[jhale]

This PR description was written with the help of an LLM (Claude), and checked by me.

In dist_symbLU (p[sdz]symbfact_distdata.c), snd_luind is never initialised. It is only allocated when the rank has something to send:

if ( nprocs > 1 && (SendCnt_l || SendCnt_u) ) {
  if (!(snd_luind = intMalloc_dist(SUPERLU_MAX(SendCnt_l, SendCnt_u))) ) {

but it is always passed as the send buffer to MPI_Alltoallv. With ParSymbFact = YES on a non-power-of-two number of processes, the processes outside the symbolic factorisation have nothing to send. They pass whatever is on the stack, with all send counts zero.

That is usually harmless, but if the garbage happens to equal MPI_IN_PLACE ((void *) 1 in Open MPI), that process does an in-place all-to-all while the others do not. The job then hangs, or aborts with:

*** An error occurred in MPI_Alltoallv
*** MPI_ERR_TRUNCATE: message truncated

Backtraces from a hung 3-process run show ranks 0 and 1 in ompi_coll_base_alltoallv_intra_pairwise and rank 2 in mca_coll_base_alltoallv_intra_basic_inplace, all called from dist_symbLU via ddist_psymbtonum.

This PR initialises snd_luind to NULL next to rcv_luind, in all three precisions.

How it was found and reproduced

DOLFINx CI started hanging after we made ParSymbFact = YES the default. It hung on 3 processes and passed on 2 and 4. Only builds with 64-bit int_t were affected, presumably because the stack layout is different. It reproduces without DOLFINx. A C driver solving a 7-point Laplacian on a k^3 grid (ColPerm = PARMETIS, ParSymbFact = YES, one pdgssvx call) hangs intermittently on 3 processes. The setup was SuperLU_DIST 9.2.1 +int64+parmetis, ParMETIS 4.0.3 and Open MPI 5.0.10 on Linux x86_64 (Spack). It hung for k = 8, 12 and 16, and aborted with MPI_ERR_TRUNCATE for k = 4. The same binary passed on another runner, as you would expect for an uninitialised read.

Clang points at the same variable:

SRC/double/pdsymbfact_distdata.c:456:6: warning: variable 'snd_luind' may be uninitialized when used here [-Wconditional-uninitialized]
SRC/double/pdsymbfact_distdata.c:117:57: note: initialize the variable 'snd_luind' to silence this warning

With this change the warning is gone in all three files. The C driver above, built against this branch with XSDK_INDEX_SIZE=64 and ParMETIS 4.0.3, passes on 2 to 7 processes for k = 4, 8 and 16. Because the bug depends on stack contents, the passing runs show that nothing regressed, not that the hang is fixed. The fix follows from the code and the backtraces.

DOLFINx issue: FEniCS/dolfinx#4666

[jhale]

This is unrelated to #230. That issue is a separate ParSymbFact bug, a deterministic segfault in symbfact_dist when refactoring with Fact=SamePattern, and this PR doesn't fix it.