[libcxx] adds `std::ranges::iter_move` and `std::iter_rvalue_reference_t`

[ldionne]

Summary

Implements parts of:

• P0896R4 The One Ranges Proposal`

Depends on D99863.

[cjdb]

rebases to activate CI

[miscco]

libcxx/include/iterator

2476

I am slightly concerned about the number of overloads that are introduced and the instantiation cost, which is quite high on clang right now due to lack of memoization of concepts.

I would rather put everything in a single operator() and then use if constexpr to return the right thing.

That would enable us to simplify the concept definitions, as we do not need to exclude the previous concepts

[cjdb]

libcxx/include/iterator

2476

which is quite high on clang right now due to lack of memoization of concepts.

How long ago did Clang disable memoisation?

I would rather put everything in a single operator() and then use if constexpr to return the right thing.

I'm aware of this optimisation from GCC Concepts TS days, but I'd like to see evidence that Clang will get a significant performance boost from this too.
It'll also mean that we can't have conditional noexcept, something I'm reluctant to give up.

[miscco]

I can only speak from experience working on the MSVC ranges machinery. The join_view tests take about 5 min to compile.

I think we would need to call in @saar.raz or @rsmith

[tcanens]

libcxx/include/iterator

2439

This doesn't match the standard. The "poison pill" that doesn't poison anything is there solely to clarify that we only do ADL.

2442

This is missing the "E has class or enumeration type" part of the spec. In particular, we don't want to do ADL for pointers.

2443

This is not in the spec?

2445

nor is the __can_reference here.

2460

__dereferenceable is stronger than the spec, which only requires *E to be well-formed, not that it returns something referenceable.

2476

Conditional noexcept is required for conformance since the wording uses expression-equivalent.

[cjdb]

fixes iter_move per @tcanens' feedback
moves __class_or_enum out of __swap scope since it's useful elsewhere

[miscco]

libcxx/include/iterator

2476

Note that we do have conditional noexcept with MSVC and I believe libstdc++

You have have a look at the _Choice machinery @CaseyCarter build. It is super helpful

2498

What is it with those strange !

[cjdb]

libcxx/include/iterator

2476

I think it would be best to benchmark before adopting a more complex solution in the name of optimisation.

[cjdb]

libcxx/include/iterator

2498

They indicate there are non-visible code points in the text (e.g. zero-width joiners). I think it's a part of the LaTeX translation process for the standard. I'm committed to hunting all of these down, so thanks for the heads up.

[cjdb]

rebases to activate CI

[zoecarver]

Just some super small things. But LGTM!

libcxx/include/concepts

452

Can't this just be = is_lvalue_reference_v<_Tp>;?

455

Same as above but for rvalue reference. Then we could get rid of __core_reference.

457

Nit: want to use your new macro here ;) ?

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

25

Given the discussion about unions and enums, want to add a super simple type traits test for unqualified lookup on both an enum and union?

96

I'm OK with this implementation. You might consider a static global to track this stuff in the future, though. Especially if you need more than a boolean. As I said, though, I'm OK with this implementation here.

139

I think this is a really good test!

187

No need to call assert here.

[cjdb]

rebases to activate CI

[cjdb]

@miscco's suggestion has been gnawing at me for a little while, so I checked it out with someone extremely knowledgeable about both core C++ and Clang. They encouraged benchmarking to be sure, but said they expect if constexpr to be faster than an overload set. They also said it isn't unlikely that the speed boost would be significant, and that there's a good chance the memory usage would be lower too.

With all that in mind, I think the added complexity is justified.

libcxx/include/concepts

455

Subsumption means that we need to have all three if we plan to use both __lvalue_reference and __rvalue_reference (see integral, signed_integral, and unsigned_integral). I can technically get away with just introducing __lvalue_reference, but I'd rather hedge that __rvalue_reference will be necessary down the line and get it right on the first go.

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

187

Don't we want to check the run-time value as well?

[zoecarver]

With all that in mind, I think the added complexity is justified.

I actually think this if-constexpr implementation isn't too bad. While it does add some complexity, it is still readable, and much better than I expected, tbh. But please add a comment above the operator() explaining why we're using this implementation.

