an empty :vis doesn't work in trait declaration

[limira]

I want to create a macro that generate fns that use in both impl Item... and trait TraitName. In trait, pub must be omitted, but in impl SomeStruct I need pub, so the macro need a $vis:vis like this:

macro_rules! create_method {
    ($vis:vis $name:ident) => {
        $vis fn $name(&self) {
            println!("{}", stringify!($name));
        }
    }
}

trait T1 {
    // This not works
    create_method!(method_of_t1);
}

trait T2 {
    fn method_of_t2(&self);
}

struct Struct;

impl T2 for Struct {
    // This works
    create_method!(method_of_t2);
}

impl Struct {
    // I want this method to be `pub`, so the macro need a `:vis`
    // for receiving the visibility keyword like this:
    create_method!(pub another_method);
}

Expected result: create_method!(method_of_t1); in trait T1 should works.

(Edit: a better comment about the :vis in impl Struct)

[petrochenkov]

@Centril
This looks relevant to #64910 (comment) and the parser work you've been doing recently.

[Centril]

    // The macro need a `:vis` to generate `fn` for a normal `impl`

This bit doesn't seem right; when I remove the pub it works (which is expected because the impl T2 for Struct { works).

Looking at fn parse_trait_item_ there seems to be something odd going on. In particular we have:

• self.eat_bad_pub();

• } else if let Some(mac) = self.parse_assoc_macro_invoc("trait", None, &mut false)? {

  (note the None here -- for fn parse_impl_method we instead have:

  if let Some(mac) = self.parse_assoc_macro_invoc("impl", Some(vis), at_end)? {

  This suggests this is somehow by design? Odd...!

[Centril]

...actually it's not that odd. This code was likely written before we had the vis matcher.

I think what we want to do is to use let vis = self.parse_visibility(false)?; inparse_trait_item_ as well and then do a check right after to see if the visibility is not "empty" (VisibilityKind::Inherited) -- that would probably fix the issue here and allow us to discuss semantic vs. syntactic errors for pub fn foo in trait defs later.

[limira]

This bit doesn't seem right;

Sorry for the confusion caused by the comment. I edit the post, hope it is better now!

[Centril]

We discussed this on the language team meeting today. The conclusion was that we regarded the issue here as a sort of bug that should be fixed (which I will do soon according to my notes above).

We also took the opportunity to discuss the grand unification of the 3 item grammars that @petrochenkov raised in #64910 (comment). We found the idea of "grand unification" and moving to semantic errors to be generally appealing as well.

[Centril]

Finally got around to this. I have a fix in #66183.