Token Goblin — munches your tokens, forges charms

token-goblin is a proc-macro library for defining inline proc-macros directly inside your crate, without a separate proc-macro target.

It is inspired by crates like crabtime and inline-proc, but aims to provide a more polished, flexible, and ergonomic API.

Getting started

Add token-goblin to your crate:

[dependencies]
token-goblin = "0.3.0"

Then try:

#[token_goblin::munch]
fn foo(input: TokenStream) -> TokenStream {
    input
}

This generates a new macro, or charm, named foo!:

foo!(bar baz); // will expand to `bar baz`

In other words, #[munch] turns the function into a new macro.

Note: because token-goblin::munch macros generate macros, the term charm is used for generated macros in the docs for clarity (and a little bit of lore).

Use Cases

A well-fed goblin is a productive goblin. Here is what it does once it has chewed through your tokens.

Simple string-based API like in crabtime

Some users don't want to mess with the proc-macro API; they find it foreign and confusing. crabtime showed another way to write macros: a simple string-based API that allows String and Vec<String> to be used directly as macro input.

Example adapted from the crabtime docs:

#[token_goblin::munch]
fn generate_enums(components: CommaSeparated<Token>) {
    let components: Vec<String> = components.into();
    for dim in 1..=components.len() {
        let cons = components[0..dim].join(",");
        output_str! {
            "#[derive(Debug)]
            enum Enum{dim} {{
                {cons}
            }}"
        }
    }
}

generate_enums!["X", "Y", "Z", "W", "V", "U", "T", "S", "R", "Q"];

which will expand to:

enum Enum1 { X }
// ... up to
enum Enum10 { X, Y, Z, W, V, U, T, S, R, Q }

Note: while this is inspired by crabtime, token-goblin adopts the approach without hardcoding String or Vec<String> handling. Instead, input is expected to implement the syn::parse::Parse trait. So CommaSeparated<Token> is just two wrappers in the token-goblin-runtime crate that provide the required syn::parse::Parse implementation.

Inline proc-macro

The string-based API is simple, but it loses span information and reduces IDE/diagnostics quality.

If you don't want to lose span information, but are still annoyed that implementing a simple proc-macro requires a separate crate, token-goblin provides a classic proc-macro2 API as well:

#[token_goblin::munch]
fn foo(input: TokenStream) -> TokenStream {
    // ..
}

And even better, it supports syn-based types as input parameters:

#[token_goblin::munch]
fn stringify(input: syn::Ident) -> TokenStream {
    let v = input.to_string();
    quote! {
        #v
    }
}

Or, you can define multiple charms in one module and extend the input parameters.

#[token_goblin::munch]
mod macros {
    struct StructParam {
        // ..
    }
    impl syn::parse::Parse for MyStruct {
        //..
    }
    /// Note: ALL `pub fn`/`pub(crate) fn` are considered as entrypoints.
    /// Note2: No need to write `#[token_goblin::munch]` before each `pub fn`, it's already implied.
    pub fn generate_enums(components: CommaSeparated<Token>) -> TokenStream {
        // ..
    }
    pub fn generate_structs(param: StructParam) -> TokenStream {
        // ..
    }
}

macros::generate_enums!["X", "Y", "Z", "W", "V", "U", "T", "S", "R", "Q"];
macros::generate_structs!{Foo};

Probes and Evals

Sometimes the goblin just sits by the fire and counts things in its head, so you don't have to at runtime.

Another common use case for macros is to precompute some data. crabtime provides an eval macro for this purpose.

But with token-goblin, you can implement it yourself:

macro_rules! eval {
    ($($expr:tt)*) => {
        {
            #[token_goblin::munch(lazy)]
            fn eval_inner(_: TokenStream) -> TokenStream {
                use std::str::FromStr;
                let x = $($expr)*;
                quote!{ #x }
            }
            eval_inner!($($expr)*)
        }
    };
}

fn main() {
    // Example from crabtime docs:
    let x = eval!((std::f32::consts::PI.sqrt() * 10.0).round() as usize);
    println!("x: {x}");
}
// prints:
// x: 18

Note: any expression is embedded into the charm as code and cannot use external variables or call functions from your crate.

