The ByteTree format is a binary format to efficiently serialize and deserialize trees. It was designed to serialize the syntax tree in
libSyntax but the framework allows serialisation of arbitrary trees. It currently offers a serialiser written in C++ and a deserialiser written in Swift.
The ByteTree format consists of two different constructs: objects and scalars. A scalar is a raw sequence of binary data. Scalars are untyped and the meaning of their binary data needs to be inferred by the client based on their position in the tree. An object consists of multiple fields, indexed by their position within the object, which again can be either objects or scalars.
A scalar is encoded as its size followed by the data. Size is a
uint_32 that represents the size of the data in bytes in little endian order. It always has its most significant bit set to 0 (to distinguish objects from scalars, see Forwards compatibility).
For example, the string “Hello World” would be encoded as
"Hello World", or in hex
0B 00 00 00 48 65 6C 6C 6F 20 57 6F 72 6C 64.
An object consists of its size, measured in the number of fields and represented as a
uint_32t in little endian order, followed by the direct concatenation of its fields. Because each field is again prefixed with its size, no delimits are necessary in between the fields.
To distinguish scalars and objects, the size of objects has its most-significant bit set to 1. It must be ignored to retrieve the number of fields in the object.
Arrays are modelled as objects whose fields are all of the same type and whose length is variadic (and is indicated by the object's size).
The ByteTree format is prepended by a 4-byte protocol version number that describes the version of the object tree that was serialized. Its exact semantics are up to each specific application, but it is encouraged to interpret it as a two-component number where the first component, consisting of the three most significant bytes, is incremented for breaking changes and the last byte is incremented for backwards-compatible changes.
Fields may be added to the end of objects in a backwards compatible manner (older deserialisers will still be able to deserialise the format). It does so by skipping over all fields that are not read during deserialisation. Newer versions of the deserialiser can detect if recently added fields are not present in the serialised data by inspecting the
numFields property passed during deserialisation.
Since all fields in objects are accessed by their index, issues quickly arise if a new field is accidentally added at the beginning of an object. To prevent issues like this, the ByteTree serialiser and deserialiser requires the explicit specification of each field's index within the object. These indices are never serialised. Their sole purpose is to check that all fields are read in the correct order in assertion builds.