Offer copy-on-GC sliced arrays
It's sometimes useful to split an array and send each part on to a different computation. Slicing works very well for this sort of thing, as long as both computations run to completion or become garbage at about the same time. It can work badly, however, if one computation is completed or abandoned and the other is retained—it may only hold a tiny slice, but that's enough to keep the entire array alive.
The alternative, currently, is to copy the array to form two new arrays. This gets rid of the retention problem, but introduces a performance problem.
One way to solve these problems might be to offer sliced arrays supported by the RTS. On collection, the garbage collector would copy each slice separately, turning it into an independent (and independently collectible) array.