partialdiscovery: avoid `undecided` related computation sooner than necessary Changeset 1d30be90c move the update of the `undecided` set within the `partialdiscovery` object in order to clarify the API. The update to the `undecided` set was unconditional in 1d30be90c and the first access to the `self.undecided` property triggered the initial computation of the set of undecided revisions. As a result, the set was computed much earlier, at a time where less information is available, immediately followed by an update of this set to remove common revisions. To fix this regression, we ignore the `undecided` related logic in `addcommons` when that `undecided` set has not been computed yet. Code that actually needs to know the `undecided` set will trigger its computation later. The change has no effects on semantic because the initial computation `undecided` set takes all knowns `common` into account. Example performance running `hg debugdiscovery` from a pypy repo missing 10 changesets: 870a89c6909d: 52.3ms (regression parent) 1d30be90c9dc: 72.0ms (regression) 5a5f504a7175: 64.8ms (this fix parent) this fix: 52.6ms

--- a/mercurial/setdiscovery.py	Mon Jan 21 22:14:29 2019 +0900
+++ b/mercurial/setdiscovery.py	Wed Jan 23 18:07:42 2019 -0500
@@ -183,7 +183,8 @@
     def addcommons(self, commons):
         """registrer nodes known as common"""
         self._common.addbases(commons)
-        self._common.removeancestorsfrom(self.undecided)
+        if self._undecided is not None:
+            self._common.removeancestorsfrom(self._undecided)
 
     def addmissings(self, missings):
         """registrer some nodes as missing"""