Adding years with + and sys_days yields different result than += with year_month_day

[QuantDevHacks]

I'm noticing a one-day discrepancy when adding years to a date, depending on whether a year_month_day is converted to sys_days with the + operator defined (incorrect result), or if += is applied to the original year_month_day object (correct).

A repro is as follows (Visual Studio compiler, ISO C++20 Standard):

using namespace std::chrono;

using std::cout;
using std::endl;

year_month_day ymd{ year{2002}, month{11}, day{14} };

auto add_years = sys_days(ymd) + years(20);   // Result is 2022-11-13 20:24:00 (incorrect)
cout << add_years  << endl;                   //  type is std::chrono::time_point<...>

ymd += years{ 20 };           // Result is 2022-11-14 (correct)
cout << ymd << endl;          // type is year_month_day

One more thing to note: the 20:24:00 time stamp in the 1st case is constant -- I tried this at different times but still got the same result.

[Erroneous1]

When you work with sys_days, you are using a time_point, which is a duration of time since a given epoch as defined by the given clock (not including leap seconds in the case of system_clock). On average there are 7304.85 days in 20 years, so you are getting the expected result there.

year_month_day is a calendar time instead of a time_point. Adding 20 years to 2002 produces 2022 and leaves the month and day alone, which is what you expected.

[HowardHinnant]

Here is an analogous description of adding months: https://stackoverflow.com/a/43018120/576911

It describes how there are both chronological and calendrical algorithms to do this arithmetic. Though it describes the situation with months, the situation with years is the same.

[QuantDevHacks]

Thanks very much to both of you for the follow-up. That makes it reasonably clear now.