The proposal to present a datatype as a character string seems the
most useful and the easiest to use for me: you get all the information
in a single call, and you get it in a representation that's universally
readable.
I have two applications that would use datatype access functions.
One is an implementation of MPI-IO, and the other is a visualization
tool for MPI type maps. For the latter tool in particular, it's
very valuable to have a portable representation of datatypes.
Although the tool currently works only for types that the user has
generated by filling in entries in a GUI, a character-based
representation of a type would make it trivial for this tool to
generate displays directly from types created in real application
programs.
However, I do have one concern about the proposal. It states that
the representation returned should use the language of the calling
function. I would also like to have a representation in the language
that created the datatype. This would allow a C-language function to
get an accurate description of a type that was originally generated
by a Fortran call, or a type that was generated in C from Fortran
base types. (Assuming that Fortran base types are available in C;
I'm not sure if they are.)
John