After one hour and various unit tests, I came up with this solution:
public static Timestamp diff (java.util.Date t1, java.util.Date t2)
{
// Make sure the result is always > 0
if (t1.compareTo (t2) < 0)
{
java.util.Date tmp = t1;
t1 = t2;
t2 = tmp;
}
// Timestamps mix milli and nanoseconds in the API, so we have to separate the two
long diffSeconds = (t1.getTime () / 1000) - (t2.getTime () / 1000);
// For normals dates, we have millisecond precision
int nano1 = ((int) t1.getTime () % 1000) * 1000000;
// If the parameter is a Timestamp, we have additional precision in nanoseconds
if (t1 instanceof Timestamp)
nano1 = ((Timestamp)t1).getNanos ();
int nano2 = ((int) t2.getTime () % 1000) * 1000000;
if (t2 instanceof Timestamp)
nano2 = ((Timestamp)t2).getNanos ();
int diffNanos = nano1 - nano2;
if (diffNanos < 0)
{
// Borrow one second
diffSeconds --;
diffNanos += 1000000000;
}
// mix nanos and millis again
Timestamp result = new Timestamp ((diffSeconds * 1000) + (diffNanos / 1000000));
// setNanos() with a value of in the millisecond range doesn't affect the value of the time field
// while milliseconds in the time field will modify nanos! Damn, this API is a *mess*
result.setNanos (diffNanos);
return result;
}
Unit tests:
Timestamp t1 = new Timestamp (0);
Timestamp t3 = new Timestamp (999);
Timestamp t4 = new Timestamp (5001);
// Careful here; internally, Java has set nanos already!
t4.setNanos (t4.getNanos () + 1);
// Show what a mess this API is...
// Yes, the milliseconds show up in *both* fields! Isn't that fun?
assertEquals (999, t3.getTime ());
assertEquals (999000000, t3.getNanos ());
// This looks weird but t4 contains 5 seconds, 1 milli, 1 nano.
// The lone milli is in both results ...
assertEquals (5001, t4.getTime ());
assertEquals (1000001, t4.getNanos ());
diff = DBUtil.diff (t1, t4);
assertEquals (5001, diff.getTime ());
assertEquals (1000001, diff.getNanos ());
diff = DBUtil.diff (t4, t3);
assertEquals (4002, diff.getTime ());
assertEquals (2000001, diff.getNanos ());