A set of bison skeleton files that can be used to generate a Bison grammar that is written in Rust.
Technically it's more like a Bison frontend for Rust.
- Rust
- Bison
3.7.3
or higher (or maybe a bit lower, it's unknown, better get the latest version)
bison
executable must be available in $PATH
.
Bison is a parser generator, and in fact it doesn't really care what's your programming language.
Under the hood it takes your .y
file, parses it, extracts all derivations and then constructs a bunch of tables.
Then, this data is passed to a template that is called skeleton
. Simply treat it as JSX/ERB/Handlebars/etc view template.
This skeleton is a special file written in M4 language (that is not really a programming language, it's closer to a macro engine) that
(once rendered) prints your .rs
file. As simple as that.
Just like in C/C++/Java/D templates the following directives can be configured:
%expect N
whereN
is a number of expected conflicts. Better set it to 0%define api.parser.struct {Parser}
whereParser
is the name of your parser struct. Optional,Parser
is the default name.%define api.value.type {Value}
whereValue
is the name of the derivation result (and a stack item) struct. Optional,Value
is the default name.%code use { }
allows you to specify a block of code that will be at the top of the file. Can be a multi-line block, optional, has no default value.%code parser_fields { }
allows you to specify additional custom fields for yourParser
struct. Can be a multi-line block, optional, has no default value.%define api.parser.check_debug { /* expr */ }
allows you to configure printing debug information,self
is an instance of your parser, so use something like this if you want to turn it into configurable field:
%code parser_fields {
debug: bool
}
%define api.parser.check_debug { self.debug }
All other directives that available in Bison can be configured too, read official Bison docs.
This skeleton generates an LALR(1) parser, and so parser has a stack. This stack is represented as Vec<Value>
where Value
is an enum that must be defined by you. The name of this enum must be set using %define api.value.type {}
directive.
Let's build a simple calculator that handles lines like 1 + (4 - 3) * 2
.
First, let's define a boilerplate in src/parser.y
:
%expect 0
%define api.parser.struct {Parser}
%define api.value.type {Value}
%define parse.error custom
%define parse.trace
%code use {
// all use goes here
use crate::Loc;
}
%code parser_fields {
// custom parser fields
}
%token
tPLUS "+"
tMINUS "-"
tMUL "*"
tDIV "/"
tLPAREN "("
tRPAREN ")"
tNUM "number"
%left "-" "+"
%left "*" "/"
%%
// rules
%%
impl Parser {
// parser implementation
}
enum Value {
// variants to define
}
Currently this grammar has no rules, but it's a good start.
This code (once compiled) defines a Parser
struct at the top of the file that looks like this:
#[derive(Debug)]
pub struct Parser {
pub yylexer: Lexer,
yy_error_verbose: bool,
yynerrs: i32,
yyerrstatus_: i32,
/* "%code parser_fields" blocks. */
}
Keep in mind that Parser
auto-implements std::fmt::Debug
, and so all custom fields also should implement it.
Value
enum is what is returned by derivations and what's stored in the stack of the parser. This enum must be defined by you and it has to have the following variants:
Uninitialized
- a variant that is stored in$$
by default (and what's overwritten by you)Stolen
- a variant that stack value is replaced with when you get it from the stack by writing$<N>
Token(TokenStruct)
- a variant that is used when shift if performed, holds yourTokenStruct
that is returned by a lexer
Additionally you can have as many variants as you want, however they must represent what you return from derivation rules.
In our case we want variants Number
(to represent a numeric expression) and None
(this is actually required to represent return value of the top-level rule).
#[derive(Clone, Debug)]
pub enum Value {
None,
Uninitialized,
Stolen,
Token(Token),
Number(i32),
}
impl Default for Value {
fn default() -> Self {
Self::Stolen
}
}
It must implement Clone
, Debug
and Default
(.take()
is used under the hood that swaps &mut Value
with Value::default()
, so default()
must return Stolen
variant).
