rustc_next_trait_solver/solve/

trait_goals.rs

//! Dealing with trait goals, i.e. `T: Trait<'a, U>`.

use rustc_type_ir::data_structures::IndexSet;
use rustc_type_ir::fast_reject::DeepRejectCtxt;
use rustc_type_ir::inherent::*;
use rustc_type_ir::lang_items::SolverTraitLangItem;
use rustc_type_ir::solve::{
    AliasBoundKind, CandidatePreferenceMode, CanonicalResponse, ExternalConstraintsData, MaybeInfo,
    NoSolutionOrRerunNonErased, OpaqueTypesJank, QueryResultOrRerunNonErased, RerunNonErased,
    RerunReason, RerunResultExt, SizedTraitKind,
};
use rustc_type_ir::{
    self as ty, ClausePolarity, ExistentialPredicate, FieldInfo, Interner, MayBeErased, Movability,
    Region, TraitClause, TraitRef, TypeVisitableExt as _, TypingMode, Unnormalized, Upcast as _,
    elaborate,
};
use tracing::{debug, instrument, trace, warn};

use crate::delegate::SolverDelegate;
use crate::solve::assembly::structural_traits::{self, AsyncCallableRelevantTypes};
use crate::solve::assembly::{
    self, AllowInferenceConstraints, AssembleCandidatesFrom, Candidate, FailedCandidateInfo,
};
use crate::solve::inspect::ProbeKind;
use crate::solve::{
    BuiltinImplSource, CandidateSource, Certainty, EvalCtxt, Goal, GoalSource, MaybeCause,
    MergeCandidateInfo, NoSolution, ParamEnvSource, StalledOnCoroutines, flatten_answer_tree,
    has_only_region_constraints,
};

impl<D, I> assembly::GoalKind<D> for TraitClause<I>
where
    D: SolverDelegate<Interner = I>,
    I: Interner,
{
    fn self_ty(self) -> I::Ty {
        self.self_ty()
    }

    fn trait_ref(self, _: I) -> ty::TraitRef<I> {
        self.trait_ref
    }

    fn with_replaced_self_ty(self, cx: I, self_ty: I::Ty) -> Self {
        self.with_replaced_self_ty(cx, self_ty)
    }

    fn trait_def_id(self, _: I) -> I::TraitId {
        self.def_id()
    }

    fn consider_additional_alias_assumptions(
        _ecx: &mut EvalCtxt<'_, D>,
        _goal: Goal<I, Self>,
        _alias_ty: ty::AliasTy<I>,
    ) -> Vec<Candidate<I>> {
        vec![]
    }

    fn consider_impl_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, TraitClause<I>>,
        goal_trait_ref: TraitRef<I>,
        impl_def_id: I::ImplId,
        then: impl FnOnce(&mut EvalCtxt<'_, D>) -> QueryResultOrRerunNonErased<I>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        let cx = ecx.cx();

        let impl_trait_ref = cx.impl_trait_ref(impl_def_id);
        if !DeepRejectCtxt::relate_rigid_infer(ecx.cx())
            .args_may_unify(goal_trait_ref.args, impl_trait_ref.skip_binder().args)
        {
            return Err(NoSolution.into());
        }

        // For every `default impl`, there's always a non-default `impl` that will *also* apply.
        // There's no reason to register a candidate for this impl, since it is *not* proof that
        // the trait goal holds.
        if cx.impl_is_default(impl_def_id) {
            return Err(NoSolution.into());
        }

        match (cx.impl_polarity(impl_def_id), goal.predicate.polarity) {
            // Impl matches polarity
            (ty::ImplPolarity::Positive, ty::ClausePolarity::Positive)
            | (ty::ImplPolarity::Negative, ty::ClausePolarity::Negative) => {}

            // Impl doesn't match polarity
            (ty::ImplPolarity::Positive, ty::ClausePolarity::Negative)
            | (ty::ImplPolarity::Negative, ty::ClausePolarity::Positive) => {
                return Err(NoSolution.into());
            }
        }

        if ecx.typing_mode().is_reflection() && !cx.is_fully_generic_for_reflection(impl_def_id) {
            return Err(NoSolution.into());
        }

        ecx.probe_trait_candidate(CandidateSource::Impl(impl_def_id)).enter(|ecx| {
            let impl_args = ecx.fresh_args_for_item(impl_def_id.into());
            ecx.record_impl_args(impl_args);
            let impl_trait_ref = impl_trait_ref.instantiate(cx, impl_args).skip_norm_wip();

            ecx.eq(goal.param_env, goal_trait_ref, impl_trait_ref)?;
            let where_clause_bounds = cx
                .clauses_of(impl_def_id.into())
                .iter_instantiated(cx, impl_args)
                .map(Unnormalized::skip_norm_wip)
                .map(|clause| goal.with(cx, clause));
            ecx.add_goals(GoalSource::ImplWhereBound, where_clause_bounds)?;

            // We currently elaborate all supertrait outlives obligations from impls.
            // This can be removed when we actually do coinduction correctly, and prove
            // all supertrait obligations unconditionally.
            ecx.add_goals(
                GoalSource::Misc,
                cx.impl_super_outlives(impl_def_id)
                    .iter_instantiated(cx, impl_args)
                    .map(Unnormalized::skip_norm_wip)
                    .map(|pred| goal.with(cx, pred)),
            )?;

            then(ecx)
        })
    }

    fn consider_error_guaranteed_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        _goal: Goal<I, Self>,
        _guar: I::ErrorGuaranteed,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    fn fast_reject_assumption(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        goal_lang_item: Option<SolverTraitLangItem>,
        assumption: I::Clause,
    ) -> Result<(), NoSolution> {
        fn trait_def_id_matches<I: Interner>(
            cx: I,
            clause_def_id: I::TraitId,
            goal_def_id: I::TraitId,
            polarity: ClausePolarity,
            goal_lang_item: Option<SolverTraitLangItem>,
        ) -> bool {
            clause_def_id == goal_def_id
            // PERF(sized-hierarchy): Sizedness supertraits aren't elaborated to improve perf, so
            // check for a `MetaSized` supertrait being matched against a `Sized` assumption.
            //
            // `PointeeSized` bounds are syntactic sugar for a lack of bounds so don't need this.
                || (polarity == ClausePolarity::Positive
                    && goal_lang_item == Some(SolverTraitLangItem::MetaSized)
                    && cx.is_trait_lang_item(clause_def_id, SolverTraitLangItem::Sized))
        }

