Hmm.  You could get the kind of flexibility & efficiency you're
   looking for using C++ templates...

Efficiency, perhaps.  Flexibility, definitely not.  Using C++
templates and inlining, I would end up duplicating much of the code
for each bus/device type.  The point of the `millicode' approach is
that the generic code only depends on the device itself, not the means
of accessing the device, and thus it can be shared.

Besides that, the thought of putting a C++ run-time in the kernel
makes me somewhat nauseous.