The reason is multidict.add(key, val) is ten times slower than dict[key] = val.

This is because multidict stores data internally as a list of cythonized _Item objects.
But creation of python (ever cythonized) object is too expensive for our use case.

The solution is using C structs for internal data but Cython has no support for visiting values stored in these structures: tp_visit and tp_clear slots.

Thus for sake of speed we need pure C implementation.