jazz-lang / jlight Goto Github PK

Ieiunium is incremental stop the world mark&sweep&compact garbage collector. This GC is already really fast and can get big boost when you will enable parallel version of this GC, but this may take long time to mark entire heap and then sweep so we need concurrent implementation of mark&sweep (compacting will require stop-the-world).

1. Bytecode should be more optimized in size because bytecode is on the heap for entire program live and may use a lot of memory if program is really huge.
2. More metadata should be stored in instructions e.g types of binary operations and similar things.
3. Instead of manually declaring each instruction we may use Python or Ruby to generate it.

Interpreter that just have big switch..case ( match in Rust) may be slow since it may not fit into CPU cache. To improve this we can implement:

1.Threaded interpreter: SImpliest solution since it doesn't require knowledge of assembly and writing big assembly files for each platform (will require some assembly to implement something like goto for threading, but it's much easier than ll-interpreter). Threaded interpreters usually much faster than just using switch...case.
2. Low-Level Interpreter: Low-level interpreter should be written in assembly, but it will give very big boost to performance since it will use real machine stack instead of "fake" stack that you allocate in heap for bytecode program.