implement higher-order unification for type constructor inference

[scabug]

implement what's described in "Partial polymorphic type inference and higher-order unification"

object Test {
  def meh[M[_], A](x: M[A]): M[A] = x
  meh{(x: Int) => x} // should solve ?M = [X] X => X and ?A = Int ... 
}

[scabug]

Imported From: https://issues.scala-lang.org/browse/SI-2712?orig=1
Reporter: @adriaanm
Other Milestones: 2.12.0-M5
See #4719
Blocks #5993

[scabug]

Harrison Klaperman (hlklaperman) said:
You can find that paper here: http://www.cs.cmu.edu/~fp/papers/lfp88.pdf. I hope this is useful to others who stumble upon this issue. Now if only I understood what the paper were talking about...

[scabug]

@Blaisorblade said:
Further work on that problem is described in "Unification via explicit substitutions: The case of higher-order patterns" - http://www.cs.cmu.edu/~fp/papers/INRIA-RR-3591.pdf, from Dowek, Hardin, Kirchner and Pfenning (Pfenning is the author of the above paper). I am not an expert, but this work is cited in Pierce's TAPL (Sec. 23.6, pag. 355), which explains that unlike the algorithm above it is terminating and always finds a unique solution; the paper suggests that this work applies to an important subcase. However, I cannot fully judge myself the respective merits of the two papers.

[scabug]

@milessabin said:
With dependent method types we have a somewhat cumbersome, but just about usable workaround. The key idea is to use implicits to destructure a type, and then pick out and use its component parts as dependent types.

// Template we will unpack into
trait UnpackM[MA] {
  type M[_]
  type A
}
  
// Destructuring implicits
implicit def unpackM1[M0[_], A0] = new UnpackM[M0[A0]] {
  type M[X] = M0[X]
  type A = A0
}

implicit def unpackM2[M0[_, _], A0, B0] = new UnpackM[M0[A0, B0]] {
  type M[X] = M0[A0, X]
  type A = B0
}

def meh[MA](x : MA)(implicit u : UnpackM[MA]) : (u.M[String], List[u.A]) = null
//      ^^          ^^^^^^^^^^^^^^^^^^^^^^^^     ^^^^^^^^^^^       ^^^
//      (1)                   (2)                    (3)           (3)
// 1: Type variable being destructured
// 2: Destructuring implicit
// 3: Uses of (dependent) component types
  
val m = meh{(x: Int) => x}
implicitly[m.type <:< (Int => String, List[Int])]

See also here for a slightly more elaborate scalaz inspired example.

[scabug]

@Blaisorblade said:
Milles Sabin, your workaround is interesting:
But unfortunately the example you link still needs these two lines:

type EitherInt[A] = Either[Int, A]
implicit object EitherTC extends TC[EitherInt]

Moreover, your code does not even correctly solve the original problem - it only infers that Int => Int matches M0[A0, B0] with M0 = Function1, not that it matches M1[A1] with M1[X] = X => X.

[scabug]

@retronym said:
See also #6744, where we would like to infer these sort of types in pattern matching.

[scabug]

@pchiusano said:
Would it be any easier to just look for partial applications of existing type constructors in left-to-right order? Haskell does roughly this (actually, type constructors are just curried, see below), it is tractable, and people don't seem to have an issue with the limitation, though occasionally you do have to introduce a new type just to flip the order of some type parameters.

Here's an example:

case class State[S,A](run: S => (A,S))
object Blah {
  def foo[M[_],A](m: M[A]): M[A] = m
  ...
  val blah = foo(State((i: Int) => ("hi",i+1)))
}

Another way to think about this is that type constructors are curried, so State[S,A] is actually represented as Ap(Ap(State, S), A), and likewise the M[A] is Ap(freevar, A) - these unify with freevar = Ap(State, S).

[scabug]

Dan Rosen (mergeconflict) said:
To Paul's comment, I agree that looking at the type parameters as curried, partially applied from left to right (#4719) is probably usually good enough.

If State[S, A] has pseudo-Haskell-ish kind (*, *) -> *, then State[S][A] would have kind * -> * -> *, and State[S] would have kind * -> *.

[scabug]

@Blaisorblade said:
It seems somehow I never mentioned higher-order pattern unification by Miller, which is what Agda uses — the paper I linked ("Unification via explicit substitutions: The case of higher-order patterns") is indeed about that. For whatever reason, everybody steers away from the original Huet algorithm, including modern implementation of lambdaProlog (which is what "Partial Polymorphic Type Inference and Higher-Order Unification" uses).

[scabug]

@adriaanm said:
Oh, jira.

[scabug]

@SethTisue said:
#2712 closed!! Dewey defeats Truman!!

[scabug]

@milessabin said:
That was cruel and heartless ...

[scabug]

@etorreborre said:
My heart rate went up for a moment...

[scabug]

Stephen Compall (s11001001) said:
long, wistful sigh

[scabug]

Jed Wesley-Smith (jedws) said:
well, that was (briefly) exciting!