Test with PYTHON_JULIACALL_HANDLE_SIGNALS=yes

[cjdoris]

Hopefully fixes this issue: #641 (comment)

[cjdoris]

If this works, I don't fully understand why, because we never seem to actually ever dereference a whole PyTypeObject, we use UnsafePointers to access individual fields, which should be fine, since these fields are available in all supported versions of Python.

Perhaps it's a bug in UnsafePointers.jl.

[cjdoris]

OK good - what I did didn't actually fix the problem, the bus error persists, so I have reverted the commit. Testing all minor versions of Python to see how bad this is.

[MilesCranmer]

Python 3.8 is EOL, likely good to remove anyways? pip will resolve to an earlier version of juliacall anyways, should someone want it. And that version would still be compatible with the old python type.

i.e., https://endoflife.date/python

[cjdoris]

Some trivial changes resulted in segfaults on linux on Python 3.9 and 3.11. Reverting those to see what happens.

[cjdoris]

Python 3.8 is EOL, likely good to remove anyways?

Yep agreed - I do aim to only support supported Python versions, just hadn't spotted that 3.8 was EOL now.

[cjdoris]

After reverting, everything passed. Undoing the revert again...

[MilesCranmer]

julia> Cushort === UInt16
true

[cjdoris]

Not necessarily on all systems.

[MilesCranmer]

Oh I see, sorry.

[MilesCranmer]

Out of curiosity, which systems would this actually not be true for?

[giordano]

In Julia the "C*" aliases are best guesses based on observed experience of common compilers (mainly gcc and clang), and in this specific case Cushort is indeed always UInt16: https://github.com/JuliaLang/julia/blob/11eeed32e8fe3cd24c8439fbe2220d3f4183f306/base/ctypes.jl#L27. But the reality is that the C standard doesn't specify fixed size for types, and leaves the implementation up to compilers which can do what the heck they want. "short" is only expected to be at least 16 bits, but a compiler making "short" 17 bits could be fully standard-compliant.

[cjdoris]

Now just Ubuntu Python 3.9 failing (segfault). Rerunning...

[cjdoris]

All passed after rerunning. I think the real issue is just that we didn't add PYTHON_JULIACALL_HANDLE_SIGNALS=yes when we made the tests multithreaded - and this is nothing to do with the recent issues with PyTypeObject.

[MilesCranmer]

Cool; makes sense