Siu Kwan Lam
Siu Kwan Lam
The problem seems to have existed for as far back as numba 0.35.0. That's the earliest accessible version in the current version of conda.
BFID: `llvmlite_310`
The fix will be simliar to https://github.com/numba/numba/pull/10052/files#diff-e5763a80bf7607850298f9a6512dd5607b74d800f070585f6b8c428e8c536d54R3740-R3752 because ``` >>> issubclass(np.complex64, complex) False >>> issubclass(np.complex128, complex) True ```
From triage discussion: - `find_potential_aliases()` in DCE is written assuming the `baseobj` (in `getattr(baseobj, attrname)`) can ever only be `numpy.ndarray`. See https://github.com/numba/numba/blob/1ba9c54e395e611bbe8d2dbe7a726488f5a75fbd/numba/core/ir_utils.py#L605. It needs to be more conservative. All non-numpy...
@kc611 I've fixed those files. I'm minimizing the changes to azure config since we will need to update it very soon for new numpy and python.
@tasansal, do you have a `.coveragerc` for the run?
Isn't it caused by https://github.com/TGSAI/segy/blob/main/pyproject.toml#L126C1-L126C36 ? This new release of Numba adds code coverage to numba compiled code. Coverage is not seeing Numba before.
Hm... numba is probably still doing something wrong here with the coverage.
I can pretty much confirm this is Numba's fault and have been working on a fix.
IIRC, to reproduce, use numba.jit to compile a function that uses typed list. Then run coverage on it and make sure you are out of numba's source tree (so it...