        if let Some(trait_clause) = assumption.as_trait_clause()
            && trait_clause.polarity() == goal.predicate.polarity
            && trait_def_id_matches(
                ecx.cx(),
                trait_clause.def_id(),
                goal.predicate.def_id(),
                goal.predicate.polarity,
                goal_lang_item,
            )
            && DeepRejectCtxt::relate_rigid_rigid(ecx.cx()).args_may_unify(
                goal.predicate.trait_ref.args,
                trait_clause.skip_binder().trait_ref.args,
            )
        {
            return Ok(());
        } else {
            Err(NoSolution)
        }
    }

    fn match_assumption(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        assumption: I::Clause,
        then: impl FnOnce(&mut EvalCtxt<'_, D>) -> QueryResultOrRerunNonErased<I>,
    ) -> QueryResultOrRerunNonErased<I> {
        let trait_clause = assumption.as_trait_clause().unwrap();

        // PERF(sized-hierarchy): Sizedness supertraits aren't elaborated to improve perf, so
        // check for a `Sized` subtrait when looking for `MetaSized`. `PointeeSized` bounds
        // are syntactic sugar for a lack of bounds so don't need this.
        // We don't need to check polarity, `fast_reject_assumption` already rejected non-`Positive`
        // polarity `Sized` assumptions as matching non-`Positive` `MetaSized` goals.
        if ecx.cx().is_trait_lang_item(goal.predicate.def_id(), SolverTraitLangItem::MetaSized)
            && ecx.cx().is_trait_lang_item(trait_clause.def_id(), SolverTraitLangItem::Sized)
        {
            let meta_sized_clause =
                trait_predicate_with_def_id(ecx.cx(), trait_clause, goal.predicate.def_id());
            return Self::match_assumption(ecx, goal, meta_sized_clause, then);
        }

        let assumption_trait_pred = ecx.instantiate_binder_with_infer(trait_clause);
        ecx.eq(goal.param_env, goal.predicate.trait_ref, assumption_trait_pred.trait_ref)?;

        then(ecx)
    }

    fn consider_auto_trait_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        let cx = ecx.cx();
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        if let Some(result) = ecx.disqualify_auto_trait_candidate_due_to_possible_impl(goal) {
            return result;
        }

        // Only consider auto impls of unsafe traits when there are no unsafe
        // fields.
        if cx.trait_is_unsafe(goal.predicate.def_id())
            && goal.predicate.self_ty().has_unsafe_fields()
        {
            return Err(NoSolution.into());
        }

        // We leak the implemented auto traits of opaques outside of their defining scope.
        // This depends on `typeck` of the defining scope of that opaque, which may result in
        // fatal query cycles.
        //
        // We only get to this point if we're outside of the defining scope as we'd otherwise
        // be able to normalize the opaque type. We may also cycle in case `typeck` of a defining
        // scope relies on the current context, e.g. either because it also leaks auto trait
        // bounds of opaques defined in the current context or by evaluating the current item.
        //
        // To avoid this we don't try to leak auto trait bounds if they can also be proven via
        // item bounds of the opaque. These bounds are always applicable as auto traits must not
        // have any generic parameters. They would also get preferred over the impl candidate
        // when merging candidates anyways.
        //
        // See tests/ui/impl-trait/auto-trait-leakage/avoid-query-cycle-via-item-bound.rs.
        if let ty::Alias(is_rigid, ty::AliasTy { kind: ty::Opaque { def_id }, args, .. }) =
            goal.predicate.self_ty().kind()
        {
            debug_assert!(is_rigid == ty::IsRigid::Yes);
            return ecx.consider_auto_trait_candidate_for_opaque_ty(goal, def_id, args);
        }

        // We need to make sure to stall any coroutines we are inferring to avoid query cycles.
        if let Some(cand) = ecx.try_stall_coroutine(goal.predicate.self_ty()) {
            return cand;
        }

        ecx.probe_and_evaluate_goal_for_constituent_tys(
            CandidateSource::BuiltinImpl(BuiltinImplSource::Misc),
            goal,
            structural_traits::instantiate_constituent_tys_for_auto_trait,
        )
    }

    fn consider_trait_alias_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let cx = ecx.cx();

        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc).enter(|ecx| {
            let nested_obligations = cx
                .clauses_of(goal.predicate.def_id().into())
                .iter_instantiated(cx, goal.predicate.trait_ref.args)
                .map(Unnormalized::skip_norm_wip)
                .map(|c| goal.with(cx, c));
            // While you could think of trait aliases to have a single builtin impl
            // which uses its implied trait bounds as where-clauses, using
            // `GoalSource::ImplWhereClause` here would be incorrect, as we also
            // impl them, which means we're "stepping out of the impl constructor"
            // again. To handle this, we treat these cycles as ambiguous for now.
            ecx.add_goals(GoalSource::Misc, nested_obligations)?;
            ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
        })
    }

    fn consider_builtin_sizedness_candidates(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        sizedness: SizedTraitKind,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        ecx.probe_and_evaluate_goal_for_constituent_tys(
            CandidateSource::BuiltinImpl(BuiltinImplSource::Trivial),
            goal,
            |ecx, ty| {
                structural_traits::instantiate_constituent_tys_for_sizedness_trait(
                    ecx, sizedness, ty,
                )
            },
        )
    }

    fn consider_builtin_copy_clone_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        // We need to make sure to stall any coroutines we are inferring to avoid query cycles.
        if let Some(cand) = ecx.try_stall_coroutine(goal.predicate.self_ty()) {
            return cand;
        }

        ecx.probe_and_evaluate_goal_for_constituent_tys(
            CandidateSource::BuiltinImpl(BuiltinImplSource::Misc),
            goal,
            structural_traits::instantiate_constituent_tys_for_copy_clone_trait,
        )
    }

