dmitry-chirkov-dremio commented on code in PR #53325:
URL: https://github.com/apache/arrow/pull/53325#discussion_r4560167131
##########
cpp/src/arrow/compute/kernels/scalar_arithmetic.cc:
##########
@@ -1126,6 +1135,368 @@ void
AddArithmeticFunctionDurationTime(std::shared_ptr<ScalarFunction> func) {
std::move(exec_4)));
}
+// ----------------------------------------------------------------------
+// Timestamp +/- interval arithmetic
+//
+// An interval is split into a calendar part (months and days) and a physical
+// part (the milliseconds of a day_time_interval or the nanoseconds of a
+// month_day_nano_interval). The calendar part is applied to the local
+// (wall clock) time of the timestamp, so that for example adding one day
+// keeps the same wall clock time across a DST transition. The physical part
+// is a fixed duration.
+//
+// `add` applies months, then days, then the physical part. `subtract`
+// applies the physical part first, so that subtracting e.g. "1 day 24 hours"
+// undoes adding it. Months are applied before days in both cases, as in
+// Pandas' DateOffset and PostgreSQL.
+//
+// Adding months clamps the day of month to the last day of the resulting
+// month (e.g. Jan 31 + 1 month = Feb 28). An error is returned if the
+// resulting local time is ambiguous or does not exist, as when adding a
+// DateOffset to a timezone-aware Pandas Series.
+
+struct CalendarInterval {
+ int32_t months;
+ int32_t days;
+ int64_t nanoseconds;
+};
+
+CalendarInterval ToCalendarInterval(int32_t months) { return {months, 0, 0}; }
+
+CalendarInterval ToCalendarInterval(DayTimeIntervalType::DayMilliseconds v) {
+ return {0, v.days, static_cast<int64_t>(v.milliseconds) * 1000000};
+}
+
+CalendarInterval ToCalendarInterval(MonthDayNanoIntervalType::MonthDayNanos v)
{
+ return {v.months, v.days, v.nanoseconds};
+}
+
+int64_t FloorDiv(int64_t x, int64_t y) {
+ const int64_t q = x / y;
+ return (x % y != 0 && ((x < 0) != (y < 0))) ? q - 1 : q;
+}
+
+// Proleptic Gregorian calendar conversions, using 64-bit years so that the
+// whole range of second-resolution timestamps is covered.
+// See https://howardhinnant.github.io/date_algorithms.html
+int64_t DaysFromCivil(int64_t y, int64_t m, int64_t d) {
+ y -= m <= 2;
+ const int64_t era = FloorDiv(y, 400);
+ const int64_t yoe = y - era * 400;
+ const int64_t doy = (153 * (m > 2 ? m - 3 : m + 9) + 2) / 5 + d - 1;
+ const int64_t doe = yoe * 365 + yoe / 4 - yoe / 100 + doy;
+ return era * 146097 + doe - 719468;
+}
+
+void CivilFromDays(int64_t z, int64_t* y, int64_t* m, int64_t* d) {
+ z += 719468;
+ const int64_t era = FloorDiv(z, 146097);
+ const int64_t doe = z - era * 146097;
+ const int64_t yoe = (doe - doe / 1460 + doe / 36524 - doe / 146096) / 365;
+ const int64_t doy = doe - (365 * yoe + yoe / 4 - yoe / 100);
+ const int64_t mp = (5 * doy + 2) / 153;
+ *d = doy - (153 * mp + 2) / 5 + 1;
+ *m = mp < 10 ? mp + 3 : mp - 9;
+ *y = yoe + era * 400 + (*m <= 2);
+}
+
+int64_t LastDayOfMonth(int64_t y, int64_t m) {
+ if (m == 2) {
+ const bool is_leap = y % 4 == 0 && (y % 100 != 0 || y % 400 == 0);
+ return is_leap ? 29 : 28;
+ }
+ return (m == 4 || m == 6 || m == 9 || m == 11) ? 30 : 31;
+}
+
+// Add a number of months to a day count, clamping the day of month.
+int64_t AddMonthsToDays(int64_t days, int64_t months) {
+ int64_t y, m, d;
+ CivilFromDays(days, &y, &m, &d);
+ const int64_t total_months = y * 12 + (m - 1) + months;
+ y = FloorDiv(total_months, 12);
+ m = total_months - y * 12 + 1;
+ return DaysFromCivil(y, m, std::min(d, LastDayOfMonth(y, m)));
+}
+
+// Converts between UTC and local time (in seconds) for a timestamp's timezone.
+class IntervalLocalizer {
+ public:
+ static Result<IntervalLocalizer> Make(const std::string& timezone) {
+ IntervalLocalizer localizer;
+ if (timezone.empty()) return localizer;
+ ARROW_ASSIGN_OR_RAISE(ArrowTimeZone tz, LocateZone(timezone));
+ if (auto zone = std::get_if<const time_zone*>(&tz)) {
+ localizer.zone_ = *zone;
+ } else {
+ const auto& offset_zone = std::get<OffsetZone>(tz);
+ localizer.fixed_offset_ =
+
offset_zone.get_info(sys_time<std::chrono::seconds>{}).offset.count();
+ }
+ return localizer;
+ }
+
+ // The bundled date library only represents years in [-32767, 32767], and
+ // throws when asked to convert an instant outside that range. Clamp to a
+ // slightly smaller range (about year +/-31700) before calling get_info().
+ static constexpr int64_t kMaxTzSeconds = 1000000000000;
Review Comment:
The epoch starts in 1970 :)
1,000,000,000,000 from epoch puts us in year 33658 outside the limit of
32767.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