[cjdb]

But please add a comment above the operator() explaining why we're using this implementation.

I think the explanation should go elsewhere, with a small comment directing people to that. This way we don't need to repeat the explanation for every CPO and niebloid in the library.
I'm thinking at the bottom of the doc you created; WDYT?

[cjdb]

s/std/_VSTD/
s/is_noexcept/__is_noexcept/

[ldionne]

I disagree with the choice to use if constexpr over overload resolution for compile-time performance reasons. I think we should continue to strive to keep the code simple. No offense, but the latest revision of this patch with if constexpr is *significantly* more complicated than a naive implementation like this (which is similar to what Chris suggested originally, but a bit simpler IMO):

namespace ranges::__iter_move {
void iter_move();

template<class _Ip>
concept __unqualified_iter_move = requires(_Ip&& __i) {
    iter_move(_VSTD::forward<_Ip>(__i));
};

// [iterator.cust.move]/1
// The name ranges​::​iter_­move denotes a customization point object. 
// The expression ranges​::​iter_­move(E) for a subexpression E is 
// expression-equivalent to:
struct __fn {
  // [iterator.cust.move]/1.1
  // iter_­move(E), if E has class or enumeration type and iter_­move(E) is a 
  // well-formed expression when treated as an unevaluated operand, [...]
  template<class _Ip>
    requires __class_or_enum<remove_cvref_t<_Ip>> && __unqualified_iter_move<_Ip>
  constexpr decltype(auto) operator()(_Ip&& __i) const 
    noexcept(iter_move(_VSTD::forward<_Ip>(__i))) 
  {
    return iter_move(_VSTD::forward<_Ip>(__i));
  }

  // [iterator.cust.move]/1.2
  // Otherwise, if the expression *E is well-formed:
  //  1.2.1 if *E is an lvalue, std​::​move(*E);
  //  1.2.2 otherwise, *E.
  template<class _Ip>
    requires !(__class_or_enum<remove_cvref_t<_Ip>> && __unqualified_iter_move<_Ip>) &&
    requires(_Ip&& __i) { *_VSTD::forward<_Ip>(__i); }
  constexpr decltype(auto) operator()(_Ip&& __i) const 
    noexcept(*_VSTD::forward<_Ip>(__i))
  {
    if constexpr (is_lvalue_reference_v<decltype(*_VSTD::forward<_Ip>(__i))>) {
      return _VSTD::move(*_VSTD::forward<_Ip>(__i));
    } else {
      return *_VSTD::forward<_Ip>(__i);
    }
  }

  // [iterator.cust.move]/1.3
  // Otherwise, ranges​::​iter_­move(E) is ill-formed.
};
} // namespace ranges::__iter_move

Bending the implementation backwards to improve compilation times when the problem is actually at the language level (overload resolution being so expensive) and at the specification level (they could have decided to go for much simpler rules with a lot less expensive SFINAE) is not the right way to go. If we want to improve the compile-times of Ranges in libc++, I'm all for finding ways to do that, but not by making our implementation significantly more complex and working around the language. Also, complexifying the implementation is not only a matter of "oh but it's hard to read". It's primarily about shipping a correct implementation, which I think is more important than shipping one that compiles fast.

Don't get me wrong, I spent like 3 years of my life trying to improve the compilation times in the C++ ecosystem by doing things like Boost.Hana and http://metaben.ch. I'm very sensitive to that problem. I just think that throwing if constexpr at the problem here is too little too late.

libcxx/include/concepts

250–253

I really think the remove_cvref_t doesn't belong here. It belongs to the callers of this concept.

Even though this concept is just a private helper, I think we should strive to make it meaningful on its own.

libcxx/include/iterator

59

Missing // Since C++20

2435

Would it make sense to add this in __ranges/iter_move.h?

2457

This is redundant since it's already included in the check for !__lvalue_iter_move<_Ip>.

2463

Not used anywhere.

2513

More strange !s

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

2

Can you please add a test with an iterator that can't be iter_moved to confirm that ranges::iter_move(E) SFINAEs away? With the if constexpr implementation, I don't think that's the case.

