fix world age errors for AbstractBindingOverlay on v1.12
#66
Conversation
Starting from v1.12.1, the implementation of `isconst/getglobal` became more accurate, and when called in a generator's world that was defined before the target method table, it now correctly determines that the method table is not yet defined. So, from v1.12 onwards, it seems necessary to use functions like `[getglobal|isconst]_at_world` as implemented here. This might be something that should be provided by Base itself though.
|
This should look at the world validity bounds of the binding and merge those into the world bounds of the returned CodeInfo. |
|
What do you think on a3e583d? |
|
Looks reasonable, though as a general comment, it looks like we may be missing the world bounds constraint and the forward edge for the method we're looking up? We fixed a lot of the correctness around this in 1.12, so very possible it just wasn't feasibly to write this correctly before. |
|
Are you saying we should also include the generator's own world age itself? I think I still don't understand which world age we're missing. |
|
We're missing the worldage of the method table query of the signature we're overlaying and the corresponding forward edge (but that can be a separate PR). |
|
Ah, I see. I can work on that in a separate PR. |
This implements feedback from the GitHub issue comment at #66 (comment). Specifically, we now use the world age from method matches by directly utilizing `Base.Compiler.findall`. However, the updated tests pass even without this change, which suggests there may be issues with the test coverage.
Starting from v1.12.1, the implementation of
isconst/getglobalbecame more accurate, and when called in a generator's world that was defined before the target method table, it now correctly determines that the method table is not yet defined.So, from v1.12 onwards, it seems necessary to use functions like
[getglobal|isconst]_at_worldas implemented here. This might be something that should be provided by Base itself though.