[concepts] CTAD does not propagate constructor constraints to deduction guides

[williamspatrick]

Sample:

#include <iostream>
#include <string>

template <bool A = false, bool B = false>
struct foo
{
    constexpr foo() requires (!A && !B) = default;
    constexpr foo() requires (A || B) = delete;

    std::string value = "Hello World!";
};

int main()
{
    foo<> works{};
    std::cout << works.value << std::endl;

    foo fails{};

    return 0;
}

https://godbolt.org/z/coxs3PvbE

This fails with:

<source>:18:9: error: ambiguous deduction for template arguments of 'foo'
    foo fails{};
        ^
<source>:7:15: note: candidate function [with A = false, B = false]
    constexpr foo() requires (!A && !B) = default;
              ^
<source>:8:15: note: candidate function [with A = false, B = false]
    constexpr foo() requires (A || B) = delete;

The same code passes with GCC 12.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

[royjacobson]

I tried to understand what the standard says here, and AFAICT the overload resolution in the CTAD here happens in terms of implicitly generated deduction guides which don't propagate the trailing requires from the constructors (see [over.match.class.deduct]). So I think this is technically ambiguous.

Relevant GCC discussion from last year: https://www.mail-archive.com/gcc-patches@gcc.gnu.org/msg257689.html

Wish they'd file a CWG issue for this.

[royjacobson]

Great, this is also a duplicate of almost 3-years old #43829.

[royjacobson]

This is now CWG2528. I don't see a reason not to implement the suggested fix, since GCC/MSVC already do it.

[royjacobson]

https://reviews.llvm.org/D134145