Test `enforce-checks-test.cc` fails after implementing P2499

[strega-nil-ms]

Hi! Friendly neighborhood STL implementor here 😸

After implementing P2499R0 in microsoft/stl#2947, the test enforce-checks-test now fails.

Compiler output

C:\Users\nimazzuc\projects\fmt\include\fmt\core.h(2742): error C2039: 'parse': is not a member of 'fmt::v9::detail::fallback_formatter<stripped_type,char_type,void>'
C:\Users\nimazzuc\projects\fmt\include\fmt\core.h(2741): note: see declaration of 'fmt::v9::detail::fallback_formatter<stripped_type,char_type,void>'
C:\Users\nimazzuc\projects\fmt\include\fmt\core.h(2947): note: see reference to function template instantiation 'const char *fmt::v9::detail::parse_format_specs<std::vector<char,std::allocator<char>>,fmt::v9::detail::compile_parse_context<Char,ErrorHandler>>(ParseContext &)' being compiled
        with
        [
            Char=char,
            ErrorHandler=fmt::v9::detail::error_handler,
            ParseContext=fmt::v9::detail::compile_parse_context<char,fmt::v9::detail::error_handler>
        ]
C:\Users\nimazzuc\projects\fmt\include\fmt\core.h(2947): note: while compiling class template member function 'fmt::v9::detail::format_string_checker<Char,fmt::v9::detail::error_handler,std::vector<char,std::allocator<char>>>::format_string_checker(fmt::v9::basic_string_view<Char>,ErrorHandler)'
        with
        [
            Char=char,
            ErrorHandler=fmt::v9::detail::error_handler
        ]
C:\Users\nimazzuc\projects\fmt\include\fmt\core.h(3141): note: see reference to class template instantiation 'fmt::v9::detail::format_string_checker<Char,fmt::v9::detail::error_handler,std::vector<char,std::allocator<char>>>' being compiled
        with
        [
            Char=char
        ]
.\test\enforce-checks-test.cc(55): note: see reference to function template instantiation 'fmt::v9::basic_format_string<char,std::vector<char,std::allocator<char>> &>::basic_format_string<test_range::<lambda_1>::()::FMT_COMPILE_STRING,0>(const S &)' being compiled
        with
        [
            S=test_range::<lambda_1>::()::FMT_COMPILE_STRING
        ]

Before, it worked because basic_string_view<char> was convertible from any contiguous range of char; now, it is only constructible from any contiguous range of char.

[vitaut]

This is surprising because vector<char> doesn't need to be convertible to string_view to be formattable as a range of chars. See e.g. https://godbolt.org/z/4oPK6dxM5.

[strega-nil-ms]

Yeah, I dunno; I think the metaprogramming is broken somehow, but it slipped under the radar because vector<char> was convertible to string_view

[QuellaZhang]

Any updates or patch for this?

[vitaut]

I don't have a windows machine at hand and the issue doesn't repro on godbolt. If someone is able to repro and can investigate this issue it would be appreciated.

[strega-nil-ms]

So, interestingly, it does work for vector<int>, but not for vector<char>:

#include <fmt/format.h>
#include <fmt/ranges.h>
#include <vector>

using namespace std;

int main() {
  vector<CHAR_OR_INT> hello{'h', 'e', 'l', 'l', 'o'};
  fmt::print("{}\n", hello);
}

works when -DCHAR_OR_INT=int, but doesn't when -DCHAR_OR_INT=char

[strega-nil-ms]

It looks like vector<char> is not considered a range for some reason:

#include <fmt/format.h>
#include <fmt/ranges.h>
#include <vector>

using namespace std;

int main() {
  static_assert(fmt::is_range<vector<char>, char>::value); // FAILS
  static_assert(fmt::is_range<vector<int>, char>::value); // SUCCEEDS
}

[strega-nil-ms]

I think I figured out the problem:

template <typename T, typename Char> struct is_range {
  static constexpr const bool value =
      detail::is_range_<T>::value && !detail::is_std_string_like<T>::value &&
      !std::is_convertible<T, std::basic_string<Char>>::value &&
      !std::is_constructible<detail::std_string_view<Char>, T>::value; // <--- problem
};

string_view is constructible from vector<char>, but vector<char> is not implicitly convertible to string_view; therefore, vector<char> is not a range, but because it's not convertible, it doesn't get formatted as a string_view either.

(I think this may be the issue in #3050)

[vitaut]

Thanks for investigating, @strega-nil-ms. Does https://github.com/fmtlib/fmt/commit/e4f4fc77884196163b45b528875bf65d1c9b922e fix the issue?

[strega-nil-ms]

@vitaut it should, but it's still incorrect - what you have is

   static constexpr const bool value =
       detail::is_range_<T>::value && !detail::is_std_string_like<T>::value &&
       !std::is_convertible<T, std::basic_string<Char>>::value &&
-      !std::is_constructible<detail::std_string_view<Char>, T>::value;
+      !std::is_convertible<detail::std_string_view<Char>, T>::value;

but what you want is

   static constexpr const bool value =
       detail::is_range_<T>::value && !detail::is_std_string_like<T>::value &&
       !std::is_convertible<T, std::basic_string<Char>>::value &&
-      !std::is_constructible<detail::std_string_view<Char>, T>::value;
+      !std::is_convertible<T, detail::std_string_view<Char>>::value;

note the different parameter ordering of is_constructible and is_convertible

[vitaut]

the different parameter ordering of is_constructible and is_convertible

Ugh.

Thanks!

[vitaut]

Should be fixed in https://github.com/fmtlib/fmt/commit/4a8e2949bb8d803a7607b62b632ec6d8669b7995.