./mill selective.prepare __ on the Mill codebase takes >60s

[lihaoyi]

It's only ~80,000 tasks, and ./mill resolve __ returns instantly so there is no reason it should take particularly long. Not sure if it's upstream non-input tasks taking time to evaluate (#4121), or just algorithmic slowness in the computation

[lihaoyi]

Seems like the slowness is coming from the ./mill plan __ logic that selective.prepare uses to find the transitive upstream codebase. And a lot of the slowness is in ResolveCore.instantiateModule

[lihaoyi]

Perhaps the reason resolve __ is fast while plan __ is slow is that resolve __ doesn't actually instantiate all the modules and tasks; it looks up the module and task names via reflection on the JVM metadata, and only instantiates the few modules necessary to determine Cross keys.