[libc++] Comment preconditions for __wrap_iter; make it work with C++20 to_address

[• Quuxplusone]

Summary

LWG3545 "std::pointer_traits should be SFINAE-friendly" is maybe related.
It's currently unclear what the difference is between "contiguous iterator"
(for which `std::to_address` is required to work) and "fancy pointer"
or "pointer-like type" (for which both `pointer_traits` and
`std::to_address` are required to work).

The precondition for __wrap_iter is that it's for use only with fancy pointers,
i.e., things that should in C++20 be unambiguously contiguous iterators...
but because of legacy code, we can't actually _enforce_ this precondition.
We sanity-check it only as far as verifying that the fancy-pointer type
we're given is at least a random-access iterator.
We explicitly do _not_ try to instantiate `pointer_traits<_Iter>`, because
legacy user-defined fancy pointer types might not support that. (The metaprogramming
around `pointer_traits` is vastly different and more stringent in C++20 from how it
was in C++11/14/17.) Instead, we compute `element_type` directly from `reference`.

Meanwhile, `__is_cpp17_contiguous_iterator` is used (only) for optimizations
in <algorithm>, so in C++17 we must be very careful not to apply those optimizations
to arbitrary user-defined types (in particular we mustn't assume that any old
random-access iterator with an `element_type` is contiguous).
In C++20 we can liberally assume that the optimizations should be applied to
any user-defined iterator types that advertise themselves as contiguous via
`iterator_category` and/or `iterator_concept` (which automatically includes
`__wrap_iter`).

Our `__wrap_iter` now correctly advertises itself as a contiguous iterator in C++20,
and produces the correct `pointer_traits<WI>::element_type`; this is now tested.

Refs D101396

[cjdb]

This comment was removed by cjdb.

[cjdb]

Realised the test will be done by D101396. 🚢 it!

[zoecarver]

Should there be a test for string too?

Eh, I think it's fine. If you feel strongly I'll add it. This will implicitly be covered for "all" iterators by the tests in D101396.

[• Quuxplusone]

If this PR is the last thing needed to unblock landing something for C++20, then OK ship it, I don't think it's introducing any wrong behavior. However, if we have time, I would like to see further investigation of how this patch fits into the general scheme of D94807, and how much of that scheme is still relevant this month. In particular:

• Can we make __is_cpp17_contiguous_iterator more similar to std::contiguous_iterator, perhaps by having __is_cpp17_contiguous_iterator look at whether element_type exists?

• Can we now remove the specialization of __is_cpp17_contiguous_iterator for __wrap_iter?

• Can we now remove the overload of _VSTD::__to_address for __wrap_iter? I think no, but I'm not sure (and it sure would be nice if we could).

• You're providing element_type unconditionally, which means that __wrap_iter<It> will be treated as pointer-like even when It is not a (fancy) pointer type itself; e.g. __wrap_iter<std::list<int>::iterator>. My impression is that we actually only ever use __wrap_iter to wrap (fancy) pointers — is this accurate? If it is not accurate, then more digging is warranted. If it is accurate, then we should document that invariant, and maybe even search-and-replace _Iter to _Ptr throughout __wrap_iter.

D101003 is related. (I'm eliminating span::iterator's dependence on __wrap_iter<T*>.)

libcxx/include/iterator

1289

If my comments are correct, then this could+should use pointer_traits<_Iter> instead of pointer_traits<pointer>...
...and the _If on line 1284 is unconditionally true, and can be simplified.
Please try that simplification (at least locally) and see if it passes the test suite.

[zoecarver]

@Quuxplusone I didn't look at that patch, but now I have lots of thoughts 😁. First, after C++17 we should probably use std::contiguous_iterator to implement __is_cpp17_contiguous_iterator. Second, pre-C++20 AFAICT __is_cpp17_contiguous_iterator is the same as is_pointer_v + a specialization for __iter_wrap, right? I think we need to make this at least check for pointer_traits so that we can handle fancy pointers as well.

Can we now remove the specialization of is_cpp17_contiguous_iterator for wrap_iter?

Can we just remove __is_cpp17_contiguous_iterator all together? It's only used in two places in the whole codebase.

Can we now remove the overload of _VSTD::to_address for wrap_iter? I think no, but I'm not sure (and it sure would be nice if we could).

