Why shouldn't we replace `_Ref_fn` with `reference_wrapper`?

[frederick-vs-ja]

In #4025 there's attemption to strengthen the exception specification of _Ref_fn::operator().

Given the standard class template std::reference_wrapper is already a invocable wrapper and its operator() has a propagating exception specification (thanks to LWG-3764), it seems OK to replace the uses of _Ref_fn with reference_wrapper.

Pros:

• user codes should only use reference_wrapper, so we might be able to instantiate less specializations.

Cons:

• reference_wrapper is slightly heavier than _Ref_fn;
• reference_wrapper is assignable, which might be unwanted for the internal details of the implementation.

[CaseyCarter]

IMO we should not replace _Ref_fn with reference_wrapper, because reference_wrapper is substantially heavier. _Ref_fn is an aggregate so it can be enregistered; reference_wrapper is not. This was the original motivation for creating _Ref_fn; "we can't accidentally assign to it" is nice, but easily achievable with const reference_wrapper.

[frederick-vs-ja]

Casey Carter (@CaseyCarter) Thanks for explonation! Is this design choice documented anywhere?

[CaseyCarter]

Sadly no. Our implementation documentation practices were nearly non-existent before open sourcing. The team was so small (1-3 maintainers) that everyone either wrote or reviewed every change, so we didn't really need documentation. Let's turn this into a documentation issue, and I'll submit a quick comment change to fix it.