gc_inline.h

gc_inline.h

[Andy Wingo]

It seems that the BDW GC, since version 7 or so, supports inline
allocation with thread-local freelists.  See gc_inline.h.  We can
probably improve our allocation speed significantly by using this
allocator in Guile.  It would require adding a freelist to scm_i_thread,
and passing the thread around wherever we would want inline allocation,
but that is already the case in the VM.

Could be a neat little project for someone -- mostly on master, where we
require BDW-GC 7.2.

Cheers,

Andy
-- 
http://wingolog.org/

Re: gc_inline.h

[Andy Wingo]

On Fri 22 Nov 2013 10:09, Andy Wingo <wingo@pobox.com> writes:

> It seems that the BDW GC, since version 7 or so, supports inline
> allocation with thread-local freelists.  See gc_inline.h.  We can
> probably improve our allocation speed significantly by using this
> allocator in Guile.  It would require adding a freelist to scm_i_thread,
> and passing the thread around wherever we would want inline allocation,
> but that is already the case in the VM.
>
> Could be a neat little project for someone -- mostly on master, where we
> require BDW-GC 7.2.

Not sure if I mentioned this, but I did this on master.  It's a minor
speed boost, just a few percent -- mostly because we avoid TLS lookups
to get the thread-local freelist within GC_malloc() and friends.  If we
passed the scm_i_thread* pointer to some new kind of gsubr, then
primitives could also do this; future work.

Andy
-- 
http://wingolog.org/