Not only can we, but we should. This is causing problems because it makes something work that really shouldn't. There's no reason __wrap_iter can't conform to contiguous_iterator via the standard method (by providing a template or element_type).

Side note: the reason we can't use the template to deduce the element type is that the template is a pointer (hence the "wrap" part of the name).

Anyway, simply providing element_type as this patch does is sufficient for to_address to start working.

You're providing element_type unconditionally, which means that wrap_iter<It> will be treated as pointer-like even when It is not a (fancy) pointer type itself; e.g. wrap_iter<std::list<int>::iterator>. My impression is that we actually only ever use wrap_iter to wrap (fancy) pointers — is this accurate? If it is not accurate, then more digging is warranted. If it is accurate, then we should document that invariant, and maybe even search-and-replace _Iter to _Ptr throughout wrap_iter.

I think we could make this guarantee and I don't see any reason not to, especially if it makes our implementation simpler/better.

Note: the only place static_assert(sizeof(typename pointer_traits<_Iter>::element_type), "") fails in the whole codebase is in the __is_cpp17_contiguous_iterator test: test/libcxx/iterators/contiguous_iterators.pass.cpp.

libcxx/include/iterator

1289

Alternatively we could add element_type with an _If (or maybe inheritance?) to make this work for non-pointer-like iterators.

[• Quuxplusone]

Can we just remove __is_cpp17_contiguous_iterator altogether? It's only used in two places

Maybe, but I doubt it. The places it's used are specifically to implement the lowering of std::copy into memmove: https://godbolt.org/z/5P9TjGza4
For performance, we really want to be able to detect "this specific iterator is known contiguous" in C++17/14/11/03 mode.
Now, can we change __is_cpp17_contiguous_iterator to rely on std::contiguous_iterator in C++20 and later? I would say yes. But we can't just make it false in C++17-and-earlier; it's still got to do the right thing for known-contiguous iterators in all modes. Unless we decide we don't care about that optimization. There's definitely (and IMHO unfortunately) precedent for libc++ saying "optimization isn't our department, let's just ask Clang to optimize better," but at least I wouldn't want to unilaterally decide that we don't care about this specific optimization.

Note: the only place [...] fails in the whole codebase is in [...] test/libcxx/iterators/contiguous_iterators.pass.cpp.

This line, right?:

static_assert((!std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::reverse_iterator<T *> > >::value), "");

I'd be 100% cool with eliminating that line from the test, because I think it violates the unstated precondition on __wrap_iter's _Iter argument. (And then we can document that precondition and enforce it.)

[zoecarver]

Maybe, but I doubt it. The places it's used are specifically to implement the lowering of std::copy into memmove: https://godbolt.org/z/5P9TjGza4
For performance, we really want to be able to detect "this specific iterator is known contiguous" in C++17/14/11/03 mode.
Now, can we change __is_cpp17_contiguous_iterator to rely on std::contiguous_iterator in C++20 and later? I would say yes. But we can't just make it false in C++17-and-earlier; it's still got to do the right thing for known-contiguous iterators in all modes. Unless we decide we don't care about that optimization. There's definitely (and IMHO unfortunately) precedent for libc++ saying "optimization isn't our department, let's just ask Clang to optimize better," but at least I wouldn't want to unilaterally decide that we don't care about this specific optimization.

The thing I'm more annoyed about is that it is used in one place and spread across four files. Here's what I suggest: we "move" __is_cpp17_contiguous_iterator to algorithm and rename it to something like __is_pointer_like_iterator or __is_any_contiguous_iterator and define it as so (pseudo code):

#if C++20
template<class T>
constexpr inline bool __is_pointer_like_iterator = std::contiguous_iterator<T>;
#else
template<class T>
constexpr inline bool __is_pointer_like_iterator = requires { typename pointer_traits<T>::element_type };
#endif

WDYT?

This line, right?:

Yes, and looking more closely the only lines it fails on are similar assertions that check what types aren't contiguous iterators.

I'd be 100% cool with eliminating that line from the test, because I think it violates the unstated precondition on __wrap_iter's _Iter argument. (And then we can document that precondition and enforce it.)

:) Great, I agree.

[tcanens]

libcxx/include/iterator

1289

It's surprising to me that this is even needed. What breaks without this change?

