CLOS implementations have had to deal with the problem of extreme dynamism for a long time, so there has been quite a bit of optimization work associated with it. Polymorphic inline caches help a lot, for example.
That said, a good tracing JIT is probably better; but on the other hand, SBCL has profited from many years of optimization work on CMU CL. It would be interesting to see some real-world performance comparisons between SBCL and PyPy.
Wouldn't SBCLs type inference(+ type hints) help with this? I'm no compiler expert, but i think clojure(with the possible exception of multimethods, which are slow even on the JVM) would map well to CL(maybe even better than the JVM).
Basically it reflection (aka dynamic dispatch) is a major performance killer if you don't implement a tracing jit.