I really enjoyed the zero cost abstractions in Standard ML achieved by the whole-program optimizing MLton compiler (and I have a t-shirt to prove it). It's amazing how one can define clean, abstract, type-safe module hierarchies that still boil down to very, very efficient machine code that's just inches away from hand-written C code in terms of raw performance.

The zero runtime cost, of course, was paid in compilation time. With some tricks, it wasn't a big problem in practice, though.