`cstree`

A library for generic lossless syntax trees

`cstree` is a generic library for creating and working with concrete syntax trees (CSTs). "Traditional" abstract syntax trees (ASTs) usually contain different types of nodes which represent different syntactical elements of the source text of a document and reduce its information to the minimal amount necessary to correctly interpret it. In contrast, CSTs are lossless representations of the entire input where all tree nodes are represented homogeneously (i.e., the nodes are _untyped_), but are tagged with a `RawSyntaxKind` to determine the kind of grammatical element they represent. One big advantage of this representation is that it cannot only recreate the original source exactly, but also lends itself very well to the representation of _incomplete or erroneous_ trees and is thus highly suited for usage in contexts such as IDEs or any other application where a user is _editing_ the source text. The concept of and the data structures for `cstree`'s syntax trees are inspired in part by Swift's [libsyntax](https://github.com/apple/swift/tree/5e2c815edfd758f9b1309ce07bfc01c4bc20ec23/lib/Syntax). Trees consist of two layers: the inner tree (called _green_ tree) contains the actual source text as position independent green nodes. Tokens and nodes that appear identically at multiple places in the source are deduplicated in this representation in order to store the tree efficiently. This means that a green tree may not actually structurally be a tree. To remedy this, the real syntax tree is constructed on top of the green tree as a secondary tree (called the _red_ tree), which models the exact source structure. As a possible third layer, a strongly typed AST [can be built] on top of the red tree. [can be built]: #ast-layer The `cstree` implementation is a fork of the excellent [`rowan`](https://github.com/rust-analyzer/rowan/), developed by the authors of [rust-analyzer](https://github.com/rust-analyzer/rust-analyzer/) who wrote up a conceptual overview of their implementation in [their repository](https://github.com/rust-analyzer/rust-analyzer/blob/master/docs/dev/syntax.md#trees). Notable differences of `cstree` compared to `rowan`: - Syntax trees (red trees) are created lazily, but are persistent. Once a red node has been created by `cstree`, it will remain allocated. In contrast, `rowan` re-creates the red layer on the fly on each traversal of the tree. Apart from the trade-off discussed [here](https://github.com/rust-analyzer/rust-analyzer/blob/master/docs/dev/syntax.md#memoized-rednodes), this helps to achieve good tree traversal speed while helping to provide the following: - Syntax (red) nodes are `Send` and `Sync`, allowing to share realized trees across threads. This is achieved by atomically reference counting syntax trees as a whole, which also gets rid of the need to reference count individual nodes. - `SyntaxNode`s can hold custom data. - `cstree` trees are trees over interned strings. This means `cstree` will deduplicate the text of tokens with the same source string, such as identifiers with the same name. In this position, `rowan` stores each token's text together with its metadata as a custom DST (dynamically-sized type). - `cstree` includes some performance optimizations for tree creation: it only allocates space for new nodes on the heap if they are not in cache and avoids recursively hashing subtrees by pre-hashing them. - `cstree` includes some performance optimizations for tree traversal: persisting red nodes allows tree traversal methods to return references instead of cloning nodes, which involves updating the tree's reference count. You can still `clone` the reference to obtain an owned node, but you only pay that cost when you need to. - The downside of offering thread safe syntax trees is that `cstree` cannot offer any mutability API for its CSTs. Trees can still be updated into new trees through replacing nodes, but cannot be mutated in place. ## Getting Started If you're looking at `cstree`, you're probably looking at or already writing a parser and are considering using concrete syntax trees as its output. We'll talk more about parsing below -- first, let's have a look at what needs to happen to go from input text to a `cstree` syntax tree: 1. Define an enumeration of the types of tokens (like keywords) and nodes (like "an expression") that you want to have in your syntax and implement `Syntax` 2. Create a `GreenNodeBuilder` and call `start_node`, `token` and `finish_node` from your parser 3. Call `SyntaxNode::new_root` or `SyntaxNode::new_root_with_resolver` with the resulting `GreenNode` to obtain a syntax tree that you can traverse Let's walk through the motions of parsing a (very) simple language into `cstree` syntax trees. We'll just support addition and subtraction on integers, from which the user is allowed to construct a single, compound expression. They will, however, be allowed to write nested expressions in parentheses, like `1 - (2 + 5)`. ### Defining the language First, we need to list the different part of our language's grammar. We can do that using an `enum` with a unit variant for any terminal and non-terminal. The `enum` needs to be convertible to a `u32`, so we use the `repr` attribute to ensure it uses the correct representation. ```rust #[derive(Debug, Clone, Copy, PartialEq, Eq)] #[repr(u32)] enum SyntaxKind { /* Tokens */ Int, // 42 Plus, // + Minus, // - LParen, // ( RParen, // ) /* Nodes */ Expr, Root, } ``` For convenience when we're working with generic `cstree` types like `SyntaxNode`, we'll also give a name to our syntax as a whole and add a