Re: [patch] Move loop structures to gc memory

["Daniel Berlin" <dberlin@dberlin.org>]

On 5/17/07, Zdenek Dvorak <rakdver@kam.mff.cuni.cz> wrote:
> Hello,
>
> > The only reason i've seen is "well, it's hard".
> >
> > Speed wise, profiles of the GC have never shown the actual root list
> > walking to be slow, only the actual mark lookups and sets.  Adding
> > dynamic roots would make the root list walking slightly slower, but if
> > we have *less* things in GC overall, we will do less mark setting,
> > which is where most of GC time goes.
> > So it should be a win overall.
>
> something like the patch below?  It allows you:
>
> -- to specify a custom marking routine for a particular type, that is
>    called instead of the one provided by garbage collector.  This
>    enables the objects of this type not to be allocated by gc, and
>    still survive ggc_collect even if gc collected things point to
>    them
> -- adds dynamic roots to ggc-common
> -- implements a particular instance of considering all objects allocated
>    in an alloc pool to be roots
> -- and applies it to make basic blocks pool allocated
>
> I guess this covers most of the profitable cases (basic blocks, edges,
> loop structures), but it is easy to extend for other modes of
> allocation.  For simplicity, the patch allocates all basic blocks from
> a single pool; it might be better to use per-function pools instead
> (that would however need some more changes, in particular the way
> omp code moves basic blocks between functions would break).
>
> Of course, the patch is not quite finished (in particular, I am fairly
> sure it breaks pch), but I would like to know whether we are both
> thinking about the same thing, or if you imagine something different.
>
No, this is exactly what i'm thinking about.

You may want to change the subject line so that it gets noticed by
people not reading this hellhole of a thread :)