specialize on type in getindex for types

[KristofferC]

Fixes #18021. From 3.65 -> 0.25 seconds in the benchmark code in the linked issue.

Ref #17798 for similar correction

Fix found by eavesdropping on a discussion by @vtjnash and @yuyichao :)

[KristofferC]

Not sure we have benchmarks for this but @nanosoldier runbenchmarks(ALL, vs = ":master")

[yuyichao]

Just discuss about something and someone else implements the fix. I hope fixing issues is always as easy as this. =)

Also ref for a similar change #18009

[KristofferC]

It might be worth doing a search through all the .jl files for ::Type,and examine which ones are worth specializing on.

[andreasnoack]

This now seems to be the general 0.5 performance trick. Has the specialization heuristics generally changed since 0.4? As I read it, this method was basically the same in 0.4.

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @jrevels

[yuyichao]

The inline heuristic might have changed.

[yuyichao]

There also seems to be some type inference/codegen issue related to Expr(:invoke), the llvm ir and asm of v32 has jl_invoke in it but the jl_invoke disappears from llvm ir after calling v32. In this particular case, jl_invoke doesn't hurt performance too much but it still cause generation of a large GC frame and boxing of all parameters passed to getindex.

[KristofferC]

Is there an issue for that @yuyichao?

[vtjnash]

This now seems to be the general 0.5 performance trick. Has the specialization heuristics generally changed since 0.4? As I read it, this method was basically the same in 0.4.

No, this fix didn't address what is broken or why, but it just addresses one of the symptoms. The actual problem is that the fast path in apply_type hasn't been optimized and is still much too slow.