Can't detect binary incompatibility when trait extends abstract class

[xuwei-k]

• scala version 2.12.6
• mima version 0.3.0
• complete example repository https://github.com/xuwei-k/mima-trait-extends-class

library x v1

package com.example

trait A

library x v2

package com.example

abstract class B {
  def foo: Int = 42
}

trait A extends B

mima said library x v1 => v2 binary compatible.

another library y

• depends on library x v1

package com.example

object C {
  val a = new com.example.A {}
}

main

• depends on library y
• also depends on library x v2 (override x v1 transitive dependency from y)

package com.example

object Main {
  def main(args: Array[String]): Unit = println(C.a.foo)
}

run main

[error] (run-main-0) java.lang.ClassCastException: com.example.C$$anon$1 cannot be cast to com.example.B
[error] java.lang.ClassCastException: com.example.C$$anon$1 cannot be cast to com.example.B
[error] 	at com.example.Main$.main(Main.scala:6)
[error] 	at com.example.Main.main(Main.scala)
[error] 	at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
[error] 	at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)

[xuwei-k]

FYI, similar (but maybe different) issue #118

[dwijnand]

Just re-reading this, I wonder if the fact that class B is abstract is relevant.

[dwijnand]

The fact that class B is abstract is not relevant: the reproduction fails the same and MiMa doesn't detect it the same.

The fact that it's a class and not a trait is relevant, when it's a trait MiMa doesn't report it for Scala 2.12+, which is correct because Scala 2.12+ implements it as default methods. That's what #118 was about.

[dwijnand]

Well it looks like the problem might be that in

class B
trait A extends B

A has no superclass?!

[szeiger]

scala> trait A1
defined trait A1

scala> :javap -c A1
Compiled from "<console>"
public interface $line8.$read$$iw$$iw$A1 {
}

scala> trait A2 extends B
defined trait A2

scala> :javap -c A2
Compiled from "<console>"
public interface $line5.$read$$iw$$iw$A2 {
}

This is as expected. An interface cannot extend a class. Similar to a self-type, the fact that A2 extends B is not visible to Java but unlike a self-type it is externally visible in Scala so the compiler may make use of it.

The incompatibility in the example is really the trait instantiation in C. main is a convenient way to expose it but C is already broken.

I can't think of a way of detecting this short of examining the Scala signatures, which means MiMa needs to be able to read those for all supported Scala versions.

[adriaanm]

(one of the motivations for scala/scala-dev#601)

[dwijnand]

I see. Makes sense, but it's also tragic. So public traits that don't extend Object are binary hazards. How do we help users like Yoshida-san from running into these problems?

@sjrd, as a fellow binary compatibility advocate, does that change your stance on scala/scala-dev#601?

[dwijnand]

Forgot to say: thanks Stefan for the explanation.

Tbh this has kind of blown my mind. If declare a class extends the trait does scalac gives it the traits (non) superclasses? I guess it makes type A not a subtype of B even though it extends class B. Confusion.

[adriaanm]

i don't follow entirely, but think of a trait extending a class as a downsteam obligation for any classes extending that trait (same as with a self type) -- the earliest opportunity, we'll add that trait's declared superclass as a parent for a class that mixes in the trait

[adriaanm]

so, any instance of the trait will also be an instance of its declared superclass

[dwijnand]

Yeah I get the mechanics now, but it breaks my mental model (invariant) that extends means parent and supertype. Still learning fundamentals in Scala, it's incredible/scary.

[adriaanm]

Agreed -- again, the surprise element is why I think we should get rid of it, but I guess I should argue that over at the other ticket :-)