3 releases
0.1.2 | Mar 17, 2019 |
---|---|
0.1.1 | Mar 11, 2019 |
0.1.0 | Mar 9, 2019 |
#2119 in Parser implementations
285KB
4K
SLoC
Kul
A unique textual notation that can be used as both a data format and a markup language and that has powerful extensibility of both lexical syntax and semantics, and a Rust library for parsing it. It is inspired by the little-known Curl language, but is only a notation like Curl's, not a programming language like Curl.
The syntax is very minimal and exists only to delimit forms of nested text that you sometimes define your own parsing and semantics of by binding operator sub-forms to combiner functions which are similar to macros. As such, by itself, it is not a complete format for most applications and they will define their own different extensions. However, it enables having a common base syntax that can still be parsed, without any extensions, into the basic structure which can be used across applications that do not need to fully handle the inner syntax of all of the nested text forms. Furthermore, shared or standardized extensions could be created in the future.
See the motivation section for why the notation is designed the way it is.
See the objectives section for what the library design achieves.
This is only an experimental exploration and is my first Rust project, but the implementation is fairly capable and featureful.
Examples of the notation
As markup of free-form text for embedding structured data types, such as a
{date 2019-03-05 15:15 PST}, which might be rendered as widgets; or for {italic
{bold styling} the text}. Only the \{, \}, and \\ characters require escaping.
{config {# As a data structure. (This form is a comment.)}
{things {list 9, "blah", {map foo: 8.7, bar: asdf}}}
{logging
{err = yes}
{warn = {maybe, if 1 + x = 3}}
}
{greeting: We can have text in structures without ugly quoting!}
{knobs this := {that}; other := {}}
{{compound-operator with arguments} 7/2 > 𝜋}
}
Using non-default delimiters:
⟪⟪source-code Rust⟫
use kul::common::inmem::parse_str;
fn main() {
let input = "Escaped the {bold non-default} delimiters: ␛⟪, ␛⟫, ␛␛";
dbg!(parse_str(input));
}
⟫
Motivation
The notation has the same advantage as Curl's design of this level, which is to have a very extensible language for both markup and data with a minimum of syntax. Unlike most markup languages, you can have your own markup forms with their own syntax, and unlike most data formats, you can have your own rich data types with their own syntax, instead of being limited to only what is provided by most others.
The flip side of this is that you have to implement your own parsing and handling of your custom forms, but premade libraries that implement form extensions could be provided by others. Because most other notations are limited and fixed, applications sometimes end up adding their own custom processing to achieve what they want anyway, which can feel like inconsistent hacks. Instead, we design the notation to fundamentally support extension so it is much cleaner and more capable in what you can express directly in it. This helps you create different domain-specific languages (DSLs) with their own lexical syntax, instead of being locked into one particular syntax like with other DSL designs. This also enables embedding multiple DSLs' different syntaxes within the same outer form or file. (Lisp languages are similar in allowing embedding multiple DSLs but they all have to use the same S-expression lexical syntax.)
At first, this might seem like it would lead to being overwhelmed with a huge variety of syntaxes and dialects. But the inherent complexity of reality that we want to model cannot be avoided and when you push that complexity out of the notation it often re-manifests in how you later try to express complex compositions. Such as jamming DSLs into the string type, or into other types like arrays and maps, or into identifiers and function signatures, because your primary notation is too limited, and then you are back to having to parse and handle your extensions but with a syntax and approach that was not designed for such extension.
The Kul notation and parser is an experiment of trying the Curl approach that is designed for extension.
For more about the motivation, see the linked paper about Curl and the other published papers about it.
Objectives
The library is designed to offer:
-
Zero-copy parsing of both entirely-in-memory inputs and of streamed inputs (if you arrange buffering into chunks appropriately), including zero-copy exclusion of the escape character. This minimizes memory usage. All achieved via a generic design of chunked text.
-
Premade
char
-and-position iterator used for all chunked text types that only borrows their internal state (to avoid copying). This greatly simplifies defining your ownText
traitimpl
s (if needed) and assists the premade ones provided. Achieved via a clever design of generically borrowing associated types with the correct lifetimes without using Rust's unstablegeneric_associated_types
feature. -
Support for very-deep huge trees of the AST type (e.g. long lists and deep nests). This enables working with deep trees produced by parsing or by your own construction. Achieved via custom
Drop
andPartialEq
implementations that avoid stack overflows (which would otherwise happen) when dropping or comparing deep trees. -
No
unsafe
code. -
No external dependencies.
