rwessel (robertwessel.delete@this.yahoo.com) on January 14, 2020 3:12 pm wrote:

>
> That's true, actual 16-bit accesses, which did happen with some frequency, would have needed more
> logic. The practice was very prevalent on things like display adapters (at least for the video
> buffer). On the other hand, many of the 8-bit I/O devices really were a collection of 8-bit ports,
> and there weren't often opportunities to do a 16-bit access to modify two ports.
>
> Still, not a huge thing, and IBM could have defined it away to an extent, by simply asserting that
> 16-bit accesses to 8-bit devices weren't supported (and for things like display adapters, actually
> making the memory 16-bits wide). Or do what they did on the AT to emulate that behavior.
>
>

The 6845 CRT controller chip was addressed as a "command" port and a "data" port (at address+1)
So, to write a register you wrote the address to the command port, and the value to the data port.
With a 16 bit OUT instruction, you could easily do one instruction that updated both, in the right order.
If you used the OUTSW instruction (available with the 286), you could update the whole chip with one instruction (excluding setup). Lots of games used this technique, so they really did have to turn one 16 bit I/O bus cycle into two 8 bit cycles for compatibility.