False positive in `new_ret_no_self` if self type carries lifetime annotation?

[llogiq]

Example (from lib_json):

impl<'a> CharIter<'a> {
    pub fn new<'b>( s: &'b str ) -> CharIter<'b> { .. }
    ..
}

While the warning can be fixed by changing it to pub fn new(s: &'a str) -> CharIter<'a> { .. }, I am unsure if the lifetime difference is really relevant here.

[mcarton]

I should probably not compare types with ==. I’ll look into that tomorrow.

[mcarton]

If I use ctxt::erase_regions and InferCtx::can_equate instead of simply ==, it fixes lifetime problems. But I still have a problem with:

impl<T> CharIter<T> {
    pub fn new<U>( s: &str ) -> CharIter<U> { .. }
    ..
}

Any idea how to compare types with type parameters? 🆘

[llogiq]

This is a bit complex, because the types can literally come from anywhere. But in general, we should consider two unbound type variables in the same position equal, regardless there name, so

impl<T> CharIter<T> {
    pub fn new<U>(..) -> CharIter<U> { .. }
    // should be considered equal to
    pub fn new2(..) -> CharIter<T> { .. }
}

Both T and U are introduced as unbound type variables, one by the inherent impl, the other by the function.

[mcarton]

But in general, we should consider two unbound type variables in the same position equal, regardless there name

I know, but any idea what function does that in rustc?

[llogiq]

I don't think there's a function for this in rustc (because for rustc the types are not equal at all). We'll have to walk the type's TyParams and compare their bounds. If the bounds are equal (either unbounded or bounded by some trait)), we should deem the types equal.

[phansch]

triage:

The first example is fixed (I'm going to add a testcase for that).
The second example still triggers, while it shouldn't:

https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=6a4da2c0118b2f6dfac4456041c2b332

[Jarcho]

This was fixed in #6952