css (Rust)

A parser that reads CSS into a faithful abstract syntax tree (the reworkcss/css model): ordered, typed nodes that preserve declaration order, duplicate properties, rule types, and comments.

This is the Rust port of @tabnas/css. The TypeScript package is canonical, the Go module tracks it, and so does this crate. All three are plugins on the tabnas engine, read the same grammar file and are held to the same shared conformance fixtures, so the same stylesheet gives the same tree in every one.

Where a port does differ, the difference is a row of ../test/divergent.tsv with a cell per runtime, and each suite asserts its own cell. There are five rows. Four are Go against TypeScript, and this crate follows TypeScript on every one, including the column arithmetic it would be tidier to round off. The fifth is this crate's own: the engine's tree form (see the engine's own API) refuses nesting deeper than TREE_RULE_DEPTH open rules, where TypeScript and Go have no limit. parse and Css::parse have no limit either.

Install

The crate is published on crates.io, where it depends on the engine, tabnas-parser, and on tabnas-jsonic:

cargo add tabnas-css

It re-exports both, as tabnas_css::tabnas and tabnas_css::tabnas_jsonic, so the examples on this page need no other entry. Version 0.5.9 on crates.io is a different implementation: it has no dependencies, and none of the plugin API this page describes.

The Cargo.toml in this repository differs from the published one. Its own dependencies are named by relative path to sibling checkouts (path = "../../parser/rs"), so a build here needs tabnas/parser, tabnas/json, tabnas/jsonic and, for the tests, tabnas/debug to be cloned beside it. The release workflow swaps those paths for crates.io versions when it publishes the crate. A crate taken from git does not get that swap.

The crate needs Rust 1.85 or later, the engine's own floor.

One example

parse is the one-call entry point: pass source, get the AST and an error:

let ast = tabnas_css::parse("a { color: red }").unwrap();
assert_eq!(
    ast.to_json(),
    r#"{"type":"stylesheet","rules":[{"type":"rule","selectors":["a"],"declarations":[{"type":"declaration","property":"color","value":"red"}]}]}"#
);

Every node is a Node, an insertion-ordered map with a type key. Fields such as rules, declarations and selectors are Value::List. Read them with Node::get and the Value accessors, or call Value::to_json and hand the text to whatever consumes it:

use tabnas_css::Value;

let ast = tabnas_css::parse("h1, h2 { margin: 0 }").unwrap();
let rule = ast.as_node().unwrap()
    .get("rules").and_then(Value::as_list).unwrap()[0]
    .as_node().unwrap();
assert_eq!(Some("rule"), rule.node_type());

parse reuses one process-wide parser and is safe to call from several threads. Building a parser installs the grammar on a new engine, which costs more than a parse does, so a hot loop that needs options should hold a Css rather than call parse_with each time.

Options

Options has two fields, both false by default:

• lowercase_properties. Normalise property names to lower case. Selectors and values are left untouched.
• position. Attach a position (1-based start and end, each a line and a column) to every node.

use tabnas_css::{Css, Options};

let css = Css::with_options(Options { position: true, ..Options::default() });
let ast = css.parse("a { color: red }").unwrap();
// Every node gains a position; the stylesheet's spans the source.
assert!(ast.to_json().ends_with(
    r#""position":{"start":{"line":1,"column":1},"end":{"line":1,"column":17}}}"#
));

Columns count UTF-16 code units, which is what a JavaScript string index counts, so a column here is the column the canonical port reports for the same source.

