proc-macro-error

This crate aims to make error reporting in proc-macros simple and easy to use. Migrate from panic!-based errors for as little effort as possible!

Also, there's ability to append a dummy token stream to your errors.

[dependencies]
proc-macro-error = "0.4"

Supports rustc 1.31 and up

Documentation and guide

What emitted errors look like

error: multiple error part: multi2

  = note: help message test
  = help: Option help test
  = note: I see what you did here...

 --> $DIR/multi-error.rs:4:18
  |
4 | make_fn!(multi1, multi2, _, multi3);
  |                  ^^^^^^

Examples

Panic-like usage

use proc_macro_error::*;
use proc_macro::TokenStream;
use syn::{DeriveInput, parse_macro_input};
use quote::quote;

// This is your main entry point
#[proc_macro]
// this attribute *MUST* be placed on top of the #[proc_macro] function
#[proc_macro_error]
pub fn make_answer(input: TokenStream) -> TokenStream {
    let input = parse_macro_input!(input as DeriveInput);

    if let Err(err) = some_logic(&input) {
        // we've got a span to blame, let's use it
        // This immediately aborts the proc-macro and shows the error
        abort!(err.span, "You made an error, go fix it: {}", err.msg);
    }

    // `Result` has some handy shortcuts if your error type implements
    // `Into<MacroError>`. `Option` has one unconditionally.
    more_logic(&input).expect_or_abort("What a careless user, behave!");

    if !more_logic_for_logic_god(&input) {
        // We don't have an exact location this time,
        // so just highlight the proc-macro invocation itself
        abort_call_site!(
            "Bad, bad user! Now go stand in the corner and think about what you did!");
    }

    // Now all the processing is done, return `proc_macro::TokenStream`
    quote!(/* stuff */).into()
}

`proc_macro::Diagnostic`-like usage

use proc_macro_error::*;
use proc_macro::TokenStream;
use syn::{spanned::Spanned, DeriveInput, ItemStruct, Fields, Attribute , parse_macro_input};
use quote::quote;

fn process_attrs(attrs: &[Attribute]) -> Vec<Attribute> {
    attrs
        .iter()
        .filter_map(|attr| match process_attr(attr) {
            Ok(res) => Some(res),
            Err(msg) => {
                emit_error!(attr.span(), "Invalid attribute: {}", msg);
                None
            }
        })
        .collect()
}

fn process_fields(_attrs: &Fields) -> Vec<TokenStream> {
    // processing fields in pretty much the same way as attributes
    unimplemented!()
}

#[proc_macro]
#[proc_macro_error]
pub fn make_answer(input: TokenStream) -> TokenStream {
    let input = parse_macro_input!(input as ItemStruct);
    let attrs = process_attrs(&input.attrs);

    // abort right now if some errors were encountered
    // at the attributes processing stage
    abort_if_dirty();

    let fields = process_fields(&input.fields);

    // no need to think about emitted errors
    // #[proc_macro_error] will handle them for you
    //
    // just return a TokenStream as you normally would
    quote!(/* stuff */).into()
}

Limitations

Warnings are emitted only on nightly, they're ignored on stable.
“help” suggestions cannot have their own span info on stable, (they inherit parent span).
If a panic occurs somewhere in your macro no errors will be displayed. This is not a technical limitation but intentional design, panic is not for error reporting.

MSRV policy

proc_macro_error will always be compatible with proc-macro holy trinity: proc_macro2, syn, quote crates. In other words, if the trinity is available to you than proc_macro_error is available too.

Motivation

Error handling in proc-macros sucks. There's not much of a choice today: you either “bubble up” the error up to the top-level of your macro and convert it to a compile_error! invocation or just use a good old panic. Both these ways suck:

Former sucks because it's quite redundant to unroll a proper error handling just for critical errors that will crash the macro anyway so people mostly choose not to bother with it at all and use panic. Almost nobody does it, simple .expect is too tempting.
Also, if you do decide to implement this Result-based architecture in your macro you're going to have to rewrite it entirely once proc_macro::Diagnostic is finally stable. Not cool.
Later sucks because there's no way to carry out span info via panic!. rustc will highlight the whole invocation itself but not some specific token inside it.
Furthermore, panics aren't for error-reporting at all; panics are for bug-detecting (like unwrapping on None or out-of-range indexing) or for early development stages when you need a prototype ASAP and error handling can wait. Mixing these usages only messes things up.
There is proc_macro::Diagnostic which is awesome but it has been experimental for more than a year and is unlikely to be stabilized any time soon.
This crate's API is intentionally designed to be compatible with proc_macro::Diagnostic and delegates to it whenever possible. Once Diagnostics is stable this crate will always delegate to it, no code changes will be required on user side.

That said, we need a solution, but this solution must meet these conditions:

It must be better than panic!. The main point: it must offer a way to carry span information over to user.
It must take as little effort as possible to migrate from panic!. Ideally, a new macro with the same semantics plus ability to carry out span info. We should also keep in mind the existence of proc_macro::Diagnostic .
It must be usable on stable.

This crate aims to provide such a mechanism. All you have to do is annotate your top-level #[proc_macro] function with #[proc_macro_errors] attribute and change panics to abort!/abort_call_site! where appropriate, see the Guide.

Disclaimer

Please note that this crate is not intended to be used in any way other than proc-macro error reporting, use Result and ? for anything else.