[• Quuxplusone]

Commandeer. Add tests for to_address-of-STL-iterator-types and pointer_traits-of-STL-iterator-types.

[• Quuxplusone]

libcxx/include/iterator

1289

In the old code, pointer_traits<std::vector<int>::iterator>::element_type would be int*, not int. https://godbolt.org/z/PTPvn11bf
Providing a user-provided element_type fixes that (in both C++17 and C++20).

Interestingly, libstdc++ and MSVC STL also report pointer_traits<std::vector<int>::iterator>::element_type as something other than int. I don't know if this is a conscious decision ("contiguous iterators needn't support pointer_traits") or if they just haven't gotten around to that part of C++20 yet. @CaseyCarter ?

[CaseyCarter]

libcxx/include/iterator

1289

https://godbolt.org/z/ba6z9cEbe demonstrates fairly clearly that pointer_traits<vector<T>::iterator>::element_type is T for MSVC STL. It looks like GCC didn't need to specialize pointer_traits for the to_address requirement: presumably operator-> always works with their pointer-wrapper type. This is not the case for our _Vector_iterator which asserts in debug mode if you call operator-> on a non-dereferenceable iterator.

[zoecarver]

libcxx/include/__iterator/iterator_traits.h

431

Heh, thanks for fixing this :P

467

Can we make this use the concept? I guess we could make that a follow up nfc.

libcxx/include/iterator

469

Please keep this as per D101098:

All sub-headers should be included by their parent header. To use the example above, `<ranges> must include the sub-header __ranges/size.h, even if it doesn't use ranges::size itself. This inclusion is mandatory because the standard requires the top level header to provide these declarations. This rule falls out of the more general "Include What You Use." You should never rely on some other header to transitively include something that you need. If your header needs a definition of iterator_traits, then you should either #include <iterator> (exactly as user code does), or, to improve compile time, #include <__iterator/iterator_traits.h>. You should never expect iterator_traits` to be "pulled in" by any other header.

(Excuse the formatting).

1281

Who gets broken by making this __is_cpp17_contiguous_iterator? This is only used to wrap pointers and internal iterators. We should just make it work.

1292

If this is now unconditionally contiguous_iterator_tag we absolutely need to change the above assertion.

1296

Why can't we use pointer_traits with this? That would seem more correct to me. Especially if we're advertising this as a "known trivial wrapper around a pointer type."

The whole reason we want to have an element type is so that pointer_traits is correct for __warp_iter. If our wrapped type doesn't support pointer_traits then we really shouldn't either, because we'll be falsely "supporting" something that we don't actually "support." However, as I have said elsewhere, I don't think we actually wrap any types that don't support pointer_traits, so I think we are safe to use that (maybe with a static assert).

If nothing else, let's make this remove_pointer<iterator_traits::pointer>.

1478

Why is this only before C++20? Doesn't that mean we loose the optimization in C++20?

1483

Do we still need this? I think not.

[zoecarver]

Also the added tests are great, thank you!

[tcanens]

libcxx/include/iterator

1289

Except that nothing in the standard requires pointer_traits<std::vector<int>::iterator>::element_type to be meaningful, only that it doesn't explode. So what was broken?

If you want to claim that it is better QOI, that's fine (but since lots of - even most of - contiguous iterators are not even Cpp17NullablePointers I find using pointer_traits with them rather debatable), but that has nothing to do with to_address really.

[zoecarver]

libcxx/include/iterator

1289

That's a good point. Our implementation of to_address shouldn't care what element_type is so long as it exists. For __wrap_iter it should be using operator-> because pointer_traits<__wrap_iter>::to_address doesn't exist either way. Maybe this patch isn't the right fix (though I think the cleanups should be applied regardless).

We should investigate why we can't use to_address with the following type:

struct T {
  using element_type = int*;
  int *operator->();
};

https://godbolt.org/z/ed7Pqzd8n

@Quuxplusone thoughts?

[• Quuxplusone]

libcxx/include/__iterator/iterator_traits.h

467

Absolutely not, because that would break all C++17 fancy pointers when compiled in C++20 mode. We can't assume that everyone running the compiler in C++20 mode is actually able to use contiguous_iterator_tag.

libcxx/include/iterator

469

Yes, I know. Check alphabetization and duplication. (This trivial diff will disappear soon because I'll have landed it.)

