They ain't gonna fix themselves... well, not yet.
This is an outcome for this summer's GSoC! - http://lists.llvm.org/pipermail/cfe-dev/2019-August/063195.html
Found one real crash, added a couple tests:
Addressed 17 style issues that aren't real crashes - mostly misuses of dyn_cast / getAs (i.e., "if the dyn_ part is really necessary here, then you crash with a null dereference a few lines below because you didn't check the result"). These changes indeed made the code cleaner and shorter.
Suppressed 4 false positives; i found these assertions/refactorings nice anyway, except one (which i'd prefer to have as a post-condition on a certain function but we didn't invent post-conditions yet in C++).
That leaves us with only one remaining false positive that isn't really actionable; i'll have to investigate.