CSS nesting

A style rule or an at-rule may appear inside another style rule's declaration block. Nested nodes are appended to the parent's declarations in source order, interleaved with declarations:

let ast = tabnas_css::parse("a { color: red; & b { top: 0 } }").unwrap();
// The rule's declarations: a declaration, then the nested rule.
assert_eq!(
    ast.to_json(),
    r#"{"type":"stylesheet","rules":[{"type":"rule","selectors":["a"],"declarations":[{"type":"declaration","property":"color","value":"red"},{"type":"rule","selectors":["& b"],"declarations":[{"type":"declaration","property":"top","value":"0"}]}]}]}"#
);

A plugin on the tabnas engine

The TypeScript and Go ports install the grammar on the tabnas engine, layered on the relaxed-JSON grammar jsonic, and this crate does the same in Rust: it runs on tabnas-parser (imported as tabnas) and tabnas-jsonic, and re-exports both, as tabnas_css::tabnas and tabnas_css::tabnas_jsonic, so that a caller names their types without a second copy of either. jsonic reads css-grammar.jsonic, and the plugin installs its rules with the canonical option overrides and the cssToken lex matcher. Those overrides switch jsonic's own rules and value matchers off, so {a:1} is refused here as it is in the other two ports.

The engine's own API

plugin() is the grammar as an engine plugin, make() builds a jsonic engine with it installed, and Css::tabnas() is the engine a Css runs. Their parse returns the engine's own tabnas::Value, the same tree as the canonical port's plain objects, with these differences from Css::parse:

• the engine writes a whole number, such as a line in a position, as 1.0 in JSON;
• an engine error's column counts Unicode scalars, where this crate's Error counts UTF-16 code units (tabnas_css::Error::from converts one);
• the tree is bounded at TREE_RULE_DEPTH (768) open rules, and one level more fails with cancel;
• with the engine's recovery on, the partial stylesheet is not always the canonical port's (the concepts page says where they part).

The how-to guide has the recipe, and the reference the details.

Building from a checkout

The engine, jsonic, the strict-JSON grammar jsonic takes, and the debug plugin the tests use are sibling checkouts, the standard tabnas development model. Clone parser, json, jsonic and debug next to this repository, and cargo test in rs/ builds against them.

Untrusted input

A parsed stylesheet is data, never instructions. CSS arrives from outside the system, so treat every selector, value, and comment as hostile text. Parsing is not sanitising: this crate returns the raw text the stylesheet contained, and escaping it for HTML, SQL or a shell remains the caller's job. A url(...) in a declaration value is untrusted text, not a link to fetch.

A parse holds memory, and takes time, in proportion to its input, and nesting costs the most. Measured in a release build, 100,000 flat rules (1.7 MB of CSS) peaked at 210 MiB, or 580 MiB with positions on, and 100,000 nested rules (0.6 MB) at 678 MiB, about 7 KiB per open level. The densest nesting, a{ repeated, holds about 2.8 KiB and takes about 6 µs per byte of input. A host that parses untrusted CSS should cap the input's size.

The engine's parse.recover and lex.relex modes are the exception to that proportion: in the engine this crate builds on, both take time quadratic in the input, and a valid 16 KB stylesheet takes about 16 s with recovery on. parse and Css::parse use neither.

parse and Css::parse have no depth limit: 20,000 nested rules parse, and the result drops, clones, compares and prints without recursion, on a 2 MiB thread. The engine's tree form stops at TREE_RULE_DEPTH open rules, as described above.

When debug assertions are on, the engine checks its whole rule stack against a shadow copy on every step, which makes a debug build's parse quadratic in nesting depth. Measured in a debug build, 1,000 nested rules took 8.3 s, 2,000 took 32.9 s and 4,000 took 134 s. This crate's manifest turns the check off for its own builds, but a profile setting applies only to the root package, so a crate that parses untrusted, possibly deep CSS in its own debug builds sets the same key in its own manifest:

[profile.dev.package.tabnas-parser]
debug-assertions = false

Documentation

Full documentation follows the Diátaxis framework:

• Tutorial. A guided first parse, start to finish.
• How-to guide. Short recipes for individual tasks.
• Reference. The public API, every option, and the complete AST node reference.
• Concepts. How the parser is built, and how the Rust port differs from TypeScript.

For the canonical TypeScript implementation, see ../ts/README.md. For the Go port, see ../go/README.md.

License

MIT.