MPI_T ABI

[jeffhammond]

Problem

The constants in the MPI tools interface need to be defined in the ABI.

Proposal

Somebody from the Tools WG needs to provide values to e.g. https://github.com/mpiwg-abi/header_and_stub_library/blob/main/mpi.h#L493.

Changes to the Text

TODO

Impact on Implementations

Necessary for ABI.

Impact on Users

Necessary for ABI.

References and Pull Requests

[mahermanns]

As per discussion in the Tools WG call on Mar 04, 2024:

• MPI_T_ERR_XXX need to adhere to the restrictions of error codes (i.e., ranges of values, etc.)
• It might be of benefit to have different constant groups "own" different bitmasks (i.e., make any MPI_T_BIND_xxx distinguishable from MPI_T_PVAR_CLASS_xxx ... but this would be an interesting benefit, not strictly required
• The MPI_T_CB_xxx constants form a hierarchy, so in principle one could make them some form of bitfield, but it is also not strictly required, as they are not used in the bitfield-ish way of ORing multiple of those constants anyway. Again, this may add comfort, but as long as they are strictly distinct, I think, this should be OK.

We didn't get to checking for all MPI_T constants in the call and more discussion seems to be in order.

[mahermanns]

As @jeffhammond read the C ABI at the September 2024 meeting with all feedback of our group included, and not further changes anticipated, we close this ticket.