google-protobuf 3.22.5 → 3.24.3

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: ed447d83d06d36c581e489caa90787e14eb0ef197e9a1f8019f1d8e39ab3930e
-  data.tar.gz: 4ccbc995a7bc56cfd96d171d50fd81eac19b20f0fb10969cae0d59c5a424716b
+  metadata.gz: 802cae742cd0aa258920ff9bde883ab655092d76cde65b4b047dd60a5a6ebd24
+  data.tar.gz: 55ca2a83f2a774be049602510b19b1bd53fcc7a5b10988c8a5020cf38aa354c3
 SHA512:
-  metadata.gz: 136713eb6ec3326893e85b49861f5d399c137b222e7eaee77f028dc9195f3e9850f0a591735ffe2c01af0ee6f3213992e83f7a10ddeb0eb086051b748b57bba0
-  data.tar.gz: 118a5c4657e91b3276ea2f8ca5c42526df71a49861e130dac13901019f9ed2391f3a1c2f7da150897d72e63bf56a0af1d4e75c264e53ba0dd9278955db944249
+  metadata.gz: fd69b0c91d8bf5597c5ec98b77b597af7182386d99d78f4a84b4fde0c352a930005ad8bbba83650b5d54069cbb5503be4b302622229edee04d7e2fe8ee790804
+  data.tar.gz: 74c4c1057144c20b2e5efc49f665fd066e7507a32b77222154b3a4bd1dc7c6c1439a12398363a95be45244cc59461b651563c456c1c3916a51ee506de01d447f

data/ext/google/protobuf_c/convert.h

@@ -31,8 +31,6 @@
 #ifndef RUBY_PROTOBUF_CONVERT_H_
 #define RUBY_PROTOBUF_CONVERT_H_
 
-#include <ruby/ruby.h>
-
 #include "protobuf.h"
 #include "ruby-upb.h"
 

data/ext/google/protobuf_c/defs.c

@@ -73,6 +73,8 @@ static VALUE rb_str_maybe_null(const char* s) {
 // -----------------------------------------------------------------------------
 
 typedef struct {
+  // IMPORTANT: WB_PROTECTED objects must only use the RB_OBJ_WRITE()
+  // macro to update VALUE references, as to trigger write barriers.
   VALUE def_to_descriptor;  // Hash table of def* -> Ruby descriptor.
   upb_DefPool* symtab;
 } DescriptorPool;
@@ -97,7 +99,7 @@ static void DescriptorPool_free(void* _self) {
 static const rb_data_type_t DescriptorPool_type = {
     "Google::Protobuf::DescriptorPool",
     {DescriptorPool_mark, DescriptorPool_free, NULL},
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED,
 };
 
 static DescriptorPool* ruby_to_DescriptorPool(VALUE val) {
@@ -125,11 +127,9 @@ static VALUE DescriptorPool_alloc(VALUE klass) {
   self->def_to_descriptor = Qnil;
   ret = TypedData_Wrap_Struct(klass, &DescriptorPool_type, self);
 
-  self->def_to_descriptor = rb_hash_new();
+  RB_OBJ_WRITE(ret, &self->def_to_descriptor, rb_hash_new());
   self->symtab = upb_DefPool_New();
-  ObjectCache_Add(self->symtab, ret);
-
-  return ret;
+  return ObjectCache_TryAdd(self->symtab, ret);
 }
 
 /*
@@ -223,6 +223,8 @@ static void DescriptorPool_register(VALUE module) {
 
 typedef struct {
   const upb_MessageDef* msgdef;
+  // IMPORTANT: WB_PROTECTED objects must only use the RB_OBJ_WRITE()
+  // macro to update VALUE references, as to trigger write barriers.
   VALUE klass;
   VALUE descriptor_pool;
 } Descriptor;
@@ -238,7 +240,7 @@ static void Descriptor_mark(void* _self) {
 static const rb_data_type_t Descriptor_type = {
     "Google::Protobuf::Descriptor",
     {Descriptor_mark, RUBY_DEFAULT_FREE, NULL},
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED,
 };
 
 static Descriptor* ruby_to_Descriptor(VALUE val) {
@@ -280,7 +282,7 @@ static VALUE Descriptor_initialize(VALUE _self, VALUE cookie,
              "Descriptor objects may not be created from Ruby.");
   }
 
