Let Option derive #[must_use]

[dhardy]

Let Option and Box derive #[must_use] from their generic parameter T.

Rendered

[burdges]

Is there some reason #[must_use] should not be a marker trait?

[Lokathor]

I would say that actually must_use is more a property of a particular function, not really of types, and sticking it on types is a shorthanded way to put it on a lot of functions "for free" cognitively. I'd say it's best if we, as an ecosystem, stick to having must_use on functions/methods rather than types.

[dlight]

I would say that actually must_use is more a property of a particular function, not really of types, and sticking it on types is a shorthanded way to put it on a lot of functions "for free" cognitively. I'd say it's best if we, as an ecosystem, stick to having must_use on functions/methods rather than types.

So #[must_use] could be a marker trait on the function's type where it is applied (each function has a unique type) and on each type it is applied. Then you add a blanked impl on functions whenever their output type impls the trait.

(it could be two traits, one for types you apply #[must_use] and another for functions, so that returning the function doesn't trigger it)

[dhardy]

I would say that actually must_use is more a property of a particular function, not really of types

Yes, but #[must_use] on types already has stable support (since 1.0 I think).

So #[must_use] could be a marker trait

Perhaps, but that would be quite a significant revision. Definitely out-of-scope for this RFC.

[dlight]

Agreed! For this RFC a trait is not necessary. If extending this feature to custom types is desired, a trait would be useful

[kennytm]

Note that it is already possible to mark a trait as #[must_use] (which is how rustc can warn about unused Future). It does not help this RFC though.

[soqb]

why "magically"?

why not permit #[must_use] on the parameters themselves, e.g.

pub struct Option<#[must_use] T> {
    Some(T),
    None,
}

"desugars"¹ to

impl<T: MustUse> MustUse for Option<T> {}

Footnotes

1. There will not necessarily be an actual MustUse trait in the implementation. ↩

[dhardy]

impl<T: MustUse> MustUse for Option {}

#[must_use] is not currently a trait. If you want to make an RFC for that, please go ahead. It may be a good idea, but it is an idea for another RFC, not this one.

Since #[must_use] is not currently a trait, we lack the language to describe derive behaviour, hence use of "magically" here.

Option<#[must_use] T>

Use of #[must_use] on parameters is novel syntax that I don't particularly like.

[soqb]

sorry, i thought i was clear in the original comment that "the MustUse trait" was mentioned purely as an analogy, and not as a normative suggestion, but perhaps not. fwiw i think a MustUse trait would be bad since

1. trait solver is slow and heavy compared to just looking at some simple hard-coded rules based on named types/traits.
2. it gives perhaps too much flexibility to the user. do we really want people to be able to write impl<T: Copy> MustUse for Thing<T>, for instance.

i think users should have access to more fine-grained #[must_use] annotations, but that doesn't per-se have to be part of this change since Option and Box alone would be clear usability improvements in my book.

i don't want to overlitigate, but i still reckon this RFC doesn't really justify the "magic" design as written: it's clearly conceptually easier to special case these two generic types, but i don't see it as better from a maintenance perspective if we do want to rip it out for something holistic. for instance is_ty_must_use is already more complicated than ideally it should be (sidebar: and, seemingly, incomplete; see this playground). should we keep adding additional branches ad infinitum?

[kennytm]

Use of #[must_use] on parameters is novel syntax that I don't particularly like.

Attribute on type parameter is already in use since #3621 (ref), it is no longer a novel syntax.

#[derive(CoercePointee)]
#[repr(transparent)]
struct MySmartPointer<#[pointee] T: ?Sized, U> {
    ptr: Box<T>,
    _phantom: PhantomData<U>,
}

[dhardy]

Option<#[must_use] T>

Something like this would be clearer (though I do find derive's "magic" handling of generic parameters unfortunate relative to making this explicit, as with autoimpl):

#[derive(must_use)]
enum Option<T> { /* ... */ }

That may be besides the point though if this RFC adds more complexity than value. (I had no idea that Result<T, Infallible> would not trigger a "must use" warning.)

[dhardy]

Since there have been a number of suggestions about how to generalise this RFC, let me be clear: this RFC is extremely specific on purpose:

• This RFC affects only two std types in a way which should be compatible with any more general "must_use derive" mechanism
• It addresses what I see as the largest short-comings of #[must_use] without the complexity of a general "attribute derive mechanism" or turning #[must_use] into a trait

[tmccombs]

It seems strange to me for it to magically apply to these two types, without a more general mechanism.

And why not Result? Or Pin?

[burdges]

Result is #[must_use] now. The issue here is Option<Result<..>> and Box<Result<..>> are not #[must_use] now, so you must manually put a #[must_use] on fns returning them.

Afaik, no pointer types are #[must_use] so Pin does not come up, but sure Pin<Box<Result<..>>> becomes a question if Box<Result<..>> becomes #[must_use].