Pointless... Because at the time you want to suck out performance of any of those, you would have to stick your guns to the provided macros or inline functions

Pointless... Because at the time you want to suck out performance of any of those, you would have to stick your guns to the provided macros or inline functions from Intel, ARM, IBM (PowerPC/AltiVec), etc. - and reimplement this using them - SSE2, AVX, whatever. This is not even funny.... And packing three doubles in structures - good programming technique? Okay.. That's enough (it might be good sometimes, but not always)