r208664 - Patched clang to emit x86 blends as shufflevectors.

[Rafael Espíndola]

> +// FIXME: We should lower as many of these as possible to LLVM IR without
> +// intrinsics which will allow us to remove redundant intrinsics in the
> +// future. Taking care that we manage to optimize them afterwards and emit
> +// the actual instruction (or better code).
> +

What I meant was that we should at some point also lower a raw
__builtin_ia32_pblendw128 to a shuffle in clang. That way we can
delete the corresponding intrinsic from llvm.

Cheers,
Rafael