-  self->descriptor_pool = descriptor_pool;
+  RB_OBJ_WRITE(_self, &self->descriptor_pool, descriptor_pool);
   self->msgdef = (const upb_MessageDef*)NUM2ULL(ptr);
 
   return Qnil;
@@ -390,7 +392,7 @@ static VALUE Descriptor_lookup_oneof(VALUE _self, VALUE name) {
 static VALUE Descriptor_msgclass(VALUE _self) {
   Descriptor* self = ruby_to_Descriptor(_self);
   if (self->klass == Qnil) {
-    self->klass = build_class_from_descriptor(_self);
+    RB_OBJ_WRITE(_self, &self->klass, build_class_from_descriptor(_self));
   }
   return self->klass;
 }
@@ -417,6 +419,8 @@ static void Descriptor_register(VALUE module) {
 
 typedef struct {
   const upb_FileDef* filedef;
+  // IMPORTANT: WB_PROTECTED objects must only use the RB_OBJ_WRITE()
+  // macro to update VALUE references, as to trigger write barriers.
   VALUE descriptor_pool;  // Owns the upb_FileDef.
 } FileDescriptor;
 
@@ -430,7 +434,7 @@ static void FileDescriptor_mark(void* _self) {
 static const rb_data_type_t FileDescriptor_type = {
     "Google::Protobuf::FileDescriptor",
     {FileDescriptor_mark, RUBY_DEFAULT_FREE, NULL},
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED,
 };
 
 static FileDescriptor* ruby_to_FileDescriptor(VALUE val) {
@@ -463,7 +467,7 @@ static VALUE FileDescriptor_initialize(VALUE _self, VALUE cookie,
              "Descriptor objects may not be created from Ruby.");
   }
 
-  self->descriptor_pool = descriptor_pool;
+  RB_OBJ_WRITE(_self, &self->descriptor_pool, descriptor_pool);
   self->filedef = (const upb_FileDef*)NUM2ULL(ptr);
 
   return Qnil;
@@ -519,6 +523,8 @@ static void FileDescriptor_register(VALUE module) {
 
 typedef struct {
   const upb_FieldDef* fielddef;
+  // IMPORTANT: WB_PROTECTED objects must only use the RB_OBJ_WRITE()
+  // macro to update VALUE references, as to trigger write barriers.
   VALUE descriptor_pool;  // Owns the upb_FieldDef.
 } FieldDescriptor;
 
@@ -532,7 +538,7 @@ static void FieldDescriptor_mark(void* _self) {
 static const rb_data_type_t FieldDescriptor_type = {
     "Google::Protobuf::FieldDescriptor",
     {FieldDescriptor_mark, RUBY_DEFAULT_FREE, NULL},
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED,
 };
 
 static FieldDescriptor* ruby_to_FieldDescriptor(VALUE val) {
@@ -570,7 +576,7 @@ static VALUE FieldDescriptor_initialize(VALUE _self, VALUE cookie,
              "Descriptor objects may not be created from Ruby.");
   }
 
-  self->descriptor_pool = descriptor_pool;
+  RB_OBJ_WRITE(_self, &self->descriptor_pool, descriptor_pool);
   self->fielddef = (const upb_FieldDef*)NUM2ULL(ptr);
 
   return Qnil;
@@ -595,7 +601,7 @@ upb_CType ruby_to_fieldtype(VALUE type) {
 
 #define CONVERT(upb, ruby)                \
   if (SYM2ID(type) == rb_intern(#ruby)) { \
-    return kUpb_CType_##upb;                \
+    return kUpb_CType_##upb;              \
   }
 