1281

All C++17 fancy pointers get broken by that. C++17 fancy pointers do not advertise themselves as contiguous iterators, because they cannot; there is no way to advertise yourself as a contiguous iterator in C++17-or-earlier.

1289

@CaseyCarter: I was looking at https://godbolt.org/z/f4zbjMMoz and seeing that pointer_traits<std::vector<int>::iterator>::element_type doesn't work at all.
Similarly with the static_assert you posted, https://godbolt.org/z/jGGEEaE5a , if compiled in any mode other than C++20 mode.
Ditto for GCC's libstdc++: https://godbolt.org/z/75sqxnKKa
However, I'm still inclining toward making everything Just Work for libc++. Otherwise, it's too confusing for my brain to keep track of what "works," what "doesn't work on purpose," what "doesn't work not on purpose," etc.

1296

As the comment says, "Prior to C++20, there was no "standard" way to mark a type as a fancy pointer; after C++20, we would prefer all fancy pointers to have contiguous_iterator_tag, but we can't assume that they will, because of legacy code."
In particular, it is not safe to use pointer_traits<FancyPointer> in C++20 mode, because not all C++17 FancyPointers are safe to use with pointer_traits in C++20 mode. This is tested (sorta) via the test for __wrap_iter<my_random_access_iterator> in libcxx/test/libcxx/iterators/contiguous_iterators.pass.cpp.

1478

In C++20 and later, __is_cpp17_contiguous_iterator<__wrap_iter<_It>> is already true, because __wrap_iter<_It> advertises itself as a contiguous iterator in C++20. The specialization is needed only in C++17-and-earlier, because in those modes there is no established way for __wrap_iter to advertise itself as a contiguous iterator.

1483

We do.

[• Quuxplusone]

libcxx/include/iterator

1350–1357

https://reviews.llvm.org/harbormaster/unit/view/565840/
It turns out that __wrap_iter::operator->() assert-fails in situations where we need std::to_address not to assert-fail. So, since I want to_address to work — and since we can't implement to_address in terms of operator-> — the only remaining possibility is to implement to_address in terms of a manually specialized pointer_traits<__wrap_iter>::to_address.

(The specification of to_address should have done it as an ordinary ADL customization point, but they messed that up royally and now it's too late to fix it.)

So, since we have to specialize pointer_traits<__wrap_iter> anyway, we probably don't need any of the element_type stuff anymore. I'll look into this again soonish.

[• Quuxplusone]

@zoecarver @ldionne: I've lost track of what's going on with contiguous/to_address anymore, but I'm pretty sure this PR was obsoleted by D101638. (It looks like we still need to make __wrap_iter properly C++20-contiguous in debug mode, so that we can enable "libcxx/test/std/utilities/memory/pointer.conversion/to_address_std_iterators.pass.cpp", but that's maybe all that's left.)

libcxx/include/__iterator/iterator_traits.h

struct _LIBCPP_TEMPLATE_VIS iterator_traits<_Tp*>
     typedef random_access_iterator_tag iterator_category;
 #if _LIBCPP_STD_VER > 17
     typedef contiguous_iterator_tag    iterator_concept;
 #endif
 };
 
 template <class _Tp, class _Up, bool = __has_iterator_category<iterator_traits<_Tp> >::value>
 struct __has_iterator_category_convertible_to
-    : _BoolConstant<is_convertible<typename iterator_traits<_Tp>::iterator_category, _Up>::value>
+    : is_convertible<typename iterator_traits<_Tp>::iterator_category, _Up>
 {};
 
 template <class _Tp, class _Up>
 struct __has_iterator_category_convertible_to<_Tp, _Up, false> : false_type {};
 
-template <class _Tp, class _Up, bool = __has_iterator_concept<_Tp>::value>
-struct __has_iterator_concept_convertible_to
-    : _BoolConstant<is_convertible<typename _Tp::iterator_concept, _Up>::value>
-{};
+template <class _Tp, class _Up, class = void>
+struct __has_iterator_concept_convertible_to : false_type {};
 
 template <class _Tp, class _Up>
-struct __has_iterator_concept_convertible_to<_Tp, _Up, false> : false_type {};
+struct __has_iterator_concept_convertible_to<_Tp, _Up, typename __void_t<_ITER_CONCEPT<_Tp> >::type>
+    : is_convertible<_ITER_CONCEPT<_Tp>, _Up> {};
 
