Preleminary Branch Profiling Results

[Andreas Jaeger <aj at suse dot de>]

Jan Hubicka got in the last weeks feedback based profiling for
branches running.  He's still working on improving the branch
heuristics and algorithms but for details better ask him yourself.

I've run SPECint2000 on a current CVS version (from 2 July) on my
Athlon 1.133 Ghz box.

The base runs used -O2 -march=athlon -malign-double.  I used the same
flags for the peak runs but did also the following (actually the SPEC
scripts did it for me after configuring it correctly):

- Add -fprofile-arcs as additional flags
- Run the programs on some "train" data to get profile information
- Rebuild the programs with -fbranch-probabilities as additonal
  compiler flag to use the profile information.
and then used those binaries for the peak runs.

The tests showed some small improvements for some tests, the biggest
improvement is for 176.gcc with a 6 % speed improvement.  For other
tests there's no improvement or improvement between 1% and 3%.

That's not much of an improvement yet.

On my Pentium III 500 I noticed four tests with a speedup of more than
5 % but those were done with an older GCC version.  So it might be
that the Athlon has much better branch heuristics, the compiler
changed in between - and definitly Honza will continue to work on
this.

The complete numbers are available at http://www.suse.de/~aj/CINT2000.045.pdf.

I'll redo those experiments later and let you know about further
progress.

Andreas
-- 
 Andreas Jaeger
  SuSE Labs aj@suse.de
   private aj@arthur.inka.de
    http://www.suse.de/~aj