   CONVERT(Float, float);
@@ -618,7 +624,7 @@ upb_CType ruby_to_fieldtype(VALUE type) {
 
 static VALUE descriptortype_to_ruby(upb_FieldType type) {
   switch (type) {
-#define CONVERT(upb, ruby)        \
+#define CONVERT(upb, ruby)   \
   case kUpb_FieldType_##upb: \
     return ID2SYM(rb_intern(#ruby));
     CONVERT(Float, float);
@@ -703,7 +709,7 @@ static VALUE FieldDescriptor_label(VALUE _self) {
   FieldDescriptor* self = ruby_to_FieldDescriptor(_self);
   switch (upb_FieldDef_Label(self->fielddef)) {
 #define CONVERT(upb, ruby) \
-  case kUpb_Label_##upb:    \
+  case kUpb_Label_##upb:   \
     return ID2SYM(rb_intern(#ruby));
 
     CONVERT(Optional, optional);
@@ -884,6 +890,8 @@ static void FieldDescriptor_register(VALUE module) {
 
 typedef struct {
   const upb_OneofDef* oneofdef;
+  // IMPORTANT: WB_PROTECTED objects must only use the RB_OBJ_WRITE()
+  // macro to update VALUE references, as to trigger write barriers.
   VALUE descriptor_pool;  // Owns the upb_OneofDef.
 } OneofDescriptor;
 
@@ -897,7 +905,7 @@ static void OneofDescriptor_mark(void* _self) {
 static const rb_data_type_t OneofDescriptor_type = {
     "Google::Protobuf::OneofDescriptor",
     {OneofDescriptor_mark, RUBY_DEFAULT_FREE, NULL},
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED,
 };
 
 static OneofDescriptor* ruby_to_OneofDescriptor(VALUE val) {
@@ -936,7 +944,7 @@ static VALUE OneofDescriptor_initialize(VALUE _self, VALUE cookie,
              "Descriptor objects may not be created from Ruby.");
   }
 
-  self->descriptor_pool = descriptor_pool;
+  RB_OBJ_WRITE(_self, &self->descriptor_pool, descriptor_pool);
   self->oneofdef = (const upb_OneofDef*)NUM2ULL(ptr);
 
   return Qnil;
@@ -988,6 +996,8 @@ static void OneofDescriptor_register(VALUE module) {
 
 typedef struct {
   const upb_EnumDef* enumdef;
+  // IMPORTANT: WB_PROTECTED objects must only use the RB_OBJ_WRITE()
+  // macro to update VALUE references, as to trigger write barriers.
   VALUE module;           // begins as nil
   VALUE descriptor_pool;  // Owns the upb_EnumDef.
 } EnumDescriptor;
@@ -1003,7 +1013,7 @@ static void EnumDescriptor_mark(void* _self) {
 static const rb_data_type_t EnumDescriptor_type = {
     "Google::Protobuf::EnumDescriptor",
     {EnumDescriptor_mark, RUBY_DEFAULT_FREE, NULL},
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED,
 };
 
 static EnumDescriptor* ruby_to_EnumDescriptor(VALUE val) {
@@ -1042,7 +1052,7 @@ static VALUE EnumDescriptor_initialize(VALUE _self, VALUE cookie,
              "Descriptor objects may not be created from Ruby.");
   }
 
-  self->descriptor_pool = descriptor_pool;
+  RB_OBJ_WRITE(_self, &self->descriptor_pool, descriptor_pool);
   self->enumdef = (const upb_EnumDef*)NUM2ULL(ptr);
 
