And in particular, sixteen ADC instructions in four chains shouldn't run any slower than sixteen ADD instructions (assuming four execution ports and a large eno

And in particular, sixteen ADC instructions in four chains shouldn't run any slower than sixteen ADD instructions (assuming four execution ports and a large enough reorder buffer), but a: Haswell sucks, and b: even on properly-designed CPUs, making parallelization easier is almost always some kind of win, because general-purpose CPUs always have dispatch overhead.