Compilation fails for powerpc with --enable-mpi-cxx

[amckinstry]

Hi,

Enabling --enable-mpi-cxx leads to:

make[2]: Entering directory '/home/mckinstry/openmpi-2.0.1/ompi/mpi/cxx'
/bin/bash ../../../libtool  --tag=CXX   --mode=compile g++ -DHAVE_CONFIG_H -I. -I../../../opal/include -I../../../ompi/include -I../../../oshmem/include -I../../../opal/mca/hwloc/hwloc1112/hwloc/include/private/autogen -I../../../opal/mca/hwloc/hwloc1112/hwloc/include/hwloc/autogen -I../../../ompi/mpiext/cuda/c  -DOMPI_BUILDING_CXX_BINDINGS_LIBRARY=1 -DOMPI_SKIP_MPICXX=1 -I../../.. -I../../../orte/include -Wdate-time -D_FORTIFY_SOURCE=2   -I/usr//include -I/home/mckinstry/openmpi-2.0.1/opal/mca/event/libevent2022/libevent -I/home/mckinstry/openmpi-2.0.1/opal/mca/event/libevent2022/libevent/include  -DNDEBUG -g -O2 -fdebug-prefix-map=/home/mckinstry/openmpi-2.0.1=. -fstack-protector-strong -Wformat -Werror=format-security -finline-functions -pthread -c -o mpicxx.lo mpicxx.cc
libtool: compile:  g++ -DHAVE_CONFIG_H -I. -I../../../opal/include -I../../../ompi/include -I../../../oshmem/include -I../../../opal/mca/hwloc/hwloc1112/hwloc/include/private/autogen -I../../../opal/mca/hwloc/hwloc1112/hwloc/include/hwloc/autogen -I../../../ompi/mpiext/cuda/c -DOMPI_BUILDING_CXX_BINDINGS_LIBRARY=1 -DOMPI_SKIP_MPICXX=1 -I../../.. -I../../../orte/include -Wdate-time -D_FORTIFY_SOURCE=2 -I/usr//include -I/home/mckinstry/openmpi-2.0.1/opal/mca/event/libevent2022/libevent -I/home/mckinstry/openmpi-2.0.1/opal/mca/event/libevent2022/libevent/include -DNDEBUG -g -O2 -fdebug-prefix-map=/home/mckinstry/openmpi-2.0.1=. -fstack-protector-strong -Wformat -Werror=format-security -finline-functions -pthread -c mpicxx.cc  -fPIC -DPIC -o .libs/mpicxx.o
In file included from ../../../opal/class/opal_object.h:128:0,
                 from ../../../ompi/errhandler/errhandler.h:32,
                 from mpicxx.cc:37:
../../../opal/threads/thread_usage.h: In function 'int32_t opal_thread_swap_32(volatile int32_t*, int32_t)':
../../../opal/threads/thread_usage.h:134:76: error: 'opal_atomic_swap_32' was not declared in this scope
         return opal_atomic_swap_ ## suffix ((volatile type *) ptr, newvalue); \

This fails because in opal/include/opal/sys/atomic.h:

 #if defined(c_plusplus) || defined(__cplusplus)
/* We no longer support inline assembly for C++ as OPAL is a C-only interface */
#define opal_atomic_swap_32 0
#define OPAL_DEC_INLINE_ASSEMBLY 0
#define OPAL_XLC_INLINE_ASSEMBLY 0

and opal_atomic_swap_32 is conditionally only included if opal_atomic_swap_32 is set

[jsquyres]

@hjelmn Is this related to #2051 and/or #2044?

[gpaulsen]

@nysal - can you or someone on your team look at this?

[hjelmn]

This isn't PowerPC specific. I used C99 in a header exposed to c++. I need to figure out a way to protect internal headers from c++ until the bindings are removed. will look at this today.

[jsquyres]

@hjelmn and I chatted on the call today -- he has an interesting proposal to fix this: require a C++11 compiler for the C++ MPI bindings.

My gut reaction is "we can't do that!", but Nathan raises the valid point that the MPI C++ bindings have been removed from the MPI 3.0 spec. So raising the bar a little to compile the MPI C++ bindings may not be horrible. Need to think about this a little bit...

[hjelmn]

Yeah, if we do C++11 I think it is ok to include any of our internal headers to C++. My understanding is many of the C99 features we use in Open MPI are available in C++11.

[hjelmn]

I can't get g++ 6.1 to accept C99 subobject naming. Guess I was wrong that it is supported. Working on another workaround now.