MPI_Reduce op x datatype parameter check needs improvement

[jjhursey]

Consider the following:

MPI_Reduce(&send, &recv, LIMIT*sizeof(int), MPI_BYTE, MPI_SUM, root, MPI_COMM_WORLD);

The MPI standard (MPI 3.1 - Section 5.9.2) says that this is not allowed, so this would be an invalid MPI program. MPI_BYTE can only be used with the MPI_BAND, MPI_BOR, MPI_BXOR MPI_op types.

Open MPI does not check this in it's ompi_op_is_valid param check here. This results in a wrong answer for the user. But they have an incorrect MPI program so all bets are off. However, Open MPI could implement some stronger parameter checking in the MPI routines that use MPI_Op datatypes.

I have a test program here if someone was interested in working on this (it just tests MPI_Reduce):

• https://gist.github.com/jjhursey/ad4e629e67649cbe2343e3754bef97b8

[bosilca]

The problem is that we convert the MPI datatype into the basic OPAL datatype before the check. As a result the MPI_BYTE gets converted to OMPI_DATATYPE_MPI_UINT8_T), which has acceptable values for all predefined op. The solution might be to use the MPI datatype ID instead of the OPAL one, and have larger and more complex ops arrays (as defined in op/op.c).