It occurred to me that Garrett D'Amore wrote in gmane.os.netbsd.devel.pkgsrc.user:
> Perry E. Metzger wrote:
>>
>> FYI, the kernel does not use floating point, and is not supposed to
>> use the floating point registers. I would suggest that there is no
>> point in setting -mmmx -msse etc in the kernel.
>>
> AFAIK, MMX and SSE have _nothing_ to do with floating point. 

They have a *lot* to do with floating point. They share registers with floating
point, and an SSE2 unit is actually able to do a pretty decent floating point
implementation of its own. I believe there's a gcc option to use SSE2 floating
point over i387 floating point, but I may be mistaken there.