Fix #3149: Fix pickling of child annotations to local classes #3202
Conversation
|
That one was really tricky to get right. |
|
We could also disallow local classes directly extending a sealed trait since it kinds of defeat the purpose of sealed traits :) |
|
Actually, here are use cases. PatternMatcher.scala in Dotty is one. Also, Allan told me it was a minimized test case from some community project. |
|
Actually the test case came to me when I was playing with #3145. |
Where is this used in PatternMatcher.scala ? |
|
We fail in several PRs with the following. Is this a bug in the buld script? [info] Resolving key references (15718 settings) ... |
|
Should go away if you rebase. |
There was a problem hiding this comment.
LGTM!
It seems we can handle anonymous class using this same trick. See #3145 .
| case Apply(TypeApply(_, List(tpTree)), _) => tpTree.symbol | ||
| } | ||
| } | ||
| val children = tp.classSymbol.children |
| @@ -702,6 +702,10 @@ class TreeUnpickler(reader: TastyReader, nameAtRef: NameRef => TermName, posUnpi | |||
| if (!sym.isType) { // Only terms might have leaky aliases, see the documentation of `checkNoPrivateLeaks` | |||
| sym.info = ta.avoidPrivateLeaks(sym, tree.pos) | |||
| } | |||
| if ((sym.isClass || sym.is(CaseVal)) && sym.isLocal) | |||
There was a problem hiding this comment.
Why use isLocal instead of isInaccessibleChildOf here?
There was a problem hiding this comment.
Because we don't have a parent yet against which to test with inAccessibleChildOf
i3149.scala is an interesting test case: A sealed top-level class `Foo` has a child which is a local (i.e., term-owned) definition in a different top-level class. It is then impossible to establish a symbolic reference to the child class from `Foo`, and it is also impossible to refer to `Foo` using a path. We deal with this by avoiding pickling such child annotations and re-establishing the annotation when the body of the child class is unpickled.
It turns out we have to know a priori whether we should register a local class as a child and we need to be precise about it.
Now that we have defined in centrally, we can avoid the duplication in Space.
i3149.scala is an interesting test case: A sealed top-level class
Foohas a child which is a local (i.e., term-owned) definition in a different top-level class.
It is then impossible to establish a symbolic reference to the child class
from
Foo, and it is also impossible to refer toFoousing a path. We deal with thisby avoiding pickling such child annotations and re-establishing the annotation
when the body of the child class is unpickled.