BUG: scipy/_lib/_numpy_compat: get_randint

[jor-]

Fix for problem mentioned in #9486. The low and high values might be chosen in some cases out of the integer type range. This is fixed.

[perimosocordiae]

Makes sense, +1 from me.

[tylerjereddy]

Is the proposal here to remove the test skips for 32-bit arch when NumPy < 1.11? Otherwise, we aren't testing for this anyway iirc.

If there's a really high confidence that this does what we want then the test skip should be removed. Ideally, I'd like to see the wheel-build matrix all-green ahead of time but I guess that's a pain to do for individual PRs in the main repo.

[jor-]

Yes, the test should work for 32-bit arch when NumPy < 1.11. I removed the test skip.

[jor-]

There is one failure:

=================================== FAILURES ===================================
______ build/testenv/site-packages/scipy/signal/tests/test_signaltools.py ______
[gw1] linux -- Python 3.5.3 /usr/local/bin/pypy3
Worker 'gw1' crashed while running 'build/testenv/site-packages/scipy/signal/tests/test_signaltools.py::TestFFTConvolve::test_random_data_multidim_axes[axes5]'

I think this is unrelated to this PR.

[rgommers]

I think this is unrelated to this PR.

yep, failure is unrelated.

[rgommers]

Merged, thanks @jor-

[jor-]

Thanks for merging.