From: HOLZHEU@XXXXXXXXXX
X-Lotus-FromDomain: IBMDE
To: gropp@XXXXXXXXXXX
cc: mpi-bind@XXXXXXXXXXX,mpi-core@XXXXXXXXXXX,mpi-io@XXXXXXXXXXX,Message-ID: <C1256763.005934CB.00@XXXXXXXXXXXXXXXXXXX>
Date: Fri, 30 Apr 1999 18:14:19 +0200
Subject: Re: C++ MPI constants
Mime-Version: 1.0
Content-type: text/plain; charset=us-ascii
Content-Disposition: inline
X-UIDL: 668f6ba832ff149fa71aa6bdc67ff2cf




Bill,

Thanks for your comments.

I would like you to consider the following issues:

4. Remove them and use the respective C versions (MPI_SEEK_SET etc.)
====================================================================

This will probably work for the "SEEK" constants since they have
the same type (int) for the C and C++ bindings.

Note, MPI constants do not have the same type in C and C++ in general.
Using C constants with the C++ bindings will work in the most cases
because of the language interoperability operators which are defined
for the C++ MPI classes.

However if you plan to remove all the MPI C++ constants (because we do not
know how
many additional offending constants exist) this probably will lead to some
type conflicts.

For example:

  /* MPI C-bindings */
  ====================

  typedef int MPI_Datarep_conversion_function(void*, MPI_Datatype, int,
void*, MPI_Offset, void*);
  MPI_Register_datarep(..., >>>> MPI_Datatype_conversion_function <<<<<
*read_conversion_fh, ...);

  MPI_CONVERSION_FN_NULL: Probably of type
"MPI_Datarep_conversion_function" in most implementations!

  /* MPI C++-bindings */
  ======================

  typedef int MPI::Datarep_conversion_function(void*, MPI::Datatype&, int,
void*, MPI::Offset, void*);
  MPI::Register_datarep(..., >>>> MPI::Datarep_conversion_function <<<<<
*read_conversion_fh, ...);

  MPI::CONVERSION_FN_NULL: Probably of type
"MPI::Datarep_conversion_function" in most implementations!


So the C and C++ types of these constants are different and the C constants
cannot be used here
for the C++ bindings.

Other examples are:

- MPI_DUP_FN
- MPI_NULL_COPY_FN
- MPI_NULL_DELETE_FN
- MPI_COMM_NULL_COPY_FN
- etc.

Using this solution (nr. 4) we would have the two alternatives:
---------------------------------------------------------------

1. Only the known offending Constants MPI::SEEK_SET etc. are removed

      + We can keep the standard almost as it is
      - We do not know how many offending constants exist
        and will come in the future
      - This decreases the orthogonality of the design of the C++ bindings
        because adhoc exceptions are introduced (we tell the programmer
        that he can use the C++ constants every time, except in the case
        of MPI_SEEK_SET, MPI_SEEK_CUR etc.)

2. All C++ constants are removed

      + orthogonal solution
      + The user has only to deal with one 'kind' of constants
      - The type problem I mentioned above (How do we solve this?)



       Michael

------------------------------------------------------------------------
OS/390 USS Parallel Environment Development
Phone: +49-7031-16-2360, Fax -4240, Bld 71032-14-126
Email: holzheu@XXXXXXXXXX



William Gropp <gropp@XXXXXXXXXXX> on 04/29/99 11:05:54 PM
Please respond to William Gropp <gropp@XXXXXXXXXXX>
To:   Jeremy Siek <jsiek@XXXXXXXXXX>
cc:   mpi-bind@XXXXXXXXXXX,mpi-core@XXXXXXXXXXX,mpi-io@XXXXXXXXXXX,      Michael Holzheu/Germany/IBM,lumesdaine.1@XXXXXX,
jsquyres@XXXXXXXXXX,jsiek@XXXXXXXXXX,treumann@XXXXXXXXXX, Reinhard
      Buendgen/Germany/IBM
Subject:  Re: C++ MPI constants





At 09:28 AM 4/29/99 -0500, jeremy siek wrote:
>
>Michael Holzheu has brought to our attention a name conflict between
>some macros in stdio.h (SEEK_SET, etc.) and the C++ MPI constants.
>
>He proposes two solutions.
>1. Change all the C++ constants to lowercase, which would be more in
> conformance with the style of the C++ standard libraries.
>
>2. Change just the offending constants by adding a "CPP_" prefix.
>
>Andrew Lumsdaine and I suggest that alternative #1 is the better choice.

That's because #2 is so awful.  Here's a more reasonable alternative

3. Change just the offending constants by adding an "MPI_" prefix.  Of
course, this makes the constants the same as the C versions, which doesn't
seem so bad.  That really makes the choices these:

1. Change to lowercase.

4. Remove them and use the respective C versions (MPI_SEEK_SET etc.)

I'd prefer 4 slightly over 1.

Bill