Fortran: disable real128 when compiler does not support

[keszybz]

Fails with v3.3.1, v.3.2.1 is OK:

src/binding/fortran/use_mpi_f08/mpi_f08_types.f90:355:16:

  355 |     real(real128),dimension(..) :: x
      |                1
Error: Kind -1 not supported for type REAL at (1)
src/binding/fortran/use_mpi_f08/mpi_f08_types.f90:385:19:

  385 |     complex(real128),dimension(..) :: x
      |                   1
Error: Kind -1 not supported for type COMPLEX at (1)
src/binding/fortran/use_mpi_f08/mpi_f08_types.f90:383:36:

  383 | subroutine MPI_Sizeof_xcomplex128 (x, size, ierror)
      |                                    1
Error: Symbol ‘x’ at (1) has no IMPLICIT type
src/binding/fortran/use_mpi_f08/mpi_f08_types.f90:353:33:

  353 | subroutine MPI_Sizeof_xreal128 (x, size, ierror)
      |                                 1
Error: Symbol ‘x’ at (1) has no IMPLICIT type
src/binding/fortran/use_mpi_f08/mpi_f08_types.f90:353:30:

  353 | subroutine MPI_Sizeof_xreal128 (x, size, ierror)
      |                              1
......
  383 | subroutine MPI_Sizeof_xcomplex128 (x, size, ierror)
      |                                 2
Error: Ambiguous interfaces in generic interface 'mpi_sizeof' for ‘mpi_sizeof_xreal128’ at (1) and ‘mpi_sizeof_xcomplex128’ at (2)
make: *** [Makefile:48733: src/binding/fortran/use_mpi_f08/mpi_f08_types.stamp] Error 1

Fails in Fedora rawhide, F31, F30.

[keszybz]

I now built it with a simple patch which removes those MPI_Sizeof_xcomplex128, MPI_Sizeof_xreal128. Not sure if it'll cause any problems down the road ;)

[nmorey]

I'm also seeing this issue on SUSE Tumbleweed. This also impacts mvapich2 2.3.2.

@keszybz Would you mind sharing the patch ?

[keszybz]

https://src.fedoraproject.org/rpms/mpich/blob/3b497cff602f88dd1664ba31a34504d40da44156/f/0001-Drop-real128.patch

[keszybz]

... but I'm only applying it on arm (https://src.fedoraproject.org/rpms/mpich/blob/master/f/mpich.spec#_111). It'd have to be conditionalized to apply everywhere.

[nmorey]

@keszybz Thanks. Will do the same for now until a proper fix is available.
Note: this applies cleanly to mvapich2 2.3.1 aswell

[harisubramoni]

@keszybz @nmorey - Thanks for the discussion and the patch.

@raffenet - do you recommend that I take this into the MVAPICH2 codebase in the interim while you guys work on a full solution?

Thx,
Hari.

[hzhou]

The only side effects removing MPI_Sizeof_xreal128 is on system that do support real128, which is real*16 I believe, and the user write code call MPI_Sizeof(x, size_x, ierror) where x is of that long double type, the compilation will fail. A proper fix probably need check configure.

In the interim, removing them essentially just removed the support of real128 for the MPI_Sizeof. Both the usage of real128 and MPI_Sizeof as well as use mpi_f08 are rare anyway, so the impact is likely minimal.

[hzhou]

Currently we generate the fortran bindings at autogen time, before configure. In order to use configure, we generate .in files with autoconf macros in them and run them through configure. This is convoluted and limited. The proper way is have configure test features, then run buildiface as ac_config_commands so it can generate interfaces that reflect feature options directly.

[hzhou]

Tagging PR #3819

[raffenet]

The proper way is have configure test features, then run buildiface as ac_config_commands so it can generate interfaces that reflect feature options directly.

Yes! This way --disable-fortran would actually save buildiface from running.

[swepettax]

I ran in to the same problem today, using Arch Linux on my RPI. Any solution yet?

[everythingfunctional]

Found this issue today trying to build a Pi cluster for testing parallel Fortran code. Just made the fix myself, but it would be nice if the configure step could do it 😉