[ldionne]

libcxx/include/concepts

252

Why are we checking for union here? Is it because the spec says has class or enumeration type, and a union is technically a class type even though is_class returns false for it? @tcanens Do you know what the intent of the spec is in that regard?

[• Quuxplusone]

libcxx/include/concepts

250–253

Re this and your above comment pinging @tcanens: I believe the intent of the spec is that "class [including struct and union] or enumeration types" are exactly the set of types which are allowed to...

(in the case of iter_move) ...have their own overloaded operator*. (Enums can have a free-function operator* found by ADL. Primitive types cannot have an overloaded operator* at all.)

(in the case of swap) ...have their own free-function swap that might reasonably do something different from the normal swap. (Pointer types can have associated namespaces which end up producing associated swap functions, but those swap functions aren't semantically allowed to do anything except swap the pointers. If this weren't true, our strategy of specializing std::sort for pointer types would be non-conforming. https://godbolt.org/z/G8d6GKbME )

It might be reasonable to rename this concept/trait to something like __causes_adl — but I think __class_or_enum is reasonable too, in that it matches the Standardese term-of-art for "the kind of thing that suffers from ADL." I also think, under this rationale, it makes perfect sense to want to move the cvref-dequalification into the concept/trait itself, because the cvref-qualification of a type doesn't have any effect on whether it suffers from ADL.

[ldionne]

Regarding if constexpr-with-a-single-overload vs normally written code: I just spoke to Chris offline and I'm going to try and chat with Casey about his previous experience with that. I don't want us to make our implementation super complex unless we know for sure it will be a serious problem, and that using if constexpr is going to make a significant difference. I think it is also possible to try to be conscious about reducing the number of overloads and the complexity of the implementation without necessarily going all the way and enforcing a single overload all the time. Let's put this on standby for a few hours until I've gotten more information.

[ldionne]

After talking to Chris, Casey and Richard offline, I believe the following approach is the most sensible:

1. Write things as simply as possible, without purposefully pessimizing compile-times upfront
2. Try to be cognizant of the fact that overloads are not cheap, especially in the presence of complicated SFINAE (which is pretty much all the time with concepts)
3. Go back with more field experience and optimize the compile-time of the implementation based on what we see is truly slowing things down

That way, we're not going to cargo-cult one way of doing things before we have actual data to justify an approach. We'll also be able to optimize just the things that matter in practice, which is far from clear at this early stage.

Based on offline discussion with Chris, I'm going to commandeer this now to go back to the pre constexpr if approach.

[ldionne]

Apply reviewer's comments.

[ldionne]

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

187

Yes, I think we do. I'm leaving the assert in place.

[zoecarver]

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

187

Why do we want to check the runtime value? The end of check_iter_move is return true; we're not "checking" anything.

[cjdb]

libcxx/include/__iterator/iter_move.h

59 ↗

(On Diff #338215)

This noexcept specifier doesn't account for a throwing move.

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

187

Sure, but the contents of check_iter_move need to be evaluated at both compile-time and run-time.

[zoecarver]

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

187

I agree. But why can't you just call check_iter_move normally? Why do you need the assert?

[• Quuxplusone]

libcxx/include/__iterator/iter_move.h

59 ↗

(On Diff #338215)

"move doesn't move," and this function returns decltype(auto), so I think we're okay here. This function either returns an rvalue reference in the if, or URVOs a prvalue in the else.

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

187

+1 @zoecarver

[cjdb]

libcxx/include/__iterator/iter_move.h

59 ↗

(On Diff #338215)

@rsmith is the one who pointed it out to me, so I'm not so sure. Granted, the tests cover seem to be passing, but a looking over them again makes me realise I only tested noexcept specifiers in one direction. We'll need to plug those holes at the very least.

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

187

Oh, you're talking about literally removing assert, not what's inside it. I'm okay with that!

[tcanens]

libcxx/include/concepts

252

More or less what Arthur said below. There is no reason to treat unions differently from other classes in the contexts at issue.

[ldionne]

libcxx/include/__iterator/iter_move.h

59 ↗

(On Diff #338215)

This function either returns an rvalue reference in the if, or URVOs a prvalue in the else.

