ARM NEON optimization for float->int conversion. Not optimal, but since lrintf is incredibly slow on ARM it is > 10x faster than the old code. A fallback solution that (incorrectly) defines lrintf(x) as (int)(x) might make sense to avoid these kind of issues for the pure C code, however we would still need to know whether lrintf is slow or not. Though maybe the better solution was if all architectures provided a non-braindead implementation of lrintf.