During dictionary generation, rootcling and rootcling_stage1 seem to randomly crash when generating dictionaries (ususally MultiProc, but sometimes also when generating libThread - which are both ususally the first dictionaries to get generated for each rootcling and rootcling_stage1). The stackstrace also seems to be at random places (from some internal LLVM function to some random function in libc). However, the stacktrace only changes when cleaning the build-directory and rebuilding, so rerunning the crashing rootcling command always yields the same result.
This happens for me on my Arch Linux installation, on a fresh ubuntu 16.04 virtual machine and on the Fedora 25. However, it doesn't happen on my other Arch Linux machine.
Another observation is that when running Mozilla's rr on the failing rootcling command, it reports that the process image is not initialized (e.g. global constructors aren't run), but I'm not sure if this is a false positive or not.
I've checked out the root master from now and from around two weeks ago, and both randomly crash.
I've created a minimal virtual machine that only contains root source code, the failing build and some LXDE desktop here (you can import this into VirtualBox. See the text file on the desktop on how to reproduce the bug. Password for the main user is 'test123'):