improve (restore) performance of fieldname by moving out error paths in separate functions

[KristofferC]

This makes it inline again (maybe new inference change to not infer error paths made it not inline anymore) and also improves performance relative 1.5.

Benchmark:

julia> using BenchmarkTools

julia> struct A
              a::Int
              b::Float64
              c::String
              end

julia> function ff(::Type{T}, i) where {T}
              return fieldname(T, i)
              end;

julia> @btime ff(A, 2)

Before (master): 245.346 ns
Before (1.5.2): 4.747 ns
PR: 2.727 ns

Fixes #38920

cc @quinnj

[timholy]

On my slightly-outdated build (avoiding the JuliaInterpreter breakage...), Base.print_statement_costs(stdout, fieldname, (DataType, Int)) suggests the reason it didn't inline is because _fieldnames is not inferrable and hence length(names) cannot have its dispatch resolved at compile time. This doesn't fix that, which leaves me puzzled about why this solves the inlining problem.

But certainly getting the string interpolation out into separate calls is a good idea.

[KristofferC]

Base.print_statement_costs(stdout, fieldname, (DataType, Int))

Doesn't it specialize on the type passed in? So what you get from that are the costs for the abstract DataType but not the actual type (A in this case).

[Keno]

Can we get a nanosoldier test for this?