`<ranges>`: `views::repeat(...) | views::take(...)` is not always a valid range

[JMazurkiewicz]

Command-line test case

#include <ranges>

using namespace std;

const auto rice = views::repeat("🌾"sv, 100zu); // difference_type is _Signed128
const auto taken_rice = rice | views::take(3);
static_assert(ranges::range<decltype(taken_rice)>); // error!

PS D:\stl-playground> clang-cl /std:c++latest /Zs .\iota_diff.cpp
In file included from .\iota_diff.cpp:1:
In file included from D:\stl\Out\build\x64\out\inc\ranges:16:
In file included from D:\stl\Out\build\x64\out\inc\iterator:12:
In file included from D:\stl\Out\build\x64\out\inc\xutility:12:
D:\stl\Out\build\x64\out\inc\__msvc_iter_core.hpp(96,1): error: no type named 'difference_type' in 'std::incrementable_traits<std::_Signed128>'
using iter_difference_t = typename conditional_t<_Is_from_primary<iterator_traits<remove_cvref_t<_Ty>>>,
^~~~~
[...]

The main reason of this error is that _Signed128 does not model weakly_incrementable.

Expected behavior
This code should compile.

Additional context
The truth is that standard does not require _Signed128 to model weakly_incrementable ([range.iota.view]/1), but it would be very nice if it did:

#include <ranges>

using namespace std;

using X = ranges::_Iota_diff_t<size_t>;
static_assert(!weakly_incrementable<X>);
using Y = ranges::_Iota_diff_t<X>; // <= compile error

[frederick-vs-ja]

I've verified that ranges::repeat_view<T, _Signed128> and ranges::repeat_view<T, _Unsigned128> are both broken.

I guess the title of this issue should be changed to "<ranges>: ranges::repeat_view<T, IntegerClass> is broken".

---

This is probably a defect in current specification, and an LWG issue seems needed.

[range.repeat.iterator] says

    using difference_type = conditional_t<is-signed-integer-like<index-type>,
        index-type,
        IOTA-DIFF-T(index-type)>;

which always instantiates IOTA-DIFF-T(index-type), and thus possibly makes the program ill-formed when index-type (same as Bound in such case) is an integer-class type.

I think the intent is when index-type is signed-integer-like, it is used as-is without instantiating IOTA-DIFF-T, which fix cases for signed integer-class types. But this may be still broken when _Bound is an unsigned integer-class type (_Unsigned128 in MSVC STL).

Edit: I've mailed LWG Chair to submit the issue.

Edit: this is now LWG-3875.

[frederick-vs-ja]

Oh, there seems another issue.

It's arguably better if views::take & views::drop preserve the bound type of repeat_view whenever possible. When E is an unbounded repeat_view, I guess views::take(E, F) should return a repeat_view whose bound type is either ptrdiff_t or decltype(auto(F)) (in which case this type must be integer-like).

Edit: such type promotion only occurs for unsigned integer-like types. I can live with it.