Message-Id: <3.0.32.19990708113153.010d2380@pop.xs4all.nl> X-Sender: diep AT pop DOT xs4all DOT nl X-Mailer: Windows Eudora Pro Version 3.0 (32) Date: Thu, 08 Jul 1999 11:31:59 +0100 To: pcg AT goof DOT com, pgcc AT delorie DOT com From: Vincent Diepeveen Subject: Re: K7 potentials Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Reply-To: pgcc AT delorie DOT com At 10:10 PM 7/7/99 +0200, you wrote: >On Wed, Jul 07, 1999 at 03:23:55PM +0100, Vincent Diepeveen wrote: >> I'm also looking forward to a motherboard that supports 4 or 8 K7s. >> Considering its expected selling prize, and the fact that we finally >> get competition for the Xeon now, this really will be a big step forward! > >I believe this will be just as expensive as their intel counterparts. And >worse, such an OpenPIC board would be a totally new design, with totally >new implementation errors ;) But my opinion doesn't really count anything >here ;) It might take some time here we agree, but there is COMPETITION! Competition is better than not having competition! >> For the program crafty (crafty is in some official benchmarks by >> the way, so it makes sense to test it; the command 'bench' >> gives you after a few minutes a benchmark in crafty; ftp ftp.cis.uab.edu > >However, you need to modify crafty a bit to get deterministic results >(crafty varies moves). > >> for crafty. linux versions of crafty are compiled using pgcc) >But only with -O, as this gives the best results with long longs (sadly). Yeah bob told me that. i wonder why. what's -O2 doing above -O? Note that crafty has big bunches of inline assembler, so perhaps that's the reason? >> See specs Merced, note that merced has quite some >> registers; work to do for the compilerfreaks! > >I never get tired of announcing the EPIC page at >http://www.goof.com/pcg/epic/, which offers quite a bit information on >that subject. page not found! >> When compiling my program for some RISC processors i get the impression >> registers don't get used very smartly (gcc for the alpha > >gcc is especially bad at using registers, yes. Especially bad on x86, >but still not perfect on other architectures. There is big room for >improvements. > >> So i guess compilers not using the major part of the merced registers >> will be commercial suicide! > >There are quite a bit larger problems with the merced than just using >registers. You need a totally different model, and heavy data as well as >control speculation. I still can't wait for a compiler using all those instructions which now don't get used. For example just using Pentium pro instruction already now would give huge speedup, like 20% (internal intel compiler normally is 20% slower, but using pentium pro instructions it's suddenly only 2.5% slower than msvc!) >-- > -----==- | > ----==-- _ | > ---==---(_)__ __ ____ __ Marc Lehmann +-- > --==---/ / _ \/ // /\ \/ / pcg AT goof DOT com |e| > -=====/_/_//_/\_,_/ /_/\_\ XX11-RIPE --+ > The choice of a GNU generation | > | > >