<filesystem>: `path`'s comparison operators are IF-NDR

[statementreply]

Describe the bug

After #2000, path's comparison operators are IF-NDR (ill-formed, no diagnostic required). See Arthur O'Dwyer (@Quuxplusone)'s blogpost and Jonathan Wakely (@jwakely)'s comment for a detailed analysis of similar issues in libc++ and libstdc++.

Command-line test case

Copied from Arthur O’Dwyer's blogpost.

D:\Temp>type repro.cpp

#include <filesystem>
#include <ranges>

static_assert(std::ranges::range<std::filesystem::path>);
static_assert(std::ranges::range<const std::filesystem::path>);

D:\Temp>clang-cl /c /EHsc /W4 /std:c++latest /D_HAS_CXX23=1 repro.cpp
repro.cpp(5,1): error: static_assert failed
static_assert(std::ranges::range<const std::filesystem::path>);
^                          ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
repro.cpp(5,28): note: because 'const std::filesystem::path' does not satisfy 'range'
static_assert(std::ranges::range<const std::filesystem::path>);
                           ^
D:\repos\STL\out\\build\x64\out\inc\xutility(2233,9): note: because '::std::ranges::begin(__r)' would be invalid: no
      matching function for call to object of type 'const _Begin::_Cpo'
        _RANGES begin(__r);
        ^
D:\repos\STL\out\\build\x64\out\inc\yvals_core.h(1411,20): note: expanded from macro '_RANGES'
#define _RANGES    ::std::ranges::
                   ^
1 error generated.

STL version

3c2fd04

[cpplearner]

In addition to path's comparison operators (both operator== and operator<=>), the following lines also cause clang-cl to attempt implicit conversion from const path& to wstring_view, and thus affect the result of range<const path>:

STL/stl/inc/filesystem

Line 971 in 3c2fd04

return operator+=(_Replacement);

STL/stl/inc/filesystem

Lines 1739 to 1740 in 3c2fd04

	} else if (_Elem == _Dot) { // skip filename elements that are dot, N4810 29.11.7.4.11 [fs.path.gen]/4.2
	} else if (_Elem == _Dot_dot) {

There are similar uses in templates, but uses in uninstantiated templates are probably fine.

[StephanTLavavej]

Casey Carter (@CaseyCarter) agrees that this is definitely a bug, and that we need to ensure that path::iterator is defined before its semantics affect the behavior of the program (by defining path::iterator earlier, or the dependent stuff later).

[fsb4000]

There are similar uses in templates, but uses in uninstantiated templates are probably fine.

S. B. Tam (@cpplearner) thank you for your analytics. Templates influenced this too. I had to do inline templated operator+= for the static_asserts to be successful