Re: Some benchmarks comparing gcc-3.4.3 and 4.0.0 on PowerPC G5

[Steven Bosscher <stevenb at suse dot de>]

On Sunday 02 January 2005 08:27, Bradley Lucier wrote:
> I've put some benchmark results comparing compile and run times for C
> code generated by the Gambit-C Scheme->C compiler at
>
> http://www.math.purdue.edu/~lucier/GNATS/GNATS-18/results.html
>
> There are a fair number of tests where the compile times for 4.0.0 are
>  > 1.5 times the corresponding compile times for 3.4.3.  (Both were
> built without checking.)  The run-time comparisons are more difficult
> to interpret; these tests were run as much to compare -mcpu and -mtune
> options on G5 as to compare compiler performance.
>
> The rest of the directory has the raw data, but I'm not sure it will
> tell you much more than the table.

Three comments:

1) The 4.0 compilers are probably slower because all but one of the
   tree-ssa optimizations are enabled at -O1.  It would be interesting
   to see results for -O2.  We should see about turning off some of
   the expensive things in tree-ssa, such as iterating DOM, at -O1.

2) You should probably compare columns with matching flags.  Comparing
   "4.0.0 (no options)" with "3.4.3 -mcpu=G4 -mtune=G5" is not really
   useful IMHO.

3) You say "each compiler was built without checking".  Do you mean
   built as in "make" instead of "make bootstrap"?  I compared compile
   times of GCC4 built with GCC 3.4.2 vs. GCC4 built by itself, and 
   of GCC4 built with GCC 3.3.5 vs. GCC4 built by itself.  I found that
   GCC4 optimizes itself a lot better than those older GCCs.  So it is
   important to know if you did a plain "make", or a "make bootstrap".

Gr.
Steven