Relit: Typed Literal Macros for Reason

Reason is an increasingly popular alternative syntax for OCaml designed by engineers at Facebook to make OCaml more notationally comfortable for contemporary programmers. However, Reason, following OCaml, builds in literal notation for only a few common data structures, e.g. list literals like [x, y, z], array literals like [|x, y, z|], and JSX literals, which support an extension of HTML notation. This approach is unsatisfying because there are many other possible data structures for which literal notation might be useful, e.g. for finite maps, regular expressions, SQL queries, syntax tree representations, and chemical structures expressed using SMILES notation, to name just a few possibilities.

In our ICFP 2018 paper (.bib, video, slides), we address this problem by introducing typed literal macros (TLMs) into Reason. TLMs allow the programmer to define new literal notation, of nearly arbitrary design, for expressions and patterns of any type at all.

Tutorial: Regex Notation

For example, say that we have defined a recursive datatype Regex.t classifying simple regular expressions:

  module Regex = {
    type t =
      | Empty
      | AnyChar
      | Str(string)
      | Seq(t, t)
      | Or(t, t)
      | Star(t);
  };

Applying these constructors directly is notationally costly, so let's define a TLM named $regex (pronounced "lit regex") that implements the familiar POSIX-style regex notation. The definition of $regex, which we place by convention into a module named Regex_notation, is outlined below (note that GitHub does not yet know how to highlight our extensions to Reason).

module Regex_notation = {
  notation $regex at Regex.t { /* ... full definition given under "TLM Definitions" below ... */ }
};

The client programmer can apply Regex_notation.$regex as follows to construct a value of type Regex.t:

let r = Regex_notation.$regex `(a*bb(b|a)b)`;

The applied TLM is responsible at compile-time for parsing and expanding the literal body, here a*bb(b|a)b, to an OCaml expression. The literal body can be any character sequence as long as any occurrences of the outer delimiters, `( and )`, are balanced. In this case, the expansion is the following expression, which is clearly more notationally costly (by a variety of measures) than the TLM application above:

Regex.(Seq(Star(Str("a")), Seq(Str("b"), Seq(Str("b"), Seq(Or(Str("b"), Str("a")), Str("b")))))

Abbreviations

To make TLM applications even more concise, we can open Regex_notation as usual to bring $regex into scope:

open Regex_notation;
let r = $regex `(a*bb(b|a)b)`;

or define the abbreviation $r for Regex_notation.$regex:

notation $r = Regex_notation.$regex;
let r = $r `(a*bb(b|a)b)`;

or implicitly apply Regex_notation.$regex to all bare literals in scope of the open notation directive:

open notation Regex_notation.$regex;
let r = `(a*bb(b|a)b)`;

or use the alternative parenthesis-delimited version of open notation for the same purpose:

let r = Regex_notation.$regex.( `(a*bb(b|a)b)` )

Splicing

Sometimes we want to construct a regex value compositionally, i.e. by "splicing together" other values. To support this, $regex recognizes the notation $(e) for a spliced regex value, and $$(e) for a spliced string value, where e is a Reason expression of arbitrary form (so e might even itself apply TLMs). For example, we can splice one regex, DNA.any_base, into another, bisI (BisI is a restriction enzyme, see here), as follows:

open notation Regex_notation.$regex;
module DNA = {
  let any_base = `(A|T|G|C)`;
};
let bisI = `(GC$(DNA.any_base)GC)`;

Each TLM decides for itself how it recognizes spliced expressions.

Keep in mind that the literal body is expanded at compile-time, so using TLMs together with composite representations of data structures like regexes and SQL queries can help programmers avoid string injection attacks without giving up the notational benefits of string representations.

Splicing is also sometimes called interpolation because it generalizes string interpolation as featured in many languages. Splicing is also sometimes called unquotation or antiquotation because it generalizes the unquotation forms in code quotation systems, like those in various Lisp dialects and many other languages.

