This is pretty good actually considering the low hanging optimizer optimizations left and that the alternative is JS which generally performs 2-10x slower.

I think vectorization support will narrow the aggregate difference here as a lot of SPEC benefits from auto vectorization if I recall correctly.