[PATCH STABLE V2] hgweb: cache fctx.parents() in annotate command (issue5414)

[Yuya Nishihara]

On Sun, 6 Nov 2016 17:01:05 +0000, Jun Wu wrote:
> Excerpts from Yuya Nishihara's message of 2016-11-06 11:31:04 +0900:
> > Perhaps fctx.parents() can be property-cached, but we'll need to drop
> > uninteresting chains of parents in fctx.annotate().
> 
> If we go the property-cache approach, I think it's better to cache
> "_adjustedlinkrev". It's at a lower level and covers both "parents"
> and "introrev". Caching "parents" may increase memory usage unintentionally.
> 
> I don't fully get what "uninteresting chains of parents" means here.
> In the annotate case, let's say f1, f2 = f0.parents().
> Both f1 and f2 have _descendantrev set to f0's adjusted linkrev.

As you said, what's in my mind was the memory usage. Caching fctx.parents()
would mean annotate() builds a full link from self to root nodes. Some of
these intermediate nodes aren't useful for hgweb.

> Suppose there is a global cache dict: {(path, filenode, srcrev): linkrev}, I
> think if srcrev=_descendantrev (it's true for f1, f2) and _descendantrev is
> adjusted from the direct child (f0), then it is "interesting" and can be
> cached. This is similar to what marmoute said during the sprint - for the
> log -f or annotate case, once the first fctx's introrev is known, the cache
> can be used to calculate the ancestors' adjusted linkrevs.

Given we have ugly hacks to pass ancestry data around fctx objects, a global
cache might be useful.