Currently IR is mostly consisting of the data objects defined in QV4::CompiledData , which is rather low-level representation. It's optimized for the size because of the serialization and de-serialization of CUs, however that could be just corresponding methods / Serializer & Deserializer "visitors"
The benefits would hopefully be easier reuse of the structures and transformation/mapping between them (for example AST -> HIR )
This will hopefully also increase reliability by moving away from the multiple raw-pointer manipulations and outdated structs like "PoolList"-s etc.