This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

g77/x86 unroll-loops performance


The performance of -O2 has improved significantly in recent
snapshots, emphasizing some of the cases where the compiler
stumbles over unrolling at -O2.  Is there an adjustable switch
which would reduce the tendency to unroll when good
performance is obtained without unrolling?  Kernel 18 is really
more extreme than it appears, as big gains from unrolling one
part are offset by big losses in the other parts.  Pentium II shows
similar comparisons, except that it saturates L1 cache
performance in several kernel tests.

 *********************************************
 THE LIVERMORE  FORTRAN KERNELS "MFLOPS" TEST:
 *********************************************
Run on Pentium Pro 200 Mhz
Compiled on egcs-19990502/g77 -O2 -malign-double
-march=pentiumpro

     -funroll-loops  not unrolled

 KERNEL MFLOP/SEC
------ ---------
  1      86.081       82.372
  2      83.863       77.847
  3     104.491       93.196
  4     132.464       94.817
  5      47.104       46.898
  6      78.903       61.202
  7      74.986       94.291
  8      74.830       79.672
  9      71.458       87.320
 10      18.145       22.260
 11      59.176       61.011
 12      71.013       62.006
 13       6.022        6.012
 14      15.843       17.106
 15      20.360       20.658
 16      34.316       34.711
 17      44.569       40.450
 18      58.901       60.349
 19      48.101       47.506
 20      25.388       27.017
 21     107.685       85.799
 22       9.608        9.631
 23      99.967      100.153
 24      20.315       21.297
Dr. Timothy C. Prince
Consulting Engineer
Solar Turbines, a Caterpillar Company
alternate e-mail: tprince@computer.org


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]