I thought Rust doesn't have reflection...?

This crate explores what it could look like to tackle the 80% use case of custom derive macros through a programming model that resembles compile-time reflection.

Motivation

My existing syn and quote libraries approach the problem space of procedural macros in a super general way and are a good fit for maybe 95% of use cases. However, the generality comes with the cost of operating at a relatively low level of abstraction. The macro author is responsible for the placement of every single angle bracket, lifetime, type parameter, trait bound, and phantom data. There is a large amount of domain knowledge involved and very few people can reliably produce robust macros with this approach.

The design explored here focuses on what it would take to make all the edge cases disappear -- such that if your macro works for the most basic case, then it also works in every tricky case under the sun.

Programming model

The idea is that we expose what looks like a boring straightforward runtime reflection API such as you might recognize if you have used reflection in Java or reflection in Go.

The macro author expresses the logic of their macro in terms of this API, using types like reflect::Value to retrieve function arguments and access fields of data structures and invoke functions and so forth. Importantly, there is no such thing as a generic type or phantom data in this model. Everything is just a reflect::Value with a type that is conceptually its monomorphized type at runtime.

Meanwhile the library is tracking the control flow and function invocations to build up a fully general and robust procedural implementation of the author's macro. The resulting code will have all the angle brackets and lifetimes and bounds and phantom types in the right places without the macro author thinking about any of that.

The reflection API is just a means for defining a procedural macro. The library boils it all away and emits clean Rust source code free of any actual runtime reflection. Note that this is not a statement about compiler optimizations -- we are not relying on the Rust compiler to do heroic optimizations on shitty generated code. Literally the source code authored through the reflection API will be what a seasoned macro author would have produced simply using syn and quote.

From the perspective of the person that ends up calling the macro, everything about how it is called is the same as if the macro were written the old fashioned way without reflection, and their code compiles exactly as fast and performs exactly as fast. The advantage is to the macro author for whom developing and maintaining a robust macro is greatly simplified.

Demo

This project contains a proof of concept of a compile-time reflection API for defining custom derives.

The tests/debug/ directory demonstrates a working compilable implementation of #[derive(Debug)] for structs with named fields. The corresponding test case shows what code we emit when deriving Debug for a struct Point with two fields; it is equivalent to the code that a handwritten derive(Debug) macro without reflection would emit for the same data structure.

The macro implementation begins with a DSL declaration of the types and functions that will be required at runtime:

reflect::library! {
    extern crate std {
        mod fmt {
            type Formatter;
            type Result;
            type DebugStruct;

            trait Debug {
                fn fmt(&self, &mut Formatter) -> Result;
            }

            impl Formatter {
                fn debug_struct(&mut self, &str) -> DebugStruct;
            }

            impl DebugStruct {
                fn field(&mut self, &str, &Debug) -> &mut DebugStruct;
                fn finish(&mut self) -> Result;
            }
        }
    }
}

There may be additional extern crate blocks here if we need to use types from outside the standard library. For example Serde's #[derive(Serialize)] macro would want to list the serde crate, the Serialize and Serializer types, and whichever of their methods will possibly be invoked at runtime.

Throughout the rest of the macro implementation, all type information is statically inferred based on the signatures given in this library declaration.

Next, the macro entry point is an ordinary proc_macro_derive function just as it would be for a derive macro defined any other way.

Once again the reflection API is just a means for defining a procedural macro. Despite what it may look like below, everything written here executes at compile time. The reflect library spits out generated code in an output TokenStream that is compiled into the macro user's crate. This token stream contains no vestiges of runtime reflection.

use proc_macro::TokenStream;

// Macro that is called when someone writes derive(MyDebug) on a data structure.
// It returns a fragment of Rust source code (TokenStream) containing an
// implementation of Debug for the input data structure. The macro uses
// compile-time reflection internally, but the generated Debug impl is exactly
// as if this macro were handwritten without reflection.
#[proc_macro_derive(MyDebug)]
pub fn derive(input: TokenStream) -> TokenStream {
    // Feed the tokens describing the data structure into the reflection library
    // for parsing and analysis. We provide a callback that describes what trait
    // impl(s) the reflection library will need to generate code for.
    reflect::derive(input, |ex| {
        // Instruct the library to generate an impl of Debug for the derive
        // macro's target type / Self type.
        ex.make_trait_impl(RUNTIME::std::fmt::Debug, ex.target_type(), |block| {
            // Instruct the library to compile debug_fmt (a function shown
            // below) into the source code for the impl's Debug::fmt method.
            block.make_function(RUNTIME::std::fmt::Debug::fmt, debug_fmt);
        });
    })
}

The following looks like a function that does runtime reflection. It receives function arguments which have the type reflect::Value and can pass them around, pull out their fields, inspect attributes, invoke methods, and so forth.

use reflect::*;