But you are not limited to simple expressions. In fact, you can do any compile-time execution, like evaluating bytecode or even downloading something from the internet (using external state in a macro is not recommended though).

e.g. from example_readme/examples/brainfuck.rs

#[token_goblin::munch]
mod brainfuck {
    pub fn execute(input: ProgramInput) -> TokenStream {
        // ..
    }

    pub fn request_and_execute(input: ProgramInput) -> TokenStream {
        // Handle program field as URL.
        let url = String::from_utf8(input.program.value()).unwrap();
        let program = reqwest::blocking::get(url).unwrap().text().unwrap();
        execute(ProgramInput {
            program: syn::LitByteStr::new(&program.as_bytes(), Span::call_site()),
            input: input.input,
        })
    }
}

    let result = brainfuck::request_and_execute!(b"https://gist.githubusercontent.com/vldm/f796f0d6235a608c0bed5957d146f8c0/raw/a068d4a8b2764fbc02b909322f31321b1b7eb7fc/reverse.bf", b"\n!dlroW olleH");
    println!("result: {result}");
    // downloads: ">,[>,]<[.<]" program that reverses input
    // prints:
    // result: Hello World!

Executing a brainfuck program is purely functional and therefore fits proc-macro purposes well, but using system APIs and requesting external data is clearly misuse. The whole crate is an experiment around proc-macro, so I think it's fun to showcase it as well.

Note: while token-goblin itself doesn't cache the output of charms, Rust itself might cache them, especially when -Zcache-proc-macros is enabled.

Note: there is a plan to implement wasm as a feature that will enforce sandboxing of charms.

Reflection?

In computer science, reflective programming or reflection is the ability of a process to examine, introspect, and modify its own structure and behavior.

Reflection is a powerful feature that allows code to be generated dynamically without knowing the exact types by observing their structure. Zig has the comptime keyword, which allows code to execute at compile time and observe the structure of the code. In Rust, we only have derives. They can replace some kinds of reflection, e.g. by providing a way to generate some traits based on struct fields. The missing piece is that they are not extendable. For example, the person who writes struct Foo defines the list of derived traits, and this list is not extendable.

So if you want to extend some type with your custom trait, you need to duplicate the Foo definition somewhere in some form. Reflection could solve this problem by providing the shape of the type and then generating the trait based on it.

token-goblin has a similar feature called Snif, which allows you to collect information about a type and pass it to another macro.

#[derive(token_goblin::Snif)]
struct Foo {
    x: i32,
}
#[token_goblin::munch(lazy)]
fn generate_getters(input: SnifedEntries) -> TokenStream {
    let syn::Item::Struct(item) = &input.entries[0].item else {
        return syn::Error::new(input.span(), "Expected struct").to_compile_error();
    };
    let name = &item.ident;
    let (fields, types): (Vec<syn::Ident>, Vec<syn::Type>) = item
        .fields
        .iter()
        .cloned()
        .map(|field| (field.ident.unwrap(), field.ty))
        .unzip();
    quote! {
        impl #name {
            #(
                pub fn #fields(&self) -> &#types {
                    &self.#fields
                }
            )*
        }
    }
}

token_goblin::snif!(Foo in generate_getters!(extra args));

generate_getters!() will receive input in this format: [Foo => { struct Foo { x : i32, } }] [ extra args]

It can then generate code based on the information about types (in this example, it generates getters for Foo).

This example can be found in example_readme/examples/generate_getters.rs.

A more complete example (MultiArrayList-like from Zig) that converts an array of structs into a struct of arrays can be found in token-goblin/examples/struct_of_arrays.rs.

Multiple Small Derives

Sometimes in big projects, you need to define multiple small derives, e.g. parsing/emitting/printing functionality is distinct and should be separated. Placing all of them in one "macro" crate might not be the best choice. Instead, token-goblin allows you to split the logic into multiple "macro" crates and use them as dependencies.

#[derive(token_goblin::Snif)]
struct Foo {
    x: i32,
}

#[token_goblin::munch]
fn generate_parser(input: SnifedEntries) -> TokenStream {
   // ..
}

#[token_goblin::munch]
fn generate_emitter(input: SnifedEntries) -> TokenStream {
}

token_goblin::snif!(Foo in generate_parser!());
token_goblin::snif!(Foo in generate_emitter!());

This approach is partially shown in token-goblin/examples/struct_of_arrays.rs. I also use it in a real project, where I want to extend my type with additional metadata while keeping derive logic separated. It looks like this:

#[derive(token_goblin::Snif)]
enum Expr {
    #[snif(mnemonic = "lit")]
    #[snif(arity = 0 -> 1)]
    Lit(syn::Lit),
    #[snif(mnemonic = "add")]
    #[snif(arity = 2 -> 1)]
    Add(Box<Expr>, Box<Expr>),
}
mod printer {
  trait Printer {}
  snif!(Expr in generate_printer!());
}
// ..

