Comment by quectophoton
13 hours ago
Having the continuation bytes always start with the bits `10` also make it possible to seek to any random byte, and trivially know if you're at the beginning of a character or at a continuation byte like you mentioned, so you can easily find the beginning of the next or previous character.
If the characters were instead encoded like EBML's variable size integers[1] (but inverting 1 and 0 to keep ASCII compatibility for the single-byte case), and you do a random seek, it wouldn't be as easy (or maybe not even possible) to know if you landed on the beginning of a character or in one of the `xxxx xxxx` bytes.
Right. That's one of the great features of UTF-8. You can move forwards and backwards through a UTF-8 string without having to start from the beginning.
Python has had troubles in this area. Because Python strings are indexable by character, CPython used wide characters. At one point you could pick 2-byte or 4-byte characters when building CPython. Then that switch was made automatic at run time. But it's still wide characters, not UTF-8. One emoji and your string size quadruples.
I would have been tempted to use UTF-8 internally. Indices into a string would be an opaque index type which behaved like an integer to the extent that you could add or subtract small integers, and that would move you through the string. If you actually converted the opaque type to a real integer, or tried to subscript the string directly, an index to the string would be generated. That's an unusual case. All the standard operations, including regular expressions, can work on a UTF-8 representation with opaque index objects.
PyCompactUnicodeObject was introduced with Python 3.3, and uses UTF-8 internally. It's used whenever both size and max code point are known, which is most cases where it comes from a literal or bytes.decode() call. Cut memory usage in typical Django applications by 2/3 when it was implemented.
https://peps.python.org/pep-0393/
I would probably use UTF-8 and just give up on O(1) string indexing if I were implementing a new string type. It's very rare to require arbitrary large-number indexing into strings. Most use-cases involve chopping off a small prefix (eg. "hex_digits[2:]") or suffix (eg. "filename[-3:]"), and you can easily just linear search these with minimal CPU penalty. Or they're part of library methods where you want to have your own custom traversals, eg. .find(substr) can just do Boyer-Moore over bytes, .split(delim) probably wants to do a first pass that identifies delimiter positions and then use that to allocate all the results at once.
You usually want O(1) indexing when you're implementing views over a large string. For example, a string containing a possibly multi-megabyte text file and you want to avoid copying out of it, and work with slices where possible. Anything from editors to parsing.
I agree though that usually you only need iteration, but string APIs need to change to return some kind of token that encapsulates both logical and physical index. And you probably want to be able to compute with those - subtract to get length and so on.
4 replies →
> PyCompactUnicodeObject was introduced with Python 3.3, and uses UTF-8 internally.
UTF8 is used for C level interactions, if it were just that being used there would be no need to know the highest code point.
For Python semantics it uses one of ASCII, iso-8859-1, ucs2, or ucs4.
This is Python; finding new ways to subscript into things directly is a graduate student’s favorite pastime!
In all seriousness I think that encoding-independent constant-time substring extraction has been meaningful in letting researchers outside the U.S. prototype, especially in NLP, without worrying about their abstractions around “a 5 character subslice” being more complicated than that. Memory is a tradeoff, but a reasonably predictable one.
Your solution is basically what Swift does. Plus they do the same with extended grapheme clusters (what a human would consider distinct characters mostly), and that’s the default character type instead of Unicode code point. Easily the best Unicode string support of any programming language.
Indices into a Unicode string is a highly unusual operation that is rarely needed. A string is Unicode because it is provided by the user or a localized user-facing string. You don't generally need indices.
Programmer strings (aka byte strings) do need indexing operations. But such strings usually do not need Unicode.
They can happen to _be_ Unicode. Composition operations (for fully terminated Unicode strings) should work, but require eventual normalization.
That's the other part of the resume UTF8 strings mid way, even combining broken strings still results in all the good characters present.
Substring operations are more dicey; those should be operating with known strings. In pathological cases they might operate against portions of Unicode bits... but that's as silly as using raw pointers and directly mangling the bytes without any protection or design plans.
Variable width encodings like UTF-8 and UTF-16 cannot be indexed in O(1), only in O(N). But this is not really a problem! Instead of indexing strings we need to slice them, and generally we read them forwards, so if slices (and slices of slices) are cheap, then you can parse textual data without a problem. Basically just keep the indices small and there's no problem.
> If you actually converted the opaque type to a real integer, or tried to subscript the string directly, an index to the string would be generated.
