So when do we get started?
If he's talking about the overhead added by the physical, link, network and transport layer, I can't help but think he's a moron. The physical layer adds no overhead. The link layer adds no real overhead as it changes from node to node anyway. And the network and transport layer add what, 64 bytes of overhead?
What would it be more than just an AND to guess character width and detect surrogates?
actually it would be extremely complicated.
Imagine the following scenario: you start reading a UTF-8 string. You get a full word from the main memory (MM).
You read the first bytes: the first character is 3 bytes long. Okay. You process the first three bytes. Now you check the fourth and last byte you've loaded. The next character is 2 bytes long. So you have to shift your current word by 24 bits, then load another word, then remember you only have to process the highest 16 bits of the word you currently have in your register.
I see. That's still not that difficult. There's an instruction which does most of the work already (SHLD), and if you're proposing a fast way to read chunks of data from streams while reading in words, that would probably be better done as generic instructions for that purpose (so that compression algorithms could take advantage of it).
I think this is the first image on the WWW.
...
I wish they would account for Africa-tier speeds and hardware.