AnyRefMap.getOrElseUpdate is faulty

[scabug]

The dotty compiler crashed with an "ArrayIndexOutofBounds" exception in "getOrElseUpdate" on an AnyRefMap. I could not minimize it, but think I know where the failure comes from. Here's the code of getOrElseUpdate:

  override def getOrElseUpdate(key: K, defaultValue: => V): V = {
    val h = hashOf(key)
    val i = seekEntryOrOpen(h, key)
    if (i < 0) {
      val value = defaultValue
      _size += 1
      val j = i & IndexMask
      _hashes(j) = h
      _keys(j) = key.asInstanceOf[AnyRef]
      _values(j) = value.asInstanceOf[AnyRef]
      if ((i & VacantBit) != 0) _vacant -= 1
      else if (imbalanced) repack()
      value
    }
    else _values(i).asInstanceOf[V]
  }

The error happens if the defaultValue computation causes a size increase and rehash of the map. In that case the value of i seems to be no longer valid.
The actual call causing the error was:

    typedTree.getOrElseUpdate(expanded(tree), typer.typed(tree, pt))

Indeed the default value expression typer.typed(tree, pt) could easily create enough new typedTree entries to cause a rehash.

[scabug]

Imported From: https://issues.scala-lang.org/browse/SI-8213?orig=1
Reporter: @odersky
Affected Versions: 2.11.0-M7

[scabug]

@Ichoran said (edited on Jan 30, 2014 6:58:17 PM UTC):
I hadn't considered that defaultValue would alter the map itself. Easy fix, though.

[scabug]

@Ichoran said:
scala/scala#3434

[SethTisue]

fwiw, I think this could plausibly have simply been closed as "wontfix". blowing up if you change the map out from under itself like this doesn't seem like a bug to me, necessarily. /cc @hrhino

[Ichoran]

One could wontfix it, but it has the unfriendly property then that m.getOrElseUpdate(k, f()) is not the same as the apparently logically equivalent if (!(m contains k)) m(k) = f(); m(k). Having to maintain the mental distinction is an unwelcome burden on the user.