Re: IRA performance regressions on PPC

[Luis Machado <luisgpm at linux dot vnet dot ibm dot com>]

On Fri, 2008-09-05 at 12:36 -0400, Vladimir Makarov wrote:
> Luis Machado wrote:
> > Hi Vladimir,
> >
> > I was just going through some benchmarks on PPC and noticed that your
> > patch from 08/26 (http://gcc.gnu.org/ml/gcc-cvs/2008-08/msg01152.html)
> > caused a significant regression on both facerec (~17%) and applu (~4%)
> > for 64-bit PPC.
> >
> > There are other degradations that i'm still working on isolating the
> > cause, just to give you a heads up on the problem.
> >   
> 
> Thanks for testing IRA, Luis.  Could you give me more details:

Yes, of course.
> 
>   What machine you are using for this?

This is a Power6 4.7Ghz (altivec supported)

>   What options (especially march or mtune) you are using?  IRA is very 
> sensitive to correct times of ld/st/moves in machine description.

I'm currently using two tuning setups.

base flags: -m64 -O2 -mcpu=power4
peak flags: -m64 -O3 -mcpu=power4 -ffast-math -ftree-loop-linear -funroll-loops -fpeel-loops

>   What is overall IRA regression on SPEC2000?

I don't have that information on this box yet. But i'll have it soon and
will let you know. Right now i only focused on those two degraded
benchmarks.

> You could use the same version of the compiler with IRA (default) and 
> old RA (-fno-ira).

Thanks for the tip. Is it a good idea to go through the ira-merge branch
as well? Or would this suffice?

Thanks,
Luis