I adjusted the code a bit more to look like the 16 coding including
restoring some very useful comments that had been lost, and pushed.
Thanks very much to Alexander and Noah for (independently) chasing
this down and reporting, testing etc, and Alvaro and Robert for review
comments.
(Passing thought: it's a bit weird that we need to zero pages at all
before restoring FPIs or initialising them. Perhaps there is some way
to defer marking the buffer valid until after the caller gets a chance
to initialise? Or something like that...)