 template <class _Tp>
 struct __is_cpp17_input_iterator : public __has_iterator_category_convertible_to<_Tp, input_iterator_tag> {};
 
 template <class _Tp>
 struct __is_cpp17_forward_iterator : public __has_iterator_category_convertible_to<_Tp, forward_iterator_tag> {};
 
 template <class _Tp>
 struct __is_cpp17_bidirectional_iterator : public __has_iterator_category_convertible_to<_Tp, bidirectional_iterator_tag> {};
 
 template <class _Tp>
 struct __is_cpp17_random_access_iterator : public __has_iterator_category_convertible_to<_Tp, random_access_iterator_tag> {};
 
-// __is_cpp17_contiguous_iterator determines if an iterator is contiguous,
-// either because it advertises itself as such (in C++20) or because it
-// is a pointer type or a known trivial wrapper around a pointer type,
-// such as __wrap_iter<T*>.
+// __is_cpp17_contiguous_iterator determines if an iterator is known by
+// libc++ to be contiguous, either because it advertises itself as such
+// (in C++20) or because it is a pointer type or a known trivial wrapper
+// around a pointer type, such as __wrap_iter<T*>. Such iterators receive
+// special "contiguous" optimizations in std::copy and std::sort.
 //
 #if _LIBCPP_STD_VER > 17
 template <class _Tp>
 struct __is_cpp17_contiguous_iterator : _Or<
     __has_iterator_category_convertible_to<_Tp, contiguous_iterator_tag>,
     __has_iterator_concept_convertible_to<_Tp, contiguous_iterator_tag>
 > {};
 #else
-template <class _Tp>
-struct __is_cpp17_contiguous_iterator : false_type {};
+template <class _Tp> struct __is_cpp17_contiguous_iterator : false_type {};
+template <class _Up> struct __is_cpp17_contiguous_iterator<_Up*> : true_type {};
 #endif
 
-// Any native pointer which is an iterator is also a contiguous iterator.
-template <class _Up>
-struct __is_cpp17_contiguous_iterator<_Up*> : true_type {};
-
-
 template <class _Tp>
 struct __is_exactly_cpp17_input_iterator
     : public integral_constant<bool,
          __has_iterator_category_convertible_to<_Tp, input_iterator_tag>::value &&
         !__has_iterator_category_convertible_to<_Tp, forward_iterator_tag>::value> {};
 
 #ifndef _LIBCPP_HAS_NO_DEDUCTION_GUIDES
 template<class _InputIterator>

libcxx/include/iterator

 #include <concepts> // Mandated by the Standard.
 #include <cstddef>
 #include <initializer_list>
 #include <__iterator/concepts.h>
 #include <__iterator/incrementable_traits.h>
 #include <__iterator/iter_move.h>
 #include <__iterator/iterator_traits.h>
 #include <__iterator/readable_traits.h>
-#include <__iterator/concepts.h>
 #include <__memory/addressof.h>
 #include <__memory/pointer_traits.h>
 #include <version>
 
 #include <__debug>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #pragma GCC system_header

 template <class _Ip, class _Op> _Op _LIBCPP_INLINE_VISIBILITY _LIBCPP_CONSTEXPR_AFTER_CXX17 copy(_Ip, _Ip, _Op);
 template <class _B1, class _B2> _B2 _LIBCPP_INLINE_VISIBILITY _LIBCPP_CONSTEXPR_AFTER_CXX17 copy_backward(_B1, _B1, _B2);
 template <class _Ip, class _Op> _Op _LIBCPP_INLINE_VISIBILITY _LIBCPP_CONSTEXPR_AFTER_CXX17 move(_Ip, _Ip, _Op);
 template <class _B1, class _B2> _B2 _LIBCPP_INLINE_VISIBILITY _LIBCPP_CONSTEXPR_AFTER_CXX17 move_backward(_B1, _B1, _B2);
 
