False reference ambiguity with redundant import and opaque type

[pikinier20]

Compiler version

Reproduced on newest nightly, 3.1.2-RC1, 3.1.2-RC2, 3.0.0

Minimized code

To reproduce this bug, we need three files:

package bar

package baz {
  import util.HTML._

  class TrashRenderer:
      raw("")

}

package bar

package foo {
  import util.HTML._
  import bar.util.HTML._

  object Bazzz:
    raw("asd").toString
}

package bar
package util {
  object HTML:
    opaque type AppliedTag = StringBuilder

    def raw(content: String): AppliedTag = new AppliedTag(content)
    def raw(content: StringBuilder): AppliedTag = content
}

I found out that there are some conditions that must be satisfied for this bug to occur:

• There must be opaque type in object HTML
• There must be nested package in the file where the raw method is used so that we can import it by absolute and relative path
• There must by second usage of raw somewhere

Output

[error] -- [E049] Reference Error: /Users/fzybala/Documents/random/ambiguous-reference-bug/src/main/scala/dotty/tools/scaladoc/tasty/comments/markdown/bug.scala:8:4 
[error] 8 |    raw("asd").toString
[error]   |    ^^^
[error]   |    Reference to raw is ambiguous,
[error]   |    it is both imported by import bar.util.HTML._
[error]   |    and imported subsequently by import bar.util.HTML._

Expectation

Should compile

This problem caused CI to fail on one of my branches in Scaladoc: #14669. Of course double-import is not a good thing but usually compiler knows that it's the same thing.

[pikinier20]

I tried to look into compiler where exactly this message is thrown and I got to TypeComparer. There's a method isSameRef that compares if two types are the same refs. When we don't satisfy one of conditions that I mentioned, the method is called with tp1 that shows: <overloaded bar.util.HTML.raw> and tp2 that shows: <overloaded bar.util.HTML.raw> and therefore it returns true. In the example that I've provided, the method gets other parameters: tp1 : <overloaded bar.util.HTML.raw> and tp2 :

(bar.util.HTML.raw : 
  (content: String): bar.util.HTML.AppliedTag
)

and therefore, the check in method:

tp1.alternatives forall (isSubRef(_, tp2))

returns false.

So, something changes the tp2 from overloaded method to one of its concrete alternatives.

[odersky]

I guess it's some strange completion order that makes the compiler not see one of the alternatives.