What conversion rule do you want to use, though? You either reject some values outright, bump those up or down, or else start with a character index that requires an O(N) translation to a byte index.
"Unicode" aka "wide characters" is the dumbest engineering debacle of the century.
> ascii and codepage encodings are legacy, let's standardize on another forwards-incompatible standard that will be obsolete in five years > oh, and we also need to upgrade all our infrastructure for this obsolete-by-design standard because we're now keeping it forever
VLQ/LEB128 are a bit better than the EBML's variable size integers. You test the MSB in the byte - `0` means it's the end of a sequence and the next byte is a new sequence. If the MSB is `1`, to find the start of the sequence you walk back until you find the first zero MSB at the end of the previous sequence (or the start of the stream). There are efficient SIMD-optimized implementations of this.
The difference between VLQ and LEB128 is endianness, basically whether the zero MSB is the start or end of a sequence.
It's not self-synchronizing like UTF-8, but it's more compact - any unicode codepoint can fit into 3 bytes (which can encode up to 0x1FFFFF), and ASCII characters remain 1 byte. Can grow to arbitrary sizes. It has a fixed overhead of 1/8, whereas UTF-8 only has overhead of 1/8 for ASCII and 1/3 thereafter. Could be useful compressing the size of code that uses non-ASCII, since most of the mathematical symbols/arrows are < U+3FFF. Also languages like Japanese, since Katakana and Hiragana are also < U+3FFF, and could be encoded in 2 bytes rather than 3.
Unfortunately, VLQ/LEB128 is slow to process due to all the rolling decision points (one decision point per byte, with no ability to branch predict reliably). It's why I used a right-to-left unary code in my stuff: https://github.com/kstenerud/bonjson/blob/main/bonjson.md#le...
The full value is stored little endian, so you simply read the first byte (low byte) in the stream to get the full length, and it has the exact same compactness of VLQ/LEB128 (7 bits per byte).
Even better: modern chips have instructions that decode this field in one shot (callable via builtin):
https://github.com/kstenerud/ksbonjson/blob/main/library/src...
After running this builtin, you simply re-read the memory location for the specified number of bytes, then cast to a little-endian integer, then shift right by the same number of bits to get the final payload - with a special case for `00000000`, although numbers that big are rare. In fact, if you limit yourself to max 56 bit numbers, the algorithm becomes entirely branchless (even if your chip doesn't have the builtin).
https://github.com/kstenerud/ksbonjson/blob/main/library/src...
It's one of the things I did to make BONJSON 35x faster to decode/encode compared to JSON.
https://github.com/kstenerud/bonjson
If you wanted to maintain ASCII compatibility, you could use a 0-based unary code going left-to-right, but you lose a number of the speed benefits of a little endian friendly encoding (as well as the self-synchronization of UTF-8 - which admittedly isn't so important in the modern world of everything being out-of-band enveloped and error-corrected). But it would still be a LOT faster than VLQ/LEB128.
We can do better than one branch per byte - we can have it per 8-bytes at least.
We'd use `vpmovb2m`[1] on a ZMM register (64-bytes at a time), which fills a 64-bit mask register with the MSB of each byte in the vector.
Then process the mask register 1 byte at a time, using it as an index into a 256-entry jump table. Each entry would be specialized to process the next 8 bytes without branching, and finish with conditional branch to the next entry in the jump table or to the next 64-bytes. Any trailing ones in each byte would simply add them to a carry, which would be consumed up to the most significant zero in the next eightbytes.
[1]:https://www.intel.com/content/www/us/en/docs/intrinsics-guid...
2 replies →
That's assuming the text is not corrupted or maliciously modified. There were (are) _numerous_ vulnerabilities due to parsing/escaping of invalid UTF-8 sequences.
Quick googling (not all of them are on-topic tho):
https://www.rapid7.com/blog/post/2025/02/13/cve-2025-1094-po...
https://www.cve.org/CVERecord/SearchResults?query=utf-8
I was just wondering a similar thing: If 10 implies start of character, doesn't that require 10 to never occur inside the other bits of a character?
Generally you can assume byte-aligned access. So every byte of UTF-8 either starts with 0 or 11 to indicate an initial byte, or 10 to indicate a continuation byte.
10 never implies the start of a character; those begin with 0 or 11.
UTF-8 encodes each character into a whole number of bytes (8, 16, 24, or 32 bits), and the 10 continuation marker is only at the start of the extra continuation bytes, it is just data when that pattern occurs within a byte.
