Pentium 4 Optimization Complimenting icc 5.0

[Tim Prince]

On Wednesday 06 February 2002 06:36, Walton, Shane wrote:
> I have been benchmarking Dual 1GHz Pentium 3 vs. Dual 1.7GHz Pentium 4
> Xeon systems with
> known UNIX/LINUX benchmark suites such as "ubench" and "unixbenchmark".
Probably not many people here familiar with redhat's own gcc versions and 
with those benchmarks.  I'm neither, but my news feed hasn't brought me any 
expert answers to your post.
> My findings are
> almost unbelievable.  The P3 system outperforms the P4 system by 20%.
>
> I have my hands on an Intel C++ Compiler (5.0) for Linux optimized for
> Pentium 4.  Rebuilding
> the benchmark utils with this compiler skewed the results so that the P4
> was 30% better than
> the P3 system.
>
> All this comes down to using a "gcc" compiler vs. the Intel "icc".  Is
> there something that I don't
> know about gcc, such as a patch or optimization flags that would give me
> the expected performance
> from a gcc built binary on a P4 system?  If not, is there any plans to
> add such support?  Thanks.
Why not try gcc-3.1,with -march=pentium4 and whatever other options you 
favor, or at least make sure that you are running a compiler built so that 
-march=pentiumpro invokes costs so that 3 adds are preferred over a shift, 
some larger number of adds over imul,...?  You can't expect a gcc about which 
little is known by some of us except that it predates P4, to keep up with a 
compiler which has P4 switches.
>
> Here is the basic config for testing environment:
>
> RedHat 6.2 and RedHat 7.2
I've read here recently that rh6.2 does not account for SSE2 registers in 
context switching, so you would not be safe using P4 options with it.
> egcs 2.91.66 (egcs-1.1.2 release)
Wasn't that some time ago?
> gcc 2.96
a redhat specific version
>