This is the mail archive of the fortran@gcc.gnu.org mailing list for the GNU Fortran project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Perfomance regression


On Tuesday 05 April 2005 09:13, Jerry DeLisle wrote:
> Steve Kargl wrote:
> >>Does anybody have ideas what can affect performance of attached test in
> >>these conditions?
> >
> > Disable hyperthreading.  It's a piece of crap, but probably
> > isn't important here.
> >
> > Is this code integer or floating point intensive?
> >
> >>From what I know about xeon cpu's, you should have bought opteron.

I agree with you of course, opteron rules ;-)  But I don't think
hyperthreading is to blame here.
(Did you know, btw, that the intel compilers can use hyperthreading
for dynamic prefetching?  It can be used for interesting things like
that...)

> Need to instrument the sample.f program and see if the computed results
> are the same for g77 and gfortran.  The sample program has some long
> complicated expressions with many continuation lines.  Might be hitting
> a limit on string length of the expressions, or the optimizers are just
> not handling it.

It could very well be a register presure issue or something like
that, if you have really large expressions.  We have this problem
in SPECfp mgrid as well.  In a very large expression, we move a
lot of loop invariant code out of inner loops in the tree passes,
only to find out that your whopping 6 registers on x86 are simply
not enough.

It would be great if you can find out which subprogram is slowed
down the most, and can create a test case from it...

Gr.
Steven


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]