   return Qnil;
@@ -1081,7 +1091,7 @@ static VALUE EnumDescriptor_name(VALUE _self) {
 static VALUE EnumDescriptor_lookup_name(VALUE _self, VALUE name) {
   EnumDescriptor* self = ruby_to_EnumDescriptor(_self);
   const char* name_str = rb_id2name(SYM2ID(name));
-  const upb_EnumValueDef *ev =
+  const upb_EnumValueDef* ev =
       upb_EnumDef_FindValueByName(self->enumdef, name_str);
   if (ev) {
     return INT2NUM(upb_EnumValueDef_Number(ev));
@@ -1100,7 +1110,8 @@ static VALUE EnumDescriptor_lookup_name(VALUE _self, VALUE name) {
 static VALUE EnumDescriptor_lookup_value(VALUE _self, VALUE number) {
   EnumDescriptor* self = ruby_to_EnumDescriptor(_self);
   int32_t val = NUM2INT(number);
-  const upb_EnumValueDef* ev = upb_EnumDef_FindValueByNumber(self->enumdef, val);
+  const upb_EnumValueDef* ev =
+      upb_EnumDef_FindValueByNumber(self->enumdef, val);
   if (ev) {
     return ID2SYM(rb_intern(upb_EnumValueDef_Name(ev)));
   } else {
@@ -1138,7 +1149,7 @@ static VALUE EnumDescriptor_each(VALUE _self) {
 static VALUE EnumDescriptor_enummodule(VALUE _self) {
   EnumDescriptor* self = ruby_to_EnumDescriptor(_self);
   if (self->module == Qnil) {
-    self->module = build_module_from_enumdesc(_self);
+    RB_OBJ_WRITE(_self, &self->module, build_module_from_enumdesc(_self));
   }
   return self->module;
 }

data/ext/google/protobuf_c/defs.h

@@ -31,8 +31,6 @@
 #ifndef RUBY_PROTOBUF_DEFS_H_
 #define RUBY_PROTOBUF_DEFS_H_
 
-#include <ruby/ruby.h>
-
 #include "protobuf.h"
 #include "ruby-upb.h"
 

data/ext/google/protobuf_c/map.c

@@ -93,7 +93,6 @@ VALUE Map_GetRubyWrapper(upb_Map* map, upb_CType key_type, TypeInfo value_type,
   if (val == Qnil) {
     val = Map_alloc(cMap);
     Map* self;
-    ObjectCache_Add(map, val);
     TypedData_Get_Struct(val, Map, &Map_type, self);
     self->map = map;
     self->arena = arena;
@@ -103,6 +102,7 @@ VALUE Map_GetRubyWrapper(upb_Map* map, upb_CType key_type, TypeInfo value_type,
       const upb_MessageDef* val_m = self->value_type_info.def.msgdef;
       self->value_type_class = Descriptor_DefToClass(val_m);
     }
+    return ObjectCache_TryAdd(map, val);
   }
 
   return val;
@@ -319,7 +319,9 @@ static VALUE Map_init(int argc, VALUE* argv, VALUE _self) {
 
   self->map = upb_Map_New(Arena_get(self->arena), self->key_type,
                           self->value_type_info.type);
-  ObjectCache_Add(self->map, _self);
+  VALUE stored = ObjectCache_TryAdd(self->map, _self);
+  (void)stored;
+  PBRUBY_ASSERT(stored == _self);
 
   if (init_arg != Qnil) {
     Map_merge_into_self(_self, init_arg);

data/ext/google/protobuf_c/map.h

@@ -31,8 +31,6 @@
 #ifndef RUBY_PROTOBUF_MAP_H_
 #define RUBY_PROTOBUF_MAP_H_
 
-#include <ruby/ruby.h>
-
 #include "protobuf.h"
 #include "ruby-upb.h"
 

data/ext/google/protobuf_c/message.c

@@ -53,6 +53,8 @@ VALUE MessageOrEnum_GetDescriptor(VALUE klass) {
 // -----------------------------------------------------------------------------
 
 typedef struct {
+  // IMPORTANT: WB_PROTECTED objects must only use the RB_OBJ_WRITE()
+  // macro to update VALUE references, as to trigger write barriers.
   VALUE arena;
   const upb_Message* msg;  // Can get as mutable when non-frozen.
   const upb_MessageDef*
@@ -65,9 +67,9 @@ static void Message_mark(void* _self) {
 }
 
 static rb_data_type_t Message_type = {
-    "Message",
+    "Google::Protobuf::Message",
     {Message_mark, RUBY_DEFAULT_FREE, NULL},
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED,
 };
 