Do I need to rewrite declarative macros to the proc-macro API?

While the proc-macro API is more Rust-like and powerful, one might want to rewrite all declarative macros to the proc-macro API. But working with TokenStream introduces some boilerplate, and some macros should be kept declarative.

TTs muncher is a technique for writing recursive declarative macros to parse complex input.

If we take the example from the link above (slightly modified):

macro_rules! trace {
    () => {};

    ($name:ident; $($tail:tt)*) => {{
        println!("{} = {:?}", stringify!($name), $name);
        trace!($($tail)*);
    }};

    ($name:ident = $value:expr; $($tail:tt)*) => {{
        let $name = $value;
        println!("{} = {:?}", stringify!($name), $name);
        trace!($($tail)*);
    }};
}

It expects input in this format:

let a = 10;
trace! {
    x = 2 + 3;
    y = x * 10;
    x;
    y;
}

expands to something like:

{
    let x = 2 + 3;
    println!("x = {:?}", x);
    {
        let y = x * 10;
        println!("y = {:?}", y);
        {
            println!("x = {:?}", x);
            {
                println!("y = {:?}", y);
            }
        }
    }
}

and produces this output in the console:

x = 5
y = 50
x = 5
y = 50

Rewriting it as a proc-macro TokenStream API would increase the amount of code and add a lot of boilerplate:

#[token_goblin::munch]
fn trace_cycle(input: TokenStream) {
    let mut iter = input.into_iter().peekable();

    while iter.peek().is_some() {
        let Some(TokenTree::Group(g)) = iter.next() else {
            panic!("Expected group");
        };
        let Some(TokenTree::Ident(ident)) = iter.next() else {
            panic!("Expected ident");
        };
        let mut expr = (&mut iter)
            .take_while(|token| !matches!(token, TokenTree::Punct(p) if p.as_char() == ';'))
            .collect::<Vec<_>>();

        let let_stmt = if expr.is_empty() {
            quote! {}
        } else {
            quote! {
                let #ident  #(#expr)*;
            }
        };
        let ident_str = ident.to_string();
        output! {
            #let_stmt;
            writeln!(#g, "{} = {:?}", #ident_str, #ident).ok();
        }
    }
    if iter.peek().is_some() {
        panic!("Expected end of input");
    }
}

Using syn with syn-derive might help with the main logic:

pub fn trace_syn(input: TraceInput) -> TokenStream {
    let mut out = TokenStream::new();

    for TraceStmt {
        writer,
        ident,
        value,
    } in input.0
    {
        let ident_str = ident.to_string();

        let let_stmt = match value {
            TraceValue::Some { expr, .. } => quote! { let #ident = #expr; },
            TraceValue::None => quote! {},
        };

        out.extend(quote! {
            #let_stmt
            writeln!(#writer, "{} = {:?}", #ident_str, #ident).ok();
        });
    }

    out
}

It still requires defining TraceInput and TraceStmt structs, and syn::parse::Parse implementation for them. See example_readme/examples/ttmunch-replace.rs for more details.

With token-goblin, you don't need to choose, since it allows you to combine both approaches.

For example, you can write a declarative macro as a facade that checks patterns and computes results in the proc-macro API.

#[token_goblin::munch]
pub fn stringify_any(input: TokenStream) -> TokenStream {
    let string = input.to_string();
    quote! {
        #string
    }
}

macro_rules! stringify_ident {
    ($ident:ident) => {
        stringify_any!($ident)
    };
}

fn main() {
    // this will fail at compile time, due to wrong input pattern
    // let result = stringify_ident!("non ident");
    // let result = stringify_ident!(foo asd);
    let result = stringify_ident!(foo);
    println!("result: {result}");
}

Uncommenting non-ident expansions will fail at compile time:

There is still the old but good proc-macro-rules crate, which allows you to use declarative macro patterns directly in the proc-macro API.

Questions

Why is it named Token Goblin?

While tinkering with the name, ChatGPT 5.5 suggested this variant among others:

I found it ridiculous, especially after I saw OpenAI's post on how they are fighting "goblin" overuse by ChatGPT.

Also, the idea of "some magical entity that eats tokens" seems like a good metaphor for macros.

Why entrypoint macros named munch and spit?

1. Because munch and spit fit well in "goblin" lore.
2. I think that #[munch] fn would be a good replacement for the existing TTs muncher, a technique for writing recursive declarative macros to parse complex input.

