reference member

[BjarneStroustrup]

I think we need a rule banning reference members. All we have is a note: (Note that using a reference member is almost always wrong.)

[shaneasd]

Related/Duplicate: #1707

[hsutter]

We have this in C.12 so perhaps a PR to propose updates to that?

[Ivarun]

I’m curious why this is desirable? Aren’t reference members (passed as constructor parameters) a perfectly valid way to do dependency injection in C++?

[eyalroz]

(I'm ignoring the note - which is not a guideline; see #1707 about that one)

While references as members can make a class somewhat unwieldy - non-assignable, non-movable - I can't say I understand why @BjarneStroustrup believes such members should be banned. I'm not even convinced they should be discouraged. When one wants a reference-type member, the three main options seem to be:

• a pointer
• a reference
• an std::reference_wrapper
• (a gsl::non_null pointer)

The appeal of a reference relative to the other two, is that it hides - when used - its referential nature: You use it as you would a value-type. It is often convenient or appealing to me as a class designer that my users don't have to say mystruct.foo but *(mystruct.bar). Also, it is non-nullable, unlike a pointer, which would require the extra hassle of non_null.

Personally, I would avoid references in a class I design as part of, say, the public API of a library. But for something internal which doesn't have to be super robust and has limited scope - I wouldn't strongly oppose their use.

[GuillaumeDua]

From my perspective, a guideline should discourage the use of reference members, as it makes lifetime management harder and may lead to dangling reference;
But when needed (like, for DI for instance), then std::reference_wrapper seems the best option to meet std::semiregular requirements.

[eyalroz]

@GuillaumeDua : Well, discourage is not as drastic than "almost always wrong"...

But let me push back a little...

• std::reference_wrapper is not default-constructible, so - not semi-regular.
• How do std::reference_wrapper or pointers reduce the chance of having a dangling reference? I'd say they would be just as likely to dangle.

[jwakely]

std::reference_wrapper is for passing references through APIs that would otherwise decay them, such as std::bind and std::thread. It's not supposed to be a general purpose semiregular reference type.

[GuillaumeDua]

@eyalroz @jwakely Silly/tired me, you're right. I meant from a copyable and movable perspective.
std::reference_wrapper does not reduce changes of dangling ref.
My point was "avoid to use ref as members, but if you really want/need to, then use ref-wrapper.