Regression in 1.26 with misplaced attributes

[pietroalbini]

All the regressed crates have a misplaced #[derive] in the codebase (such as a #[derive(Debug)] above a variable declaration). In Rust 1.25 the compiler accepts the code, but in Rust 1.26 it throws an error and then crashes.

• bitlab regressed from stable to beta (build log) cc @kkayal
• promise-rs regressed from stable to beta (build log) cc @stan-kondrat
• serde_bencode regressed from stable to beta (build log) cc @toby
• paulkramme/rustybrainfuck regressed from stable to beta (build log) cc @paulkramme

Even if we keep the error (which could be turned into a warn-by-default lint) the ICE should be fixed.

[pietroalbini]

This also happens with misplaced attributes on functions:

• chiisai regressed from stable to beta (build log) cc @mgattozzi
• foundation regressed from stable to beta (build log) cc @mgattozzi
• libimagentrylink regressed from stable to beta (build log) cc @matthiasbeyer
• seq_io regressed from stable to beta (build log) cc @markschl
• indeyets/rust-nquads regressed from stable to beta (build log) cc @indeyets

[Mark-Simulacrum]

cc @alexcrichton @dtolnay -- proc-macro related changes presumably

[mgattozzi]

I have a feeling I know where it's broken in one of my macros for chiisai based off the build log. I slapped an #[inline(always)] above a closure that gets passed to it in that macro. I'll need to confirm though. It's a non proc_macro just the regular macro_rules!

[alexcrichton]

My guess is that this isn't related to procedural macros but rather #49291, cc @tejom and @petrochenkov

[tejom]

The change that effected inline and closures was brought up in issue #49632

I don't mind making a change to make the original pull request a warning if that is preferred?

[tejom]

The ice with the derive check doesn't seem to be caused by #49291

fn main() {
        #[derive(Debug)]
        let _a = 5;

        #[repr(always)]
        let _x = 12;
}

This compiles with stable.
That pull request I made was in nightly on 2018-04-01, and that only has an error for the incorrect repr above _x
With nightly from 2018-04-10 The compiler crashed after showing an error about the incorrect derive.

[pietroalbini]

Did a bisect run, and #49124 is the source of the ICE. cc @petrochenkov @abonander

[abonander]

I saw this issue earlier and meant to comment that I was pretty sure it was my PR, so I could have saved you a bisect. However, I haven't been able to build Rust for at least a week so I'd be mostly relying on Travis to test any fixes.

[abonander]

Ironically I'm able to work on beta and not nightly.

However, compile-fail seems to eat the ICE because while I can reproduce it on master, I have to manually call the built rustc to get it because the compile-fail suite passes because it still produces the expected error.

[nikomatsakis]

triage: P-high

Has a pending fix.