BUG: ensure static_string.buf is never NULL for a non-null string

[ngoldbaum]

Right now if you do something like:

npy_static_string s = {0, NULL};
int is_null = NpyString_load(allocator, packed_string, &s)

Then it is possible to end up in a state where s.buf is NULL even though is_null is 0. It's much simpler to reason about things if that can only happen if packed_string is a null string.

This is causing intermittent CI failures due to UBSAN not liking this:

numpy/_core/tests/test_stringdtype.py::TestStringLikeCasts::test_void_casts[unset-True-strings1] ../numpy/_core/src/multiarray/stringdtype/casts.c:1464:9: runtime error: null pointer passed as argument 2, which is declared to never be null

See e.g. this run. It's not at all clear to me why that run hang for six hours instead of immediately crashing and also why the error from UBSAN never gets reported in the CI output. I was able to reproduce the failure locally though.

I'm hoping that explicitly specifying that we don't want to halt on UBSAN errors will avoid confusing CI hangs like the one above, but can't confirm.

[charris]

C literal strings are NUL terminated, I assume that is taken into account.

[ngoldbaum]

Yes, it just needs to be a valid non-null pointer. The size is always zero in this case. This is mainly to avoid passing NULL to e.g. memcpy, which is UB, or adding NULL checks before doing memcpy. It also makes things easier to reason about IMO.

[charris]

Thanks Nathan.