long long performance

[Rob Willis]

This may not be the best place to post this comment...  Please tell me
where to redirect if necessary.

I've been running some benchmarks of the basic operations ( inc, add,
mult, assign, static_cast ) for different numeric types (long, double,
long long).

Most of the results are what i expected.  The really bizarre result was
for the multiply op for "long long", which was 6 times slower than that
for a long or double multiply.  This seemed like an excessive
performance hit - can anyone explain why this might be??  The "long
long" inc, add, and assign operations were okay - performance was
somewhere between that for just normal longs and doubles.  The other
anomally, casting a double to a "long long" took even longer than the
multiply op!!  Has anyone seen similar results?  Are there standard
benchmarking tools out there? - I built a custom tool for testing these
runtimes.

My system:  P3/600 w/Redhat 6.2.  I'm using an older snapshot of the
compiler and library: egcs-20000612, so this problem may have been fixed
more recently.

My configure options:
./configure --enable-libstdc++-v3 --enable-long-long
--enable-cshadow-headers --enable-namespaces --enable-threads=posix 
--prefix=/usr/local

I compiled my benchmark without any optimizations.  

I was considering taking advantage of the "long long" type, but now i'm
not sure if that is a good idea.  

-Rob

-- 
Rob Willis
rob@e-critical.com