Printing std::chrono::time_point requires std::to_chars<long double>

[georgthegreat]

MWE (using gtest here) is as follows:

#include <chrono>
#include <format>
#include <sstream>


TEST(StdFormatShould, FormatTimestamp) {
    using Ts = std::chrono::time_point<std::chrono::system_clock, std::chrono::milliseconds>;
    Ts ts{0};
    std::stringstream ss;
    ss << ts;
}

Attempting to compile this on macOS with the latest SDK results in

In file included from test_format.cpp:1:
In file included from /usr/include/c++/v1/chrono:837:
In file included from /usr/include/c++/v1/__chrono/formatter.h:23:
In file included from /usr/include/c++/v1/__chrono/ostream.h:30:
In file included from /usr/include/c++/v1/__format/format_functions.h:29:
/usr/include/c++/v1/__format/formatter_floating_point.h:73:30: error: 'to_chars' is unavailable: introduced in macOS 13.3
   73 |   to_chars_result __r = std::to_chars(__first, __last, __value, __fmt, __precision);
      |                              ^
/usr/include/c++/v1/__format/formatter_floating_point.h:392:34: note: in instantiation of function template specialization 'std::__formatter::__to_buffer<float>' requested here
  392 |   __result.__last = __formatter::__to_buffer(__integral, __buffer.end(), __value, chars_format::general, __precision);
      |                                  ^
/usr/include/c++/v1/__format/formatter_floating_point.h:459:27: note: in instantiation of function template specialization 'std::__formatter::__format_buffer_general_lower_case<float, float>' requested here
  459 |       return __formatter::__format_buffer_general_lower_case(__buffer, __value, __buffer.__precision(), __first);
      |                           ^
/usr/include/c++/v1/__format/formatter_floating_point.h:657:42: note: in instantiation of function template specialization 'std::__formatter::__format_buffer<float, float>' requested here
  657 |   __float_result __result = __formatter::__format_buffer(
      |                                          ^
/usr/include/c++/v1/__format/formatter_floating_point.h:763:25: note: in instantiation of function template specialization 'std::__formatter::__format_floating_point<float, char, std::format_context>' requested here
  763 |     return __formatter::__format_floating_point(__value, __ctx, __parser_.__get_parsed_std_specifications(__ctx));
      |                         ^
/usr/include/c++/v1/__format/format_functions.h:287:42: note: in instantiation of function template specialization 'std::__formatter_floating_point<char>::format<float, std::format_context>' requested here
  287 |             __ctx.advance_to(__formatter.format(__arg, __ctx));
      |                                          ^
/usr/include/c++/v1/__format/format_functions.h:427:15: note: (skipping 9 contexts in backtrace; use -ftemplate-backtrace-limit=0 to see all)
  427 |   return std::__vformat_to(std::move(__out_it), __fmt, __args);
      |               ^
/usr/include/c++/v1/__format/format_arg_store.h:249:19: note: in instantiation of function template specialization 'std::__format::__create_packed_storage<std::format_context, const std::chrono::time_point<std::chrono::system_clock, std::chrono::duration<long long, std::ratio<1, 1000>>>>' requested here
  249 |         __format::__create_packed_storage(__storage.__types_, __storage.__values_, __args...);
      |                   ^
/usr/include/c++/v1/__format/format_functions.h:70:10: note: in instantiation of member function 'std::__format_arg_store<std::format_context, const std::chrono::time_point<std::chrono::system_clock, std::chrono::duration<long long, std::ratio<1, 1000>>>>::__format_arg_store' requested here
   70 |   return _VSTD::__format_arg_store<_Context, _Args...>(__args...);
      |          ^
/usr/include/c++/v1/__config:897:17: note: expanded from macro '_VSTD'
  897 | #  define _VSTD std
      |                 ^
/usr/include/c++/v1/__format/format_functions.h:610:59: note: in instantiation of function template specialization 'std::make_format_args<std::format_context, const std::chrono::time_point<std::chrono::system_clock, std::chrono::duration<long long, std::ratio<1, 1000>>>>' requested here
  610 |   return std::vformat(std::move(__loc), __fmt.get(), std::make_format_args(__args...));
      |                                                           ^
/usr/include/c++/v1/__chrono/ostream.h:48:23: note: in instantiation of function template specialization 'std::format<const std::chrono::time_point<std::chrono::system_clock, std::chrono::duration<long long, std::ratio<1, 1000>>> &>' requested here
   48 |   return __os << std::format(__os.getloc(), _LIBCPP_STATICALLY_WIDEN(_CharT, "{:L%F %T}"), __tp);
      |                       ^
test_format.cpp:10:5: note: in instantiation of function template specialization 'std::chrono::operator<<<char, std::char_traits<char>, std::chrono::duration<long long, std::ratio<1, 1000>>>' requested here
   10 |         ss << ts;
      |            ^

This is probably related to the format used here which uses %F:

f, F: Produces the output as if by calling
std::to_chars(first, last, value, std::chars_format::fixed, precision) where precision is the specified precision, or 6 if precision is not specified.

despite being explicitly annotated with !treat_as_floating_point_v<typename _Duration::rep>

macOS 13.3 was releases just 2 years ago and we can no raise -mmacos-system-min to this value just to fix the test.

@ldionne, could you, please, take a look?
There is probably nothing that can be done to fix current SDKs already available (except weird patches applied to the client code), but new versions might have this fixed.

[mordante]

Format indeed uses std::to_chars internally. Since the floating-point tables are large we put them in the dylib.
This indeed only makes them available on newer MacOS versions. I don't think there is a solution.

[georgthegreat]

But there is nothing regarding floating point formatting in the case, as the time_point being formatted is integral!
What is the point of using %F there? Isn't this a perfomance issue, @mordante?

[mordante]

This matches the specification in the standard [time.clock.system.nonmembers]/2

Effects: Equivalent to:
  return os << format(os.getloc(), STATICALLY-WIDEN<charT>("{:L%F %T}"), tp);

In this case %F is equivalent to %Y-%m-%d and not a floating point format.

I missed the part that you didn't use a floating-point representation. Unfortunately that does not change the issue. For some reason not having floating-point support for to_chars disables more formatting on MacOS. For example

llvm-project/libcxx/test/std/time/time.syn/formatter.year.pass.cpp

Line 13 in 3f8e280

// TODO FMT This test should not require std::to_chars(floating-point)

test has the same issue. I have look at it with @ldionne in the past and we haven't found a solution yet.

[georgthegreat]

Well, this is definitely a degradation which prevents us from switching to a newer SDK / libc++ version (thus increasing techdebt / the price of a future switching).
I might propose a workaround to googletest, but I do not believe it will be accepted.

[mordante]

Can you explain what you mean with "this is definitely a degradation"?

[georgthegreat]

Prior to SDK upgrade Google Test code above was able to compile without depending on newer macOS.
The fact that newer SDK brings dependency on newer macOS is a degradation.

[mordante]

Thanks for the explanation it was not clear to me the code stopped working after an SDK update. What is the working version of the SDK?

[georgthegreat]

I am updating from Xcode 15.1 (15C65).
The problem is probably caused by gtest dispatching its printing to std::format whenever it is available.