propagate bounds from class type params to existential quantifiers

[scabug]

see attachments

In java you can write

Circle circle = CircleBuilder.create().centerX(100).centerY(115).radius(100).build();

This is a builder pattern which make use of existential f-bounds
The problem is that in scala the skolems get erased after two method calls and so the next method doesn't exist.

if I javap the java code

const #33 = NameAndType #51:#52;//  create:()Ljavafx/scene/shape/CircleBuilder;
const #34 = NameAndType #53:#54;//  centerX:(D)Ljavafx/scene/shape/CircleBuilder
;
const #35 = NameAndType #55:#54;//  centerY:(D)Ljavafx/scene/shape/CircleBuilder
;
const #36 = NameAndType #56:#54;//  radius:(D)Ljavafx/scene/shape/CircleBuilder;

const #37 = NameAndType #57:#58;//  build:()Ljavafx/scene/shape/Circle;

then I can see that the names and types are stored in bytecode (so a sort of runtime type info / refied generics)

In scala:

val circle = CircleBuilder.create().centerX(100).centerY(115).radius(100).build()

causes

C:\scala-2.10.0-M6\myexamples>scalac javafxbuilder.scala
javafxbuilder.scala:19: error: value centerY is not a member of ?0
      val circle = CircleBuilder.create().centerX(100).centerY(115).radius(100).
build()
                                                       ^
one error found

The only way to solve it in scala is to do a downcast after two methods (.asInstanceOf[CircleBuilder[_]]) or implicit conversion or to generate reified skolems _$x beforehand.

val circlebuilder = CircleBuilder.create().asInstanceOf[CircleBuilder[_ <: CircleBuilder[_ <: CircleBuilder[_ <: CircleBuilder[_]]]]]

which generates $_1, $_2 and $_3 for calling 6 methods (2 per skolem)

This looks a bit hackish to me and very unscala.

associated thread:
https://groups.google.com/forum/?hl=en&fromgroups#!topic/scala-user/Op79HcAoj2M

[scabug]

Imported From: https://issues.scala-lang.org/browse/SI-6169?orig=1
Reporter: DaveScala (davescala)
Affected Versions: 2.10.0-M6
See #8198, #1786, #4745
Duplicates #2385
Attachments:

• JavafxBuilder.java (created on Aug 2, 2012 11:16:29 AM UTC, 573 bytes)
• javafxbuilder.scala (created on Aug 2, 2012 11:16:29 AM UTC, 804 bytes)

[scabug]

DaveScala (davescala) said:
Test files

[scabug]

@retronym said (edited on Mar 30, 2013 9:29:03 PM UTC):
Actually, it's not a Java-interop problem, the equivalent Scala code also fails.

object Main {
  def foo(b: CircleBuilderScala[_]) =
    b.foo.foo  // error: value foo is not a member of _$1
}

class CircleBuilderScala[B <: CircleBuilderScala[B]] {
  def foo: B = ???
}

See also #4745 and other tickets about f-bounded type parameters + type inference.

[scabug]

@retronym said:
Here's a way you can use the techniques touched on in #4745 to centralize your casting pain.

scala> type CircleBuilderF = ({type CB <: CircleBuilder[CB]; type CB1 = CB with CircleBuilder[CB]})#CB1
defined type alias CircleBuilderF

scala> def newCircleBuilder: CircleBuilderF = CircleBuilder.create.asInstanceOf[CircleBuilderF]
newCircleBuilder: CircleBuilderF

scala> newCircleBuilder.centerX(0).centerX(0).build()
res28: javafx.scene.shape.Circle = Circle@38b9a306

Now, if we can just figure out how to roll this into the compiler...

