如何将enrich-my-library模式应用于Scala集合？

Question

One of the most powerful patterns available in Scala is the enrich-my-library* pattern, which uses implicit conversions to appear to add methods to existing classes without requiring dynamic method resolution. For example, if we wished that all strings had the method spaces that counted how many whitespace characters they had, we could:

class SpaceCounter(s: String) {
  def spaces = s.count(_.isWhitespace)
}
implicit def string_counts_spaces(s: String) = new SpaceCounter(s)

scala> "How many spaces do I have?".spaces
res1: Int = 5

Unfortunately, this pattern runs into trouble when dealing with generic collections. For example, a number of questions have been asked about grouping items sequentially with collections . There is nothing built in that works in one shot, so this seems an ideal candidate for the enrich-my-library pattern using a generic collection C and a generic element type A :

class SequentiallyGroupingCollection[A, C[A] <: Seq[A]](ca: C[A]) {
  def groupIdentical: C[C[A]] = {
    if (ca.isEmpty) C.empty[C[A]]
    else {
      val first = ca.head
      val (same,rest) = ca.span(_ == first)
      same +: (new SequentiallyGroupingCollection(rest)).groupIdentical
    }
  }
}

except, of course, it doesn't work . The REPL tells us:

<console>:12: error: not found: value C
               if (ca.isEmpty) C.empty[C[A]]
                               ^
<console>:16: error: type mismatch;
 found   : Seq[Seq[A]]
 required: C[C[A]]
                 same +: (new SequentiallyGroupingCollection(rest)).groupIdentical
                      ^

There are two problems: how do we get a C[C[A]] from an empty C[A] list (or from thin air)? And how do we get a C[C[A]] back from the same +: line instead of a Seq[Seq[A]] ?

* ^{Formerly known as pimp-my-library.}

Answer 1

The key to understanding this problem is to realize that there are two different ways to build and work with collections in the collections library. One is the public collections interface with all its nice methods. The other, which is used extensively in creating the collections library, but which are almost never used outside of it, is the builders.

Our problem in enriching is exactly the same one that the collections library itself faces when trying to return collections of the same type. That is, we want to build collections, but when working generically, we don't have a way to refer to "the same type that the collection already is". So we need builders .

Now the question is: where do we get our builders from? The obvious place is from the collection itself. This doesn't work . We already decided, in moving to a generic collection, that we were going to forget the type of the collection. So even though the collection could return a builder that would generate more collections of the type we want, it wouldn't know what the type was.

Instead, we get our builders from CanBuildFrom implicits that are floating around. These exist specifically for the purpose of matching input and output types and giving you an appropriately typed builder.

So, we have two conceptual leaps to make:

1. We aren't using standard collections operations, we're using builders.
2. We get these builders from implicit CanBuildFrom s, not from our collection directly.

Let's look at an example.

class GroupingCollection[A, C[A] <: Iterable[A]](ca: C[A]) {
  import collection.generic.CanBuildFrom
  def groupedWhile(p: (A,A) => Boolean)(
    implicit cbfcc: CanBuildFrom[C[A],C[A],C[C[A]]], cbfc: CanBuildFrom[C[A],A,C[A]]
  ): C[C[A]] = {
    val it = ca.iterator
    val cca = cbfcc()
    if (!it.hasNext) cca.result
    else {
      val as = cbfc()
      var olda = it.next
      as += olda
      while (it.hasNext) {
        val a = it.next
        if (p(olda,a)) as += a
        else { cca += as.result; as.clear; as += a }
        olda = a
      }
      cca += as.result
    }
    cca.result
  }
}
implicit def iterable_has_grouping[A, C[A] <: Iterable[A]](ca: C[A]) = {
  new GroupingCollection[A,C](ca)
}

Let's take this apart. First, in order to build the collection-of-collections, we know we'll need to build two types of collections: C[A] for each group, and C[C[A]] that gathers all the groups together. Thus, we need two builders, one that takes A s and builds C[A] s, and one that takes C[A] s and builds C[C[A]] s. Looking at the type signature of CanBuildFrom , we see

CanBuildFrom[-From, -Elem, +To]

which means that CanBuildFrom wants to know the type of collection we're starting with--in our case, it's C[A] , and then the elements of the generated collection and the type of that collection. So we fill those in as implicit parameters cbfcc and cbfc .

