Re: first bisection step takes quite a while

> > >>  * The "this is good enough" logic currently allows us to be within
> > >>    0.1% of the real halfway point.  Until the candidate set becomes
> > >>    small enough, we could loosen the criteria to allow larger, say
> > >>    3%, slack.  This code is written but not enabled (with "0 &&").
> >
> > The above follows the same reasoning why we chose "division by 1024"
> > in the first place.  The illustration patch postulates that we could
> > be way more aggressive than 0.1% while the set is large by dividing
> > 64, without wanting to loosen the criteria near the end of the
> > bisection session when the remaining set is reasonably small like
> > 1000 commits.  So we cannot rely on integer division truncating.
>
> The code you posted above uses 10000 as the threshold, not 1000:
>
> 10000 < nr && abs(diff) < nr / 64) || abs(diff) < nr / 1024)