The particular opcode was a very low cost one - most of the win came from memory savings (it’s basically float->int conversion with overridden round mode)