Unreproducible ICEs on Cortex A15

[dann frazier]

Any significantly large build (mysql-5.5, the kernel) using gcc-4.7 will eventually fail due to an unreproducible ICE.

    [ 6%] Building C object strings/CMakeFiles/strings.dir/ctype-uca.c.o
    /tmp/buildd/mysql-5.5-5.5.29/strings/ctype-uca.c: In function 'create_tailoring':
    /tmp/buildd/mysql-5.5-5.5.29/strings/ctype-uca.c:8070:1: internal compiler error: Segmentation fault
    Please submit a full bug report,
    with preprocessed source if appropriate.
    See <file:///usr/share/doc/gcc-4.7/README.Bugs> for instructions.
    The bug is not reproducible, so it is likely a hardware or OS problem.

As the error message states, this is usually due to an OS or hardware defect and I have no hard evidence to the contrary. However, there is some evidence that points away from the hardware/OS.

1) The issue seems specific to gcc-4.7
PASS: gcc-4.6 (4.6.3-10ubuntu1)
FAILS: gcc-4.7 (4.7.2-2ubuntu1)
FAILS: gcc-4.7 (4.7.2-20ubuntu1)
PASS: gcc-snapshot (20130121-0ubuntu1)
2) There are no other signs of instability - gcc-4.7 is the only thing we see fail

On the other hand, this has only been seen on the unsupported exynos5-based arndale platform, so it could be a platform bug that only gcc-4.7 happens to tickle. If we can demonstrate this, we could reassign this bug to the arndale support project (http://launchpad.net/arndale) and ask for some linaro assistance. I could use some gcc-debugging help to make progress on that - e.g., tips on getting a debuggable coredump.

[James M. Leddy (jm-leddy)]

http://moxielogic.org/blog/?p=482

[Adam Conrad (adconrad)]

What kernel is this? Since we don't actually ship any A15 kernels in the archive, I assume this is a custom deal?

[Adam Conrad (adconrad)]

Also, is this a default ubuntu armhf toolchain being called with default options, or are you using a different -march/cpu/tune?

[James M. Leddy (jm-leddy)]

Hi Dann,

What are the specific reproduction steps? For example, are you re-compiling mysql from source, or using debuild on the ubuntu specific branch?

Also, to get a core on ubuntu you can follow these steps:

$ sudo sysctl -w kernel.core_pattern=core
$ ulimit -c unlimited
$ gcc...

Please include a build log as well.

[dann frazier (dannf)]

On Wed, Feb 06, 2013 at 08:36:54PM -0000, Adam Conrad wrote:
> What kernel is this? Since we don't actually ship any A15 kernels in
> the archive, I assume this is a custom deal?

The kernel is available in this PPA:
  https://launchpad.net/~arndale-team/+archive/ppa

git tree is here:
  http://kernel.ubuntu.com/git?p=hwe/ubuntu-quantal-arndale.git;a=summary

[dann frazier (dannf)]

On Wed, Feb 06, 2013 at 08:40:17PM -0000, Adam Conrad wrote:
> Also, is this a default ubuntu armhf toolchain being called with default
> options, or are you using a different -march/cpu/tune?

All defaults.

[dann frazier (dannf)]

On Wed, Feb 06, 2013 at 08:51:50PM -0000, James M. Leddy wrote:
> Hi Dann,
>
> What are the specific reproduction steps? For example, are you re-
> compiling mysql from source, or using debuild on the ubuntu specific
> branch?

Just rebuilding the package, e.g.:

  sudo apt-get build-dep mysql-5.5
  apt-get source -b mysql-5.5

> Also, to get a core on ubuntu you can follow these steps:
>
> $ sudo sysctl -w kernel.core_pattern=core
> $ ulimit -c unlimited
> $ gcc...

Yeah, tried that - and it wasn't working for me. I wondered if gcc was
doing something to block this, but I didn't investigate further.

> Please include a build log as well.

Will do.

[dann frazier (dannf)]

Here's a reduced reproducer:

ubuntu@s5:~$ while /usr/lib/gcc/arm-linux-gnueabihf/4.7/cc1 repro.c -quiet -march=armv7-a -mfloat-abi=hard -mfpu=vfpv3-d16 -O2 -fPIC --param ssp-buffer-size=4 -o - > /dev/null; do echo -n .; done
/home/ubuntu/mysql/mysql-5.5-5.5.29/strings/dtoa.c: In function ‘my_gcvt’:
/home/ubuntu/mysql/mysql-5.5-5.5.29/strings/dtoa.c:441:1: internal compiler error: Segmentation fault
Please submit a full bug report,
with preprocessed source if appropriate.
See <file:///usr/share/doc/gcc-4.7/README.Bugs> for instructions.

I started with the original compilation command (see below), created a preprocessed file using -E, then removed various flags after verifying I could still reproduce with them omitted.

// /usr/lib/gcc/arm-linux-gnueabihf/4.7/cc1 -quiet -I /home/ubuntu/mysql/mysql-5.5-5.5.29/builddir/include -I /home/ubuntu/mysql/mysql-5.5-5.5.29/include -imultilib . -imultiarch arm-linux-gnueabihf -D _FORTIFY_SOURCE=2 -D DBUG_OFF -D HAVE_CONFIG_H -D DISABLE_MYSQL_THREAD_H -D BIG_JOINS=1 /home/ubuntu/mysql/mysql-5.5-5.5.29/strings/dtoa.c -quiet -dumpbase dtoa.c -march=armv7-a -mfloat-abi=hard -mfpu=vfpv3-d16 -mthumb -mtls-dialect=gnu -auxbase-strip CMakeFiles/strings.dir/dtoa.c.o -O2 -Wformat -Wformat-security -Werror=format-security -Wall -fstack-protector -fno-strict-aliasing -fPIC --param ssp-buffer-size=4 -fstack-protector -o - -frandom-seed=0

[Matthias Klose (doko)]

the preprocessed source works for me

[James M. Leddy (jm-leddy)]

Hi Dann,

Thanks for the reduced test case. I'm sure it will make diagnosing and potentially bisecting a lot easier.

[Manoj Iyer (manjo)]

I was able to re-create the bug with this testcase. ICE happens with 4.7 compiler and not with 4.6.

gcc -O5 -c ./job.i
./tools/src/make-3.82/job.c: In function ‘construct_command_argv_internal.isra.1’:
./tools/src/make-3.82/job.c:3073:1: internal compiler error: Segmentation fault
Please submit a full bug report,
with preprocessed source if appropriate.
See <file:///usr/share/doc/gcc-4.7/README.Bugs> for instructions.
The bug is not reproducible, so it is likely a hardware or OS problem.

[Matthias Klose (doko)]

Am 14.02.2013 21:34, schrieb Manoj Iyer:
> I was able to re-create the bug with this testcase. ICE happens with 4.7
> compiler and not with 4.6.

no, you are not:

> gcc -O5 -c ./job.i
> ./tools/src/make-3.82/job.c: In function ‘construct_command_argv_internal.isra.1’:
> ./tools/src/make-3.82/job.c:3073:1: internal compiler error: Segmentation fault
> Please submit a full bug report,
> with preprocessed source if appropriate.
> See <file:///usr/share/doc/gcc-4.7/README.Bugs> for instructions.
> The bug is not reproducible, so it is likely a hardware or OS problem.

"The bug is not reproducible, so it is likely a hardware or OS problem".

means, that the issue wasn't reproducible in a second run.

[James M. Leddy (jm-leddy)]

I'm taking this out of the OEM queue for now. Please reopen if you want help with this bug.