Bug 23164

Created attachment 16225 [details]
This source shows the bug (I hope). If not, it shows SOME bug.

I was just about to submit this same bug as occurring in gcc-2.96.81 (standard
RHL v7.1) and gcc-2.96.85 (Rawhide).

In fact the bug is even easier to reproduce: just replace the default
"-march=i386 -mcpu=i686 -O2" with "-march=i386 -mcpu=k6 -O2".  That's all it
takes to generate an ICE while compiling file ~/gcc/libgcc2.c (in function
__divdi3).

Why, after 5 months, is this bug still marked as "New"?

Because it is not reproduceable.
I've tried it back in January with both -march=k6 and -mcpu=athlon build
and it worked, today I did the same (patching gcc.spec so that it set
RPM_OPT_FLAGS="-O2 -march=i386 -mcpu=k6") and it built without problems.
Can you reproduce it always on the same file?

I can reproduce the problem with the same source file (though the function
reported in the error msg varies).

These instructions describe the attempt to build the GCC from Rawhide (latest
publicly available) on a newly-installed RedHat v7.1 system.  This system is
based on an AMD K6-2/333MHz with 160MB RAM.  Installation type was Custom (I
opted for "Install Everything").

Do this:

  1. rpm -ivh ~/gcc-2.96-85.src.rpm
  2. In the spec file, replace $RPM_OPT_FLAGS /w "-march=i386 -mcpu=k6 -O2"
  3. rpm -bc /usr/src/redhat/SPECS/gcc.spec

To get this result:

for d in libgcc; do \
  if [ -d $d ]; then true; else mkdir $d; fi \
done
if [ -f stmp-dirs ]; then true; else touch stmp-dirs; fi
./xgcc -B./ -B/usr/i386-redhat-linux/bin/ -isystem
/usr/i386-redhat-linux/include -O2   -DIN_GCC    -W -Wall  -Wwrite-strings
-Wstrict-prototypes -Wmissing-prototypes  -Wno-long-long -O2 -march=i386
-mcpu=k6 -O2 -isystem ./include  -fPIC -g1 -DHAVE_GTHR_DEFAULT -DIN_LIBGCC2
-D__GCC_FLOAT_NOT_NEEDED  -I. -I../../gcc -I../../gcc/config
-I../../gcc/../include  -DL_muldi3 -c ../../gcc/libgcc2.c -o libgcc/./_muldi3.o
../../gcc/libgcc2.c: In function `__muldi3':
../../gcc/libgcc2.c:199: Internal error: Segmentation fault.
Please submit a full bug report.
See <URL:http://bugzilla.redhat.com/bugzilla/> for instructions.
make[3]: *** [libgcc/./_muldi3.o] Error 1
make[3]: Leaving directory
`/usr/src/redhat/BUILD/gcc-2.96-20000731/obj-i386-redhat-linux/gcc'ptmdlPioInstance_t
make[2]: *** [libgcc.a] Error 2
make[2]: Leaving directory
`/usr/src/redhat/BUILD/gcc-2.96-20000731/obj-i386-redhat-linux/gcc'
make[1]: *** [bootstrap-lean] Error 2
make[1]: Leaving directory
`/usr/src/redhat/BUILD/gcc-2.96-20000731/obj-i386-redhat-linux/gcc'
make: *** [bootstrap-lean] Error 2
error: Bad exit status from /var/tmp/rpm-tmp.33662 (%build)

This problem is all too reproducable.  Please let me know if I can provide any
addional info.

Please disregard that trailing text "ptmdlPioInstance_t".  It somehow snuck into
my cut 'n' paste of the GCC error output.  It is unrelated to the problem being
described.

Yet another clarification.

Above I wrote that while the ICE ocurred in the same file, the function in which
it occurres varies.  The latter part of this statement is not true.

Given the *same* set of source files, the ICE occures in the same place.  I saw
it in 2 different functions because I used 2 different sets of source, first GCC
from RHL v7.1, now in the GCC SRPM gotten from Rawhide.

My experience: given invariant source files, the ICE will also be invariant. 
Now that I'm testing exclusively /w gcc-2.96-85.src.rpm, the ICE is always
reported as being in __muldi3().

Created attachment 21205 [details]
Feel the bug, man!!!

A patch to GCC v3.0 has been submitted to address this problem (the PR2767
referred to is in GNATS): http://gcc.gnu.org/ml/gcc/2001-06/msg01357.html

I haven't tried it yet so YMMV.

Sorry, forgot to update.
This was fixed in 2.96-93 and above by simply not using loop insns anywhere
(similarly to the 3.0.x solution).

As noted, worked around for 2.96 and 3.0; fixed for real at least
as of gcc 3.2.  Confirmed works in gcc-3.2.3-20, gcc-3.3.2-1, 
gcc 3.4.2 20040907, and gcc 4.0.0 20040930.