Fixing SystemError when nb_bool/nb_nonzero sets a Python exception in type_caster<bool>::load

[YannickJadoul]

If the error indicator is set by a failing conversion in type_caster<bool>::load (e.g., by trying to convert a numpy.ndarray of 2 or more values, resulting in ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()), the error is not noticed and Python complains that a non-NULL value is returned while an error has been set (SystemError: [...] returned a result with an error set).

This PR should fix that.

NOTE: Without the changes to cast.h, the newly added test fails:

terminate called after throwing an instance of 'pybind11::error_already_set'
what(): SystemError: <class 'type'> returned a result with an error set

Is that not unexpected as well? Shouldn't all error_already_set exceptions be caught by pybind11?

[YannickJadoul]

(Thanks to @tdegeus for reporting on Gitter and sticking along while debugging, btw!)

[YannickJadoul]

Is that not unexpected as well? Shouldn't all error_already_set exceptions be caught by pybind11?

OK, this is because py::repr and py::str are called in cpp_function::dispatcher when trying to report that the function could not be called, but this is outside the big try-catch. Somewhere in there, Python notices that there's been an error, returns NULL as result of PyObject_Repr, and pybind11 throws an error_already_set, causing the Python interpreter to crash.
Not entirely clean/expected, but independent from the problem in this PR.

Ready to merge, as far as I'm concerned, then!

[tdegeus]

Thanks @YannickJadoul !

[wjakob]

What strikes me as odd about this PR is why this is something specific to the bool caster? Can't these types of errors crop up elsewhere as well?

[YannickJadoul]

@wjakob The main problem is calling into the C API without checking the result (i.e., if it's -1, in this case). If that call goes wrong, PyErr_Occurred() will be not-NULL, but pybind11 continues because it expects errors to throw a error_already_set exception.

When other casters call into the C API, they do check the result for errors and clear if necessary. For example:

integer-types caster, type_caster<T, enable_if_t<std::is_arithmetic<T>::value && !is_std_char_type<T>::value>>:

pybind11/include/pybind11/cast.h

Line 1024 in a6355b0

bool py_err = py_value == (py_type) -1 && PyErr_Occurred();

string_caster<typename StringType, bool IsView = false>:

pybind11/include/pybind11/cast.h

Line 1219 in a6355b0

if (!temp) { PyErr_Clear(); returnfalse; }

[wjakob]

Ok, fair enough -- merged, thanks!

[YannickJadoul]

The two Travis CI jobs are failing because something (brew in one case) times out; it shouldn't be a problem with this PR.