3 ms·

The world of text encodings is pretty insane, especially when you start getting into the realm of what seems like endless variations on multi-level encodings, l

• ben-schaaf 5y ago
  I recently implemented GB18030, which is a 1,2 or 4 byte encoding with a giant code page covering the entirety of unicode.
• dhosek 5y ago
  Back in the 90s I had a 4-inch binder stuffed with every encoding registered with ECMA. I'll take UTF-8 with all of its warts over the mess that was the pre-Unicode encoding hell.
  • kmeisthax 5y ago
    UTF-8 actually isn't the warty bit of Unicode. It's actually fairly reasonable[0], with the only downside being that you can't efficiently index strings by integer. If you exclusively use UTF-8 everywhere in your application, you will remain wart free. UTF-16 is the warty bit. Why? Because it was designed to be a fixed-width encoding, under the assumption that 65,536 code points would cover all languages we needed. Well, spoiler alert: this wasn't enough, even with the mess that was UniHan. So they carved out a chunk of those code points from the 16-bit part of Unicode, called them "surrogate code points", and used them to extend UTF-16 into a variable-width encoding. Thus, they could have 16 planes of 65,536 characters each. They weren't valid characters beforehand, so obviously we could just repurpose them as flags for "astral" characters outside the Basic Multilingual Plane. Simple, right? Except that plenty of code to work with UTF-16 strings already existed, and assumed UTF-16 was a fixed-width encoding. If you chucked an astral character at it, it'd think it's two rather than one. It'll also happily split or rearrange astral characters into invalid sequences of codepoints. Furthermore, if you ask these systems for UTF-8, it'll happily encode the surrogates, leading to two nested layers of variable encodings. Oh, and the two biggest users of UTF-16 were Windows NT and JavaScript. So we're absolutely stuck with fixed-length UTF-16 and the mangled WTF-8[1] encodings they will output until the end of time. Furthermore, if we ever do fill up the astral planes, we'll have to define superastral characters, which means having supersurrogate code points that will break modern UTF-16 again and make WTF-8 handling even more complicated. [0] Unless you use MySQL, where someone decided to break UTF-8 early on and now there's "utf8mb4" to un-break it. [1] UTF-8 which contains encoded UTF-16 surrogates, as defined in https://simonsapin.github.io/wtf-8/ https://simonsapin.github.io/wtf-8/
    • bruce511 5y ago
      You are completely right. Utf-16 is either 2 or 4 bytes long per code point (not per character) but a lot of older programs (and older programmers) see it as a "2 byte per code point" string. There is a 2 byte per code point encoding, its called UCS2 so its important to make sure when something says utf-16 they don't mean UCS2. I am not familiar enough with UCS2 to know if it also supports multiple code points per char - but I expect it does - so the mental model of 2 bytes per char breaks down even there.
