On Wed, Jun 27, 2012 at 11:59 AM, N.M. Maclaren <[email protected]> wrote:

It's not the compiler that is the issue, because MPI needs only the C facilities that are available in the C subset of C++.

But C (90 or otherwise) isn't a subset of C++, though I don't see it being relevant.

It's what is defined in the various standards.

I was hoping to get time to respond in a more carefully thought-out message, but am in a WG5 meeting, so this is in haste.

Using C99 is NOT a solution, largely because the requirement is for use from C++ and std::complex and std::bool do not have equivalents in C99 (_Complex ,and _Bool are semantically different, and are not required to match those.

So in C++ sizeof(bool) is not required to be 1. It's not clear to me whether the integer rank inequalities in C99 require it to be equal to 1 (like char). In either case, there is no guarantee that these have the same size, thus we seem to need a MPI_CXX_BOOL for completeness (though no science will ever be impeded by lack of support for bool). It is certainly the case that C99 _Complex has the same data representation as the corresponding std::complex type, cf. C99 6.2.5.13 and C++11 26.4 paragraph 4. Note that MPI does not care about differences in conversion semantics that account for the most commonly encountered incompatibilities between C99 and C++ types.

Also, the required language is C90, because C++03 inherits from it, and C99 has been rejected (for VERY good reasons) by about half of the C-using communities.

While C++11 nominally inherits from C99, it has not included _Complex or _Bool, and has kept a lot of C90 semantics in cases where C99 and C90 are incompatible. Most of those don't affect this, but one aspect does.

Someone mentioned that C99 _Complex has better semantics than std::complex. I am afraid that is the converse of the truth, especially when you bring in the utterly appalling Annex G.

What is the problem with Annex G (which is not normative)? C++ doesn't specify anything in particular, nor even suggest that division ought to be implemented in a stable way.

C++11 has included the function calls for IEEE 754 support, but neither the pragmas nor the semantics, and the combination is therefore totally undefined.

Sure, so C99 semantics conform to C++ requirements because C++ doesn't bother specifying anything.

Yes, this is an ungodly mess. I don't agree with the forum's decision, but let's accept it. The only reasonable approach is the one of having MPI types for C++'s types, which could still be used via the C interface from a C++ compiler. I believe that is what you were referring to, and what is intended.

As in my first email, I think that MPI_CXX_COMPLEX et al should be defined, but I haven't seen anything to indicate that an implementation would not be able to simply make them aliases to MPI_C_COMPLEX.