[dict-proto] C++ implementation of SwissNameDictionary, pt. 6

This CL is part of a series that adds the C++ implementation of
SwissNameDictionary, a deterministic property backing store based on
Swiss Tables.

This CL adds the Add, DeleteEntry, Rehash, and Shrink operations, as
well as some helpers required by those.

Bug: v8:11388
Change-Id: I81a181e99e679095c2acc877a293e81b6a581db4
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2715192Reviewed-by: Igor Sheludko <ishell@chromium.org>
Reviewed-by: Marja Hölttä <marja@chromium.org>
Commit-Queue: Frank Emrich <emrich@google.com>
Cr-Commit-Position: refs/heads/master@{#73020}

src/objects/swiss-name-dictionary-inl.h

@@ -317,6 +317,21 @@ PropertyDetails SwissNameDictionary::DetailsAt(InternalIndex entry) {
   return DetailsAt(entry.as_int());
 }
 
+// static
+template <typename LocalIsolate>
+Handle<SwissNameDictionary> SwissNameDictionary::EnsureGrowable(
+    LocalIsolate* isolate, Handle<SwissNameDictionary> table) {
+  int capacity = table->Capacity();
+
+  if (table->UsedCapacity() < MaxUsableCapacity(capacity)) {
+    // We have room for at least one more entry, nothing to do.
+    return table;
+  }
+
+  int new_capacity = capacity == 0 ? kInitialCapacity : capacity * 2;
+  return Rehash(isolate, table, new_capacity);
+}
+
 swiss_table::ctrl_t SwissNameDictionary::GetCtrl(int entry) {
   DCHECK_LT(static_cast<unsigned>(entry), static_cast<unsigned>(Capacity()));
 
@@ -352,6 +367,25 @@ void SwissNameDictionary::SetCtrl(int entry, ctrl_t h) {
   ctrl[copy_entry] = h;
 }
 
+// static
+inline int SwissNameDictionary::FindFirstEmpty(uint32_t hash) {
+  // See SwissNameDictionary::FindEntry for description of probing algorithm.
+
+  auto seq = probe(hash, Capacity());
+  while (true) {
+    Group g{CtrlTable() + seq.offset()};
+    auto mask = g.MatchEmpty();
+    if (mask) {
+      // Note that picking the lowest bit set here means using the leftmost
+      // empty bucket in the group. Here, "left" means smaller entry/bucket
+      // index.
+      return seq.offset(mask.LowestBitSet());
+    }
+    seq.next();
+    DCHECK_LT(seq.index(), Capacity());
+  }
+}
+
 void SwissNameDictionary::SetMetaTableField(int field_index, int value) {
   // See the STATIC_ASSERTs on |kMax1ByteMetaTableCapacity| and
   // |kMax2ByteMetaTableCapacity| in the .cc file for an explanation of these
@@ -453,6 +487,57 @@ bool SwissNameDictionary::ToKey(ReadOnlyRoots roots, InternalIndex entry,
   return ToKey(roots, entry.as_int(), out_key);
 }
 