      • kmeisthax 5y ago
        It depends on which UCS-2 you're talking about. Originally, UCS-2 was variable-width, and then it was fixed-width, and then UTF-16 was proposed to make it variable-width again. The reason for this is actually because there used to be a different standard called UCCS, which was in competition with Unicode at one point before being withdrawn and replaced by it. But the actual historical drama is very relevant for why UTF-16 is such a warty mess. ISO proposed UCCS (then ISO 10646), where you had 31-bit code points. Each codepoint was organized by it's bytes into "groups" (high byte), "planes", "rows", and "cells" (low byte). Furthermore, you couldn't have a group, plane, row, or cell that was an ASCII or ISO control code (00-1F/80-9F), so everything was in Group 20 and Plane 20. The standard then defined 1, 2, and 4-byte encodings of these 31-bit codepoints. These were UCS-4, UCS-2, and UTF-1[0]. UCS-2 as defined in UCCS very much supports multiple code points per character; because UCS-4 was supposed to be the fixed-width encoding. The extension mechanism is very different from UTF-16 surrogates and is actually the same as defined in ISO 2022. That is, when you wanted to use a character from a different group or plane, you would actually encode special code-switching words that said you wanted to change all further characters in the string to that new group or plane. This is also why control codes were forbidden in UCCS code points, for the same reason why UTF-16 surrogates are forbidden in Unicode control points. The underlying problem with UCS-2 (and, for that matter, UTF-1) was that it was not self-synchronizing. In other words, if I take a UCS-2 string and delete one byte from it, all the characters onwards flip their byte order. If I delete a word that's part of a code-switch, then all characters onwards will be misinterpreted in the wrong group or plane. This is a terrible property for text encodings where erroneous deletions or mutations can occur. A couple of things happened that killed UCCS: - Software companies vehemently opposed 31-bit codepoints as wasteful and insisted on doing everything in 16-bits. They promoted a competing standard called Unicode which defined one 16-bit, fixed-width (and, thus, self-synchronizing) encoding just called "Unicode". - When designing Plan 9[1], the designers of that OS realized they could make UTF-1 self-synchronizing, and proposed UTF-8 as an alternative encoding for UCCS. This also had the advantage of not requiring integer division to decode, which made it faster on contemporary processors. ISO ultimately caved and withdrew the original version of ISO 10646. Unicode 1.1 was altered slightly to please ISO, but ISO 10646 was altered drastically to match Unicode. The resulting standard became ISO 10646-1:1993; and future versions of ISO 10646 are more or less identical to specific Unicode versions. At this point our nomenclature changes, because in Unicode 2.0 they define or reference several encodings[2][3]: - UCS-2: Fixed-width 16-bit words. At this point, UCS-2's code-switching mechanism had been withdrawn from ISO 10646, making it identical to just an array of Unicode 1.1 codepoints. - UCS-4: Still specified by ISO 10646 to support 31-bit codepoints, but with the control code restriction withdrawn. Since they were now using group 0 and plane 0 exclusively, the Unicode 2.0 spec says to treat this as codepoints with extra zero padding. At some point in the future this would become UTF-32 with a restriction to 16-bit + astral codepoints. - UCS Transformation Format, 7-bit form; or UTF-7: Backwards compatibility for MIME, because e-mail is evil. - UCS Transformation Format, 8-bit form; or UTF-8: Backwards compatibility for US-ASCII and UNIX filesystems. This is also referred to as "File System Safe UTF", "FSS-UTF", or "UTF-2", but caps characters at 4 bytes, prohibiting the use of 31-bit codepoints. - UTF-16: The historical mess I mentioned in the previous comment. Interestingly enough, several parts of the 2.0 spec differ as to whether or not "Unicode" means UCS-2 or UTF-16. In the UTF-8 appendix they specify 4-byte UTF-8 characters as having a "Unicode value" consisting of two UTF-16 surrogates. But in other parts of the spec, such as Table C-1, they list "UCS-2" and "Unicode" as equivalent. I have a feeling that surrogates were a last-minute compromise between ISO and Unicode, given that the UTF-16 spec warns against using existing private-use superastrals[4] already allocated in ISO 10646. [0] Wikipedia calls this UTF-1 but I don't have access to the old versions of ISO 10646 to check if that standard actually called it that or UCS-1. The Unicode spec calls UTF-8 "UCS Transformation Format-8" so it's possible that the "UTF" nomenclature actually came from ISO first. [1] Plan 9 is a really fascinating OS, because, while it wasn't very commercially successful, it wound up being highly influential. Things like /proc in Linux were ripped wholesale from Plan 9. It even had C extensions that the designers would later copy in Go (which they also made). [2] https://www.unicode.org/versions/Unicode2.0.0/appA.pdf https://www.unicode.org/versions/Unicode2.0.0/appA.pdf [3] https://www.unicode.org/versions/Unicode2.0.0/appC.pdf https://www.unicode.org/versions/Unicode2.0.0/appC.pdf [4] Term I coined in the previous comment for characters beyond U+10FFFF. Encoding them in UTF-16 would require defining a second level of astral surrogate pairs, themselves being encoded as pairs of surrogates.