Here's my first draft of the ARM atomics overhaul. The results should
be ideal for v7+, but I'm not entirely happey with the code for
supporting fallbacks. Right now the concept is to have multiple
fallback versions of the code laid out exactly like the kuser helper
page, and just select one of these. A relative address is used so that
the default of zero at early load time results in using safe "fake
atomics" that don't depend on kuser helper page or v6/v7 instructions.
But there's a lot of overhead loading the address to make the indirect
call that happens on each atomic. It probably doesn't perform that
bad, but it bloats the code size a good bit.

An alternative that might be less ugly is making direct calls to
external functons written in asm, and having the latter do the
dispatch to different versions of the code. I've tried this a few
times though and I wasn't too happy with where it was going.

Other ideas/comments are welcome!

Rich