[mvapich-discuss] Compilation error for mvapich2-2.0.1 with disabled C++ bindings.

[Jonathan Perkins]

On Tue, Jan 13, 2015 at 03:36:35PM -0600, Chris Green wrote:
> Hi,
> 
> We are attempting to build a non-C++ mvapich2 that can be used safely by
> people writing C++ code using multiple compilers. Using the following
> configuration line on SLF6.6 (Scientific Linux Fermi -- a close derivative
> of RHEL6.6), with the OS-provided InfiniBand support and development
> libraries:
> 
> ./configure --prefix=/usr/local/mvapich2-2.0.1 --enable-fast=O3,ndebug --enable-f77 --enable-fc \
>             --disable-cxx --enable-romio --enable-versioning --enable-threads=runtime \
>             --enable-registration-cache --enable-rsh --enable-shared --enable-static \
>             --enable-yield=sched_yield --enable-rdma-cm
> 
> a subsequent compilation fails attempting to compile
> src/mpi/coll/ch3_shmem_coll.c with:
> 
> <snip/>
> make[2]: Entering directory `/home/greenc/work/cet-is/sources/ssi-build/mvapich-ssi-build/mvapich2-2.0.1'
>   CC       src/mpi/coll/ch3_shmem_coll.lo
> src/mpi/coll/ch3_shmem_coll.c: In function 'mv2_shm_zcpy_reduce':
> src/mpi/coll/ch3_shmem_coll.c:3829:33: error: 'is_cxx_uop' undeclared (first use in this function)
>                        datatype, is_cxx_uop);
>                                  ^
> src/mpi/coll/ch3_shmem_coll.c:3829:33: note: each undeclared identifier is reported only once for each function it appears in
> make[2]: *** [src/mpi/coll/ch3_shmem_coll.lo] Error 1
> <snip/>
> 
> The following is the minimum patch required for successful compilation:
> 
>  [greenc at woof] ~/work/cet-is/sources/ssi-build/mvapich-ssi-build/mvapich2-2.0.1/src/mpi/coll $ diff -bu ch3_shmem_coll.c{~,}
> --- ch3_shmem_coll.c~   2014-10-29 13:07:22.000000000 -0500
> +++ ch3_shmem_coll.c    2015-01-13 13:33:25.406662528 -0600
> @@ -3826,7 +3826,13 @@
>         /* Now, do the intra-node shm-slot reduction */
>         mv2_shm_tree_reduce(shmem, in_buf, len, count, intra_node_root, uop,
> -                      datatype, is_cxx_uop);
> +                      datatype,
> +#ifdef HAVE_CXX_BINDING
> +                      is_cxx_uop
> +#else
> +                      0
> +#endif
> +);
> 
>         shmem->queue[shmem->local_rank].shm_slots[windex]->psn = shmem->write;
>         shmem->queue[shmem->local_rank].shm_slots[windex]->tail_psn = (volatile uint32_t *)
> @@ -3901,7 +3907,13 @@
>         }
>      } else {
>         mv2_shm_tree_reduce(shmem, in_buf, len, count, intra_node_root, uop,
> -                      datatype, is_cxx_uop);
> +                      datatype,
> +#ifdef HAVE_CXX_BINDING
> +                      is_cxx_uop
> +#else
> +                      0
> +#endif
> +);
>         shmem->queue[shmem->local_rank].shm_slots[windex]->psn = shmem->write;
>      }
> 
> We believe this is a clear-cut error and simple fix, but if we do in fact
> have something wrong with our configure line, we'd be grateful for the
> education.

Chris, thanks for pointing out this issue and providing a patch.  It
appears that we pass the value of is_cxx_uop into functions regardless
of whether HAVE_CXX_BINDING is defined or not.

The fix is simple and I did not see any problems with your configure
line (some default values are explicitly set but that may have value in
your case).

-- 
Jonathan Perkins