Avoid unnecessary division in interval_cmp_value().

Splitting the time field into days and microseconds is pretty
useless when we're just going to recombine those values.
It's unclear if anyone will notice the speedup in real-world
cases, but a cycle shaved is a cycle earned.

Discussion: https://postgr.es/m/2629129.1632675713@sss.pgh.pa.us

Author: tglsfdc

src/backend/utils/adt/timestamp.c

@@ -2352,20 +2352,17 @@ static inline INT128
 interval_cmp_value(const Interval *interval)
 {
 	INT128		span;
-	int64		dayfraction;
 	int64		days;
 
 	/*
-	 * Separate time field into days and dayfraction, then add the month and
-	 * day fields to the days part.  We cannot overflow int64 days here.
+	 * Combine the month and day fields into an integral number of days.
+	 * Because the inputs are int32, int64 arithmetic suffices here.
 	 */
-	dayfraction = interval->time % USECS_PER_DAY;
-	days = interval->time / USECS_PER_DAY;
-	days += interval->month * INT64CONST(30);
+	days = interval->month * INT64CONST(30);
 	days += interval->day;
 
-	/* Widen dayfraction to 128 bits */
-	span = int64_to_int128(dayfraction);
+	/* Widen time field to 128 bits */
+	span = int64_to_int128(interval->time);
 
 	/* Scale up days to microseconds, forming a 128-bit product */
 	int128_add_int64_mul_int64(&span, days, USECS_PER_DAY);