Hide some internal functions and 13 variables from public ABI

[hppritcha]

Add OMPI_HIDDEN attribute to internal symbols in libmpi.so and libmpi_abi.so to prevent them from being exported in the public API.

Hidden functions (27):

• 9 ABI C callback functions (ABI_C_MPI_*_FN)
• 12 MPI_T functions (ompit_*, including ompi_mpit_lock/unlock)
• 4 Fortran ABI functions (ompi_abi_get/set_fortran_*)
• 2 Sendrecv functions (ompi_sendrecv, ompi_isendrecv)

Hidden variables (13):

• 9 ABI variables (ompi_abi_mpi_*)
• 2 MPI_T variables (ompi_mpit_big_lock, ompi_mpit_init_failed)
• 2 Class instances (ompi_isendrecv_context_t_class, ompi_isendrecv_replace_context_t_class)

Files modified:

• ompi/mpi/c/attr_fn_abi.c
• ompi/mpi/c/abi.h
• ompi/mpi/c/bindings.h
• ompi/mpi/c/ompi_abi_fortran.c
• ompi/mpi/c/ompi_isendrecv.c
• ompi/mpi/c/ompi_sendrecv.c
• ompi/mpi/tool/mpit_common.c
• ompi/mpi/tool/mpit-internal.h
• ompi/test/general/abi_fortran.c
• ompi/mpi/bindings/ompi_bindings/c.py
• ompi/mpi/c/isendrecv_replace.c.in

Also updated AGENTS.md with a note about changing visibility of internal symbols which tests might be using.

Related to #14336

[hppritcha]

This PR doesn't address the mysterious

10 Removed function symbols not referenced by debug info:

  [D] MPI_Attr_delete
  [D] MPI_Attr_get
  [D] MPI_Attr_put
  [D] MPI_Keyval_create
  [D] MPI_Keyval_free
  [D] PMPI_Attr_delete
  [D] PMPI_Attr_get
  [D] PMPI_Attr_put
  [D] PMPI_Keyval_create
  [D] PMPI_Keyval_free

mentioned in the issue. That will be investigated/addressed in a different PR>

[devreal]

I cannot find another place where we have used OMPI_HIDDEN in the ompi source tree (only in the third-party tree). Why is it necessary for internal ABI and tools functions? Would OMPI_DECLSPEC not achieve the same goal?

[hppritcha]

@dalcinl how important do you think it is for there to be abidiff between the stub implementation and ompi libmpi_abi?

[dalcinl]

how important do you think it is for there to be abidiff between the stub implementation and ompi libmpi_abi?

Well, it depends on what kind of differences we are talking about.

For sure there should be nothing missing, that would be a bug. That's why I pointed out the missing MPI_(Attr|Keyval)_* functions.

However, I would not complain if OMPI exposes additional functions. For example, OMPI may want to expose MPIX_* functions related to the current in-development fault tolerance APis. I think MPICH is currently doing that, that is exposing its MPIX_ symbols.

What I'm not so OK about is carelessly exposing internal functions and variables without a clear purpose. I think that is the case of the 27 functions and 13 variables I reported in #14336.

We could think of adding a CI test using abidiff. A suppression file would be enough to ignore extra symbols not in the MPI standard, e.g. the MPIX_ stuff, if OMPI ever decides to expose these things in libmpi_abi.so.

[dalcinl]

I cannot find another place where we have used OMPI_HIDDEN in the ompi source tree (only in the third-party tree). Why is it necessary for internal ABI and tools functions? Would OMPI_DECLSPEC not achieve the same goal?

Well, OMPI's ABI mpi.h does not use OMPI_DECLSPEC. Maybe it should, and libmpi_abi sources should then be build with -fvisibility=hidden ? IIRC, MPICH does something like this. The installed, user-facing ABI mpi.h is plain, no visibility attributes on functions, just like OMPI. However, the sources are built with a different, autogenerated header that do have the equivalent of OMPI_DECLSPEC.