fix: use MPI_STATUS_SIZE=6 for Open MPI (avoids 4-byte stack overflow) - #137

Summary

• Open MPI 5.x uses a 6-int (24-byte) MPI_Status, but the Fortran binding hardcoded MPI_STATUS_SIZE = 5. This causes the native MPI_Recv (and any other status-writing call) to write 4 bytes past the end of the Fortran tmp_status buffer allocated inside our wrappers, corrupting the caller's stack.
• Depending on the caller's exact stack layout, this manifests as a hard segfault deep inside a subsequent runtime call (e.g. _lcompilers_string_format_fortran) at a suspicious address such as 0x7fff00000000 — the upper half of a stack pointer left over after the 4-byte overflow.
• Gate MPI_STATUS_SIZE on the existing -DOPEN_MPI preprocessor symbol (already used elsewhere in mpi_c_bindings.f90) so it matches the underlying C header:
  • Open MPI → 6
  • MPICH → 5 (unchanged)

Rationale

Verified against Open MPI 5.0.10 that sizeof(MPI_Status) == 24 and the Fortran binding shipped by mpi.mod exposes MPI_STATUS_SIZE = 6. Our own binding must agree, otherwise we hand C a smaller buffer than it expects.

Verification

• Repro (before): tests/recv_1.f90 crashes with Segmentation fault: address not mapped to object at address 0x7fff00000000 inside print_into_string of the LFortran runtime.
• After this fix: the full tests/run_tests.sh suite passes (57/57 test-rank combinations, including recv_1) at the currently pinned upstream commit 31033d3 using LFortran built with LLVM 15 on Linux/x86_64 with Open MPI 5.0.10.

Test plan

• FC="lfortran --cpp" ./tests/run_tests.sh passes end-to-end with Open MPI 5.
• Re-check under MPICH (unchanged path) — no source changes on that leg beyond additional comments.

Made with Cursor