performance regression in findfirst due to inlining

[JeffBezanson]

julia> a = rand(10000000); z = a[end];

julia> @time findfirst(equalto(z), a)
  0.411288 seconds (10.01 M allocations: 153.062 MiB, 36.84% gc time)

In 0.6 there are no allocations. The reason for this is that findfirst is a simple wrapper for findnext, and so it does not get specialized on its function argument (since it doesn't call the function itself). However, findnext is being inlined into it anyway. I don't think we should inline in this case (perhaps whenever the argument types are not is_cacheable_sig for the target method, and/or because the target method has a loop).

[KristofferC]

Perhaps #26418 has the same cause?

[vtjnash]

I vaguely thought we already had an issue for not inlining a called parameter into the unspecialized wrapper, but I can't find it. There is #25116, which is related though at least.

[vtjnash]

Fixed now?

julia> @time findfirst(isequal(z), a)
  0.019936 seconds (6 allocations: 192 bytes)
10000000

[JeffBezanson]

Yes #27857 fixes this.