He also eliminated all the stack operations from machineForth except DUP, DROP, and OVER. Not even SWAP!
As I understand it, the A register, introduced in the MuP21, solved some problems having to do with on-die fanout, fanin, and path length that limited the performance of the Novix NC4016 and the Harris RTX 2000. It also allows you to do operations like memset and memcpy without having to fetch and store the pointers to RAM (at fixed addresses!) for every byte/word read or written, which improves performance dramatically, makes the code a lot more readable (especially with its incrementing addressing modes), and makes the code more compact.
None of these are problems that exist in the PostScript world.
You can't use the A register in the MuP21 or F18A as a local variable, because it isn't local; it's shared with whatever subroutines you call and whatever subroutine called you. Moreover, the A register is the only way to read and write memory, so they can't avoid overwriting it if they need to read or write memory.
[I've programmed in PostScript, and I've written a little Forth, but I've never programmed any of the chips mentioned above.]