GCC and Fortran eat GCJ for lunch at number crunching

[Mark Wielaard]

Hi,

On Thu, 2003-01-02 at 16:32, Per Bothner wrote:
> Erik Poupaert wrote:
> > http://www.coyotegulch.com/reviews/almabench.html
> > 
> > """
> > Another surprise was delivered by the weak showing of the GNU gcj compiler.
> > For numerical work, at least, there doesn't look to be any reason to use gcj
> > (which is only a partial Java implementation) unless you need a bit of extra
> > performance in your Pentium III-based Java code. Perhaps I'm missing
> > something obvious; clearly this result requires further investigation.
> > """
> 
> A couple of people have pointed out on slashdot that
> compiling with --fast-math makes a huge difference.

There is also a difference between the last released 3.2 and the 3.3
from CVS. With only the following flags "-march=athlon-xp
-mcpu=athlon-xp -O3" there was already the following difference for his
almabench on my machine:

3.2: 92 seconds
3.3: 66 seconds

So when 3.3 is released it should already have a positive impact on this
benchmark and hopefully the test will be repeated then.

Cheers,

Mark