Text Editor Data Structures
cdacamar.github.io
cdacamar.github.io
What would be the data structures for that? I can only think of trying to replicate something like the HTML DOM but i have a feeling something like Write for Windows 3.1 used a simpler data structure.
The real complexity is rendering all of unicode properly, and supporting international fonts, bidi layout, vertical text, etc.
5 bytes? In what encoding?
> So one Unicode character can be up to 5 bytes long and take up the same canvas space as 3 characters.
FWIW, I didn't read that as suggesting an upper bound of 5 bytes, but rather as an example using arbitrary numbers: N bytes of code units could, depending on the font providing the glyph(s) for the respective grapheme(s), could be rendered at M times the size of, say, the letter A, where N != M -- despite the font otherwise being monospaced. Which is just another way of saying that you must consult the font for the character widths involved.
I think you're reading that quote as an assertion that:
For any grapheme G, G can be encoded in at most 5 bytes.
There exists a grapheme G, where G is encoded in 5 bytes, and the respective glyph happens to be displayed at 3 times a single character (e.g. the letter A), despite the font otherwise being monospaced. Therefore you *must* consult the font for each glyph to correctly determine character widths.> You also need to read ahead as there are combination characters, for example a smiley combined with the color brow becomes a brown smiley.
Emphasis mine. Clearly combination characters are being treated separately.
Frankly I think it's crazy to read "up to 5 bytes" and not think that it suggest an upper bound. I think you're reaching for a highly questionably interpretation of a totally unambiguous clause. If the author meant to express what you're saying, they would certainly have written: "Some Unicode characters are 5 bytes long and take up the same canvas space as 3 characters". Which would still look incorrect if they followed it with the sentence "You also need to read ahead as there are combination characters...".
It is far more likely that the author is simply mistaken and should have said 4 bytes, and perhaps used the word "codepoint" instead of "character" in the original sentence. That's a perfectly understandable technical error, while the reinterpretation you're putting together would imply an error of colloquial language.
An idea for a variable width encoding of 1 to 3 bytes: Read the MSB of each byte: If it's 0, don't read any more bytes. If it's 1, read the next byte. Do the same (up to 3 times). The non MSB bits of each byte then make up the codepoint.
0xxxxxxx (ASCII)
1xxxxxxx 0xxxxxxx (0x0080 - 0x3FFF)
1xxxxxxx 1xxxxxxx 0xxxxxxx (0x4000 - 0x1FFFFF)
1xxxxxxx 1xxxxxxx 1xxxxxxx 0xxxxxxx (0x200000 - 0x10000000)
Another drawback is that it is not self-synchronizing, which is one of the benefits of UTF-8.
It also has the issue that you can represent some codepoints with more than one encoding: eg, put ASCII characters into 2 or 3 bytes. So you would need rules to use the minimal encoding for each codepoint.
As a space-saving technique, it may offer better density than UTF-8 or UTF-16 on some texts.
You could also use a fixed-width encoding of 24-bits to avoid the problem of reading it serially, but as computers typically work in powers of 2, you would align 24-bit values at 32-bit addresses and load them into 32-bit+ registers, so there's nothing to really gain in terms of performance here over UTF-32, but you could save a bit of space.
In UTF-8, a 3-byte encoding uses 8-bits as part of the encoding, a full byte worth of bits for the encoding itself, leaving only 16-bits for the codepoint. If you need higher code-points you need to use 4 bytes, where 11 bytes are the encoding and 21 bytes are the codepoint.
So UTF-8 is space efficient for ASCII, but ~1/3 of the bits are used for the encoding in for non-ASCII, versus a fixed 1/8 of the bits used for the encoding above for all 1-3 bytes. The above has a fixed 12.5% space overhead over raw codepoints. UTF-8 has 12.5% only for ASCII, and ~33% overhead for everything else.
Although it is not self-synchronizing like UTF-8, you can synchronize a reliable stream by holding a buffer of the previous byte. If the previous byte's value is >=0x80, the current byte is part of the same character. If it's <0x80, the current byte is the start of a new character, so it's still possible to do substring matching etc, fairly efficiently but slightly less efficiently than UTF-8. It makes it suitable for file storage, but not ideal for transmission.
