We spend a lot of time debating what non standard behaviours are safe to rely on because other software relies on it. Someone actually went and surveyed some major projects about what behaviours are being counted on.
Three results aren't as conclusive as I would have expected but even so some of then are still interesting. http://www.cl.cam.ac.uk/~pes20/cerberus/notes50-2015-05-24-survey-discussion.html
