SMS scheduling

[Mostafa Hagog]

Richard Henderson <rth@redhat.com> wrote on 30/09/2004 19:03:35:

> On Thu, Sep 30, 2004 at 10:51:31AM +0200, Mostafa Hagog wrote:
> > > On Thu, Sep 30, 2004 at 10:03:37AM +0200, Mostafa Hagog wrote:
> > > > For 4.0 SMS intra-loop memory dependencies are detected using
> > > > the alias.c (as in haifa-sched) for inter-loop memory dependencies
> > > > we are very conservative and assume that all memory accesses are
> > > > aliased.
> > >
> > > alias.c is flow insensative, and thus as conservative as you
> > > could want.  Why aren't you using it all the time?
> >
> > The answer is within your question; flow insensitive.
> > In SMS we are interested in loop-carried dependancies
> > and this is not provided by alias.c.
>
> Certainly, but surely flow insensitive data is better than
> assuming all memory accesses are aliased.

Yes sure, but the only useful information that we can get
is the alias sets. If two memory accesses are in different alias
sets then they don't alias, otherwise we must assume they are
aliased across iterations. This is one of the patches that
are pending to post-gcc4.0.

Mostafa.