ICE (Missing tokens for nt) on nightly with `stmt_expr_attributes` and `proc_macro_hygiene`

I tried this code:

use proc_macro::TokenStream;

#[proc_macro_attribute]
pub fn breaks_rust(_: TokenStream, input: TokenStream) -> TokenStream {
    input
}

#![feature(stmt_expr_attributes, proc_macro_hygiene)]

use the_proc_macro_crate::breaks_rust;

#[breaks_rust]
fn main() {
    let _break;
    let rust = ();
    #[breaks_rust]
    _break = rust; // line 10
}

I expected the code to compile sucessfully, but got an ICE instead:

thread 'rustc' panicked at 'Missing tokens for nt at tests/ice.rs:10:5: 10:11 (#0): "_break"', compiler/rustc_parse/src/lib.rs:295:9

(This is procedural-macro-related. I can't reproduce without a procedural macro.)

This used to work and regressed in #79472.

Backtrace

   0: rust_begin_unwind
             at /rustc/b122908617436af187252572ed5db96850551380/library/std/src/panicking.rs:493:5
   1: std::panicking::begin_panic_fmt
             at /rustc/b122908617436af187252572ed5db96850551380/library/std/src/panicking.rs:435:5
   2: rustc_parse::nt_to_tokenstream
   3: rustc_expand::base::Annotatable::into_tokens
   4: rustc_expand::expand::MacroExpander::fully_expand_fragment
   5: rustc_expand::expand::MacroExpander::expand_crate
   6: rustc_interface::passes::configure_and_expand_inner
   7: rustc_interface::passes::configure_and_expand::{{closure}}
   8: rustc_data_structures::box_region::PinnedGenerator<I,A,R>::new
   9: rustc_interface::passes::configure_and_expand
  10: rustc_interface::queries::Queries::expansion
  11: rustc_interface::queries::<impl rustc_interface::interface::Compiler>::enter
  12: rustc_span::with_source_map
  13: rustc_interface::interface::create_compiler_and_run

cc #15701 #54727

[camelid]

If this code only works on nightly, then I don't think it's a regression-from-stable-to-nightly. However, we don't have a regression-from-nightly-to-nightly label, so ...

[Aaron1011]

This will be fixed by #81286, which handles attributes on expressions in a more systematic way.

As a side note, #[breaks_rust] _break = rust; is currently parsed as (#[breaks_rust] _break) = rust;

[camelid]

This will be fixed by #81286, which handles attributes on expressions in a more systematic way.

Do you want to add Fixes #81555 to that PR then?