Adding holder_caster typename SFINAE = void hooks to help work around the current lack of smart-pointer interoperability

[rwgk]

The two new SFINAE hooks are provided to help work around the current lack of support for smart-pointer interoperability. Please consider it an implementation detail that may change in the future, as formal support for smart-pointer interoperability is added into pybind11.

Changelog not needed.

[YannickJadoul]

For the (literal) record: @rwgk pointed me to these two use cases: https://github.com/pybind/pybind11_protobuf/blob/a451639e0ac6ca66f1ea1bdddb401e264ac27fc8/fast_cpp_proto_casters.h#L321 & https://github.com/pybind/pybind11_protobuf/blob/a451639e0ac6ca66f1ea1bdddb401e264ac27fc8/proto_utils.h#L38:

E.g.:

// All protos use shared pointers as their holder, so automatically convert a
// unique_ptr to a shared pointer. Note that this requires a patch to the core
// pybind11 to add a SFINAE template parameter to move_only_holder_caster.
template <typename ProtoType, typename HolderType>
struct move_only_holder_caster<
    ProtoType, HolderType, std::enable_if_t<google::is_proto_v<ProtoType>>> {
  static handle cast(HolderType&& src, return_value_policy policy,
                     handle handle) {
    return copyable_holder_caster<ProtoType, std::shared_ptr<ProtoType>>::cast(
        std::shared_ptr(std::move(src)), policy, handle);
  }
  static constexpr auto name = type_caster_base<type>::name;
};

And yes, specializing type_caster should still be the only way, but pybind11 already uses the SFINAE of type_caster in the PYBIND11_DECLARE_HOLDER_TYPE macro, so any straightforward attack on this issue would result in an ambiguous specialization, I believe:

pybind11/include/pybind11/cast.h

Lines 1627 to 1634 in 23c3edc

	#define PYBIND11_DECLARE_HOLDER_TYPE(type, holder_type, ...) \
	namespace pybind11 { namespace detail { \
	template <typename type> \
	struct always_construct_holder<holder_type> : always_construct_holder<void, ##__VA_ARGS__> { }; \
	template <typename type> \
	class type_caster<holder_type, enable_if_t<!is_shared_ptr<holder_type>::value>> \
	: public type_caster_holder<type, holder_type> { }; \
	}}

So once again, our interface is fuzzier than it seems. But the addition of this big fat warning should give us the freedom to blame users if they misuse this part of the internal API.

And I do mean that. Don't dare blaming pybind11 if you rely on this and things change.

[YannickJadoul]

I'm still not entirely happy, but it does seem to be the simplest way of making these use cases work (and the change is totally innocuous to the good users that keep their hands off of pybind11's internal).

[rwgk]

Thanks @YannickJadoul!