Previously, we chose a rev based on numeric ordering, which could
cause "the same merge" in topologically identical but numerically
different repos to choose different merge bases.
We now choose the lexically least node; this is stable across
different revlog orderings.
Instead of walking all the way to the root of the DAG, we generate
a set of candidate GCA revs, then figure out which ones will win
the race to the root (usually without needing to traverse all the
way to the root).
In the common case of nodes that are close to each other in both
revision number and topology, this is usually a big win: it makes
"hg --time debugancestors" up to 9 times faster than the more general
ancestor function when measured on heads of the linux-2.6 hg repo.
Victory is not assured, however. The older function can still win
by a large margin if one node is much closer to the root than the
other, or by a much smaller amount if one is an ancestor of the
other.
For now, we've also got a small paranoid harness function that calls
both ancestor functions on every input and ensures that they give
equivalent answers.
Even without the checker function, the old ancestor function needs
to stay alive for the time being, as its generality is used by
context.filectx.merge.
This also makes the perfancestorset command use lazy membership testing. In a
linear repository with over 400,000 commits, without this patch, hg
perfancestorset takes 0.80 seconds no matter how far behind we're looking.
With this patch, hg perfancestorset -- X takes:
Rev X Time
-1 0.00s
-4000 0.01s
-20000 0.04s
-80000 0.17s
-200000 0.43s
-300000 0.69s
0 0.88s
Thus, for revisions close to tip, we're up to several orders of magnitude
faster. At 0 we're around 10% slower.
One of the major reasons rebase is slow in large repositories is
the computation of the detach set: the set of ancestors of the
changesets to rebase not in the destination parent. This is currently
done via a revset that does two walks all the way to the root of
the DAG. Instead of doing that, to find ancestors of a set <revs>
not in another set <common> we walk up the tree in reverse revision
number order, maintaining sets of nodes visited from <revs>, <common>
or both.
For the common case where the sets are close both topologically and
in revision number (relative to repository size), this has been
found to speed up rebase by around 15-20%. When the nodes are farther
apart and the DAG is highly branching, it is harder to say which
would win.
Here's how long computing the detach set takes in a linear repository
with over 400000 changesets, rebasing near tip:
Rebasing across 4 changesets
Revset method: 2.2s
New algorithm: 0.00015s
Rebasing across 250 changesets
Revset method: 2.2s
New algorithm: 0.00069s
Rebasing across 10000 changesets
Revset method: 2.4s
New algorithm: 0.019s
When computing the DAG depth, we walk through all ancestors: this commit adds
memoization during that first step. Then, the memorized parents are fetched
from a dict instead of calling parents() on each vertex.
This tweak, according to Kcachegrind, improves ancestor() performance by 16%
when cloning a big repository.
Make depth calculation non-recursive
Add simple shortcut for linear ancestry
Convert context to use ancestor function
make memoized parents function
Convert revlog to use ancestor function