You are correct that it never occurs at the start of a byte that isn’t a continuation bytes: the first byte in each encoded code point starts with either 0 (ASCII code points) or 11 (non-ASCII).
It's not uncommon when you want variable length encodings to write the number of extension bytes used in unary encoding
https://en.wikipedia.org/wiki/Unary_numeral_system
and also use whatever bits are left over encoding the length (which could be in 8 bit blocks so you write 1111/1111 10xx/xxxx to code 8 extension bytes) to encode the number. This is covered in this CS classic
https://archive.org/details/managinggigabyte0000witt
together with other methods that let you compress a text + a full text index for the text into less room than text and not even have to use a stopword list. As you say, UTF-8 does something similar in spirit but ASCII compatible and capable of fast synchronization if data is corrupted or truncated.
This is referred to as UTF-8 being "self-synchronizing". You can jump to the middle and find a codepoint boundary. You can read it backwards. You can read it forwards.
also, the redundancy means that you get a pretty good heuristic for "is this utf-8". Random data or other encodings are pretty unlikely to also be valid utf-8, at least for non-tiny strings
Wouldn't you only need to read backwards at most 3 bytes to see if you were currently at a continuation byte? With a max multi-byte size of 4 bytes, if you don't see a multi-byte start character by then you would know it's a single-byte char.
I wonder if a reason is similar though: error recovery when working with libraries that aren't UTF-8 aware. If you slice naively slice an array of UTF-8 bytes, a UTf-8 aware library can ignore malformed leading and trailing bytes and get some reasonable string out of it.
It’s not always possible to read backwards.
Okay so you seek by 3 less bytes.
Or you accept that if you're randomly losing chunks, you might lose an extra 3 bytes.
The real problem is that seeking a few bytes won't work with EMBL. If continuation bytes store 8 payload bits, you can get into a situation where every single byte could be interpreted as a multi-byte start character and there are 2 or 3 possible messages that never converge.
This isn't quite right. In invalid UTF8, a continuation byte can also emit a replacement char if it's the start of the byte sequence. Eg, `0b01100001 0b10000000 0b01100001` outputs 3 chars: a�a. Whether you're at the beginning of an output char depends on the last 1-3 bytes.
> outputs 3 chars
You mean codepoints or maybe grapheme clusters?
Anyways yeah it’s a little more complicated but the principle of being able to truncate a string without splitting a codepoint in O(1) is still useful
Yah, I was using char interchangeably with code point. I also used byte instead of code unit.
> truncate a string without splitting a codepoint in O(1) is still useful
Agreed!
> Having the continuation bytes always start with the bits `10` also make it possible to seek to any random byte, and trivially know if you're at the beginning of a character or at a continuation byte like you mentioned, so you can easily find the beginning of the next or previous character.
Given four byte maximum, it's a similarly trivial algo for the other case you mention.
The main difference I see is that UTF8 increases the chance of catching and flagging an error in the stream. E.g., any non-ASCII byte that is missing from the stream is highly likely to cause an invalid sequence. Whereas with the other case you mention the continuation bytes would cause silent errors (since an ASCII character would be indecipherable from continuation bytes).
Encoding gurus-- am I right?
[dead]
so you replace one costly sweeping with a costly sweeping. i wouldn't call that an advantage in any way over junping n bytes.
what you describe is the bare minimum so you even know what you are searching for while you scan pretty much everything every time.
What do you mean? What would you suggest instead? Fixed-length encoding? It would take a looot of space given all the character variations you can have.
UTF-16 is both simpler to parse and more compact than utf-8 when writing non-english characters.
UTF-8 didn't win on technical merits, it won becausw it was mostly backwards compatible with all American software that previously used ASCII only.
When you leave the anglosphere you'll find that some languages still default to other encodings due to how large utf-8 ends up for them (Chinese and Japanese, to name two).
14 replies →
> so you can easily find the beginning of the next or previous character.
It is not true [1]. While it is not UTF-8 problem per se, it is a problem of how UTF-8 is being used.
[1] https://paulbutler.org/2025/smuggling-arbitrary-data-through...
Parent means “character” as defined here in Unicode: https://www.unicode.org/versions/Unicode17.0.0/core-spec/cha..., effectively code points. Meanings 2 and 3 in the Unicode glossary here: https://www.unicode.org/glossary/#character