The main guy behind CompCert is Xavier Leroy (already known for OCaml).
I still had a jaw-drop moment the first time I heard about this; Coq is not a pleasant language, and proving this thing must have been a nightmare.
A C compiler doesn't have to promise much of anything about the behavior of the code produced for a program with undefined behavior.
This statement could be vacuous for a program p with undefined behavior if (Cstrategy.exec_program p beh) isn't true for any behavior beh, or only true for a behavior which makes (not_wrong beh) false.
If they don't always, then they actually give you a guarantee for some bad programs as well. Nothing in the spec says running a bad program has to make demons fly out your nose.
http://cristal.inria.fr/~xleroy/
I keep this link in a feel-good folder, next to Bellard's page. Sometimes I need a feel-good break after reading US-centric news.