The branchhead computation is one of the more heavy operations for
bigger repositories as it has to scan all changesets and potentially
involves the expensive computation of the ancestor sets. Redo the
computation to handle the common cases directly and use tighter
conditions for when the ancestor scan is necessary. Most importantly,
avoid it completely if the non-continous branches are processed in one
update as seen in the initial computation after a clone.
For the Mercurial repository, it gives a small 2-3% performance boost.
For the NetBSD test repository, it cuts the time in half.
Not that in theory, turning cl.ancestors into a set can be quite expensive. The return of ancestors is a expected to be lazy ancestors object expected to be good at membership test. So I would expect something like that to be faster:
(and if it is not, we should fix lazy ancestors)