Optimize Commissioner callbacks

[marcandre]

Callback building was the single most time consuming method in all RuboCop, at around 8% of overall processing.

This is because we currently build a new Commissionner for every file (could be optimized but not easy as config can change) and a commissioner must know which cops need responding to which node types.

Previous method was thus O(files * cops * types) where types is the number of node types encountered in a typical file.
Testing on files == 200 typical RuboCop files, I calculated types ~22.5. Currently cops == 422.

This new algorithm is the same O but types is instead the number of types that a cop responds to (on average). For the cops that we run that is ~1.8.

The optimized cache building is almost 6x faster, for an overall gain of ~6.6%

It also has the advantage that adding new callbacks has zero impact on the cache building. Note: I'd like to add on_after_send etc (see #8844).

The only downside is that the list of callbacks is (by default) cached per Cop class.
This means that any Cop that somehow relies on adding callbacks it responds to at runtime is incompatible.

There's a single cop that does this (see #8881). Solutions include: not doing that (as in the PR), or only modifying the callbacks (in this case it could have been done by adding empty methods on_send, etc., even though they would be overriden by the extending modules), or overriding Cop::Base#callbacks_needed (although I marked the api as private for now).

How I tested performance:

$ stackprof tmp/stackprof.dump --text --method 'RuboCop::Cop::Commissioner#cops_callbacks_for'
  samples:  1039 self (7.5%)  /   1104 total (8.0%)

$ stackprof tmp/stackprof.dump --text --method 'RuboCop::Cop::Commissioner#initialize_callbacks'
  samples:    25 self (0.2%)  /    180 total (1.4%)

[bbatsov]

Great work!

There's a single cop that does this (see #8881). Solutions include: not doing that (as in the PR), or only modifying the callbacks (in this case it could have been done by adding empty methods on_send, etc., even though they would be overriden by the extending modules), or overriding Cop::Base#callbacks_needed (although I marked the api as private for now).

I think it's fine to just assume cops shouldn't do this. As you've mentioned yourself - only cop does this, so no need to overthink it.

[marcandre]

I should probably wait until after we make the next bugfix release to merge this

[bbatsov]

Agreed.

[bbatsov]

You can update the changelog and merge this.

[marcandre]

@bbatsov talking about ChangeLog conflicts, I've been through all stages of adding changelog entries, merging them (with an existing entry) and release in rubocop-ast... May I copy that over here too or you would rather wait?

[bbatsov]

May I copy that over here too or you would rather wait?

You can copy it over if it seems robust to you at this point. Just don't forget to update the contribution docs as well, otherwise. Maybe we should also add some note to the top of the changelog that it's preferrable to not update it directly.