Yeah, the only unsafe code right now is using this trait to prevent some duplicate bounds checking when reading/writing to memories that wasn't being optimized by the rust compiler. Shouldn't be too much of a performance difference for code that's not too heavy on memory access, made about a 5-10% difference for the selfhosted benchmark if I remember correctly though.