Incorrect Floating Point comparisons with NaN

And what if the "correct behavior" degrades the peformance of every floating point comparison in your program?