Consider the future impact of OxCaml

[bluddy]

Jane Street's OxCaml implements support for unboxed floats of different kinds as well as many other optimizations. While Raven cannot diverge from upstream OCaml, as features make their way to upstream, some things may change. For example, it may turn out that unboxed float arrays together with OxCaml's SIMD support will be much faster than a BigArray-based approach. It's worth keeping this in mind to make sure Raven doesn't tie itself to BigArrays too strongly.

[tmattio]

Agreed - I'd love to see our OCaml backend superseeding the C backend if OxCaml allows us to get similar performances.

I'd very much welcome experimentations on that, I don't know if Dune allows conditional compilation depending on the compiler toolchain, but if so, we could have maintain an experimental oxcaml backend.

[tmattio]

I probably won't have time to work on a complete OxCaml backend for some time, but if anyone is interested in playing with this, I've created some setup and boilerplate in https://github.com/raven-ml/raven/tree/main/dev/nx-oxcaml