`--depwarn=error` silences the "world prior to its definition world" warning, instead of making it an error

[nsajko]

If I run Julia master like ./julia, I get this:

julia> using LinearSolve
Info Given LinearSolve was explicitly requested, output will be shown live 
WARNING: Detected access to binding `LinearSolve.defaultalg_symbol` 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.
To make this warning an error, and hence obtain a stack trace, use `julia --depwarn=error`.
Precompiling LinearSolve finished.
  64 dependencies successfully precompiled in 59 seconds. 45 already precompiled.
  1 dependency had output during precompilation:
┌ LinearSolve
│  [Output was shown above]
└

After deleting .julia/compiled/v1.13, running it like ./julia --depwarn=error, I get no warning or error during precompilation:

julia> using LinearSolve
Precompiling LinearSolve finished.
  77 dependencies successfully precompiled in 88 seconds. 32 already precompiled.

[Keno]

Possibly there was a try/catch around it?

[nsajko]

There doesn't seem to be any try in the LinearSolve.jl repo:

• https://github.com/search?q=repo%3ASciML%2FLinearSolve.jl%20path%3A%2F%5Esrc%5C%2F%2F%20try&type=code

[ViralBShah]

I just saw this in 1.12.0-beta4.

[Keno]

The reason it doesn't show up with --depwarn=error is that this is inside the generator of a generated function (which has an implicit try/catch around it) called from inference. There's also a bug with this error making it only fire once. This is a bit unfortunate. The way that error handling in generated functions is supposed to work is that they're supposed to be suppressed during inference and then if you actually reach it you get the error. I think the most sensible thing (other than fixing the bug) is to always force the UndefVarError when called from generators, even without --depwarn=error. That's marginally more breaking, but I think preferable to the whackamole of inference-triggered depwarns.

[vtjnash]

Sounds good. Accessing any mutable state from there is an effects violation, so we it was already UB to access any non-const globals

[Keno]

(#58651)