Why not use crabtime or inline-proc?

They both look unmaintained.

inline-proc uses syn 1.0 and has had no updates for ~5-6 years. It doesn't compile anymore on modern Rust versions.

I have tried to contribute to crabtime at https://github.com/wdanilo/crabtime/issues?q=author%3Avldm, but it looks like the author is not interested in maintaining it anymore. There are still issues related to the build cache.

token-goblin combines all the features from both crabtime and inline-proc, like:

• using dylib to load proc-macro definition
• support for workspace dependencies
• support for attributes and derive macro helpers
• mod and fn entrypoints

It also adds some extra:

• Emit ide helper for Rust-Analyzer completion ide-helper
• Allow span information to be preserved in output span_recovery
• Convert any panic to compile error panic
• Extensible interface for input and output ux
• A "reflection"-like macro to store tokens of some items and use them as input to another macro snif.

And more is planned:

• Mapping panics/compile errors to compile_error! should show any error at the right source location.
• Support for wasm as a feature that will enforce sandboxing of charms.

Testing

Most tests are implemented as regular integration tests or doctests directly in the macro library. Fixtures represent tests that need to be run with different environments (currently only toolchain or Cargo config).

Fixtures can be run with:

cargo test -p token-goblin --test fixtures

Usage recommendations

Some hints and recommendations for using token-goblin in your projects.

IDE support

As with offline builds, it is recommended to add token-goblin-runtime to [dev-dependencies] in your Cargo.toml. This will help rust-analyzer find the needed crate and provide important semantic information for your macros.

Laziness

token-goblin::munch provides a lazy attribute that allows enforcing laziness of charm compilation.

By default, all charms generated by token-goblin::munch are eager. This means that a charm is compiled during expansion of the #[munch] attribute, and users of the charm only use the compiled dylib.

This setup is faster, since charm is compiled only once, and every user (expansion of charm itself) skips the compilation step.

But during development, flycheck could call cargo check on broken code and spam errors. In VS Code + Lens, this can slow down IDE performance. Therefore, you can set lazy to true in #[token_goblin::munch] attributes.

#[token_goblin::munch(lazy)] // or #[token_goblin::munch(lazy = true)]
fn foo(input: TokenStream) -> TokenStream {
    // ..
}

With this setup, #[munch] will not compile the charm; instead, compilation will be triggered during foo expansion. Note: for the same code, compilation is only triggered once, since token-goblin caches the compiled dylib.

Debugging

You can also set the TOKEN_GOBLIN_PRINT_LEVEL environment variable to 1-4 to enable debug prints.

1 - print basic info
2 - print timings
3 - print input and output of internal macros
4 - print environment variables

You can also use println / eprintln / dbg and other macros to debug your charms.

Share cache or not?

By default, all charms generated by token-goblin::munch share the same build-cache directory. Sharing a cache forces Cargo to lock the directory, so one "slow" charm can slow down the whole compilation process.

To avoid this, you can set split_cache to true in #[token_goblin::munch] attributes.

#[token_goblin::munch(split_cache)] // or #[token_goblin::munch(split_cache = true)]
fn foo(input: TokenStream) -> TokenStream {
    // ..
}

This will force the charm to use a separate build-cache directory, so it will not be affected by other charms.

I recommend using split_cache only for "big" charms that require a lot of dependencies or take a lot of time to build. This is because a charm with a separate cache can be compiled in parallel with other charms.

Caveats

Even a helpful goblin has its quirks. Mind these before you let it loose.

• only proc-macro2::fallback is used (no proc-macro API is available) in generated crates, which introduces some limitations.
• mixed_site is not supported by proc_macro2::fallback.
• we use dev-dependencies for charm dependencies, which cannot be optional by design of the Cargo resolver, so one small macro may increase compile time by rebuilding all dev-dependencies.
• name in #[munch] fn name should not be proc-macro generated and is expected to have a local source file.
• on macOS, loading dylibs (newly generated charms) may take more time than compilation itself (~300ms). This is a known issue related to XProtect. See the link above for a workaround.
• rust-analyzer will not analyze "optional" dependencies and will emit "unresolved external crate" errors on charms. To disable IDE support for charms, use the no_ide_helper attribute: #[token_goblin::munch(dependencies = [..],no_ide_helper)].

Offline build

Note: token-goblin-runtime is a hardcoded dependency of generated crates and might not be downloaded by cargo fetch or cargo vendor. To build offline, add token-goblin-runtime to [dev-dependencies] in your Cargo.toml.