benchmark result

[Mathieu Lacage]

On Tue, 2004-12-07 at 14:35 -0700, Tom Tromey wrote:
> >>>>> "Mathieu" == Mathieu Lacage <Mathieu.Lacage@sophia.inria.fr> writes:
> 
> Mathieu>  11.77    251.00   251.00                             Jv_CheckCast
> 
> We have a plan to remove redundant cast checks, see
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15525

added myself to CC list.

> 
> Eliminating checks when doing an array store will require more type
> machinery in the gcc optimizers.  We've talked about this a few times
> but I wouldn't expect anything before GCC 4.1.
> 
> Whether this would help with your program depends on how many of
> those checks are actually redundant.

any idea on how I might measure that ?

> 
> Mathieu>   0.75   2039.00    16.00                             _ieee754_pow
> Mathieu>   0.33   2056.00     7.00                             _ieee754_sqrt
> 
> Interesting to see these here.  I would have thought that at least
> sqrt would be turned into an instruction -- that was one hope from
> turning Math.sqrt calls into calls to __builtin_sqrt.
> 
> It was also interesting to see how low _Jv_InitClass was on the list.
> I won't worry as much about that one for a while :-)
> 
> 
> Benchmark information like this is very valuable.  I'd like us to find
> a way not to lose it.

I do not mind filing a bug report if you feel the need to. I will also
probably keep an eye on the evolution of the performance of that code
so, I might come back later with new results. Currently, I get some
really strange segfaults when running the code on "interesting"
scenarios so I decided to keep on using the sun jdk. 

As a side note, is it "normal" for me to see SIGPWR and SIGXCPU signals
when I run the program in gdb ? I did use the "handle" command to ignore
these and pass them to the application but it looks like something is
getting really confused since I randomly get a SIGABRT a bit later. Of
course Valgrinding the code is a bit useless with the GC behind the
scenes so I gave up on that issue.

The code I work on will be released as open source probably within 2 or
3 months at most. If you feel the need to look at the code, and since I
own the copyright on the code, I am sure we can arrange something to be
done earlier.

Mathieu
--