 template <class _Iter>
 class __wrap_iter
 {
+    // Prior to C++20, there was no "standard" way to mark a type as a fancy pointer;
+    // after C++20, we would prefer all fancy pointers to have contiguous_iterator_tag,
+    // but we can't assume that they will, because of legacy code.
+    // Since C++03, every fancy pointer should at least advertise itself as a
+    // random-access iterator. But we can't assume fancy pointers satisfy any particular
+    // C++20 concept (e.g. std::contiguous_iterator), again because legacy code exists.
+    //
+    static_assert(__is_cpp17_random_access_iterator<_Iter>::value,
+        "__wrap_iter cannot wrap anything but a fancy pointer");
+
 public:
     typedef _Iter                                                      iterator_type;
     typedef typename iterator_traits<iterator_type>::value_type        value_type;
     typedef typename iterator_traits<iterator_type>::difference_type   difference_type;
     typedef typename iterator_traits<iterator_type>::pointer           pointer;
     typedef typename iterator_traits<iterator_type>::reference         reference;
     typedef typename iterator_traits<iterator_type>::iterator_category iterator_category;
 #if _LIBCPP_STD_VER > 17
-    typedef _If<__is_cpp17_contiguous_iterator<_Iter>::value,
-                contiguous_iterator_tag, iterator_category>            iterator_concept;
+    typedef contiguous_iterator_tag                                    iterator_concept;
 #endif
+    typedef typename remove_reference<
+            typename iterator_traits<iterator_type>::reference
+        >::type                                                        element_type;
 
 private:
     iterator_type __i;
 public:
     _LIBCPP_INLINE_VISIBILITY _LIBCPP_CONSTEXPR_IF_NODEBUG __wrap_iter() _NOEXCEPT
 #if _LIBCPP_STD_VER > 11
                 : __i{}
 #endif
#endif
     _LIBCPP_INLINE_VISIBILITY _LIBCPP_CONSTEXPR_IF_NODEBUG reference operator*() const _NOEXCEPT
     {
 #if _LIBCPP_DEBUG_LEVEL == 2
         _LIBCPP_ASSERT(__get_const_db()->__dereferenceable(this),
                        "Attempted to dereference a non-dereferenceable iterator");
 #endif
         return *__i;
     }
     _LIBCPP_INLINE_VISIBILITY _LIBCPP_CONSTEXPR_IF_NODEBUG pointer  operator->() const _NOEXCEPT
     {
 #if _LIBCPP_DEBUG_LEVEL == 2
         _LIBCPP_ASSERT(__get_const_db()->__dereferenceable(this),
                        "Attempted to dereference a non-dereferenceable iterator");
 #endif
         return (pointer)_VSTD::addressof(*__i);
     }
     _LIBCPP_INLINE_VISIBILITY _LIBCPP_CONSTEXPR_IF_NODEBUG __wrap_iter& operator++() _NOEXCEPT
     {
 #if _LIBCPP_DEBUG_LEVEL == 2
         _LIBCPP_ASSERT(__get_const_db()->__dereferenceable(this),
                        "Attempted to increment non-incrementable iterator");
 #endif
         ++__i;
         return *this;
#endif
     template <class _Ip, class _Op> friend _LIBCPP_CONSTEXPR_AFTER_CXX17 _Op copy(_Ip, _Ip, _Op);
     template <class _B1, class _B2> friend _LIBCPP_CONSTEXPR_AFTER_CXX17 _B2 copy_backward(_B1, _B1, _B2);
     template <class _Ip, class _Op> friend _LIBCPP_CONSTEXPR_AFTER_CXX17 _Op move(_Ip, _Ip, _Op);
     template <class _B1, class _B2> friend _LIBCPP_CONSTEXPR_AFTER_CXX17 _B2 move_backward(_B1, _B1, _B2);
 };
 
 #if _LIBCPP_STD_VER <= 17
 template <class _It>
-struct __is_cpp17_contiguous_iterator<__wrap_iter<_It> > : __is_cpp17_contiguous_iterator<_It> {};
-#endif
+struct __is_cpp17_contiguous_iterator<__wrap_iter<_It> > : true_type {};
 
 template <class _Iter>
 _LIBCPP_CONSTEXPR
-_EnableIf<__is_cpp17_contiguous_iterator<_Iter>::value, decltype(_VSTD::__to_address(declval<_Iter>()))>
+decltype(_VSTD::__to_address(declval<_Iter>()))
 __to_address(__wrap_iter<_Iter> __w) _NOEXCEPT {
     return _VSTD::__to_address(__w.base());
 }
+#endif
 
 template <class _Iter1, class _Iter2>
 inline _LIBCPP_INLINE_VISIBILITY _LIBCPP_CONSTEXPR_IF_NODEBUG
 bool
 operator==(const __wrap_iter<_Iter1>& __x, const __wrap_iter<_Iter2>& __y) _NOEXCEPT
 {
     return __x.base() == __y.base();
 }

libcxx/test/libcxx/iterators/contiguous_iterators.pass.cpp

public:
     friend bool operator!=(const Self&, const Self&);
     friend bool operator<(const Self&, const Self&);
     friend bool operator>(const Self&, const Self&);
     friend bool operator<=(const Self&, const Self&);
     friend bool operator>=(const Self&, const Self&);
 };
 #endif
 