Having realized this, that's most of the work. We can use our CanBuildFrom s to give us builders (all you need to do is apply them). And one builder can build up a collection with += , convert it to the collection it is supposed to ultimately be with result , and empty itself and be ready to start again with clear . The builders start off empty, which solves our first compile error, and since we're using builders instead of recursion, the second error also goes away.

One last little detail--other than the algorithm that actually does the work--is in the implicit conversion. Note that we use new GroupingCollection[A,C] not [A,C[A]] . This is because the class declaration was for C with one parameter, which it fills it itself with the A passed to it. So we just hand it the type C , and let it create C[A] out of it. Minor detail, but you'll get compile-time errors if you try another way.

Here, I've made the method a little bit more generic than the "equal elements" collection--rather, the method cuts the original collection apart whenever its test of sequential elements fails.

Let's see our method in action:

scala> List(1,2,2,2,3,4,4,4,5,5,1,1,1,2).groupedWhile(_ == _)
res0: List[List[Int]] = List(List(1), List(2, 2, 2), List(3), List(4, 4, 4), 
                             List(5, 5), List(1, 1, 1), List(2))

scala> Vector(1,2,3,4,1,2,3,1,2,1).groupedWhile(_ < _)
res1: scala.collection.immutable.Vector[scala.collection.immutable.Vector[Int]] =
  Vector(Vector(1, 2, 3, 4), Vector(1, 2, 3), Vector(1, 2), Vector(1))

It works!

The only problem is that we don't in general have these methods available for arrays, since that would require two implicit conversions in a row. There are several ways to get around this, including writing a separate implicit conversion for arrays, casting to WrappedArray , and so on.

---

Edit: My favored approach for dealing with arrays and strings and such is to make the code even more generic and then use appropriate implicit conversions to make them more specific again in such a way that arrays work also. In this particular case:

class GroupingCollection[A, C, D[C]](ca: C)(
  implicit c2i: C => Iterable[A],
           cbf: CanBuildFrom[C,C,D[C]],
           cbfi: CanBuildFrom[C,A,C]
) {
  def groupedWhile(p: (A,A) => Boolean): D[C] = {
    val it = c2i(ca).iterator
    val cca = cbf()
    if (!it.hasNext) cca.result
    else {
      val as = cbfi()
      var olda = it.next
      as += olda
      while (it.hasNext) {
        val a = it.next
        if (p(olda,a)) as += a
        else { cca += as.result; as.clear; as += a }
        olda = a
      }
      cca += as.result
    }
    cca.result
  }
}

Here we've added an implicit that gives us an Iterable[A] from C --for most collections this will just be the identity (eg List[A] already is an Iterable[A] ), but for arrays it will be a real implicit conversion. And, consequently, we've dropped the requirement that C[A] <: Iterable[A] --we've basically just made the requirement for <% explicit, so we can use it explicitly at will instead of having the compiler fill it in for us. Also, we have relaxed the restriction that our collection-of-collections is C[C[A]] --instead, it's any D[C] , which we will fill in later to be what we want. Because we're going to fill this in later, we've pushed it up to the class level instead of the method level. Otherwise, it's basically the same.

Now the question is how to use this. For regular collections, we can:

implicit def collections_have_grouping[A, C[A]](ca: C[A])(
  implicit c2i: C[A] => Iterable[A],
           cbf: CanBuildFrom[C[A],C[A],C[C[A]]],
           cbfi: CanBuildFrom[C[A],A,C[A]]
) = {
  new GroupingCollection[A,C[A],C](ca)(c2i, cbf, cbfi)
}

where now we plug in C[A] for C and C[C[A]] for D[C] . Note that we do need the explicit generic types on the call to new GroupingCollection so it can keep straight which types correspond to what. Thanks to the implicit c2i: C[A] => Iterable[A] , this automatically handles arrays.

