Date: Fri, 26 Mar 1999 09:31:15 -0500 (EST)
From: Nicholas Nevin - Sun HPC High Performance Computing <nevin@XXXXXXXXXXXXXXXXX>
Reply-To: Nicholas Nevin - Sun HPC High Performance Computing <nevin@XXXXXXXXXXXXXXXXX>
Subject: Re: clarification on inter-language interoperability of  attributes
To: gropp@XXXXXXXXXXX
Cc: mpi-core@XXXXXXXXXXX
MIME-Version: 1.0
Content-Type: TEXT/plain; charset=us-ascii
Content-MD5: SnVfQtdUlxzACsd6XLHBTg==
X-Mailer: dtmail 1.2.1 CDE Version 1.2.1 SunOS 5.6 sun4u sparc 
X-UIDL: 13b6da43a7a30fbc2941e59029b26225


> X-Sender: gropp@localhost
> Date: Fri, 26 Mar 1999 07:47:53 -0600
> To: Nicholas Nevin - Sun HPC High Performance Computing <nevin@XXXXXXXXXXXXXXXXX>> From: William Gropp <gropp@XXXXXXXXXXX>> Subject: Re: clarification on inter-language interoperability of attributes
> Cc: mpi-core@XXXXXXXXXXX
> 
> At 04:37 PM 3/25/99 -0500, Nicholas Nevin - Sun HPC High Performance
> Computing wrote:
> >
> >Hi, All. I need some clarification on the behaviour of MPI_Comm_get_attr
> >in the case of an architecture with 32 bit ints and 64 bit pointers.
> >
> >Section 4.12.7 of the MPI-2 standard states that MPI will store
> >internally *address* sized attributes. It also states that an integer
> >valued attribute when accessed from C via MPI_Comm_get_attr will return
> >the address of a pointer to the integer valued attribute. 
> >
> >This seems to imply that MPI_Comm_get_attr will return a pointer to
> >a 64 bit quantity. This hoWhen an integer valued attribute is accessed from C or C++, then
MPI_xxx_get_attr will return the address of (a pointer to) the integer valued attribute.wever will break codes which do this.
> 
> No, it means that the attribute is stored as the address itself.  The thing

This might work for predefined attributes where the address can
point to the integer value in static storage, but won't work for
attributes set by the user in Fortran code.  If you store a pointer
to the integer as the attribute it may point to a temporary which
might no longer exist when you try and access it later.

> that the attribute points to may be an int, a long long, or a structure.
> That's why an MPI_TAG_UB is returned as the address of an int.  We've been
> doing this for a long time with MPICH on many architectures with 32 bit C
> ints and 64 bit pointers (it is one of the reasons why Get_attr returns a
> pointer to the integer containing the tag UB, rather than the value of the
> UB itself).
> 

Unfortunately I left out some parentheses in my paraphrase of the
standard, it actually reads "When an integer valued attribute is
accessed from C or C++, then MPI_xxx_get_attr will return the
address of (a pointer to) the integer valued attribute."  This is
what I had meant to say.

So my question still stands.  Assuming Fortran INTEGER is 32 bits
and MPI_ADDRESS_KIND is 8, do I get back from MPI_xxx_get_attr a
pointer to a 32 bit integer or a 64 bit integer?

I see three cases and the answer may not be the same for all of
them.

1. predefined attributes. The answer may vary according to the 
   attribute (window size is of type MPI_ADDRESS_KIND in Fortran).
2. attributes set in Fortran with MPI_ATTR_PUT
3. attributes set in Fortran with MPI_COMM_SET_ATTR

Of course it would be nice if the answer was the same for all the above cases
which would require that a pointer to a 64 bit integer be returned 
in order to avoid possible data truncation.

-nick.