GADT unsoundness: We can infer sufficient constraints when a union type is on the lhs of a subtype check

[smarter]

This compiles, but shouldn't:

sealed trait Expr[+T]
final case class FooExpr() extends Expr[1 | 2]

object Test {
  def foo[T](x: Expr[T]): T = x match {
    case x: FooExpr =>
      3
  }

  val x: 1 | 2 = foo(FooExpr())
}

I think the problem is that when a union type is on the lhs, we can infer sufficient constraints regardless of whether we're in GadtConstraintInference mode or not, for example:
https://github.com/lampepfl/dotty/blob/b5a171536a8ce4294c2107770f47cb97dfc148f5/compiler/src/dotty/tools/dotc/core/TypeComparer.scala#L450
If I rewrite this to:

          isSubType(tp1.widenSingletons, tp2, approx.addLow)

Then the example above doesn't compile anymore:

-- [E007] Type Mismatch Error: try/gadtx.scala:7:6 -----------------------------
7 |      3
  |      ^
  |  Found:    (3 : Int)
  |  Required: T
  |
  |  where:    T is a type in method foo with bounds >: (1 : Int) | (2 : Int)

But I haven't checked what other consequences this would have or if there are other cases that would need to be changed like this too.

/cc @odersky

[abgruszecki]

Ah, it seems pretty clear that we need to add approx in that case, given the definition of widenSingletons (not to be confused with widenSingleton!). I can have a look at other calls to that function in TypeComparer.

[smarter]

It's not just widenSingletons I think, for example there's recur(tp1.join, tp2) in joinOK