Hi guys!
As we know that mixing have some penalties but some of instructions are very sweet. MOVHPD for example, Isn't some small penalty better than extra instruction to load/unload upper integer qword to regular register? Don't we have to modify MXCSR before use to prevent NaN exception or it checks for arithmetics only?

Personally I use many times instructions as movhlps, movlhps, movhps for integer data. I guess overall performance depend upon instuctions above and below mentioned. - If they are for floats or integer, too.

Yes. It depends upon many factors.

If you have a performance problem then profile/benchmark and use that to direct your attention to the problem areas.

I suggest you read Agner Fog's optimization manuals because it depends on the CPU in question.

Far more useful than benchmarking which (1) is only for CPUs you have access to and (2) you're unlikely to get the precise results someone else meticulously analyzed.

In short: it depends on the CPU. Some have extra latency when crossing the "domain" and some don't, for example. (usually newer ones don't so you can freely mix)