As Colin says towards the end, it seems like a combination of "all registers are callee saved" would help (combined __noinline__) but I agree you'd still have to find a way to make sure you could reliably zero out the xmm registers following the AESni runs. I would hope that an extreme dose of volatile with mm_load_ps and maybe some OR'ing into the result will defeat gcc and clangs ability to optimize away the clearing, but that's precisely what Colin is complaining about: you can't really express your demand in C and be certain it sticks.
About the AESNI runs, a comment on the post suggested to run some AESNI instructions on random numbers to clear (but not zero) the registers. With my very limited knowledge of all this, I can't shoot any holes in it. Is it insufficient? (for just the XMM register problem)
You'd still have to find a way to get the compiler not to eliminate the whole thing. So if you attempt to trash the registers with stuff but don't "mix it in" to the return value, the compiler is free to elude the whole thing.
Perhaps if you write it out to a dummy output parameter though, combined with noinline that would sufficiently defeat this.
memset_s, at least in C++, is guaranteed to run for this reason