Re: New GC concept

[Andrea Corallo]

Daniel Colascione <dancol@dancol.org> writes:

> Emacs has had the same GC for a decent amount of time now (since the
> 1980s, really). I spent some time in 2020 rewriting it from scratch. I
> haven't had time to work on the new GC recently, but I figure I'd
> throw it out here to get some feedback on the general concept.
>
> Check out
> https://github.com/dcolascione/emacs-1/blob/newgc-wip/src/alloc.c,
> specifically the big doc comment on top
>
> The new GC basically replaces alloc.c and a few other things. It has a
> few cool features:
>
> * fully copying and compacting
>
> * special treatment of sxhash to preserve object identify even while
>   we move it around in memory
>
> * generational
>
> * contiguous storage of mark bits separately from the data heap
>
> * concurrent (in design, not current implementation): idea is that we
>   do concurrent marking and barely pause for sweep
>
> * small string optimization
>
> * bump pointer allocation of new objects
>
> * heap enumeration support
>
> * hard requirement on pdumper
>
> * specialized GC spaces for conses, strings, arrays, and so on: no
>   stupid header word for cons cells bloating memory use by 50%!
>
> * cool modern C implementation that relies heavily on compiler
>   inlining and constant propagation
>
> The current implementation is deficient in many ways. Honestly, I'm
> not even sure whether that specific revision compiles. But like I
> said, I haven't had time recently to continue work on it.
>
> Still, I'm still curious about what people think of the overall
> effort. It might work nicely with the new native compilation stuff,
> giving us a managed code execution environment kind-of, sort-of on par
> with the big modern managed-code runtimes.

Sounds cool!

The only comment I've so far is that IMO *the* important feature for a
new Emacs GC is to have it concurrent (or say concurrent as much as
possible).

Emacs user experience is often dictated by its reactivity, we need to
head towards a GC that is concurrent prioritizing in the design this
feature over others, I wouldn't mind sacrificing some efficiency for
that.

I like the idea of a moving/generational GC but possibily porting what
we have to a tri-color mark and sweep would solve already the problem
with less impact.  This is what I would have tried if I had time.

Thanks for this work!

  Andrea