This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: gcc300 vs gcc295.3 benchmarking?
- To: ronis at onsager dot chem dot mcgill dot ca
- Subject: Re: gcc300 vs gcc295.3 benchmarking?
- From: Paolo Carlini <pcarlini at unitus dot it>
- Date: Tue, 17 Jul 2001 18:25:36 +0200
- CC: pete at ltoi dot iap dot physik dot tu-darmstadt dot de, gcc at gcc dot gnu dot org
- References: <15188.23546.812869.404242@ronispc.chem.mcgill.ca>
Hi,
David Ronis wrote:
> While 3.0 shouldn't give worse results than 2.95.3, note that stack
> alignment only works as well as your libraries and startup modules are
> aligned. Most installed libc's aren't, and hence, no matter how well
> 3.0 maintains alignment, chances are your code isn't aligned.
I find your observation very interesting. I learned the hard way what you
describe: http://gcc.gnu.org/ml/gcc/2001-07/msg00177.html
However, I also learned (during some investigations carried out in
collaboration with Peter Schorsch) that compiling with an appropriate
-mpreferred-stack-boundary may help, perhaps does not restore the optimal
behaviour, but helps. For instance, compiling that benchmark ("flops")
with -O2 -mpreferred-stack-boundary=2 -fomit-frame-pointer I managed to
beat VC6 :-)
Why?
Cheers,
Paolo.