Re: [PATCH][RFC] Add versioning for constant strides for vectorization

[Richard Guenther <rguenther at suse dot de>]

On Tue, 27 Jan 2009, Dorit Nuzman wrote:

> Richard Guenther <rguenther@suse.de> wrote on 27/01/2009 14:42:44:
> 
> > On Tue, 27 Jan 2009, Dorit Nuzman wrote:
> >
> ...
> > > about:
> > >
> > > > +   /* ???  Delay this change until after versioning or
> > > > +      preserve the original step somewhere.  */
> >
> > Ah, this comment is a left-over.  The change here is definitely needed.
> >
> > > When we take this forward to support outer-loop vectorization with
> unknown
> > > inner-loop strides, we'd indeed want to
> > > delay this change.
> >
> > I guess outer-loop vectorization should simply deal with non-constant
> > DR_STEP then, even when the versioning for non-constant strides is not
> > done?
> 
> yes, exactly
> 
> > So, the data-ref change should already support this?
> >
> 
> Sebastian suggested a patch for data-ref in
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33113#c5. Note that dependence
> analysis does not support unknown strides, but we don't always need
> dependence analysis (e.g. when we have a reduction and no stores, or when
> we obviously deal with different arrays).,
> 
> > I have SPEC2006 tested the patch and it doesn't show any improvements
> > or regressions.
> >
> 
> (wonder how many times we actually applied this versioning in SPEC?)

It happens quiet often in tonto, but mostly on never called
subroutines.

Richard.