    fn consider_builtin_fn_ptr_trait_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        let self_ty = goal.predicate.self_ty();
        match goal.predicate.polarity {
            // impl FnPtr for FnPtr {}
            ty::ClausePolarity::Positive => {
                if self_ty.is_fn_ptr() {
                    ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc).enter(|ecx| {
                        ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
                    })
                } else {
                    Err(NoSolution.into())
                }
            }
            //  impl !FnPtr for T where T != FnPtr && T is rigid {}
            ty::ClausePolarity::Negative => {
                // If a type is rigid and not a fn ptr, then we know for certain
                // that it does *not* implement `FnPtr`.
                if !self_ty.is_fn_ptr() && self_ty.is_known_rigid() {
                    ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc).enter(|ecx| {
                        ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
                    })
                } else {
                    Err(NoSolution.into())
                }
            }
        }
    }

    fn consider_builtin_fn_trait_candidates(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        goal_kind: ty::ClosureKind,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let cx = ecx.cx();
        let Some(tupled_inputs_and_output) =
            structural_traits::extract_tupled_inputs_and_output_from_callable(
                cx,
                goal.predicate.self_ty(),
                goal_kind,
            )?
        else {
            return ecx.forced_ambiguity(MaybeInfo::AMBIGUOUS);
        };
        let (inputs, output) = ecx.instantiate_binder_with_infer(tupled_inputs_and_output);

        // A built-in `Fn` impl only holds if the output is sized.
        // (FIXME: technically we only need to check this if the type is a fn ptr...)
        let output_is_sized_pred =
            ty::TraitRef::new(cx, cx.require_trait_lang_item(SolverTraitLangItem::Sized), [output]);

        let pred =
            ty::TraitRef::new(cx, goal.predicate.def_id(), [goal.predicate.self_ty(), inputs])
                .upcast(cx);
        Self::probe_and_consider_implied_clause(
            ecx,
            CandidateSource::BuiltinImpl(BuiltinImplSource::Misc),
            goal,
            pred,
            [(GoalSource::ImplWhereBound, goal.with(cx, output_is_sized_pred))],
        )
    }

    fn consider_builtin_async_fn_trait_candidates(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        goal_kind: ty::ClosureKind,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let cx = ecx.cx();
        let (tupled_inputs_and_output_and_coroutine, nested_preds) =
            structural_traits::extract_tupled_inputs_and_output_from_async_callable(
                cx,
                goal.predicate.self_ty(),
                goal_kind,
                // This region doesn't matter because we're throwing away the coroutine type
                Region::new_static(cx),
            )?;
        let AsyncCallableRelevantTypes {
            tupled_inputs_ty,
            output_coroutine_ty,
            coroutine_return_ty: _,
        } = ecx.instantiate_binder_with_infer(tupled_inputs_and_output_and_coroutine);

        // A built-in `AsyncFn` impl only holds if the output is sized.
        // (FIXME: technically we only need to check this if the type is a fn ptr...)
        let output_is_sized_pred = ty::TraitRef::new(
            cx,
            cx.require_trait_lang_item(SolverTraitLangItem::Sized),
            [output_coroutine_ty],
        );

        let pred = ty::TraitRef::new(
            cx,
            goal.predicate.def_id(),
            [goal.predicate.self_ty(), tupled_inputs_ty],
        )
        .upcast(cx);
        Self::probe_and_consider_implied_clause(
            ecx,
            CandidateSource::BuiltinImpl(BuiltinImplSource::Misc),
            goal,
            pred,
            [goal.with(cx, output_is_sized_pred)]
                .into_iter()
                .chain(nested_preds.into_iter().map(|pred| goal.with(cx, pred)))
                .map(|goal| (GoalSource::ImplWhereBound, goal)),
        )
    }

    fn consider_builtin_async_fn_kind_helper_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        let [closure_fn_kind_ty, goal_kind_ty] = *goal.predicate.trait_ref.args.as_slice() else {
            panic!();
        };

        let Some(closure_kind) = closure_fn_kind_ty.expect_ty().to_opt_closure_kind() else {
            // We don't need to worry about the self type being an infer var.
            return Err(NoSolution.into());
        };
        let goal_kind = goal_kind_ty.expect_ty().to_opt_closure_kind().unwrap();
        if closure_kind.extends(goal_kind) {
            ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
                .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
        } else {
            Err(NoSolution.into())
        }
    }

    /// ```rust, ignore (not valid rust syntax)
    /// impl Tuple for () {}
    /// impl Tuple for (T1,) {}
    /// impl Tuple for (T1, T2) {}
    /// impl Tuple for (T1, .., Tn) {}
    /// ```
    fn consider_builtin_tuple_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        if let ty::Tuple(..) = goal.predicate.self_ty().kind() {
            ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
                .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
        } else {
            Err(NoSolution.into())
        }
    }

    fn consider_builtin_pointee_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    fn consider_builtin_future_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let ty::Coroutine(def_id, _) = goal.predicate.self_ty().kind() else {
            return Err(NoSolution.into());
        };

        // Coroutines are not futures unless they come from `async` desugaring
        let cx = ecx.cx();
        if !cx.coroutine_is_async(def_id) {
            return Err(NoSolution.into());
        }

        // Async coroutine unconditionally implement `Future`
        // Technically, we need to check that the future output type is Sized,
        // but that's already proven by the coroutine being WF.
        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    fn consider_builtin_iterator_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let ty::Coroutine(def_id, _) = goal.predicate.self_ty().kind() else {
            return Err(NoSolution.into());
        };

        // Coroutines are not iterators unless they come from `gen` desugaring
        let cx = ecx.cx();
        if !cx.coroutine_is_gen(def_id) {
            return Err(NoSolution.into());
        }

        // Gen coroutines unconditionally implement `Iterator`
        // Technically, we need to check that the iterator output type is Sized,
        // but that's already proven by the coroutines being WF.
        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    fn consider_builtin_fused_iterator_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let ty::Coroutine(def_id, _) = goal.predicate.self_ty().kind() else {
            return Err(NoSolution.into());
        };

        // Coroutines are not iterators unless they come from `gen` desugaring
        let cx = ecx.cx();
        if !cx.coroutine_is_gen(def_id) {
            return Err(NoSolution.into());
        }

        // Gen coroutines unconditionally implement `FusedIterator`.
        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    fn consider_builtin_async_iterator_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let ty::Coroutine(def_id, _) = goal.predicate.self_ty().kind() else {
            return Err(NoSolution.into());
        };

        // Coroutines are not iterators unless they come from `gen` desugaring
        let cx = ecx.cx();
        if !cx.coroutine_is_async_gen(def_id) {
            return Err(NoSolution.into());
        }