// This function will get compiled into Debug::fmt, which has this signature:
//
//     fn fmt(&self, formatter: &mut fmt::Formatter) -> fmt::Result
//
fn debug_fmt(f: MakeFunction) -> Value {
    let receiver: reflect::Value = f.arg(0);  // this is `self`
    let formatter: reflect::Value = f.arg(1);

    // The input value may be any of unit struct, tuple struct, ordinary braced
    // struct, or enum.
    match receiver.data() {
        Data::Struct(receiver) => match receiver {
            Struct::Unit(receiver) => unimplemented!(),
            Struct::Tuple(receiver) => unimplemented!(),
            Struct::Struct(receiver) => {
                /* implemented below */
            }
        },
        // For an enum, the active variant of the enum may be any of unit
        // variant, tuple variant, or struct variant.
        Data::Enum(receiver) => receiver.match_variant(|variant| match variant {
            Variant::Unit(variant) => unimplemented!(),
            Variant::Tuple(variant) => unimplemented!(),
            Variant::Struct(variant) => unimplemented!(),
        }),
    }
}

In the case of a struct with named fields we use reflection to loop over fields of the struct and invoke methods of the standard library Formatter API to append each field value into the debug output.

Refer to the DebugStruct example code in the standard library API documentation for what this is supposed to do at runtime.

Paths beginning with RUNTIME:: refer to library signatures declared by the library! { ... } snippet above.

let builder = RUNTIME::std::fmt::Formatter::debug_struct
    .INVOKE(formatter, type_name)
    .reference_mut();

for field in receiver.fields() {
    RUNTIME::std::fmt::DebugStruct::field.INVOKE(
        builder,
        field.get_name(),
        field.get_value(),
    );
}

RUNTIME::std::fmt::DebugStruct::finish.INVOKE(builder)

The reflection library is able to track how reflect::Value objects flow from one INVOKE to another, and contains a compiler that can compile this data flow into strongly typed Rust source code in a robust way. In the case of the Debug derive macro from this demo, when invoked on a braced struct with two fields,

#[derive(MyDebug)]
struct Point {
    x: i32,
    y: i32,
}

the reflection library would emit a trait impl that looks like this:

// expands to:
impl ::std::fmt::Debug for Point {
    fn fmt(&self, _arg1: &mut ::std::fmt::Formatter) -> ::std::fmt::Result {
        match *self {
            Point { x: ref _v0, y: ref _v1 } => {
                let mut _v2 = ::std::fmt::Formatter::debug_struct(_arg1, "Point");
                let _ = ::std::fmt::DebugStruct::field(&mut _v2, "x", _v0);
                let _ = ::std::fmt::DebugStruct::field(&mut _v2, "y", _v1);
                let _v3 = ::std::fmt::DebugStruct::finish(&mut _v2);
                _v3
            }
        }
    }
}

This generated code is what ends up running at runtime. Notice that there is no reflection. In fact this is pretty much identical to what the standard library's built-in derive(Debug) macro produces for the same data structure.

Robustness and how things go wrong

I mentioned above about how implementing robust macros simply using syn and quote is quite challenging.

The example I like to use is taking a single struct field and temporarily wrapping it in a new struct. This is a real life use case drawn from how serde_derive handles serialize_with attributes. Conceptually:

let input: DeriveInput = syn::parse(...).unwrap();

// Pull out one of the field types.
let type_of_field_x: syn::Type = /* ... */;

quote! {
    // Very not robust.
    struct Wrapper<'a> {
        x: &'a #type_of_field_x,
    }

    Wrapper { x: &self.x }
}

Making the quote! part of this simply generate compilable code for all possible values of type_of_field_x is extremely involved. The macro author needs to consider and handle all of the following in order to make this work reliably:

• Lifetime parameters used by type_of_field_x,
• Type parameters used by type_of_field_x,
• Associated types used by type_of_field_x,
• Where-clauses on input that constrain any of the above,
• Similarly, trait bounds on type parameters of input,
• Where-clauses or bounds affecting any other fields of input,
• Type parameter defaults on input that need to be stripped.

In contrast, the reflect library will be able to get it right every single time with much less thought from the macro author. Possibly as trivial as:

let wrapper: reflect::Type = reflect::new_struct_type();

wrapper.instantiate(vec![input.get_field("x").reference()])

Remaining work

In its current state the proof of concept generates just barely working code for our simple Debug derive. The reflect library needs more work to produce robust code in the presence of lifetimes and generic parameters, and for library signatures involving more complicated types.

Crucially all remaining work should happen without touching the code of our Debug derive. The promise of reflect is that if the macro works for the most basic cases (which the code above already does) then it also works in all the edge cases. From here it is reflect's responsibility to compile the dead simple reflection-like reflect::Value object manipulations into a fully general and robust procedural macro.

License