1. Make the feature test work by excluding known-deficient processors, so
we don't have to maintain a complete list of all the processors that support
REV and REV16.
2. Don't abuse 'register' to get an effect similar to GCC's +l constraint,
but which was unnecessarily restrictive.
3. Fix __swap64md so _x isn't clobbered, breaking 64-bit swaps.
4. Make <byteswap.h> (which declars bswap_16 and friends) use <endian.h>
rather than <sys/endian.h>, so we get the machine-dependent implementations.
(git cherry-pick of 3cf53d1a7814e1520df09d24b009c16f4f27db0d)
570cb24498