That was my reasoning when writing that.

but a looking over them again makes me realise I only tested noexcept specifiers in one direction. We'll need to plug those holes at the very least.

Sorry, I'm not quite following, what do you mean by testing noexcept specifiers only in one direction? I'll add the test coverage, but I'd like to understand what's missing.

[cjdb]

libcxx/include/__iterator/iter_move.h

59 ↗

(On Diff #338215)

The tests in this commit exclusively assume noexcept(std::ranges::iter_move(lvalue)) == false and noexcept(std::ranges::iter_move(rvalue)) == true. That was an oversight in my design testing (and what I get for not knowing how to do TDD properly in 2019).

[cjdb]

LGTM pending test update

[cjdb]

libcxx/include/__iterator/iter_move.h

45 ↗

(On Diff #338215)

Please make sure all overloads are [[nodiscard]] before submitting too.

[ldionne]

Address review comments.

[ldionne]

Fix stray comment.

[ldionne]

Will ship it once CI is green per agreement with reviewers.

[cjdb]

🚢

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.nodiscard.verify.cpp

35 ↗

(On Diff #338994)

This is nice :-)

[zoecarver]

A few comments, sorry. Nothing blocking, though, so feel free to commit.

libcxx/include/__iterator/iter_move.h

15 ↗

(On Diff #338994)

Why are these (or at least why does it appear that these) are both importing __class_or_enum? (I think only the latter should have the comment.)

49 ↗

(On Diff #338994)

Non blocking nit: _LIBCPP_NOEXCEPT_RETURN?

libcxx/include/concepts

446–457

Non blocking, nit: _LIBCPP_HAS_NO_RANGES

libcxx/include/iterator

453

I'd rather have this be such that we can import any libc++ header and it will guard against the correct stdlib version, etc. internally. This is how we do it for other "versioning" features, I think we should do it with this, too.

In other words, I'd rather have the _LIBCPP_HAS_NO_RANGES check inside of iter_move.h.

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

24

This header is only used in one place, right? At least, I'd like to see it in this directory and not the parent one. But I don't see any reason it can't just be inlined into this file, if nothing else is using it.

187

Heh, sorry I wan't clear. Yes, that's what I mean :)

[cjdb]

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

24

The intention is for it to be shared with iter_swap tests.

[zoecarver]

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

24

Fair enough. In that case, disregard my comment.

[cjdb]

libcxx/include/iterator

453

+10. This is already causing me issues downstream where I only import headers that expose the necessary names.

[ldionne]

libcxx/include/__iterator/iter_move.h

15 ↗

(On Diff #338994)

It was a typo, fixed.

49 ↗

(On Diff #338994)

I'm really not a big fan of that macro, and it's only defined locally when defining __apply_tuple_impl. Not doing.

libcxx/include/concepts

446–457

Actually, this line isn't touched anymore after rebase, so not changing anything.

libcxx/include/concepts

 // [concept.copyconstructible]
 template<class _Tp>
 concept copy_constructible =
   move_constructible<_Tp> &&
   constructible_from<_Tp, _Tp&> && convertible_to<_Tp&, _Tp> &&
   constructible_from<_Tp, const _Tp&> && convertible_to<const _Tp&, _Tp> &&
   constructible_from<_Tp, const _Tp> && convertible_to<const _Tp, _Tp>;
 
+template<class _Tp>
+concept __class_or_enum =
+  is_class_v<remove_cvref_t<_Tp>> ||
+  is_union_v<remove_cvref_t<_Tp>> ||
+  is_enum_v<remove_cvref_t<_Tp>>;
+
 // [concept.swappable]
 namespace ranges::__swap {
   // Deleted to inhibit ADL
   template<class _Tp>
   void swap(_Tp&, _Tp&) = delete;
 
-  template<class _Tp>
-  concept __class_or_enum = is_class_v<_Tp> || is_union_v<_Tp> || is_enum_v<_Tp>;
 
