Do you suspect the bug to be in clang (the front-end) itself, or in LLVM's optimizer or back-end? For the latter cases, bugpoint is included in the LLVM source tree and does a good job of narrowing things down, so long as you provide a program whose output should be deterministic (this may involve some work on your part if your program usually depends on data from the outside world). It is described here http://llvm.org/docs/Bugpoint.html . It may be useful for Clang bugs too, I'm not sure.
People may help you by suggesting strategies for finding a small test case that tickles the same problem if you post to the mailing list here http://lists.cs.uiuc.edu/mailman/listinfo/cfe-dev . I've posted a couple times and found people surprisingly responsive and helpful.
That looks like an interesting tool, I'll have to try it out. The program is deterministic given a fixed seed.
The bug depends on whether or not optimization is on, so it might be in the backend. What I've heard is that the C++ FE in particular still does have some troubles with miscompilations, so that is why I suspect it.
People may help you by suggesting strategies for finding a small test case that tickles the same problem if you post to the mailing list here http://lists.cs.uiuc.edu/mailman/listinfo/cfe-dev . I've posted a couple times and found people surprisingly responsive and helpful.