Investigate lower-subreg

Lower subreg has changed upstream to have a new costs based mechanism to determine when it's not worth splitting some larger types into smaller word type operations. However this has a dependence on the costs in the ARM backend and they would need to be carefully audited along with the way in which this new algorithm behaves. There are a couple of cases where this helps especially with the fix for PR48941 in upstream bugzilla but it makes things worse for some other code sequences involving other modes of copies. This looks like a meaty chunk of work that could do with some more investigation