New restrictions on generated functions in 1.12?

[KristofferC]

Consider the following generated function:

@generated function constructorof(::Type{T}) where T
    getfield(parentmodule(T), nameof(T))
end

This should (at least from what I can see) be an ok generated function considering the current restrictions the docs places on them. It only calls functions defined before the function and it only reads global constants. However, it currently gives a warning

julia> struct Foo end

julia> constructorof(Foo)
WARNING: Detected access to binding `Main.Foo` in a world prior to its definition world.
  Julia 1.12 has introduced more strict world age semantics for global bindings.
  !!! This code may malfunction under Revise.
  !!! This code will error in future versions of Julia.
Hint: Add an appropriate `invokelatest` around the access to this binding.
Foo

So it seems that there is an additional restriction to generated functions now in that it is only allowed to read a global constant defined before the generated function. If that is the case, the documentation should be updated.

A lot of code in e.g. ConstructionBase uses this pattern:

https://github.com/JuliaObjects/ConstructionBase.jl/blob/72bb19dad11695fd94de66c738db534f89b23878/src/ConstructionBase.jl#L29-L31

https://github.com/JuliaObjects/ConstructionBase.jl/blob/72bb19dad11695fd94de66c738db534f89b23878/src/functions.jl#L11

[Keno]

This was undefined behavior before, the compiler just didn't have the ability to check. E.g. something like:

@generated function foo()
       try; x; return :(1.0) catch; return 1; end
       end

would give you type confusion and segfaults. In fact, with the new 1.12 semantics, this is no longer UB (so it actually removes a restriction on generated functions), you just always get the value (or error of x) at definition time. That said, we do still have the backdate code of course, which does make this UB again, but in the fullness of time, this will no longer be UB.

[Keno]

In addition, 1.12 gives you the ability to actually write a legal version of that function that does what the authors want, although you do have to manually declare your world bounds and edges, which is a bit tricky and there's likely simpler versions to achieve what that package wants, but at least it's possible, which it wasn't before.

[vtjnash]

It may still be likely UB, since they are using the generated function to bypass the otherwise detected incorrectness of doing egal structural queries on a merely type-equal operator. I think they want typename(T).wrapper here instead, which is still strongly discouraged as a coding antipattern, but at least not unsound like their current version

[Keno]

Sure, the particular code in the issue may still be UB due to that issue. I was referring specifically to the access to a possibly undefined constant

[vtjnash]

Working on a fix for the package here: JuliaObjects/ConstructionBase.jl#99