[MERGE] Improve knit extraction time

[John Arbash Meinel]

John Arbash Meinel wrote:
...

> With bzr.dev this takes around                10.0s
> If I never collapse this takes                 9.2s
> If I collapse when num links > 2000 (2 times)  7.8s
>               when num links > 1000 (3 times)  7.0s
>               when num links > 750  (4 times)  6.7s
>               when num links > 500  (7 times)  6.5s
>               when num links > 250  (13 times) 6.6s
>               when num links > 100  (28 times) 6.7s
> 
> So I picked 500 because it seemed like a nice low spot.
> 
> I'm going to be running some more benchmarks, doing something like
> extracting every inventory text from bzr.dev. Just to make sure it is a
> general improvement, and not just an improvement when num lines is huge
> and num deltas is bigger than normal.

Well, I can show that this code is a big win for really large files with
lots of patches.

But it turns out that for smaller files (fewer lines), the time to munge
a list is pretty cheap, so the overhead of the extra tracking hurts
performance. So I'm going to be looking into it a little bit more to see
if I can strike a balance.

John
=:->