I am a friend of simple garbage collectors. Let initial placement be decided by running code.

Let later placement be decided by the order in which the heap is scavenged. That is what e.g. re-unifies array-of-pointer target instances. You scavenge the array of pointers and in the default case you line up the instances the pointers point to beautifully one after another. Even if the initial allocation had interleaving other memory. In that case your code gets faster after GC than before the data is first GCed.

I don't see how we can know whether ordering things in memory by breadth-first or depth-first traversal will make things faster, though? It seems like it depends whether it's more common to iterate over the array or access single elements of it. And when iterating over the array, how deep does the code go when looking at each array element?