Mail Archives: pgcc/1999/05/11/18:31:17
Marc Lehmann wrote:
> > True, it is almost as close. -O6 however, is slower than -O2 on my K6-III. I
> > went back and rebuilt XFree with -O6 to see if it made any difference, and ran
> > 'x11perf' through most of the tests so that I could compare it against what I
> > had from previous runs. Attached is the output from 'x11perfcomp -ro' showing
> > the relative performance of a 'stock' RH5.2 XFree86, a '-march=k6 -O2' version
> > compiled with pgcc, and a '-march=k6 -O6' version compiled with pgcc. The
> > k6/O2 version is a bit faster for some things, slower on others. However,
> > from seeing the output of the k6/O6 version, I can't see any improvement.
>
> It would be interesting to find out which optimizations cause the
> slowdown. Its much work to do this, however (and xfree is not a good
> candidate for optimizing, at the code often is optimized for specific
> hardware timings, and "better" code can result in worse memory performance
> for the todays machines). I'd rather note that xfree86+k6+-O6 is not good,
> maybe find out what it is, and benchmark some other program, like gimp or
> povray ;)
I'll try a few other things out here and see which ones it seems to make a
difference on and which ones don't seem to have much boost in them. I tried
doing one test compile with 'gzip' as '-O6 -march=k6' to see what that one
did, and got around a 5% improvement when compressing a 15MB file this time.
Definately better than the degredation that I saw with '-O2 -march=k6'
previously.
> > Don't take this as a slag, I could totally understand K6 support not being as
> > widespread or ferverously developed as standard Pentium support. I'm really
>
> Its simple: I've never touched a K6, thats one of the problems ;)
Understandable. If my box here can be of much use as a test machine for
trying out any other K6 tweaks that come along with time, I'd be willing to
take them for a run...
> > just more curious to find out if the results that I'm seeing are more or less
> > what people expect to see out of these options.
>
> A significant slowdown is definitely not expected, but there might be bugs...
True, as is the case with any software. I'll see what I can fiddle with that
might help figure out which ones of the options that are enabled are actually
making a difference. Will try to get together a Perl script to run it through
some combinations of enabled optimizations and see which ones are actually
making a difference and which ones seem to degrade performance here. When
I've finally got some results on that together I'll send them off to the list
here, although it may have to wait until I get back from Europe depending on
how things go this week (note that I'll be gone for until mid-June). Will try
to get it off sooner though.
--
Graham TerMarsch
// -----------------------------------------------------------------
// In God we trust; all else we walk through.
// -----------------------------------------------------------------
- Raw text -