C optimization tests - almabench results

[Uros Bizjak <uros at kss-loka dot si>]

Hello!

I was trying to analyze, what is wrong with gcc's trigonometric 
functions in alma benchmark. However, I can't confirm long execution 
times with alma benchmark [alma benchmark was downloaded from 
http://www.coyotegulch.com/download/almabench-1.0.1.tar.gz.

With CVS mainline gcc (3. May 2004), my timings on 3.2G pentium4, 2GB 
RAM are:

[uros@omen5 cpp]$ g++ -O3 -march=pentium4 -lm almabench.cpp
[uros@omen5 cpp]$ time ./a.out

real    0m25.875s
user    0m25.668s
sys     0m0.053s

[uros@omen5 cpp]$ g++ -O3 -ffast-math -march=pentium4 almabench.cpp
[uros@omen5 cpp]$ time ./a.out

real    0m22.616s
user    0m22.564s
sys     0m0.035s

[uros@omen5 cpp]$ g++ -O3 -ffast-math -D__NO_MATH_INLINES 
-march=pentium4 almabench.cpp
[uros@omen5 cpp]$ time ./a.out

real    0m15.215s
user    0m15.172s
sys     0m0.041s

I don't think that 15% faster clock can produce such a difference...

BTW: With -O3 -ffast-math -D__NO_MATH_INLINES, these instructions should 
be generated:
objdump -S a.out | grep fsincos
80487ae:       d9 fb                   fsincos
80487ce:       d9 fb                   fsincos
8048804:       d9 fb                   fsincos
8048839:       d9 fb                   fsincos
804889a:       d9 fb                   fsincos
80488ea:       d9 fb                   fsincos
8048945:       d9 fb                   fsincos
8048950:       d9 fb                   fsincos
8048959:       d9 fb                   fsincos
8048981:       d9 fb                   fsincos

Uros.