As far as I understand, the SBCL compiles each function separately, and can fall back to very slow codepaths in doing so. Getting peak performance out of Common Lisp takes some effort.