fix: null pointer dereference in attr_with_type_hint

[InvincibleRMC]

Description

Fixes regression caught in #5575

Suggested changelog entry:

    Fix bug in ``attr_with_type_hint`` to allow objects to be in ``attr_with_type_hint``

[rwgk]

@timohl looks good to me (the suggestions I posted are pretty minor), could you please take a quick look, too?

[kvarg]

I believe theres still a dereference issue since the function record is also set to a nullptr. This line will be hit if the user has not yet created bindings for the class:

pybind11/include/pybind11/pybind11.h

Line 168 in 97022f8

} else if (func_rec->is_new_style_constructor && arg_index == 0) {

I guess the correct behaviour would be a pybind11 failure?

[rwgk]

I believe theres still a dereference issue since the function record is also set to a nullptr. This line will be hit if the user has not yet created bindings for the class:

pybind11/include/pybind11/pybind11.h

Line 168 in 97022f8

} else if (func_rec->is_new_style_constructor && arg_index == 0) {

I guess the correct behaviour would be a pybind11 failure?

Thanks @kvarg for pointing this out, I totally overlooked that func_rec can be nullptr. I believe we need additional guards in multiple places. We have is_annotation, but that's currently used in only one place.

Ideally, for each additional guard we're adding, we should have a corresponding unit test.

[InvincibleRMC]

I believe theres still a dereference issue since the function record is also set to a nullptr. This line will be hit if the user has not yet created bindings for the class:

pybind11/include/pybind11/pybind11.h

Line 168 in 97022f8

} else if (func_rec->is_new_style_constructor && arg_index == 0) {

I guess the correct behaviour would be a pybind11 failure?

I got it to have the same error as using m.attr() without creating bindings. "AttributeError: 'module' object has no attribute 'foo4'

[InvincibleRMC]

Currently doing the following results in the following.

struct foo4 {};
 m.attr_with_type_hint<foo4>("foo4") = 3;

It makes the type hint test_submodule_pytypes(module_&)::foo4. This is the same result as doing it with m.def() so I assume this is the intended type.

[rwgk]

This seems fine as a fix for a regression.

But I realize now, it's generally not a great pattern to create a dummy object (the function record) to reuse a complex function. Careful factoring would be better, so that generate_type_signature<T>() does not involve a function record at all.

[rwgk]

I fixed a tiny typo: 07c8453

Waiting for GitHub Actions to finish, then I'll merge.