Re: guile and libgccjit

[tomas]

On Sun, Oct 10, 2021 at 09:57:13AM +0200, Linus Björnstam wrote:
> The the current JIT is, in my understanding, more or less a stepping stone 
> towards native compilation. The current JIT is very small and lightweight, 
> without any tracing, meaning things like register allocation is flexible 
> going forward.
> 
> I am not a maintainer, nor heavily involved, but I believe libgccjit would be 
> a step sideways considering the above. 

I think there's some (very readable!) rationale on why Andy chose a
relatively "simple" JIT in his blog [1]. It's a fork of GNU lightning.
The post explains why the powerful optimisations built into new
versions of Lightning seem contraproductive for Guile. Since libgccjit
will have even more powerful optimisations (it's gcc, after all!), I
guess the same arguments hold.

Note that Emacs is currently integrating an Elisp compiler based on
libgccjit. But as far as I can see, it's more an AOT compiler than a
JIT.

Cheers
[1] https://wingolog.org/archives/2019/05/24/lightening-run-time-code-generation
 - t

signature.asc
Description: Digital signature