+struct fake_deque_iterator : std::deque<int>::iterator {
+    using element_type = int;
+};
+static_assert(std::__is_cpp17_random_access_iterator<fake_deque_iterator>::value, "");
+static_assert(!std::__is_cpp17_contiguous_iterator<fake_deque_iterator>::value, "");
+
+#if TEST_STD_VER >= 20
+struct fake2_deque_iterator : std::deque<int>::iterator {
+    using iterator_concept = std::contiguous_iterator_tag;
+    using element_type = int;
+};
+static_assert(std::__is_cpp17_random_access_iterator<fake2_deque_iterator>::value, "");
+static_assert(std::__is_cpp17_contiguous_iterator<fake2_deque_iterator>::value, "");
+#endif
+
 int main(int, char**)
 {
 //  basic tests
     static_assert(( std::__is_cpp17_contiguous_iterator<char *>::value), "");
     static_assert(( std::__is_cpp17_contiguous_iterator<const char *>::value), "");
     static_assert(( std::__is_cpp17_contiguous_iterator<int *>::value), "");
     static_assert(( std::__is_cpp17_contiguous_iterator<int **>::value), "");
     static_assert(( std::__is_cpp17_contiguous_iterator<T *>::value), "");

 #endif
 
     static_assert(( std::__is_cpp17_contiguous_iterator<std::__wrap_iter<char *> >::value), "");
     static_assert(( std::__is_cpp17_contiguous_iterator<std::__wrap_iter<const char *> >::value), "");
     static_assert(( std::__is_cpp17_contiguous_iterator<std::__wrap_iter<int *> >::value), "");
 
     static_assert(( std::__is_cpp17_contiguous_iterator<std::__wrap_iter<T *> >::value), "");
     static_assert(( std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::__wrap_iter<T *> > >::value), "");
-    static_assert((!std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::reverse_iterator<T *> > >::value), "");
 
-    static_assert((!std::__is_cpp17_contiguous_iterator<std::__wrap_iter<my_random_access_iterator> >::value), "");
-    static_assert((!std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::__wrap_iter<my_random_access_iterator> > >::value), "");
-    static_assert((!std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::reverse_iterator<my_random_access_iterator> > >::value), "");
+    static_assert(( std::__is_cpp17_contiguous_iterator<std::__wrap_iter<my_random_access_iterator> >::value), "");
+    static_assert(( std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::__wrap_iter<my_random_access_iterator> > >::value), "");
 
 #if TEST_STD_VER >= 20
     static_assert(( std::__is_cpp17_contiguous_iterator<std::__wrap_iter<my_contiguous_iterator> >::value), "");
     static_assert(( std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::__wrap_iter<my_contiguous_iterator> > >::value), "");
-    static_assert((!std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::reverse_iterator<my_contiguous_iterator> > >::value), "");
 #endif
 
 //  iterators in the libc++ test suite
     static_assert((!std::__is_cpp17_contiguous_iterator<output_iterator       <char *> >::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<input_iterator        <char *> >::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<forward_iterator      <char *> >::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<bidirectional_iterator<char *> >::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<random_access_iterator<char *> >::value), "");
//  string
     static_assert((!std::__is_cpp17_contiguous_iterator<std::wstring::reverse_iterator>      ::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<std::wstring::const_reverse_iterator>::value), "");
 
 //  deque is random-access but not contiguous
     static_assert((!std::__is_cpp17_contiguous_iterator<std::deque<int>::iterator>                   ::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<std::deque<int>::const_iterator>             ::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<std::deque<int>::reverse_iterator>           ::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<std::deque<int>::const_reverse_iterator>     ::value), "");
-    static_assert((!std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::deque<int>::iterator> >::value), "");
 
