25 releases (13 breaking)

0.16.0 Jun 6, 2024
0.14.0 Apr 24, 2024
0.12.0 Apr 17, 2023
0.11.0 Nov 13, 2022
0.9.0 Jul 19, 2022

#167 in Rust patterns

MIT license

150KB
3K SLoC

Parsel, the Zero-Code Parser Generator

MSRV

Parsel is a library for generating parsers directly from syntax tree node types.

The main entry point is the #[derive(Parse)] custom derive proc-macro, which generates an implementation of the syn::parse::Parse trait for the annotated AST node type. Adding #[derive(FromStr)] also implements the standard FromStr trait for the type, by simply forwarding to its Parse impl.

In addition, a #[derive(ToTokens)] macro is provided, for easily obtaining the source representation of a specific AST node via the quote crate. This in turn helps with getting its Span due to the blanket impl<T: ToTokens> Spanned for T. Adding #[derive(Display)] also implements the standard Display trait for the type, by simply forwarding to its ToTokens impl.

Furthermore, the ast module provides a number of helper types for common needs, such as optional productions, repetition, parenthesization, and grouping. These are mostly lightweight wrappers around parsing collections and parsing logic already provided by syn. However, some very useful syn types, such as Option<T: Parse> and Punctuated, have multiple, equally valid parses, so they don't implement Parse in order to avoid amibiguity. Parsel handles this ambiguity at the type level, by splitting the set of valid parses into multiple, unambiguously parseable types.

Examples and How It Works

The fundamental idea behind Parsel is the observation that structs and enums directly correspond to sequences and alternation in grammars, and that they are composable: one does not need to know the exact implementation of sub-expressions in order to produce a parser for the current rule.

AST nodes that have a struct type correspond to sequences: every field (whether named or numbered) will be parsed and populated one after another, in the order specified in the source.

AST nodes having an enum type correspond to alternation: their variants will be tried in order, and the first one that succeeds will be returned. Fields of tuple and struct variants are treated in the same sequential manner as struct fields.

Accordingly, you define your grammar by specifying the fields and variants of AST nodes, and Parsel will generate a parser from them. Let's see what this looks like in the context of the parser and the printer for a simple, JSON-like language:

use core::iter::FromIterator;
use core::convert::TryFrom;
use parsel::{Parse, ToTokens};
use parsel::ast::{Bracket, Brace, Punctuated, LitBool, LitInt, LitFloat, LitStr};
use parsel::ast::token::{Comma, Colon};

mod kw {
    parsel::custom_keyword!(null);
}

#[derive(PartialEq, Eq, Debug, Parse, ToTokens)]
enum Value {
    Null(kw::null),
    Bool(LitBool),
    Int(LitInt),
    Float(LitFloat),
    Str(LitStr),
    Array(
        #[parsel(recursive)]
        Bracket<Punctuated<Value, Comma>>
    ),
    Object(
        #[parsel(recursive)]
        Brace<Punctuated<KeyValue, Comma>>
    ),
}

#[derive(PartialEq, Eq, Debug, Parse, ToTokens)]
struct KeyValue {
    key: LitStr,
    colon: Colon,
    value: Value,
}

let actual: Value = parsel::parse_quote!({
    "key1": "string value",
    "other key": 318,
    "recursive": [
        1.6180,
        2.7182,
        3.1416,
        null
    ],
    "inner": {
        "nested key": true,
        "hard to write a parser": false
    }
});
let expected = Value::Object(Brace::from(Punctuated::from_iter([
    KeyValue {
        key: LitStr::from("key1"),
        colon: Colon::default(),
        value: Value::Str(LitStr::from("string value")),
    },
    KeyValue {
        key: LitStr::from("other key"),
        colon: Colon::default(),
        value: Value::Int(LitInt::from(318)),
    },
    KeyValue {
        key: LitStr::from("recursive"),
        colon: Colon::default(),
        value: Value::Array(Bracket::from(Punctuated::from_iter([
            Value::Float(LitFloat::try_from(1.6180).unwrap()),
            Value::Float(LitFloat::try_from(2.7182).unwrap()),
            Value::Float(LitFloat::try_from(3.1416).unwrap()),
            Value::Null(kw::null::default()),
        ]))),
    },
    KeyValue {
        key: LitStr::from("inner"),
        colon: Colon::default(),
        value: Value::Object(Brace::from(Punctuated::from_iter([
            KeyValue {
                key: LitStr::from("nested key"),
                colon: Colon::default(),
                value: Value::Bool(LitBool::from(true)),
            },
            KeyValue {
                key: LitStr::from("hard to write a parser"),
                colon: Colon::default(),
                value: Value::Bool(LitBool::from(false)),
            },
        ]))),
    },
])));

assert_eq!(actual, expected);

Recursive AST Nodes and Cyclic Constraints

Most useful real-world grammars are recursive, i.e., they contain productions that refer to themselves directly (direct recursion) or indirectly (mutual recursion). This results in AST node types that contain pointers to the same type. Even more importantly, it leads to cyclic constraints in the implementations of Parse and ToTokens. These cyclic constraints are trivially satisfied and resolvable, but the constraint solver of the Rust compiler is currently struggling with them due to Issue #48214.

Thus, one must break such constraint cycles when deriving the implementations of Parse and ToTokens. Parsel supports this use case by providing the attribute #[parsel(recursive)], or an equivalent spelling, #[parsel(recursive = true)]. Adding this attribute to a field of a struct or a variant of an enum has the effect of omitting all FieldType: Parse and FieldType: ToTokens constraints from the where clause of the generated Parse and ToTokens impls, breaking the constraint cycle, and thus allowing the code to compile.

It is sufficient to break each constraint cycle on one single type (practically on the one that requires adding the smallest number of #[parsel(recursive)] annotations). However, if the grammar contains several self-referential cycles, it is necessary to break each of them. Furthermore, if breaking a cycle requires omitting a constraint on a type which appears in multiple fields of a struct or a variant, then it is necessary to add #[parsel(recursive)] to all of those fields.

As an example, consider the following grammar for simple Boolean operations and the accompanying comments:

use parsel::{Parse, ToTokens};
use parsel::ast::{Paren, LitBool};
use parsel::ast::token::{Or, And, Not};

#[derive(PartialEq, Eq, Debug, Parse, ToTokens)]
enum Expr {
    Or {
        lhs: Conjunction,
        op: Or,
        #[parsel(recursive)] // break direct recursion
        rhs: Box<Expr>,
    },
    Conjunction(Conjunction),
}

#[derive(PartialEq, Eq, Debug, Parse, ToTokens)]
enum Conjunction {
    And {
        lhs: Term,
        op: And,
        #[parsel(recursive)] // break direct recursion
        rhs: Box<Conjunction>,
    },
    Term(Term),
}

#[derive(PartialEq, Eq, Debug, Parse, ToTokens)]
enum Term {
    Literal(LitBool),
    Not(
        Not,
        #[parsel(recursive)] // break direct recursion
        Box<Term>,
    ),
    Group(
        #[parsel(recursive)] // break mutual recursion
        Paren<Box<Expr>>
    ),
}

let expr: Expr = parsel::parse_str("true & (false | true & true) & !false").unwrap();

assert_eq!(
    expr,
    Expr::Conjunction(Conjunction::And {
        lhs: Term::Literal(LitBool::from(true)),
        op: And::default(),
        rhs: Box::new(Conjunction::And {
            lhs: Term::Group(Paren::from(Box::new(Expr::Or {
                lhs: Conjunction::Term(Term::Literal(LitBool::from(false))),
                op: Or::default(),
                rhs: Box::new(Expr::Conjunction(Conjunction::And {
                    lhs: Term::Literal(LitBool::from(true)),
                    op: And::default(),
                    rhs: Box::new(Conjunction::Term(Term::Literal(LitBool::from(true)))),
                }))
            }))),
            op: And::default(),
            rhs: Box::new(Conjunction::Term(Term::Not(
                Not::default(),
                Box::new(Term::Literal(LitBool::from(false))),
            ))),
        })
    })
);

Dealing with Left Recursion

If you carefully examine the grammar, you can notice it's right-recursive, i.e., the subexpression with identical precedence appears on the right-hand side, while the left-hand side descends one level to the next tightest-binding subexpression. This in turn means that consecutive operations of equal precedence will associate to the right. The reason for this is that recursive descent parsers, such as the ones generated by Parsel, fall into infinite recursion if they attempt parsing a left-recursive grammar. For instance, if our top-level expression were defined as

expr = expr '|' conjunction
     | conjunction

then the code generated for expr would immediately and unconditionally try to call itself again.

While it is fine to rewrite the grammar as right-recursive in the case of simple Boolean expressions (since they are associative), it is generally not possible to just omit left recursion altogether from a grammar. Operations which are not associative care a lot about how they are grouped, and even e.g. basic algebraic operations such as subtraction and division are defined to be left-associative by widespread convention. Thus, it is required that Parsel support associating terms to the left. There are two ways to achieve this goal:

  1. Side-step the problem by simply not representing associativity in the AST. This is done by using a helper type capable of expressing explicit repetition of arbitrary length (e.g., Separated), instead of creating binary AST nodes. The repeated AST nodes will be sub-expressions at the next highest precedence level. This approach puts off the question of associativity until evaluation/codegen, that is, until tree-walking time.

  2. Use the LeftAssoc helper type. This solves the problem of infinite recursion by parsing iteratively (just like Separated). It then transforms the resulting linear list of subexpressions into a properly left-associative (left-leaning) tree of AST nodes.

    Note that there is an analogous RightAssoc type as well. Strictly speaking, this is not necessary, because right recursion makes progress and terminates just fine. However, deriving the parse tree in an iterative manner has the advantage of recursing less, and including the right-leaning counterpart is preferable for reasons of symmetry/consistency.

Span and Error Reporting

Types that implement ToTokens get an automatic impl Spanned for T: ToTokens. This means that by default, all types deriving ToTokens will also report their span correctly, and parse errors will have useful span information.

However, there is an important caveat regarding alternations (enums) in the grammar. The way alternations can be parsed in a fully automatic and deceptively simple way is by attempting to parse each alternative production, one after the other, and pick the first one that parses successfully. However, if none of them parses, then it is not obvious to the parser which of the errors it should report.

The heuristic we use to solve this problem is that we use Span information to select the variant that got furthest in the token stream before having failed. This works because most "nice" context-free grammars are constant lookahead, or even better, LL(1), i.e. single-token lookahead. This means that if a production involving more than one token fails in the middle, it will have advanced further in the stream than other productions, which failed right at the very first token.

However, if span information is not available or not useful (i.e., when every production is spanned to the same Span::call_site() source location), then this heuristic breaks down, and it will select an arbitrary production, resulting in subpar error messages. This means that you should try to preserve spans as much as possible. This in turn implies that using syn::parse_str() for parsing code outside procedural macros is preferable to using syn::parse2(), because the former will result in a usefully-spanned AST, while the latter will not, at least not when used on e.g. a TokenStream obtained via quote!() or parse_quote!().

Roadmap, TODOs

  • Document all of the public API
  • Document all of the non-public API as well
  • Allow specifying custom error messages for each production/AST node type
    • Allow conditions/sorting criteria/other customization for discovering the best production/error message to report when parsing alternation
  • enum Either AST helper type for basic binary alternation
  • Any AST helper type for parsing until a given production succeeds. Unlike Many, it doesn't require the productions to extend until end-of-input.
  • Implement AsRef, Deref, and Borrow consistently for wrapper types (e.g., Paren, Bracket, Brace)
  • Make the error reporting heuristic for alternation (based on the furthest parsing production) work even when span information is not useful. Nota bene: this absolutely shouldn't be done by just counting the number of tokens/bytes in the remaining input, because that will result in an accidentally quadratic parsing performance!

Dependencies

~3.5MB
~69K SLoC