Message-Id: <199805131737.TAA19543@router.dresearch.de>
From: "Michael Beck" <beck AT dresearch DOT de>
To: "beastium" <beastium-list AT Desk DOT nl>
Subject: Re: pgcc vs AMD K6
Date: Wed, 13 May 1998 19:39:04 +0200
MIME-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 7bit
Sender: Marc Lehmann <pcg AT goof DOT com>
Status: RO
Content-Length: 1587
Lines: 34

> NB: the results for the p6 was done with the snapshot pgcc-980406, and
> the flags -mamdk6 -O9 -funroll-all-loops -fexpensive-optimizations.
> The flag -mamdk6 for the p6 is not a joke ! I usually get the better
> results with this flag, sometime it's 15%-20% faster !!! I don't know if
> someone else also observed this behavior, and I don't know it this is
> relevant for the "Pentium team". I also don't know why -O9 get a little
> improvement on the result. Note also that the flag -pentiumpro is a
> disaster for my p6 ...(Yes, I'm sure that there is a  p6 in the
> computer, and the motherboard is a QDI Legend V)

Yes... First you should know, that a PentiumPro and your 'P6' is mostly the
same,
Intel added MMX instructions, a twice-as-large-but-twice-as-slow second
level cache
and segment register caches (these makes 16bit code faster !).
While the Pentium has very hard restriction how to combine instructions for
using
both pipelines, P6 (and this means PPro too) work in a fully different
manner.
A K6 is more a pentium than a P6, but has register renaming and some more
features
that allow using it's both pipelines more efficiently. So specifying
-mamdk6 generates code
that works on P6 good too, but may be bad on a P5. However, optimizing
specially for P6 
architecture should gain even more. Try -mpentiumpro -fno-risc, this was at
least some times 
ago needed for best P6 code.

cheers,
--
Michael Beck,                        email: beck AT dresearch DOT de 
DResearch GmbH, D-10319 Berlin,  Otto-Schmirgal-Strasse 3
Fon.: +49 (030) 515932 224 Fax.: +49 (030) 515932 299