refactor: Replace h.get_type() with py::type(h)

[henryiii]

With the addition of #2364, a very natural extension to py::type<T>() would be py::type(h), mimicking py::cast. I would propose we depreciate and eventually remove h.get_type() in favor of py::type(h). Reasons:

• Free functions are generally better than member functions in C++; they are simpler, composable, etc.
• It would simplify the mixin, which currently gets added to several different classes, and keeps the API closer to Python.
• h.str() is already deprecated in favor of py:str(h).

It is only used internally, not publicly in any docs or tests, so to me it seems it would read better (more python like) and simplify the mixin.

For example, randomly picking one of the 10 places it's currently used:

if (!obj.get_type().is(matrix_type))
// Becomes
if (!py::type(obj).is(matrix_type))

For comparison, the Python would be:

if not type(obj) is matrix_type:

(Discussion moved from #2364).

[YannickJadoul]

👍 Yup, you already convinced me in #2364! ;-)

[YannickJadoul]

We could make this similar to py::str, though, and solve #1692 at the same time. (#2364 hasn't been released yet, so we can still change that interface?)

[henryiii]

Technically, I think that issue would be solved by py::type<std::string>()(), if we supported str/int/etc. (which we don't yet, in that PR)? I don't think there was a need for a typeobject type expressed there, but rather a way to get the type as a handle?

[YannickJadoul]

Right, but I'm thinking we could create a class type : public object {, just like with str? That would allow to e.g. use it in signatures, call py::insinstance<py::type>(...), etc.

[henryiii]

Then we would lose py::type<T>(), which only works if py::type is a function. I'm not sure we would really gain much from it being a class. Also, you'd run in to the same issue with str, where it can't be a handle since it's based on object.

[YannickJadoul]

Agh, I had naively thought/hoped you could make make a templated template <typename T> type() {} constructor, but apparently that's not allowed for constructors :-(

Best I can then think of is:

class type {
public:
    template <typename T>
    static type of() { return type(typeid(T)); }

    explicit type(const std::type_info &type_info) : m_type_info(type_info) {}

    void print() {
        std::cout << m_type_info.name() << std::endl;
    }

private:
    const std::type_info& m_type_info;
};

This of<T> pattern already is present for dtype, I believe.

Not sure if that's worth it, but I do like the conformity of having a py::type class for PyType_Type (just like we have py::int_, py::dict, etc).
Then again, I'm not sure if this is now sooo crucial. I don't see if people would use it that much, so I can live with "just" making this a function. But worth mentioning/discussing, at least, before we fix this API and can never use py::type again for anything else.

[henryiii]

I'd be okay with py::type::of<T>(); not quite as nice, but py::type for PyType_Type sounds good.

[rwgk]

FWIW: I looked in the Google codebase: I see .get_type() used, but only in 8 files, 14 hits total (we have ~360 pybind11 extensions). — This looks like a fairly easy change, if there is a deprecation period.

[YannickJadoul]

I'd be okay with py::type::of<T>(); not quite as nice, but py::type for PyType_Type sounds good.

I know :-/ I'm quite conflicted as well. I quite like how py::type<T>() looks, and it makes sense (similar to py::cast, etc), but ...
Of course, there's probably nothing more py::type (wrapping PyType_Type) would do, compared to py::object, apart from py::isinstance?

@wjakob, any thoughts on this? Having py::type<T>() and py::type(handle) as functions (conform py::cast), vs. having py::type as a class wrapping PyType_Type with a py::type::of<T>() and a py::type(handle) converting constructor (conform e.g. py::str and py::dtype) ?

[wjakob]

+1 for this interface in general. I'm kind of torn about the specifics of naming it as well, but leaning slightly towards consistency with py::str, which also has the advantage of subsuming everything under a common prefix. (Otherwise we'd have find some different identifier for PyType_Type, which seems like it's going to become very confusing).