Update MPI_Status definition to give implementers flexibility

[dalcinl]

No description provided.

[jeffhammond]

i'll merge after MPI Forum discussion

[dalcinl]

I'm afraid we cannot give a lot of flexibility to implementations. They will have to work with what what they are provided, otherwise alignment issues will break binary interoperability across implementations.

Rather than keeping discussing it in the abstract, let me trough some code to it to make my point.

#include <stdio.h>
#include <stdint.h>

typedef int64_t MPI_Count;

typedef struct {
  int MPI_SOURCE;
  int MPI_TAG;
  int MPI_ERROR;
#if FLAG == 0
  int MPI_internal[5];
#else
  int       MPI_internal_cancelled;
  MPI_Count MPI_internal_count;
  char      MPI_internal_unused[2*sizeof(int)];
#endif
} MPI_Status;

int main() {
  printf("size:  %d\n", (int)sizeof(MPI_Status));
  printf("align: %d\n", (int)_Alignof(MPI_Status));
}

$ cc -DFLAG=0 status.c; ./a.out
size:  32
align: 4

$ cc -DFLAG=1 status.c; ./a.out
size:  32
align: 8

Unfortunately, looks like we cannot give implementations absolute freedom about what to put in MPI_Status. They can have some, but they should at guarantee that whatever they put, the final MPI_Status has the same alignment as int.
Therefore, unless proven wrong with actual evidence, I stand by the original proposal of using int MPI_internal[5];

Implementations, welcome to the new world order!

[jeffhammond]

we can use C11 https://en.cppreference.com/w/c/language/_Alignas to force 8 byte alignment