 static Message* ruby_to_Message(VALUE msg_rb) {
@@ -105,8 +107,10 @@ upb_Message* Message_GetMutable(VALUE msg_rb, const upb_MessageDef** m) {
 void Message_InitPtr(VALUE self_, upb_Message* msg, VALUE arena) {
   Message* self = ruby_to_Message(self_);
   self->msg = msg;
-  self->arena = arena;
-  ObjectCache_Add(msg, self_);
+  RB_OBJ_WRITE(self_, &self->arena, arena);
+  VALUE stored = ObjectCache_TryAdd(msg, self_);
+  (void)stored;
+  PBRUBY_ASSERT(stored == self_);
 }
 
 VALUE Message_GetArena(VALUE msg_rb) {
@@ -978,7 +982,7 @@ static VALUE Message_decode(int argc, VALUE* argv, VALUE klass) {
         rb_hash_lookup(hash_args, ID2SYM(rb_intern("recursion_limit")));
 
     if (depth != Qnil && TYPE(depth) == T_FIXNUM) {
-      options |= UPB_DECODE_MAXDEPTH(FIX2INT(depth));
+      options |= upb_DecodeOptions_MaxDepth(FIX2INT(depth));
     }
   }
 
@@ -1096,7 +1100,7 @@ static VALUE Message_encode(int argc, VALUE* argv, VALUE klass) {
         rb_hash_lookup(hash_args, ID2SYM(rb_intern("recursion_limit")));
 
     if (depth != Qnil && TYPE(depth) == T_FIXNUM) {
-      options |= UPB_DECODE_MAXDEPTH(FIX2INT(depth));
+      options |= upb_DecodeOptions_MaxDepth(FIX2INT(depth));
     }
   }
 
@@ -1162,6 +1166,12 @@ static VALUE Message_encode_json(int argc, VALUE* argv, VALUE klass) {
                               Qfalse))) {
       options |= upb_JsonEncode_EmitDefaults;
     }
+
+    if (RTEST(rb_hash_lookup2(hash_args,
+                              ID2SYM(rb_intern("format_enums_as_integers")),
+                              Qfalse))) {
+      options |= upb_JsonEncode_FormatEnumsAsIntegers;
+    }
   }
 
   upb_Status_Clear(&status);

data/ext/google/protobuf_c/message.h

@@ -31,8 +31,6 @@
 #ifndef RUBY_PROTOBUF_MESSAGE_H_
 #define RUBY_PROTOBUF_MESSAGE_H_
 
-#include <ruby/ruby.h>
-
 #include "protobuf.h"
 #include "ruby-upb.h"
 

data/ext/google/protobuf_c/protobuf.c

@@ -171,6 +171,8 @@ void StringBuilder_PrintMsgval(StringBuilder *b, upb_MessageValue val,
 
 typedef struct {
   upb_Arena *arena;
+  // IMPORTANT: WB_PROTECTED objects must only use the RB_OBJ_WRITE()
+  // macro to update VALUE references, as to trigger write barriers.
   VALUE pinned_objs;
 } Arena;
 
@@ -190,10 +192,11 @@ static VALUE cArena;
 const rb_data_type_t Arena_type = {
     "Google::Protobuf::Internal::Arena",
     {Arena_mark, Arena_free, NULL},
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED,
 };
 
-static void* ruby_upb_allocfunc(upb_alloc* alloc, void* ptr, size_t oldsize, size_t size) {
+static void *ruby_upb_allocfunc(upb_alloc *alloc, void *ptr, size_t oldsize,
+                                size_t size) {
   if (size == 0) {
     xfree(ptr);
     return NULL;
@@ -233,7 +236,7 @@ void Arena_Pin(VALUE _arena, VALUE obj) {
   Arena *arena;
   TypedData_Get_Struct(_arena, Arena, &Arena_type, arena);
   if (arena->pinned_objs == Qnil) {
-    arena->pinned_objs = rb_ary_new();
+    RB_OBJ_WRITE(_arena, &arena->pinned_objs, rb_ary_new());
   }
   rb_ary_push(arena->pinned_objs, obj);
 }
@@ -250,164 +253,48 @@ void Arena_register(VALUE module) {
 // Object Cache
 // -----------------------------------------------------------------------------
 
