clarify what functions can be called on uncommitted datatypes

[jeffhammond]

This text

4.1.9 Commit and Free
A datatype object has to be committed before it can be used in a communication. 
As an argument in datatype constructors, uncommitted and also committed datatypes 
can be used. There is no need to commit basic datatypes. They are “pre-committed.”

leaves some ambiguity since the functions MPI_TYPE_SIZE(_X) and MPI_TYPE_GET_{CONTENTS,ENVELOPE,EXTENT(_X),TRUE_EXTENT(_X)} lie outside of the union of "communication" and "datatype constructors".

It would be helpful to users if the text was unambiguous about the functions that can and cannot be called on uncommitted datatypes.

[bosilca]

The text clearly states what the major constraint is: datatype must be committed before being used in a communication. All other cases do not require the commit stage. However, I agree with you that the second part is confusing, there is no need to over-clarify that a non-committed datatype can be used as an argument to datatype constructors, or if we do we should define it as an example.

[jeffhammond]

@bosilca My goal is to clarify the case of functions that are neither communication functions nor datatype constructors. It may be obvious to the Forum, but I believe a power user was confused about this, which is why I created the original issue.

[bosilca]

@jeffhammond we are over-specifying. The text should be simple: only datatypes used in communications should be committed, all other usages do not require a committed datatype.

[jeffhammond]

@bosilca It sounds like we should replace "datatype constructors" with "other functions" or "non-communication MPI functions".

[bosilca]

@jeffhammond we should be careful about how we define "non-communication functions". Are I/O included? They should but one can argue that they might not qualify from a user perspective as "communication".

[jeffhammond]

@bosilca Hence my original proposal: It would be helpful to users if the text was unambiguous about the functions that can and cannot be called on uncommitted datatypes.

[thender]

Should uncommitted datatypes be freed via MPI_TYPE_FREE()? My use-case is nested datatypes where "leaf" datatypes need not be committed (per 4.1.9).

[wgropp]

Yes, they will eventually need to be freed. They can be freed once the type is no longer needed explicitly, even if they are used within another type. In fact, in many cases, this is the correct approach - create the datatypes, free all but the ones that are committed and used in subsequent communication.

[schulzm]

Austin June 2018 meeting: we should revisit this and propose a clarification - Wesley, Martin and Jeff (?) will work on this and propose for BCN meeting