Re: loop optimizations, mark 9

[Toon Moene <toon at moene dot indiv dot nluug dot nl>]

Yesterday, I wrote:

>  I'm now running a snapshot CVS'd 199805222000UTC with
>  your mark 9 patches applied.

[ It bootstrapped successfully ]

I actually ran a complete HIRLAM run with it:

0SUPOBS TOOK :       19.85659790039
0DATACH TOOK :      9907.8935546875		*)
0ANAEVA TOOK :     11532.0419921875
0GRPEVA TOOK :     22830.1308593750
0HUMSUP TOOK :        1.65625000000
0DATACH TOOK :       561.8164062500
0HUMEVA TOOK :       289.7539062500
0GRPEVA TOOK :       385.2695312500
 PREPARATIONS TOOK         50.7206 SECONDS
 FORECAST TOOK        289.4115 SECONDS
 PREPARATIONS TOOK         54.0532 SECONDS
 FORECAST TOOK      32278.8711 SECONDS

Contrast with (snapshot of 199805210900UTC):

0SUPOBS TOOK :       19.66827392578
0DATACH TOOK :      9275.4941406250		*)
0ANAEVA TOOK :     11315.5214843750
0GRPEVA TOOK :     23142.7968750000
0HUMSUP TOOK :        1.60937500000
0DATACH TOOK :       581.2851562500
0HUMEVA TOOK :       289.7539062500
0GRPEVA TOOK :       385.1562500000
 PREPARATIONS TOOK         48.5590 SECONDS
 FORECAST TOOK        278.5487 SECONDS
 PREPARATIONS TOOK         53.5047 SECONDS
 FORECAST TOOK      32280.9922 SECONDS

Note the timing with *)  Definitely slower with your changes.

Basically, there are two things that are not well, at least on  
m68k-next-nextstep3:

1. I use the compile time option -fforce-addr, because that'll
   assure me of the optimal use of John Carr's alias analysis
   (private communication, May '96).
   However, when I compile `resid' with -O2 -fforce-addr, I clearly
   see that your stuff isn't effective (when comparing with -O2).

2. Rerunning loop optimisation still seems to be useful for my code;
   however, this also could be a result of (1.)

HTH,
Toon.