E0726 makes invalid mid-keyword suggestion with proc-macro derived on struct

[Zannick]

I tried this code:

extern crate sort_by_derive;

use sort_by_derive::SortBy;
use std::borrow::Cow;

pub trait Ctx: Clone {}

#[derive(Clone, SortBy)]
pub struct ContextWrapper<'a, T>
where T: Ctx,
{
    ctx: Cow<'a, T>,
    #[sort_by]
    elapsed: i32,
}

impl<'a, T: Ctx> ContextWrapper<'a, T> {
    pub fn new(ctx: T) -> ContextWrapper<'a, T> {
        ContextWrapper {
            ctx: Cow::Owned(ctx),
            elapsed: 0,
        }
    }
}

and in Cargo.toml:

[dependencies]
sort_by_derive = "0.1.10"

I expected either:

1. the code compiles with cargo check
2. any errors point out what needs to be changed

Instead, this error appeared:

error[E0726]: implicit elided lifetime not allowed here
  --> src/lib.rs:12:1
   |
12 | pub struct ContextWrapper<'a, T>
   | ^^^^^^^^^^^^^^^^^^^^^^^^^ expected lifetime parameter
   |
   = note: assuming a `'static` lifetime...
help: indicate the anonymous lifetime
   |
12 | p'_, ub struct ContextWrapper<'a, T>
   |  +++

For more information about this error, try `rustc --explain E0726`.

I'm guessing the error being pointed to is in the procedural macro generated code, but that's not immediately clear as it points to my struct definition. This code compiles when I remove SortBy and its attribute.

Meta

rustc --version --verbose:

rustc 1.67.1 (d5a82bbd2 2023-02-07) (Fedora 1.67.1-1.fc37)
binary: rustc
commit-hash: d5a82bbd26e1ad8b7401f6a718a9c57c96905483
commit-date: 2023-02-07
host: x86_64-unknown-linux-gnu
release: 1.67.1
LLVM version: 15.0.7

(Issue template requested a backtrace but RUST_BACKTRACE=1 cargo build did not produce additional output.)

[estebank]

sort_by_derive Is effectively using the wrong span on their expanded code. If they were doing nothing then #109082 would eliminate the issue, but the best bet we have here is to modify sort_by_derive itself, instead of adding more heuristics to rustc to detect that the macros of the Token Stream coming out of a proc-macro might be misleading (at least in the short term).

[Zannick]

If the lifetimes suggestion is meant to be in the expanded code, then the error message pointing to the expanded code/the derive macro would be sufficient for me, since I would be able to pinpoint the issue there is in the other crate and bring it to them, and probably figure out that the spans are wrong. (The original issue was that I couldn't tell where to look until I started removing code.) If rustc run on sort_by_derive itself can warn that the spans are wrong, that's a bonus.

[valsteen]

hi, crate author here 👋

Indeed the problem is on my side and lifetimes were missing from the expanded code, so I fixed that and released the fix. I should probably check if I can highlight the derive macro itself when it generated something invalid, but it's going to be of limited value to the users, except indeed signaling that it's not their fault.