[graphite] Fix for pr38492 and pr38498

Jack Howarth howarth@bromo.med.uc.edu
Wed Jan 7 14:23:00 GMT 2009


On Wed, Jan 07, 2009 at 06:06:51AM -0800, Jan Sjodin wrote:
> > Sebastian,
> >    Interestingly, if I add -floop-block to
> > "-ffast-math -funroll-loops -msse3 -O3
> > -fgraphite-identity",
> > I find that the capacita benchmark fails again with the
> > same NaN errors as before Jan's patch. Is it
> > possible that Jan's patch needs to extended to handle
> > -floop-block?
> >                  Jack
> 
> Yes, blocking can introduce similar errors to the ones that were fixed. Wrong answers can be introduced by interchange and segmentation faults are more likely caused by strip-mining since the loop bounds are modified. 
> 
> - Jan
Jan,
   So is it straight forward to implement similar fixes to the loop blocking
code?
                 Jack



More information about the Gcc-patches mailing list