        // Gen coroutines unconditionally implement `Iterator`
        // Technically, we need to check that the iterator output type is Sized,
        // but that's already proven by the coroutines being WF.
        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    fn consider_builtin_coroutine_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let self_ty = goal.predicate.self_ty();
        let ty::Coroutine(def_id, args) = self_ty.kind() else {
            return Err(NoSolution.into());
        };

        // `async`-desugared coroutines do not implement the coroutine trait
        let cx = ecx.cx();
        if !cx.is_general_coroutine(def_id) {
            return Err(NoSolution.into());
        }

        let coroutine = args.as_coroutine();
        Self::probe_and_consider_implied_clause(
            ecx,
            CandidateSource::BuiltinImpl(BuiltinImplSource::Misc),
            goal,
            ty::TraitRef::new(cx, goal.predicate.def_id(), [self_ty, coroutine.resume_ty()])
                .upcast(cx),
            // Technically, we need to check that the coroutine types are Sized,
            // but that's already proven by the coroutine being WF.
            [],
        )
    }

    fn consider_builtin_discriminant_kind_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        // `DiscriminantKind` is automatically implemented for every type.
        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    fn consider_builtin_destruct_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        // `Destruct` is automatically implemented for every type in
        // non-const environments.
        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    fn consider_builtin_transmute_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        use rustc_type_ir::transmute::Answer;

        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        // `rustc_transmute` does not have support for type or const params
        if goal.predicate.has_non_region_placeholders() {
            return Err(NoSolution.into());
        }

        // Match the old solver by treating unresolved inference variables as
        // ambiguous until `rustc_transmute` can compute their layout.
        if goal.has_non_region_infer() {
            return ecx.forced_ambiguity(MaybeInfo::AMBIGUOUS);
        }

        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc).enter(
            |ecx| -> Result<_, NoSolutionOrRerunNonErased> {
                let assume = ecx.structurally_normalize_const(
                    goal.param_env,
                    goal.predicate.trait_ref.args.const_at(2),
                )?;

                let answer = ecx.is_transmutable(
                    goal.predicate.trait_ref.args.type_at(0),
                    goal.predicate.trait_ref.args.type_at(1),
                    assume,
                )?;

                let certainty = match answer {
                    Answer::No(_) => return Err(NoSolution.into()),
                    Answer::Yes => Certainty::Yes,
                    Answer::If(cond) => {
                        let tcx = ecx.cx();
                        ecx.add_goals(
                            GoalSource::Misc,
                            flatten_answer_tree(tcx, goal.predicate, cond)
                                .into_iter()
                                .map(|predicate| Goal::new(tcx, goal.param_env, predicate)),
                        )?;

                        Certainty::Yes
                    }
                };

                ecx.evaluate_added_goals_and_make_canonical_response(certainty)
            },
        )
    }

    /// NOTE: This is implemented as a built-in goal and not a set of impls like:
    ///
    /// ```rust,ignore (illustrative)
    /// impl<T> BikeshedGuaranteedNoDrop for T where T: Copy {}
    /// impl<T> BikeshedGuaranteedNoDrop for ManuallyDrop<T> {}
    /// ```
    ///
    /// because these impls overlap, and I'd rather not build a coherence hack for
    /// this harmless overlap.
    ///
    /// This trait is indirectly exposed on stable, so do *not* extend the set of types that
    /// implement the trait without FCP!
    fn consider_builtin_bikeshed_guaranteed_no_drop_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let cx = ecx.cx();
        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc).enter(|ecx| {
            let ty = goal.predicate.self_ty();
            match ty.kind() {
                // `&mut T` and `&T` always implement `BikeshedGuaranteedNoDrop`.
                ty::Ref(..) => {}
                // `ManuallyDrop<T>` always implements `BikeshedGuaranteedNoDrop`.
                ty::Adt(def, _) if def.is_manually_drop() => {}
                // Arrays and tuples implement `BikeshedGuaranteedNoDrop` only if
                // their constituent types implement `BikeshedGuaranteedNoDrop`.
                ty::Tuple(tys) => {
                    ecx.add_goals(
                        GoalSource::ImplWhereBound,
                        tys.iter().map(|elem_ty| {
                            goal.with(cx, ty::TraitRef::new(cx, goal.predicate.def_id(), [elem_ty]))
                        }),
                    )?;
                }
                ty::Array(elem_ty, _) => {
                    ecx.add_goal(
                        GoalSource::ImplWhereBound,
                        goal.with(cx, ty::TraitRef::new(cx, goal.predicate.def_id(), [elem_ty])),
                    )?;
                }

                // All other types implement `BikeshedGuaranteedNoDrop` only if
                // they implement `Copy`. We could be smart here and short-circuit
                // some trivially `Copy`/`!Copy` types, but there's no benefit.
                ty::FnDef(..)
                | ty::FnPtr(..)
                | ty::Error(_)
                | ty::Uint(_)
                | ty::Int(_)
                | ty::Infer(ty::IntVar(_) | ty::FloatVar(_))
                | ty::Bool
                | ty::Float(_)
                | ty::Char
                | ty::RawPtr(..)
                | ty::Never
                | ty::Pat(..)
                | ty::Dynamic(..)
                | ty::Str
                | ty::Slice(_)
                | ty::Foreign(..)
                | ty::Adt(..)
                | ty::Alias(..)
                | ty::Param(_)
                | ty::Placeholder(..)
                | ty::Closure(..)
                | ty::CoroutineClosure(..)
                | ty::Coroutine(..)
                | ty::UnsafeBinder(_)
                | ty::CoroutineWitness(..) => {
                    ecx.add_goal(
                        GoalSource::ImplWhereBound,
                        goal.with(
                            cx,
                            ty::TraitRef::new(
                                cx,
                                cx.require_trait_lang_item(SolverTraitLangItem::Copy),
                                [ty],
                            ),
                        ),
                    )?;
                }

                ty::Bound(..)
                | ty::Infer(
                    ty::TyVar(_) | ty::FreshTy(_) | ty::FreshIntTy(_) | ty::FreshFloatTy(_),
                ) => {
                    panic!("unexpected type `{ty:?}`")
                }
            }

            ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
        })
    }

    /// ```ignore (builtin impl example)
    /// trait Trait {
    ///     fn foo(&self);
    /// }
    /// // results in the following builtin impl
    /// impl<'a, T: Trait + 'a> Unsize<dyn Trait + 'a> for T {}
    /// ```
    fn consider_structural_builtin_unsize_candidates(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Vec<Candidate<I>>, RerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Ok(vec![]);
        }

        let result = ecx.probe(|_| ProbeKind::UnsizeAssembly).enter(
            |ecx| -> Result<Vec<Candidate<I>>, NoSolutionOrRerunNonErased> {
                let a_ty = goal.predicate.self_ty();
                // We need to normalize the b_ty since it's matched structurally
                // in the other functions below.
                let b_ty = ecx.structurally_normalize_ty(
                    goal.param_env,
                    goal.predicate.trait_ref.args.type_at(1),
                )?;

                let goal = goal.with(ecx.cx(), (a_ty, b_ty));
                match (a_ty.kind(), b_ty.kind()) {
                    (ty::Infer(ty::TyVar(..)), ..) => panic!("unexpected infer {a_ty:?} {b_ty:?}"),

                    (_, ty::Infer(ty::TyVar(..))) => {
                        Ok(vec![ecx.forced_ambiguity(MaybeInfo::AMBIGUOUS)?])
                    }

                    // Trait upcasting, or `dyn Trait + Auto + 'a` -> `dyn Trait + 'b`.
                    (ty::Dynamic(a_data, a_region), ty::Dynamic(b_data, b_region)) => Ok(ecx
                        .consider_builtin_dyn_upcast_candidates(
                            goal, a_data, a_region, b_data, b_region,
                        )),

                    // `T` -> `dyn Trait` unsizing.
                    (_, ty::Dynamic(b_region, b_data)) => Ok(vec![
                        ecx.consider_builtin_unsize_to_dyn_candidate(goal, b_region, b_data)?,
                    ]),

                    // `[T; N]` -> `[T]` unsizing
                    (ty::Array(a_elem_ty, ..), ty::Slice(b_elem_ty)) => {
                        Ok(vec![ecx.consider_builtin_array_unsize(goal, a_elem_ty, b_elem_ty)?])
                    }

                    // `Struct<T>` -> `Struct<U>` where `T: Unsize<U>`
                    (ty::Adt(a_def, a_args), ty::Adt(b_def, b_args))
                        if a_def.is_struct() && a_def == b_def =>
                    {
                        Ok(vec![ecx.consider_builtin_struct_unsize(goal, a_def, a_args, b_args)?])
                    }

                    _ => Err(NoSolution.into()),
                }
            },
        );

        match result.map_err_to_rerun()? {
            Ok(resp) => Ok(resp),
            Err(NoSolution) => Ok(vec![]),
        }
    }

    fn consider_builtin_try_as_dyn_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }
        let cx = ecx.cx();

        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc).enter(|ecx| {
            let self_ty = goal.predicate.self_ty();
            let ty_lifetime = goal.predicate.trait_ref.args.region_at(1);
            match self_ty.kind() {
                ty::Dynamic(bounds, lifetime) => {
                    for bound in bounds.iter() {
                        match bound.skip_binder() {
                            ExistentialPredicate::Trait(_) => {}
                            // FIXME(try_as_dyn): check what kind of projections we can allow
                            ExistentialPredicate::Projection(_) => return Err(NoSolution.into()),
                            // Auto traits do not affect lifetimes outside of specialization,
                            // which is disabled in reflection.
                            ExistentialPredicate::AutoTrait(_) => {}
                        }
                    }
                    ecx.add_goal(
                        GoalSource::Misc,
                        goal.with(cx, ty::OutlivesClause(ty_lifetime, lifetime)),
                    )?;
                    ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
                }

                ty::Bound(..)
                | ty::Infer(
                    ty::TyVar(_) | ty::FreshTy(_) | ty::FreshIntTy(_) | ty::FreshFloatTy(_),
                ) => {
                    panic!("unexpected type `{self_ty:?}`")
                }

                _ => Err(NoSolution.into()),
            }
        })
    }

    fn consider_builtin_field_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        if goal.predicate.polarity != ty::ClausePolarity::Positive {
            return Err(NoSolution.into());
        }

        let ty::Adt(def, args) = goal.predicate.self_ty().kind() else {
            return Err(NoSolution.into());
        };

        let Some(FieldInfo { base, ty, .. }) = def.field_representing_type_info(ecx.cx(), args)
        else {
            return Err(NoSolution.into());
        };

        match base.kind() {
            ty::Adt(def, _) if def.is_struct() && !def.is_packed() => {}
            ty::Adt(..) => return Err(NoSolution.into()),
            ty::Tuple(..) => {}
            _ => unreachable!(),
        }

        ecx.probe_builtin_trait_candidate(BuiltinImplSource::Misc).enter(|ecx| {
            let sized_trait = ecx.cx().require_trait_lang_item(SolverTraitLangItem::Sized);

            // FIXME: add better support for builtin impls of traits that check for the bounds
            // on the trait definition in std.
            //
            // NOTE: these bounds have to be kept in sync with the definition of the `Field`
            // trait in `library/core/src/field.rs` as well as the old trait solver `fn
            // assemble_candidates_for_field_trait` in
            // `compiler/rustc_trait_selection/src/traits/select/candidate_assembly.rs`.
            ecx.add_goal(
                GoalSource::ImplWhereBound,
                Goal {
                    param_env: goal.param_env,
                    predicate: TraitRef::new(ecx.cx(), sized_trait, [base]).upcast(ecx.cx()),
                },
            )?;
            ecx.add_goal(
                GoalSource::ImplWhereBound,
                Goal {
                    param_env: goal.param_env,
                    predicate: TraitRef::new(ecx.cx(), sized_trait, [ty]).upcast(ecx.cx()),
                },
            )?;

            ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
        })
    }
}

/// Small helper function to change the `def_id` of a trait predicate - this is not normally
/// something that you want to do, as different traits will require different args and so making
/// it easy to change the trait is something of a footgun, but it is useful in the narrow
/// circumstance of changing from `MetaSized` to `Sized`, which happens as part of the lazy
/// elaboration of sizedness candidates.
#[inline(always)]
fn trait_predicate_with_def_id<I: Interner>(
    cx: I,
    clause: ty::Binder<I, ty::TraitClause<I>>,
    did: I::TraitId,
) -> I::Clause {
    clause
        .map_bound(|c| TraitClause {
            trait_ref: TraitRef::new_from_args(cx, did, c.trait_ref.args),
            polarity: c.polarity,
        })
        .upcast(cx)
}

impl<D, I> EvalCtxt<'_, D>
where
    D: SolverDelegate<Interner = I>,
    I: Interner,
{
    /// Trait upcasting allows for coercions between trait objects:
    /// ```ignore (builtin impl example)
    /// trait Super {}
    /// trait Trait: Super {}
    /// // results in builtin impls upcasting to a super trait
    /// impl<'a, 'b: 'a> Unsize<dyn Super + 'a> for dyn Trait + 'b {}
    /// // and impls removing auto trait bounds.
    /// impl<'a, 'b: 'a> Unsize<dyn Trait + 'a> for dyn Trait + Send + 'b {}
    /// ```
    fn consider_builtin_dyn_upcast_candidates(
        &mut self,
        goal: Goal<I, (I::Ty, I::Ty)>,
        a_data: I::BoundExistentialPredicates,
        a_region: Region<I>,
        b_data: I::BoundExistentialPredicates,
        b_region: Region<I>,
    ) -> Vec<Candidate<I>> {
        let cx = self.cx();
        let Goal { predicate: (a_ty, _b_ty), .. } = goal;

        let mut responses = vec![];
        // If the principal def ids match (or are both none), then we're not doing
        // trait upcasting. We're just removing auto traits (or shortening the lifetime).
        let b_principal_def_id = b_data.principal_def_id();
        if a_data.principal_def_id() == b_principal_def_id || b_principal_def_id.is_none() {
            responses.extend(self.consider_builtin_upcast_to_principal(
                goal,
                CandidateSource::BuiltinImpl(BuiltinImplSource::Misc),
                a_data,
                a_region,
                b_data,
                b_region,
                a_data.principal(),
            ));
        } else if let Some(a_principal) = a_data.principal() {
            for (idx, new_a_principal) in
                elaborate::supertraits(self.cx(), a_principal.with_self_ty(cx, a_ty))
                    .enumerate()
                    .skip(1)
            {
                responses.extend(self.consider_builtin_upcast_to_principal(
                    goal,
                    CandidateSource::BuiltinImpl(BuiltinImplSource::TraitUpcasting(idx)),
                    a_data,
                    a_region,
                    b_data,
                    b_region,
                    Some(new_a_principal.map_bound(|trait_ref| {
                        ty::ExistentialTraitRef::erase_self_ty(cx, trait_ref)
                    })),
                ));
            }
        }

        responses
    }

    fn consider_builtin_unsize_to_dyn_candidate(
        &mut self,
        goal: Goal<I, (I::Ty, I::Ty)>,
        b_data: I::BoundExistentialPredicates,
        b_region: Region<I>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        let cx = self.cx();
        let Goal { predicate: (a_ty, _), .. } = goal;

        // Can only unsize to an dyn-compatible trait.
        if b_data.principal_def_id().is_some_and(|def_id| !cx.trait_is_dyn_compatible(def_id)) {
            return Err(NoSolution.into());
        }

        self.probe_builtin_trait_candidate(BuiltinImplSource::Misc).enter(|ecx| {
            // Check that the type implements all of the predicates of the trait object.
            // (i.e. the principal, all of the associated types match, and any auto traits)
            ecx.add_goals(
                GoalSource::ImplWhereBound,
                b_data.iter().map(|pred| goal.with(cx, pred.with_self_ty(cx, a_ty))),
            )?;

            // The type must be `Sized` to be unsized.
            ecx.add_goal(
                GoalSource::ImplWhereBound,
                goal.with(
                    cx,
                    ty::TraitRef::new(
                        cx,
                        cx.require_trait_lang_item(SolverTraitLangItem::Sized),
                        [a_ty],
                    ),
                ),
            )?;

            // The type must outlive the lifetime of the `dyn` we're unsizing into.
            ecx.add_goal(GoalSource::Misc, goal.with(cx, ty::OutlivesClause(a_ty, b_region)))?;
            ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
        })
    }

    fn consider_builtin_upcast_to_principal(
        &mut self,
        goal: Goal<I, (I::Ty, I::Ty)>,
        source: CandidateSource<I>,
        a_data: I::BoundExistentialPredicates,
        a_region: Region<I>,
        b_data: I::BoundExistentialPredicates,
        b_region: Region<I>,
        upcast_principal: Option<ty::Binder<I, ty::ExistentialTraitRef<I>>>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        let param_env = goal.param_env;

        // We may upcast to auto traits that are either explicitly listed in
        // the object type's bounds, or implied by the principal trait ref's
        // supertraits.
        let a_auto_traits: IndexSet<I::TraitId> = a_data
            .auto_traits()
            .into_iter()
            .chain(a_data.principal_def_id().into_iter().flat_map(|principal_def_id| {
                self.cx()
                    .supertrait_def_ids(principal_def_id)
                    .filter(|def_id| self.cx().trait_is_auto(*def_id))
            }))
            .collect();

        // More than one projection in a_ty's bounds may match the projection
        // in b_ty's bound. Use this to first determine *which* apply without
        // having any inference side-effects. We process obligations because
        // unification may initially succeed due to deferred projection equality.
        let projection_may_match =
            |ecx: &mut EvalCtxt<'_, D>,
             source_projection: ty::Binder<I, ty::ExistentialProjection<I>>,
             target_projection: ty::Binder<I, ty::ExistentialProjection<I>>|
             -> Result<bool, RerunNonErased> {
                if source_projection.item_def_id() != target_projection.item_def_id() {
                    return Ok(false);
                }
                match ecx.probe(|_| ProbeKind::ProjectionCompatibility).enter(|ecx| {
                    let target_projection = ecx.deeply_resolve_ignoring_regions(target_projection);
                    ecx.enter_forall_with_assumptions(
                        target_projection,
                        param_env,
                        |ecx, target_projection| {
                            let source_projection =
                                ecx.instantiate_binder_with_infer(source_projection);
                            ecx.eq(param_env, source_projection, target_projection)?;
                            ecx.try_evaluate_added_goals()
                        },
                    )
                }) {
                    Ok(_) => Ok(true),
                    Err(NoSolutionOrRerunNonErased::NoSolution(_)) => Ok(false),
                    Err(NoSolutionOrRerunNonErased::RerunNonErased(rerun)) => Err(rerun),
                }
            };

