Description
The canonical encoding of list in scala 2 would be:
sealed trait List[+A]
object List {
case object Empty extends List[Nothing]
case class NonEmpty[A](head: A, tail: List[A]) extends List[A]
}
but this implementation involves a lot of pointer chasing (to iterate N times into the List we need to dereference N times).
We know from work on immutable vectors that having blocks of arrays that we copy on change can give considerable improvements due to cache locality (and fewer derefs).
So one can imagine another list:
sealed trait List[+A] {
def uncons: Option[(A, List[A])]
def prepend(a: A): List[A]
}
object List {
private val BlockSize = 16 // experiment on this
private case object Empty extends List[Nothing] {
def uncons = None
def prepend(a: A) = {
val ary = new Array[Any](BlockSize)
val offset = BlockSize - 1
ary(offset) = a
Impl(offset, ary, Empty)
}
}
private case class Impl[A](offset: Int, block: Array[Any], tail: List[A]) extends List[A] {
def uncons = {
val nextOffset = offset + 1
val next = if (nextOffset == block.length) tail else Impl(nextOffset, block, tail)
Some((block(offset).asInstanceOf[A], next)
}
def prepend(a: A) = {
val ary = new Array[Any](BlockSize)
if (offset > 0) {
// copy the right side
System.arraycopy(block, offset, ary, offset, BlockSize - offset)
val nextOffset = offset - 1
ary(nextOffset) = a
Impl(nextOffset, ary, tail)
}
else {
val offset = BlockSize - 1
ary(offset) = a
Impl(offset, ary, this)
}
}
}
or something like this...
The idea is to copy the block on write. Of course, apply
, map
, foldLeft
, foreach
, iterator
, etc... could be significantly optimized since once you have a handle on the block iterating to the next value will be in cache.
With some appropriate benchmarks you could set the size of the block to some optimal value that I bet would be bigger than 1. It could be that this beats naive List in almost all cases, and when the List isn't too large it could compete with Vector (which isn't as fast as List for stack uses).