[mpich-discuss] MPI_Get on the same memory location

[Balaji, Pavan]

On Aug 22, 2014, at 7:59 AM, Balaji, Pavan <balaji at anl.gov> wrote:

> On Aug 21, 2014, at 8:42 PM, Nick Radcliffe <nradclif at cray.com> wrote:
> 
>>> If buffers overlap, use memmov and not memcpy:
>> 
>> Good point, you would have to use memmov in general for overlapping source and target buffers. I was thinking of the special case where the source and target buffers are identical.
>> 
>>> I don’t think so.  It’s an incorrect program to get from a buffer to itself.  MPICH is being helpful is catching this user error.
>> 
>> Do you know where in the standard this is specified? I'm having trouble finding it.
> 
> The parameter in PUT is a “const void *”.  It’s read-only for MPI.

It’s less clear for GET, but the standard says it’s equivalent to sending and receiving, which should imply overlapping buffers is invalid.  Either way, I can’t imagine that not being the intention of the RMA working group (I’m a part of it).

In MPI-2, this was not a problem at all since we couldn’t send data from the local public window to a remote public window (two simultaneous locks were not allowed; FENCE didn’t allow load/store and PUT/GET in the same epoch, etc.).  In MPI-3 we allowed this, but forgot to make it clear that they shouldn’t overlap.

We should try to make it clearer in the standard, though I don’t think it’s valid even in MPI-3.

  — Pavan

--
Pavan Balaji  ✉️
http://www.mcs.anl.gov/~balaji