[vchuravy]

MPI.Buffer(data) has a method only for Array (plus Ref, some SubArrays and, via extensions, GPU arrays). Since #941, Base.cconvert(MPIPtr, x) works for any AbstractArray through mpi_ptr_type, so the remaining blocker for using other dense arrays directly as communication buffers is Buffer itself. For example, MPI.Send(m, ...) or MPI.Allreduce!(m, +, comm) with m::Memory{Float64} or a FixedSizeArray from FixedSizeArrays.jl throws a MethodError.

This PR changes Buffer(arr::Array) to Buffer(arr::DenseArray). A DenseArray stores its elements contiguously in column-major order, so sending length(arr) elements of eltype(arr) is correct. Pointer conversion is unchanged and still has to be provided by the array type (mpi_ptr_type and Base.cconvert/Base.unsafe_convert). Types that can't provide a pointer fail at the ccall as before.

BitArray is a DenseArray{Bool}, but it stores packed bits, so it gets an explicit ArgumentError. Before this PR, Buffer(::BitArray) was a MethodError.

The existing Buffer methods for CuArray, ROCArray and oneArray in the extensions are now redundant, since those are DenseArrays. I left them in place to keep this PR minimal.

Motivation

Trixi.jl is replacing unsafe_wrap(Array, pointer(v), dims) with FixedSizeArrays that wrap resize!able vectors, and some of those arrays are passed to MPI. A package extension in FixedSizeArrays.jl (JuliaArrays/FixedSizeArrays.jl#198) covers released MPI.jl v0.20.x; with this PR and #941, no extension would be needed.

Tests

The new test/test_buffer_densearray.jl uses a minimal DenseArray wrapper type and Memory (on Julia ≥ 1.11). It checks the Buffer fields, a ring Isend/Irecv!, Allreduce! and Bcast!, and that BitArray throws. It runs only for the CPU ArrayType.

Locally (Julia 1.12.6, MPICH_jll), with mpiexec -n 2 and -n 3:

• The new test passes, and fails without the change (MethodError for MPI.Buffer).
• test_sendrecv, test_gatherv, test_scatterv, test_allreduce, test_bcast, test_subarray, test_reduce and test_onesided pass.
• Test.detect_ambiguities(MPI) reports no ambiguities involving Buffer.

I haven't run the full suite or the GPU backends.

Opening as a draft to check whether this direction is welcome. Happy to also remove the now-redundant GPU Buffer methods or add a NEWS/docs entry.

This PR was prepared with substantial assistance from Claude Code (Anthropic's AI coding assistant). I reviewed the changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MEJYUviAUshj5fZt8LTe8A