Re: [tree-ssa] performance regression after yesterdays mainline merge

[Richard Guenther <rguenth at tat dot physik dot uni-tuebingen dot de>]

Fariborz Jahanian wrote:
> On Apr 2, 2004, at 1:09 AM, Richard Guenther wrote:
>
> > Hi!
> >
> > The automatic tester at
> > http://www.tat.physik.uni-tuebingen.de/~rguenth/gcc/monitor- summary.html
> >
> > caught a performance regression of the compiled tramp3d-v3 code. It's a
> > 7.43s to 7.64s (per iteration) aka 3% regression.
> >
> > As this is between a mainline merge I suspect something from there  
> > causing
> > this regression (I don't monitor mainline, maybe I should).  Usually
> > regressins in this test are caused by FP loop pessimizations (as nearly
> > all time is spent there), and as this is ia64 I suspected
> >
> > 2004-03-29  Fariborz Jahanian <fjahanian@apple.com>
> >
> >         * fold-const.c (fold): Reassociate multiply expression
> >         with an adjacent non-multiply expression to use
> >         architecture's multiply-add instruction.
> >
> > to be the cause of the regression, but just reverting the above patch
> > causes another regression to 7.69s - so I'm lost.  Diffs of assembly  are
>
>
>  From above numbers, it seems to me that above patch is actually helping
> performance. It lowers the time to 7.64s which would otherwise have been
> 7.69s. I don't have access to an ia64 machine but I think investigation  
> should
> start with why this benchmark regressed to 7.69s (without above patch).

Yes, I read the numbers the same.  But maybe I screwed something up 
there - I cannot find another possible candidate from the ChangeLog. 
And as I'm ignorant of ia64 assembly trying to figure out what happened 
from the assembly side proved unsuccessful.

Oh well - maybe tomorrow its magically fixed...

Richard.

> > not helpful, as all non-label differences seem to be complete different
> > compiled routines.
> >
> > But you may notice the raise in bootstrap time which might (didnt  check)
> > be caused by a regression in gengtype which consumes quite a lot of
> > bootstrap time on ia64.
> >
> > Richard.
> >
> > --
> > Richard Guenther <richard dot guenther at uni-tuebingen dot de>
> > WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/