Re: K6 Optimized Debian Binaries
On Wed, Feb 16, 2000 at 08:28:11PM -0800, Ross Boylan generated a stream of 1s and 0s:
> This reminds me: what are the best compiler options for AMD chips (I have a
> K6-2)? I principally thinking of gcc or egcs. If they have AMD specific
> options great, otherwise, what type of Intel is it best to pretend to be?
>
> pgcc is news to me. I take it this is an AMD targetted compiler. I would
> have thought AMD support would just be folded into gcc. Is there a reason
> it's not? (Or is it just slower and safer)?
>
>
> --
> Unsubscribe? mail -s unsubscribe debian-user-request@lists.debian.org < /dev/null
>
Wrong. GCC 2.95.2 (formerly EGCS) does optimize for K6, Pentium and PPro, try -march=k6 -O3. What does Mike Lehmann do? He takes GCC code as it is, and patches
with his own Pentium/PPro/K6 optimizations in addition to those in gcc. What's the result? Well he says on average you can expect a 5% improvement with Pentiums/Pentium Pros. Last time I checked he doesn't have a K6 machine, so the optimizations are based on some documentation he has. With some code (like compression i.e. gzip/bzip2) improvement is up to 30 percent. With PGCC improvement is mostly with integer code. It actually makes floating point code worse. What truly sucks about K6 is it's sensitivity to memory alignments. This means some software might perform differently every time it's ran, sometimes slow, sometimes fast. PGCC primarily targets P5/P6. Some distributions compiled with pgcc: Mandrake, Stampede, Eunuch. Rarely, but still happens pgcc miscompiles code. I think the main thing about PGCC is that performance difference is NOTICEABLE. If you get a benchmarking software like nbench, and compile it with PGCC you'll see a noticeable improvement. Distributions like Mandrake do run snappier on Pentium but slower on other CPUs because of those optimizations. AFAIK all other distros like Debian are uhm, if optimized at all are for 386! More info: http://www.goof.com/pcg
--
Get the truth or risk frying your brains! --> www.truthinlabeling.org <--
Reply to: