Make trait refs & assoc ty paths properly induce trait object lifetime defaults

[fmease]

View all comments

Trait Object Lifetime Defaults

Primer & Definitions

You could read this section in the Reference but it has several issues (see rust-lang/reference#1407). Here's a small explainer by me that only mentions the parts relevant to this PR:

Basically, given dyn Trait (≠ dyn Trait + '_) we want to deduce its trait object lifetime bound from context without relying on normal region inference as we might not be in a body¹. The "context" means the closest – what I call – (eligible) container C<X0, …, Xn> that wraps this trait object type. A container is to be understood as a use site of a "parametrized definition" (more general than type constructors). Currently eligible are ADTs, type aliases, traits and enum variants.

So if we have C<dyn Trait> (e.g., &'r dyn Trait or Struct<'r, dyn Trait>), D<C<dyn Trait>> or C<N<dyn Trait>> (e.g., Struct<'r, (dyn Trait,)>), we use the explicit² outlives-bounds on the corresponding type parameter of C to determine the trait object lifetime bound. Here, C & D denote (eligible) containers and N denotes a generic type that is not an eligible container. E.g., given struct Struct<'a, T: 'a + ?Sized>(…);, we elaborate Struct<'r, dyn Trait> to Struct<'r, dyn Trait + 'r>.

Finally, we call lifetime bounds used as the default for constituent trait object types of an eligible container C the trait object lifetime defaults (induced by C). These defaults may of course end up getting shadowed in parts of the type by the defaults induced by any inner eligible containers.

Changes Made

These changes are theoretically breaking.

1. Make resolved associated type paths / projections eligible containers.
   • <Y0 as TraitRef<X0, …, Xn>>::AssocTy<Y1, …, Ym> now induces trait object lifetime defaults for constituents Y0 to Ym (TraitRef is considered a separate container, see also list item (3)).
   • Notably, for the self type Y0 of (resolved) projections we now look at the bounds on the Self type param of the relevant trait (e.g., given trait Outer<'a>: 'a { type Proj; } or trait Outer<'a> where Self: 'a { type Proj; } we elaborate <dyn Inner as Outer<'r>>::Proj to <dyn Inner + 'r as Outer<'r>>::Proj).
   • Example breakages:

     trait Outer<'a> { type Ty<T: ?Sized + 'a>; }
     impl<'a> Outer<'a> for () { type Ty<T: ?Sized + 'a> = &'a T; }
     trait Inner {}
     
     fn f<'r>(x: <() as Outer<'r>>::Ty<dyn Inner>) {
     //                                ~~~~~~~~~
     //                                this branch:  dyn Inner + 'r       (due to bound `'a` on `T`)
     //                                stable/main:  dyn Inner + 'static  (due to item signature fallback)
         let _: <() as Outer<'r>>::Ty<dyn Inner + 'static> = x;
     //         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
     //         this branch:  error: lifetime may not live long enough // `'r` must outlive `'static`
     //         stable/main:  OK
     }

     trait Outer { type Ty; }
     trait Inner {}
     
     impl<'a> Outer for dyn Inner + 'a { type Ty = &'a (); }
     
     fn f<'r>(x: *mut &'r <dyn Inner as Outer>::Ty) {
     //                    ~~~~~~~~~
     //                    this branch:  dyn Inner + 'static  (due to lack of bounds on `Outer`; the assoc type path shadows the default induced by the type ctor `&`)
     //                    stable/main:  dyn Inner + 'r       (due to bound `'a` on `T` in (pseudo) `builtin type &<'a, T: 'a + ?Sized>;`)
         let _: *mut &'r <dyn Inner + 'r as Outer>::Ty = x;
     //         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
     //         this branch:  error: lifetime may not live long enough // `'r` must outlive `'static`
     //         stable/main:  OK
     }

2. In type-relative paths Y0::Name<Y1, …, Ym> consider the trait object lifetime default indeterminate
   • Meaning if we're in an "item context" / "item signature" / "non-body" (& the principal trait isn't bounded by any outlives-bounds which would take precedence over the default) we will reject any implicit trait object lifetime bounds that would take on that default
   • Reason: Limitations of the current implementation which can't be easily overcome
     • RBV (which resolves trait object lifetime defaults by recursing into the local crate "in one sitting") would require the resolution of type-relative paths in order to look up the generics but these paths are only resolved in HIR ty lowering (that can selectively lower local items) which depends on the results of RBV (cyclic dependency!)
     • While one might be able to resolve type-relative paths in RBV in an ad-hoc fashion, it would require a lot of duplication with HIR ty lowering and its impl would be very brittle (RTN does something like that in RBV but we require a more sophisticated resolver)
     • I did attempt that but it got too gnarly and brittle and would've likely been incomplete anyway
     • See also this GH thread
     • See also #t-types/meetings > 2025-09-16 weekly @ 💬
   • This should still be maximally forward compatible and allow us to implement the desired behavior in the future.
   • Example breakage:

     trait Outer { type Ty<'a, T: 'a + ?Sized>; }
     trait Inner {}
     
     fn f<'r, T: Outer>(x: T::Ty<'r, dyn Inner>) {}
     //                              ~~~~~~~~~
     //                              this branch:  error: indeterminate  (reservation)
     //                              stable/main:  dyn Inner + 'static   (due to item signature fallback)

