pbwire 0.1.7

Dynamic protobuf over the pbwire C ABI, schema at runtime, no generated code.

pbwire Dart binding

bindings/dart/lib/pbwire.dart is the Dart binding: dart:ffi over the C ABI in include/pbwire.h. A FileDescriptorSet (protoc --descriptor_set_out) becomes the schema at run time; there is no generated code and no C glue. The wire format lives in the C core, so a Dart message and a C message are the same bytes by construction.

import 'package:pbwire/pbwire.dart';

final schema = Schema.load('tests/data/matrix3.desc');
final ty = schema.message('structtest.Scalars3');
final innerTy = schema.message('structtest.Inner3');

final msg = ty.newMessage()
  ..setInt('i32', -42)
  ..setString('str', 'alice');
for (final tag in ['a', 'bb']) {
  msg.setString('tags', tag);                       // repeated: appends
}
msg.setMessage('inner')                             // nested: created in place
  ..setString(innerTy.field('name'), 'kid')
  ..setInt('value', 42);

final wire = msg.encode();
final decoded = ty.decode(wire);                    // unknown fields preserved
assert(decoded.at('tags', 1)?.asString() == 'bb');
assert(_equals(decoded.encode(), wire));            // byte-exact round trip

Fields are addressed by name ('i32') or by a Field handle from ty.field('i32'); the handle skips the name lookup and is what a hot loop should reuse.

Requirements

The Dart SDK (3.x) and make shared (build/libpbwire.dylib / .so). PBWIRE_LIB points at the library; otherwise it is found by walking up from the working directory or the entry script. The only package dependency is package:ffi, for malloc and toNativeUtf8/toDartString.

Build and run

dart pub get
dart run tests/selftest.dart                        # self-check vs pinned protoc bytes
dart run examples/roundtrip.dart                    # example

Benchmarks

cd benchmarks && dart pub get && ./bench.sh         # vs the protobuf Dart runtime, four tiers

Needs dart pub global activate protoc_plugin (it provides protoc-gen-dart; add ~/.pub-cache/bin to PATH). The four tiers are small / medium / large / xlarge (the repeated fields scaled x0 / x1 / x100 / x3000). Copied from BENCHMARKS.md, which is the single source for all performance numbers; higher is better, MB/s over wire bytes:

payload	op	pbwire-dart (MB/s, x protobuf-dart)	protobuf-dart (MB/s)
136 B	encode	775.8 (3.7x)	207.8
136 B	decode	383.8 (1.8x)	218.3
174 B	encode	709.9 (3.8x)	185.6
174 B	decode	395.5 (2.0x)	196.2
2469 B	encode	1113.7 (8.7x)	127.4
2469 B	decode	492.3 (3.2x)	151.6
74913 B	encode	1273.8 (12.2x)	104.0
74913 B	decode	555.9 (3.3x)	166.3

Strings cross as NUL-terminated C strings, so setString() stops at the first NUL byte: use setBytes() for binary payloads (a string field accepts them too, which is how protobuf's wire format works).

Ownership and threads

A Schema and a root Message own native memory: close() frees it, and a Finalizer catches the ones that are not closed. A Type keeps its Schema (and a Message its Type) reachable, so garbage collection cannot free the schema under a type or message; an explicit close() is still the caller's promise that nothing else uses it. A nested message is a view of its parent, holds it alive, and never frees anything. dart:ffi state is per-isolate, and a message tree is not shared: one tree belongs to one isolate, like the C core.

Typed models

protoc-gen-pbwire-typed (in ../typed) generates Dart classes with typed fields, Uint8List encode(schema) and static decode. The committed output is tests/gen/ and examples/gen/; examples/roundtrip.dart assigns typed fields only. See ../typed/README.md.