Typing, Hygiene and Segmentation

User-defined notation is great when you are familiar with it, but what about when you encounter an unfamiliar notation? TLMs were carefully designed to be uniquely reasonable in this situation. In particular, you do not need to peek at the generated expansion or the details of the parser to reason about types and binding in a program that uses TLMs. Instead, the system maintains the following important abstract reasoning principles:

• Expansion Typing: Each notation definition specifies a type annotation—at Regex.t on $regex above—that determines the type of the generated expansion.

• Context Independence: The expansion is guaranteed to be context independent, meaning that it does not make any assumptions about which variables (including module variables) are in scope at the application site. Therefore, clients can rename variables and manage imports without thinking about the expansion's dependencies. For example, the Regex module can be shadowed, or even out of scope entirely, when applying $regex, even though the expansion uses the constructors defined in the Regex module (see below for more on how dependencies are managed).

• Capture Avoidance: Spliced expressions are capture avoiding, meaning that any variables that appear in a spliced expression cannot capture bindings internal to the expansion. Consider the following example:

  let tmp = DNA.any_base;
  let bisI = $regex `(GC$(tmp)GC)`

  Even if the expansion generated by the TLM above happens to bind a variable named tmp for internal use, the system ensures that the reference to tmp in the spliced expression will always refer to the binding of tmp on the first line.

  The context independence and capture avoidance principles together are referred to as the hygiene principles. Relit is strictly hygienic—there is no way for a TLM to opt out of these restrictions.

• Segmentation: Spliced expressions must be separated by at least one character. This ensures that there is always a unique segmentation of every literal body into spliced expressions and characters parsed in some other way by the TLM.

• Segment Typing: Each spliced expression is also labeled with an expected type by the applied TLM. This information is currently used when reporting type errors. In the future, we expect to convey the segmentation and segment typing information interactively within the program editor.

The ICFP paper investigates these reasoning principles in formal detail (i.e. with a typed lambda calculus and proofs).

TLM Definitions

Let us now consider the full definition of Regex_notation.$regex, given below.

module Regex_notation = {
  notation $regex at Regex.t {
    lexer Regex_parser.Lexer
    parser Regex_parser.Parser.start
    in package regex_parser;
    dependencies = {
      module Regex = Regex;
    };
  };
};

Scoping

A TLM definition can appear anywhere a module definition can appear, and TLM definitions follow the same scoping rules as modules (internally, they are implemented as modules with singleton signatures; see the paper).

Lexing and Parsing

Each TLM must specify a lexer, here Regex_parser.Lexer, and a parser, here Regex_parser.Parser.start, where start is the name of the starting non-terminal. The lexer and parser will be loaded and invoked at compile-time. To cleanly facilitate this, the lexer and parser must be packaged into a named ocamlfind package, here indicated by in package regex_parser.

The lexer must be generated by (or satisfy the same interface as lexers generated by) ocamllex and the parser must be generated by (or satisfy the same interface as parsers generated by) Menhir, which is a modernized derivative of ocamlyacc. These are the most popular and mature lexer and parser generators within the OCaml ecosystem, and notably, Reason itself is implemented using these same generators. Chapter 16 of Real World OCaml nicely introduces both.

We will not detail the regex lexer and parser definitions here, but the ICFP paper (Sec. 2.2) does cover them. The full definitions can be found alongside the rest of the definitions above in the example_tlms/regex_example directory. For the most part, they are entirely standard lexer and parser definitions. The only interesting bit has to do with splicing: the paper describes how splicing is implemented at the level of the lexer by invoking a helper function, Relit.Segment.read_to, in the relit_helper package. Ultimately, the parser generates standard OCaml parse trees, with a special representation for tracking spliced expressions (see paper). We rely on the metaquot library to make the generation of parse trees notationally tractable. (In the future, we might switch to TLMs as suggested in the paper, but the existing library is more mature.)

Dependencies

Each TLM definition also provides a listing of expansion dependencies, i.e. types and modules from the definition site that expansions generated by the parser might need access to (other than Pervasives, which are implicitly opened). In the example above, there is a single dependency on the Regex module, which expansions can refer to internally also as Regex (the internal name can differ in general). The system ensures that the dependencies are available at all application sites, including those where Regex might be unbound or bound to a different module. This maintains the context independence principle described above.

