Internalize strings in-place

src/objects.cc

 // Copyright 2015 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/objects.h"
 
 #include <cmath>
 #include <iomanip>
 #include <memory>
 #include <sstream>
@@ skipping 2313 matching lines @@
                   resource->data(),
                   resource->length() * sizeof(smart_chars[0])) == 0);
   }
 #endif  // DEBUG
   int size = this->Size();  // Byte size of the original string.
   // Abort if size does not allow in-place conversion.
   if (size < ExternalString::kShortSize) return false;
   Heap* heap = GetHeap();
   bool is_one_byte = this->IsOneByteRepresentation();
   bool is_internalized = this->IsInternalizedString();
-  bool has_pointers = this->IsConsString() || this->IsSlicedString();
+  bool has_pointers =
+      this->IsConsString() || this->IsSlicedString() || this->IsThinString();

[Igor Sheludko, 2016/12/20 23:46:58]

Suggestion for follow-up CL: 
bool IsIndirectString() {
  return (map()->instance_type() & kIsIndirectStringMask) ==
kIsIndirectStringTag;
}

[Jakob Kummerow, 2017/01/04 12:45:06]

Done (StringShape supports this already).

 
   // Morph the string to an external string by replacing the map and
   // reinitializing the fields.  This won't work if the space the existing
   // string occupies is too small for a regular  external string.
   // Instead, we resort to a short external string instead, omitting
   // the field caching the address of the backing store.  When we encounter
   // short external strings in generated code, we need to bailout to runtime.
   Map* new_map;
   if (size < ExternalString::kSize) {
     new_map = is_internalized
@@ skipping 51 matching lines @@
     DCHECK(memcmp(smart_chars.start(),
                   resource->data(),
                   resource->length() * sizeof(smart_chars[0])) == 0);
   }
 #endif  // DEBUG
   int size = this->Size();  // Byte size of the original string.
   // Abort if size does not allow in-place conversion.
   if (size < ExternalString::kShortSize) return false;
   Heap* heap = GetHeap();
   bool is_internalized = this->IsInternalizedString();
-  bool has_pointers = this->IsConsString() || this->IsSlicedString();
+  bool has_pointers =
+      this->IsConsString() || this->IsSlicedString() || this->IsThinString();

[Igor Sheludko, 2016/12/20 23:46:58]

Same here.

[Jakob Kummerow, 2017/01/04 12:45:06]

Done.

 
   // Morph the string to an external string by replacing the map and
   // reinitializing the fields.  This won't work if the space the existing
   // string occupies is too small for a regular  external string.
   // Instead, we resort to a short external string instead, omitting
   // the field caching the address of the backing store.  When we encounter
   // short external strings in generated code, we need to bailout to runtime.
   Map* new_map;
   if (size < ExternalString::kSize) {
     new_map = is_internalized
@@ skipping 8344 matching lines @@
     while (
         right > left &&
         unicode_cache->IsWhiteSpaceOrLineTerminator(string->Get(right - 1))) {
       right--;
     }
   }
 
   return isolate->factory()->NewSubString(string, left, right);
 }
 
-bool String::LooksValid() {
-  if (!GetIsolate()->heap()->Contains(this)) return false;
-  return true;
-}
-
+bool String::LooksValid() { return GetIsolate()->heap()->Contains(this); }
 