But wait, what if we want to use strings? Now we're in trouble, because you can't have a "string of strings". This is where the extra abstraction helps: we can call D something that's suitable to hold strings. Let's pick Vector , and do the following:

val vector_string_builder = (
  new CanBuildFrom[String, String, Vector[String]] {
    def apply() = Vector.newBuilder[String]
    def apply(from: String) = this.apply()
  }
)

implicit def strings_have_grouping(s: String)(
  implicit c2i: String => Iterable[Char],
           cbfi: CanBuildFrom[String,Char,String]
) = {
  new GroupingCollection[Char,String,Vector](s)(
    c2i, vector_string_builder, cbfi
  )
}

We need a new CanBuildFrom to handle the building of a vector of strings (but this is really easy, since we just need to call Vector.newBuilder[String] ), and then we need to fill in all the types so that the GroupingCollection is typed sensibly. Note that we already have floating around a [String,Char,String] CanBuildFrom, so strings can be made from collections of chars.

Let's try it out:

scala> List(true,false,true,true,true).groupedWhile(_ == _)
res1: List[List[Boolean]] = List(List(true), List(false), List(true, true, true))

scala> Array(1,2,5,3,5,6,7,4,1).groupedWhile(_ <= _) 
res2: Array[Array[Int]] = Array(Array(1, 2, 5), Array(3, 5, 6, 7), Array(4), Array(1))

scala> "Hello there!!".groupedWhile(_.isLetter == _.isLetter)
res3: Vector[String] = Vector(Hello,  , there, !!)

Answer 2

As of this commit it's a lot easier to "enrich" Scala collections than it was when Rex gave his excellent answer. For simple cases it might look like this,

import scala.collection.generic.{ CanBuildFrom, FromRepr, HasElem }
import language.implicitConversions

class FilterMapImpl[A, Repr](val r : Repr)(implicit hasElem : HasElem[Repr, A]) {
  def filterMap[B, That](f : A => Option[B])
    (implicit cbf : CanBuildFrom[Repr, B, That]) : That = r.flatMap(f(_).toSeq)
}

implicit def filterMap[Repr : FromRepr](r : Repr) = new FilterMapImpl(r)

which adds a "same result type" respecting filterMap operation to all GenTraversableLike s,

scala> val l = List(1, 2, 3, 4, 5)
l: List[Int] = List(1, 2, 3, 4, 5)

scala> l.filterMap(i => if(i % 2 == 0) Some(i) else None)
res0: List[Int] = List(2, 4)

scala> val a = Array(1, 2, 3, 4, 5)
a: Array[Int] = Array(1, 2, 3, 4, 5)

scala> a.filterMap(i => if(i % 2 == 0) Some(i) else None)
res1: Array[Int] = Array(2, 4)

scala> val s = "Hello World"
s: String = Hello World

scala> s.filterMap(c => if(c >= 'A' && c <= 'Z') Some(c) else None)
res2: String = HW

And for the example from the question, the solution now looks like,

class GroupIdenticalImpl[A, Repr : FromRepr](val r: Repr)
  (implicit hasElem : HasElem[Repr, A]) {
  def groupIdentical[That](implicit cbf: CanBuildFrom[Repr,Repr,That]): That = {
    val builder = cbf(r)
    def group(r: Repr) : Unit = {
      val first = r.head
      val (same, rest) = r.span(_ == first)
      builder += same
      if(!rest.isEmpty)
        group(rest)
    }
    if(!r.isEmpty) group(r)
    builder.result
  }
}

implicit def groupIdentical[Repr : FromRepr](r: Repr) = new GroupIdenticalImpl(r)

Sample REPL session,

scala> val l = List(1, 1, 2, 2, 3, 3, 1, 1)
l: List[Int] = List(1, 1, 2, 2, 3, 3, 1, 1)

scala> l.groupIdentical
res0: List[List[Int]] = List(List(1, 1),List(2, 2),List(3, 3),List(1, 1))

scala> val a = Array(1, 1, 2, 2, 3, 3, 1, 1)
a: Array[Int] = Array(1, 1, 2, 2, 3, 3, 1, 1)

scala> a.groupIdentical
res1: Array[Array[Int]] = Array(Array(1, 1),Array(2, 2),Array(3, 3),Array(1, 1))

scala> val s = "11223311"
s: String = 11223311

scala> s.groupIdentical
res2: scala.collection.immutable.IndexedSeq[String] = Vector(11, 22, 33, 11)

Again, note that the same result type principle has been observed in exactly the same way that it would have been had groupIdentical been directly defined on GenTraversableLike .

