[QUESTION] How to efficiently stop iteration?

I defined an iterable class in pybind11 by creating a __getitem__ method which throws a std::out_of_range exception if the index is out of range. While this successfully stops the iteration in python (without raising a python exception), it appears that stopping the iteration takes about 100us vs 1us for a native implementation, so ~100x slower. I tried the following things:

A. PyErr_SetString(PyExc_StopIteration, "..."); throw error_already_set();: takes ~25us, so ~4x faster than std::out_of_range but still ~25x slower than the native version.
B. PyErr_SetString(PyExc_StopIteration, "..."); and then propagating the error manually, without throwing a C++ exception. This takes ~5us, so ~20x faster than std::out_of_range and "only" ~5x slower than the native version.

However, option B requires returning a NULL PyObject to python, which the dispatcher doesn't like- it triggers the error here: https://github.com/pybind/pybind11/blob/master/include/pybind11/pybind11.h#L924-L928. In order to run option 2, I have to patch those lines to check PyErr_Occurred() and only set an error if none is already set.

so, my questions:

1. Is there any equally or more efficient way to define an iterable object in C++ using pybind11 than __getitem__ and option B above?
2. If the answer to that is "no", can cpp_function::dispatcher be patched to allow setting a python error directly, without throwing a C++ exception?

Start with the cpp_function::dispatcher code in include/pybind11/pybind11.h around lines 924-928, then compare the reported __getitem__ approaches and their timings. Determine whether direct propagation of an existing Python error can be supported without a C++ exception, and verify that iteration stops without replacing the pending error.