collections

Arrays

Array is a special kind of collection in Scala. On the one hand, Scala arrays correspond one-to-one to Java arrays. That is, a Scala array Array[Int] is represented as a Java int[], an Array[Double] is represented as a Java double[] and a Array[String] is represented as a Java String[]. But at the same time, Scala arrays offer much more than their Java analogues. First, Scala arrays can be generic. That is, you can have an Array[T], where T is a type parameter or abstract type. Second, Scala arrays are compatible with Scala sequences - you can pass an Array[T] where a Seq[T] is required. Finally, Scala arrays also support all sequence operations. Here’s an example of this in action:

scala> val a1 = Array(1, 2, 3)
a1: Array[Int] = Array(1, 2, 3)
scala> val a2 = a1 map (_ * 3)
a2: Array[Int] = Array(3, 6, 9)
scala> val a3 = a2 filter (_ % 2 != 0)
a3: Array[Int] = Array(3, 9)
scala> a3.reverse
res1: Array[Int] = Array(9, 3)

Given that Scala arrays are represented just like Java arrays, how can these additional features be supported in Scala? In fact, the answer to this question differs between Scala 2.8 and earlier versions. Previously, the Scala compiler somewhat “magically” wrapped and unwrapped arrays to and from Seq objects when required in a process called boxing and unboxing. The details of this were quite complicated, in particular when one created a new array of generic type Array[T]. There were some puzzling corner cases and the performance of array operations was not all that predictable.

The Scala 2.8 design is much simpler. Almost all compiler magic is gone. Instead the Scala 2.8 array implementation makes systematic use of implicit conversions. In Scala 2.8 an array does not pretend to be a sequence. It can’t really be that because the data type representation of a native array is not a subtype of Seq. Instead there is an implicit “wrapping” conversion between arrays and instances of class scala.collection.mutable.WrappedArray, which is a subclass of Seq. Here you see it in action:

scala> val seq: Seq[Int] = a1
seq: Seq[Int] = WrappedArray(1, 2, 3)
scala> val a4: Array[Int] = seq.toArray
a4: Array[Int] = Array(1, 2, 3)
scala> a1 eq a4
res2: Boolean = true

The interaction above demonstrates that arrays are compatible with sequences, because there’s an implicit conversion from arrays to WrappedArrays. To go the other way, from a WrappedArray to an Array, you can use the toArray method defined in Traversable. The last REPL line above shows that wrapping and then unwrapping with toArray gives the same array you started with.

There is yet another implicit conversion that gets applied to arrays. This conversion simply “adds” all sequence methods to arrays but does not turn the array itself into a sequence. “Adding” means that the array is wrapped in another object of type ArrayOps which supports all sequence methods. Typically, this ArrayOps object is short-lived; it will usually be inaccessible after the call to the sequence method and its storage can be recycled. Modern VMs often avoid creating this object entirely.

The difference between the two implicit conversions on arrays is shown in the next REPL dialogue:

scala> val seq: Seq[Int] = a1
seq: Seq[Int] = WrappedArray(1, 2, 3)
scala> seq.reverse
res2: Seq[Int] = WrappedArray(3, 2, 1)
scala> val ops: collection.mutable.ArrayOps[Int] = a1
ops: scala.collection.mutable.ArrayOps[Int] = [I(1, 2, 3)
scala> ops.reverse
res3: Array[Int] = Array(3, 2, 1)

You see that calling reverse on seq, which is a WrappedArray, will give again a WrappedArray. That’s logical, because wrapped arrays are Seqs, and calling reverse on any Seq will give again a Seq. On the other hand, calling reverse on the ops value of class ArrayOps will give an Array, not a Seq.

The ArrayOps example above was quite artificial, intended only to show the difference to WrappedArray. Normally, you’d never define a value of class ArrayOps. You’d just call a Seq method on an array:

scala> a1.reverse
res4: Array[Int] = Array(3, 2, 1)

The ArrayOps object gets inserted automatically by the implicit conversion. So the line above is equivalent to

scala> intArrayOps(a1).reverse
res5: Array[Int] = Array(3, 2, 1)

where intArrayOps is the implicit conversion that was inserted previously. This raises the question how the compiler picked intArrayOps over the other implicit conversion to WrappedArray in the line above. After all, both conversions map an array to a type that supports a reverse method, which is what the input specified. The answer to that question is that the two implicit conversions are prioritized. The ArrayOps conversion has a higher priority than the WrappedArray conversion. The first is defined in the Predef object whereas the second is defined in a class scala.LowPritoryImplicits, which is inherited from Predef. Implicits in subclasses and subobjects take precedence over implicits in base classes. So if both conversions are applicable, the one in Predef is chosen. A very similar scheme works for strings.