+// static
+template <typename LocalIsolate>
+Handle<SwissNameDictionary> SwissNameDictionary::Add(
+    LocalIsolate* isolate, Handle<SwissNameDictionary> original_table,
+    Handle<Name> key, Handle<Object> value, PropertyDetails details,
+    InternalIndex* entry_out) {
+  DCHECK(original_table->FindEntry(isolate, *key).is_not_found());
+
+  Handle<SwissNameDictionary> table = EnsureGrowable(isolate, original_table);
+
+  int nof = table->NumberOfElements();
+  int nod = table->NumberOfDeletedElements();
+  int new_enum_index = nof + nod;
+
+  int new_entry = table->AddInternal(*key, *value, details);
+
+  table->SetNumberOfElements(nof + 1);
+  table->SetEntryForEnumerationIndex(new_enum_index, new_entry);
+
+  if (entry_out) {
+    *entry_out = InternalIndex(new_entry);
+  }
+
+  return table;
+}
+
+int SwissNameDictionary::AddInternal(Name key, Object value,
+                                     PropertyDetails details) {
+  DisallowHeapAllocation no_gc;
+
+  DCHECK(key.IsUniqueName());
+  DCHECK_LE(UsedCapacity(), MaxUsableCapacity(Capacity()));
+
+  uint32_t hash = key.hash();
+
+  // For now we don't re-use deleted buckets (due to enumeration table
+  // complications), which is why we only look for empty buckets here, not
+  // deleted ones.
+  int target = FindFirstEmpty(hash);
+
+  SetCtrl(target, swiss_table::H2(hash));
+  SetKey(target, key);
+  ValueAtPut(target, value);
+  DetailsAtPut(target, details);
+
+  // Note that we do not update the number of elements or the enumeration table
+  // in this function.
+
+  return target;
+}
+
 template <typename LocalIsolate>
 void SwissNameDictionary::Initialize(LocalIsolate* isolate,
                                      ByteArray meta_table, int capacity) {

src/objects/swiss-name-dictionary.cc

@@ -10,6 +10,90 @@
 namespace v8 {
 namespace internal {
 
+// static
+Handle<SwissNameDictionary> SwissNameDictionary::DeleteEntry(
+    Isolate* isolate, Handle<SwissNameDictionary> table, InternalIndex entry) {
+  // GetCtrl() does the bounds check.
+  DCHECK(IsFull(table->GetCtrl(entry.as_int())));
+
+  int i = entry.as_int();
+
+  table->SetCtrl(i, Ctrl::kDeleted);
+  table->ClearDataTableEntry(isolate, i);
+  // We leave the PropertyDetails unchanged because they are not relevant for
+  // GC.
+
+  int nof = table->NumberOfElements();
+  table->SetNumberOfElements(nof - 1);
+  int nod = table->NumberOfDeletedElements();
+  table->SetNumberOfDeletedElements(nod + 1);
+
+  // TODO(v8:11388) Abseil's flat_hash_map doesn't shrink on deletion, but may
+  // decide on addition to do an in-place rehash to remove deleted elements. We
+  // shrink on deletion here to follow what NameDictionary and
+  // OrderedNameDictionary do. We should investigate which approach works
+  // better.
+  return Shrink(isolate, table);
+}
+
+// static
+template <typename LocalIsolate>
+Handle<SwissNameDictionary> SwissNameDictionary::Rehash(
+    LocalIsolate* isolate, Handle<SwissNameDictionary> table,
+    int new_capacity) {
+  DCHECK(IsValidCapacity(new_capacity));
+  DCHECK_LE(table->NumberOfElements(), MaxUsableCapacity(new_capacity));
+  ReadOnlyRoots roots(isolate);
+
+  Handle<SwissNameDictionary> new_table =
+      isolate->factory()->NewSwissNameDictionaryWithCapacity(
+          new_capacity, Heap::InYoungGeneration(*table) ? AllocationType::kYoung
+                                                        : AllocationType::kOld);
+
+  DisallowHeapAllocation no_gc;
+
+  int new_enum_index = 0;
+  new_table->SetNumberOfElements(table->NumberOfElements());
+  for (int enum_index = 0; enum_index < table->UsedCapacity(); ++enum_index) {
+    int entry = table->EntryForEnumerationIndex(enum_index);
+
+    Object key;
+
+    if (table->ToKey(roots, entry, &key)) {
+      Object value = table->ValueAtRaw(entry);
+      PropertyDetails details = table->DetailsAt(entry);
+
+      int new_entry = new_table->AddInternal(Name::cast(key), value, details);
+
+      // TODO(v8::11388) Investigate ways of hoisting the branching needed to
+      // select the correct meta table entry size (based on the capacity of the
+      // table) out of the loop.
+      new_table->SetEntryForEnumerationIndex(new_enum_index, new_entry);
+      ++new_enum_index;
+    }
+  }
+
+  new_table->SetHash(table->Hash());
+  return new_table;
+}
+
+// static
+Handle<SwissNameDictionary> SwissNameDictionary::Shrink(
+    Isolate* isolate, Handle<SwissNameDictionary> table) {
+  // TODO(v8:11388) We're using the same logic to decide whether or not to
+  // shrink as OrderedNameDictionary and NameDictionary here. We should compare
+  // this with the logic used by Abseil's flat_hash_map, which has a heuristic
+  // for triggering an (in-place) rehash on addition, but never shrinks the
+  // table. Abseil's heuristic doesn't take the numbere of deleted elements into
+  // account, because it doesn't track that.
+
+  int nof = table->NumberOfElements();
+  int capacity = table->Capacity();
+  if (nof >= (capacity >> 2)) return table;
+  int new_capacity = std::max(capacity / 2, kInitialCapacity);
+  return Rehash(isolate, table, new_capacity);
+}
+
 // The largest value we ever have to store in the enumeration table is
 // Capacity() - 1. The largest value we ever have to store for the present or
 // deleted element count is MaxUsableCapacity(Capacity()). All data in the