 // static
 MaybeHandle<String> Name::ToFunctionName(Handle<Name> name) {
   if (name->IsString()) return Handle<String>::cast(name);
   // ES6 section 9.2.11 SetFunctionName, step 4.
   Isolate* const isolate = name->GetIsolate();
   Handle<Object> description(Handle<Symbol>::cast(name)->name(), isolate);
   if (description->IsUndefined(isolate)) {
     return isolate->factory()->empty_string();
   }
@@ skipping 113 matching lines @@
   StringShape shape(this);
   String* string = this;
   int offset = 0;
   if (shape.representation_tag() == kConsStringTag) {
     ConsString* cons = ConsString::cast(string);
     if (cons->second()->length() != 0) {
       return FlatContent();
     }
     string = cons->first();
     shape = StringShape(string);
+  } else if (shape.representation_tag() == kThinStringTag) {
+    ThinString* thin = ThinString::cast(string);
+    string = thin->actual();
+    shape = StringShape(string);
   }
   if (shape.representation_tag() == kSlicedStringTag) {
     SlicedString* slice = SlicedString::cast(string);
     offset = slice->offset();
     string = slice->parent();
     shape = StringShape(string);
     DCHECK(shape.representation_tag() != kConsStringTag &&
            shape.representation_tag() != kSlicedStringTag);
   }
   if (shape.encoding_tag() == kOneByteStringTag) {
@@ skipping 74 matching lines @@
     case kSeqStringTag:
       return SeqTwoByteString::cast(this)->SeqTwoByteStringGetData(start);
     case kExternalStringTag:
       return ExternalTwoByteString::cast(this)->
         ExternalTwoByteStringGetData(start);
     case kSlicedStringTag: {
       SlicedString* slice = SlicedString::cast(this);
       return slice->parent()->GetTwoByteData(start + slice->offset());
     }
     case kConsStringTag:
+    case kThinStringTag:
       UNREACHABLE();
       return NULL;
   }
   UNREACHABLE();
   return NULL;
 }
 
 
 const uc16* SeqTwoByteString::SeqTwoByteStringGetData(unsigned start) {
   return reinterpret_cast<uc16*>(
@@ skipping 246 matching lines @@
       }
     } else {
       return string->Get(index);
     }
   }
 
   UNREACHABLE();
   return 0;
 }
 
+uint16_t ThinString::ThinStringGet(int index) { return actual()->Get(index); }
 
 uint16_t SlicedString::SlicedStringGet(int index) {
   return parent()->Get(offset() + index);
 }
 
 
 template <typename sinkchar>
 void String::WriteToFlat(String* src,
                          sinkchar* sink,
                          int f,
@@ skipping 74 matching lines @@
         }
         break;
       }
       case kOneByteStringTag | kSlicedStringTag:
       case kTwoByteStringTag | kSlicedStringTag: {
         SlicedString* slice = SlicedString::cast(source);
         unsigned offset = slice->offset();
         WriteToFlat(slice->parent(), sink, from + offset, to + offset);
         return;
       }
+      case kOneByteStringTag | kThinStringTag:
+      case kTwoByteStringTag | kThinStringTag:
+        source = ThinString::cast(source)->actual();
+        break;
     }
   }
 }
 
 
 
 template <typename SourceChar>
 static void CalculateLineEndsImpl(Isolate* isolate,
                                   List<int>* line_ends,
                                   Vector<const SourceChar> src,
@@ skipping 201 matching lines @@
 };
 
 
 bool String::SlowEquals(String* other) {
   DisallowHeapAllocation no_gc;
   // Fast check: negative check with lengths.
   int len = length();
   if (len != other->length()) return false;
   if (len == 0) return true;
 
+  // Fast check: if at least one ThinString is involved, dereference it/them
+  // and restart.
+  if (this->IsThinString() || other->IsThinString()) {
+    if (other->IsThinString()) other = ThinString::cast(other)->actual();
+    if (this->IsThinString()) {
+      return ThinString::cast(this)->actual()->Equals(other);
+    } else {
+      return this->Equals(other);
+    }
+  }
+
   // Fast check: if hash code is computed for both strings
   // a fast negative check can be performed.
   if (HasHashCode() && other->HasHashCode()) {
 #ifdef ENABLE_SLOW_DCHECKS
     if (FLAG_enable_slow_asserts) {
       if (Hash() != other->Hash()) {
         bool found_difference = false;
         for (int i = 0; i < len; i++) {
           if (Get(i) != other->Get(i)) {
             found_difference = true;
@@ skipping 21 matching lines @@
   return comparator.Equals(this, other);
 }
 
 
 bool String::SlowEquals(Handle<String> one, Handle<String> two) {
   // Fast check: negative check with lengths.
   int one_length = one->length();
   if (one_length != two->length()) return false;
   if (one_length == 0) return true;
 
