feruca is a basic implementation of the Unicode Collation Algorithm in Rust. It's current with Unicode version 14.0. The name of the library is a portmanteau of Ferris 🦀 and UCA.
No unsafe is used directly in this library. It relies on the well-vetted
bstr to accept input (in the form of
either &str or &[u8]), to perform UTF-8 validation, and to generate a list
of Unicode scalar values, which can then be processed for collation. The idea is
to be tolerant of input that may not be entirely kosher UTF-8.
In describing feruca as a "basic implementation," I have a few things in mind.
First, I don't expect that it will win any awards for performance. My rough
attempts at benchmarking suggest that this is on the order of 8–10x slower than
ucol from icu4c. (On the other hand,
that isn't as bad as one might imagine, considering the incredible degree of
optimization achieved in the ICU libraries.) My initial priority was to pass the
official
conformance tests.
feruca also passes the conformance tests for the
CLDR root collation order.
Second, support for tailoring is minimal (so far). You can choose between two tables of character weights: the Default Unicode Collation Element Table (DUCET), or the CLDR variation thereof. The CLDR table then becomes the starting point for actual collation tailoring based on language/locale. I have added only one tailoring, intended for use with Arabic-script languages. It shifts letters in the Arabic script so that they sort before the Latin script. This is enough for my own work with Persian and Arabic texts. The CLDR table in its unmodified form—i.e., the root collation order—works out-of-the-box for several other languages. I do plan to add more tailorings, but it will be a gradual process, and driven by demand. Realistically, feruca will never have the kind of all-encompassing, flexible support for tailoring that is provided by ICU. My feeling is that there's a place for less sophisticated solutions, with simpler APIs, smaller dependency trees, etc. (If you have thoughts on this, I would be interested in hearing them.)
Apart from locale tailoring, you can choose between the "non-ignorable" and "shifted" strategies for handling variable-weight characters—with the latter being the default.
Third, this library has effectively[0] just one public method, collate,
belonging to a struct, Collator, which sets a few options. collate accepts
two string references or byte slices, and returns an Ordering value. It is
designed to be passed as a comparator to the standard library method sort_by
(or sort_unstable_by). See "Example usage" below.
For many people and use cases, UCA sorting will not work properly without being
able to specify a certain locale. Again, however, it is worth emphasizing the
usefulness of the CLDR root collation order on its own. When defining a
Collator, you can set the default options (see below), which specify the use
of the CLDR table with the "shifted" strategy. I think this is a good starting
point.
[0]: There is also a variant form, collate_no_tiebreak, which will return
Ordering::Equal for any two strings that produce the same UCA sort key. (The
normal version will fall back on byte-value comparison in such cases.)
```rust use feruca::Collator;
fn main() { let mut uca = [ "چنگیز", "Éloi", "Ötzi", "Melissa", "صدام", "Mélissa", "Overton", "Elrond", ];
let mut naive = uca;
let collator = Collator::default();
uca.sort_by(|a, b| collator.collate(a, b));
naive.sort();
for item in uca {
println!("{}", item);
}
// Éloi
// Elrond
// Melissa
// Mélissa
// Ötzi
// Overton
// چنگیز
// صدام
// An empty line (in case you run this verbatim)
println!();
for item in naive {
println!("{}", item);
}
// Elrond
// Melissa
// Mélissa
// Overton
// Éloi
// Ötzi
// صدام
// چنگیز
} ```
The UCA conformance tests can be run with the command cargo test --release.
Please note that, as a result of this library's reliance on bstr for UTF-8
validation, any
surrogate code points
found in input to the collate method will be converted to the standard
"replacement character," U+FFFD. Conformant implementations of the UCA are
explicitly allowed to follow this approach. It does mean, however, that a
handful of lines (out of hundreds of thousands) in the conformance tests need to
be skipped. If you look at the conformance function in the tests module,
you'll see that any line containing a surrogate code point is passed over.
The binary files included with feruca represent hash tables of Unicode data. They are generated in a separate repository, feruca-mapper, and serialized using bincode. You can rebuild them yourself, if you prefer.
The text files in the test-data directory are covered by the
Unicode License Agreement. Everything
else is MIT-licensed.