[SUPERSEDED] Inherited members have same binding level as package level [ci: last-only]

[som-snytt]

As with names introduced by package statements, base classes
can introduce definitions that shadow local definitions and
import statements.

Let inherited members have the same name-binding precedence
as package-level definitions: highest when defined in the current
compilation unit, and lowest otherwise. This means that a base
class defined elsewhere cannot introduce a name that shadows a
local definition.

The inherited name can be referred to unambiguously using this;
a conflicting local variable or parameter must be renamed or
aliased; a conflicting member using a C.this prefix.

Fixes scala/bug#11921

[dwijnand]

(Can't [ci: last-only] post-creation, sadly. Might actually change that, it's so annoying.)

[odersky]

The problem is that this way inherited members also cause ambiguities local imports which is too strict, IMO.

[som-snytt]

@dwijnand I have syntax fixes to come, I thought it had worked previously? I'll find out.

@odersky I liked the simplicity of the rule, but it wreaks havoc in the cake.

[som-snytt]

Sample confusing thing:

trait Aliases { _: Context => def TypeTag[T](tpe: Type) = universe.TypeTag[T](...) }
trait ExprUtils { _: Context => import universe._ ; def f = TypeTag[String](...) }

which TypeTag is intended? I have no way of seeing if it's provided by the self-type or the import, and in fact I would assume the import.

Here Aliases is a misnomer, because the signatures differ. Its TypeTag method adapts universe.TypeTag. Context is also a misnomer, because it means here Everything, or everything supplemental to universe: Global.

For purposes of naming, it's not obvious that a conflicting definition somewhere in Context is different from a conflicting definition somewhere in an enclosing package. The package-scoped definition is at least scoped: it suggests that all clients in this package use this definition. But the mixed-in definition could come from anywhere. (Arguably, "In Scala 3 we don't compose like that any more.")

The proposed precedence rule makes it harder to write a "grab-bag" import; Java calls it "on-demand".

[som-snytt]

@dwijnand it seems to respect last-only, whew. I wasn't prepared to squash yet.

Some people sit at home sewing surgical masks during quarantine.

The experience here was similar to updating the implementation previously to spec: the troublesome code was hard to read (scaladoc packagings).

I thought the cross-talk with Any members and wildcard import would doom this approach, but now I'm not sure. Today's scaladoc example is inherited def toString(link: LinkTo) subverted by wildcard import of toString().

"Root" imports don't import Any members; possibly importing without rename could be disallowed, but that seems draconian, like locking down entire nations in a pandemic.

➜  scala git:(issue/11921) skala
Welcome to Scala 2.13.2-20200329-011423-83a8963 (OpenJDK 64-Bit Server VM, Java 11.0.3).
Type in expressions for evaluation. Or try :help.

scala> object X ; locally { import X._ ; toString() }
object X
val res0: String = X$@7e7e962d

scala>
:quit
➜  scala git:(issue/11921) scala
Welcome to Scala 2.13.1 (OpenJDK 64-Bit Server VM, Java 11.0.3).
Type in expressions for evaluation. Or try :help.

scala> object X ; locally { import X._ ; toString() }
                                         ^
       error: reference to toString is ambiguous;
       it is both defined in class Object and imported subsequently by
       import X._

scala> :quit

Anyway, this implementation isn't quite correct; I'll take a second look to see if it's worth a third look.

[dwijnand]

@dwijnand it seems to respect last-only, whew. I wasn't prepared to squash yet.

Yeah. I guess it's per-push. I think this is right: if you open a PR with 10 commits and then edit the title, then it will still consider all 10 commits combined; but if you push again then it switches to last only.

[som-snytt]

I'll resubmit this for the next pandemic.

[som-snytt]

It's still lockdown.

[som-snytt]

Closing for pandemic.

[som-snytt]

The superseding linked PR attempts to follow Scala 3 semantics.