-// A pointer -> Ruby Object cache that keeps references to Ruby wrapper
-// objects.  This allows us to look up any Ruby wrapper object by the address
-// of the object it is wrapping. That way we can avoid ever creating two
-// different wrapper objects for the same C object, which saves memory and
-// preserves object identity.
-//
-// We use WeakMap for the cache. For Ruby <2.7 we also need a secondary Hash
-// to store WeakMap keys because Ruby <2.7 WeakMap doesn't allow non-finalizable
-// keys.
-//
-// We also need the secondary Hash if sizeof(long) < sizeof(VALUE), because this
-// means it may not be possible to fit a pointer into a Fixnum. Keys are
-// pointers, and if they fit into a Fixnum, Ruby doesn't collect them, but if
-// they overflow and require allocating a Bignum, they could get collected
-// prematurely, thus removing the cache entry. This happens on 64-bit Windows,
-// on which pointers are 64 bits but longs are 32 bits. In this case, we enable
-// the secondary Hash to hold the keys and prevent them from being collected.
-
-#if RUBY_API_VERSION_CODE >= 20700 && SIZEOF_LONG >= SIZEOF_VALUE
-#define USE_SECONDARY_MAP 0
-#else
-#define USE_SECONDARY_MAP 1
-#endif
-
-#if USE_SECONDARY_MAP
-
-// Maps Numeric -> Object. The object is then used as a key into the WeakMap.
-// This is needed for Ruby <2.7 where a number cannot be a key to WeakMap.
-// The object is used only for its identity; it does not contain any data.
-VALUE secondary_map = Qnil;
-
-// Mutations to the map are under a mutex, because SeconaryMap_MaybeGC()
-// iterates over the map which cannot happen in parallel with insertions, or
-// Ruby will throw:
-//   can't add a new key into hash during iteration (RuntimeError)
-VALUE secondary_map_mutex = Qnil;
-
-// Lambda that will GC entries from the secondary map that are no longer present
-// in the primary map.
-VALUE gc_secondary_map_lambda = Qnil;
-ID length;
-
-extern VALUE weak_obj_cache;
-
-static void SecondaryMap_Init() {
-  rb_gc_register_address(&secondary_map);
-  rb_gc_register_address(&gc_secondary_map_lambda);
-  rb_gc_register_address(&secondary_map_mutex);
-  secondary_map = rb_hash_new();
-  gc_secondary_map_lambda = rb_eval_string(
-      "->(secondary, weak) {\n"
-      "  secondary.delete_if { |k, v| !weak.key?(v) }\n"
-      "}\n");
-  secondary_map_mutex = rb_mutex_new();
-  length = rb_intern("length");
-}
-
-// The secondary map is a regular Hash, and will never shrink on its own.
-// The main object cache is a WeakMap that will automatically remove entries
-// when the target object is no longer reachable, but unless we manually
-// remove the corresponding entries from the secondary map, it will grow
-// without bound.
-//
-// To avoid this unbounded growth we periodically remove entries from the
-// secondary map that are no longer present in the WeakMap. The logic of
-// how often to perform this GC is an artbirary tuning parameter that
-// represents a straightforward CPU/memory tradeoff.
-//
-// Requires: secondary_map_mutex is held.
-static void SecondaryMap_MaybeGC() {
-  PBRUBY_ASSERT(rb_mutex_locked_p(secondary_map_mutex) == Qtrue);
-  size_t weak_len = NUM2ULL(rb_funcall(weak_obj_cache, length, 0));
-  size_t secondary_len = RHASH_SIZE(secondary_map);
-  if (secondary_len < weak_len) {
-    // Logically this case should not be possible: a valid entry cannot exist in
-    // the weak table unless there is a corresponding entry in the secondary
-    // table. It should *always* be the case that secondary_len >= weak_len.
-    //
-    // However ObjectSpace::WeakMap#length (and therefore weak_len) is
-    // unreliable: it overreports its true length by including non-live objects.
-    // However these non-live objects are not yielded in iteration, so we may
-    // have previously deleted them from the secondary map in a previous
-    // invocation of SecondaryMap_MaybeGC().
-    //
-    // In this case, we can't measure any waste, so we just return.
-    return;
-  }
-  size_t waste = secondary_len - weak_len;
-  // GC if we could remove at least 2000 entries or 20% of the table size
-  // (whichever is greater).  Since the cost of the GC pass is O(N), we
-  // want to make sure that we condition this on overall table size, to
-  // avoid O(N^2) CPU costs.
-  size_t threshold = PBRUBY_MAX(secondary_len * 0.2, 2000);
-  if (waste > threshold) {
-    rb_funcall(gc_secondary_map_lambda, rb_intern("call"), 2, secondary_map,
-               weak_obj_cache);
-  }
-}
-
-// Requires: secondary_map_mutex is held by this thread iff create == true.
-static VALUE SecondaryMap_Get(VALUE key, bool create) {
-  PBRUBY_ASSERT(!create || rb_mutex_locked_p(secondary_map_mutex) == Qtrue);
-  VALUE ret = rb_hash_lookup(secondary_map, key);
-  if (ret == Qnil && create) {
-    SecondaryMap_MaybeGC();
-    ret = rb_class_new_instance(0, NULL, rb_cObject);
-    rb_hash_aset(secondary_map, key, ret);
-  }
-  return ret;
-}
-
-#endif
-
-// Requires: secondary_map_mutex is held by this thread iff create == true.
-static VALUE ObjectCache_GetKey(const void *key, bool create) {
-  VALUE key_val = (VALUE)key;
-  PBRUBY_ASSERT((key_val & 3) == 0);
-  VALUE ret = LL2NUM(key_val >> 2);
-#if USE_SECONDARY_MAP
-  ret = SecondaryMap_Get(ret, create);
-#endif
-  return ret;
-}
-
 // Public ObjectCache API.
 
