This is the mail archive of the
fortran@gcc.gnu.org
mailing list for the GNU Fortran project.
Re: Polyhedron performance regression
- From: "Paul Richard Thomas" <paul dot richard dot thomas at gmail dot com>
- To: "FX Coudert" <fxcoudert at gmail dot com>, "Uros Bizjak" <ubizjak at gmail dot com>
- Cc: stevenb dot gcc at gmail dot com, richard dot guenther at gmail dot com, paolo dot bonzini at lu dot unisi dot ch, fortran at gcc dot gnu dot org
- Date: Mon, 13 Nov 2006 18:56:21 +0100
- Subject: Re: Polyhedron performance regression
- Domainkey-signature: a=rsa-sha1; q=dns; c=nofws; s=beta; d=gmail.com; h=received:message-id:date:from:to:subject:cc:in-reply-to:mime-version:content-type:content-transfer-encoding:content-disposition:references; b=o/nDJlOOlcu4IU46q+EgOA/dMK5ZSyAA5bboPIHJjSYwjd8mMlz2mD7SQQ07WXsa+SyXiMw/s7VKP+lP46wU4ZITaiXIBEwTkEe5+eZhddb7A3M2JbVd3+VOiE1bjvmxoneztcB+5BwTkHA1MY3dijGRvCrYJz2KcoTWqUAp3qI=
- References: <339c37f20611120416j32a7798ft318d63b604d18a56@mail.gmail.com> <FAE7FB3C-8DC3-4DAE-95A4-565D1B9B027C@gmail.com>
FX,
I'm ashamed that revision 118492
(http://gcc.gnu.org/viewcvs?view=rev&revision=118492), which
introduced this performance regression, is the MOD/MODULO patch by
Paul Thomas and myself.
I realised from looking at the code for ac.f90, whilst on the plane last
night, that this must be the case. ac.f90 hinges on the random number
generator, which uses MOD.
I don't think that we should be ashamed about it. 32bit systems are
fine; I checked the Polyhedron suite before submitting and was
disappointed that it seemed to have little or no effect, other than
correcting the original PR!
Evidently there is something not very clever with the 64bit library.
Can Uros advise, please? In fact, is there a -ffast-math version
available or is it resorting to the ordinary library version? Detecting
the presence or not of the fast version, was what stopped me from
implementing this patch for so long.
Still one thing I don't understand is the difference between Paul
results and mine for the fwprop: it looks like it impacted with -O3
-ftree-vectorize -fno-fast-math.
No, I did not understand it either. However, I was in a burning hurry
and had to resort to using bash > time ./a.out to do the comparison.
Maybe the figures are too dependent on whatever background stuff was
happening.
We have to make a decision quickly as to whether to wait for the library
to be put right, if it can be, or to comment out the platch for the time
being. Please note that I am unable to do a thing to trunk or 4.2,
since the firewall of the site that I am at will not have anything to do
with ssh.
Regards
Paul