The paper further motivates this design decision, but briefly, explicit dependencies serve to ensure that renamings need not propagate into the parse trees constructed in the parser (where variables are represented using strings), and it also serves to maintain the abstraction discipline of the OCaml module system (making all bindings at the definition site implicitly available at application sites would require violating abstraction).

More Examples and Tests

We've got an examples_tlms directory that is the home of several example notations we've defined using Relit.

To see some simple uses of these examples, look under simple_tests. To build and run them:

make {simple_example,spliced_ocaml,splice_in_splice,my_first_timeline}

Run make to run the full test suite. Those tests are pretty clear to read! Check out test/test.ml.

There is also a standlone repository that defines a TLM for lists and can be used as a project template for TLM providers.

Installation

Requirements

Make sure you have opam and OCaml 4.07.x by running opam switch.

via opam

Run opam update && opam install ppx_relit. This will pull Relit from opam, which will also install Relit's opam dependencies.

via git

An alternative way to install Relit is via git. This allows you to run the tests and examples in this repository.

git clone https://github.com/cyrus-/relit
cd relit
make deps
make install
make test

Usage

To use Relit in your own project, you just have to include the Relit ppx, ppx_relit, in your build process. With dune (which we strongly recommend), use (preprocess (staged_pps ppx_relit)) as you can see in simple_tests/dune. Note that you need to use the staged_pps option.

What about OCaml??

Relit does not support OCaml at the moment. There's nothing particularly Reason-specific, aside from the fact that we have only forked the Reason parser and not the OCaml one. We're planning to add OCaml support sometime soon. 😄

How It Works

The ppx execution starts in ppx_relit/ppx_relit.ml at the very last line of the file. Generally reading up from there will give you a good idea of what's going on, and specifically the function relit_expansion_pass is supposed to provide a high-level overview.

There's also a four page talk proposal and slides that cover the implementation in a little more detail. We're happy to discuss the implementation with anyone who wants to work on it or implement TLMs in another language!

Debug Mode

Relit does provide a way to peek at the underlying expansion of macros when the need arises.

Setting the environment variable RELIT_DEBUG=true within the build environment will trigger the Relit PPX to print its fully-expanded AST to stderr. For example, a file that looks like :

open Regex_example;

let regex = Regex_notation.$regex `(a|b|c)`;

let () = print_endline(Regex.show(regex));

will cause the Relit PPX to print out the following (comments added here for clarity):

open Regex_example;

/* type annotation (moved to the let binding by the pretty-printer) ensures
   that the expansion is of the expected type */
let regex: Regex_notation.RelitInternalDefn_regex.t =
  (
    /* open this first to avoid edge case where path to TLM is through a 
       shadowed module in Pervasives */
    [@warning "-33"] /* suppresses warnings when a dependency is not used */
    Regex_notation.RelitInternalDefn_regex.( 
      [@warning "-33"]
      Pervasives.(
        [@warning "-33"]
        Dependencies.( /* open the dependencies */
          () => /* no spliced expressions, so this is an empty argument list */
            /* the generated expansion itself */
            Regex.Or(
              Regex.Or(Regex.Str("a"), Regex.Str("b")),
              Regex.Str("c"),
            )
        )
      )
    )
  )();

let () = print_endline(Regex.show(regex));

This ends up showing a lot of the implementation details of Relit. Relit is designed to ensure that TLM readers and users should rarely have to look at the expansion of a TLM application. Debug mode is mainly targeted towards authors of TLM definitions: it allows TLM writers to debug their parsers easily.

Current Limitations

• Relit does not yet implement pattern TLMs.

• Relit does not yet implement parametric TLMs. As a somewhat awkward workaround, you can use parameterized modules (functors), e.g. see this example of a TLM for lists.

• Relit doesn't work within the toploop/rtop/utop.

• The following warning will come up a lot:

  [WARNING] Interface topdirs.cmi occurs in several directories: /home/cyrus/.opam/4.06.1/lib/ocaml/compiler-libs, /home/cyrus/.opam/4.06.1/lib/ocaml
  

  It is harmless. It is due to a bug in OCaml.

Contributors

• Charles Chamberlain
• Cyrus Omar