glibc2.32 build error for powerpc64-unknown-linux-gnu
Chris Packham
Chris.Packham@alliedtelesis.co.nz
Sun Sep 6 09:35:49 GMT 2020
On 5/09/20 1:33 am, Adhemerval Zanella wrote:
>
> On 03/09/2020 17:58, Chris Packham wrote:
>> On 3/09/20 10:53 pm, Adhemerval Zanella wrote:
>>> On 03/09/2020 06:09, Chris Packham wrote:
>>>> On 3/09/20 2:39 pm, Chris Packham wrote:
>>>>> On 3/09/20 9:14 am, Chris Packham wrote:
>>>>>> On 3/09/20 2:20 am, Adhemerval Zanella wrote:
>>>>>>> On 02/09/2020 05:05, Chris Packham wrote:
>>>>>>>> On 2/09/20 7:40 am, Adhemerval Zanella wrote:
>>>>>>>>> On 30/08/2020 07:40, Chris Packham via Libc-help wrote:
>>>>>>>>>> Hi,
>>>>>>>>>>
>>>>>>>>>> crosstool-ng was recently updated to add glibc-2.32. We're seeing
>>>>>>>>>> the
>>>>>>>>>> following error building for powerpc64
>>>>>>>>>>
>>>>>>>>>> [ALL ] In file included from
>>>>>>>>>> ../sysdeps/powerpc/powerpc32/power4/multiarch/wordcopy-ppc32.c:26,
>>>>>>>>>> [ALL ] from
>>>>>>>>>> ../sysdeps/powerpc/powerpc64/multiarch/wordcopy-ppc64.c:18:
>>>>>>>>>> [ALL ] ../string/wordcopy.c: In function
>>>>>>>>>> '_wordcopy_fwd_aligned':
>>>>>>>>>> [ERROR] ../string/wordcopy.c:98:26: error: 'a1' may be used
>>>>>>>>>> uninitialized in this function [-Werror=maybe-uninitialized]
>>>>>>>>>> [ALL ] 98 | ((op_t *) dstp)[0] = a1;
>>>>>>>>>> [ALL ] | ~~~~~~~~~~~~~~~~~~~^~~~
>>>>>>>>>>
>>>>>>>>>> Looking at the code in question I think it's a spurious warning.
>>>>>>>>>> There
>>>>>>>>>> even appears to be an attempt to turn off the warning in
>>>>>>>>>> sysdeps/generic/Makefile but the file is #included for powerpc so
>>>>>>>>>> the
>>>>>>>>>> flags applied.
>>>>>>>>>>
>>>>>>>>>> Bug report with full logs
>>>>>>>>>> https://github.com/crosstool-ng/crosstool-ng/issues/1380
>>>>>>>>>>
>>>>>>>>> I haven't see it with gcc version 9.3.1, so it might be something
>>>>>>>>> related
>>>>>>>>> to gcc 10 or newer. I will build a gcc 10 and check if I can
>>>>>>>>> reproduce it.
>>>>>>>> I just tried a crosstool-ng config selecting gcc 9.3.0 and I still see
>>>>>>>> the same error.
>>>>>>>>
>>>>>>> I just tried to reproduce it building powerpc64-linux-gnu,
>>>>>>> powerpc-linux-gnu
>>>>>>> and powerpc-linux-gnu-power4 (with --with-cpu-power4) with both gcc
>>>>>>> 9.3.1
>>>>>>> and 10, but without success. Are you using any special flags or
>>>>>>> patch for gcc?
>>>>>> There are a collection of patches
>>>>>> https://github.com/crosstool-ng/crosstool-ng/tree/master/packages/gcc/10.2.0
>>>>>> mostly just brought forward from older gcc versions. I'll try
>>>>>> disabling them.
>>>>> Even without any of the glibc or gcc patches and I still see the same
>>>>> error.
>>>> I think I might know why you can't see the error. crosstool-ng sets
>>>> -Werror when building glibc. So it's a warning that gets promoted to an
>>>> error by -Werror.
>>>>
>>>> I'll update crosstool-ng to stop doing that for powerpc64 + glibc-2.32.
>>> We add -Werror by default since 2.21 and have added --disable-werror to
>>> disable it. I am trying to understand why you are seeing this error in
>>> your environment because this specific code has not been touched in a long
>>> time and constantly built *all* supported ABI some different gcc versions [1].
>>>
>>> And the last report for gcc 9 [1] and gcc 10 has not shown any issue, neither
>>> any testing I did with gcc 8, gcc9, or gcc10. I will try to build some
>>> powerpc permutations with --with-cpu and/or --disable-multi-arch.
