ABI: Fortran ABI TODO

[hzhou]

The next step for MPI ABI after swappable libmpi.so is to make libmpifort.so swappable. This issue is for discussion and tracking todo list.

TODO list

• Add to Fortran binding scripts to generate ABI versions based on ABI header mpi_abi.h
• F77
  • MPI Calls are directly linked to named-mangled symbols -- nothing to be done, but double check
  • Need standardize common blocks names. Currently it is like

00000000000d3570 B mpifcmb5_
00000000000d3560 B mpifcmb9_
00000000000d3550 B mpifcmba_

• F90
  • Double check
• F08
  • Remove the internal modules
    The symbols are tied to the module name, e.g. __mpi_f08_callbacks_MOD_mpi_comm_dup_fn. It need be __mpi_f08_MOD_mpi_comm_dup_fn

Discussion

Hmm, likely there will be more issues. Fortran ABI will likely require standardizing Fortran binding implementation. I think it is more sensible to release a stand-alone libmpifort_abi.so that is built on top of libmpi_abi.so

[dalcinl]

it is more sensible to release a stand-alone libmpifort_abi.so that is built on top of libmpi_abi.so

Definitely. Moreover, you better do not mess with the current internal Fortran module names or you will break the MPICH ABI for Fortran users in a really bad way (unless Fortran is not covered by the MPICH ABI initiative?).

[hzhou]

it is more sensible to release a stand-alone libmpifort_abi.so that is built on top of libmpi_abi.so

Definitely. Moreover, you better do not mess with the current internal Fortran module names or you will break the MPICH ABI for Fortran users in a really bad way (unless Fortran is not covered by the MPICH ABI initiative?).

That's okay. The new ABI is an opt-in.

[hzhou]

Closing this issue. Follow #6953 for Fortran binding on top of MPI ABI.