3. Fixes trait object lifetime defaults inside trait refs TraitRef<X0, …, Xn> (this fell out from the previous changes). They used to be completely broken due to a nasty off-by-one error for not accounting for the implicit Self type param of traits which lead to cases like
   • Outer<'r, dyn Inner> (with trait Outer<'a, T: 'a + ?Sized> {}) getting rejected as indeterminate (it tries to access a lifetime at index 1 instead 0) (playground)
   • Outer<'r, 's, dyn Inner> (with trait Outer<'a, 'b, T: 'a + ?Sized> {}) elaborating dyn Inner to dyn Inner + 's instead of dyn Inner + 'r(!) which subsequently gets rejected of course since 's isn't known to outlive 'r (playground)
   • The same applies to trait alias refs (feature trait_alias)
   • Example breakage:

     trait Outer<'a, 'b, T: 'a + ?Sized> {}
     trait Inner {}
     
     struct F<'r, T>
     where
         T: Outer<'r, 'static, dyn Inner>
     //                        ~~~~~~~~~
     //                        this branch:  dyn Inner + 'r       (correctly mapping `'a` => `'r`)
     //                        stable/main:  dyn Inner + 'static  (incorrectly mapping `'a` => `'static` due to off-by-one)
     {
         g: G<'r, T>,
     //     ~~~~~~~~
     //     this branch:  error: mismatched types
     //                          expected: `Outer<'r, 'static, (dyn Inner + 'static)>`
     //                             found: `Outer<'r, 'static, (dyn Inner + 'r)>`
     //     stable/main:  OK
     }
     
     struct G<'r, T>(&'r (), T)
     where
         T: Outer<'r, 'static, dyn Inner + 'static>;

4. In associated type binding TraitRef<AssocTy<X0, …, Xn> = Y> consider the trait object lifetime default indeterminate (in X0, …, Xn and Y) if X0, …, Xn contains any lifetime arguments.
   • Meaning if we're in an item context (& the principal trait isn't bounded) we will reject any implicit trait object lifetime bounds that would take on that default
   • This reserves us the right to (1) take into account the item bounds of AssocTy in the future when computing the default for Y (2) take into account the parameter bounds of AssocTy in the future when computing the defaults for X0, …, Xn.
   • This extends a preexisting hack that – given TraitRef<X0, …, Xn, AssocTy<Y0, …, Ym> = Z> – treats the default indeterminate in Y0, …, Ym and Z if X0, …, Xn contains any lifetime arguments.
   • Rephrased, this hack / reservation previously didn't account for GAT args, only trait ref args, which is insufficient
   • See also this GH comment of mine
   • Example breakages:

     trait Outer { type Ty<'a>: ?Sized; }
     trait Inner {}
     
     fn f<'r>(_: impl Outer<Ty<'r> = dyn Inner>) {}
     //                              ~~~~~~~~~
     //                              this branch:  error: indeterminate  (reservation)
     //                              stable/main:  dyn Inner + 'static   (forced)

     trait Outer { type Ty<'a, T: ?Sized + 'a>; }
     trait Inner {}
     
     fn f<'r>(_: impl Outer<Ty<'r, dyn Inner> = ()>) {}
     //                            ~~~~~~~~~
     //                            this branch:  error: indeterminate  (reservation)
     //                            stable/main:  dyn Inner + 'static   (forced)

Motivation

Both trait object lifetime default RFCs (599 and 1156) never explicitly specify what constitutes a — what I call — (eligible) container but it only makes sense to include anything that can be parametrized by generics and can be mentioned in places where we don't perform full region inference … like associated types. So it's only consistent to make this change.

Breakages

These changes are theoretically breaking because they can lead to different trait object lifetime bounds getting deduced compared to main which is obviously user observable. Moreover, we're now explicitly rejecting implicit trait object lifetime bounds inside type-relative paths (excl. the self type) and on the RHS of assoc type bindings if the assoc type has lifetime params.
However, the latest crater run found 0 non-spurious regressions (see here and here).

---

Fixes #115379.
Fixes #140710.
Fixes #141997.

Footnotes

1. If we are in a body we do use to normal region inference as a fallback. ↩

2. Indeed, we don't consider implied bounds (inferred outlives-bounds). ↩