That said, most sane protocols will prefix a string with a length (in bytes), so self-synchronization is not always an issue.
Saying a ZWJ sequence can be "up to 5 bytes" is like saying "the current generation of Intel processors run at clock speeds of up to 2 GHz".
If they were referring to ZWJ sequences (I don't think they were; I think they were just misremembering the maximum encoded length of a codepoint) and they had said "up to 35 bytes", then I might agree with you. It's still not technically accurate, but it's a reasonable colloquial usage, like saying "human males can grow up to seven feet tall".
I believe UTF-8 reserved up to six bytes for a single character.
So what is "up to 5"?
The "styling a range of text" is something i thought but you still need to somehow associate the text with the range - and vice versa - and this doesn't handle things like inserting images and other types of objects since these aren't text.
You could have a document be a series of "paragraphs", each being a series of "elements" with each "element" being something like "text" (with a style), "image", etc. But then once tables enter the picture, you need to expand paragraphs to be of "table" type and each table cell is itself a self-contained "series of paragraphs" - and then start thinking about nested tables or images in tables!
Generalize that enough to avoid special cases inside special cases and you end up with more of a tree-like structure representing a DOM and less with a linear structure with range-based styling.
(of course, then again, i don't remember Write for Windows 3.1 having tables in the first place :-P but i'm interested if there are alternative approaches anyway)
EDIT: one thing i forgot to mention - and why i am curious about non-DOM-based approaches - is that one problem with the DOM approach is the selection: with a linear/range-based structure the selection is just one or two indices inside the range, but with the DOM the selection can start from a node with node-specific subrange (e.g. character in a text node) and end with another node and both being very unrelated to each other (i.e. only having some distant common ancestor and not necessarily at the same level).
A plaintext document is an array of chars, a richtext document is tree, which may or may not be well-formed.
Think about someone trying to bold semi-half of_a sentence_, and how MS Frontpage was made by smart people, it’s just really hard.
The most interesting thing lately is the HTML attribute `contenteditable`, and how it almost just kinda works! You still have to be full-stack to make something good, but that was an amazing improvement to the browser.
You can move a lot of that stuff to reusable methods but personally i find the whole "editing" aspect to be more involved than the "drawing" side - and also the one more likely to be different than a plain text editor - when dealing with DOM-like structures. Hence why i am interested to see what alternatives there are.
Quill works but is basically dead since 2017. Almost anything foss is in a similarly ambiguous boat.
And then people like us, defeated, eventually buy something when we actually need it.
There are just so many ways that users try to use it. It’s a tough problem!
> with the DOM the selection can start from a node with node-specific subrange (e.g. character in a text node) and end with another node and both being very unrelated to each other
I'd just store the range as character indices, using those the right nodes in the tree can be accessed pretty quickly as needed.
Other approaches would probably require the selection to be a tree of its own, I can't really say whether that's simpler overall or not.
Though that is basically another way to represent what i wrote above with having a pair of node pointers and a subrange (well, an index actually, the other end of the subrange is implicit if the node pointers are different). This is basically what the old HTML editing control Microsoft had back in the 90s used and that worked with the DOM tree (also what i used in a test editor i wrote some time ago). And yeah it isn't simple.
It's basically how the classic RTF format [1] works, and things like VT100/ANSI escape codes in terminals. It's kind of like the difference between imperative code and declarative code: "this character sequence means toggle the state of bold" versus "this node of characters is bold".
The docs for ProseMirror are a brilliant insight into how many of these editors are designed. ProseMirror maintains its own document model in parallel to the html DOM and rectifies one to the other as either changes.
For realtime collaborative rich text editors take a look at PeriText, they have a brilliant article explaining the CRDT structures used.
Here are the Win32 docs: https://learn.microsoft.com/en-us/windows/win32/controls/ric...
The more I read about this control, the more I learn about its insane feature set! Microsoft continues to make significant improvements to a version that is only shipped with Microsoft Office -- not available from a barebones Win7/10/11 install. Read more here: https://devblogs.microsoft.com/math-in-office/using-richedit...
