Tracking Issue for attributes changing "Minimal Complete Definition" of a trait

[WaffleLapkin]

This is a tracking issue for attributes changing "Minimal Complete Definition" of a trait. "Minimal Complete Definition" is effectively a set of rules you need to follow (by implementing items of the trait) in order for a trait implementation to be "Complete" (and compile). Normally there is only one rule — implement all items that do not have defaults. However, sometimes it is meaningful to add additional restrictions.

As a simple example consider a trait with functions equal and not_equal, both of them can be implemented as !the_other_function(...), but only if the other one has a meaningful implementation. Thus, it may be beneficial for the library design to make Minimal Complete Definition be "you must implement at least one of equal and not equal".

A more realistic example would probably be performance related — one function is easier to implement, while the other theoretically allows a more performant implementation (see Read::{read, read_buf}).

Current status

This is currently implemented as an internal unstable rustc attribute #[rustc_must_implement_one_of]. It accepts a list of identifiers of trait items and adds a requirement that at least one of the items from the list is implemented. A usage example:

// `read` and `read_buf` are mutually recursive, it would be bad to implement neither
#[rustc_must_implement_one_of(read, read_buf)]
pub trait Read {
    fn read(&mut self, buf: &mut [u8]) -> Result<usize> {
        let mut buf = ReadBuf::new(buf);
        self.read_buf(&mut buf)?;
        Ok(buf.filled_len())
    }

    fn read_buf(&mut self, buf: &mut ReadBuf<'_>) -> Result<()> {
        default_read_buf(|b| self.read(b), buf)
    }
}

impl Read for Ty0 {} 
//^ This will fail to compile even though all `Read` methods have default implementations

// Both of these will compile just fine
impl Read for Ty1 {
    fn read(&mut self, buf: &mut [u8]) -> Result<usize> { /* ... */ }
}
impl Read for Ty2 {
    fn read_buf(&mut self, buf: &mut ReadBuf<'_>) -> Result<()> { /* ... */ }
}

It is hopefully not the final form of the attribute, but just a "MVP". Some ideas that we might want to explore in the future:

• A way to specify multiple sets of items (i.e. a is mutually recursive with b, c is mutually recursive with d, so MCD is (a | b) & (c | d))
• Add support for non-function items (recursive constants? is this useful?)
• Consider if the trait-level attribute is a good solution
  • Would something like item-level #[requires(...)] be easier to use/reason about?
• Iterate on rustdoc output
• Consider how we can support more complicated cases like the Iterator, where a lot of methods can be implemented in terms of each other
  • Would it make sense to have multiple default implementations of an item and choose one of them depending on whatever another item is implemented?
• Write an RFC (!)

Stability

While the attribute itself is very unstable and "MVP", we can still use it in stable and unstable APIs, if we are sure that this is something we wish to support in the future. Thus, we need to be able to control the stability of default implementations (to be able to test if we even need to change MCD of a trait, for example).

For this we have 2 tools. We can either require an unstable method (#[rustc_must_implement_one_of(something_stable, something_unstable)], we can add a default implementation to something_stable without worrying — to use the default implementation one would need to opt-in into unstable feature to implement something_unstable), or use the #[rustc_default_body_unstable] attribute.

#[rustc_default_body_unstable] attribute, as with any other stability attributes, allows to set a feature gate to using a default body. For example:

// in std
pub trait Trait {
    #[rustc_default_body_unstable(feature = "feat", isssue = "none")]
    fn item() {}
}

// in a user crate

impl Trait for Type {} // <-- does not provide an `item` implementation, so it uses the default one
//~^ error: not all trait items implemented, missing: `item`
//~| note: default implementation of `item` is unstable
//~| note: use of unstable library feature 'feat'
//~| help: add `#![feature(constant_default_body)]` to the crate attributes to enable

// this is fine
impl Trait for AnotherType {
    fn item() { println!("hehe"); }
}

About tracking issues

Tracking issues are used to record the overall progress of implementation.
They are also used as hubs connecting to other relevant issues, e.g., bugs or open design questions.
A tracking issue is however not meant for large scale discussion, questions, or bug reports about a feature.
Instead, open a dedicated issue for the specific matter and add the relevant feature gate label.

Steps

• Implement an MVP
• Write an RFC
• Implement the RFC
• Adjust documentation (see instructions on rustc-dev-guide)
• Stabilization PR (see instructions on rustc-dev-guide)

Unresolved Questions

See the "Current status".

