Illegal cyclic reference regression in 2.13

[ybasket]

In typelevel/squants#344 we face the following problem when compiling with 2.13 (2.10-2.12 work fine) - blocking squants to be released for 2.13:

[error] squants/shared/src/main/scala/squants/motion/Acceleration.scala:25:25: illegal cyclic reference involving class Velocity
[error]     with TimeDerivative[Velocity]

Some work fixing such issues was already done in #11640, but even the integration builds of the compiler containing the fix (I used now 2.13.1-bin-d602ff4) still fail to compile squants. If I add -Ybreak-cycles -Yrecursion 2147483647 compiler flags the error message changes a bit:

[error] squants/shared/src/main/scala/squants/motion/Velocity.scala:24:13: self constructor arguments cannot reference unconstructed `this`
[error]     extends Quantity[Velocity]

Extracting a minimal example has not worked so far, even when extracting larger parts of the class hierarchy, so I can't provide more than a branch of squants to test for now: https://github.com/ybasket/squants/tree/scala-2.13

One possibility to investigate might be the alternative solution to #11640 proposed by @adriaanm in #11640 (comment).

Let me know if any more details are needed or there are things to be tried.

[ybasket]

I was able to find a work around for squants and by that got more insights on this bug. It seems to be related to the wildcard import of type aliases defined in the squants package object which shadow the fact that the classes referencing each other as type arguments to parent classes are in the same package. Apparently scalac gives up on resolving this cycle through the type aliases though the cycles are actually compilable when defined without aliases (or as-is on 2.11/2.12).

I pushed a minimal working example to https://github.com/ybasket/squants-compiler-bug-example – happy to help if I can.

[SethTisue]

It seems to be related to the wildcard import of type aliases

I wonder if scala/scala#6589 (see also #11593) is an ingredient in this mix. Does it seem relevant to you? (I haven't looked at your example code, apologies in advance if the question is off-base.)

[ybasket]

I wonder if scala/scala#6589 (see also #11593) is an ingredient in this mix. Does it seem relevant to you? (I haven't looked at your example code, apologies in advance if the question is off-base.)

Seems quite relevant.

So it might even be a desired breaking change in the Scala compiler causing it. I wonder whether in this case the cycle resolving logic should be updated to also work "through" type aliases. But my knowledge of the spec and the possibilities to implement such behaviour ends here, so I have to leave this discussion to more informed people.

[Facsimiler]

I'm not sure whether it's related to the OP's issue, but I get this same error in a very simple case (that didn't cause an error in Scala 2.12) using Scala 2.13.1:

class X {
  // ...
  class Y extends Ordered[X#Y] { // <- Illegal cyclic reference involving class Y
    // ...
  }
}

I can't see any reason why this should be illegal code, but I'm certainly willing to hear an explanation...

[SethTisue]

fwiw, Dotty accepts @Facsimiler's code:

scala> class X { class Y extends Ordered[X#Y] { def compare(that: X#Y) = ??? } }
// defined class X

[markusa380]

I'm experiencing this issue too, restructuring the packages solved it for me.