UTF-8 history (2003)
cl.cam.ac.uk
cl.cam.ac.uk
Rob's initial claim is that Ken came up with UTF-8 entirely from scratch, without even looking at IBM's proposal.
But Ken's oldest document from Sep 2, 1992 has his changes simply appended after the original FSS-UTF document from IBM (starting at "We define 7 byte types", as mentioned in the email), and notes the changes from original spec.
Then the final document that he sent out (Sep 8) is basically the FSS-UTF document with his changes applied (this is also mentioned in the email!).
There are two changes, basically:
1. Use 10 instead of 1 as the prefix for continuation bytes, so you can synchronize from an arbitrary location.
2. Once the bits are reassembled, use the value as-is instead of adding a bias constant, which simplifies the code at the cost of a tiny bit of packing efficiency.
Given the documentation provided, I would say that the fairest description of the development of UTF-8 is that IBM came up with the initial design, and Plan 9 made two improvements on it to produce the final UTF-8 design.
This means UTF-8 encoded values of more than 4 bytes can never represent a valid unicode codepoint even if they produce a valid 32 bit numerical value.
I'm of two minds about this. On the one hand, such an ability is pretty useless in modern systems filled with checksums at every stage, and reduces the bit efficiency of UTF-8. On the other hand, this was the only valid-y reason at the time to rewrite a UTF implementation from scratch.
If not for that requirement, we would have just had UTF-8 implemented as regular VLQs.
Edit: Actually, now that I think about it, VLQ already does satisfy the synchronization requirement. Just scan for the next cleared high bit. At most 1 character consumed, and far less bit wastage.
There are other useful properties of UTF-8 encoding. For one, it is easy to identify invalid UTF-8 sequences, and valid UTF-8 sequences are unlikely to appear in written text encoded as e.g. ISO 8859-1. This must have been more useful in the past when fixed 8-bit encodings were more common. With a simple VLQ your guess whether a stream complies to your encoding won't be nearly as informed, and the means to provide a fallback is diminished. Allowing graceful transition from ASCII and extended ASCII encodings was a very important property for the sake of adoption.
It's also useful that any non-ASCII sequence has the msbit set through the whole sequence. You can easily discard all sequences that can not be displayed as ASCII by throwing bytes >= 0x80, or for example replace every contiguous sequence of >= 0x80 bytes as question marks in an ASCII-only display system and still display all ASCII compatible sequences perfectly.
EDIT: It's also very much not the case that the days of starting read mid-stream are over.
Imagine a text editor displaying a large CSV file. Wouldn't you mind having to read everything betseen two locations if you jump forward from the one to the other? Or read everything from the start if you jumping backwards? Even if the text editor stores its own synchronization points, it has to read the file completely at least once, which can be annyoing for very large files.
Also, many of the simpler text tools that only look for ASCII bytes wouldn't work. For example, printing the last lines.
Also, I don't know that other sytem, but what about robustness if there's one bad byte somewhere in the file?
There is a requirement that 7-bit ASCII character codes would not appear as a part of non-ASCII character encoding, so NULL-terminated strings, slash path separation and other similar issues could be handled by existing charset-independent code. That would not work with regular VLQs.
The only remaining mystery then would be why they bothered encoding it like this:
0xxxxxxx
110xxxxx 10xxxxxx
1110xxxx 10xxxxxx 10xxxxxx
11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
0xxxxxxx
11xxxxxx 11xxxxxx ... 10xxxxxx
0xxxxxxx
100xxxxx 1xxxxxxx
101xxxxx 1xxxxxxx 1xxxxxxx
110xxxxx 1xxxxxxx 1xxxxxxx 1xxxxxxx
111xxxxx 1xxxxxxx 1xxxxxxx 1xxxxxxx 1xxxxxxx4) The first byte should indicate the number of bytes to follow in a multibyte sequence.
This is actually a pretty smart requirement for efficiency. This way the parser can know if the input bytes it has constitute a UTF-8 sequence by only looking at the first byte. That saves a lot of unnecessary processing.