Implementation history

• Implement #[rustc_must_implement_one_of] attribute #92164
• Check that #[rustc_must_implement_one_of] is applied to a trait #93386
• Implement #[rustc_default_body_unstable] #96478
• rustdoc: Add support for #[rustc_must_implement_one_of] #99235
• Tweak rustc_must_implement_one_of diagnostic output #105506

(potential) Uses

• Allow only implementing Read::read_buf #106643

[the8472]

Would it make sense to have multiple default implementations of an item and choose one of them depending on whatever another item is implemented?

Imo that is almost necessary to have that. Otherwise the default impls are forced to form a cyclic graph so that the dependencies always hit an implemented one. But for larger sets of alternatives those chains are likely suboptimal implementations.
Even if there's only 2 options one might still want the other default bodies that aren't part of the cycle to directly choose which one they use.

[joshtriplett]

Tracking issue looks great! Can we get a draft PR adding this (as must_implement_one_of) to the reference?

(To clarify, that's not a blocker for using this on nightly, that's a blocker for stabilizing read_buf with this attribute applied to Read.)

[WaffleLapkin]

@joshtriplett I'm not sure it's worth adding to the reference, as I said in my opinion it's quite incomplete.

[Sky9x]

must_implement_one_of could also be used for the Wake trait, avoiding dilemma of if wake or wake_by_ref should be defaulted.

[the8472]

(potential) Uses

Another one is Iterator. Currently next is the mandatory method to implement. With this attribute implementations could also offer try_fold or next_chunk as their base methods.

[WaffleLapkin]

@the8472 sadly... this is not actually possible. try_fold requires Self: Sized, so you can't call it in next (play).

Unless you remove the bound from try_fold, which I think is a breaking change? (and also makes the vtable bigger, but that's unlikely to be a problem)

[Sky9x]

try_fold requires Self: Sized because it has a generic and can't be used with dynamic dispatch (dyn Iterator). However, it might still be possible to implement next in terms of it, but I don't know how viable or useful that would be.

[the8472]

This circles back to

Would it make sense to have multiple default implementations of an item and choose one of them depending on whatever another item is implemented?

next would then only be implementable in terms of next_fold if try_fold is implemented, e.g. because Sized is always true.

[chernoivanenkoofficial]

The core neccessity for this feature , in my opinion, first of all comes from the default implementation validness assertion, which compiler cannot (at the moment), and from the design standpoint even shouldn't track. The need for restriction of trait's MCD is just byproduct of such requirement.

Instead of looking at "completeness" of trait implementation from outer, trait "POV", we can always say, that for trait to be minimally implemented, each of its items should be minimally implemented.

Currently, this perspective view change leads to next requirement (as stated in issue description) for member minimal implementation - Member should have default implementation or Member should be implemented in impl block.

This default requirement is constructed from three conditions:

1. Member should have default implementation
2. Member should be implemented in impl block.
3. And top level condition Either condition 1 or condition 2 should be satisfied.

We could change condition 1 to be Member should have valid default implementation, where validness of default implementation is defined by additional requirements in the form of attributes, e.g. Default implementation requires another member to be implemented in the impl block in the form of

trait Foo {
    #[requires_impl(b)] // or even #[default_requires_impl(b)] for more clarity
    fn a() {
        Self::b()
    }

    #[requires_impl(a)]
    fn b() {
        Self::a()
    }
}

This allows not only to understand minimal requirements when implementing trait as a whole, but also provides a little perspective into design of intended behavior of each trait member and their interconnection, which could be handy e.g. in the case of big "internally co-dependent" traits like Iterator.

As for multiple default implementations for member, I am not sure of cases, where it can be usefull, but if it was to be done, member-level approach would be even more appropriate, as each default implementation could be treated as separate "pseudo-members", each with their own requirements, which then can be "eliminated" and coerced to a single MCD of the member.

But for this to work compiler should be able to ensure directly at the trait definition site, that no ambiguity between available default implementations is possible, if member-wise requirements were satisfied. This task seems to be much more complicated, as it requires "negative requirements", but on the other hand, there's no need fot it to be part of this proposition. And if it is needed later, this feature could be built on top of the core "member-wise requirements" (or even "member-wise requirement tree") assertion for MCD.

[WaffleLapkin]

@chernoivanenkoofficial I sympathize with your idea (and I think I've heard similar ideas before, so you are not alone thinking this). I'm a bit concerned though that it will make checking for breaking changes even harder, since the requirements will be spread throughout the trait.