It was observed that a for comprehension and foreach showed different behavior when iterating over a sorted map. This led to the discover that the tuple pattern in

for ((x, y) <- map) println(x)

is not presently considered irrefutable, due to null. It was irrefutable in 2.8.1, but that was more out of the absence of implementation than out of intention.

// scala 2.8.1
scala> (null: Any) match { case List(_*) => true }
res4: Boolean = true

So the filter call being added in 2.9 (in addition to impacting performance and increasing bytecode size) loses the sortedness of the map.

Irrefutable specification says:

> > A pattern p is irrefutable for a type T , if one of the following applies:
> > 1. pisavariablepattern,
> > 2. pisatypedpatternx:T′,andT<:T′,
> > 3. pisaconstructorpatternc(p1,...,pn),thetypeTisaninstanceofclassc,the primary constructor (�5.3) of type T has argument types T1, ..., Tn, and each pi isirrefutableforTi.

I suggest it wouldn't hurt to explicitly account for null somewhere in the above lines. A typed pattern _: T' or a constructor pattern T(..) does not match null, but conditions 2 and 3 above quietly add in null. (I'm not suggesting changing the specified behavior -- in the absence of first class support for NotNull-ness, if null undid irrefutability there would be little left to be irrefutable.)

So I am proceeding under the belief that the 2.8.1 behavior shown below is what we want.

Scala 2.8.1:

scala> val xs: List[(Int, Int)] = List((2, 2), null)
xs: List[(Int, Int)] = List((2,2), null)

scala> for ((x, y) <- xs) yield x
scala.MatchError: null
	at $$anonfun$$1.apply(<console>:7)

Scala 2.9.0:

scala> val xs: List[(Int, Int)] = List((2, 2), null)
xs: List[(Int, Int)] = List((2,2), null)

scala> for ((x, y) <- xs) yield x
res0: List[Int] = List(2)