osc/base: Detect unsupported data types and abort

[jjhursey]

Using MPI_MINLOC or MPI_MAXLOC with the following data types
leads to data corruption:

• MPI_DOUBLE_INT
• MPI_LONG_INT
• MPI_SHORT_INT
• MPI_LONG_DOUBLE_INT

Detect this print a error message and abort.
This workaround should be removed once the following issue is resolved:

• MPI_Accumulate Error when using datatype "MPI_SHORT_INT" #1666

Signed-off-by: Joshua Hursey jhursey@us.ibm.com
(cherry picked from commit 94f92f6)
Signed-off-by: Joshua Hursey jhursey@us.ibm.com

[jjhursey]

Refs PR #2832

[hjelmn]

These datatypes and ops need to die :-/.

[jsquyres]

Please consolidate the opal_output and opal_show_help into a single, descriptive/useful show_help message.

[nysal]

@jjhursey I think this patch will only work for osc/pt2pt. osc/rdma will probably need a similar check in ompi_osc_rdma_rget_accumulate_internal()

[gpaulsen]

Can you put this in now, and create a new PR for ompi_osc_rdma_rget_accumulate_internal().

[hjelmn]

osc/pt2pt should be trivial to update to support for minloc/maxloc. It is a slight modification to osc_pt2pt_accumulate_buffer. The hetero #ifdef needs to be removed and the if needs to check if the datatype is non-contiguous.

I have no interest in making either MPI_MINLOC or MPI_MAXLOC work with osc/rdma (outside just treating MPI_SHORT_INT as contiguous) so I welcome a check that makes it error.

[jjhursey]

@nysal I updated the PR to include a commit for ompi_osc_rdma_rget_accumulate_internal() - can you and/or @hjelmn confirm that this is the correct/best placement for that logic in the osc/rdma component?

[jsquyres]

@hppritcha Good to go