@nospecialize for identity

[Tokazama]

No description provided.

[aviatesk]

We generally don't need to add @nospecialize annotation if a generic function only contains a single method (and identity will rarely be extended), but approved nevertheless as it doesn't hurt anything anyway.

[Tokazama]

I know there's a point where specialization doesn't occur, but I figure that we never wanted it to specialize and that even this was more than we wanted:

julia> using MethodAnalysis

julia> using StaticArrays

julia> length(methodinstances(identity))
18

julia> identity(SVector(1, 2, 3));

julia> length(methodinstances(identity))
19

julia> foo(x) = identity(x)
foo (generic function with 1 method)

julia> foo(SVector(1, 2, 3, 4));

julia> length(methodinstances(identity))
20

[aviatesk]

Please note that @nospecialize influences dispatch only, does not intervene into inference in anyway. See #41931 for more explanation.

[Tokazama]

The goal here was just to avoid codegen. I'm assuming that if this made it so we couldn't even infer what type identity returned that it would be a bad idea.

[aviatesk]

"codegen" would be avoided even as is assuming the method body of identity is inlined always.

[Tokazama]

I was considering additional method instances code generation (which I can see being considered an overgeneralization). Wouldn't we want to avoid generating those addition methods?

[aviatesk]

It's simply that that information was used for Julia-level inference but it doesn't necessarily imply whole codegen process.

[Tokazama]

Gotcha. Thanks for the clarification.