simplify: Drop mapping tparams to tvars #18786
Conversation
Mapping type params back to type vars is dangerous for match types because it can lead to one use-site's type var being cached as the reduction of the match type in a match alias defintion. It looks like mapping type parameters to type vars is only important when getting the instance type of a type parameter, because the bounds could make references to other parameters, including as type arguments. So, as long as we map then, we can drop this mapping from general simplification. There is, however, something going on with capture checking, which I documented, but perhaps it's due to some underlying bug. The situation occurs because boxed A *: T does seem to match case x *: _. So I kept the fully-defined forcing and normalization case from simplify.
|
3.3.2-RC1 is already underway, yet this regression is left open. What concerns me most is that the compiler is undergoing drastic changes which could cause more regressions, but we have no way of knowing before first solving the known regressions and bringing the OpenCB to a green status. |
|
I tested the issue on 3.3.2-RC1 and it seems to be resolved. Was the original change reverted or did something else resolve the bug? |
|
I assume by your comment (#18171 (comment)) that it's still reverted. |
It was a different PR that caused the regression, IIUC. |
|
In any case, I now tested under the latest 3.4 nightly |
Mapping type params back to type vars is dangerous for match types
because it can lead to one use-site's type var being cached as the
reduction of the match type in a match alias defintion.
It looks like mapping type parameters to type vars is only important
when getting the instance type of a type parameter, because the bounds
could make references to other parameters, including as type arguments.
So, as long as we map then, we can drop this mapping from general
simplification.
There is, however, something going on with capture checking, which I
documented, but perhaps it's due to some underlying bug. The situation
occurs because boxed A *: T does seem to match case x *: _. So I kept
the fully-defined forcing and normalization case from simplify.