-
Non-panicking API. (
panic!
s should never happen but this is not proven currently.) -
A
no_std
core crate usable on its own with only stack allocation. The parser's dynamic allocation can be done from fixed-size arrays. This can also be used to enforce memory-consumption limiting. -
A
std
-using crate with more convenientVec
s,String
s,Box
es, etc. that builds on the core crate. -
Very generically parameterized in most aspects to allow maximal reuse for diverse applications.
Status
Version 0.1.2
: experimental and unstable. Builds fine and passes all tests
and lints.
Rust version
At least 1.33
required. This library will always require only the stable
version of Rust (not the nightly one).
Usage of kul
It would require some length to describe the full range of possible usages, given how very generically parameterized the library is.
See the README
of the kul_core
crate for a usage example relevant to
no_std
applications where all allocation is done from the stack only.
For common basic applications, the following example shows how input strings can
be parsed, with and without custom extensions. It prints the resulting AST
(abstract syntax tree) structures, collected into a vector per parse, that you
would work with, so you can see the Datum
variants that correspond to the
different forms of the notation. (As an introduction, this seems clearer than
showing a lot of matching and destructuring of the variants.)
use std::{time::SystemTime, str::FromStr, iter::FromIterator};
use kul::{
common::inmem::{
parse_str, parse_str_with,
Text, OperatorBindings, DatumAllocator,
},
Combiner, Datum, datum::{BoxDatum, DatumBox}, Error, Text as _,
};
/// Parse without any bound operators and print results. This shows that the
/// common base syntax can always be parsed without knowing about possible
/// extensions.
fn no_extension() {
dbg!(parse_str("λ"));
dbg!(parse_str(r"es\{c\}ap\\es"));
dbg!(parse_str("{}"));
dbg!(parse_str("{▷}"));
dbg!(parse_str("Surrounding {{▷} λ {}} text."));
}
/// Parse with some bound operators and print results. This shows that the
/// syntax and semantics of particular forms can be extended in custom ways.
fn with_extensions()
{
/// Extends the types that may occur in the returned ASTs.
#[derive(Hash, Eq, PartialEq, Debug)]
enum MyDatumVariants {
Time(SystemTime),
Integer(i128),
}
/// Extends the types that may occur in errors returned by our custom form
/// processing.
#[derive(Debug)]
enum MyCombinerError<'input> {
Oops,
Darnit(MyDatum<'input>),
}
// Convenient type aliases.
type MyDatum<'input> = BoxDatum<Text<'input>, MyDatumVariants>;
type MyOperatorBindings<'input> =
OperatorBindings<'input, MyDatumVariants, MyCombinerError<'input>>;
type MyDatumAllocator<'input> = DatumAllocator<'input, MyDatumVariants>;
type AllocArg<'a> = &'a mut MyDatumAllocator<'static>;
// The functions that process our custom forms. Using closures can be nicer
// because at least some type inference of the argument and return types can
// be gained. The operator and allocator arguments are always ignored for
// this example, as they often are in real programs.
let comment = |_operator, _operands, _: AllocArg<'_>| {
Ok(None)
};
let pass_thru = |_operator, operands, _: AllocArg<'_>| {
Ok(Some(operands))
};
let current_time = |_operator, operands, _: AllocArg<'_>| {
if let Datum::EmptyList = operands {
Ok(Some(Datum::Extra(MyDatumVariants::Time(SystemTime::now()))))
} else {
Err(Error::FailedCombiner(MyCombinerError::Darnit(operands)))
}
};
let int = |_operator, operands: Text<'_>, _: AllocArg<'_>| {
// Must convert the operands text into a `&str`, to be able to use other
// parsing functions/libraries that take string slices. (When the other
// parsing functionality can instead take `Iterator`s of `char`s, this
// conversion is unneeded.)
let i = i128::from_str(&String::from_iter(operands.chars()))
.map_err(|_| Error::FailedCombiner(MyCombinerError::Oops))?;
Ok(Some(Datum::Extra(MyDatumVariants::Integer(i))))
};
// Establish bindings of particular operator sub-forms to our processing
// functions. Other more declarative and concise ways of doing this are
// possible, but, for this example, this shows the basic nature that other
// ways could build on.
