This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch, rfc] Make store motion use alias oracle
Hello,
> >I have some doubts about its usefulness, though. The usage of alias
> >oracle makes sm more expensive -- for each "nice" memory reference that
> >would previously be refused because of may-alias according to vops, we
> >now need to check all the memory references with that it might conflict
> >using the alias oracle;
>
> Well, actually, you can do significantly better than this by using
> schemes that enable you to only have to check a small subset of the
> memory references (those with a base that could alias, those with the
> same type if it's dereferenced, etc).
this basically seems to be the info you get from vops, so we already
use that. We can do a bit more, obviously, but I would not be as
optimistic regarding the "significantly" part. I am not really worried
about the number of tests we do (according to the measurements I sent, this
number seems to be quite small), though; but in any case, there is some
overhead over the implementation that uses purely vops.
That being said, I did not really measure compile time impacts of the
patch (there do not seem to be any significant difference for gcc
compilation, but the same does not need to be the case for more
sm-intensive benchmarks).
Zdenek
> That said, i agree that more experimentation is necessary before we go
> ahead with this.