Type inference doesn't work for implicitly converted functions

[scabug]

The Scala compiler does not infer the argument types for function expressions (e.g. x => x+1) when the function is passed to a method that requires it to be implicitly converted to another type, even though it does infer the types of other objects in a similar situation. (Thanks to Heather Miller for pointing out the second case.) For example, take a look at the attached test2.scala:

class MyFunc[-A, +B] {}

class MyCollection[A] {
  def map[B](f: MyFunc[A, B]): MyCollection[B] = new MyCollection[B]
}

class OtherFunc[-A, +B] {}

object Test {
  implicit def functionToMyFunc[A, B](f: A => B): MyFunc[A, B] = new MyFunc

  implicit def otherFuncToMyFunc[A, B](f: OtherFunc[A, B]): MyFunc[A, B] = new MyFunc

  def main(args: Array[String]) {
    val col = new MyCollection[Int]

    // Doesn't compile: error: missing parameter type for expanded function ((x$1) => x$1.toString)
    // println(col.map(_.toString)) 

    // Doesn't compile: error: missing parameter type
    // println(col.map(x => x.toString))

    // Does compile
    println(col.map((x: Int) => x.toString))
    
    // Does compile (even though type params of OtherFunc not given)
    println(col.map(new OtherFunc))
  }
}

This is a problem for cases where you want to represent functions through some richer class, or to apply a macro to function expressions passed to a set of methods. For example, in Spark, I would like to apply a macro to all closure expressions passed to map(), filter(), groupBy(), and other operations in order to properly bind variables, and I would've liked to do this through an implicit conversion macro from Function1[A, B] to ShippableFunction[A, B]. Because of this bug, this can't be done without losing type inference. It would be nice to add it in this case since it seems that the types for other objects can be inferred anyway.

[scabug]

Imported From: https://issues.scala-lang.org/browse/SI-6221?orig=1
Reporter: Matei Zaharia (matei)
Affected Versions: 2.9.1, 2.10.0-M6
Attachments:

• test2.scala (created on Aug 12, 2012 2:20:26 PM UTC, 860 bytes)

[scabug]

@heathermiller said:
Update on this after digging around a bit.

The following seems to happen.

In the case of col.map(new OtherFunc), the expected type of map is MyFunc[Int,?]. Meanwhile, the type parameters of OtherFunc are inferred independently of the implicit conversion to be OtherFunc[Any, Nothing].

In the case of col.map(_.toString), the expected type of map is also MyFunc[Int,?].
However, in this case, implicit search seems to try to find a conversion from Function1[<error>,<error>] to the above expected type.

-Ydebug -Ylog:typer output for the col.map(_.toString) case:

[log typer] inferring view from scala.this.Function1[<error>,<error>] to <empty>.this.MyFunc[scala.this.Int,?]
[log typer] infer view from scala.this.Function1[<error>,<error>] to <empty>.this.MyFunc[scala.this.Int,?]

In both cases, implicit search does not have an influence on inferring the type of the argument of map. This doesn't seem to be spec'd, but I agree that this would be very nice to have.

[scabug]

@hubertp said:
I can make it to work if MyFunc extends Function1 as even that doesn't work now (https://github.com/hubertp/scala-dev/tree/2.10.x-issue/6221) but I realize that isn't entirely what you are asking.
The implicit search involving ErrorType => ErrorType is spurious and shouldn't be happening at all.

[scabug]

@heathermiller said:
Awesome, I've been messing around on a fix in the same place. You're right that the ticket is phrased to address a more general problem. However, your fix works for our use case-- we would ultimately extend Function1 anyway. Though, regardless, yes, the more general implicit conversion + type inference issue actually illustrated above still remains.

[scabug]

Matei Zaharia (matei) said:
Hubert, if that fix can make it into some version of 2.10, that would be awesome. It would really make it much simpler to use macros to preprocess closures, which has a lot of DSL applications.

[scabug]

@hubertp said:
Might be a bit too late for 2.10.0, but it should make it into one of the inevitable minor 2.10 releases.

[scabug]

@xeno-by said:
Seems related #7187

[scabug]

@paulp said:
scala/scala#2650