Include top-level symbols from same file in outer ambiguity error

[lrytz]

When checking if an inherited definition is ambiguous with an outer definition, include top-level outer definitions defined in the same compilation unit.

Follow-up for #8622.

[odersky]

I am a it worried about the added complexity (both runtime and conceptual) this introduces. In the original PR there was this spec:

It is an error if an identifier x is available as an inherited member
in an inner scope and the same name x is defined in an outer scope, unless

• the inherited member (has an overloaded alternative that) coincides with
  (an overloaded alternative of) the definition x
• x is global, i.e. a member of a package.

I assume the last clause would have to be changed to this:

• x is global, i.e. a member of a package, and it is defined not in the same source as the inner scope.

I am not sure we need this additional complication however. Have we observed cases in the wild where this is a problem that should be diagnosed?

[lrytz]

I did come across such a case when deploying the 2.12 backport of this change at our customer, that's how I noticed.

The following is not ambiguous:

File 1

trait T { class X }

File 2

package p {
  class X
  class C extends T {
    def test = new X // resolves to this.X
  }
}

Changing package p to object p makes the X reference ambiguous, which I find surprising.

[odersky]

OK, I think we can change the spec as follows (and it would be good to include it as a doc comment):

It is an error if an identifier x is available as an inherited member in an inner scope
and the same name x is defined in an outer scope in the same source file, unless
the inherited member (has an overloaded alternative that) coincides with
(an overloaded alternative of) the definition x.

Then we can argue that it's a simplification, spec-wise.

[odersky]

Otherwise LGTM

[odersky]

It seems this change caused a lot of regressions in the open community build. See #17433. We should probably revert it.

[odersky]

If this is the error that regressed in 3.3.1 and succeeds in nightly it should be backported.

[lrytz]

This PR was reverted in main (#17438), but it seems that revert is not in the 3.3.1 branch (https://github.com/lampepfl/dotty/blob/release-3.3.1/compiler/src/dotty/tools/dotc/typer/Typer.scala#L437).

Follow-up PR is here: #17439 (comment)