scala-reflect produces inconsistent results for type boundaries obtained through etaExpand on scala 2.13.0

[Paul S.]

Let's consider this example (you may just copypaste it into scala shell):

import scala.language.experimental.macros
import scala.reflect.macros.blackbox
import scala.reflect.runtime.universe._

trait LightTypeTag { /*TODO*/ }

def makeTag[T: c.WeakTypeTag](c: blackbox.Context): c.Expr[LightTypeTag] = {
  import c.universe._
  val tpe = implicitly[WeakTypeTag[T]].tpe
  println((tpe, tpe.hashCode, System.identityHashCode(tpe)))

  def test() = {
  tpe.etaExpand.resultType.dealias.typeArgs.map(_.dealias.resultType.typeSymbol.typeSignature).map {
    case tpe: TypeBoundsApi =>
      tpe.hi
  }.foreach {
    tpe =>
    println((tpe, tpe.hashCode, System.identityHashCode(tpe)))
  }
  }
  test()
  test()
  test()
  c.Expr[LightTypeTag](q"null")
}

def materialize1[T[_]]: LightTypeTag = macro makeTag[T[Nothing]]

materialize1[Enum]

On scala < 2.13 it prints

(Enum,-2104744931,1129119773)
(Enum[E],2055684080,910505843)
(Enum[E],2055684080,910505843)
(Enum[E],2055684080,910505843)

Which is expected.

On scala 2.13 it prints

(Enum,847238925,1831829645)
(Enum[E],-1086966789,342809115)
(Enum[E],-37304610,678306695)
(Enum[E],-500995075,606816495)

So, each invocation returns it's own, unique type boundary. And =:= check fails.

It works without etaExpand though.

[Seth Tisue]

there is some discussion at https://gitter.im/scala/contributors?at=5d4dde99d03a7c63e6fc4071 on whether this qualifies as a regression, or whether it's simply the kind of compiler-internals thing that might be expected to change in a minor release, let alone a major release

[Seth Tisue]

the bisect script from adriaanm/binfu#5 should be useful for identifying the scala/scala PR where this behavior changed, which I think would be the single most important piece of information to have in hand, in order to discuss whether this is a regression or simply a behavior change

[Paul S.]

Some time ago it (magically) started working again.