jeffhammond commented 9 months ago

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 commented 8 months ago

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 commented 1 month ago

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.

mpi-forum / mpi-issues