Amigas evidently also get to suffer the ills of ISA.

> A pci-to-isa bridge is part of [CHRP] ...
> So we (future) CHRP users will be interested in isa support too, and 
> these problems, and thus it's a bit bigger than machine-dep. Though I do 
> think it's probably just the isa code, not the whole OS, which is ill.

The actual devices you mentioned as being on the ISA bus by default do not
use ISA busmaster DMA, hence are not limited (necessarily) to 16MB (the
soundblaster, at any rate, requires an external DMA engine, and they'll
presumably just spec the functionality of the PC DMA chip, which is a piece
of junk but at least allows more than 24 address lines into memory).  Also,
the PCI bus is pretty feature-rich, and though I kind of doubt it, it's
at least conceivable that PCI-to-ISA bridges could include mapping registers,
meaning that the model "isa == bounce-buffers" would not be correct.

Hence I repeat my suggestion that someone in "core" give consideration to
stealing one of NT's few good ideas ;-).