for comprehension fails when using an intermediate variable

Question

Scastie version

With this infra-structure:

trait Pat[A]
object Pat {
  def apply[A](elems: A*): Pat[A] = ???
}

implicit class PatOps[A](p: Pat[A]) {
  def ++ (that: Pat[A]): Pat[A] = ???

  def bubble: Pat[Pat[A]] = ???

  def grouped(size: Pat[Int]): Pat[Pat[A]] = ???
}

implicit class PatPatOps[A](p: Pat[Pat[A]]) {
  def map[B](f: Pat[A] => Pat[B]): Pat[Pat[B]] = ???

  def flatMap[B](f: Pat[A] => Pat[B]): Pat[B] = ???

  def flatten: Pat[A] = ???
}

It is possible to write the following for-comprehension:

trait Test1 {
  val lPat = Pat(1, 2, 3)
  val xs = for {
    len     <- lPat.bubble
    cantus  <- Pat(4, 40, 3).grouped(len)
  } yield {
    cantus ++ Pat(-1)
  }
  xs.flatten
}

But this one, using an intermediate variable, fails:

trait Test2 {
  val lPat = Pat(1, 2, 3)
  val xs = for {
    len     <- lPat.bubble  // XXX
    brown = Pat(4, 40, 3)
    cantus  <- brown.grouped(len)
  } yield {
    cantus ++ Pat(-1)
  }
  xs.flatten
}

The error for the line marked XXX is:

type mismatch;
 found   : (Playground.this.Pat[Int], Playground.this.Pat[Int])
 required: Playground.this.Pat[?]

Scala is 2.12.4

Answer 1

This happens when you define map with overly restrictive signature map[B](f: Pat[A] => Pat[B]). Recall that usually, it is supposed to accept functions with arbitrary result type B, that is, it's supposed to be rather something like:

map[B](f: A => B): <stuff>

Now, your for-comprehension with intermediate helper variable brown

  val xs = for {
    len     <- lPat.bubble
    brown = Pat(4, 40, 3)
    cantus  <- brown.grouped(len)
  } yield {
    cantus ++ Pat(-1)
  }

is rewritten using a map into

  val xs = lPat.bubble.map(((len) => {
    val brown = Pat(4, 40, 3);
    scala.Tuple2(len, brown)
  })).flatMap(((x$1) => x$1: @scala.unchecked match {
    case scala.Tuple2((len @ _), (brown @ _)) => 
      brown.
        grouped(len).
        map(((cantus) => cantus.$plus$plus(Pat(-1))))
  }))

as described in the documentation or in my overly detailed answer here.

Note how the return type of the implicitly generated map is now something like (Pat[A], Pat[Int]) (the type of the tuple (len, brown)), and doesn't match the pattern Pat[B] from your declaration.

I don't see any workarounds. Just do whatever you can to avoid defining map as map[B](f: Pat[A] => Pat[B]), otherwise it will behave way too strangely. Avoid breaking functoriality of map. If your Pat[X] cannot map f: X => Y to a Pat[Y] for arbitrary X and Y, then don't call it map.

---

Edit: there is always a work-around...

One thing you could do is to introduce some kind of implicitly supplied CanPatFrom:

trait CanPatFrom[X, A] extends (X => Pat[A])

and then

...
def map[X, B](f: Pat[A] => X)(implicit cpf: CanPatFrom[X, B]) = {
  val pb: Pat[B] = cpf(f(...))
  /* do your stuff here with `Pat[B]` instead of 
   * generic `X`
   */
  ...
}

Assuming that your Pat carries some kind of cartesian-monoidal structure, you could define

• CanPatFrom[Pat[A], Pat[A]],
• CanPatFrom[(Pat[A], Pat[B]), Pat[(A, B)]],
• CanPatFrom[(Pat[A], Pat[B], Pat[C]), Pat[(A, B, C)]],
• ...

and thereby obtain a map that can at least cope with the case that the return type is a tuple.