Antti Haapala <antti.haapala@iki.fi> writes:
> It's also Debian (3.0).
On investigation the interval_time() function was completely broken for
the --enable-integer-datetimes case --- it was reducing the interval
value modulo one second, rather than modulo one day as intended.
I also noticed that neither the integer nor float case behaved
rationally for negative intervals. I've applied the attached patch
to 7.3 and HEAD ...
regards, tom lane
*** src/backend/utils/adt/date.c.orig Thu Feb 13 12:04:24 2003
--- src/backend/utils/adt/date.c Mon Jun 16 14:56:53 2003
***************
*** 999,1004 ****
--- 999,1009 ---- /* interval_time() * Convert interval to time data type.
+ *
+ * This is defined as producing the fractional-day portion of the interval.
+ * Therefore, we can just ignore the months field. It is not real clear
+ * what to do with negative intervals, but we choose to subtract the floor,
+ * so that, say, '-2 hours' becomes '22:00:00'. */ Datum interval_time(PG_FUNCTION_ARGS)
***************
*** 1007,1021 **** TimeADT result; #ifdef HAVE_INT64_TIMESTAMP result = span->time;
! if ((result >= INT64CONST(86400000000))
! || (result <= INT64CONST(-86400000000)))
! result -= (result / INT64CONST(1000000) * INT64CONST(1000000)); #else
- Interval span1;
- result = span->time;
! TMODULO(result, span1.time, 86400e0); #endif PG_RETURN_TIMEADT(result);
--- 1012,1034 ---- TimeADT result; #ifdef HAVE_INT64_TIMESTAMP
+ int64 days;
+ result = span->time;
! if (result >= INT64CONST(86400000000))
! {
! days = result / INT64CONST(86400000000);
! result -= days * INT64CONST(86400000000);
! }
! else if (result < 0)
! {
! days = (-result + INT64CONST(86400000000-1)) / INT64CONST(86400000000);
! result += days * INT64CONST(86400000000);
! } #else result = span->time;
! if (result >= 86400e0 || result < 0)
! result -= floor(result / 86400e0) * 86400e0; #endif PG_RETURN_TIMEADT(result);