+  // Fast check: if at least one ThinString is involved, dereference it/them
+  // and restart.
+  if (one->IsThinString() || two->IsThinString()) {
+    if (one->IsThinString()) one = handle(ThinString::cast(*one)->actual());
+    if (two->IsThinString()) two = handle(ThinString::cast(*two)->actual());
+    return String::Equals(one, two);
+  }
+
   // Fast check: if hash code is computed for both strings
   // a fast negative check can be performed.
   if (one->HasHashCode() && two->HasHashCode()) {
 #ifdef ENABLE_SLOW_DCHECKS
     if (FLAG_enable_slow_asserts) {
       if (one->Hash() != two->Hash()) {
         bool found_difference = false;
         for (int i = 0; i < one_length; i++) {
           if (one->Get(i) != two->Get(i)) {
             found_difference = true;
@@ skipping 5129 matching lines @@
   Handle<Object> AsHandle(Isolate* isolate) override {
     // Internalize the string if possible.
     MaybeHandle<Map> maybe_map =
         isolate->factory()->InternalizedStringMapForString(string_);
     Handle<Map> map;
     if (maybe_map.ToHandle(&map)) {
       string_->set_map_no_write_barrier(*map);
       DCHECK(string_->IsInternalizedString());
       return string_;
     }
+    // External strings get special treatment, to avoid copying their contents.
+    if (string_->IsExternalOneByteString()) {
+      return isolate->factory()
+          ->InternalizeExternalString<ExternalOneByteString>(string_);
+    } else if (string_->IsExternalTwoByteString()) {
+      return isolate->factory()
+          ->InternalizeExternalString<ExternalTwoByteString>(string_);
+    }
     // Otherwise allocate a new internalized string.
     return isolate->factory()->NewInternalizedStringImpl(
         string_, string_->length(), string_->hash_field());
   }
 
   static uint32_t StringHash(Object* obj) {
     return String::cast(obj)->Hash();
   }
 
+ private:
   Handle<String> string_;
 };
 
 
 template<typename Derived, typename Shape, typename Key>
 void HashTable<Derived, Shape, Key>::IteratePrefix(ObjectVisitor* v) {
   BodyDescriptorBase::IteratePointers(this, 0, kElementsStartOffset, v);
 }
 
 
@@ skipping 914 matching lines @@
   uint32_t hash_;
 };
 
 
 MaybeHandle<String> StringTable::InternalizeStringIfExists(
     Isolate* isolate,
     Handle<String> string) {
   if (string->IsInternalizedString()) {
     return string;
   }
+  if (string->IsThinString()) {
+    return handle(Handle<ThinString>::cast(string)->actual(), isolate);
+  }
   return LookupStringIfExists(isolate, string);
 }
 
 
 MaybeHandle<String> StringTable::LookupStringIfExists(
     Isolate* isolate,
     Handle<String> string) {
   Handle<StringTable> string_table = isolate->factory()->string_table();
   InternalizedStringKey key(string);
   int entry = string_table->FindEntry(&key);
@@ skipping 26 matching lines @@
 
 void StringTable::EnsureCapacityForDeserialization(Isolate* isolate,
                                                    int expected) {
   Handle<StringTable> table = isolate->factory()->string_table();
   // We need a key instance for the virtual hash function.
   InternalizedStringKey dummy_key(isolate->factory()->empty_string());
   table = StringTable::EnsureCapacity(table, expected, &dummy_key);
   isolate->heap()->SetRootStringTable(*table);
 }
 
+namespace {
+
+template <class StringClass>
+void MigrateExternalStringResource(Isolate* isolate, Handle<String> from,
+                                   Handle<String> to) {
+  Handle<StringClass> cast_from = Handle<StringClass>::cast(from);
+  Handle<StringClass> cast_to = Handle<StringClass>::cast(to);
+  const typename StringClass::Resource* to_resource = cast_to->resource();
+  if (to_resource == nullptr) {
+    // |to| is a just-created internalized copy of |from|. Migrate the resource.
+    cast_to->set_resource(cast_from->resource());
+    // Zap |from|'s resource pointer to reflect the fact that |from| has
+    // relinquished ownership of its resource.
+    cast_from->set_resource(nullptr);
+  } else if (to_resource != cast_from->resource()) {
+    // |to| already existed and has its own resource. Finalize |from|.
+    isolate->heap()->FinalizeExternalString(*from);
+  }
+}
+
+}  // namespace
 
