[1.30 beta] Regression in cfg attribute parsing

[emilyalbini]

Parsing of the following snippet is failing on Rust 1.30 beta:

#![cfg_attr(feature = "nightly", feature(plugin_registrar, rustc_private,quote,),)]

• command-macros 0.1.11 regressed from stable to beta (build log) cc @krdln

cc @petrochenkov

[petrochenkov]

Most probably caused by #53893.
I wonder why the previous crater run didn't caught this.

[eddyb]

@petrochenkov According to @pietroalbini, before and after had the same error (for #53893).
Is it possible a (different?) regression landed on master before that crater run?

[emilyalbini]

The regression is caused by #53293, which didn't go through a Crater run.

[petrochenkov]

(plugin_registrar, rustc_private,quote,), after the attribute path feature was parsed as an arbitrary token stream containing two token trees - parenthesized group (...) and comma ,.

Arbitrary token streams were never supposed to work on stable channel, but this wasn't properly enforced.
In particular, the unstable syntax was usable in code that's cfg-ed out.
The point of #53293 was exactly to fix this hole, so I think this issue is WONTFIX, as stated by rust-lang/rfcs#2405, especially given that this is the only regression of this kind and the code in command-macros was fixed long ago.

Note, that #![cfg_attr(feature = "nightly", feature(plugin_registrar, rustc_private,quote,),)] doesn't compile on stable if feature = "nightly" is actually enabled.

[petrochenkov]

We can accept trailing comma in cfg_attr specifically though.
It will need to be accepted anyway when rust-lang/rfcs#2539 is implemented.

[petrochenkov]

Fixed in #54581