So now you know how arrays can be compatible with sequences and how they can support all sequence operations. What about genericity? In Java you cannot write a T[] where T is a type parameter. How then is Scala’s Array[T] represented? In fact a generic array like Array[T] could be at run-time any of Java’s eight primitive array types byte[], short[], char[], int[], long[], float[], double[], boolean[], or it could be an array of objects. The only common run-time type encompassing all of these types is AnyRef (or, equivalently java.lang.Object), so that’s the type to which the Scala compiler maps Array[T]. At run-time, when an element of an array of type Array[T] is accessed or updated there is a sequence of type tests that determine the actual array type, followed by the correct array operation on the Java array. These type tests slow down array operations somewhat. You can expect accesses to generic arrays to be three to four times slower than accesses to primitive or object arrays. This means that if you need maximal performance, you should prefer concrete over generic arrays. Representing the generic array type is not enough, however, There must also be a way to create generic arrays. This is an even harder problem, which requires a little bit of help from you. To illustrate the problem, consider the following attempt to write a generic method that creates an array.

// this is wrong!
def evenElems[T](xs: Vector[T]): Array[T] = {
  val arr = new Array[T]((xs.length + 1) / 2)
  for (i <- 0 until xs.length by 2)
    arr(i / 2) = xs(i)
  arr
}

The evenElems method returns a new array that consist of all elements of the argument vector xs which are at even positions in the vector. The first line of the body of evenElems creates the result array, which has the same element type as the argument. So depending on the actual type parameter for T, this could be an Array[Int], or an Array[Boolean], or an array of some of the other primitive types in Java, or an array of some reference type. But these types have all different runtime representations, so how is the Scala runtime going to pick the correct one? In fact, it can’t do that based on the information it is given, because the actual type that corresponds to the type parameter T is erased at runtime. That’s why you will get the following error message if you compile the code above:

error: cannot find class manifest for element type T
  val arr = new Array[T]((arr.length + 1) / 2)
        ^ What's required here is that you help the compiler out by providing some runtime hint what the actual type parameter of `evenElems` is. This runtime hint takes the form of a class manifest of type `scala.reflect.ClassManifest`. A class manifest is a type descriptor object which describes what the top-level class of a type is. Alternatively to class manifests there are also full manifests of type `scala.reflect.Manifest`, which describe all aspects of a type. But for array creation, only class manifests are needed.

The Scala compiler will construct class manifests automatically if you instruct it to do so. “Instructing” means that you demand a class manifest as an implicit parameter, like this:

def evenElems[T](xs: Vector[T])(implicit m: ClassManifest[T]): Array[T] = ...

Using an alternative and shorter syntax, you can also demand that the type comes with a class manifest by using a context bound. This means following the type with a colon and the class name ClassManifest, like this:

// this works
def evenElems[T: ClassManifest](xs: Vector[T]): Array[T] = {
  val arr = new Array[T]((xs.length + 1) / 2)
  for (i <- 0 until xs.length by 2)
    arr(i / 2) = xs(i)
  arr
}

The two revised versions of evenElems mean exactly the same. What happens in either case is that when the Array[T] is constructed, the compiler will look for a class manifest for the type parameter T, that is, it will look for an implicit value of type ClassManifest[T]. If such a value is found, the manifest is used to construct the right kind of array. Otherwise, you’ll see an error message like the one above.

Here is some REPL interaction that uses the evenElems method.

scala> evenElems(Vector(1, 2, 3, 4, 5))
res6: Array[Int] = Array(1, 3, 5)
scala> evenElems(Vector("this", "is", "a", "test", "run"))
res7: Array[java.lang.String] = Array(this, a, run)

In both cases, the Scala compiler automatically constructed a class manifest for the element type (first, Int, then String) and passed it to the implicit parameter of the evenElems method. The compiler can do that for all concrete types, but not if the argument is itself another type parameter without its class manifest. For instance, the following fails:

scala> def wrap[U](xs: Array[U]) = evenElems(xs)
<console>:6: error: could not find implicit value for
 evidence parameter of type ClassManifest[U]
     def wrap[U](xs: Array[U]) = evenElems(xs)
                                      ^ What happened here is that the `evenElems` demands a class manifest for the type parameter `U`, but none was found. The solution in this case is, of course, to demand another implicit class manifest for `U`. So the following works:

scala> def wrap[U: ClassManifest](xs: Array[U]) = evenElems(xs)
wrap: [U](xs: Array[U])(implicit evidence$1: ClassManifest[U])Array[U]

This example also shows that the context bound in the definition of U is just a shorthand for an implicit parameter named here evidence$1 of type ClassManifest[U].

In summary, generic array creation demands class manifests. So whenever creating an array of a type parameter T, you also need to provide an implicit class manifest for T. The easiest way to do this is to declare the type parameter with a ClassManifest context bound, as in [T: ClassManifest].

blog comments powered by Disqus