Continued work on 64-bit shifts. I've tidied up most of the nastiness I had in my first attempt at this work, unified the logic into fewer places, and fixed the one logic error that I knew about. I had a mysterious reload bug that just wouldn't go away, but having stared at it for some time I've decided that what I was doing wasn't the best option in any case, and the new way seems to work. I still have the problem that register-allocation prefers to allocate the values in core-registers, even when it would seem to me to work better in NEON. This maybe because the shift amount must be negated in some cases, and there isn't a neon instruction for that.
Continued trying to use LAVA to do builds. Unfortunately this seems to have hit against a lava-test bug. Filed as lp:915915.