Fix #5077: avoid pattern-bound type for selectors

[liufengyun]

Fix #5077: avoid pattern-bound type for selectors

Given the following definition:

trait Is[A]
case object IsInt extends Is[Int]
case object IsString extends Is[String]
case class C[A](is: Is[A], value: A)

and the pattern match:

(x: C[_]) match
case C(IsInt, i) => ...

The typer (enhanced with GADT) will infer C[A$1] as the type of the
pattern, where A$1 =:= Int.

The patternMatcher generates the following code:

case val x15: C[A$1] =
  C.unapply[A$1 @ A$1](x9.$asInstanceOf$[C[A$1]])
case val x16: Is[A$1] = x15._1
case val x17: A$1 = x15._2        // erase to `Int`
if IsInt.==(x16) then
  {
    case val i: A$1 = x17
...
  }

Note that x17 will have the erased type Int. This is incorrect: it
may only assume the type Int if the test is true inside the block.
If the test is false, we will get an type cast exception at runtime.

To fix the problem, we replace pattern-bound types by Any for
selector results:

case val x15: C[A$1] =
  C.unapply[A$1 @ A$1](x9.$asInstanceOf$[C[A$1]])
case val x16: Is[A$1] = x15._1
case val x17: Any = x15._2
if IsInt.==(x16) then
  {
    case val i: A$1 = x17.$asInstanceOf$[A$1]
    ...
  }

The patternMatcher will then use a type cast to pass the selector
value for nested unapplys or assign it to bound variables.

[dwijnand]

Isn't this wrong because == is user-definable?

[liufengyun]

Isn't this wrong because == is user-definable?

Do you mean the first pattern match? Yes, it should be allowed. #5077 (comment) . However, I think it's also fair for the compiler to be incomplete and reject the code. See #9740.

[odersky]

Otherwise LGTM