[scabug]

@retronym said (edited on Mar 31, 2013 10:40:42 AM UTC):
You can actually centralize this hack as follows:

scala> implicit class RichFBoundBuilder[B[X <: B[X]]](val _builder: B[_]) {
  type B1 = X forSome { type X <: B[X] }
  def asScala: B1 = _builder.asInstanceOf[B1]
}
warning: there were 2 feature warning(s); re-run with -feature for details
defined class RichFBoundBuilder

scala> import javafx.scene.shape._
import javafx.scene.shape._

scala> CircleBuilder.create.asScala.centerX(0).centerX(0).build()
res0: javafx.scene.shape.Circle = Circle@6c60efe7

[scabug]

Stephen Compall (s11001001) said:
RichFBoundBuilder doesn't require the proof it looks like it requires, though:

Welcome to Scala version 2.10.1 (OpenJDK Client VM, Java 1.6.0_27).
Type in expressions to have them evaluated.
Type :help for more information.

scala> :paste
// Entering paste mode (ctrl-D to finish)

implicit class RichFBoundBuilder[B[X <: B[X]]](val _builder: B[_]) {
  type B1 = X forSome { type X <: B[X] }
  def asScala: B1 = _builder.asInstanceOf[B1]
}

// Exiting paste mode, now interpreting.

defined class RichFBoundBuilder

scala> (x: List[Int]) => x.asScala
res0: List[Int] => X forSome { type X <: List[X] } = <function1>

[scabug]

@retronym said:
I can't figure out a way to tighten this up, unfortunately. You are probably best making a library to wrap all the builder creation calls to localize the use of this hack.

[scabug]

Stephen Compall (s11001001) said (edited on Apr 2, 2013 7:03:48 PM UTC):
Yes; I've written casts for just the builders we use in Scala.

I don't see how 6169 can be fixed, interop-wise, if #7318 isn't a bug, though.

[scabug]

@retronym said:
I would suggest to go for this to tighten up the availability of asScala:

scala> import javafx.scene._, shape._import javafx.scene._
import shape._

scala> implicit class RichJavaFXBuilder[B[X <: B[X]] <: NodeBuilder[_]](val _builder: B[_]) {
     |   type B1 = X forSome { type X <: B[X] }
     |   def asScala: B1 = _builder.asInstanceOf[B1]
     | }
warning: there were 2 feature warning(s); re-run with -feature for details
defined class RichJavaFXBuilder

scala> CircleBuilder.create().asScala.centerX(0).centerX(0).build()
res3: javafx.scene.shape.Circle = Circle@2becc08c

scala> List(1).asScala
<console>:20: error: value asScala is not a member of List[Int]
              List(1).asScala
                      ^

[scabug]

Stephen Compall (s11001001) said:
#1786 almost seems that it should have fixed this, but the F-bound case in particular doesn't seem to have benefited from the inferred existential bound introduced there.

(in 2.11.0-M3, a variant of 1786's sample code):

abstract class SomeClass2[T] {def tValue: T}
class MyClass2[T <: SomeClass2[T]](val myValue:T) 

def myMethod(i:MyClass2[_]) {
   i.myValue.tValue      // << error: value tValue is not a member of _$1
}

def myMethod[T <: SomeClass2[T]](i:MyClass2[T]) {
   i.myValue.tValue      // << works
}

[scabug]

Stephen Compall (s11001001) said:
...but with better justification.

[scabug]

@adriaanm said:
Unassigning and rescheduling to M6 as previous deadline was missed.

[scabug]

Stephen Compall (s11001001) said:
An F-bound isn't necessary to exercise this. In 2.11.0-M8:

public final class It<T extends String>
{
  private final T _it;
  public It(T it){_it = it;}

  public static It<?> instance() {
    return new It<String>("42");
  }

  public static It<? extends String> instance2() {
    return instance();
  }

  public T it(){return _it;}
}

scala> It.instance.it
res1: Any = 42

scala> It.instance2.it
res2: String = 42

[scabug]

@adriaanm said:
I don't see what's wrong here. It<?> corresponds to the existential type It[_], so the instance method packages a string into a box that hides its contents, and when you open up the box you get Any.

[scabug]

@adriaanm said:
I hadn't considered that It's type parameter is bounded by String, so the It<?> in instance's return type is implicitly It[_ <: String].