 Handle<String> StringTable::LookupString(Isolate* isolate,
                                          Handle<String> string) {
+  if (string->IsThinString()) {
+    DCHECK(Handle<ThinString>::cast(string)->actual()->IsInternalizedString());
+    return handle(Handle<ThinString>::cast(string)->actual(), isolate);
+  }
   if (string->IsConsString() && string->IsFlat()) {
-    string = String::Flatten(string);
+    string = handle(Handle<ConsString>::cast(string)->first(), isolate);
     if (string->IsInternalizedString()) return string;
   }
 
   InternalizedStringKey key(string);
   Handle<String> result = LookupKey(isolate, &key);
 
-  if (string->IsConsString()) {
-    Handle<ConsString> cons = Handle<ConsString>::cast(string);
-    cons->set_first(*result);
-    cons->set_second(isolate->heap()->empty_string());
-  } else if (string->IsSlicedString()) {
-    STATIC_ASSERT(ConsString::kSize == SlicedString::kSize);
+  if (string->IsExternalString()) {
+    if (result->IsExternalOneByteString()) {
+      MigrateExternalStringResource<ExternalOneByteString>(isolate, string,
+                                                           result);
+    } else if (result->IsExternalTwoByteString()) {
+      MigrateExternalStringResource<ExternalTwoByteString>(isolate, string,
+                                                           result);
+    }
+  }
+
+  // The LookupKey() call above tries to internalize the string in-place.
+  // In cases where that wasn't possible (e.g. new-space strings), turn them
+  // into ThinStrings referring to their internalized versions now.
+  if (!string->IsInternalizedString()) {
     DisallowHeapAllocation no_gc;
     bool one_byte = result->IsOneByteRepresentation();
-    Handle<Map> map = one_byte ? isolate->factory()->cons_one_byte_string_map()
-                               : isolate->factory()->cons_string_map();
-    string->set_map(*map);
-    Handle<ConsString> cons = Handle<ConsString>::cast(string);
-    cons->set_first(*result);
-    cons->set_second(isolate->heap()->empty_string());
+    Handle<Map> map = one_byte ? isolate->factory()->thin_one_byte_string_map()
+                               : isolate->factory()->thin_string_map();
+    int old_size = string->Size();
+    DCHECK(old_size >= ThinString::kSize);
+    string->synchronized_set_map(*map);
+    Handle<ThinString> thin = Handle<ThinString>::cast(string);
+    thin->set_actual(*result);
+    Address thin_end = thin->address() + ThinString::kSize;
+    int size_delta = old_size - ThinString::kSize;
+    Heap* heap = isolate->heap();

[Igor Sheludko, 2016/12/20 23:46:58]

Maybe wrap below lines in "if (size_delta) {" to simplify small-string case
(because AdjustLiveBytes() looks heavy).

[Jakob Kummerow, 2017/01/04 12:45:06]

Done.

+    heap->CreateFillerObjectAt(thin_end, size_delta, ClearRecordedSlots::kNo);
+    heap->AdjustLiveBytes(*thin, -size_delta);
   }
   return result;
 }
 
 
 Handle<String> StringTable::LookupKey(Isolate* isolate, HashTableKey* key) {
   Handle<StringTable> table = isolate->factory()->string_table();
   int entry = table->FindEntry(key);
 
   // String already in table.
@@ skipping 2648 matching lines @@
       // depend on this.
       return DICTIONARY_ELEMENTS;
     }
     DCHECK_LE(kind, LAST_ELEMENTS_KIND);
     return kind;
   }
 }
 
 }  // namespace internal
 }  // namespace v8