Use unicode-normalization instead of unicode-normalization-alignments

tokenizers/Cargo.toml

@@ -54,7 +54,7 @@ rayon = "1.10"
 rayon-cond = "0.4"
 serde = { version = "1.0", features = [ "derive" ] }
 serde_json = "1.0"
-unicode-normalization-alignments = "0.1"
+unicode-normalization = "^0.1.25"
 unicode_categories = "0.1"
 unicode-segmentation = "1.11"
 indicatif = {version = "0.18", optional = true}
@@ -74,6 +74,7 @@ monostate = "0.1.12"
 ahash = { version = "0.8.11", features = ["serde"] }
 dary_heap = { version = "0.3.6", features = ["serde"] }
 compact_str = { version = "0.9", features = ["serde"] }
+smallvec = "1"
 
 [features]
 default = ["progressbar", "onig", "esaxx_fast"]

tokenizers/src/lib.rs

@@ -145,6 +145,8 @@ pub mod pre_tokenizers;
 pub mod processors;
 pub mod tokenizer;
 
+pub(crate) mod unicode_normalization_alignments;
+
 // Re-export from tokenizer
 pub use tokenizer::*;
 

tokenizers/src/normalizers/strip.rs

@@ -1,7 +1,7 @@
 use crate::tokenizer::{NormalizedString, Normalizer, Result};
+use crate::unicode_normalization_alignments::char::is_combining_mark;
 use crate::utils::macro_rules_attribute;
 use serde::{Deserialize, Serialize};
-use unicode_normalization_alignments::char::is_combining_mark;
 
 #[derive(Copy, Clone, Debug, Deserialize, Serialize)]
 #[serde(tag = "type")]
@@ -61,7 +61,7 @@ mod tests {
     use crate::normalizer::NormalizedString;
     use crate::normalizers::Lowercase;
     use crate::normalizers::NFKD;
-    use unicode_normalization_alignments::UnicodeNormalization;
+    use crate::unicode_normalization_alignments::UnicodeNormalization;
 
     #[test]
     fn test_strip_accents() {

tokenizers/src/tokenizer/normalizer.rs

@@ -1,7 +1,7 @@
 use crate::pattern::Pattern;
+use crate::unicode_normalization_alignments::UnicodeNormalization;
 use crate::{Offsets, Result};
 use std::ops::{Bound, RangeBounds};
-use unicode_normalization_alignments::UnicodeNormalization;
 
 use serde::{Deserialize, Serialize};
 

tokenizers/src/unicode_normalization_alignments/decompose.rs

@@ -0,0 +1,161 @@
+use smallvec::SmallVec;
+use std::fmt::{self, Write};
+use std::iter::Fuse;
+use std::ops::Range;
+use unicode_normalization::char::{
+    canonical_combining_class, decompose_canonical, decompose_compatible,
+};
+
+#[derive(Clone)]
+enum DecompositionType {
+    Canonical,
+    Compatible,
+}
+
+/// External iterator for a string decomposition's characters.
+#[derive(Clone)]
+pub struct Decompositions<I> {
+    kind: DecompositionType,
+    iter: Fuse<I>,
+
+    // This buffer stores pairs of (canonical combining class, character),
+    // pushed onto the end in text order.
+    //
+    // It's divided into up to three sections:
+    // 1) A prefix that is free space;
+    // 2) "Ready" characters which are sorted and ready to emit on demand;
+    // 3) A "pending" block which stills needs more characters for us to be able
+    //    to sort in canonical order and is not safe to emit.
+    buffer: SmallVec<[(u8, char, isize); 4]>,
+    ready: Range<usize>,
+}
+
+#[inline]
+pub fn new_canonical<I: Iterator<Item = char>>(iter: I) -> Decompositions<I> {
+    Decompositions {
+        kind: self::DecompositionType::Canonical,
+        iter: iter.fuse(),
+        buffer: SmallVec::new(),
+        ready: 0..0,
+    }
+}
+
+#[inline]
+pub fn new_compatible<I: Iterator<Item = char>>(iter: I) -> Decompositions<I> {
+    Decompositions {
+        kind: self::DecompositionType::Compatible,
+        iter: iter.fuse(),
+        buffer: SmallVec::new(),
+        ready: 0..0,
+    }
+}
+
+impl<I> Decompositions<I> {
+    #[inline]
+    fn push_back(&mut self, ch: char, first: bool) {
+        let class = canonical_combining_class(ch);
+
+        if class == 0 {
+            self.sort_pending();
+        }
+
+        self.buffer.push((class, ch, if first { 0 } else { 1 }));
+    }
+
+    #[inline]
+    fn sort_pending(&mut self) {
+        // NB: `sort_by_key` is stable, so it will preserve the original text's
+        // order within a combining class.
+        self.buffer[self.ready.end..].sort_by_key(|k| k.0);
+        self.ready.end = self.buffer.len();
+    }
+
+    #[inline]
+    fn reset_buffer(&mut self) {
+        // Equivalent to `self.buffer.drain(0..self.ready.end)` (if SmallVec
+        // supported this API)
+        let pending = self.buffer.len() - self.ready.end;
+        for i in 0..pending {
+            self.buffer[i] = self.buffer[i + self.ready.end];
+        }
+        self.buffer.truncate(pending);
+        self.ready = 0..0;
+    }
+
+    #[inline]
+    fn increment_next_ready(&mut self) {
+        let next = self.ready.start + 1;
+        if next == self.ready.end {
+            self.reset_buffer();
+        } else {
+            self.ready.start = next;
+        }
+    }
+}
+
+impl<I: Iterator<Item = char>> Iterator for Decompositions<I> {
+    type Item = (char, isize);
+
+    #[inline]
+    fn next(&mut self) -> Option<(char, isize)> {
+        while self.ready.end == 0 {
+            match (self.iter.next(), &self.kind) {
+                (Some(ch), &DecompositionType::Canonical) => {
+                    let mut first = true;
+                    decompose_canonical(ch, |d| {
+                        self.push_back(d, first);
+                        first = false;
+                    });
+                }
+                (Some(ch), &DecompositionType::Compatible) => {
+                    let mut first = true;
+                    decompose_compatible(ch, |d| {
+                        self.push_back(d, first);
+                        first = false;
+                    });
+                }
+                (None, _) => {
+                    if self.buffer.is_empty() {
+                        return None;
+                    } else {
+                        self.sort_pending();
+
+                        // This implementation means that we can call `next`
+                        // on an exhausted iterator; the last outer `next` call
+                        // will result in an inner `next` call. To make this
+                        // safe, we use `fuse`.
+                        break;
+                    }
+                }
+            }
+        }
+
+        // We can assume here that, if `self.ready.end` is greater than zero,
+        // it's also greater than `self.ready.start`. That's because we only
+        // increment `self.ready.start` inside `increment_next_ready`, and
+        // whenever it reaches equality with `self.ready.end`, we reset both
+        // to zero, maintaining the invariant that:
+        //      self.ready.start < self.ready.end || self.ready.end == self.ready.start == 0
+        //
+        // This less-than-obviously-safe implementation is chosen for performance,
+        // minimizing the number & complexity of branches in `next` in the common
+        // case of buffering then unbuffering a single character with each call.
+        let (_, ch, size) = self.buffer[self.ready.start];
+        self.increment_next_ready();
+        Some((ch, size))
+    }
+
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        let (lower, _) = self.iter.size_hint();
+        (lower, None)
+    }
+}
+
+impl<I: Iterator<Item = char> + Clone> fmt::Display for Decompositions<I> {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        for c in self.clone() {
+            f.write_char(c.0)?;
+        }
+        Ok(())
+    }
+}

