[Re: optimization/8412: Code generation problem with -march=k6

[Jim Bray]

  I think I'm seeing this same bug, with 
gcc version 3.2.1 20021207 (Gentoo Linux 3.2.1-20021207)
 I've noticed considerable instability with multimedia things including
xine, and have been studying an easily reproducible problem with
xmms. I'd assumed it was 3dnow related, but testing has zeroed me in
on a combination of -funroll-all-loops and -O[23].

  Here are my crude notes to date. Not yet a good bug-report,

The test is compiling and running xmms and doing a particular
operation with it. My system is a k6-2.

-finline-functions is used in all tests.

CFLAGS='-march=k6-2 -g -O1' works.
CFLAGS='-march=k6-2 -g -O1 -ffast-math' works.
CFLAGS='-march=k6-2 -g -O1 -ffast-math -funroll-all-loops' works.
CFLAGS='-march=k6-2 -g -O2 -ffast-math -funroll-all-loops' FAILS.
CFLAGS='-march=k6-2 -g -O2 -ffast-math' works.
CFLAGS='-march=k6-2 -O2 -ffast-math' works.

  The xmms configuration normally uses -funroll-all-loops, and
I was seeing consistent failure with any variant of -m{cpu,arch}=
{k6,k6-2} -O[23], which is not reflected above. The problem went
away when the k6-specific flags were removed. I did not start
getting any k6-specific build successes until I started dropping
below -O2 and -funroll-all-loops.

  I did some studying of this with gdb, which seemed to be zeroing me
into an area of code which became an unrolled loop.

 Bug I 'sent email to interested parties' from:

http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view%20audit-trail&database=gcc&pr=8412


-- 
Jim Bray <jb@as220.org>
-- 
Jim Bray <jb@as220.org>