rust/docs/RFC.md

• Feature Name: libsyntax2.0
• Start Date: 2017-12-30
• RFC PR: (leave this empty)
• Rust Issue: (leave this empty)

I think the lack of reusability comes in object-oriented languages, not functional languages. Because the problem with object-oriented languages is they’ve got all this implicit environment that they carry around with them. You wanted a banana but what you got was a gorilla holding the banana and the entire jungle.

If you have referentially transparent code, if you have pure functions — all the data comes in its input arguments and everything goes out and leave no state behind — it’s incredibly reusable.

Joe Armstrong

Summary

The long-term plan is to rewrite libsyntax parser and syntax tree data structure to create a software component independent of the rest of rustc compiler and suitable for the needs of IDEs and code editors. This RFCs is the first step of this plan, whose goal is to find out if this is possible at least in theory. If it is possible, the next steps would be a prototype implementation as a crates.io crate and a separate RFC for integrating the prototype with rustc, other tools, and eventual libsyntax removal.

Note that this RFC does not propose to stabilize any API for working with rust syntax: the semver version of the hypothetical library would be 0.1.0. It is intended to be used by tools, which are currently closely related to the compiler: rustc, rustfmt, clippy, rls and hypothetical rustfix. While it would be possible to create third-party tools on top of the new libsyntax, the burden of adopting to breaking changes would be on authors of such tools.

Motivation

There are two main drawbacks with the current version of libsyntax:

• It is tightly integrated with the compiler and hard to use independently

• The AST representation is not well-suited for use inside IDEs

IDE support

There are several differences in how IDEs and compilers typically treat source code.

In the compiler, it is convenient to transform the source code into Abstract Syntax Tree form, which is independent of the surface syntax. For example, it's convenient to discard comments, whitespaces and desugar some syntactic constructs in terms of the simpler ones.

In contrast, IDEs work much closer to the source code, so it is crucial to preserve full information about the original text. For example, IDE may adjust indentation after typing a } which closes a block, and to do this correctly, IDE must be aware of syntax (that is, that } indeed closes some block, and is not a syntax error) and of all whitespaces and comments. So, IDE suitable AST should explicitly account for syntactic elements, not considered important by the compiler.

Another difference is that IDEs typically work with incomplete and syntactically invalid code. This boils down to two parser properties. First, the parser must produce syntax tree even if some required input is missing. For example, for input fn foo the function node should be present in the parse, despite the fact that there is no parameters or body. Second, the parser must be able to skip over parts of input it can't recognize and aggressively recover from errors. That is, the syntax tree data structure should be able to handle both missing and extra nodes.

IDEs also need the ability to incrementally reparse and relex source code after the user types. A smart IDE would use syntax tree structure to handle editing commands (for example, to add/remove trailing commas after join/split lines actions), so parsing time can be very noticeable.

Currently rustc uses the classical AST approach, and preserves some of the source code information in the form of spans in the AST. It is not clear if this structure can full fill all IDE requirements.

Reusability

In theory, the parser can be a pure function, which takes a &str as an input, and produces a ParseTree as an output.

This is great for reusability: for example, you can compile this function to WASM and use it for fast client-side validation of syntax on the rust playground, or you can develop tools like rustfmt on stable Rust outside of rustc repository, or you can embed the parser into your favorite IDE or code editor.

This is also great for correctness: with such simple interface, it's possible to write property-based tests to thoroughly compare two different implementations of the parser. It's also straightforward to create a comprehensive test suite, because all the inputs and outputs are trivially serializable to human-readable text.

Another benefit is performance: with this signature, you can cache a parse tree for each file, with trivial strategy for cache invalidation (invalidate an entry when the underling file changes). On top of such a cache it is possible to build a smart code indexer which maintains the set of symbols in the project, watches files for changes and automatically reindexes only changed files.

Unfortunately, the current libsyntax is far from this ideal. For example, even the lexer makes use of the FileMap which is essentially a global state of the compiler which represents all know files. As a data point, it turned out to be easier to move rustfmt into the main rustc repository than to move libsyntax outside!

Guide-level explanation

Not applicable.

Reference-level explanation

It is not clear if a single parser can accommodate the needs of the compiler and the IDE, but there is hope that it is possible. The RFC proposes to develop libsynax2.0 as an experimental crates.io crate. If the experiment turns out to be a success, the second RFC will propose to integrate it with all existing tools and rustc.

Next, a syntax tree data structure is proposed for libsyntax2.0. It seems to have the following important properties:

• It is lossless and faithfully represents the original source code, including explicit nodes for comments and whitespace.

• It is flexible and allows to encode arbitrary node structure, even for invalid syntax.

• It is minimal: it stores small amount of data and has no dependencies. For instance, it does not need compiler's string interner or literal data representation.

• While the tree itself is minimal, it is extensible in a sense that it possible to associate arbitrary data with certain nodes in a type-safe way.

It is not clear if this representation is the best one. It is heavily inspired by PSI data structure which used in IntelliJ based IDEs and in the Kotlin compiler.

Untyped Tree

The main idea is to store the minimal amount of information in the tree itself, and instead lean heavily on the source code for the actual data about identifier names, constant values etc.

All nodes in the tree are of the same type and store a constant for the syntactic category of the element and a range in the source code.

Here is a minimal implementation of this data structure with some Rust syntactic categories

#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
pub struct NodeKind(u16);

pub struct File {
	text: String,
	nodes: Vec<NodeData>,
}

struct NodeData {
	kind: NodeKind,
	range: (u32, u32),
	parent: Option<u32>,
	first_child: Option<u32>,
	next_sibling: Option<u32>,
}

#[derive(Clone, Copy)]
pub struct Node<'f> {
	file: &'f File,
	idx: u32,
}

pub struct Children<'f> {
	next: Option<Node<'f>>,
}

impl File {
	pub fn root<'f>(&'f self) -> Node<'f> {
		assert!(!self.nodes.is_empty());
		Node { file: self, idx: 0 }
	}
}

impl<'f> Node<'f> {
	pub fn kind(&self) -> NodeKind {
		self.data().kind
	}

	pub fn text(&self) -> &'f str {
		let (start, end) = self.data().range;
		&self.file.text[start as usize..end as usize]
	}

	pub fn parent(&self) -> Option<Node<'f>> {
		self.as_node(self.data().parent)
	}

	pub fn children(&self) -> Children<'f> {
		Children { next: self.as_node(self.data().first_child) }
	}

	fn data(&self) -> &'f NodeData {
		&self.file.nodes[self.idx as usize]
	}

	fn as_node(&self, idx: Option<u32>) -> Option<Node<'f>> {
		idx.map(|idx| Node { file: self.file, idx })
	}
}

impl<'f> Iterator for Children<'f> {
	type Item = Node<'f>;

	fn next(&mut self) -> Option<Node<'f>> {
		let next = self.next;
		self.next = next.and_then(|node| node.as_node(node.data().next_sibling));
		next
	}
}

pub const ERROR: NodeKind = NodeKind(0);
pub const WHITESPACE: NodeKind = NodeKind(1);
pub const STRUCT_KW: NodeKind = NodeKind(2);
pub const IDENT: NodeKind = NodeKind(3);
pub const L_CURLY: NodeKind = NodeKind(4);
pub const R_CURLY: NodeKind = NodeKind(5);
pub const COLON: NodeKind = NodeKind(6);
pub const COMMA: NodeKind = NodeKind(7);
pub const AMP: NodeKind = NodeKind(8);
pub const LINE_COMMENT: NodeKind = NodeKind(9);
pub const FILE: NodeKind = NodeKind(10);
pub const STRUCT_DEF: NodeKind = NodeKind(11);
pub const FIELD_DEF: NodeKind = NodeKind(12);
pub const TYPE_REF: NodeKind = NodeKind(13);

Here is a rust snippet and the corresponding parse tree:

struct Foo {
	field1: u32,
	&
	// non-doc comment
	field2:
}

FILE
  STRUCT_DEF
    STRUCT_KW
    WHITESPACE
    IDENT
    WHITESPACE
    L_CURLY
    WHITESPACE
    FIELD_DEF
      IDENT
      COLON
      WHITESPACE
      TYPE_REF
        IDENT
    COMMA
    WHITESPACE
    ERROR
      AMP
    WHITESPACE
    FIELD_DEF
      LINE_COMMENT
      WHITESPACE
      IDENT
      COLON
      ERROR
    WHITESPACE
    R_CURLY

Note several features of the tree:

• All whitespace and comments are explicitly accounted for.

• The node for STRUCT_DEF contains the error element for &, but still represents the following field correctly.

• The second field of the struct is incomplete: FIELD_DEF node for it contains an ERROR element, but nevertheless has the correct NodeKind.

• The non-documenting comment is correctly attached to the following field.

Typed Tree

It's hard to work with this raw parse tree, because it is untyped: node containing a struct definition has the same API as the node for the struct field. But it's possible to add a strongly typed layer on top of this raw tree, and get a zero-cost AST. Here is an example which adds type-safe wrappers for structs and fields:

// generic infrastructure

pub trait AstNode<'f>: Copy + 'f {
	fn new(node: Node<'f>) -> Option<Self>;
	fn node(&self) -> Node<'f>;
}

pub fn child_of_kind<'f>(node: Node<'f>, kind: NodeKind) -> Option<Node<'f>> {
	node.children().find(|child| child.kind() == kind)
}

pub fn ast_children<'f, A: AstNode<'f>>(node: Node<'f>) -> Box<Iterator<Item=A> + 'f> {
	Box::new(node.children().filter_map(A::new))
}

// AST elements, specific to Rust

#[derive(Clone, Copy)]
pub struct StructDef<'f>(Node<'f>);

#[derive(Clone, Copy)]
pub struct FieldDef<'f>(Node<'f>);

#[derive(Clone, Copy)]
pub struct TypeRef<'f>(Node<'f>);

pub trait NameOwner<'f>: AstNode<'f> {
	fn name_ident(&self) -> Node<'f> {
		child_of_kind(self.node(), IDENT).unwrap()
	}

	fn name(&self) -> &'f str { self.name_ident().text() }
}


impl<'f> AstNode<'f> for StructDef<'f> {
	fn new(node: Node<'f>) -> Option<Self> {
		if node.kind() == STRUCT_DEF { Some(StructDef(node)) } else { None }
	}
	fn node(&self) -> Node<'f> { self.0 }
}

impl<'f> NameOwner<'f> for StructDef<'f> {}

impl<'f> StructDef<'f> {
	pub fn fields(&self) -> Box<Iterator<Item=FieldDef<'f>> + 'f> {
		ast_children(self.node())
	}
}


impl<'f> AstNode<'f> for FieldDef<'f> {
	fn new(node: Node<'f>) -> Option<Self> {
		if node.kind() == FIELD_DEF { Some(FieldDef(node)) } else { None }
	}
	fn node(&self) -> Node<'f> { self.0 }
}

impl<'f> FieldDef<'f> {
	pub fn type_ref(&self) -> Option<TypeRef<'f>> {
		ast_children(self.node()).next()
	}
}

impl<'f> NameOwner<'f> for FieldDef<'f> {}


impl<'f> AstNode<'f> for TypeRef<'f> {
	fn new(node: Node<'f>) -> Option<Self> {
		if node.kind() == TYPE_REF { Some(TypeRef(node)) } else { None }
	}
	fn node(&self) -> Node<'f> { self.0 }
}

Note that although AST wrappers provide a type-safe access to the tree, they are still represented as indexes, so clients of the syntax tree can easily associated additional data with AST nodes by storing it in a side-table.

Missing Source Code

The crucial feature of this syntax tree is that it is just a view into the original source code. And this poses a problem for the Rust language, because not all compiled Rust code is represented in the form of source code! Specifically, Rust has a powerful macro system, which effectively allows to create and parse additional source code at compile time. It is not entirely clear that the proposed parsing framework is able to handle this use case, and it's the main purpose of this RFC to figure it out. The current idea for handling macros is to make each macro expansion produce a triple of (expansion text, syntax tree, hygiene information), where hygiene information is a side table, which colors different ranges of the expansion text according to the original syntactic context.

Implementation plan

This RFC proposes huge changes to the internals of the compiler, so it's important to proceed carefully and incrementally. The following plan is suggested:

• RFC discussion about the theoretical feasibility of the proposal, and the best representation representation for the syntax tree.

• Implementation of the proposal as a completely separate crates.io crate, by refactoring existing libsyntax source code to produce a new tree.

• A prototype implementation of the macro expansion on top of the new sytnax tree.

• Additional round of discussion/RFC about merging with the mainline compiler.

Drawbacks

• No harm will be done as long as the new libsyntax exists as an experiemt on crates.io. However, actually using it in the compiler and other tools would require massive refactorings.

• It's difficult to know upfront if the proposed syntax tree would actually work well in both the compiler and IDE. It may be possible that some drawbacks will be discovered during implementation.

Rationale and alternatives

• Incrementally add more information about source code to the current AST.

• Move the current libsyntax to crates.io as is. In the past, there were several failed attempts to do that.

• Explore alternative representations for the parse tree.

• Use parser generator instead of hand written parser. Using the parser from libsyntax directly would be easier, and hand-written LL-style parsers usually have much better error recovery than generated LR-style ones.

Unresolved questions

• Is it at all possible to represent Rust parser as a pure function of the source code? It seems like the answer is yes, because the language and especially macros were cleverly designed with this use-case in mind.

• Is it possible to implement macro expansion using the proposed framework? This is the main question of this RFC. The proposed solution of synthesizing source code on the fly seems workable: it's not that different from the current implementation, which synthesizes token trees.

• How to actually phase out current libsyntax, if libsyntax2.0 turns out to be a success?