Unfortunately the wording here breaks that advantage: https://en.cppreference.com/w/cpp/string/multibyte/mbrtoc16
if the next n bytes constitute an incomplete, but so far valid, multibyte character. Nothing is written to
The "so far valid" forces implementations to still look at the other bytes, which undoes the optimization in UTF-8. I suspect somebody was asleep at the wheel there.
Regarding your question about the wasted bits: I don't think it matters much. It certainly does not for English text like we exchange here, where the important case is the 0-0x7f case which UTF-8 handles optimally.
Your maximum compactness variant means if you start in the middle of a sequence, you can't tell you are in the middle and not at the start. UTF-8 is doing this well in my opinion. You can seek anywhere in a file and then move forwards or backwards till the beginning of a sequence without fear of misparsing the middle of a correct sequence as a different sequence.
This feels a lot like mixing transport layer metadata into the data format, potentially giving a small processing performance benefit at the cost of huge data wastage when certain languages are encoded.
Is there a language that consistently uses codepoints with more than 2 bytes?
It bothers me a bit that UTF-8 is not an infinitely extendable encoding. But that also isn't an important objection, because it is finite, but huge.
There are definitely (small) communities using scripts that lie entirely in the SMP. For example, Mru, Adlam, Takri, Pracalit, Miao, Wancho, etc. Most of these are either historic scripts that have mostly been supplanted by unified ones (esp. Devanagari) but retain usage in some areas, or languages that did not have a pre-colonial writing system that are attempting to reclaim cultural identity with a new script.
But yes, I don't think there are major communities that consistently do so. My anecdata from a few Mandarin- and Japanese-speaking friends is that SIP characters rarely occur.
Really if anything, emoji obsessives, mathematicians using bold/fraktur characters, and historical linguists/anthropologists would have the biggest savings.
https://en.wikipedia.org/wiki/Mru_language#Alphabet https://en.wikipedia.org/wiki/Adlam_script https://en.wikipedia.org/wiki/Takri_script https://en.wikipedia.org/wiki/Pracalit_script https://en.wikipedia.org/wiki/Pollard_script
Kanji, hiragana, katakana etc. At least 3 bytes with some 4-byte sequences. This compares unfavorably to for example Shift-JIS.
If there is no length encoding, start, middle, end and ASCII byte encodings must not overlap to support correct decoding of any subslice of a document.
If there is a length encoding on the start, then start, middle and ASCII must not overlap.
same with the second encoding, if so happens that one 1xxxxxxx takes the value of 11010101 how would you know it's a multibyte start or a continuation?
basically in the current solution if you read the head of a multibyte character you know it's a valid head, if you read the head of a multibyte in the proposed encoding you can't know if it's valid.
My point is that embedding transport level metadata into the data format seems like a poor tradeoff because of the sheer inflation potential of the data encoding (potentially 10%), when a single guard byte per field would solve the problem of first character truncation detection.
> .. in the event that your medium has no error detection ... add a zero synchronization byte
How would such encoding deal with non-utf8-safe editors, copy-pasting, programs truncating, then inserting previously broken sequences, etc?
Encoding obviously can't fix all errors, but it is quite useful if broken sequences are obviously broken and non-broken sequences remain valid when handling text in non-aware/non-safe applications.
I think in UTF8 two splices can generate a random character, but in a characters + splice combination, the character remains recognizable in any order and combination and a lone splice is also recognizable as an error.
The idea I heard was that to always code in 4 bytes blocks, and use some form of delta encoding. Some variations allow for less than OlogN character position search. And given that you can feed 32 wide data into NEON/SSE, and block are always 32 bit aligned, you can have that working faster than UTF-8
While not technically equivalent, synchronized byte stream allows a backward scanning that is beneficial for many Unicode-related algorithms. In fact synchronization is the simplest way to do that.
What was then the advantage of UTF8 over VLQ?
I cannot find the words to describe how much I admire these guys (ken, rob, drm, etc).
I think Rob Pike may have one of the coolest email addresses in the world (r@google.com).
such as examples/UTF-8-demo.txt a beautiful demo file for testing terminal emulators
or the wcwidth.c file I discovered was the origin of the same c function found in all OS's, i forked & maintain in python form as the public "wcwidth" module,
anyway just wanted to point out the treasure trove of the parent folder!
Some might find it more digestable than the linked page.