 VALUE weak_obj_cache = Qnil;
 ID item_get;
-ID item_set;
+ID item_try_add;
+
+static void ObjectCache_Init(VALUE protobuf) {
+  item_get = rb_intern("get");
+  item_try_add = rb_intern("try_add");
 
-static void ObjectCache_Init() {
   rb_gc_register_address(&weak_obj_cache);
-  VALUE klass = rb_eval_string("ObjectSpace::WeakMap");
-  weak_obj_cache = rb_class_new_instance(0, NULL, klass);
-  item_get = rb_intern("[]");
-  item_set = rb_intern("[]=");
-#if USE_SECONDARY_MAP
-  SecondaryMap_Init();
+#if SIZEOF_LONG >= SIZEOF_VALUE
+  VALUE cache_class = rb_const_get(protobuf, rb_intern("ObjectCache"));
+#else
+  VALUE cache_class = rb_const_get(protobuf, rb_intern("LegacyObjectCache"));
 #endif
+
+  weak_obj_cache = rb_class_new_instance(0, NULL, cache_class);
+  rb_const_set(protobuf, rb_intern("OBJECT_CACHE"), weak_obj_cache);
+  rb_const_set(protobuf, rb_intern("SIZEOF_LONG"), INT2NUM(SIZEOF_LONG));
+  rb_const_set(protobuf, rb_intern("SIZEOF_VALUE"), INT2NUM(SIZEOF_VALUE));
 }
 
