Performance loss with gfortran on 64-bit Linux

[Uros Bizjak]

Hello!

> As Steve hints, some -pg results might help us guess why you 
> apparently did not find it useful to turn on any optimizations like 
> -O2, -funroll-loops, or -ftree-vectorize, even though you chose 
> extremely aggressive settings for commercial compilers.

 From the -static profile runs, it looks like some kind of a library 
problem to me, as on the top of 64bit profiled run (run-pg.64s) we have:

 26.82    279.89   279.89                             fesetenv
 19.00    478.20   198.31                             fesetround
 15.24    637.24   159.03                             __ieee754_expf
 10.30    744.73   107.49                             __ieee754_powf
  9.35    842.30    97.57                             feholdexcept
  4.91    893.59    51.29     4000     0.01     0.02  irmars_
  3.24    927.43    33.84                             __ieee754_exp
  3.01    958.88    31.45                             expf

In 32bit run (run-pg.32s), following profile can be obtained:

 45.98    140.38   140.38     4000     0.04     0.04  irmars_
 20.22    202.09    61.72                             __ieee754_powf
 15.25    248.65    46.56  1236002     0.00     0.00  gaus8_
  9.61    277.98    29.33     4000     0.01     0.01  co2emissvt_
  3.35    288.21    10.23                             powf

Attached to this message, please find profiled runs for 32bit (run-pg.32s)

FFLAGS = -ffast-math -frecord-marker=4 -O3 -static -pg -m32

and 64bit (run-pg.64s)

FFLAGS = -ffast-math -frecord-marker=4 -O3 -static -pg -m64.

on athlon_64-linux-gnu.

HTH,
Uros.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: profiles.tar.gz
Type: application/x-gzip
Size: 7646 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20060723/90ba6b51/attachment.bin>