BUG: np.array([0, datetime.now()], dtype=object).sort() can raise on cmp(), but exception ignored

[mattip]

the various functions in quicksort.c.src do not check PyErr_Occurred() before returning, when the tp_compare fails no exception is raised. The exception is thrown later by any random C-API call, but the error originated in the sort function

python -c "import numpy as np; from datetime import datetime as dt; \
                 a = np.array([1, 2, dt.now(), 0, 3], dtype=object); \
                 np.sort(a); print a"
numpy/core/fromnumeric.py:822: RuntimeWarning: tp_compare didn't return -1 or -2 for exception
  a.sort(axis=axis, kind=kind, order=order)
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "numpy/core/numeric.py", line 1942, in array_str
    return array2string(a, max_line_width, precision, suppress_small, ' ', "")
  File "numpy/core/arrayprint.py", line 397, in wrapper
    return f(self, *args, **kwargs)
  File "numpy/core/arrayprint.py", line 529, in array2string
    separator, prefix, formatter=formatter)
  File "numpy/core/arrayprint.py", line 373, in _array2string
    _summaryEdgeItems, summary_insert)[:-1]
  File "numpy/core/arrayprint.py", line 565, in _formatArray
    for i in range(leading_items):
TypeError: can't compare datetime.datetime to int

FWIW, on PyPy we detect that a C-API function raised an exception but returns a valid value, which
is how I found the problem

[eric-wieser]

This only occurs on python 2.

The numpy code falsely assumes that comparisons cannot raise exceptions in python 2