Also skeleton defines a Lexer
struct with a bunch of constants representing token numbers, it looks like this:
// AUTO-GENERATED
impl Lexer {
/* Token kinds. */
// Token "end of file", to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const YYEOF: i32 = 0;
// Token error, to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const YYerror: i32 = 256;
// Token "invalid token", to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const YYUNDEF: i32 = 257;
// Token "+", to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const tPLUS: i32 = 258;
// Token "-", to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const tMINUS: i32 = 259;
// Token "*", to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const tMUL: i32 = 260;
// Token "/", to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const tDIV: i32 = 261;
// Token "(", to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const tLPAREN: i32 = 262;
// Token ")", to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const tRPAREN: i32 = 263;
// Token "number", to be returned by the scanner.
#[allow(non_upper_case_globals, dead_code)]
pub const tNUM: i32 = 264;
}
Thus, we can define our lexer logic:
use crate::{Loc, Value};
/// A token that is emitted by a lexer and consumed by a parser
#[derive(Clone)]
pub struct Token {
// Required field, used by a skeleton
pub token_type: i32,
// Optional field, used by our custom parser
pub token_value: i32,
// Required field, used by a skeleton
pub loc: Loc,
}
/// `Debug` implementation
impl std::fmt::Debug for Token {
fn fmt(&self, f: &mut std::fmt::Formatter<'_> /*' fix quotes */) -> std::fmt::Result {
f.write_str(&format!(
"[{}, {:?}, {}...{}]",
token_name(self.token_type()),
self.token_value,
self.loc.begin,
self.loc.end
))
}
}
impl Token {
/// Used by a parser to "unwrap" `Value::Token` variant into a plain Token value
pub(crate) fn from(value: Value) -> Token {
match value {
Value::Token(v) => v,
other => panic!("expected Token, got {:?}", other),
}
}
}
#[allow(non_upper_case_globals)]
impl Lexer {
pub fn new(src: &str) -> Self {
let mut tokens = vec![];
for (idx, c) in src.chars().enumerate() {
let (token_type, token_value) = match c {
'0' => (Self::tNUM, 0),
'1' => (Self::tNUM, 1),
'2' => (Self::tNUM, 2),
'3' => (Self::tNUM, 3),
'4' => (Self::tNUM, 4),
'5' => (Self::tNUM, 5),
'6' => (Self::tNUM, 6),
'7' => (Self::tNUM, 7),
'8' => (Self::tNUM, 8),
'9' => (Self::tNUM, 9),
'+' => (Self::tPLUS, -1),
'-' => (Self::tMINUS, -1),
'*' => (Self::tMUL, -1),
'/' => (Self::tDIV, -1),
'(' => (Self::tLPAREN, -1),
')' => (Self::tRPAREN, -1),
' ' => continue,
_ => panic!("unknown char {}", c),
};
let token = Token {
token_type,
token_value,
loc: Loc {
begin: idx,
end: idx + 1,
},
};
tokens.push(token)
}
tokens.push(Token {
token_type: Self::YYEOF,
token_value: 0,
loc: Loc {
begin: src.len(),
end: src.len() + 1,
},
});
Self { tokens }
}
pub(crate) fn yylex(&mut self) -> Token {
self.tokens.remove(0)
}
}
This lexer is not buffered and it does unnecessary work in case of a syntax error, but let's use at it's easier to understand.
Now let's define Parser
<-> Lexer
composition:
impl Parser {
pub fn new(lexer: Lexer) -> Self {
Self {
yy_error_verbose: true,
yynerrs: 0,
yyerrstatus_: 0,
yylexer: lexer,
}
}
fn next_token(&mut self) -> Token {
self.yylexer.yylex()
}
fn report_syntax_error(&self, ctx: &Context) {
eprintln!("syntax error: {:#?}", ctx)
}
}
Parser
encapsulates Lexer
and calls it in a next_token
method that is called by a skeleton.
