Hacker News Reader

Comment by sanxiyn | Hacker News Reader

> The main time you want to be able to index by code point is if you’re implementing algorithms defined in the unicode spec that operate on unicode strings (casefolding, segmentation, NFD/NFC). Most if not all of these algorithms operate on whole strings, so implementing them as an iteration pass is usually necessary anyway, so you don’t lose anything if you can’t do arbitrary code point indexing.

Another algorithm for unicode strings which wants to index by code point is unicode regular expression matching. I heard that you do lose something here if you can't assume all code points encode to same length. Unlike other algorithms you mentioned which only iterate forward, depending on implementation regex needs to backtrack.

I once heard that complexity of implementing regex directly on UTF-8 is one reason why Python will not use UTF-8 internally.

View on HN

If everything that can match in a regular expression is valid UTF-8, and everything the regular expression can generate is valid UTF-8, you can just run the regular expression on the byte string.

This requires that matches match UTF-8 substrings, only. That's easy for explicit strings, such as "abc". For more general forms, "." must match one rune, not one byte. But that's not hard.

Python's problem is that strings are random-access indexable by rune (in the Go sense). Python either needs a representation that's rune-indexable, or it needs to generate an index array for strings that need one. There's an argument for the second approach, because most loops don't need a random-access index. "for i in "abcdef: ... " does not, for example.

> Python's problem is that strings are random-access indexable by rune

Yeah, one of my points in the blog post is that random-access rune (code point) indexing isn't very valuable. Python is stuck with it, but if anyone in the future needs to make a similar choice, I hope they don't use O(1) rune indexing as a reason unless they have a very specific use case where O(1) rune indexing actually matters.

(I, for one, am very happy with the "rune" terminology, "char" is way too overloaded. It took me a minute to understand why they did that in Go when I first picked it up, but when I realized the reason I found it brilliant)

One approach for a Python implementation would be to store strings as UTF-8, and generate an index array as needed. For most operations, you don't need an index. Concatenation doesn't need one. Ordinary iteration (for c in "hello" :) doesn't need one. Regular expressions don't need an index. That covers most of the common use cases.

I like "rune" as terminology. A rune is one Unicode code point. A grapheme is a sequence of runes which should not be split.

Yeah, that could work. You don't even need a full index array, it can be a binary datastructure (I am reminded of how skip lists work), e.g. you start off with just one index for the halfway point and fill in things afterwards. There are lots of optimizations you could do.

I'm a bit wary of this because I fear indexing in Python is currently used more than this scheme can handle (i.e. too much for it to be considered a win), if erroneously. I'm not too sure. I certainly have written bad python code using indexing in the past, but that's just me (and many years ago).

String indexing is pretty rare. Most of them are s[0] or s[:n], where you don't need an index.

Typically this is found when parsing something, eg.

    if line[0] == '#':
      # skip comment lines
      continue

    prefixes = {...}
    for line in lines:
        prefix, remainder = line[:4], line[4:]
        if prefix not in prefixes:
            raise ValueError('Invalid prefix ' + prefix)
        parsed = prefixes[prefix](remainder)

Don't you need to index for `s[:n]`? You need to find the byte position of code point n. In current python this isn't a problem because the strings don't use a variable-width encoding, but we're talking about making Python use utf8 :)

It's also possible to chunk the index array: to store the byte-index of every 16th rune, say, and do a bit of cheap iteration afterward. I've had good success with this approach: it's fast, it's light on memory, and it preserves constant-time indexing by code point.

> Another algorithm for unicode strings which wants to index by code point is unicode regular expression matching.

Why? This is certainly not a requirement for at least some levels of Unicode support.

> I once heard that complexity of implementing regex directly on UTF-8 is one reason why Python will not use UTF-8 internally.

It's really not that bad. Both RE2 and Rust's regex engine operate on UTF-8 directly, and this is one of the reasons why these engines are fast. If your automaton is byte based, then you can do a lot of nice tricks in your DFA implementation. (In fact, it's a critical reason why ripgrep doesn't slow down when handling Unicode features, unlike GNU grep.)

For Rust at least, the trickier parts of generating the UTF-8 automaton are available for reuse in the utf8-ranges crate: https://docs.rs/utf8-ranges/1.0.0/utf8_ranges/ (Which is used in at least two regex implementations I've written.)

I'd say the complexity claim might be coming from some other component of the implementation, or something more related to how Python strings are generally represented, rather than it being a fundamental complexity of building a regex engine on UTF-8 directly.

Backtracking in a UTF-8 string is easy: when you need to remember a location, just record its byte offset. You have exactly the same issue in UTF-16, which is also variable-width and therefore not O(1)-indexable by code point. There are loads of regular expression engines which operate on both of these encodings, and I've never seen them go to any great pains to work around the lack of code point indexing.

Yes, you have exactly the same issue in UTF-16, I never said otherwise.

