I found a link to this now disappeared Clojure wiki page giving a pretty good explanantion of Clojure support for fast-math by Rich. I then found it again at http://clojure.org/old_news and on the original thread on the ML. The document illustrates the rationale behind some early Clojure optimisations that are still integral part of Clojure nowadays. The changes Rich describes here are mainly to enable the JVM JIT inlining of primitive operations and bring Clojure performance on primitive types (including arrays) on par with Java.
Below I also added the related IRC chat with some additional explanation.
You can consider this an extended comment to the work encompassing (roughly) git commit https://github.com/clojure/clojure/commit/5d9e87e8a30dfc99fe41c3be0a8ddc5ddd4e1d8d to