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]

i686 double alignment vs performance



I got the following geometric mean Mflops from Livermore
Kernels lfk537e.f (double precision), with my modifications.  Both
g77 and egcs are compiled with binutils-2.9.1 installed, with
-O2 -malign-double -march=pentiumpro -funroll-lloops so that I
believe the loop code alignments are reasonable and
comparable.

Compiler       PPro/200      P II/233

g77-0.5.23     14.8              50
egcs-0628     35.8              50
lf90-4.50d       42.5              55

So, the additional alignments currently generated in egcs must
give it a tremendous advantage on the pentiumpro, but the
pentium2 isn't so dependent on them.  If I compare egcs and
g77 without -funroll-loops, however, egcs often has an
advantage even on P II.  If I turn off -malign-double, performance
drops significantly in every case, but more so with egcs on PPro.

           To:                                              INTERNET - IBMMAIL


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