DTilde test causes a bus error on MacOS M2

[samcunliffe]

Not caught in CI because pathological to MacOS on ARM64. Not reproducible on any other OS we've tried so far. Nor Mac M1 (thanks @t-young31).

The first SECTION macro in the DTilde test somehow causes a bus error (access to nonexistent memory). So far it seems that the call to mxDestroyArray is responsible.

Note
Links are at 3d0fe2c so that they're permanent. Likely this was introduced in earlier commits (#159, #162).

Low priority for now. Especially given that it's likely to be made obsolete by #70, but I'm making a ticket so I don't forget to investigate and check it goes with MATLAB.

Ideas

• Update MATLAB (my Mac has 2022a and the CI runners are on b).
• Try gdb on linux. lldb is not helpful.
• Check matlab_array is non-empty somehow, intercept.

Links

Might relate to

• Memory leak in TDFieldExporter2D::export_field #129

[samcunliffe]

So adding debug output it’s the mxDestroyArray that causes it. Trail end of test_DTilde.cpp...

  // tear down
  SPDLOG_DEBUG("About to teardown");
  mxDestroyArray(test_input);
  SPDLOG_DEBUG("Called mxDestroyArray");
}

Results in:

➤ ./tdms_tests "DTilde" 
Filters: DTilde
Randomness seeded to: 1507720952
[2022-11-17 15:26:21.956] [debug] [test_DTilde.cpp:23] num_det_modes() == 0
[2022-11-17 15:26:21.957] [debug] [test_DTilde.cpp:25] !x.has_elements()
[2022-11-17 15:26:21.957] [debug] [test_DTilde.cpp:27] !y.has_elements()
[2022-11-17 15:26:21.957] [debug] [test_DTilde.cpp:30] REQUIRE macro is satisfied
[2022-11-17 15:26:21.957] [debug] [test_DTilde.cpp:96] About to teardown
zsh: bus error  ./tdms_tests "DTilde"

[willGraham01]

Check matlab_array is non-empty somehow, intercept.

mxDestroyArray is robust in the other test cases where it is called on an empty struct. What happens if on line 18 we explicitly initialise mxArray *matlab_input = nullptr? mxDestroyArray is supposedly robust in this case too, and matlab_input isn't actually set in the 1st test case.

As an asside: we need to be careful what we mean by "empty" since MATLAB distinguishes between "the struct with no fields" and a "0-by-x" dimensional struct. Both of which are seemingly empty.

[samcunliffe]

"I know, in theory, how to fix it."

-- @willGraham01 5th June 2023