Time to define rules:
%type <Number> expr number program
%%
program: expr
{
self.result = Some($<Number>1);
$$ = Value::None;
}
| error
{
self.result = None;
$$ = Value::None;
}
expr: number
{
$$ = $1;
}
| tLPAREN expr tRPAREN
{
$$ = $2;
}
| expr tPLUS expr
{
$$ = Value::Number($<Number>1 + $<Number>3);
}
| expr tMINUS expr
{
$$ = Value::Number($<Number>1 - $<Number>3);
}
| expr tMUL expr
{
$$ = Value::Number($<Number>1 * $<Number>3);
}
| expr tDIV expr
{
$$ = Value::Number($<Number>1 / $<Number>3);
}
number: tNUM
{
$$ = Value::Number($<Token>1.token_value());
}
%%
As you can see our grammar has the following rules:
program: expr
| error
expr: number
| '(' number ')'
| number '+' number
| number '-' number
| number '*' number
| number '/' number
number: [0-9]
$$
is a return value and it has type Value
. You can use $1
, $2
, etc to get items 1, 2, etc that are no unwrapped, i.e. that also have type Value
. To unwrap it you can use $<Variant>1
, but then you must have the following method:
impl Variant {
fn from(value: Value) -> Self {
match value {
Value::Variant(out) => out,
other => panic!("wrong type, expected Variant, got {:?}", other),
}
}
}
In our case we want to have only one such variant - Number
:
use crate::Value;
#[allow(non_snake_case)]
pub(crate) mod Number {
use super::Value;
pub(crate) fn from(value: Value) -> i32 {
match value {
Value::Number(out) => out,
other => panic!("wrong type, expected Number, got {:?}", other),
}
}
}
Yes, it's a mod, but that's absolutely OK. It doesn't matter what Variant
is, it's all about calling Variant::from(Value)
.
Also, as you might notice, there's a self.result = ...
assignment in the top-level rule program
. The reason why it's required is that there's no way to get value that is left on the stack because stack is not a part of the parser's state.
This is why we also need to declare it:
%code parser_fields {
result: Option<i32>,
}
// And Parser's constructor must return
fn new(lexer: Lexer) -> Self {
Self {
result: None,
// ...
}
}
Now we need a build.rs
script:
extern crate rust_bison_skeleton;
use rust_bison_skeleton::{process_bison_file, BisonErr};
use std::path::Path;
fn main() {
match process_bison_file(&Path::new("src/parser.y")) {
Ok(_) => {}
Err(BisonErr { message, .. }) => {
eprintln!("Bison error:\n{}\nexiting with 1", message);
std::process::exit(1);
}
}
}
And so after running cargo build
we should get src/parser.rs
with all auto-generated and manually written code combined into a single file.
You can find a full example in tests/src/calc.y
.
This skeleton full matches behavior of other built-in Bison skeletons:
- If you want to return an error from a derivation you can either:
- do
return Ok(Self::YYERROR);
- or just
Err(())?
- do
- If you want to completely abort execution you can:
return Ok(Self::YYACCEPT);
to abort with success-like status codereturn Ok(Self::YYABORT);
to abort with error-like status code
Once error is returned a special error
rule can catch and "swallow" it:
numbers: number
{
$$ = Value::NumbersList(vec![ $Number<1> ]);
}
| numbers number
{
$$ = Value::NumbersList( $<NumbersList>1.append($<Number>2) );
}
| error number
{
// ignore $1 and process only $<Number>2
$$ = Value::NumbersList(vec![ $Number<2> ]);
}
number: tNUM { $$ = $1 }
| tINVALID_NUM { return Ok(Self::YYERROR); }
Information about the error is automatically passed to Parser::report_syntax_error
. Context
that it takes has methods token()
and location()
, so implementation of this method can look like this:
fn report_syntax_error(&mut self, ctx: &Context) {
let token_id: usize = ctx.token().code().try_into().unwrap();
let token_name: &'static str = Lexer::TOKEN_NAMES[id];
let error_loc: &Loc = ctx.location();
eprintln!("Unexpected token {} at {:?}", token_name, loc);
}
To make Parser
generic you need to configure the following directive:
%define api.parser.generic {<T1, T2>}
This code is added to struct Parser
and impl Parser
:
struct Parser<T1, T2> {
// ...
}
impl<T1, T2> Parser<T1, T2> {
// ...
}
If you wan to specify lifetimes make sure to fix quotes with comments:
%define api.parser.generic {<'a /* 'fix quotes */, T>}
You can find a perf
example that runs a Parser
thousands times and creates a flamegraph.
$ TIMES=20000 cargo run --bin perf --release