On Wed, Jan 3, 2018 at 1:07 AM, Thomas Jahns <jahns@dkrz.de> wrote:
>
> On 01/02/18 20:52, Jeff Hammond wrote:
>>
>> If a user passes a (base,size) pair that corresponds to a range of addresses that are not allocated and/or mapped, is that a correct MPI program?
>
>
> Such a thing wouldn't be a valid object pointer in C/C++ or a variable in Fortran, i.e. both languages prohibit this.

Which of the following MPI_Win_create calls are not using a valid object pointer according to ISO C and/or C++?

char a = '\0';

int main() {
..
char * x = malloc(1);
MPI_Win_create(x,1000,1,MPI_INFO_NULL,MPI_COMM_WORLD,&win);
...
MPI_Win_create(&a,1000,1,MPI_INFO_NULL,MPI_COMM_WORLD,&win);
..
foo(); // compiler cannot prove this function does not call
         // mmap(0x10000, 1, PROT_READ|PROT_WRITE, MAP_FIXED|MAP_ANONYMOUS, -1, 4096);
..
MPI_Win_create(0x10000,1000,1,MPI_INFO_NULL,MPI_COMM_WORLD,&win);
..
}
 
>> The only text I can find that implies that buffers passed to MPI_Win_create have to be allocated is the following, assuming that one interprets "existing memory" this way.
>>
>> "Each process specifies a window of existing memory that it exposes to RMA accesses by the processes in the group of comm." (MPI 3.1 11.2.1)
>
>
> Since MPI is a library standard I would still expect all the guarantees of the language standards involved to hold and the above to only be an additional clarification?

It's only a clarification.  However, I don't think it's a good idea for the MPI standard to punt this off on ISO C, because not every implementation of MPI is an ISO C language lawyer, and very few users are.

>> The Fortran-specific text is similarly suggestive, but it only says that a user can pass a simply contiguous array, not that such a thing is required.
>>
>> "In Fortran, one can pass the first element of a memory region or a whole array, which must be ‘simply contiguous’"
>
>
> That's most probably because of the Fortran storage association "feature" which is more about how arrays were passed in pre-Fortran 90 times, i.e. it's legal for a caller C to pass some array element A(i) to a subprogram P accepting an array argument B(*) provided the effective size (that is the indices of B used in P match to valid indices of A).

The widespread use of MPI in the context of non-strict Fortran is a particularly good reason to clarify this ourselves rather than rely on ISO Fortran.  As has been discussed many times in the passed, MPI relies on extensions and/or implementation-specific behavior with Fortran through the latest standard.  Only with Fortran 2018 will this situation disappear.