type alias for it. That way, we can match against `SyntaxKind`s using the original name, but use the more informative `Node` to instantiate `cstree`'s types. ```rust type Calculator = SyntaxKind; ``` Most of these are tokens to lex the input string into, like numbers (`Int`) and operators (`Plus`, `Minus`). We only really need one type of node; expressions. Our syntax tree's root node will have the special kind `Root`, all other nodes will be expressions containing a sequence of arithmetic operations potentially involving further, nested expression nodes. To use our `SyntaxKind`s with `cstree`, we need to tell it how to convert it back to just a number (the `#[repr(u32)]` that we added) by implementing the `Syntax` trait. We can also tell `cstree` about tokens that always have the same text through the `static_text` method on the trait. This is useful for the operators and parentheses, but not possible for numbers, since an integer token may be produced from the input `3`, but also from other numbers like `7` or `12`. We implement `Syntax` on an empty type, just so we can give it a name. ```rust impl Syntax for Calculator { fn from_raw(raw: RawSyntaxKind) -> Self { // This just needs to be the inverse of `into_raw`, but could also // be an `impl TryFrom for SyntaxKind` or any other conversion. match raw.0 { 0 => SyntaxKind::Int, 1 => SyntaxKind::Plus, 2 => SyntaxKind::Minus, 3 => SyntaxKind::LParen, 4 => SyntaxKind::RParen, 5 => SyntaxKind::Expr, 6 => SyntaxKind::Root, n => panic!("Unknown raw syntax kind: {n}"), } } fn into_raw(self) -> RawSyntaxKind { RawSyntaxKind(self as u32) } fn static_text(self) -> Option<&'static str> { match self { SyntaxKind::Plus => Some("+"), SyntaxKind::Minus => Some("-"), SyntaxKind::LParen => Some("("), SyntaxKind::RParen => Some(")"), _ => None, } } } ``` #### Deriving `Syntax` To save yourself the hassle of defining this conversion (and, perhaps more importantly, continually updating it while your language's syntax is in flux), `cstree` includes a derive macro for `Syntax` when built with the `derive` feature. With the macro, the `Syntax` trait implementation above can be replaced by simply adding `#[derive(Syntax)]` to `SyntaxKind`. ### Parsing into a green tree With that out of the way, we can start writing the parser for our expressions. For the purposes of this introduction to `cstree`, I'll assume that there is a lexer that yields the following tokens: ```rust #[derive(Debug, PartialEq, Eq, Clone, Copy)] pub enum Token<'input> { // Note that number strings are not yet parsed into actual numbers, // we just remember the slice of the input that contains their digits Int(&'input str), Plus, Minus, LParen, RParen, // A special token that indicates that we have reached the end of the file EoF, } ``` A simple lexer that yields such tokens is part of the full `readme` example, but we'll be busy enough with the combination of `cstree` and the actual parser, which we define like this: ```rust pub struct Parser<'input> { // `Peekable` is a standard library iterator adapter that allows // looking ahead at the next item without removing it from the iterator yet lexer: Peekable>, builder: GreenNodeBuilder<'static, 'static, Calculator>, } impl<'input> Parser<'input> { pub fn new(input: &'input str) -> Self { Self { // we get `peekable` from implementing `Iterator` on `Lexer` lexer: Lexer::new(input).peekable(), builder: GreenNodeBuilder::new(), } } pub fn bump(&mut self) -> Option> { self.lexer.next() } } ``` In contrast to parsers that return abstract syntax trees, with `cstree` the syntax tree nodes for all element in the language grammar will have the same type: `GreenNode` for the inner ("green") tree and `SyntaxNode` for the outer ("red") tree. Different kinds of nodes (and tokens) are differentiated by their `SyntaxKind` tag, which we defined above. You can implement many types of parsers with `cstree`. To get a feel for how it works, consider a typical recursive descent parser. With a more traditional AST, one would define different AST structs for struct or function definitions, statements, expressions and so on. Inside the parser, the components of any element, such as all fields of a struct or all statements inside a function, are parsed first and then the parser wraps them in the matching AST type, which is returned from the corresponding parser function. Because `cstree`'s syntax trees are untyped, there is no explicit AST representation that the parser would build. Instead, parsing into a CST using the `GreenNodeBuilder` follows the source code more closely in that you tell `cstree` about each new element you enter and all tokens that the parser consumes. So, for example, to parse a struct definition the parser first "enters" the struct definition node, then parses the `struct` keyword and type name, then parses each field, and finally "finishes" parsing the struct node. The most trivial example is the root node for our parser, which just creates a root node containing the whole expression (we could do without a specific root node if any expression was a node, in particular if we wrapped integer literal tokens inside `Expr` nodes). ```rust pub fn parse(&mut self) -> Result<(), String> { self.builder.start_node(SyntaxKind::Root); self.parse_expr()?; self.builder.finish_node(); Ok(()) } ``` As there isn't a static AST type to return, the parser is very flexible as to what is part of a node. In the previous example, if the user is adding a new field to the struct and has not yet typed the field's type, the CST node for the struct doesn't care if there is no child node for it. Similarly, if the user is deleting fields