OpenBLAS 0.3.13 fails to build on ppc64le / gcc 8.2

[jonglezb]

I'm building OpenBLAS 0.3.13 using Spack on a POWER8 machine running Debian 10.

Build fails because of the new POWER10 dynamic support:

==> Error: ProcessError: Command exited with status 2:
    'make' '-j16' 'CC=/grid5000/spack/lib/spack/env/gcc/gcc' 'FC=/grid5000/spack/lib/spack/env/gcc/gfortran' 'MAKE_NB_JOBS=0' 'DYNAMIC_ARCH=1' 'USE_OPENMP=0' 'USE_THREAD=0' 'RANLIB=ranlib' 'libs' 'netlib' 'shared'                        

2 errors found in build log:
     1190    /grid5000/spack/lib/spack/env/gcc/gcc -O2 -DMAX_STACK_ALLOC=2048 -Wall -m64 -DF_INTERFACE_GFORT -fPIC -DDYNAMIC_ARCH -DNO_WARMUP -DMAX_CPU_NUMBER=160 -DMAX_PARALLEL_NUMBER=1 -DBUILD_SINGLE=1 -DBUILD_DOUBLE=1 -DBUILD_COMPLEX=
             1 -DBUILD_COMPLEX16=1 -DVERSION=\"0.3.13\" -Ofast -mcpu=power8 -mtune=power8 -mvsx  -fno-fast-math -UASMNAME -UASMFNAME -UNAME -UCNAME -UCHAR_NAME -UCHAR_CNAME -DASMNAME=openblas_env -DASMFNAME=openblas_env_ -DNAME=openblas_
             env_ -DCNAME=openblas_env -DCHAR_NAME=\"openblas_env_\" -DCHAR_CNAME=\"openblas_env\" -DNO_AFFINITY -I../.. -c openblas_env.c -o openblas_env.o                                                                                 
     1191    /grid5000/spack/lib/spack/env/gcc/gcc -O2 -DMAX_STACK_ALLOC=2048 -Wall -m64 -DF_INTERFACE_GFORT -fPIC -DDYNAMIC_ARCH -DNO_WARMUP -DMAX_CPU_NUMBER=160 -DMAX_PARALLEL_NUMBER=1 -DBUILD_SINGLE=1 -DBUILD_DOUBLE=1 -DBUILD_COMPLEX=
             1 -DBUILD_COMPLEX16=1 -DVERSION=\"0.3.13\" -Ofast -mcpu=power8 -mtune=power8 -mvsx  -fno-fast-math -UASMNAME -UASMFNAME -UNAME -UCNAME -UCHAR_NAME -UCHAR_CNAME -DASMNAME=dynamic_power -DASMFNAME=dynamic_power_ -DNAME=dynamic
             _power_ -DCNAME=dynamic_power -DCHAR_NAME=\"dynamic_power_\" -DCHAR_CNAME=\"dynamic_power\" -DNO_AFFINITY -I../..   -c -o dynamic_power.o dynamic_power.c                                                                       
     1192    dynamic_power.c:9:1: warning: multi-line comment [-Wcomment]
     1193     //#if (!defined __GNUC__) || ( __GNUC__ >= 11) \
     1194     ^
     1195    dynamic_power.c: In function 'gotoblas_dynamic_init':
  >> 1196    dynamic_power.c:57:6: error: cpu 'power10' is an invalid argument to builtin '__builtin_cpu_is'
     1197      if (__builtin_cpu_is("power10"))
     1198          ^~~~~~~~~~~~~~~~~~~~~~~~~~~
  >> 1199    make[1]: *** [<builtin>: dynamic_power.o] Error 1
     1200    make[1]: *** Waiting for unfinished jobs....
     1201    make[1]: Leaving directory '/var/spack/stage/root/spack-stage-openblas-0.3.13-57p57xjbiiv757vjq4uickpldoq4vcfy/spack-src/driver/others'                                                                                         
     1202    make: *** [Makefile:165: libs] Error 1

Here are the versions:

# gcc --version
gcc (Debian 8.3.0-6) 8.3.0
Copyright (C) 2018 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

# ld --version
GNU ld (GNU Binutils for Debian) 2.31.1
Copyright (C) 2018 Free Software Foundation, Inc.
This program is free software; you may redistribute it under the terms of
the GNU General Public License version 3 or (at your option) a later version.
This program has absolutely no warranty.

I browsed other issues, and as far as I can tell, POWER10 support is supposed to be disabled with old versions of GCC. Not sure what's up.

[martin-frbg]

Looks like fallout from #2994 (and yes RajalakshmiSR warned about this happening, I should not have merged antonblanchard's PR without further clarification from him probably. This did not come up in my tests back then though)

[jonglezb]

FYI, 0.3.12 builds fine (also using Spack, on the same POWER8 host).

I assume you already have access to POWER machines, but just in case, the GCC compile farm has several hw/sw POWER environments to test things: https://cfarm.tetaneutral.net/

[martin-frbg]

Thx, I'm allowed to use the OpenPOWER minicloud at unicamp.br

[RajalakshmiSR]

@martin-frbg Should we just back out 2994 for now?

[martin-frbg]

Either that, or put another gcc version check for >=11 around that __builtin_cpu_is() call. The other change in #2994 looked more important, and I can only assume I got fixated on whether "cc -Wl,--version" was actually safer than "ld --version".

[brada4]

@jonglezb You posted debian compiler version, does not matter, no released gcc would work.

https://github.com/spack/spack/blob/develop/var/spack/repos/builtin/packages/gcc/package.py

[martin-frbg]

Actually gcc 10.2 would work, as RajalakshmiSR pointed out to me in #3075 the relevant gcc patch was backported from mainline shortly before its release.

[jonglezb]

Thanks, I just tested a build from the develop branch and it builds fine!