1. CGF now has fewer bytes of state (one pointer instead of a vector).
2. The generated code is determinstic, instead of getting labels in
   'map order' based on pointer addresses.
3. Clang now emits one 'indirect goto switch' for each function, instead
   of one for each indirect goto.  This fixes an M*N = N^2 IR size issue
   when there are lots of address-taken labels and lots of indirect gotos.
4. This also makes the default cause do something useful, reducing the
   size of the jump table needed (by one).

llvm-svn: 83952