 //  vector<bool> is random-access but not contiguous
     static_assert((!std::__is_cpp17_contiguous_iterator<std::vector<bool>::iterator>                   ::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<std::vector<bool>::const_iterator>             ::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<std::vector<bool>::reverse_iterator>           ::value), "");
     static_assert((!std::__is_cpp17_contiguous_iterator<std::vector<bool>::const_reverse_iterator>     ::value), "");
-    static_assert((!std::__is_cpp17_contiguous_iterator<std::__wrap_iter<std::vector<bool>::iterator> >::value), "");
 
 #if TEST_STD_VER >= 11
     static_assert(( std::__is_cpp17_contiguous_iterator<std::initializer_list<int>::iterator>      ::value), "");
     static_assert(( std::__is_cpp17_contiguous_iterator<std::initializer_list<int>::const_iterator>::value), "");
 #endif
 
 #if TEST_STD_VER >= 17
     static_assert(( std::__is_cpp17_contiguous_iterator<std::string_view::iterator>      ::value), "");

libcxx/test/std/utilities/memory/pointer.conversion/to_address_std_iterators.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
+//
+//===----------------------------------------------------------------------===//
+
+// <memory>
+
+// UNSUPPORTED: c++03, c++11, c++14, c++17
+
+// template <class T> constexpr T* to_address(T* p) noexcept;
+// template <class Ptr> constexpr auto to_address(const Ptr& p) noexcept;
+
+#include <memory>
+
+#include <array>
+#include <cassert>
+#include <span>
+#include <string>
+#include <string_view>
+#include <valarray>
+#include <vector>
+#include "test_macros.h"
+
+template<class C>
+void test_container_iterators(C c)
+{
+    const C& cc = c;
+    assert(std::to_address(c.begin()) == c.data());
+    assert(std::to_address(c.end()) == c.data() + c.size());
+    assert(std::to_address(cc.begin()) == cc.data());
+    assert(std::to_address(cc.end()) == cc.data() + cc.size());
+}
+
+void test_valarray_iterators()
+{
+    std::valarray<int> v(100);
+    int *p = std::to_address(std::begin(v));
+    int *q = std::to_address(std::end(v));
+    assert(q - p == 100);
+}
+
+int main(int, char**) {
+    test_container_iterators(std::array<int, 3>());
+    test_container_iterators(std::vector<int>(3));
+    test_container_iterators(std::string("abc"));
+    test_container_iterators(std::string_view("abc"));
+    test_container_iterators(std::span<const char>("abc"));
+    test_valarray_iterators();
+
+    return 0;
+}

libcxx/test/std/utilities/memory/pointer.traits/std_iterators.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
+//
+//===----------------------------------------------------------------------===//
+
+// <memory>
+
+// template <class Ptr>
+// struct pointer_traits
+// {
+//     using element_type = see below;
+//     ...
+// };
+
+#include <memory>
+#include <array>
+#include <span>
+#include <string>
+#include <string_view>
+#include <valarray>
+#include <vector>
+
+#include "test_macros.h"
+
+int main(int, char**)
+{
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::array<int, 3>::iterator>::element_type, int);
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::array<int, 3>::const_iterator>::element_type, const int);
+
+#if TEST_STD_VER >= 20
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::span<int>::iterator>::element_type, int);
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::span<const int>::iterator>::element_type, const int);
+#endif
+
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::string::iterator>::element_type, char);
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::string::const_iterator>::element_type, const char);
+
+#if TEST_STD_VER >= 17
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::string_view::iterator>::element_type, const char);
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::string_view::const_iterator>::element_type, const char);
+#endif
+
+#if TEST_STD_VER >= 11
+    {
+        std::valarray<int> v;
+        auto it = std::begin(v);
+        ASSERT_SAME_TYPE(typename std::pointer_traits<decltype(it)>::element_type, int);
+        const std::valarray<int>& cv = v;
+        auto cit = std::begin(cv);
+        ASSERT_SAME_TYPE(typename std::pointer_traits<decltype(cit)>::element_type, const int);
+    }
+#endif
+
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::vector<int>::iterator>::element_type, int);
+    ASSERT_SAME_TYPE(typename std::pointer_traits<std::vector<int>::const_iterator>::element_type, const int);
+
+    return 0;
+}