i686 g77 optimization regressions, egcs-20000619

Tim Prince tprince@computer.org
Thu Jun 22 16:20:00 GMT 2000

The following anomalies occur with the attached Livermore Kernel source
code, both on linux-glibc and cygwin:

erroneous result in kernel 18 at -Os -funroll-loops, correct with all
other option combinations tested
At -Os, the first loop of kernel 18, where the error appears to occur,
is unrolled by 2.
At -O2, this loop is unrolled by 4, with several pointer spills.
Tests with previous 2.96 versions showed an error in the first loop also
with the original LFK code.

Kernel 1 performance drops 20% from -O to -Os, even more from -Os
to -O2.  This is the largest contributor to a 10% overall reduction in
performance at -O2 from version 2.95.2.

Kernels 8 and 21 have lost performance at -Os from 2.95.2, but perform
well at -O2.

Kernel 13 has lost performance from 2.95.2 at all levels of

Tim Prince
-------------- next part --------------
A non-text attachment was scrubbed...
Name: lfk537g.f.gz
Type: application/x-gzip
Size: 72374 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-bugs/attachments/20000622/8d37abfa/attachment.bin>

More information about the Gcc-bugs mailing list