Implement _mm_getcsr, _mm_setcsr, _mm_sfence

[nominolo]

This PR implements the following sse intrinsics:

• _mm_getcsr
• _mm_setcsr
• _MM_GET_EXCEPTION_STATE
• _MM_SET_EXCEPTION_STATE
• _MM_GET_EXCEPTION_MASK
• _MM_SET_EXCEPTION_MASK
• _MM_GET_ROUNDING_MODE
• _MM_SET_ROUNDING_MODE
• _MM_GET_FLUSH_ZERO_MODE
• _MM_SET_FLUSH_ZERO_MODE
• _mm_sfence

However, all the _MM_* helpers are macros in C, but functions in Rust, so I turned them into lower-case names. See #84 for the open policy discussion on this.

Testing this turned out to be a bit tricky because almost all of the small unit tests seem to get constant folded (see #86).

The _mm_sfence intrinsic doesn't quite fit in thematically. I can put it into a separate PR if that's preferred.

[nominolo]

Hm, the test failure is likely due to constant folding (https://travis-ci.org/rust-lang-nursery/stdsimd/jobs/281739237). Do Travis CI and appveyor use different LLVM versions? The rustc versions seem to be the same.

[alexcrichton]

I wonder, if this is setting a CPU register, does it work in multithreaded programs? Will this need to be tested as a separate executable? Other than that using functions for macros seems fine, and maybe test::black_box could be used to inhibit constant folding?

[nominolo]

_mm_setcsr must only affect the current thread (otherwise, that's a CPU or Kernel bug). I wrote tests in such a way that there are no asserts between modifying and restoring the state, so even if multiple tests are run in the same thread, it should be safe. I don't know if Rust uses a new thread per test.

[alexcrichton]

Awesome, thanks @nominolo!