let mut bindings = MyOperatorBindings::default();
bindings.hashmap.insert(Datum::Text(Text::from_str("#")),
Combiner::Operative(Box::new(comment)));
bindings.hashmap.insert(Datum::Text(Text::from_str("current-time")),
Combiner::Applicative(Box::new(current_time)));
bindings.hashmap.insert(Datum::Text(Text::from_str("int")),
Combiner::Operative(Box::new(int)));
let compound_operator_form =
Datum::Combination {
operator: DatumBox::new(Datum::Text(Text::from_str("compound"))),
operands: DatumBox::new(Datum::EmptyList),
};
bindings.hashmap.insert(compound_operator_form,
Combiner::Applicative(Box::new(pass_thru)));
// Parse a string that uses all of the above and print results.
dbg!(parse_str_with(
"{{compound} {current-time} {# removed} {unbound form} {int -42}}",
bindings)
);
}
fn main() {
no_extension();
with_extensions();
}
The above example can be run by doing:
cargo run --example common_basic
Documentation
The source-code has many doc comments, which are rendered as the API documentation.
View online at: http://docs.rs/kul and http://docs.rs/kul_core
Or, you can generate them yourself and view locally by doing:
cargo doc --open
See the documentation at the top of the kul_core
crate for some further
overview of the Kul language and design.
(TODO
: Some of the less-important doc comments' links are broken currently due
to refactorings changing their relative directories and names, and some just
need links to be made in the first place. I'm hoping that the
intra_rustdoc_links
feature (RFC 1946) will become
stable soon to make fixing these be much easier and much more maintainable.)
Tests
The crates have many unit tests and integration tests, which can be run by doing:
cargo test --all
That uses around 2.1G of memory running them with --test-threads=4
on my
64-bit 4-core computer. If you want it to use less, you can do:
cargo test --all -- "" tree-size=$((2**N)) # where N < 21
221 is the default for the size of the trees used to test the Drop
and PartialEq
implementations that prevent stack overflows for deep trees
(e.g. long lists and deep nests) of kul::Datum
. Below some size, it will no
longer be properly testing this because overflows would not happen without our
implementations anyway. Alternatively, you may increase N > 21
as much as you
can, to test very-deep trees. To see overflows happen, comment-out the Drop
impl
s in src/drop.rs
.
Unresolved
The following aspects are unresolved:
-
Some of the terminology chosen might be improvable.
-
The
Datum
enum type, used for the returned AST nodes and for the arguments given to extension functions, is designed to be multi-purpose so constrained applications, that do not use heap allocation and use only the coreno_std
crate, only need to provide an allocator of this one simple type, which enables just using a basic array of them on the stack. But this results in some of the API being non-ideal becauseDatum
is used where only a subset of its variants are possible, requiring matching to destructure with a branch that has to be madeunreachable!()
. I couldn't figure out a way to improve this while keeping support for constrained allocation from basic arrays of a single type. It is unresolved whether this can be improved without causing difficulties for no-heap constrained applications. -
The
Text
trait, used to represent logical concatenation of chunks of sequences ofchar
s, is designed to be multi-purpose so it can be used for AST output, and for input of streamed or in-memory sources, and for efficiently excluding escape characters, and used generically with various underlying representations. But this results in having to use iteration of its logicalchar
sequence for all operations that want to know its content, which doesn't fit perfectly with applications that want to parse its content by using other libraries that take string slices (&str
) as their inputs to parse (e.g. thestd
library'sFromStr
implementations). Creating temporaryString
s fromText
content is a reasonable way to use such libraries, but this doesn't feel great given the project's goal of supporting zero-copy as much as possible, andString
(nor any dynamic string types) is not available to no-heap constrained applications. It is unresolved whether this can be improved while keeping all the desired uses and qualities of theText
trait. (The Kul library itself preserves zero-copy while using iterators ofchar
s by using its own special iterator and related traits and types.) -
The name Kul was chosen after abandoning the name Kruvi. Kruvi is the Lojban word for curve, in homage to Curl, but it sounds too similar to the name Groovy (an existing software project) and is too long to use as a filename extension. The name Kul seems to be unclaimed as a project name and
.kul
unclaimed as a filename extension. I'm imagining that applications might use filename extensions like.whatever.kul
to indicate their particular extensions of the format as well as indicate that it can also always be parsed as the basic.kul
structure. I'm not very attached to any name and am open to changing it but do like that Kul is cool.