Rust's iterators are really nice. As far as I can tell, the idea is to do only one check on each iteration, which doubles as an array bounds check (for memory safety) and a loop condition check (for loop termination). That's much better than the usual "for" loop with an index into a checked array, which uses one check for loop termination, and two more checks upon array indexing. Why don't more languages do that? It seems like a simple enough idea, I blogged about it sometime ago.
In the case of Java, I believe it's because the underlying JVM only supports bounds-checked array indexing. So even though the iterator APIs in Java could in theory eliminate the bounds checks, the high-level information is compiled away and the JVM has to prove on its own that the bounds check is not necessary.
I wonder if the language can support similarly fast access to user-defined data structures. Though maybe it's not a very interesting question, if arrays are built in. Most user-defined data structures are already their own iterators in some sense. For example, iterating over an ML-style linked list obviously requires only one check per iteration.
