Inlining dependent on @nospecialize?

[KristofferC]

In #29133 it was noted that a @nospecialize annotation on iterate for tuples lead to a performance regression in the following (badly named becuase it isn't really testing the perf of setindex! at all) benchmark:

perf_setindex!(A, val, inds) = setindex!(A, val, inds...)

using BenchmarkTools
s = 2
A = rand(Float64, ntuple(one, s)...)
y = one(eltype(A))
i = length(A)
@btime perf_setindex!($(fill!(A, y)), $y, $i)

According to @vtjnash:

This is strange—inference isn’t supposed to look at this flag. It would be helpful to understand what inlining decisions got made and why, as this might be indicating a deeper problem.

Opening this to track.

[vtjnash]

dup #29114