`@interpret(x == 2)` and `@interpret(x) == 2` lead to different execution world ages

[timholy]

https://github.com/JuliaDebug/JuliaInterpreter.jl/blob/22f186bf2fcca6ef6b5570770035011441cb3bba/test/interpret.jl#L416-L420 is broken on Julia 1.12 and can be fixed by changing the test to @test @interpret(g_gf()) == 2. As much sense as I think that change makes, I decided to try to understand why this happens and only appeared now. It turns out that the error is thrown at https://github.com/JuliaDebug/JuliaInterpreter.jl/blob/22f186bf2fcca6ef6b5570770035011441cb3bba/src/construct.jl#L753, where we try to evaluate the arguments passed to the call, and for the current test the outer call is ==(g_gf(), 2). Thus it tries to Core.eval(Main, :(g_gf())), and this fails for reasons of world age.

Fundamentally this seems to be testing that JuliaInterpreter doesn't respect world ages properly. Should we ditch this test?

[timholy]

Yes, if that's what's intended, that would be a fine fix. But note that the test as written doesn't actually interpret g_gf(). If that's the goal, we need the parens. (EDIT: this was a response to a comment that was deleted.)

[KristofferC]

(I deleted my earlier reply because I posted it too early, sorry).

[KristofferC]

Maybe the test can be changed to

q = 0

function h()
    eval(:(q = 2))
    return q
end

@test @interpret(h() == 2)

?

[KristofferC]

I think @aviatesk has said that this whole optimization of resolving values early is invalid anyway though..

[timholy]

Yes, it doesn't seem to respect world ages. That may be more important with the bindings work in 1.12. Unfortunately, getting this fixed is probably not a small change, and the performance cost will likely be significant.

[timholy]

In an ideal world what we'd have is a whole invalidation framework for JuliaInterpreter now: resolve the bindings in the right world age, but when the world age updates you have to check for invalidations. Ugh.

[timholy]

So the easier thing is to boot out the optimization and get things at the time you need them (in the right world age).

[timholy]

I think @aviatesk has said that this whole optimization of resolving values early is invalid anyway though..

It's also been deactivated (which is partly why so much is broken)

https://github.com/JuliaDebug/JuliaInterpreter.jl/blob/822632d6f77323d338cf1e6765d1d6d5e10421c0/src/optimize.jl#L43