Revise restriction for MPI_T functions to return error classes

[mahermanns]

Problem

Before MPI 4.0 and the introduction of sessions, MPI_Error_class was not in the set of functions callable before MPI initialization. Therefore, in MPI 3.0 we introduced the restriction on functions of the Tool Information Interface (MPI_T) to return error classes instead of error codes.

The rationale for this restrictions---which is not correct for MPI 4.0 onwards---has been removed without replacement.

This issue shall provide discussion ground for a suitable clarification of the situation.

Proposal

tbd.

Changes to the Text

tbd.

Impact on Implementations

Depends on the solution we decide for.

Impact on Users

Depends on the solution we decide for.

References

• Issue in the 4.0 RC Process: Tools chapter problem: MPI_ERROR_CLASS _can_ be called early (4.0 RC review) #369
  • Corresponding PR: https://github.com/mpi-forum/mpi-standard/pull/396

[schulzm]

The non existence of error handling routines before Init was only part (or rather a side effect) of the use of error classes in MPI_T - the main goals were to have a more defined set of return values for tools to operate with and the non fatality in all cases, to not impact MPI operations of a tool return goes wrong. If I remember, we started with something completely new and then later said, might as well just call those error classes

[mahermanns]

@schulzm Revisiting this after the Sep 2022 meeting, would you propose to leave the semantics as it is (apart from changes that are necessary to fulfill #542 and close this issue?

This would mean everything would stay the same for MPI_T function, always return something of MPI_T_ERR_* sorts. Correct?

[schulzm]

I think this is indeed resolved by 542 and we can probably close this one - the rationale should in any case be deleted, but this seems to have happened in 542 already

[mahermanns]

Closing this issue as it will be resolved by the fixes for #542.