Tom Lane wrote:
> Ron Johnson <ron.l.johnson@cox.net> writes:
> > Out of curiosity: why does -ffast-math break the datetime rounding code?
>
> We dug into this last night, and it turns out that the culprit is code
> like
>
> int hour = time / 3600;
>
> where time is a double. This yields an exact result when done
> correctly, but with -ffast-math gcc will "improve" it to
>
> int hour = time * 0.000277777777777778;
>
> the constant being the nearest double value to 1.0 / 3600.0. The
> problem is that the constant is inexact and in fact is slightly too
> large; so for example if time is exactly 18000.0, you get a resulting
> hour value of 4, not 5, after truncation to integer. Repeated a couple
> more times, what should have been 5:00:00 comes out as 4:59:60 ...
Hard to imagine why anyone would want such an optimization. How much
faster could it possibly be? I guess if you were doing only complex
math approximations, it would be a win, but that isn't really proper for
a database.
--
Bruce Momjian | http://candle.pha.pa.us
pgman@candle.pha.pa.us | (610) 359-1001
+ If your life is a hard drive, | 13 Roberts Road
+ Christ can be your backup. | Newtown Square, Pennsylvania 19073
