I'm encountering a weird problem trying to use a simple ADT in Scala 2.10.
The following program shows the issue:
sealed trait Trie[+T, +V]
case object EmptyTrie extends Trie[Nothing, Nothing]
case class TrieNode[T, V](data: V, subTrie: Map[T, Trie[T, V]]) extends Trie[T, V]
case class TrieLeaf[T, V](data: V) extends Trie[T, V]
object Trie {
def emptyTrie[T, V](): Trie[T, V] = EmptyTrie
def addWith[T, V](xs: List[T], trie: Trie[T, V], update: (V => V), default: V): Trie[T, V] =
xs match {
case Nil => trie match {
case EmptyTrie => new TrieLeaf(default)
case TrieNode(value, subTrie) => new TrieNode(update(value), subTrie)
case TrieLeaf(value) => new TrieLeaf(update(value))
}
case x :: xs1 => trie match {
case EmptyTrie => new TrieNode(default, Map(x -> addWith(xs1, EmptyTrie, update, default)))
case TrieNode(value, subTrie) => new TrieNode(update(value), subTrie.updated(x, addWith(xs1, subTrie.getOrElse(x, EmptyTrie), update, default)))
case TrieLeaf(value) => new TrieNode(update(value), Map(x -> addWith(xs1, EmptyTrie, update, default)))
}
}
}
Compiling this code with Scala 2.10.x result in the following errors:
[error] /tmp/rendererkH7rTCTbYi/src/main/scala/test.scala:22: type mismatch;
[error] found : x.type (with underlying type T)
[error] required: ?T4 where type ?T4 <: T (this is a GADT skolem)
[error] case TrieNode(value, subTrie) => new TrieNode(update(value), subTrie.updated(x, addWith(xs1, subTrie.getOrElse(x, EmptyTrie), update, default)))
[error] ^
[error] /tmp/rendererkH7rTCTbYi/src/main/scala/test.scala:22: type mismatch;
[error] found : x.type (with underlying type T)
[error] required: ?T4 where type ?T4 <: T (this is a GADT skolem)
[error] case TrieNode(value, subTrie) => new TrieNode(update(value), subTrie.updated(x, addWith(xs1, subTrie.getOrElse(x, EmptyTrie), update, default)))
[error] ^
The same code compiled with Scala 2.9.3 compiles successfully and the programs works as expected.
I've worked the problem around by changing the EmptyTrie definition to:
case class EmptyTrie[T, V]() extends Trie[T, V]
and dropping the covariance specification on T and V on the trait definition.
However I really don't like having to instantiate new empty classes when a single object instance would do the job. Using the object just seems a lot cleaner.
In the end my questions are:
- Is this a regression in the scala compiler?
- Is it possible to add
some type hints to aid the compiler (all my attempts at this
miserably failed)?