: Fix formatting for unusual durations

[A. Jiang (frederick-vs-ja)]

Fixes #3644. hh_mm_ss<_Duration> seems wrong for _Hh_mm_ss_part_underflow_to_zero to me.

Need to explore how to accept duration>. Currently libstdc++ rejects it, while libc++ accepts it (Godbolt link).

Edit: the current approach for calculation of hh_mm_ss::fractional_width doesn't seem right. I'm trying to fix it.

Edit 2: now the formatting is made well-formed, but UB may happen.

[A. Jiang (frederick-vs-ja)]

Currently no implementation correctly formats duration>{40} (Godbolt link). In MSVC STL, libc++, and libstdc++, hh_mm_ss is used when formatting duration and thus duration_cast is also involved. And it's extremely easy for duration_cast to cause UB when the source type is duration>.

However, it seems that [time.format] doesn't say duration_cast or hh_mm_ss is equivalently used when formatting duration, so perhaps the current standard wording is requiring correctly formatting such duration...

[nicole mazzuca (strega-nil-ms)]

Wait, actually, I'm trying to understand the UB here... could you write up an example?

[A. Jiang (frederick-vs-ja)]

Wait, actually, I'm trying to understand the UB here... could you write up an example?

I attempted to test formatting duration>{40}. The duration needs to undergo _Hh_mm_ss_part_underflow_to_zero, and thus _Remove_duration_part. Finally duration_cast(...) needs to be called, and signed integer overflow is unavoidable.

Reduced example that contains needed runtime operations and fails to compile due to integer overflow (Godbolt link):

#include <chrono>
#include <cstdint>
#include <limits>
#include <ratio>
#include <type_traits>

using LongRatio = std::ratio<INT64_MAX - 1, INT64_MAX>;
using D = std::chrono::duration<int, LongRatio>;

constexpr auto weird40 = D{40};
constexpr auto cast_to_day = std::chrono::duration_cast(weird40);

hh_mm_ss conversion doesn't work either (Godbolt link).

[Stephan T. Lavavej (StephanTLavavej)]

Thanks! I pushed a conflict-free merge with main and trivial header inclusions for the test.

[Stephan T. Lavavej (StephanTLavavej)]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[Stephan T. Lavavej (StephanTLavavej)]

We forgot to defend _STD max against macroization - pushed.

[Stephan T. Lavavej (StephanTLavavej)]

Thanks for investigating and fixing this bug! ⌚ 🎉 😸