`*mut` "uses" type params, but `*const` (and therefore `NonNull`) does not

[eggyal]

Update(fmease): See #118976 (comment) and #118976 (comment).

---

As first mentioned on URLO, and demonstrated on quick playground:

struct Foo<T>(*mut Self); // works
struct Bar<T>(*const Self); // error[E0392]: parameter `T` is never used

Is this a bug, or is there some obscure reason for it?

[Darksonn]

Note that the use of Self is not the problem:

struct Foo<T>(*mut Foo<T>); // works
struct Bar<T>(*const Bar<T>); // error[E0392]: parameter `T` is never used

[fmease]

*mut is invariant over its pointee type while *const is covariant over it. Foo is invariant over T because invariance “overrules” bivariance (all to-be-inferred terms start out as bivariant). Bar is bivariant over T (i.e., T is “unused”) since covariance gets “overruled” by bivariance. Sorry for the super technical explanation. I don't really see a bug though 🤔

Similar example:

struct Foo<T: 'static>(&'static mut Foo<T>); // works
struct Bar<T: 'static>(&'static Bar<T>); // errors

You can remedy this with PhantomData and a wrapper type that renders T invariant, e.g., struct Bar<T>(*const Bar<T>, PhantomData<*mut T>); should be just fine.

[eggyal]

Ah, thank you. I had thought it might be variance-related and did refresh my memory with a read of the chapters on subtyping in the reference and nomicon but neither mention "bivariance" at all nor how variance relates to whether a type parameter is deemed to be "used" in a type definition.

Perhaps this is already documented somewhere that I missed, but if not maybe this is a documentation bug?

[fmease]

The rustc dev guide mentions bivariance but that one isn't really user-facing as its name suggests. Not sure why the Reference and the Nomicon don't mention it, maybe because it's considered to be an implementation detail?

[eggyal]

Can it really be an implementation detail if its effect is user facing ?

[fmease]

Yeah, makes sense. Probably worth opening an issue over at the Reference and/or the Nomicon. It's an implementation detail insofar that rustc never mentions bivariance in any error messages for example and only talks about generic parameters “that must be used”.

[eggyal]

Okay, thanks again. Will open an issue as advised. Closing here.

[QuineDot]

In short "uses" really means "could determine variance of".

Related issues:

• unused_lifetimes warning misses unused lifetime #74165
• Unused generic parameter E0392 not emitted in recursive type that is behind Cell  #92561
• "parameter T is never used" with recursive type #26283
• Recursive generic type parameters can give confusing error messages #53191
• misleading E0392 "parameter T is never used" for every containing struct #66912
• Parameter considered not used in recursive types #105740
• Recursive types in a Vec<T> with generic parameters gives ambiguous error #110466

[eggyal]

Yes, that's a helpful clarification. Perhaps that could be added into the explanation of E0392?