rustdoc's resugaring of projection predicates to type bindings is flawed when supertraits are involved

[fmease]

First discovered thanks to bevyengine/bevy#8898: There, the Output type of the cross-crate opaque impl FnMut(…) -> bool gets dropped by rustdoc resulting in impl FnMut(…). This specific bug is caused by rustdoc's opaque bound cleaner not realizing that it can & should utilize the Output assoc ty defined on FnMut's supertrait FnOnce. During further investigations I've discovered more bugs in this area.

In rustdoc there are two three major and distinct places where we resugar bounds. Both All three are flawed but each one exhibits different bugs.

• (1) clean_middle_opaque_bounds for bounds of opaques (existential impl-Trait)
  • we only check for TraitRef equality while we should also check the super predicates (incl. their substs tho!)
    • (once we do, ensure that we don't intro name clashes into the list of type bindings (this can be triggered by relatively simple user code involving supertraits))
• (2) clean_ty_generics for bounds in where-clauses and APIT (universal impl-Trait)
  • we do check super predicates but we only ever check DefIds (of traits) and never the relevant substs which is super incorrect leading rustdoc to move some projection predicates to unrelated trait bounds
  • the process mutates state by yanking those predicates rendering it order dependent (consider permutations of super and subtraits each one with own type bindings)
• (3) param_env_to_generics in auto_trait.rs for where-clauses of synthetic impls
  • update: replaced by simplify::{where_clauses,sized_bounds} in #123340
  • doesn't use simplify::merge_bounds at all unlike the other two places
  • doesn't merge two partially resugared bounds like Tr<A = …> and Tr<B = …>
  • doesn't render bound vars / higher-ranked parameters at all
  • (I haven't thoroughly checked yet how supertraits are handled but it seems like Fn-family traits are special-cased)

Probably there are some more but less important bugs to be found (e.g. related to bound vars I can imagine).

I'm already deep in the weeds working on a correct & most general fix for this.
Opening this issue to keep the description of the future PR more slender.
@rustbot claim

@rustbot label C-bug T-rustdoc A-cross-crate-reexports

[RalfJung]

The duplicate Target = [u8] here is another instance of this bug.

[fmease]

Major: OPAQ: We don't take supertrait bounds into account when reconstructing assoc item constraints (“type bindings”) from projections (#116735 would've fixed that but I've closed it in favor of a more general PR).

Source

pub fn rpit_fn() -> impl Fn() -> bool {
    || false
}

Rendered inlined cross-crate re-export

pub fn rpit_fn() -> impl Fn()

Source

pub fn f() -> impl TrA<X = ()> {}

pub trait TrA: TrB {}
pub trait TrB { type X; }

impl TrA for () {}
impl TrB for () { type X = (); }

Rendered inlined cross-crate re-export

pub fn f() -> impl TrA

[fmease]

Major: WHERE & APIT: We don't take generic arguments of super trait bounds into account when reconstructing assoc item constraints (“type bindings”) from projections, like at all!

Source

pub struct Ty<T>(T)
where
    T: TrA + TrB<i32, X = ()>;

pub trait TrA: TrB<()> {}
pub trait TrB<D> { type X; }

Rendered inlined cross-crate re-export

pub struct Ty<T>(/* private fields */)
where
    T: TrA<X = ()> + TrB<i32>;

Source

pub struct Ty<T>(T)
where
    T: TrA<X = bool> + TrB<i32, X = ()>;

pub trait TrA: TrB<()> {}
pub trait TrB<D> { type X; }

Rendered inline cross-crate re-export

pub struct Ty<T>(/* private fields */)
where
    T: TrA<X = bool, X = ()> + TrB<i32>;

[fmease]

Minor: WHERE: We don't try to reconstruct overarching / outer binders of clauses.

Source

pub struct Ty<T>(T)
where
    for<'o> T: TrA<'o> + TrB<'o>;

pub trait TrA<'a> {}
pub trait TrB<'a> {}

Rendered inlined cross-crate re-export

pub struct Ty<T>(/* private fields */)
where
    T: for<'o> TrA<'o> + for<'o> TrB<'o>;

---

Edit: Medium: WHERE: We drop the "outer" binder of type-outlives predicates:

Source

pub fn demo<T>() where for<'a> T: 'a {}

Rendered inlined cross-crate re-export

pub fn demo<T>()
where
    T: 'a,

[fmease]

Minor: WHERE: We don't necessarily preserve the order bounds that differ in LHS (bounded ty, lt). However, the HIR cleaner doesn't seem to do that either (requires a different reproducer though).

Source

pub fn f<'a, T>() where T: 'static, 'a: 'static {}

Rendered inlined cross-crate re-export

pub fn f<'a, T>()
where
    'a: 'static,
    T: 'static,

Order is lifetime bounds, then type bounds(, then equality bounds). This happens in simplify::where_clauses.

[fmease]

Minor: OPAQ: We don't necessarily preserve the order of bounds, we move traits in front of lifetime bounds.

[fmease]

Medium: WHERE & APIT: The cleaning process is order-dependent.

pub trait TrA: TrB {}
pub trait TrB { type X; }

pub fn f<T: TrA + TrB<X = ()>>() {} // rendered IXCRX of the bound: `T: TrA<X = ()> + TrB`
pub fn g<T: TrB<X = ()> + TrA>() {} // rendered IXCRX of the bound: `T: TrB<X = ()> + TrA`