Allow MPI_*_GET_NAME functions to accept MPI_*_NULL handles

[jsquyres]

Problem

The MPI_*_SET_NAME and MPI_*_GET_NAME functions were created to aid debugging. However, similar to most other MPI functions, it is (likely?) invalid to pass the MPI_*_NULL handles to the MPI_*_GET_NAME functions. For example:

MPI_Datatype dt = MPI_DATATYPE_NULL;
char name[MPI_MAX_OBJECT_NAME];
int outlen = 0;

// ...
MPI_Type_get_name(dt, name, &outlen);
printf("The datatype is: %s\n", name);

If dt is still MPI_DATATYPE_NULL, Open MPI v4.1.x and MPICH 4.0.0 will both invoke errors. Indeed, a user reported this behavior as a bug to Open MPI: open-mpi/ompi#10028

Instead, users must add an alternate code path to make the above snipit valid, perhaps something like this:

if (dt == MPI_DATATYPE_NULL) {
    strcpy(name, "MPI_DATATYPE_NULL");
} else {
    MPI_Type_get_name(dt, name, &outlen);
}
printf("The datatype is: %s\n", name);

Proposal

We propose making it permissible to pass in the MPI_*_NULL handles to the MPI_*_GET_NAME functions.

Note: we do not propose making it permissible to pass in the MPI_*_NULL handles to the MPI_*_SET_NAME functions.

Changes to the Text

Add sentences to the MPI_*_GET_NAMES definitions. See the PDF built from https://github.com/mpi-forum/mpi-standard/pull/652, pages:

• 361: MPI_COMM_GET_NAME
• 362: MPI_TYPE_GET_NAME
• 363: MPI_WIN_GET_NAME
• 1001: changelog

Impact on Implementations

Allow the MPI_*_NULL handles in the MPI_*_GET_NAME functions. This will involve relaxing the error checking of these 3 functions, and possibly special-casing the string that is returned in the case of MPI_*_NULL inputs.

Impact on Users

Users no longer need to special-case the use of these functions.

References and Pull Requests

https://github.com/mpi-forum/mpi-standard/pull/652

[wesbland]

@jsquyres I'm assuming that you're not going to be doing the reading here?

In the absence of any other owner, I'm going to assign this issue to @GuillaumeMercier since he's the CC chair for the impacted chapter.

[Wee-Free-Scot]

Curious: Why is there no equivalent GET|SET_NAME functionality for FILE objects in MPI?

Files have a name in the filesystem; I guess that is seen as sufficient? The MPI_FILE_NULL handle exists. To get the "filename" of an MPI_FILE, we need to use MPI_FILE_GET_INFO so we can look for the "filename" key-value pair. Can MPI_FILE_NULL be passed to MPI_FILE_GET_INFO?

The MPI_FILE object might benefit from a name that is not identical to its "filename" in the filesystem. For example, I might name one MPI_FILE with "input data file" (its role in my program), even though the filesystem name is "data_128x1024_molecule007.bin"

[wesbland]

This issue had a "no-no" vote on 2022-05-26, which passed:

Yes	No	Abstain
26	0	3