A JIT for D? That would be many, many man-years of work.
Wrong!
It would be quite easy. I've figured it out myself.
I've thought of using Pegged with PEG/BNF ParseTrees, made faster using
ParseTree serialization matched to the source's CRC32-encoding for
memoization, with a multi-threaded model of tree-generation when
branching in Or! steps.
Then, the interpretation stage can re-create the D environment with hash
maps of a "Entity" Variant of D types, functions, values and
instantiations added with their respective identifiers, each as a single
Entity struct containing their instructions as a simple array.
e.g. struct Entity { Entity[] scope; Entity[] instructions }
Once that's done, you should have all your types and objects in
HashMaps, ready to execute with an entry point. You move through the
instructions with a foreach loop by looking up function calls and
expecting the appropriate return type.
Call me crazy, but I think typed variants would be easier and faster to
handle than the regular casting methods that interpreted languages use
currently.
Anyway, the hardest part again would be to write a good garbage collector.
