Building llvm-gcc for specific processors

I'm working on revamping the validators a bit to exercise more of LLVM. I'd
like to build an LLVM tuned to specific microarhchitectures. One for
nehalem, one for northwood, etc. This way we can test various SSE levels and
so forth.

I understand that some of the LLVM tests are target-specific and contain
targeting information in the tests. But especially for the big multi-source
tests I'd really like to push LLVM to the limit.

Is there some magic to pass to the llvm-gcc configure script to make certain
-mcpu / -mtune flags default?