Stefan Sperling:

> > I'd be fine with this if people prefer performance over sanity checks.
> 
> Update: We should then also drop the head_ref variable.

My test case is the Git conversion of OpenBSD src,
got branch -c cb8c029a90 local, and cherrypicking the commits in
usr.bin/rsync:
got cy 57987d16c1
got cy 722539fd92
got cy fecc32f18d
got cy ede24f484b,
got cy 0345af144f

And I ran that on an APU2.  Slow by x86 standards, pretty fast by
non-x86 standards.

With HEAD:
    5m50.33s real     4m02.36s user     2m04.41s system

With your patch:
    1m04.92s real     0m28.24s user     0m36.54s system

That's a lot faster.
It's still fairly slow, though.

Backout of the same commits from the tip of master:
    1m07.46s real     0m29.40s user     0m38.02s system

-- 
Christian "naddy" Weisgerber                          naddy@mips.inka.de