Interesting to see that they're reusing what used to be the REP/REPNZ prefices for this purpose, rather than assigning a completely new (and likely 0F+) opcode. I suppose it's because the existing instruction decoder already recognises it as a prefix (they would just be ignored pre-Haswell.)
Exactly. As I understand it, HLE programs that use XACQUIRE and XRELEASE should run fine on older hardware. I suspect Intel's play here is to get the technology into mainstream compilers and libraries quickly, to make their Haswell and newer processors look good on benchmarks without special compiler flags.
Not a bad thing, just a compromise. It does restrict the feature set of HLE a little bit, and make some of its semantics a little awkward. Backward compatibility is obviously a good and useful thing. I think Intel did the right thing for their business and their customers, but probably not the right thing for the long-term prettiness of the assembly.