Having done some network code on the 16-bit Xerox Alto, it's a big pain if you can't access bytes directly. If you want to deal with characters, you need to do shifts and AND operations, depending if you want the odd byte or the even byte. It's very annoying.

Do you mean characters as in real languages? Because then bytes don't matter, but codepoints, which can be any length.

Obviously dealing with legacy formats that are bit or byte aligned should still be supported using char arrays, it just would require compiler assistance for pointers.

Well let's just switch to 32bit unicode characters and drop all this UTF8 stuff. But even then we would addressing 32 bit words with a 64bit machine.

I don't know if you lived through the early RISC days, but it was really a bit of a bother when dealing with things that weren't words. but sure, its wasn't that bad. I just get really tired of all the shift and mask stuff...seems like we should be able to do better. I guess you could paper over it with sufficient language/runtime support

It's really convenient to pass around pointers to a single GPIO pin on embedded systems that have a special region of memory where each byte corresponds to one pin rather than the usual 8. Atomic access to bits would also be convenient.

We have atomic bitwise operations already (look at glibc's mutex implementation), and the unit atomic operations work on is a 64-byte cache line. Cache lines are useful because reading 64 bytes isn't really more expensive but it improves sequential memory access by a lot.