regression in inference performance caused by CodeInstance refactor

[vtjnash]

It looks like #53219 causes some fairly extreme performance issues in inference (up to 50x longer inference times), though curiously also sometimes provides up to a 5x speed up

Originally posted by @vtjnash in #53219 (comment)

[oscardssmith]

Do we have a MWE for the regression?

[vtjnash]

Yes, sorry, I should have pointed out that this is reported by BaseBenchmarks for this commit: https://github.com/JuliaCI/NanosoldierReports/blob/master/benchmark/by_hash/1c25d93_vs_c0a93f8/report.md

[Zentrik]

FWIW, it seems like a lot of the regression has been fixed, but there's still a fairly large regression in the abstract interpretation benchmarks (in the daily benchmarks at least):

Summary

	Range	Mean	Count
Regressions	1.97%, 855.32%	119.18%	12
Improvements	-	0.00%	0
All	1.97%, 855.32%	119.18%	12

Abstract Interpretation Benchmarks

Benchmark	% Change
inference.abstract interpretation.Base.init_stdio(::Ptr{Cvoid})	855.32%
inference.abstract interpretation.REPL.REPLCompletions.completions	226.32%
inference.abstract interpretation.broadcasting	103.55%
inference.abstract interpretation.println(::QuoteNode)	65.12%
inference.abstract interpretation.rand(Float64)	49.64%
inference.abstract interpretation.sin(42)	45.25%
inference.abstract interpretation.many_invoke_calls	31.07%
inference.abstract interpretation.many_method_matches	21.72%
inference.abstract interpretation.many_global_refs	20.46%
inference.abstract interpretation.many_opaque_closures	7.69%
inference.abstract interpretation.many_const_calls	2.08%
inference.abstract interpretation.many_local_vars	1.97%

Per http://tealquaternion.camdvr.org/compare.html?stat=min-wall-time&name=inference.abst&nonRelevant=true

[Zentrik]

I bisected a 6x regression in the min run time of BaseBenchmarks.SUITE[["inference", "allinference", "Base.init_stdio(::Ptr{Cvoid})"]] to ea1a0d2, i.e. #53326.

I bisected a 5x regression in the min run time of BaseBenchmarks.SUITE[["inference", "abstract interpretation", "broadcasting"]] to 144f58b, i.e. #53580

[vtjnash]

Okay, the init_stdio regression is probably fine then, since we just significantly increased the amount of code visible to the compiler, but didn't change the compiler.