   // [1]
   template<class _Tp, class _Up>
   concept __unqualified_swappable_with =
-    (__class_or_enum<remove_cvref_t<_Tp>> || __class_or_enum<remove_cvref_t<_Up>>) &&
+    (__class_or_enum<_Tp> || __class_or_enum<_Up>) &&
     requires(_Tp&& __t, _Up&& __u) {
       swap(_VSTD::forward<_Tp>(__t), _VSTD::forward<_Up>(__u));
     };
 
   struct __fn;
 
   template<class _Tp, class _Up, size_t _Size>
   concept __swappable_arrays =

 
 // [concept.equiv]
 template<class _Rp, class _Tp, class _Up>
 concept equivalence_relation = relation<_Rp, _Tp, _Up>;
 
 // [concept.strictweakorder]
 template<class _Rp, class _Tp, class _Up>
 concept strict_weak_order = relation<_Rp, _Tp, _Up>;
-#endif //_LIBCPP_STD_VER > 17 && !defined(_LIBCPP_HAS_NO_CONCEPTS)
+
+// Obscurely named so the Clang diagnostic engine doesn't suggest it by accident.
+template<class _Tp>
+concept __core_reference = is_reference_v<_Tp>;
+
+template<class _Tp>
+concept __lvalue_reference = __core_reference<_Tp> && is_lvalue_reference_v<_Tp>;
+
+template<class _Tp>
+concept __rvalue_reference = __core_reference<_Tp> && !__lvalue_reference<_Tp>;
+
+#endif // _LIBCPP_STD_VER > 17 && !defined(_LIBCPP_HAS_NO_CONCEPTS)
 
 _LIBCPP_END_NAMESPACE_STD
 
 _LIBCPP_POP_MACROS
 
 #endif // _LIBCPP_CONCEPTS

libcxx/include/iterator

 {
     typedef ptrdiff_t difference_type;
     typedef T value_type;
     typedef T* pointer;
     typedef T& reference;
     typedef random_access_iterator_tag iterator_category;
 };
 
+namespace ranges {
+  inline namespace unspecified {
+    inline constexpr unspecified iter_move = unspecified; // since C++20
+  }
+}
+
+template<dereferenceable T>
+  requires requires(T& t) {
+    { ranges::iter_move(t) } -> can-reference;
+  }
+using iter_rvalue_reference_t
+  = decltype(ranges::iter_move(declval<T&>()));
+
 template<class Category, class T, class Distance = ptrdiff_t,
          class Pointer = T*, class Reference = T&>
 struct iterator
 {
     typedef T         value_type;
     typedef Distance  difference_type;
     typedef Pointer   pointer;
     typedef Reference reference;

 #include <concepts>
 #include <cstddef>
 #include <initializer_list>
 #include <__memory/base.h>
 #include <__memory/pointer_traits.h>
 #include <version>
 #include <concepts>
 
 #include <__debug>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #pragma GCC system_header
 #endif
 
 _LIBCPP_BEGIN_NAMESPACE_STD
 
 template <class _Iter>
if (__pred(*__iter))
       __iter = __c.erase(__iter);
     else
       ++__iter;
   }
 
