On Sat, Apr 2, 2022 at 3:08 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> Joseph Koshakow <koshy44@gmail.com> writes:
> > Ok I actually remember now, the issue is with the rounding
> > code in AdjustFractMicroseconds.
> > ...
> > I believe it's possible for `frac -= usec;` to result in a value greater
> > than 1 or less than -1 due to the lossiness of int64 to double
> > conversions.
>
> I think it's not, at least not for the interesting range of possible
> values in this code. Given that abs(frac) < 1 to start with, the
> abs value of usec can't exceed the value of scale, which is at most
> USECS_PER_DAY so it's at most 37 or so bits, which is well within
> the exact range for any sane implementation of double. It would
> take a very poor floating-point implementation to not get the right
> answer here. (And we're largely assuming IEEE-compliant floats these
> days.)
Ah, I see. That makes sense to me.
On Sat, Apr 2, 2022 at 3:10 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> Joseph Koshakow <koshy44@gmail.com> writes:
> > I took a stab at this issue and the attached patch (which would be
> > applied on top of your v10 patch) seems to fix the issue. Feel
> > free to ignore it if you're already working on a fix.
>
> You really only need to flip val/fval in one place. More to the
> point, there's also the hh:mm:ss paths to deal with; see my v11.
Good point. Thanks again for all the help!
- Joe Koshakow