Router: extract rules to Splits, use class lookup

[kitbellew]

A profiler run revealed that getSplits is the largest bottleneck of performance, despite being run only once, almost double that of the part which was previously thought to be the more expensive: BestFirstSearch.

Let's refactor away from using a massive match to several lookups by the token class.

In community-tests runs, there's a 35% reduction in running times (22 minutes to 14, 15 to 10).

[kitbellew]

this is a tricky change :) i tested on scala2+scala3+spark community repos that community-tests runs on, and it produces the same results, but conceivably could be different. if you could check on your repos, would appreciate it.

[tgodzik]

this is a tricky change :) i tested on scala2+scala3+spark community repos that community-tests runs on, and it produces the same results, but conceivably could be different. if you could check on your repos, would appreciate it.

I checked Metals and Scala CLI repos and the results seem the same, but I only changed version in .scalafmt.conf and run using the jar based runner.

[tgodzik]

LGTM though I don't really have time to look at the full changes.

[kitbellew]

LGTM though I don't really have time to look at the full changes.

Yes, alas, this change is not easily reviewable. The code is basically the same but in different places; but even with that, there were a few tricky parts:

• now there's a fixed order of rule evaluation (first we look at the left token, then the right, etc), whereas before matching could be in any order
• the location in the source file (line, to be exact) is used as a tiebreaker when comparing formatting states, and since these lines have changed, I had to add an explicit rank field.