Fathomable Rust Macros

A breakdown of the inner workings and authorship of macros in Rust

Rust macros are compile-time constructs that operate on streams of Rust language tokens.

A brief aside on compilation

What are “Rust language tokens”?

When a compiler begins compiling a program, it first reads in the source code file. For simplicity’s sake, let’s say the compiler stores that source code in a string. The next step is to step through the string, character by character, and divide it up into “tokens.”

For example, a Rust snippet like:

let foo: u32 = 30;

Might be “tokenized” into:

[
    KeywordLet,
    Identifier("foo"),
    Colon,
    Identifier("u32"),
    SingleEquals,
    NumericLiteral("30"),
    Semicolon,
]

(Note that this is a completely imaginary example.)

A macro takes a stream of tokens akin to the above as input, and also outputs a stream of tokens.1 This has some major implications:

  • Rust macros can add new code: add a trait implementation, create a new struct, write a new function, etc.
  • Rust macros cannot interact with the logic in the code (e.g. see whether a type implements a trait, call a function declared in the source, etc.), because the logic has not actually been constructed yet.

There are two main categories of Rust macros: declarative macros and procedural macros.

Declarative macros

Declarative macros can be declared and used alongside other code. They are declared using the special macro_rules! construct, and have some unique syntax:

macro_rules! my_macro {
    ($a: ident => $b: expr) => {
        fn $a() {
            println!("{}", $b);
        }
    };
    ($a: ident, $b: expr) => {
        println!("{} {}", $a, $b);
    };
}

Declarative macros accept Rust tokens as input and perform pattern matching against them. In the example above, the macro my_macro matches two different patterns:

  1. An identifier and an expression separated by a fat arrow =>, and
  2. An identifier and an expression separated by a comma ,.

This macro could be invoked like:

my_macro!(foo, 45);
my_macro!(bar => "hello");
my_macro!(baz => 9 * 8);

Nested macros and recursion

One of the most popular Rust crates, serde_json, includes a declarative macro json!(), which allows you to write JSON-like syntax in your Rust code. It returns a serde_json::Value.

json!({
    "id": 42,
    "name": {
        "first": "John",
        "last": "Zoidberg",
    },
});

As it turns out, you can put any valid Rust expression (that evaluates to a value that implements Serialize) as the value:

json!({
    "id": 21 + 21, // Computed expression
    "name": {
        "first": "John",
        "last": "Zoidberg",
    },
});

And I mean any valid Rust expression…

json!({
    "id": 21 + 21,
    "name": json!({ // This is another macro invocation!
        "first": "John",
        "last": "Zoidberg",
    }),
});

This ability extends to the code that your macro generates as well. For example, we can write a basic parser that recursively translates the logic symbols for AND (∧; ^ in code) and OR (∨; v in code) into the Rust equivalents.

macro_rules! andor {
    ($a: ident ^ $b: ident $($tail: tt)*) => {
        $a && andor!($b $($tail)*) // Recursive invocation
    };
    ($a: ident v $b: ident $($tail: tt)*) => {
        $a || andor!($b $($tail)*) // Recursive invocation
    };
    ($($a: tt)*) => {
        $($a)*
    }
}

andor!(true ^ false v false ^ true) // true && false || false && true
// => false

Because it is a potentially infinite operation, the macro recursion has a maximum depth defined by the Rust compiler.

Procedural macros

Procedural macros are written using normal Rust code (not a unique syntax), which is compiled, and then run by the compiler when invoked. For this reason, procedural macros are also sometimes called “compiler plugins.”

Because crates = compilation units, in order for a procedural macro to be compiled before its execution, procedural macros must be defined in (and subsequently exported from) a different crate from that in which they are used. These must be library crates with the following in Cargo.toml:

[lib]
proc-macro = true

Procedural macros appear in three forms, which are all invoked differently:

  • Attribute-like.
    Input: annotated item.
    Ouput replaces input. (Original input does not exist in final token stream.)

    #[my_attribute_macro]
    struct MyStruct; // This struct is the input to the macro
    
    struct AnotherStruct; // This struct is not part of the macro's input
    
  • Custom derive.
    Input: annotated item.
    Output is appended to input. (Original input still exists in final token stream.)

    #[derive(MyDeriveMacro)]
    struct MyStruct; // This struct is the input to the macro
    
    struct AnotherStruct; // This struct is not part of the macro's input
    
  • Function-like.
    Input: enclosed token stream. Delimiters are [], {}, or ().
    Output replaces input. (Original input does not exist in final token stream.)

    my_function_like_macro!(arbitrary + token : stream 00);
    // is the same as
    my_function_like_macro![arbitrary + token : stream 00];
    // is the same as
    my_function_like_macro!{arbitrary + token : stream 00};
    

In this post, I will discuss writing attribute and derive macros.

Authoring procedural macros

At first glance, writing a procedural macro from scratch can be really daunting:

use proc_macro::TokenStream;

#[proc_macro_attribute]
pub fn my_attribute_macro(attr: TokenStream, item: TokenStream) -> TokenStream {
    todo!("Good luck!")
}

This macro could be invoked like this:

#[my_attribute_macro]
struct AnnotatedItem;

In this case, the attr token stream would be empty, and the item token stream would contain the AnnotatedItem struct.

If you invoke the macro like this:

#[my_attribute_macro(attribute_tokens)]
fn my_function() {}

In this case, the attr token stream would contain attribute_tokens, and the item token stream would contain the my_function function.

Cool. We have the basic infrastructure set up, now we just have parse the input token streams.

