepeatCount(repeat_count), "repeat count");
+  }
+}
+
 void Serializer::Pad(int padding_offset) {
   // The non-branching GetInt will read up to 3 bytes too far, so we need
   // to pad the snapshot to make sure we don't read over the end.
@@ -693,37 +702,33 @@ void Serializer::ObjectSerializer::VisitPointers(HeapObject host,
     while (current < end &&
            (*current)->GetHeapObject(&current_contents, &reference_type)) {
       RootIndex root_index;
+      // Compute repeat count and write repeat prefix if applicable.
       // Repeats are not subject to the write barrier so we can only use
       // immortal immovable root members. They are never in new space.
-      if (current != start &&
+      MaybeObjectSlot repeat_end = current + 1;
+      if (repeat_end < end &&
           serializer_->root_index_map()->Lookup(current_contents,
                                                 &root_index) &&
           RootsTable::IsImmortalImmovable(root_index) &&
-          *current == *(current - 1)) {
+          *current == *repeat_end) {
         DCHECK_EQ(reference_type, HeapObjectReferenceType::STRONG);
         DCHECK(!Heap::InYoungGeneration(current_contents));
-        int repeat_count = 1;
-        while (current + repeat_count < end - 1 &&
-               *(current + repeat_count) == *current) {
-          repeat_count++;
+        while (repeat_end < end && *repeat_end == *current) {
+          repeat_end++;
         }
-        current += repeat_count;
+        int repeat_count = static_cast<int>(repeat_end - current);
+        current = repeat_end;
         bytes_processed_so_far_ += repeat_count * kTaggedSize;
-        if (repeat_count > kNumberOfFixedRepeat) {
-          sink_->Put(kVariableRepeat, "VariableRepeat");
-          sink_->PutInt(repeat_count, "repeat count");
-        } else {
-          sink_->Put(kFixedRepeatStart + repeat_count, "FixedRepeat");
-        }
+        serializer_->PutRepeat(repeat_count);
       } else {
-        if (reference_type == HeapObjectReferenceType::WEAK) {
-          sink_->Put(kWeakPrefix, "WeakReference");
-        }
-        serializer_->SerializeObject(current_contents, kPlain, kStartOfObject,
-                                     0);
         bytes_processed_so_far_ += kTaggedSize;
         ++current;
       }
+      // Now write the object itself.
+      if (reference_type == HeapObjectReferenceType::WEAK) {
+        sink_->Put(kWeakPrefix, "WeakReference");
+      }
+      serializer_->SerializeObject(current_contents, kPlain, kStartOfObject, 0);
     }
   }
 }

src/snapshot/serializer.h

@@ -209,6 +209,7 @@ class Serializer : public SerializerDeserializer {
   // Emit alignment prefix if necessary, return required padding space in bytes.
   int PutAlignmentPrefix(HeapObject object);
   void PutNextChunk(int space);
+  void PutRepeat(int repeat_count);
 
   // Returns true if the object was successfully serialized as a root.
   bool SerializeRoot(HeapObject obj, HowToCode how_to_code,