directaccess: use cached filteredrevs Before this patch we were calling directly repoview.computehidden(repo) to compute the revisions visible with direct access, without going through the caching mechanism for the filtered revisions. There was two issues with that: (1) Performance: We were not leverating the cached values of the 'visible' revs (2) Stability: If there were to be a cache inconsistency with the computation of 'visible' we would crash in the branchmap consistency check partial.validfor. Consider the scenario of rebase with bookmarks: - when we delete a bookmark on an obsolete changeset (like what rebase does when moving the bookmark after rebasing the changesets) - then this changes the value returned by repoview.computehidden(repo) as bookmarks are used as dynamic blockers in repoview.computehidden(repo) - as of now, we don't invalidate the cache in the case of bookmark change - if we have a cached value from before the bookmark change, repoview.filterrevs(repo, 'visible') considers the cached value correct and returns something different than repoview.computehidden(repo) - in turn, if we use repoview.computehidden(repo) in directaccess, the subset relationship is broken and the cache consistency assertion (parial.validfor) fails if branchmap.updatecache is called in this time window This patch leverages the caching infrastructure in place to speed up the computation of the filteredrevs for visible-directaccess-nowarn and visible-directaccess-warn. Incidentally it prevents the bug discussed in (2) from crashing when running a rebase with a bookmark. Note that there still needs to be a fix in core for the case discussed in (2). The test for this side of the fix (not core's fix for (2) is very hard to implement without introducing a lot of dependencies and does not belong here. It is much easier to have the test of the fix for the scenario (2) in core along with the fix.

inhibit: improve performance of transaction wrapping Before this patch, transaction wrapping was the most expensive part of inhibit computation wise. This patch changes the revset that we use in the transaction wrapping to make it ~50x faster to compute: revset #0: obsolete() - hidden() 0) wall 0.000214 comb 0.000000 user 0.000000 sys 0.000000 (best of 11209) vs revset #0: (not hidden()) and obsolete() 0) wall 0.010965 comb 0.010000 user 0.010000 sys 0.000000 (best of 237)