        self.probe_trait_candidate(source).enter(|ecx| {
            for bound in b_data.iter() {
                match bound.skip_binder() {
                    // Check that a's supertrait (upcast_principal) is compatible
                    // with the target (b_ty).
                    ty::ExistentialPredicate::Trait(target_principal) => {
                        let source_principal = upcast_principal.unwrap();
                        let target_principal = bound.rebind(target_principal);
                        // We might unify infer vars in previous iterations.
                        let target_principal =
                            ecx.deeply_resolve_ignoring_regions(target_principal);
                        ecx.enter_forall_with_assumptions(
                            target_principal,
                            param_env,
                            |ecx, target_principal| {
                                let source_principal =
                                    ecx.instantiate_binder_with_infer(source_principal);
                                ecx.eq(param_env, source_principal, target_principal)?;
                                ecx.try_evaluate_added_goals()
                            },
                        )?;
                    }
                    // Check that b_ty's projection is satisfied by exactly one of
                    // a_ty's projections. First, we look through the list to see if
                    // any match. If not, error. Then, if *more* than one matches, we
                    // return ambiguity. Otherwise, if exactly one matches, equate
                    // it with b_ty's projection.
                    ty::ExistentialPredicate::Projection(target_projection) => {
                        let target_projection = bound.rebind(target_projection);
                        let mut matching_projection = None;
                        for source_projection in a_data.projection_bounds() {
                            if projection_may_match(ecx, source_projection, target_projection)? {
                                if matching_projection.is_some() {
                                    return ecx.evaluate_added_goals_and_make_canonical_response(
                                        Certainty::AMBIGUOUS,
                                    );
                                }
                                matching_projection = Some(source_projection);
                            }
                        }
                        let Some(matching) = matching_projection else {
                            return Err(NoSolution.into());
                        };

                        // We might unify infer vars in previous iterations.
                        let target_projection =
                            ecx.deeply_resolve_ignoring_regions(target_projection);
                        ecx.enter_forall_with_assumptions(
                            target_projection,
                            param_env,
                            |ecx, target_projection| {
                                let source_projection = ecx.instantiate_binder_with_infer(matching);
                                ecx.eq(param_env, source_projection, target_projection)?;
                                ecx.try_evaluate_added_goals()
                            },
                        )?;
                    }
                    // Check that b_ty's auto traits are present in a_ty's bounds.
                    ty::ExistentialPredicate::AutoTrait(def_id) => {
                        if !a_auto_traits.contains(&def_id) {
                            return Err(NoSolution.into());
                        }
                    }
                }
            }

            // Also require that a_ty's lifetime outlives b_ty's lifetime.
            ecx.add_goal(
                GoalSource::ImplWhereBound,
                Goal::new(ecx.cx(), param_env, ty::OutlivesClause(a_region, b_region)),
            )?;

            ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
        })
    }

    /// We have the following builtin impls for arrays:
    /// ```ignore (builtin impl example)
    /// impl<T: ?Sized, const N: usize> Unsize<[T]> for [T; N] {}
    /// ```
    /// While the impl itself could theoretically not be builtin,
    /// the actual unsizing behavior is builtin. Its also easier to
    /// make all impls of `Unsize` builtin as we're able to use
    /// `#[rustc_deny_explicit_impl]` in this case.
    fn consider_builtin_array_unsize(
        &mut self,
        goal: Goal<I, (I::Ty, I::Ty)>,
        a_elem_ty: I::Ty,
        b_elem_ty: I::Ty,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        self.eq(goal.param_env, a_elem_ty, b_elem_ty)?;
        self.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    /// We generate a builtin `Unsize` impls for structs with generic parameters only
    /// mentioned by the last field.
    /// ```ignore (builtin impl example)
    /// struct Foo<T, U: ?Sized> {
    ///     sized_field: Vec<T>,
    ///     unsizable: Box<U>,
    /// }
    /// // results in the following builtin impl
    /// impl<T: ?Sized, U: ?Sized, V: ?Sized> Unsize<Foo<T, V>> for Foo<T, U>
    /// where
    ///     Box<U>: Unsize<Box<V>>,
    /// {}
    /// ```
    fn consider_builtin_struct_unsize(
        &mut self,
        goal: Goal<I, (I::Ty, I::Ty)>,
        def: I::AdtDef,
        a_args: I::GenericArgs,
        b_args: I::GenericArgs,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        let cx = self.cx();
        let Goal { predicate: (_a_ty, b_ty), .. } = goal;

        let unsizing_params = cx.unsizing_params_for_adt(def.def_id());
        // We must be unsizing some type parameters. This also implies
        // that the struct has a tail field.
        if unsizing_params.is_empty() {
            return Err(NoSolution.into());
        }

        let tail_field_ty = def.struct_tail_ty(cx).unwrap();

        let a_tail_ty = tail_field_ty.instantiate(cx, a_args).skip_norm_wip();
        let b_tail_ty = tail_field_ty.instantiate(cx, b_args).skip_norm_wip();

        // Instantiate just the unsizing params from B into A. The type after
        // this instantiation must be equal to B. This is so we don't unsize
        // unrelated type parameters.
        let new_a_args = cx.mk_args_from_iter(a_args.iter().enumerate().map(|(i, a)| {
            if unsizing_params.contains(i as u32) { b_args.get(i).unwrap() } else { a }
        }));
        let unsized_a_ty = Ty::new_adt(cx, def, new_a_args);

