Deal with special case for Array[Unit] in Array.{copyOf,copyAs}

[retronym]

No description provided.

[retronym]

Thanks to @mkeskells for spotting that we were filling Arrry[Unit] with null-s.

[sjrd]

That's not false, but fixing this seems like a drop in the ocean. There are many other cases where constructing an array of Units fills it with null. new Array[Unit](n), for starters:

$ ~/opt/scala-2.13.0-M4/bin/scala
Welcome to Scala 2.13.0-M4 (Java HotSpot(TM) 64-Bit Server VM, Java 1.8.0_40).
Type in expressions for evaluation. Or try :help.

scala> val a = new Array[Unit](5)
a: Array[Unit] = Array((), (), (), (), ()) // This is a lie

scala> a.asInstanceOf[Array[scala.runtime.BoxedUnit]](0)
res0: scala.runtime.BoxedUnit = null

The lie comes from ArraySeq.ofUnit.apply (or WrappedArray.ofUnit.apply in earlier versions), which returns a hard Unit, and therefore the result of array(index) is actually dropped, and a new BoxedUnit.UNIT is created in the bridge apply(int)Object generated to implement the inherited def apply(index: Int): T.

[szeiger]

Any reason not to use the usual match case x: Array[BoxedUnit]?

[retronym]

I've changed to use that. The only tricky part is that it needs to precede the match on _: Array[AnyRef] due to array covariance in the VM.

[xuwei-k]

BTW xuwei-k@a9761ba 🤔

[szeiger]

@xuwei-k The JVM doesn't have void arrays, therefore Array[Unit] is really Array[BoxedUnit]. The pattern matcher would have to explicitly check for that case when matching on Array[AnyRef]. It's not exactly a common case, favoring correctness over performance seems reasonable. I suggest filing a bug.