revset optimizer: branch() is not necessarily cheaper than outgoing()

[Greg Ward]

A small experiment with revset.  First, without touching the code:

  $ hg log -r "outgoing(<localrepo>) & branch(<branch>)" --time
  [...]
  Time: real 40.150 secs (user 37.950+0.000 sys 0.430+0.000)

Incidentally, "hg log -b <branch>" takes about 30 sec for me.

Now hack revset.py to remove the penalty for outgoing():

--- a/mercurial/revset.py
+++ b/mercurial/revset.py
@@ -545,7 +545,7 @@
         wa, ta = optimize(x[2], small)
         if f in "grep date user author keyword branch file":
             w = 10 # slow
-        elif f in "modifies adds removes outgoing":
+        elif f in "modifies adds removes":
             w = 30 # slower
         elif f == "contains":
             w = 100 # very slow

and run the same command again:

  $ hg log -r "outgoing(<localrepo>) & branch(<branch>)" --time
  Time: real 4.890 secs (user 4.580+0.000 sys 0.160+0.000)

Much better -- about what I expected in the first place.  "hg outgoing
<localrepo>" takes about 3.7 sec for me.

Thoughts:

  * for a large repo, queries that read the whole changelog (grep, date,
    user, branch, etc.) are pretty costly -- more costly than a typical
    outgoing() query that finds only tens or hundreds of changesets

  * outgoing() vs a local repo is almost certainly cheaper than outgoing()
    over ssh or http -- perhaps the optimizer should take that into account

Greg