[WIP] Use FunctionWrappers for custom reduction operators

[giordano]

My attempt at addressing #404. This is using FunctionWrappers.jl instead of @cfunction, as suggested in a recent JuliaHPC call.

Note: it doesn't work for me even on x86_64 when using more than 1 rank. Simple reproducer:

using MPI
MPI.Init()
let
    T = Int
    dims = 1
    op = (x,y) -> 2x+y-x
    send_arr = Array(zeros(T, Tuple(3 for i in 1:dims)))
    send_arr[:] .= 1:length(send_arr)
    recv_arr = Array{T}(undef, size(send_arr))
    MPI.Reduce!(send_arr, recv_arr, op, MPI.COMM_WORLD; root=0)
end

Run the script with

mpiexecjl -np 2 julia /path/to/script

The program should die with a segmentation fault (but the generated pointer should be non-null, as far as I can tell)

[giordano]

I thought that there was a segfault in the call to MPI_Reduce at

MPI.jl/src/collective.jl

Lines 598 to 600 in 4cd7118

	@mpichk ccall((:MPI_Reduce, libmpi), Cint,
	(MPIPtr, MPIPtr, Cint, MPI_Datatype, MPI_Op, Cint, MPI_Comm),
	rbuf.senddata, rbuf.recvdata, rbuf.count, rbuf.datatype, op, root, comm)

but putting some debug printing shows that the ccall should work fine. I'm wondering if something is something should be GC.@preserve'd.

Edit: "should work fine" was probably a bit optimistic, as the result of the reduction is garbage.