   return __old_size - __c.size();
 }
 
+#if !defined(_LIBCPP_HAS_NO_RANGES)
+
+namespace ranges::__iter_move {
+void iter_move();
+
+template<class _Ip>
+concept __unqualified_iter_move =
+  __class_or_enum<_Ip> &&
+  requires(_Ip&& __i) {
+    iter_move(_VSTD::forward<_Ip>(__i));
+  };
+
+template<class _Ip>
+concept __lvalue_iter_move =
+  !__unqualified_iter_move<_Ip> &&
+  __dereferenceable<_Ip> &&
+  requires(_Ip&& __i) {
+    { *_VSTD::forward<_Ip>(__i) } -> __lvalue_reference;
+  };
+
+template<class _Ip>
+concept __rvalue_iter_move =
+  !__unqualified_iter_move<_Ip> &&
+  !__lvalue_iter_move<_Ip> &&
+  requires(_Ip&& __i) {
+    *_VSTD::forward<_Ip>(__i);
+  };
+
+#define _LIBCPP_NOEXCEPT_RETURN(...) \
+  noexcept(noexcept(__VA_ARGS__))    \
+  { return __VA_ARGS__; }
+
+// The name `ranges::iter_­move` denotes a customization point object ([customization.point.object]).
+// The expression `ranges::iter_­move(E)` for a subexpression `E` is expression-equivalent to:
+struct __fn {
+  // `iter_­move(E)`, if `E` has class or enumeration type and `iter_­move(E)` is a well-formed
+  // expression when treated as an unevaluated operand, with overload resolution performed in a
+  // context that does not include a declaration of `ranges::iter_­move` but does include the
+  // declaration
+  // ```
+  // void iter_move();
+  // ```
+  template<__unqualified_iter_move _Ip>
+  constexpr decltype(auto) operator()(_Ip&& __i) const
+  _LIBCPP_NOEXCEPT_RETURN(
+    iter_move(std::forward<_Ip>(__i))
+  )
+
+  // Otherwise, if the expression `*E` is well-formed:
+  // if `*E` is an lvalue, `std::move(*E)`;
+  template<__lvalue_iter_move _Ip>
+  constexpr auto&& operator()(_Ip&& __i) const
+  _LIBCPP_NOEXCEPT_RETURN(
+    std::move(*std::forward<_Ip>(__i))
+  )
+
+  // otherwise, `*E`.
+  template<__rvalue_iter_move _Ip>
+  constexpr decltype(auto) operator()(_Ip&& __i) const
+  _LIBCPP_NOEXCEPT_RETURN(
+    *std::forward<_Ip>(__i)
+  )
+
+  // Otherwise, `ranges!::!iter_­move(E)` is ill-formed.
+  void operator()(auto&&) const = delete;
+};
+} // namespace ranges::__iter_move
+
+namespace ranges::inline __cpo {
+  inline constexpr auto iter_move = __iter_move::__fn{};
+}
+
+template<__dereferenceable _Tp>
+requires requires(_Tp& __t) { { ranges::iter_move(__t) } -> __can_reference; }
+using iter_rvalue_reference_t = decltype(ranges::iter_move(declval<_Tp&>()));
+
+#undef _LIBCPP_NOEXCEPT_RETURN
+#endif // !defined(_LIBCPP_HAS_NO_RANGES)
+
 _LIBCPP_END_NAMESPACE_STD
 
