Randomize analyzer instantiation to avoid static constructor/jit contention

[genlu]

Fixing internal bug https://devdiv.visualstudio.com/DevDiv/_workitems/edit/1880870

[CyrusNajmabadi]

Not loving that this whole algorithm was written out twice.

[333fred]

I do wonder a bit about whether this will fully fix contention (I think it may be likely that users don't have that many analyzer assemblies, and the analyzers within an assembly will share some number of helper types, so randomizing may help a bit, but still have contention in the shared helper types), but I'm willing to give it a try and see if it helps.

[genlu]

@dotnet/roslyn-compiler Could I have another compiler review please? Thanks!

[davkean]

I do wonder a bit about whether this will fully fix contention (I think it may be likely that users don't have that many analyzer assemblies, and the analyzers within an assembly will share some number of helper types, so randomizing may help a bit, but still have contention in the shared helper types), but I'm willing to give it a try and see if it helps.

I'm seeing 4 types of contention:

• Static constructors
• JITing (the same methods)
• Type loading
• Assembly loading

I'm fairly hopeful that randomization will have a pretty good impact on the top two, unsure about the last two yet.

Update: Spin waiting contention (ie CPU) was in ~60% of all solution load traces, and that doesn't factor in in sync (blocked) time.

[genlu]

@333fred Could you pls take another look? Pushed additional change to ensure deterministic (here for more details)

[genlu]

Since we are sorting here, I changed the type of analyzerTypeNames to ImmutableHashSet from ImmutableSortedSet