On Tuesday 18 November 2003 05:44, Jason Stubbs wrote:
> On Tuesday 18 November 2003 09:57, Chris Graves wrote:
> > interesting article about selecting gcc optimizations (found on
> > OSNews.com).
> >
> > http://www.coyotegulch.com/acovea/index.html
>
> Quite a good article! Shame it doesn't apply to us. I can't remember
> where I read it, but the gcc crew were talking about doing an
> overhaul (or at least an investigation) into -O1 -O2 and -O3. As
> previous informal benchmarks done by Gentooists show, with the
> current GCC compilers (including 3.3) -O2 can often produce faster
> code than -O3. This is meant to have been addressed in 3.4.

And this applys to Gentoo expecially, because the whole system is 
compiled with those flags. In the tests the only ones that was compiled 
with the said flags were only the test applications.

To make real optimizations one should first compile glibc with those 
different optimizations and then for these glibc versions the 
applications with different optimizations. Which makes it even more 
complex problem.

Oh and by the way the testing was done in steril system. How does the 
comilation of those unrollinbg affect on desktop systems where there 
are multiple programs running? So cache trashing comes even easier and 
thus slows down every process...

So now we need multiple hard disks, which every one has the exact same 
Gentoo system, but compiled with different optimizations. Then simply 
make the system boot so that the system is ran from our 2GB of memory 
so that different HD speeds doesn't affect the tests. :)

> On the other hand, the code for his tester can be downloaded and then
> modified only cover gcc33/32 optimizations. Anybody know of a few
> noninteractive benchmark programs that could be tested with? I'm
> happy to run some tests and publish results for AthlonXP using gcc33.

The most distracting point in his teest was the fact that all tests were 
running in isolated environment ie no other running processes. This 
lets the whole 11k instruction cache of P4 for the one process, so that 
the penalty of unrolling and inlining raises quite a bit from normal 
desktop systems.

I thus think that usefull tests can only be made in an environment 
resembling the real use situation.

So the test made would suit for compile or rendering farms, but not for 
desktop environment.


--
[EMAIL PROTECTED] mailing list

Reply via email to