[4.2rc1] Test failures on armel: no FPEs

Description

On the Debian “ARM 32 bit little Endian Soft Float” (“armel”) platform, I get test failures since the platform seems to handle floating point errors differently. This is a regression, but it could probably be ignored?

Expected behavior

The tests should succeed

Actual behavior

I get the following two failures:

______________________ test_linear_fit_model_set_weights _______________________

    def test_linear_fit_model_set_weights():
       […]
        with pytest.warns(RuntimeWarning,
                          match=r'invalid value encountered in true_divide'):
>           fitted_model = fitter(init_model, x,
                                  np.ma.array(y, mask=np.isclose(weights, 0)),
                                  weights=weights)
E           Failed: DID NOT WARN. No warnings of type (<class 'RuntimeWarning'>,) was emitted. The list of emitted warnings is: [].

astropy/modeling/tests/test_model_sets.py:392: Failed

and

____________________ test_blackbody_exceptions_and_warnings ____________________

    def test_blackbody_exceptions_and_warnings():
        […]
        with pytest.warns(AstropyUserWarning, match='invalid') as w:
            bb(0 * u.AA)
>       assert len(w) == 3  # 2 of these are RuntimeWarning from zero divide
E       assert 1 == 3
E        +  where 1 = len(WarningsChecker(record=True))

astropy/modeling/tests/test_physical_models.py:119: AssertionError

Steps to Reproduce

Build and test the package on armel

System Details

• Python 3.8.6, 3.9.0
• Numpy 1.19.4
• Astropy 4.2rc1
• Scipy 1.5.3
• Matplotlib 3.3.2

[olebole]

Sure, this is not a problem at all. I will just remove the check/assumption for zero divide warnings in these places. armel is (in my opinion) also a quite exotic arch for astropy (all recent Raspberries run with armhf or a variant of), and we should not put too much efforts here for an architecture where floating point is emulated by software. That is anyway not where you want to run astropy, and the problem is probably rather the emulation than astropy itself. I reported this rather since it may point to a real problem and the code author should know it. This could be tagged with “wontfix” and closed.

[olebole]

@plim (guessing the question was general), I checked my patches, and I think they are all not useful for the general. However I found some that disabled failing tests which are claimed to be fixed now – I will remove and check them tomorrow. Thank for reminidng to look!