src/objects/swiss-name-dictionary.h

@@ -72,6 +72,18 @@ class SwissNameDictionary : public HeapObject {
  public:
   using Group = swiss_table::Group;
 
+  template <typename LocalIsolate>
+  inline static Handle<SwissNameDictionary> Add(
+      LocalIsolate* isolate, Handle<SwissNameDictionary> table,
+      Handle<Name> key, Handle<Object> value, PropertyDetails details,
+      InternalIndex* entry_out = nullptr);
+
+  static Handle<SwissNameDictionary> Shrink(Isolate* isolate,
+                                            Handle<SwissNameDictionary> table);
+
+  static Handle<SwissNameDictionary> DeleteEntry(
+      Isolate* isolate, Handle<SwissNameDictionary> table, InternalIndex entry);
+
   template <typename LocalIsolate>
   inline InternalIndex FindEntry(LocalIsolate* isolate, Object key);
 
@@ -103,6 +115,11 @@ class SwissNameDictionary : public HeapObject {
   template <typename LocalIsolate>
   void Initialize(LocalIsolate* isolate, ByteArray meta_table, int capacity);
 
+  template <typename LocalIsolate>
+  static Handle<SwissNameDictionary> Rehash(LocalIsolate* isolate,
+                                            Handle<SwissNameDictionary> table,
+                                            int new_capacity);
+
   inline void SetHash(int hash);
   inline int Hash();
 
@@ -218,6 +235,10 @@ class SwissNameDictionary : public HeapObject {
   using ctrl_t = swiss_table::ctrl_t;
   using Ctrl = swiss_table::Ctrl;
 
+  template <typename LocalIsolate>
+  inline static Handle<SwissNameDictionary> EnsureGrowable(
+      LocalIsolate* isolate, Handle<SwissNameDictionary> table);
+
   // Returns table of byte-encoded PropertyDetails (without enumeration index
   // stored in PropertyDetails).
   inline uint8_t* PropertyDetailsTable();
@@ -235,6 +256,13 @@ class SwissNameDictionary : public HeapObject {
 
   inline bool ToKey(ReadOnlyRoots roots, int entry, Object* out_key);
 
+  inline int FindFirstEmpty(uint32_t hash);
+  // Adds |key| ->  (|value|, |details|) as a new mapping to the table, which
+  // must have sufficient room. Returns the entry (= bucket) used by the new
+  // mapping. Does not update the number of present entries or the
+  // enumeration table.
+  inline int AddInternal(Name key, Object value, PropertyDetails details);
+
   // Use |set_ctrl| for modifications whenever possible, since that function
   // correctly maintains the copy of the first group at the end of the ctrl
   // table.