This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: gzip performance test
- To: David Edelsohn <dje at watson dot ibm dot com>, Joe Buck <jbuck at synopsys dot COM>
- Subject: Re: gzip performance test
- From: Jean Francois Martinez <jfm2 at club-internet dot fr>
- Date: Fri, 25 May 2001 08:05:21 +0200
- Cc: gcc at gcc dot gnu dot org
- References: <200105250206.WAA23388@makai.watson.ibm.com>
- Reply-To: jfm2 at club-internet dot fr
On Friday 25 May 2001 04:06, David Edelsohn wrote:
> gzip built with which options? gzip-1.2.4a CFLAGS seems to
> default to -O.
>
> I tried gzip-1.2.4a on powerpc-ibm-aix4.3.3.0. The results depend
> on the options:
>
> With -O2, gcc-3.0 prerelease is slightly faster, matching your results.
> With -O3 -funroll-all-loops, gcc-2.95 is slightly faster.
> With -O3 -funroll-all-loops -fstrict-aliasing, they're statistically
> identical.
>
> David
I certainly wouldn't want to use unroll-all-loops on a genral basis since the
increase in TLB and cache consumption plus swapping can degrade instead of
improving performance.
Also:
1) gzip spends a fair share of its time in glibc. If you don't use
different glibcs (through LD_LOAD_PATH) your results will be flattened
and could look not significative
2) Part of gcc 3.0 improvement will be due to additional optimizations.
Until -O2 is redefined to include them you will not be getting them so it
is unfair to compare -O2 with -O2
3) At least with the Byte benchmark the unholy gcc 2.96 provides code
significantly faster than gcc 2.95 so if gcc 3.0 id not better than gcc 2.95
it means it is worse than the evil gcc 2.96
JFM