I agree that you never said otherwise, and I applaud you for your accuracy in not doing so. :-D

The reason I bring it up is that UTF-16 is a very commonly used internal encoding for strings, and regular expression engines for platforms which use UTF-16 internally are generally designed to work on that encoding without conversion to UTF-32. Examples that spring to mind are most JavaScript regular expression engines, the ICU regular expression engine, and java.util.regex. If variable-width encodings were so difficult to deal with then you would expect these to have run into some problems with it; however, as far as I can tell, they have not. This is evidence against variable-width encodings being difficult for regular expression engines to deal with.

(Sanity check for my argument above: If you read through https://swtch.com/~rsc/regexp/regexp2.html, which covers both backtracking and non-backtracking regex engines and includes code, none of the code examples would need more than trivial modification to deal with variable-width encodings.)

Note that JS isn't UTF-16, it's slightly different. In JS unpaired surrogates are allowed. String indexing is by code unit, so non-BMP code points are considered to be two "characters".

IIRC Java also allows unpaired surrogates? But does consider non-BMP codepoints to be single characters. I think. Idk.

So from JSs point of view it is a fixed-width encoding, one where all byte values are valid. `"\uD83D\uDC68".match(/\ud83d/)` works (U+D83D is a high surrogate, and "\uD83D\uDC68" is U+1F468 MAN), so the regex engine operates on code units, not code points.

I think for Java the regex engine would have to deal with it as a variable-width encoding, though.

Edit: Hacker News removes emoji from comments.

Java strings are sequences of 16-bit units and don't guarantee UTF-16 validity. There are newer standard library APIs that perform astral-aware operations, though.

So does regex matching break up paired surrogates? It seems like Java Characters can be from the higher planes. JS treats surrogates as their own kind of character, but I think Java only does this for unpaired surrogates, so the encoding is still multibyte. ICBW.

Java, like NT, was originally UCS-2 and moved to UTF-16 when it became clear that UCS-2 won't suffice. APIs that operate on units (anything involving "char"s [the type], CharSequences and so on) are legacy and, when used, are almost certainly used incorrectly. Note that newer APIs on String also use the word "char", but the type "int", these are actually operating on code points and not CUs, ie. these APIs are correct.

(Note that java.lang.Character is a char, so wrong. I think there is no CodePoint type or similar that wraps int.)

In Java regexps, the dot matches a code point or an unpaired surrogate, so if you have a proper surrogate pair, the dot matches the whole pair without splitting it. It does split grapheme clusters.

It's unclear to me what you mean by "Java Characters". A Java char is a UTF-16 code unit, and the java.lang.Character boxes a char (16 bits unsigned). However, since Java 5, java.lang.Character has static methods for astral-aware UTF-16 operations where a code point is a Java int (32 bits signed).

Yeah, by Character I meant java.lang.Character.

> the dot matches the whole pair without splitting it

Yeah, so UTF16 in Java is being used as a variable-width encoding. Nice to know.

There are many considerations of course and I might not have fully grasped your point, but in reference to your last paragraph:

By default RE2 matches directly against the UTF-8 encoding, see this neat state diagram:

https://swtch.com/~rsc/regexp/regexp3.html#step3

In terms of backtracing engines, perl also stores strings internally as UTF-8 (on most platforms) and its regexp engine runs directly on them. It can also match "eXtended grapheme clusters" with \X.

Hm, interesting. You still don't need arbitrary indexing for that; just the ability to iterate backwards, but I can see how it would complicate things. I wonder if it could be implemented on utf8 without affecting performance much.

burntsushi's utf8-ranges library used by the Rust Regex lib probably makes this easier to deal with. I wonder how it performs in relation to the Python implementation (which I assume is in native code)

UTF-8 substrings can't find a misaligned match, because of the way UTF-8 is encoded. First bytes are unambiguous. I had to write "rfind" in Rust, and here it is:

    fn rfind(s: &[&str], key: &str) -> Option<usize> {
        (0 .. s.len()).rev().find(|&i| s[i] == key)    
    }

No need to explicitly worry about rune boundaries. This, like forward find, returns a byte index into a UTF-8 string. Explicit byte indices into UTF-8 strings are troublesome.

Incidentally, that function could be s.iter().rev().position(|ss| ss == key), or drop the `rev()` and use `.rposition(...)` if one wants indices from the start of the slice. Although I'm not sure how searching a slice of strings for a specific one is related to UTF-8: the &str's are used completely opaquely. Did you mean to be searching a &str itself?

> because of the way UTF-8 is encoded.

Yes, I know, I mention this in my blog post :)

Rust's indexing methods on strings actually do operate on byte indices. You can't do `"foo"[2]`, but you can do `"foo[2..3]"`, and it will panic if you are not on a boundary. Nobody uses this (like you say, it's troublesome) :P I've seen folks just break out the byte/char iterator whenever they need something more complicated, because usually they do need to iterate anyway.

Rust internally stores slices as byte indices (well, pointer + byte length), but the compiler enforces the validity of that.

Reply on news.ycombinator.com