 #endif  // _LIBCPP_ITERATOR

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_move.pass.cpp

+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+// UNSUPPORTED: c++03, c++11, c++14, c++17
+// UNSUPPORTED: libcpp-no-concepts
+// UNSUPPORTED: gcc-10
+
+// template<class I>
+// unspecified iter_move;
+
+#include <iterator>
+
+#include <array>
+#include <algorithm>
+#include <cassert>
+#include <utility>
+
+#include "../unqualified_lookup_wrapper.h"
+
+// Wrapper around an iterator for testing `iter_move` when an unqualified call to `iter_move` isn't
+// possible.
+template <typename I>
+class iterator_wrapper {
+public:
+  iterator_wrapper() = default;
+
+  constexpr explicit iterator_wrapper(I i) noexcept : base_(std::move(i)) {}
+
+  /// \note `noexcept(false)` is used to check that this operator is called.
+  ///
+  [[nodiscard]] constexpr decltype(auto) operator*() const& noexcept(false) {
+    return *base_;
+  }
+
+  /// \note `noexcept` is used to check that this operator is called.
+  ///
+  [[nodiscard]] constexpr auto&& operator*() && noexcept {
+    return std::move(*base_);
+  }
+
+  constexpr iterator_wrapper& operator++() noexcept {
+    ++base_;
+    return *this;
+  }
+
+  constexpr void operator++(int) noexcept { ++base_; }
+
+  [[nodiscard]] constexpr bool
+  operator==(iterator_wrapper const& other) const noexcept {
+    return base_ == other.base_;
+  }
+
+private:
+  I base_ = I{};
+};
+
+class move_tracker {
+public:
+  move_tracker() = default;
+
+  constexpr move_tracker(move_tracker&& other) noexcept
+      : moves_{other.moves_ + 1} {
+    other.moves_ = 0;
+  }
+
+  constexpr move_tracker& operator=(move_tracker&& other) noexcept {
+    moves_ = other.moves_ + 1;
+    other.moves_ = 0;
+    return *this;
+  }
+
+  constexpr move_tracker(move_tracker const& other) = delete;
+  constexpr move_tracker& operator=(move_tracker const& other) = delete;
+
+  [[nodiscard]] constexpr int moves() const noexcept { return moves_; }
+
+private:
+  int moves_ = 0;
+};
+
+template <typename I>
+constexpr void unqualified_lookup_move(I first_, I last_, I result_first_,
+                                       I result_last_) {
+  auto first = unqualified_lookup_wrapper{std::move(first_)};
+  auto last = unqualified_lookup_wrapper{std::move(last_)};
+  auto result_first = unqualified_lookup_wrapper{std::move(result_first_)};
+  auto result_last = unqualified_lookup_wrapper{std::move(result_last_)};
+
+  static_assert(!noexcept(std::ranges::iter_move(first)),
+                "unqualified-lookup case not being chosen");
+
+  for (; first != last && result_first != result_last;
+       (void)++first, ++result_first) {
+    *result_first = std::ranges::iter_move(first);
+  }
+}
+
+template <typename I>
+constexpr void lvalue_move(I first_, I last_, I result_first_, I result_last_) {
+  auto first = iterator_wrapper{std::move(first_)};
+  auto last = ::iterator_wrapper{std::move(last_)};
+  auto result_first = iterator_wrapper{std::move(result_first_)};
+  auto result_last = iterator_wrapper{std::move(result_last_)};
+
+  static_assert(!noexcept(std::ranges::iter_move(first)),
+                "`operator*() const&` is not noexcept, and there's no hidden "
+                "friend iter_move.");
+
+  for (; first != last && result_first != result_last;
+       (void)++first, ++result_first) {
+    *result_first = std::ranges::iter_move(first);
+  }
+}
+
+template <typename I>
+constexpr void rvalue_move(I first_, I last_, I result_first_, I result_last_) {
+  auto first = iterator_wrapper{std::move(first_)};
+  auto last = iterator_wrapper{std::move(last_)};
+  auto result_first = iterator_wrapper{std::move(result_first_)};
+  auto result_last = iterator_wrapper{std::move(result_last_)};
+
+  static_assert(
+      noexcept(std::ranges::iter_move(std::move(first))),
+      "`operator*() &&` is noexcept, and there's no hidden friend iter_move.");
+
+  for (; first != last && result_first != result_last;
+       (void)++first, ++result_first) {
+    auto i = first;
+    *result_first = std::ranges::iter_move(std::move(i));
+  }
+}
+
+constexpr bool check_iter_move() {
+  constexpr int full_size = 100;
+  constexpr int half_size = full_size / 2;
+  constexpr int reset = 0;
+  auto v1 = std::array<move_tracker, full_size>{};
+
+  auto move_counter_is = [](auto const n) {
+    return [n](auto const& x) { return x.moves() == n; };
+  };
+
+  auto v2 = std::array<move_tracker, half_size>{};
+  unqualified_lookup_move(v1.begin(), v1.end(), v2.begin(), v2.end());
+  assert(std::all_of(v1.cbegin(), v1.cend(), move_counter_is(reset)));
+  assert(std::all_of(v2.cbegin(), v2.cend(), move_counter_is(1)));
+
+  auto v3 = std::array<move_tracker, half_size>{};
+  unqualified_lookup_move(v1.begin() + half_size, v1.end(), v3.begin(),
+                          v3.end());
+  assert(std::all_of(v1.cbegin(), v1.cend(), move_counter_is(reset)));
+  assert(std::all_of(v3.cbegin(), v3.cend(), move_counter_is(1)));
+
+  auto v4 = std::array<move_tracker, half_size>{};
+  unqualified_lookup_move(v3.begin(), v3.end(), v4.begin(), v4.end());
+  assert(std::all_of(v3.cbegin(), v3.cend(), move_counter_is(reset)));
+  assert(std::all_of(v4.cbegin(), v4.cend(), move_counter_is(2)));
+
+  lvalue_move(v2.begin(), v2.end(), v1.begin() + half_size, v1.end());
+  assert(std::all_of(v2.cbegin(), v2.cend(), move_counter_is(reset)));
+  assert(std::all_of(v1.cbegin() + half_size, v1.cend(), move_counter_is(2)));
+
+  lvalue_move(v4.begin(), v4.end(), v1.begin(), v1.end());
+  assert(std::all_of(v4.cbegin(), v4.cend(), move_counter_is(reset)));
+  assert(std::all_of(v1.cbegin(), v1.cbegin() + half_size, move_counter_is(3)));
+
+  rvalue_move(v1.begin(), v1.end(), v2.begin(), v2.end());
+  assert(std::all_of(v1.cbegin(), v1.cbegin() + half_size,
+                     move_counter_is(reset)));
+  assert(std::all_of(v2.cbegin(), v2.cend(), move_counter_is(4)));
+
+  rvalue_move(v1.begin() + half_size, v1.end(), v3.begin(), v3.end());
+  assert(std::all_of(v1.cbegin(), v1.cend(), move_counter_is(reset)));
+  assert(std::all_of(v3.cbegin(), v3.cend(), move_counter_is(3)));
+
+  return true;
+}
+
+int main(int, char**) {
+  static_assert(check_iter_move());
+  assert(check_iter_move());
+
+  return 0;
+}

