Fix random MPI crashes caused by array out-of-bounds access

[ChristopherMayes]

Fix random MPI crashes caused by array out-of-bounds access

Summary

Fixes intermittent segmentation faults when running with MPI (e.g. mpirun -n 8 ImpactZexe-mpi). The crashes are non-deterministic because they depend on how particles are distributed across ranks at runtime.

A test case that reliably reproduces the issue is attached as error-test.zip.

Root Cause

Five array out-of-bounds bugs, all exposed by compiling with -fcheck=all:

1. src/Contrl/Input.f90 — obtype(0) when parsing comments

The lattice input parser reads lines in a loop, incrementing index i only for data lines (non-comment). However, the -99 end-of-lattice check on obtype(i) was outside the if(comst.ne."!") block, so when the first line is a comment, i is still 0 and obtype(0) is accessed.

Fix: Move the obtype(i).eq.-99 check inside the data-reading branch.

2. src/Contrl/Output.f90 — glbin(0) in 12 percentile search loops

The 90th/95th/99th percentile emittance calculations use cumulative histograms. Twelve do i = 1, nbin loops access glbin(i-1), which gives glbin(0) when i=1. The cumulative sum is built starting from i=2, so glbin(0) is never initialized and is out of bounds.

Fix: Start all 12 search loops at i = 2 instead of i = 1. This is safe because the interpolation formula uses glbin(i-1) and glbin(i), and glbin(1) already holds the raw count for bin 1.

3. src/Func/Ptclmger.f90 — zero-sized temp1 allocations in particle exchange

When a rank has zero particles to send/receive in a given direction, jleft, jright, jdown, or jup can be 0. allocate(temp1(9, 0)) creates a zero-sized array, but the subsequent MPI_RECV(temp1(1,1), ...) accesses element (1,1) which is out of bounds.

Fix: Use allocate(temp1(9, max(..., 1))) for all four directions.

4. src/Func/Ptclmger.f90 — zero-sized left/right/up/down allocations

Similar to above: when Nptlocal or numbuf is 0, nsmall becomes 0, leading to zero-sized allocations for the directional send buffers. The MPI_SEND(left(1,1), ...) call then accesses out-of-bounds memory.

Fix: Use max(Nptlocal, 1) / max(numbuf, 1) when computing nsmall.

5. src/Appl/BeamBunch.f90 — rho indexed out of bounds in deposit_BeamBunch

In the charge deposition routine, particle grid indices (ix, jx, kx) can fall outside the local subdomain bounds. With 8 MPI ranks, indices like 33 were computed for an array with upper bound 3.

Fix: Add a bounds check that skips particles whose indices fall outside the local rho grid.

Testing

• Reproducer: error/ directory (attached as error-test.zip) — 10,000 particles through a 4-dipole chicane with extended diagnostics, mpirun -n 8.
• Before fix: ~1 in 5 runs crash with SIGSEGV (signal 11). 10/10 runs fail with -fcheck=all.
• After fix: 20/20 runs pass (release build), 10/10 runs pass (debug build with -fcheck=all).
• Regression: All three bundled examples (Example1, Example2, Example3) still run correctly.

error-test.zip

Files Changed

File	Change
src/Contrl/Input.f90	Move -99 check inside data branch
src/Contrl/Output.f90	12 loops: do i = 1 → do i = 2
src/Func/Ptclmger.f90	6 allocations: guard against zero size
src/Appl/BeamBunch.f90	Bounds-check particle indices before deposition

Acknowledgement

Bug diagnosis and fixes developed with AI assistance (GitHub Copilot, Claude Opus 4.6). The original problem shows up in https://christophermayes.github.io/lume-impact/examples/z/elements/csr-zeuthen/.