Add a new rustc_show_spans internal attribute

[FranciscoTGouveia]

This new internal attribute is a first step for solving #126763.

Given the code:

#![feature(rustc_attrs)]

mod x { pub const Y: u8 = 1; }

macro_rules! path  { ($a:tt, $b:tt) => { let _ = $a::$b; }; }

#[rustc_show_spans(path)]
fn no_macro() { let _ = x::Y; }

#[rustc_show_spans(path)]
fn path_bug() { path!(x, Y); }

We can now report the spans and thus show the bug of the span not falling inside the macro context:

note: showing the spans of the path
 --> test.rs:8:25
  |
8 | fn no_macro() { let _ = x::Y; }
  |                         ^^^^

note: showing the spans of the path
  --> test.rs:11:23
   |
11 | fn path_bug() { path!(x, Y); }
   |                       ^^^^

r? @petrochenkov

[rustbot]

Some changes occurred in compiler/rustc_attr_parsing

cc @jdonszelmann, @JonathanBrouwer

Some changes occurred in compiler/rustc_passes/src/check_attr.rs

cc @jdonszelmann, @JonathanBrouwer

Some changes occurred in compiler/rustc_attr_ir

cc @jdonszelmann, @JonathanBrouwer

[FranciscoTGouveia]

I was undecided about whether to consider this new attribute as Testing or Misc.
Since it seems to be temporary for the migration to the new span combining scheme, I went with the former.

View changes since the review

[FranciscoTGouveia]

Before visiting a new item, I saved the current state of ShowSpans and reset it to support nested items.
Is this necessary, or is it redundant?

View changes since the review

[mejrs]

That depends on whether this should print the path. I'd, perhaps naively, expect it to do print it, in which case resetting is wrong.

#[rustc_show_spans(path)]
fn bar() {
    fn foo() {
        let _ = a::b::c;
    }
}

(this is also something that should have a test)

[mejrs]

Can you also add some tests?

It need not be completely implemented, but its current state must be reflected in the test suite.

View changes since this review

[mejrs]

You can use parse_limited_sym in this case.

Also, no need for the item.attrs.iter().any(|attr| attr.has_name(sym::rustc_show_spans)) check, attribute parsing will check this.

[FranciscoTGouveia]

Oh, I did not know about parse_limited_sym; thanks!

The idea with item.attrs.iter().any(|attr| attr.has_name(sym::rustc_show_spans)) was to have a cheaper check to exit early. However, I am now realizing this is not a hot path, so the benefit is nil.

[mejrs]

That depends on whether this should print the path. I'd, perhaps naively, expect it to do print it, in which case resetting is wrong.

#[rustc_show_spans(path)]
fn bar() {
    fn foo() {
        let _ = a::b::c;
    }
}

(this is also something that should have a test)

[rustbot]

Reminder, once the PR becomes ready for a review, use @rustbot ready.