-void ObjectCache_Add(const void *key, VALUE val) {
-  PBRUBY_ASSERT(ObjectCache_Get(key) == Qnil);
-#if USE_SECONDARY_MAP
-  rb_mutex_lock(secondary_map_mutex);
-#endif
-  VALUE key_rb = ObjectCache_GetKey(key, true);
-  rb_funcall(weak_obj_cache, item_set, 2, key_rb, val);
-#if USE_SECONDARY_MAP
-  rb_mutex_unlock(secondary_map_mutex);
-#endif
-  PBRUBY_ASSERT(ObjectCache_Get(key) == val);
+static VALUE ObjectCache_GetKey(const void *key) {
+  VALUE key_val = (VALUE)key;
+  PBRUBY_ASSERT((key_val & 3) == 0);
+  // Ensure the key can be stored as a Fixnum since 1 bit is needed for
+  // FIXNUM_FLAG and 1 bit is needed for the sign bit.
+  VALUE new_key = LL2NUM(key_val >> 2);
+  PBRUBY_ASSERT(FIXNUM_P(new_key));
+  return new_key;
+}
+
+VALUE ObjectCache_TryAdd(const void *key, VALUE val) {
+  VALUE key_val = ObjectCache_GetKey(key);
+  return rb_funcall(weak_obj_cache, item_try_add, 2, key_val, val);
 }
 
 // Returns the cached object for this key, if any. Otherwise returns Qnil.
 VALUE ObjectCache_Get(const void *key) {
-  VALUE key_rb = ObjectCache_GetKey(key, false);
-  return rb_funcall(weak_obj_cache, item_get, 1, key_rb);
+  VALUE key_val = ObjectCache_GetKey(key);
+  return rb_funcall(weak_obj_cache, item_get, 1, key_val);
 }
 
 /*
@@ -457,11 +344,10 @@ VALUE Google_Protobuf_deep_copy(VALUE self, VALUE obj) {
 // This must be named "Init_protobuf_c" because the Ruby module is named
 // "protobuf_c" -- the VM looks for this symbol in our .so.
 __attribute__((visibility("default"))) void Init_protobuf_c() {
-  ObjectCache_Init();
-
   VALUE google = rb_define_module("Google");
   VALUE protobuf = rb_define_module_under(google, "Protobuf");
 
+  ObjectCache_Init(protobuf);
   Arena_register(protobuf);
   Defs_register(protobuf);
   RepeatedField_register(protobuf);

data/ext/google/protobuf_c/protobuf.h

@@ -31,8 +31,22 @@
 #ifndef __GOOGLE_PROTOBUF_RUBY_PROTOBUF_H__
 #define __GOOGLE_PROTOBUF_RUBY_PROTOBUF_H__
 
+// Ruby 3+ defines NDEBUG itself, see: https://bugs.ruby-lang.org/issues/18777
+#ifdef NDEBUG
+#include <ruby.h>
+#else
+#include <ruby.h>
+#undef NDEBUG
+#endif
+
+#include <ruby/version.h>
+
+#if RUBY_API_VERSION_CODE < 20700
+#error Protobuf requires Ruby >= 2.7
+#endif
+
+#include <assert.h>  // Must be included after the NDEBUG logic above.
 #include <ruby/encoding.h>
-#include <ruby/ruby.h>
 #include <ruby/vm.h>
 
 #include "defs.h"
@@ -76,10 +90,9 @@ void Arena_Pin(VALUE arena, VALUE obj);
 // being collected (though in Ruby <2.7 is it effectively strong, due to
 // implementation limitations).
 
-// Adds an entry to the cache. The "arena" parameter must give the arena that
-// "key" was allocated from.  In Ruby <2.7.0, it will be used to remove the key
-// from the cache when the arena is destroyed.
-void ObjectCache_Add(const void* key, VALUE val);
+// Tries to add a new entry to the cache, returning the newly installed value or
+// the pre-existing entry.
+VALUE ObjectCache_TryAdd(const void* key, VALUE val);
 
 // Returns the cached object for this key, if any. Otherwise returns Qnil.
 VALUE ObjectCache_Get(const void* key);
@@ -110,7 +123,9 @@ extern VALUE cTypeError;
   do {                      \
   } while (false && (expr))
 #else
-#define PBRUBY_ASSERT(expr) assert(expr)
+#define PBRUBY_ASSERT(expr) \
+  if (!(expr))              \
+  rb_bug("Assertion failed at %s:%d, expr: %s", __FILE__, __LINE__, #expr)
 #endif
 
 #define PBRUBY_MAX(x, y) (((x) > (y)) ? (x) : (y))