At the last meeting in San Jose, Rolf brought up a good point w.r.t.
ticket #18 [1]. Part of the proposal for ticket #18 is to add
MPI_AINT and MPI_OFFSET types corresponding to MPI_Aint/
MPI_ADDRESS_KIND and MPI_Offset/MPI_OFFSET_KIND language types.
Rolf's point was that it isn't defined in the standard whether or not
MPI_Aint and INTEGER(KIND=MPI_ADDRESS_KIND) are exactly the same
type. Obviously we cannot define a new predefined datatype
corresponding to two different language types. I believe Rolf's view
(correct me if I'm misrepresenting you, Rolf) was that the forum ought
to make their equivalence explicit, irrespective of what happens with
ticket #18. I'm ambivalent about this issue, but I need a resolution
on it in order to get #18 into a final voteable state.
So the larger question to the Forum is: Are these types exactly the
same between the two languages?
Looking at MPICH2 it seems that we try to make them the same size but
there are some convoluted code paths in the build logic that might
result in different sizes. I would not consider this to be a
compelling argument that they are permitted to differ because the
logic in question may simply be buggy or never exercised. Does anyone
have an example from other MPI implementations where the types do/may
differ? Do any Fortran experts out there know of instances where
MPI_ADDRESS_KIND would be different from MPI_Aint?
As for the impact on ticket #18, there are several ways to deal with
this:
1) If the Forum decides that the types are always equivalent, add text
to that effect (probably in a separate ticket) and otherwise stick
with the current proposal.
2) If the Forum decides that the type equivalence is undefined, drop
the new MPI_AINT/MPI_OFFSET types from the proposal.
3) If the Forum decides that the type equivalence is undefined, create
four types instead of just two: MPI_AINT, MPI_OFFSET, MPI_F_AINT,
MPI_F_OFFSET (or some such names).
4) If the Forum decides that the type equivalence is undefined, make
MPI_AINT/MPI_OFFSET only correspond to the C types.
I'm inclined towards either #1 or #2, depending on what the Forum
decides about the type equivalence. I also can separate the MPI_AINT/
MPI_OFFSET types out into a separate ticket if this becomes too
contentious of an issue, but I'd like to keep it together for now
since the text changes are easier to describe correctly in just one
place.
-Dave
[1] https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/18
