2 juni 2020 kl. 18.14 skrev Eli Zaretskii <eliz@gnu.org>:

> I'd prefer to fix only the symmetry bug (which AFAIU happens because
> we use bit shifts on signed integers), without introducing any other
> effects on the function's behavior and return values.  AFAIU, such a
> fix should not require any changes outside of the function itself.

Very well, it is obviously an improvement. The reason for the current asymmetry was actually that the algorithm discarded the low bits; what about fixing that as well? The improved accuracy amounts to less than 1 % of difference in the return value; no other code needs changing, and we get the symmetry for free. Proposed patch attached.