The Rust compiler hasn’t even been so kind as to create the syntax tree for us yet. We just get a token stream, and we have to somehow parse it into something sensible (like a struct definition, impl block, etc.), manipulate it in some way, and then synthesize an output that the compiler can make sense of as valid code.

That’s a lot of work to do!

Enter: syn and quote.

Communicating with the compiler

syn and quote are a pair of crates that simplify token stream manipulation. syn provides utilities for parsing token streams into syntax trees, and quote for converting Rust-like code back into token streams.

The basic use-cases for each of these crates are extremely simple—they’re very well-designed crates!

Here is a bare-bones attribute macro using syn and quote, which does absolutely nothing (it returns its input):

use proc_macro::TokenStream;
use syn::{parse_macro_input, AttributeArgs, Item};
use quote::quote;

#[proc_macro_attribute]
pub fn my_attribute_macro(attr: TokenStream, item: TokenStream) -> TokenStream {
    let _attr = parse_macro_input!(attr as AttributeArgs);
    let item = parse_macro_input!(item as Item);

    quote!{
        #item
    }.into()
}

The parse_macro_input macro tries to parse a TokenStream into a syn data structure, and produces a compiler error on failure. The syn data structures and documentation are worth perusing on your own. They will give you a pretty good idea of what a syntax tree might look like.

The Item parsed in the example above is an enum, which you can match against:

match item { // The annotated item was parsed as...
    Item::Enum(e) => {}, // ...an enum
    Item::Fn(f) => {}, // ...a function
    Item::Impl(i) => {}, // ...an impl block
    Item::Struct(s) => {}, // ...a struct

    // ...and so on and so forth
    _ => todo!(),
}

The quote macro produces a proc_macro2::TokenStream (which can be easily transformed into a normal proc_macro::TokenStream via Into::into) from some input that is like Rust code. It also supports variable interpolation, via the #identifier syntax seen above.

In the world of macro authorship, syn and quote are pretty ubiquitous. Here’s an example of syn and quote used in the popular crate thiserror.

Parameterization and configuration

The last tool in our belt for building maintainable and usable macros is darling. The crate’s stated description is:

A proc-macro library for reading attributes into structs when implementing custom derives.

However, it is useful for both custom derives and attribute macros. syn and quote are useful for parsing and manipulating streams of normal Rust tokens, and darling is useful for parsing attribute and item input streams into custom structs, attaching custom logic to process, and reporting errors, making the combination of these three crates a powerful framework for procedural macro authorship.

Practical Example

Here is an example of a very simple derive macro using all three crates, complete with error-handling, an optional configuration parameter, and some of darling’s auto-forwarded fields (data, generics, ident, fields).

use darling::{FromDeriveInput, FromVariant};
use proc_macro2::TokenStream as TokenStream2;
use quote::quote;
use syn::{parse_macro_input, DeriveInput, Path};

#[derive(Debug, FromDeriveInput)]
// The struct will be deserialized from a `#[display]` attribute on any kind of enum
#[darling(attributes(display), supports(enum_any))]
struct EnumMeta {
    // Try to optionally deserialize an item path
    pub transform: Option<Path>,

    // Forwarded attributes
    pub ident: syn::Ident,
    pub generics: syn::Generics,
    pub data: darling::ast::Data<VariantVisitor, ()>,
}

#[derive(Debug, FromVariant)]
struct VariantVisitor {
    // The name of the enum variant
    pub ident: syn::Ident,
    pub fields: darling::ast::Fields<()>,
}

fn expand(meta: EnumMeta) -> Result<TokenStream2, darling::Error> {
    let EnumMeta {
        transform,
        data,
        generics,
        ident,
    } = meta;

    let variants = data.take_enum().unwrap();

    let match_arms = variants.iter().map(|variant| {
        let i = &variant.ident;
        let name = i.to_string();
        match variant.fields.style {
            darling::ast::Style::Tuple => {
                quote! { Self :: #i ( .. ) => #name , }
            }
            darling::ast::Style::Struct => {
                quote! { Self :: #i { .. } => #name , }
            }
            darling::ast::Style::Unit => {
                quote! { Self :: #i  => #name , }
            }
        }
    });

    // Properly includes generics in output
    let (imp, ty, wher) = generics.split_for_impl();

    // Rust code output
    Ok(quote! {
        impl #imp std::fmt::Display for #ident #ty #wher {
            fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
                write!(f, "{}", #transform (
                    match self { #(#match_arms)* }
                ))
            }
        }
    })
}

// Declares the name of the macro and the attributes it supports
#[proc_macro_derive(Display, attributes(display))]
pub fn derive_display(input: proc_macro::TokenStream) -> proc_macro::TokenStream {
    let input = parse_macro_input!(input as DeriveInput);

    FromDeriveInput::from_derive_input(&input)
        .and_then(expand)
        .map(Into::into)
        // Error handling
        .unwrap_or_else(|e| e.write_errors().into())
}

This code is also available on GitHub.

This derive macro creates an implementation of Display on the targeted enum. It optionally accepts a transform attribute field, which is a path to a function which transforms the name of the variant before it is written out.

Further Reading


I’m a software engineer for NEAR Protocol, head of education at the Blockchain Acceleration Foundation, and a graduate student at the Tokyo Institute of Technology.

Connect with me on Twitter.


  1. Correction from matklad:

    The macro evaluation process is … messy, a correct thing to say is that “in rust compiler, parsing, name resolution, and macro expansion are mutually recursive procedures which happen at the same time”. Luckily, I think for the purposes of this post we don’t need to explain when macro expansion happens, it is enough to say “tokens is what is used as input or output of the macro. Macros don’t have direct access to a parsed AST, but a macro can parse input tokens itself”.

     ↩︎