Answer 3

As of this commit the magic incantation is slightly changed from what it was when Miles gave his excellent answer.

The following works, but is it canonical? I hope one of the canons will correct it. (Or rather, cannons, one of the big guns.) If the view bound is an upper bound, you lose application to Array and String. It doesn't seem to matter if the bound is GenTraversableLike or TraversableLike; but IsTraversableLike gives you a GenTraversableLike.

import language.implicitConversions
import scala.collection.{ GenTraversable=>GT, GenTraversableLike=>GTL, TraversableLike=>TL }
import scala.collection.generic.{ CanBuildFrom=>CBF, IsTraversableLike=>ITL }

class GroupIdenticalImpl[A, R <% GTL[_,R]](val r: GTL[A,R]) {
  def groupIdentical[That](implicit cbf: CBF[R, R, That]): That = {
    val builder = cbf(r.repr)
    def group(r: GTL[_,R]) {
      val first = r.head
      val (same, rest) = r.span(_ == first)
      builder += same
      if (!rest.isEmpty) group(rest)
    }
    if (!r.isEmpty) group(r)
    builder.result
  }
}

implicit def groupIdentical[A, R <% GTL[_,R]](r: R)(implicit fr: ITL[R]):
  GroupIdenticalImpl[fr.A, R] =
  new GroupIdenticalImpl(fr conversion r)

There's more than one way to skin a cat with nine lives. This version says that once my source is converted to a GenTraversableLike, as long as I can build the result from GenTraversable, just do that. I'm not interested in my old Repr.

class GroupIdenticalImpl[A, R](val r: GTL[A,R]) {
  def groupIdentical[That](implicit cbf: CBF[GT[A], GT[A], That]): That = {
    val builder = cbf(r.toTraversable)
    def group(r: GT[A]) {
      val first = r.head
      val (same, rest) = r.span(_ == first)
      builder += same
      if (!rest.isEmpty) group(rest)
    }
    if (!r.isEmpty) group(r.toTraversable)
    builder.result
  }
}

implicit def groupIdentical[A, R](r: R)(implicit fr: ITL[R]):
  GroupIdenticalImpl[fr.A, R] =
  new GroupIdenticalImpl(fr conversion r)

This first attempt includes an ugly conversion of Repr to GenTraversableLike.

import language.implicitConversions
import scala.collection.{ GenTraversableLike }
import scala.collection.generic.{ CanBuildFrom, IsTraversableLike }

type GT[A, B] = GenTraversableLike[A, B]
type CBF[A, B, C] = CanBuildFrom[A, B, C]
type ITL[A] = IsTraversableLike[A]

class FilterMapImpl[A, Repr](val r: GenTraversableLike[A, Repr]) { 
  def filterMap[B, That](f: A => Option[B])(implicit cbf : CanBuildFrom[Repr, B, That]): That = 
    r.flatMap(f(_).toSeq)
} 

implicit def filterMap[A, Repr](r: Repr)(implicit fr: ITL[Repr]): FilterMapImpl[fr.A, Repr] = 
  new FilterMapImpl(fr conversion r)

class GroupIdenticalImpl[A, R](val r: GT[A,R])(implicit fr: ITL[R]) { 
  def groupIdentical[That](implicit cbf: CBF[R, R, That]): That = { 
    val builder = cbf(r.repr)
    def group(r0: R) { 
      val r = fr conversion r0
      val first = r.head
      val (same, other) = r.span(_ == first)
      builder += same
      val rest = fr conversion other
      if (!rest.isEmpty) group(rest.repr)
    } 
    if (!r.isEmpty) group(r.repr)
    builder.result
  } 
} 

implicit def groupIdentical[A, R](r: R)(implicit fr: ITL[R]):
  GroupIdenticalImpl[fr.A, R] = 
  new GroupIdenticalImpl(fr conversion r)