>>>
>>>
>>> [1] https://sourceware.org/pipermail/libc-testresults/2020q3/thread.html
>>> [2] https://sourceware.org/pipermail/libc-testresults/2020q3/006398.html
>>> [3] https://sourceware.org/pipermail/libc-testresults/2020q3/006397.html
>> Hmm the plot thickens.
>>
>> If it helps this is the configure line crosstool-ng ends up using
>>
>> BUILD_CC='x86_64-build_pc-linux-gnu-gcc'
>> CC='powerpc64-unknown-linux-gnu-gcc -g -O2 -U_FORTIFY_SOURCE ' CFLAGS=''
>> AR='powerpc64-unknown-linux-gnu-ar'
>> RANLIB='powerpc64-unknown-linux-gnu-ranlib' '/bin/bash'
>> '/home/runner/work/crosstool-ng/crosstool-ng/.build/powerpc64-unknown-linux-gnu/src/glibc/configure'
>> '--prefix=/usr' '--build=x86_64-build_pc-linux-gnu'
>> '--host=powerpc64-unknown-linux-gnu'
>> '--cache-file=/home/runner/work/crosstool-ng/crosstool-ng/.build/powerpc64-unknown-linux-gnu/build/build-libc-final/multilib/config.cache'
>> '--without-cvs' '--disable-profile' '--without-gd'
>> '--with-headers=/home/runner/work/crosstool-ng/crosstool-ng/x-tools/powerpc64-unknown-linux-gnu/powerpc64-unknown-linux-gnu/sysroot/usr/include'
>> '--disable-debug' '--disable-sanity-checks' '--enable-obsolete-rpc'
>> '--enable-kernel=5.5.5' '--with-__thread' '--with-tls' '--enable-shared'
>> '--enable-add-ons=no' '--with-pkgversion=crosstool-NG UNKNOWN'
> Some notes you might consider to add on crosstool-ng:
>
> --without-cvs - this is a long time removed option.
> --disable-profile - it is the default option, no need to add it.
> --disable-debug - there is no such option.
> --disable-sanity-checks - this is only useful to disable the warning that
> using the default --prefix value (/usr/local) is not
> valid. Since it is already setting to the expected
> value this is a superfluous option.
> --enable-obsolete-rpc - This option was removed on 2.32. The NEWS contains
> more information.
> --with-__thread - this is another deprecated/remove option, glibc now
> requires TLS support.
> --with-tls - Ditto.
> --enable-shared - This is the default.
> --enable-add-ons - Another deprecated/remove option.
>
Thanks for the pointers. Some of these are hangovers from older libc
versions. We support down to 2.12.1 so might need to be retained for
some of these but they can be made conditional where needed.
>> It's the same between a glibc-2.31 build and a glibc-2.32 build.
>>
> I check multiple build permutation with gcc-9 and gcc-10 native (building
> on a powerpc64):
>
> powerpc64-linux-gnu-power4
> powerpc64-linux-gnu-power5
> powerpc64-linux-gnu-power5+
> powerpc64-linux-gnu-power6
> powerpc64-linux-gnu-power6x
> powerpc64-linux-gnu-power7
> powerpc64-linux-gnu-power8
> powerpc64-linux-gnu-power8-disable-multi-arch
> powerpc64-linux-gnu
> powerpc64-linux-gnu-disable-multi-arch
> powerpc64-linux-gnu-power4-disable-multi-arch
> powerpc64-linux-gnu-power5+-disable-multi-arch
> powerpc64-linux-gnu-power5-disable-multi-arch
> powerpc64-linux-gnu-power6-disable-multi-arch
> powerpc64-linux-gnu-power6x-disable-multi-arch
> powerpc64-linux-gnu-power7-disable-multi-arch
>
> The "-powerN" means using "--with-cpu=powerN" which "-disable-multi-arch" means
> using "-disable-multi-arch".
>
> I also checked on a cross build using both gcc-9 and gcc-10 on a x86_64 host
> to powerpc64-linux-gnu.
>
> I really think this is environment issue on your side. Maybe I can fix it by
> disabling with pragmas, but it would be hard to get any reviews without
> actually reproducing it :/
I'd tend to agree with you. Things might be working for glibc-2.31 by
accident. I've looked over the differences between 2.31 and 2.32 and I
agree there's nothing remotely related that could be causing this.
I've seen failures in my local environment (Docker container using
Debian buster) and in the github runners (Ubuntu 18.04). When I get a
chance I'll try manually building glibc-2.32.
More information about the Libc-help
mailing list