Rich Edit control also supports the Text Object Model, which is very powerful. Read more here: https://learn.microsoft.com/en-us/windows/win32/api/tom/nn-t...
There is a text file describing the reverse engineered file format here:
https://web.archive.org/web/20130831064118/http://msxnet.org...
The TOM stuff was added later when the MSWord people took over ownership.
I worked on Ready,Set,Go! back in the day and also wrote my own styled editor for the Mac in the 90's.
One interesting thing about this is that you could use a lazy "adjustor" to remove duplicate or overlapping styles. No need to worry about it during typing or selecting. A low priority task could analyze the runs and fix them up as needed and no one was the wiser.
IMO, the hardest part about writing a word processor back then was font management. You had to construct width tables to be able to compose lines properly. This generally involved a bunch of poorly documented APIs and peering into opaque data structures.
[1]: https://github.com/FXMisc/RichTextFX/tree/master/richtextfx/...
[2]: https://gluonhq.com/presenting-a-new-richtextarea-control/
[3]: https://github.com/gluonhq/rich-text-area/blob/master/sample...
[4]: https://github.com/DaveJarvis/keenwrite
[5]: https://www.youtube.com/watch?v=3QpX70O5S30&list=PLB-WIt1cZY...
[1] https://source.winehq.org/git/wine.git/tree/HEAD:/dlls/riche...
https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.48...
The major failing point of the Piece Tree from my benchmarks is the substring/querying time. An idea I want to try out to speed up my Piece Tree implementation is to distinguish between Concat (metadata) and Leaf (string) nodes, just as a Rope does, storing metadata in the internal nodes and Pieces at the leaves.
The reason (I hope) that will improve substring times is because, in a Piece Tree, the original string can be reconstructed through an in-order traversal of the tree's internal nodes.
So, if you specify a substring range that starts from one character before the root node and ends one character after the root node, you end up traversing to the rightmost node in the left subtree and the leftmost node in the right subtree (two O(log n) operations).
I'm hopeful the tree depth that would need to be traversed if the nodes were at the Leaves (like in a Rope) would be shorter (especially since adjacent pieces won't be O(log n) distance away) and want to try it out myself, but my intuition might be wrong. You can have a go trying that out yourself if the idea interests you.
[0]: https://github.com/microsoft/vscode/tree/main/src/vs/editor/...
Now in the world of V8 and JavaScriptCore, I don't think it's crazy at all. So much work has gone into JS runtime optimization. For heavily concurrent and memory intensive workloads, I can imagine problems though.
This would still only be a O(n) operation. The constant value might be higher, but the complexity is the same.
I don’t buy the argument that gap buffers “are bad for multiple cursors”. I get the argument on a theoretical level, but real hardware is not theoretical. There are several operations that are theoretical faster with a hashmap or b-tree than a vector, like insert and delete. But in reality the vector is usually faster in the real world except for very large inputs[1]. Gap buffers are basically vectors.
Another point with multiple cursors and gap buffers is that Chris Wellons animations show the gap moving back to the first cursor every time it needed to add a new character. But in reality you would just walk the cursors in reverse order each time, saving a trip.
I have actually written and benchmarked a very naive and unoptimized gap buffer, and the results showed that it was faster than highly optimized rope implementations on all real world benchmarks[2], including multiple cursors.
That being said, a gap buffer is still probably not the best data structure for a text editor because it has worse “worse case” performance than something like a rope or piece-tree. Even though it is faster overall, it’s the tail latency's that really matter for interactive programs.
Overall I enjoyed reading the post, I find the topics fascinating and this was well presented.
[1] http://www.goodmath.org/blog/2009/02/18/gap-buffers-or-dont-...
[2] https://github.com/CeleritasCelery/rune/issues/17#issuecomme...