libcxx/test/std/iterators/iterator.requirements/iterator.cust/iterator.cust.move/iter_rvalue_reference_t.pass.cpp

+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+// UNSUPPORTED: c++03, c++11, c++14, c++17
+// UNSUPPORTED: libcpp-no-concepts
+// UNSUPPORTED: gcc-10
+
+// template<class I>
+// using iter_rvalue_reference;
+
+#include <iterator>
+
+#include <concepts>
+#include <list>
+#include <vector>
+
+static_assert(
+    std::same_as<std::iter_rvalue_reference_t<std::vector<int>::iterator&>,
+                 int&&>);
+static_assert(
+    std::same_as<std::iter_rvalue_reference_t<std::vector<int>::const_iterator>,
+                 int const&&>);
+static_assert(
+    std::same_as<std::iter_rvalue_reference_t<std::list<int const>::iterator>,
+                 int const&&>);
+
+int main(int, char**) { return 0; }

libcxx/test/std/iterators/iterator.requirements/iterator.cust/unqualified_lookup_wrapper.h

+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+#ifndef LIBCPP_TEST_STD_ITERATOR_UNQUALIFIED_LOOKUP_WRAPPER
+#define LIBCPP_TEST_STD_ITERATOR_UNQUALIFIED_LOOKUP_WRAPPER
+
+#include <iterator>
+#include <utility>
+
+// Wrapper around an iterator for testing unqualified calls to `iter_move` and `iter_swap`.
+template <typename I>
+class unqualified_lookup_wrapper {
+public:
+  unqualified_lookup_wrapper() = default;
+
+  constexpr explicit unqualified_lookup_wrapper(I i) noexcept
+      : base_(std::move(i)) {}
+
+  [[nodiscard]] constexpr decltype(auto) operator*() const noexcept {
+    return *base_;
+  }
+
+  constexpr unqualified_lookup_wrapper& operator++() noexcept {
+    ++base_;
+    return *this;
+  }
+
+  constexpr void operator++(int) noexcept { ++base_; }
+
+  [[nodiscard]] constexpr bool
+  operator==(unqualified_lookup_wrapper const& other) const noexcept {
+    return base_ == other.base_;
+  }
+
+  // Delegates `std::ranges::iter_move` for the underlying iterator. `noexcept(false)` will be used
+  // to ensure that the unqualified-lookup overload is chosen.
+  [[nodiscard]] friend constexpr decltype(auto)
+      iter_move(unqualified_lookup_wrapper& i) noexcept(false) {
+    return std::ranges::iter_move(i.base_);
+  }
+
+private:
+  I base_ = I{};
+};
+
+#endif // LIBCPP_TEST_STD_ITERATOR_UNQUALIFIED_LOOKUP_WRAPPER