and the source code currently contains a leftover field name, this additional identifier can be a part of the struct node without any modifications to the syntax tree definition. This property is the key to why CSTs are such a good fit as a lossless input representation, which necessitates the syntax tree to mirror the user-specific layout of whitespace and comments around the AST items. In the parser for our simple expression language, we'll also have to deal with the fact that, when we see a number the parser doesn't yet know whether there will be additional operations following that number. That is, in the expression `1 + 2`, it can only know that it is parsing a binary operation once it sees the `+`. The event-like model of building trees in `cstree`, however, implies that when reaching the `+`, the parser would have to have already entered an expression node in order for the whole input to be part of the expression. To get around this, `GreenNodeBuilder` provides the `checkpoint` method, which we can call to "remember" the current position in the input. For example, we can create a checkpoint before the parser parses the first `1`. Later, when it sees the following `+`, it can create an `Expr` node for the whole expression using `start_node_at`: ```rust fn parse_lhs(&mut self) -> Result<(), String> { // An expression may start either with a number, or with an opening parenthesis that is // the start of a parenthesized expression let next_token = *self.lexer.peek().unwrap(); match next_token { Token::Int(n) => { self.bump(); self.builder.token(SyntaxKind::Int, n); } Token::LParen => { // Wrap the grouped expression inside a node containing it and its parentheses self.builder.start_node(SyntaxKind::Expr); self.bump(); self.builder.static_token(SyntaxKind::LParen); self.parse_expr()?; // Inner expression if self.bump() != Some(Token::RParen) { return Err("Missing ')'".to_string()); } self.builder.static_token(SyntaxKind::RParen); self.builder.finish_node(); } Token::EoF => return Err("Unexpected end of file: expected expression".to_string()), t => return Err(format!("Unexpected start of expression: '{t:?}'")), } Ok(()) } fn parse_expr(&mut self) -> Result<(), String> { // Remember our current position let before_expr = self.builder.checkpoint(); // Parse the start of the expression self.parse_lhs()?; // Check if the expression continues with `+ ` or `- ` let Some(next_token) = self.lexer.peek() else { return Ok(()); }; let op = match *next_token { Token::Plus => SyntaxKind::Plus, Token::Minus => SyntaxKind::Minus, Token::RParen | Token::EoF => return Ok(()), t => return Err(format!("Expected operator, found '{t:?}'")), }; // If so, retroactively wrap the (already parsed) LHS and the following RHS // inside an `Expr` node self.builder.start_node_at(before_expr, SyntaxKind::Expr); self.bump(); self.builder.static_token(op); self.parse_expr()?; // RHS self.builder.finish_node(); Ok(()) } ``` ### Obtaining the parser result Our parser is now capable of parsing our little arithmetic language, but it's methods don't return anything. So how do we get our syntax tree out? The answer lies in `GreenNodeBuilder::finish`, which finally returns the tree that we have painstakingly constructed. ```rust impl Parser<'_> { pub fn finish(mut self) -> (GreenNode, impl Interner) { assert!(self.lexer.next().map(|t| t == Token::EoF).unwrap_or(true)); let (tree, cache) = self.builder.finish(); (tree, cache.unwrap().into_interner().unwrap()) } } ``` `finish` also returns the cache it used to deduplicate tree nodes and tokens, so you can re-use it for parsing related inputs (e.g., different source files from the same crate may share a lot of common function and type names that can be deduplicated). See `GreenNodeBuilder`'s documentation for more information on this, in particular the `with_cache` and `from_cache` methods. Most importantly for us, we can extract the `Interner` that contains the source text of the tree's tokens from the cache, which we need if we want to look up things like variable names or the value of numbers for our calculator. To work with the syntax tree, you'll want to upgrade it to a `SyntaxNode` using `SyntaxNode::new_root`. You can also use `SyntaxNode::new_root_with_resolver` to combine tree and interner, which lets you directly retrieve source text and makes the nodes implement `Display` and `Debug`. The same output can be produced from `SyntaxNode`s by calling the `debug` or `display` method with a `Resolver`. To visualize the whole syntax tree, pass `true` for the `recursive` parameter on `debug`, or simply debug-print a `ResolvedNode`: ```rust let input = "11 + 2-(5 + 4)"; let mut parser = Parser::new(input); parser.parse().unwrap(); let (tree, interner) = parser.finish(); let root = SyntaxNode::::new_root_with_resolver(tree, interner); dbg!(root); ``` ## AST Layer While `cstree` is built for concrete syntax trees, applications are quite easily able to work with either a CST or an AST representation, or freely switch between them. To do so, use `cstree` to build syntax and underlying green tree and provide AST wrappers for your different kinds of nodes. An example of how this is done can be seen [here](https://github.com/rust-analyzer/rust-analyzer/blob/master/crates/syntax/src/ast/generated.rs) and [here](https://github.com/rust-analyzer/rust-analyzer/blob/master/crates/syntax/src/ast/generated/nodes.rs) (note that the latter file is automatically generated by a task). ## License `cstree` is primarily distributed under the terms of both the MIT license and the Apache License (Version 2.0). See `LICENSE-APACHE` and `LICENSE-MIT` for details.