It seems to me that MACHINE_NONCONTIG should just work, regardless.
If you have contiguous memory, okay, cool, everything _should_ be fine,
because you can always make contiguous spaces LOOK like noncontig spaces.
It's a bit harder to achieve it the other way around.

And if you don't have contiguous memory, well, the routines built to
handle it will handle it.

Of course, virtual reality and physical reality differ greatly :-), but
my premise is that, for example, on the SPARCstations, memory space is
a 64MB hole of which some fractions are filled with actually addressable
space, but if you have everything on an ss1 filled, for example,
that 64MB space is actually contiguous memory (or so it would seem,
physically, but see my above comment re realities).  Judging by the order
in which you must lay memory into a SS2 or an IPX, though, it would appear
that memory is non-contiguous no matter what.  Yet, in the ss1 case,
the pmapping is done via a noncontig algorithm, even though the memory
"appears" (to me) to be contiguous if all slots are completely full.

...I'm missing something again, aren't I?