apply_any gets specialized despite annotations to the contrary

[timholy]

After running the JuliaInterpreter tests, I get this:

julia> m = first(methods(Base.append_any))
append_any(xs...) in Base at essentials.jl:663

julia> m.specializations.func
MethodInstance for append_any(::Vararg{Any,N} where N)

julia> m.specializations.next.func
MethodInstance for append_any(::Tuple{Any}, ::Vararg{Any,N} where N)

julia> m.specializations.next.next.func
MethodInstance for append_any(::Any, ::Vararg{Any,N} where N)

This is despite the intention to avoid specialization:

julia/base/essentials.jl

Lines 662 to 663 in b073bb5

	function append_any(xs...)
	@nospecialize

Unfortunately, this isn't fixed by changing the declaration to

function append_any(@nospecialize xs...)

[JeffBezanson]

nospecialize means that separate code will not be generated for different signatures, but the function can still be inferred on different argument types, which is probably what we're seeing here. A noinfer would be more dangerous, since it would have nonlocal effects, injecting worse type info into callers that might need the performance.