My Little Operator: Friendship Is Magic

[StephanTLavavej]

This changes all of the STL's operators that are required to exist, but that aren't required to be namespace-scope functions, into "hidden friends". (For example, n + vector::iterator.)

Hidden friends are great in two different ways:

• They improve compiler throughput, because they don't pollute unqualified name lookup. Only argument-dependent lookup finds them when they're needed.
• They automatically work with named modules, when their parent class (or a function returning their parent class, etc.) is exported. (Namespace-scope operators would have to be individually exported.)

A few notes:

• span::iterator has used hidden friends since <span>: fix cross-type iterator operations #474 and currently says:

  STL/stl/inc/span

  Line 117 in ea32e86

  _NODISCARD_FRIEND constexpr _Span_iterator operator+(const difference_type _Off, _Span_iterator _Next) noexcept {

• The <iomanip> diff looks horrible, but I've structured it into a series of commits for easier reviewing.
  • The _Elem and _Elem2 template parameters have to "trade places" when becoming hidden friends. We static_assert that they're the same, so there is no potential for mistakes.
• quoted() needed extra attention due to how its operators referred to each other. I recommend diffing the "before" and "after" parts of each commit.
  • First, I made this easier to deal with by moving _Quote_out before _Quote_in.
  • In "Use friendship for _Quote_out.", I had to rename to _OsTraits to avoid a collision, and use class _QuTraits = _Traits both for clarity and to delay evaluation of is_void_v<_QuTraits>.
  • In "Use friendship for _Quote_in, part 1.", instead of calling quoted(), I directly construct a _Quote_out (thanks to the earlier code movement).
  • Finally, "Use friendship for _Quote_in, part 2." is nearly pure code movement.

🐴 🪄 🦄

[miscco]

That is soo much better

[StephanTLavavej]

I'm speculatively mirroring this to the MSVC-internal repo. Further changes can be pushed, but please notify me.

[CaseyCarter]

Ugh, <iomanip> was a mess to review.

[StephanTLavavej]

Ugh, <iomanip> was a mess to review.

Thanks for slogging through it - import-std will be much easier to review thanks to this 😻

[DanielaE]

Stephan,
regarding the second bullet point in your opening post please be aware of CWG2588. We've been discussing this in EWG just last week. The one proposed solution which got consensus is the only reasonable one IMHO. So you're probably safe but I'm not sure if the compiler already implements these semantics yet.

[StephanTLavavej]

@DanielaE Thanks for the heads up! 😸