tokenizers/src/unicode_normalization_alignments/mod.rs

@@ -0,0 +1,73 @@
+pub use unicode_normalization::char;
+
+pub mod decompose;
+pub mod recompose;
+
+pub use decompose::Decompositions;
+pub use recompose::Recompositions;
+use std::str::Chars;
+
+/// Methods for iterating over strings while applying Unicode normalizations
+/// as described in
+/// [Unicode Standard Annex #15](http://www.unicode.org/reports/tr15/).
+pub trait UnicodeNormalization<I: Iterator<Item = char>> {
+    /// Returns an iterator over the string in Unicode Normalization Form D
+    /// (canonical decomposition).
+    fn nfd(self) -> Decompositions<I>;
+
+    /// Returns an iterator over the string in Unicode Normalization Form KD
+    /// (compatibility decomposition).
+    fn nfkd(self) -> Decompositions<I>;
+
+    /// An Iterator over the string in Unicode Normalization Form C
+    /// (canonical decomposition followed by canonical composition).
+    fn nfc(self) -> Recompositions<I>;
+
+    /// An Iterator over the string in Unicode Normalization Form KC
+    /// (compatibility decomposition followed by canonical composition).
+    fn nfkc(self) -> Recompositions<I>;
+}
+
+impl<'a> UnicodeNormalization<Chars<'a>> for &'a str {
+    #[inline]
+    fn nfd(self) -> Decompositions<Chars<'a>> {
+        decompose::new_canonical(self.chars())
+    }
+
+    #[inline]
+    fn nfkd(self) -> Decompositions<Chars<'a>> {
+        decompose::new_compatible(self.chars())
+    }
+
+    #[inline]
+    fn nfc(self) -> Recompositions<Chars<'a>> {
+        recompose::new_canonical(self.chars())
+    }
+
+    #[inline]
+    fn nfkc(self) -> Recompositions<Chars<'a>> {
+        recompose::new_compatible(self.chars())
+    }
+}
+
+impl<I: Iterator<Item = char>> UnicodeNormalization<I> for I {
+    #[inline]
+    fn nfd(self) -> Decompositions<I> {
+        decompose::new_canonical(self)
+    }
+
+    #[inline]
+    fn nfkd(self) -> Decompositions<I> {
+        decompose::new_compatible(self)
+    }
+
+    #[inline]
+    fn nfc(self) -> Recompositions<I> {
+        recompose::new_canonical(self)
+    }
+
+    #[inline]
+    fn nfkc(self) -> Recompositions<I> {
+        recompose::new_compatible(self)
+    }
+}