This example triggers an assertion failure:
```
fn f() -> u32 {
#[cfg_eval] #[cfg(not(FALSE))] 0
}
```
The sequence of events:
- `configure_annotatable` calls `parse_expr_force_collect`, which calls
`collect_tokens`.
- Within that, we end up in `parse_expr_dot_or_call`, which again calls
`collect_tokens`.
- The return value of the `f` call is the expression `0`.
- This inner call collects tokens for `0` (parser range 10..11) and
creates a replacement covering `#[cfg(not(FALSE))] 0` (parser range
0..11).
- We return to the outer `collect_tokens` call. The return value of the
`f` call is *again* the expression `0`, again with the range 10..11,
but the replacement from earlier covers the range 0..11. The code
mistakenly assumes that any attributes from an inner `collect_tokens`
call fit entirely within the body of the result of an outer
`collect_tokens` call. So it adjusts the replacement parser range
0..11 to a node range by subtracting 10, resulting in -10..1. This is
an invalid range and triggers an assertion failure.
It's tricky to follow, but basically things get complicated when an AST
node is returned from an inner `collect_tokens` call and then returned
again from an outer `collect_token` node without being wrapped in any
kind of additional layer.
This commit changes `collect_tokens` to return early in some extra cases,
avoiding the construction of lazy tokens. In the example above, the
outer `collect_tokens` returns earlier because the `0` token already has
tokens and `self.capture_state.capturing` is `Capturing::No`. This early
return avoids the creation of the invalid range and the assertion
failure.
Fixes#129166. Note: these invalid ranges have been happening for a long
time. #128725 looks like it's at fault only because it introduced the
assertion that catches the invalid ranges.
