To: djgpp AT delorie DOT com
Subject: Re: Interesting benchmark results
Message-ID: <19970311.190012.8503.0.chambersb@juno.com>
References: <199703110238 DOT MAA29246 AT solwarra DOT gbrmpa DOT gov DOT au>
From: chambersb AT juno DOT com (Benjamin D Chambers)
Date: Tue, 11 Mar 1997 21:55:56 EST


On Tue, 11 Mar 1997 12:38:55 +1000 (EST) leathm AT solwarra DOT gbrmpa DOT gov DOT au
(Leath Muller) writes:
>Ummm...basically, to my knowledge anyway, and by my understanding of 
>the
>doc's, your completely wrong. :)  Code always goes I could be - I've
been wrong before :)

superscalar if it 
>can
>because its _always_ in the cache (code _has_ to be in the cache to
>execute to start with, no?) and it will go superscalar if executed the
>first time too. 
>
>Branches can be predicted following some basic rules...
Yes, but wrong predictions incur a 4-5 cycle penalty, don't they?

>
>Have your read the Pentium Programmers Guide? Its really cool if you 
>haven't,
>and if you have, where does it say this stuff?
My library system doesn't have it yet (they don't even have the 486
programmer's guide yet! :), and I don't have the money for it.  I got the
above from Abrash's Zen of Code Optimization, and it spent a lengthy
chapter just on making code go superscalar.
Of course, I haven't had the chance to time things to know for sure
whether the first iteration goes superscalar - but my timings (when I
occasionally get access to a Pentium) show that they usually do (bear in
mind, these are iterations of thousands of times, so usually is a pretty
big number).  In either case, the first iteration wouldn't make too much
difference if you have a large loop value.

One point to make though:  I have often been wrong before.  It would be
no surprise if I was wrong again :)

...Chambers