As of last friday SBCL/AMD64 can be used to compile SBCL (with the tiny caveat that purification has to be disabled in the host). The codebase is >200KLOC of lisp, so this is a pretty good test of the whole system. I also tried running Paul Dietz's ansi-tests and was rewarded with 147 out of 18248 test failures. This is now down to 60 failures compared to 51 for a SBCL/x86 of the same vintage.
Also, I believe we have another monthly Helsinki Lisp meet coming up this tuesday.