Detect pub structs never constructed and unused associated constants

[mu001999]

Lints never constructed public structs.

If we don't provide public methods to construct public structs with private fields, and don't construct them in the local crate. They would be never constructed. So that we can detect such public structs.

---

Update:

Also lints unused associated constants in traits.

[rustbot]

r? @wesleywiser

rustbot has assigned @wesleywiser.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[mu001999]

r? @pnkfelix

cc @shepmaster

[rustbot]

Some changes occurred to MIR optimizations

cc @rust-lang/wg-mir-opt

[SparrowLii]

just some nits

@pnkfelix

[bors]

☔ The latest upstream changes (presumably #125759) made this pull request unmergeable. Please resolve the merge conflicts.

[pnkfelix]

I found the state of this analysis interesting, especially the fact that it flags the struct as dead even when there's a impl Clone of it sitting there.

This makes sense: the presence of clone does not provide evidence that the struct is live, because you need an instance of the struct in order to be able to enter the clone code in the first place, in order to call the fn clone(&self) -> Self method.

But it led me to wonder: How far does this reasoning go? I.e., does it also filter out other variations of the same inductive reasoning, such as a variant trait that takes x: Self or x: &Self?

So I built a local copy, and soon determined: No, the reasoning does not go that far. That is, either of the following trait impls will suffice to treat a struct constructor as "live code":

pub trait U { fn f(_: Self) -> Self; }
impl U for S { fn f(x: Self) -> Self { S(10) } }


pub trait W { fn f(_: &Self) -> Self; }
impl W for S { fn f(x: &Self) -> Self { S(10) } }

Note: This is fine. I'd rather we start with a lint that doesn't attempt to catch such cases, than risk shipping an over-aggressive lint that ends up also accidentally treating things like fn f(x: Option<&Self>) -> Self as proof of life.

[SparrowLii]

This should be because struct_all_fields_are_public additionally checks that each field of the struct is pub or not. We may have to find a way to reduce the number of extra checks for struct fields, such as changing the return value of the function to

struct Publicness {
    ty_is_public: bool,
    ty_and_all_fields_are_public: Option<bool>,
}

Obtain whether each field of the structure is pub only when necessary.

@mu001999 @Kobzol

[mu001999]

I think there is another source of extra time like #121752 (comment), more items in the worklist means that these items may be checked repeatedly.