Follow up for a problem described in https://reviews.llvm.org/D57230
I think the ClusterAnalysis is fundamentally flawed at this point and it does not support non-base regions properly. The reason we could get away with this so far is that it is very rare to have non-base regions there.
I made an attempt to fix this and I successfully fixes the false positive described by Artem in https://reviews.llvm.org/D57230.
While I am not completely sure yet whether this is the most desirable solution, I wanted to share the work in progress solution to get some feedback about the directions.