I miss that editor so much, that I'm considering to write one some day, but I have no idea how to do so. I can invent things myself, but I guess those things were invented already back in the days computers were different.
https://github.com/arximboldi/ewig
https://github.com/arximboldi/immer
See the author instantly opening a ~1GB text file with async loading, paging through, copying/pasting, and undoing/redoing in their prototype “ewig” text editor about 27 minutes into their talk here:
https://m.youtube.com/watch?v=sPhpelUfu8Q
It’s backed by a “vector of vectors” data structure called a relaxed radix balanced tree:
https://infoscience.epfl.ch/record/169879/files/RMTrees.pdf
That original paper has seen lots of attention and attempts at performance improvements, such as:
Yes, this isn't encouraged enough. I often serialize my data structures as either JSON or Graphviz DOT files for visualization. It helps save an immense amount of time. You can also use the generated files for regression testing, i.e. diff the actual output with the serialized output and if they're different, then a bug was introduced.
It's been a while, but I believe that it represented everything in a tree of n-character chunks (n = 6?). It was probably the first editor that I used that could open files of pretty much any length.
"Idioteque" is a song => by the English rock band Radiohead <=, released on their fourth album, Kid A (2000).
If I want to then edit this text, is there an efficient algorithm for figuring out the start and end index of the highlight for the edited text?
1. The claim that the rope is inefficient for undo/redo is based on a singular example of a small edit on short strings. This isn't where ropes shine, admittedly, but they don't need to shine there, because when dealing with such small pieces of data, pretty much anything you do will be faster than the user can see. If using larger strings, the space allocated for nodes becomes background noise as the strings themselves dominate the size.
2. The chosen solution, the piece table, is more memory-efficient than the rope at first glance, but that's a surface-level efficiency. The eventually-chosen solution, a piece tree, is far less memory efficient. Sure, at first glance it is more memory-efficient, but this is at the expense of tree traversals, which in the VSCode article, are addressed with cache, which... uses more memory. In the author's implementation there's even more memory used because there's a requirement he didn't include in his list: he wants it all to be immutable. Nevermind that ropes were immutable from the start...
3. If you have a document which uses a lot (read, thousands) of very small edits, then the size of small strings might start to matter. So if you're going to optimize for this, optimize it. There are some fairly small optimizations that make the inefficiency concerns completely irrelevant. One is pointer packing: in a 64-bit system, pointers are 64 bits, but in practice, the vast majority of systems use 48 or fewer of those bits: as it turns out, there aren't many systems with more than 2^48 bytes = 256 terabytes of RAM. This means the leading 16 bits are 0s. Trivially, this means you can store strings of 7 8-bit characters in the pointer itself, using the first 8 bits to signal if it's a string or pointer (if they're all 0s, it's a pointer) and the length of the string. All the strings in the inefficiency example can fit in a 64 bit integer: "Hello", " ", and "world" are all fewer than 7 bytes, which means you're passing around 64-bit integers, by value, with no allocations necessary. In fact, this means you can either append the " " to "Hello" like "Hello ", or prepend it to "world" like " world", and still stay under 7 bytes in either case. Remember, this is now 64 bit integers being passed around by value: this is far faster than piece trees.
4. The author treats undo/redo as a stack, but all the cool editors treat it as a tree. If you make a change, then undo it, then make another change, then undo the second change, is your first change lost? In vim/emacs, the answer is no: you can go back into a tree and find it to reapply. This means that all text is not only immutable but immortal: it has to be kept for the duration of the editing session. This enables a few more optimizations: we no longer need reference counts or garbage collection since we aren't reclaiming the memory, and now we can point into existing strings since they'll never change. Consider the following string: "The quick brown box jumps over the lazy dog." You may have noticed a typo: "box" should be "fox". This change requires 0 buffer allocations: we have a pointer to "The quick brown box jumps over the lazy dog." for the original string, a pointer to the same spot for "The quick brown " (with a length), a second pointer to "ox jumps over the lazy dog.", and a packed pointer (integer) for the string "f". This is pretty key because if you're not freeing any of this memory, you need to make sure you don't allocate more than necessary!
NOTE: I'm not saying that the rope is the better structure here. There may be more requirements which weren't captured in the article which mean that piece buffers really are the right answer. All I'm saying is that the article doesn't really explore ropes deeply enough to write them off so quickly.