C.21 appears to mandate the Rule of Five -- is that desirable?

[hsutter]

See also #2164 where this issue is creating false positives in Guidelines checkers such as clang-tidy.

Problem

C.21 mandates the Rule of Five:

C.21: If you define or =delete any copy, move, or destructor function, define or =delete them all

[...]

Enforcement

(Simple) A class should have a declaration (even a =delete one) for either all or none of the copy/move/destructor functions.

Is that too strong?

• Yes, this will catch some errors, as noted: "to avoid unwanted effects like turning all potential moves into more expensive copies, or making a class move-only." That's good.
• But this will also require programmers to write redundant boilerplate for classes that don't need it, such as classes that have user-written copy operations where move cannot be more efficient than copy and the programmer knows that writing only copy will cause move to be treated as copy for their type and know that's okay. Programmers hate being told to make code changes that they know don't change meaning just to satisfy a tool, and it's the kind of thing that causes them to disable a rule as noisy / low-value.

Note: The current rule and Enforcement would flag every pre-C++11 class with a user-defined copy operation.

One potential suggestion

The main noise seems to come from requiring explicit move operations. One change that could address #2164 and pre-C++11 types and most of the required-boilerplate cases would be to not require writing out move operations. Something like this:

C.21: If you define or =delete any copy, move, or destructor function, define or =delete all copy and destructor operations

[...]

Enforcement

(Simple) A class should have a declaration (even a =delete one) for all copy/destructor functions if any copy/move/destructor function is user-declared.

We could also add a note along the lines of "and if you write a copy operation check whether you should also write a more-efficient move operation, to avoid turning all potential moves into more expensive copies" -- as a Note here, or as a separate guideline.

[bgloyer]

I'd vote that C.21 should continue to mandate the rule of five.

It's important that move operations are noexcept (C.66) which may not happen when move falls back to copy. It's hard to know how others use C++ but it seems like there aren't many cases where a user defined copy and move should be the same that wouldn't fall under the rule of zero (C.20).

However the rule should be modified to allow the copy-and-swap idiom by counting pass by value as a substitute for both const& and &&.

[eyalroz]

The specific issue brought up in #2164 could be addressed by expanding the notion of what it means to define any of the 5 methods. The assignment operator A& operator=(A other) (note the lack of reference) should count as having "defined a move assignment operator" and "defined a copy assignment operator" for purposes of the rule-of-5 - provided that actual A& and A&& assignment operators are deleted, since in that case, the class author defined exactly what happen with the assignments typically covered by the the A& and A&& assignment operators.

As for the more general question of mandating the "rule of five" - well, since a "declaration" in the form of =default or =delete is sufficient to meet the rule, I believe it's not too onerous to require programmers to be more explicit with these 5 methods. People are not required to change their class' behavior, just to add some text.