The behaviour of git bisect skip

["H. Peter Anvin" <hpa@zytor.com>]

I recently had the unhappy experience of trying to bisect a tree with a 
large region of the history obscured by auxilliary bugs.  "git bisect 
skip" will stay in the central region, thus being largely useless.

I was thinking about how to possibly do it better.  This is something I 
came up with, and thought it might be interesing.

a. we obviously cannot move the start and end (good and bad) markers, 
since they have not been shown one way or the other.

b. however, the practice of testing the centermost point is merely the 
*optimal*, corresponding to 1 bit of information per iteration.  An 
off-center test is also possible (as long as the value depends on both 
endpoints, and isn't fixed from one of the endpoints; in that case we 
have a linear search), corresponding to a smaller amount of information 
- for example, sampling at the one-quarter point corresponds to 
3/4*log2(3/4) + 1/4*log2(1/4) =~ 0.811 bits of information.

I would suggest something based on the following algorithm:

Given an interval with a certain number of skip points, subdivide the 
interval into subintervals each separated by a skip point.  Pick the 
centermost point of the *largest* of these intervals.  If there is more 
than one largest interval, choose the one centermost point that ends up 
being centermost in the overall interval.

This algorithm obviously needs some adjustment (as does plain binary 
search) in order to deal with a branched history, but I thought it might 
be an interesting starting point.  It has the desirable property that it 
can make forward progress even in the presence of skip points, and that 
it avoids needlessly searching close to skip points.

	-hpa