[libc++] std::to_chars precision loss for 80-bit long double

[alvinhochun]

std::to_char in libc++ converts long double to double:

llvm-project/libcxx/src/charconv.cpp

Lines 81 to 83 in 3c126d5

	to_chars_result to_chars(char* __first, char* __last, long double __value, chars_format __fmt, int __precision) {
	return _Floating_to_chars<_Floating_to_chars_overload::_Format_precision>(
	__first, __last, static_cast<double>(__value), __fmt, __precision);

Test shows libc++ std::to_char output being different from printf/std::ostream, and compared to libstdc++: https://godbolt.org/z/GsYfTTjq7

[mordante]

The original code was provided by Microsoft, on Windows double and long double have the same size.
It's on my todo list to improve this at some point, it's also listed at the status page.

https://libcxx.llvm.org/Status/Cxx17.html#note-p0067

[alvinhochun]

Before this gets implemented properly, would it be better to delete the to_chars overload for long double if it has a different size than double? Seems to me that having it produce the wrong result can do more harm than making it unavailable.

[JustAnotherHumanBeing]

And when will libc++ support from_chars for floating types?

[mordante]

Before this gets implemented properly, would it be better to delete the to_chars overload for long double if it has a different size than double? Seems to me that having it produce the wrong result can do more harm than making it unavailable.

During review there were no objections and we already shipped this. So we can no longer remove it. There will be differences in some cases, but I'm not sure how often that happens in practice. I'm not even sure whether anybody still uses long doubles.

And when will libc++ support from_chars for floating types?

It's on my todo list, however that list is rather long and this item doesn't have a high priority for me. So I don't expect to work on it this year. Maybe somebody else will pick it up before it's on the top of my list.

[mordante]

Since we typically don't use the bug tracker for not implemented features I'm closing this.

[WarrenWeckesser]

A very late "strong agree" with @alvinhochun's comment. That silent cast to double is an annoying "gotcha".

I'm not even sure whether anybody still uses long doubles.

The 80 bit extended precision long doubles aren't that useful. But true IEEE float128 is available on some platforms, and that can be useful.