std::string_view must be std::is_trivially_move_constructible_v

[Igor Zhukov (fsb4000)]

Thanks A. Jiang (@frederick-vs-ja) for finding the issue at #2111

We didn't have move constructor and after adding range constructors (6fe02ac) string_view was no longer trivially move constructible.

reduced example(which I wanted to send to FE team (because A. Jiang (@frederick-vs-ja) thought it was a compiler issue)): https://gcc.godbolt.org/z/qqWbfce6s

Fixes DevDiv-1510475/VSO-1374748/AB#1374748.

[A. Jiang (frederick-vs-ja)]

Oh... You're right.
Perhaps what we need is an additional constraint (!same_as<remove_cvref_t<_Range>, basic_string_view>). I've submitted an LWG issue.

[Igor Zhukov (fsb4000)]

A. Jiang (@frederick-vs-ja)
Yes, you are right. I should change range constructor.
I opened: https://en.cppreference.com/w/cpp/string/basic_string_view/basic_string_view and I didn't find

constexpr basic_string_view(basic_string_view&&) noexcept = default;

constructor.

[Jonathan Wakely (jwakely)]

Perhaps what we need is an additional constraint (!same_as<remove_cvref_t<_Range>, basic_string_view>).

FWIW libstdc++ already has that constraint.

[A. Jiang (frederick-vs-ja)]

I suspect their omission is the simplest way to specify that moves leave the source object unmodified.

Casey Carter (@CaseyCarter) I think this problem can resolved by WG21-P2251.

By the way, there're several types in the standard library whose move functions are suppressed by explicitly defaulted copy functions. The types listed here may be incomplete, some of which are able to propagate non-trivial copy operations:

• std::span
• std::basic_string_view
• std::allocator
• std::istreambuf_iterator
• std::ostream_iterator
• std::complex (specializations for float, double, and long double)
• std::chrono::leap_second
• std::istream_iterator
• std::chrono::duration

I don't know whether it's intended to suppress the implicitly declared move functions. IMO it seems better to leave copy and move special member functions implicitly declared in the specification. (std::pmr::polymorphic_allocator, std::pmr::memory_resource, std::nested_exception, and std::ios_base::Init are likely in different situations.)

[Jonathan Wakely (jwakely)]

IMO it seems better to leave copy and move special member functions implicitly declared in the specification.

I disagree. For a type where moving is identical to a copy I see no benefit to declaring a move constructor or move assignment operator. A separate move constructor makes sense if moving can do something different from a copy, define what moving means. If construction from an rvalue just does a bitwise copy, let that be done by the copy constructor.

The language does the right thing here: suppressing the declaration of the move constructor is exactly what we want.

[Casey Carter (CaseyCarter)]

IMO it seems better to leave copy and move special member functions implicitly declared in the specification.

I disagree. For a type where moving is identical to a copy I see no benefit to declaring a move constructor or move assignment operator. A separate move constructor makes sense if moving can do something different from a copy, define what moving means. If construction from an rvalue just does a bitwise copy, let that be done by the copy constructor.

I do see a benefit in not declaring any SMFs when the implicitly-defined defaults suffice. The most readable spec is the spec you don't have to read at all. I don't think it's worth spending committee time to change the existing types, but if we were standardizing basic_string_view today on a green field I'd champion removing the SMF declarations.

[Casey Carter (CaseyCarter)]

Looks great Igor Zhukov (@fsb4000), thanks for fixing my bug in the implementation. Also thanks to A. Jiang (@frederick-vs-ja) for submitting the LWG issue to fix the corresponding defect I introduced in the Standard. (Yes, I make so many mistakes I need a posse of contributors to follow me around and fix them.)

[Stephan T. Lavavej (StephanTLavavej)]

FYI Casey Carter (@CaseyCarter), Igor Zhukov (@fsb4000) pushed small changes to address my code review comments after you approved.

[Stephan T. Lavavej (StephanTLavavej)]

I'm mirroring this to an MSVC-internal PR. Please notify me if any further changes are pushed.

[Stephan T. Lavavej (StephanTLavavej)]

Thanks again! 😻 I'll concede to Casey Carter (@CaseyCarter) in our label tug-of-war and will changelog this as a bugfix.

[Casey Carter (CaseyCarter)]

Thanks again! 😻 I'll concede to Casey Carter (@CaseyCarter) in our label tug-of-war and will changelog this as a bugfix.

I'll admit that "We provide a non-standard guarantee that basic_string_view is trivially copyable" falls just short of implying that any particular operation is trivial, but I think every non-language-lawyer expects that copies and moves of a trivially copyable type are either invalid or trivial. That's really what most folk who have never attended WG21's Core Working Group think "trivially copyable" means =) So while the corrected behavior didn't break our stated contract it did break our implied contract.