CODEWITHSUNDEEP
scalatypesselftype-systems
soc
soc

Reputation: 28443

Why did the Scala compiler get more strict regarding self types in 2.10?

I have the following code:

  trait TFn1B {
    type In
    type Out
    type Apply[T <: In] <: Out
  }

  trait TFn1[I, O] extends TFn1B {
    type In = I
    type Out = O
  }

  trait >>[F1 <: TFn1[_, _], F2 <: TFn1[_, _]] extends TFn1[F1#In, F2#Out] {
    type Apply[T] = F2#Apply[F1#Apply[T]]
  }

It compiles on Scala 2.9.1 without a warning or an error.

But on the current 2.10 build, I get the following error message:

Fun.scala:12: error: illegal inheritance;
 self-type this.>>[F1,F2] does not conform to this.TFn1[_$1,_$4]'s selftype this.TFn1[_$1,_$4]
  trait >>[F1 <: TFn1[_, _], F2 <: TFn1[_, _]] extends TFn1[F1#In, F2#Out] {
                                                       ^
one error found

Is this a regression or is the code unsound and the compiler just started to catch it lately? If the code is wrong, what would be a work-around to make it work again?

Upvotes: 4

Views: 325

Answers (2)

Daniel C. Sobral
Daniel C. Sobral

Reputation: 297265

It looks like a bug to me. There's no constraints on the type parameters of the super classes, but, more to the point, the trace for -explaintypes looks suspicious to me:

scala>   trait >>[F1 <: TFn1[_, _], F2 <: TFn1[_, _]] extends TFn1[F1#In, F2#Out] {
     |     type Apply[T] = F2#Apply[F1#Apply[T]]
     |   }
<console>:9: error: illegal inheritance;
 self-type >>[F1,F2] does not conform to TFn1[_$1,_$4]'s selftype TFn1[_$1,_$4]
         trait >>[F1 <: TFn1[_, _], F2 <: TFn1[_, _]] extends TFn1[F1#In, F2#Out] {
                                                              ^
>>[F1,F2] <: TFn1[_$1,_$4]?
  TFn1[_$1,_$4] <: TFn1[_$1,_$4]?
    _$1 <: _$1?
      _$1 <: Nothing?
        <notype> <: Nothing?
        false
        Any <: Nothing?
          <notype> <: Nothing?
          false
        false
      false
      Any <: _$1?
        Any <: Nothing?
          <notype> <: Nothing?
          false
        false
      false
    false
  false
false

Specifically, I don't understand how or why it cannot prove that TFn1[_$1,_$4] <: TFn1[_$1,_$4] or even _$1 <: _$1.

Upvotes: 2

VonC
VonC

Reputation: 1327224

As seen on the Scala mailing list (posted by you), Paul Phillips answers that:

It's already open (SI-5399, SI-5400).
The regression arises from attempts to address SI-5120: rather spectacular soundness failure with abstract types.

Potential problematic commits:

Upvotes: 3

Related Questions