Old loop optimizer in 4.0?

[Steven Bosscher]

On Saturday 25 September 2004 17:14, Zdenek Dvorak wrote:

> Below are some benchmark comparisons from x86_64.  Base is -O2, peak
> is -O2 -fno-loop-optimize -fmove-loop-invariants (i.e. old loop
> optimizer disabled, loop invariant motion in new rtl loop optimizer
> enabled).


Heh.  Nice numbers.  I got some too!  :-)

I independenty tested something similar on a POWER5 box a few days ago:

cflags base: -O2 -mtune=power5
cflags peak: -O2 -floop-optimize2 -fmove-loop-invariants -mtune=power5

Total time for base compilation: 238 s
Total time for peak compilation: 239 s

Total: Base: 3255829 bytes
Total: Peak: 3295333 bytes

                     peak/base
   168.wupwise       1.017
   171.swim          1.001
   172.mgrid         1.136 (!)
   173.applu         1.011
   177.mesa          1.020
   179.art           0.868 (!)
   183.equake        1.000
   188.ammp          1.002
   301.apsi          0.958
   MEAN              0.995

The other Fortran tests failed because the runtime couldn't be found,
my bad :-/

If we figure out why the old loop optimizer helps art so much, there's 
not much reason from the ppc performance point of view to keep the ugly
beast around.

Gr.
Steven