gcc 2.95.1 bad performance on sparc

[Henrik Schmidt]

Hi,

I just installed gcc 2.95.1 on Solaris 2.6

Compared to our old gcc 2.8.1 its optimezed performance with the linpack
benchmark is much worse :

Old gcc 2.8.1 :

gcc -O3 -msupersparc :
Rolled Double Precision Linpack

    times are reported for matrices of order   400
      dgefa      dgesl      total       kflops     unit      ratio
 times for array with leading dimension of  401
       1.67       0.01       1.69      25489       0.08      30.12
       1.67       0.01       1.68      25556       0.08      30.04
       1.71       0.01       1.72      24990       0.08      30.72
       1.67       0.01       1.68      25559       0.08      30.03
 times for array with leading dimension of 400
       1.66       0.01       1.67      25750       0.08      29.81
       1.66       0.01       1.67      25705       0.08      29.86
       1.69       0.01       1.71      25183       0.08      30.48
       1.66       0.01       1.68      25634       0.08      29.94
Rolled Double  Precision 25559 Kflops ; 10 Reps 


New gcc 2.95.1 :

gcc -O3 -msupersparc :
Rolled Double Precision Linpack

    times are reported for matrices of order   400
      dgefa      dgesl      total       kflops     unit      ratio
 times for array with leading dimension of  401
       2.22       0.02       2.24      19208       0.10      39.96
       2.22       0.02       2.24      19181       0.10      40.02
       2.22       0.02       2.24      19208       0.10      39.96
       2.30       0.02       2.32      18559       0.11      41.36
 times for array with leading dimension of 400
       2.24       0.02       2.26      19031       0.11      40.33
       2.24       0.02       2.26      19001       0.11      40.40
       2.24       0.02       2.26      19033       0.11      40.33
       2.24       0.02       2.26      19008       0.11      40.38
Rolled Double  Precision 18559 Kflops ; 10 Reps 



Thats on an Ultra 1. Without optimization the speed of both versions 
is comparable ( ~7500 Kflops ). 

Any idea ? On a Supersparc architecture the difference is even worse.


Henrik
--
Henrik Schmidt                  Christian-Albrechts-Universitaet zu Kiel
Institut fuer Informatik        +49 (431) 5604 -  81 Fax or 75 Phone
Preusserstr. 1-9                http://www.ks.informatik.uni-kiel.de/~hbs/
24105 Kiel, Germany             email: hbs@ks.informatik.uni-kiel.de