Yes, but you wouldn't get to perform the IO proper, just get code data from source other than memory.
You can restrain yourself from using MOVX for several miliseconds just to get your IO work done when you have memory mapped IO and XRAM on the same ports. But what about code? Are there time windows between code read when you can perform generic IO with some external devices if you use the same ports for code? You need to fit in sub-instruction time frame so I'm really not sure how would that work.

An interesting idea for an educational project would be emulating a memory chip with a '51, with ability to arbitrarily change the contents of the memory while running...