`new_without_default_derive` False Positive

[oxalica]

It seems deriving Default on struct with generic type and lifetime restriction will cause it.
In this example, clippy 0.0.171 suggests me to add #[derive(Default)] again.

#[derive(Default)]
pub struct S<'a, T: 'a>(Option<&'a T>);

// Manually impl (remove `#[derive(Default)]`) works as expected.
// impl<'a, T: 'a> Default for S<'a, T> {
//     fn default() -> Self {
//         S::new()
//     }
// }

impl<'a, T: 'a> S<'a, T> {
    pub fn new() -> Self { 
        S(None) 
    }
}

fn main() {
    let _ = S::<i32>::default(); // works
}

[oli-obk]

I'm confused. What's the wrong behaviour?

[oxalica]

S derives Debug and it works. But clippy suggests me to add #[derive(Default)] again.

[oli-obk]

Oh... that is so weird. I'll have a look

[oli-obk]

Ok. I found the culprit:

error[E0277]: the trait bound `T: std::default::Default` is not satisfied
  --> tests/ui/new_without_default.rs:90:28
   |
90 |     pub fn new() -> Self { Default::default() }
   |                            ^^^^^^^^^^^^^^^^ the trait `std::default::Default` is not implemented for `T`
   |
   = help: consider adding a `where T: std::default::Default` bound
   = note: required because of the requirements on the impl of `std::default::Default` for `FalsePositive<'_, T>`
   = note: required by `std::default::Default::default`

Self really does not implement Default, thus we suggest to derive it when we should be suggesting to manually impl it

[oxalica]

The deriving impl of Default requires T: Default although we have Option<T>: Default for all T. Is this the expected behavior? It's quite confusing.
Without lifetimes, #[derive(Default)] pub struct S<T>(Option<T>); also has this issue.

[oli-obk]

Yea this is a limitation of derive. Rustc tracks this under rust-lang/rust#26925

[oli-obk]

Anyone interested in fixing this. Essentially you need to do the same as the missing_debug_implementations lint: https://github.com/rust-lang/rust/blob/2278506f682062c7988c9e534a94366c56693981/src/librustc_lint/builtin.rs#L588-L600 instead of the current logic in can_derive_default that calls !implements_trait(...). If there are any impls of Default for the type, irrelevant of the generics, then just not report the lint.

A future lint could detect derives that could be made more general by turning them into implementations. So essentially fixing the rust bug of bad derives by detecting bad derives and doing any of the fixes suggested in the issue.