Reported by Mikael Holmén on http://lists.llvm.org/pipermail/cfe-commits/Week-of-Mon-20180416/225349.html - a combination of rC329780 and rC300178 caused a performance regression that seemed to be a hang on the attached test case.
Reducing even further from 20 to 10 gives a ~15% further speed boost on the test, but i don't think it's worth it because such code is not common and therefore accuracy is more important.