use `java.lang.ClassValue` instead of `IO.tag`

[xuwei-k]

trafficstars

I think java.lang.ClassValue is faster than current implementation.

• benchmark code https://github.com/xuwei-k/cats-effect/commit/f7c2dabfd974bcfb57cff683a360ef7c163a4487
• https://github.com/scala/scala/pull/9632

Jmh / run -i 10 -wi 10 -f 1 -t 1 cats.effect.benchmarks.TagBenchmark

[info] Benchmark                 Mode  Cnt         Score        Error  Units
[info] TagBenchmark.classValue  thrpt   10  10348652.195 ± 116363.873  ops/s
[info] TagBenchmark.tag         thrpt   10   7336411.108 ±  91988.686  ops/s

[djspiewak]

Oh this is incredible and I didn't know about it. What this allows us to do is eliminate a megamorphic virtual dispatch. The replacement looks to be a tableswitch. Without digging into the JIT output, I can only guess about how it gets inlined, but clearly this gives the runtime a lot more information to work with. Very nice.

[djspiewak]

The linked scala/scala PR is really interesting. I did a similar set of benchmarking (though more specialized to IOFiber and obviously not including ClassValue) and found that virtual dispatch fell behind tableswitch more or less across the board. I suspect there's a subtle interaction with the inliner here.