Looks good to me, if only for the increase in clarity.

Looks like this fails for some dtypes. From Travis:

======================================================================
ERROR: numpy.core.tests.test_numeric.TestNonzero.test_count_nonzero_axis_all_dtypes
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/travis/build/numpy/numpy/builds/venv/lib/python3.6/site-packages/nose/case.py", line 198, in runTest
    self.test(*self.arg)
  File "/home/travis/build/numpy/numpy/builds/venv/lib/python3.6/site-packages/numpy/core/tests/test_numeric.py", line 1070, in test_count_nonzero_axis_all_dtypes
    assert_equal_w_dt(np.count_nonzero(m, axis=0),
  File "/home/travis/build/numpy/numpy/builds/venv/lib/python3.6/site-packages/numpy/core/numeric.py", line 413, in count_nonzero
    return a.astype(np.bool_, copy=False).sum(axis=axis, dtype=np.intp)
ValueError: invalid literal for int() with base 10: ''

@shoyer: That's because there's a bugfix this needed from #9848, as I mention in the description.

However, @ahaldane argued the bugfix needed a deprecation cycle, so I've redone this with a workaround.

This still requires #9856 to pass.

That should hopefully trigger a retest now that #9856 is in

@shoyer Tests now pass

I'm struggling to get asv to provide useful comparison benchmarks on my machine, frustratingly.

Two consecutive runs give wildly (~50%) different results on the same commit, so comparisons are meaningless. Even more concerningly, some benchmarks complete instantaneously, and it's a different set each time.

I reckon we just merge this, and see what the online asv has to say

Thanks Eric.

Massive improvement for object arrays: https://pv.github.io/numpy-bench/#bench_core.CountNonzero.time_count_nonzero_multi_axis.

Maybe the special casing of bool was worth having, but the user would do better to call sum directly in those cases anyway.

Although actually, a performance hit for very large object arrays, presumably because we now allocate a large boolean array too.