        // Finally, we require that `TailA: Unsize<TailB>` for the tail field
        // types.
        self.eq(goal.param_env, unsized_a_ty, b_ty)?;
        self.add_goal(
            GoalSource::ImplWhereBound,
            goal.with(
                cx,
                ty::TraitRef::new(
                    cx,
                    cx.require_trait_lang_item(SolverTraitLangItem::Unsize),
                    [a_tail_ty, b_tail_ty],
                ),
            ),
        )?;
        self.probe_builtin_trait_candidate(BuiltinImplSource::Misc)
            .enter(|ecx| ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes))
    }

    fn consider_auto_trait_candidate_for_opaque_ty(
        &mut self,
        goal: Goal<I, TraitClause<I>>,
        def_id: I::OpaqueTyId,
        args: I::GenericArgs,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        let cx = self.cx();
        let source = CandidateSource::BuiltinImpl(BuiltinImplSource::Misc);

        for item_bound in cx.item_self_bounds(def_id.into()).skip_binder() {
            if item_bound.as_trait_clause().is_some_and(|b| b.def_id() == goal.predicate.def_id()) {
                return Err(NoSolution.into());
            }
        }

        let candidate = self.probe_trait_candidate(source).enter(|ecx| {
            let hidden_ty = cx.type_of(def_id.into()).instantiate(cx, args).skip_norm_wip();
            ecx.add_goal(
                GoalSource::ImplWhereBound,
                goal.with(cx, goal.predicate.with_replaced_self_ty(cx, hidden_ty)),
            )?;
            ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
        })?;

        // Proving an auto trait for the hidden type must not constrain inference
        // variables, as that would leak the hidden type itself.
        if !candidate.result.value.is_identity_modulo_regions() {
            return self.forced_ambiguity(MaybeInfo::AMBIGUOUS);
        }

        let ExternalConstraintsData {
            region_constraints: _,
            ref opaque_types,
            ref normalization_nested_goals,
        } = *candidate.result.value.external_constraints;
        debug_assert!(normalization_nested_goals.is_empty());

        // New defining uses may leak an opaque's hidden type into the caller's
        // inference state. This is safe after typeck, where hidden types are already
        // fixed and only regions are inferred, but not during typeck while hidden
        // types may still contain inference variables.
        if !opaque_types.is_empty() {
            let typing_mode = self.typing_mode();

            match typing_mode {
                // We're inferring regions of opaque types, but
                // the type itself is already fully known, no way
                // to leak the hidden type.
                TypingMode::PostTypeckUntilBorrowck { .. } => {}
                // We're inferring the hidden type of opaques, could
                // leak types through it.
                TypingMode::Typeck { .. } => {
                    return self.forced_ambiguity(MaybeInfo::AMBIGUOUS);
                }
                // we never add new uses to the opaque type storage
                TypingMode::Coherence
                | TypingMode::PostBorrowck { .. }
                | TypingMode::Reflection
                | TypingMode::PostAnalysis
                | TypingMode::Codegen
                | TypingMode::ErasedNotCoherence(MayBeErased) => {
                    unreachable!(
                        "we never add new uses to opaque types in typing mode {typing_mode:?}"
                    );
                }
            }
        }

        Ok(candidate)
    }

    // Return `Some` if there is an impl (built-in or user provided) that may
    // hold for the self type of the goal, which for coherence and soundness
    // purposes must disqualify the built-in auto impl assembled by considering
    // the type's constituent types.
    fn disqualify_auto_trait_candidate_due_to_possible_impl(
        &mut self,
        goal: Goal<I, TraitClause<I>>,
    ) -> Option<Result<Candidate<I>, NoSolutionOrRerunNonErased>> {
        let self_ty = goal.predicate.self_ty();
        let check_impls = || {
            let mut disqualifying_impl = None;
            self.cx().for_each_relevant_impl(goal.predicate.trait_ref, |impl_def_id| {
                disqualifying_impl = Some(impl_def_id);
            });
            if let Some(def_id) = disqualifying_impl {
                trace!(?def_id, ?goal, "disqualified auto-trait implementation");
                // No need to actually consider the candidate here,
                // since we do that in `consider_impl_candidate`.
                return Some(Err(NoSolution.into()));
            } else {
                None
            }
        };

        match self_ty.kind() {
            // Stall int and float vars until they are resolved to a concrete
            // numerical type. That's because the check for impls below treats
            // int vars as matching any impl. Even if we filtered such impls,
            // we probably don't want to treat an `impl !AutoTrait for i32` as
            // disqualifying the built-in auto impl for `i64: AutoTrait` either.
            ty::Infer(ty::IntVar(_) | ty::FloatVar(_)) => {
                Some(self.forced_ambiguity(MaybeInfo::AMBIGUOUS))
            }

            // Backward compatibility for default auto traits.
            // Test: ui/traits/default_auto_traits/extern-types.rs
            ty::Foreign(..) if self.cx().is_default_trait(goal.predicate.def_id()) => check_impls(),

            // These types cannot be structurally decomposed into constituent
            // types, and therefore have no built-in auto impl.
            ty::Dynamic(..)
            | ty::Param(..)
            | ty::Foreign(..)
            | ty::Alias(
                ty::IsRigid::Yes,
                ty::AliasTy {
                    kind: ty::Projection { .. } | ty::Free { .. } | ty::Inherent { .. },
                    ..
                },
            )
            | ty::Placeholder(..) => Some(Err(NoSolution.into())),

            // Coroutines have one special built-in candidate, `Unpin`, which
            // takes precedence over the structural auto trait candidate being
            // assembled.
            ty::Coroutine(def_id, _)
                if self
                    .cx()
                    .is_trait_lang_item(goal.predicate.def_id(), SolverTraitLangItem::Unpin) =>
            {
                match self.cx().coroutine_movability(def_id) {
                    Movability::Static => Some(Err(NoSolution.into())),
                    Movability::Movable => Some(
                        self.probe_builtin_trait_candidate(BuiltinImplSource::Misc).enter(|ecx| {
                            ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
                        }),
                    ),
                }
            }

            // If we still have an alias here, it must be rigid. For opaques, it's always
            // okay to consider auto traits because that'll reveal its hidden type. For
            // non-opaque aliases, we will not assemble any candidates since there's no way
            // to further look into its type.
            ty::Alias(ty::IsRigid::Yes, ty::AliasTy { kind: ty::Opaque { .. }, .. }) => None,

            // For rigid types, any possible implementation that could apply to
            // the type (even if after unification and processing nested goals
            // it does not hold) will disqualify the built-in auto impl.
            //
            // We've originally had a more permissive check here which resulted
            // in unsoundness, see #84857.
            ty::Bool
            | ty::Char
            | ty::Int(_)
            | ty::Uint(_)
            | ty::Float(_)
            | ty::Str
            | ty::Array(_, _)
            | ty::Pat(_, _)
            | ty::Slice(_)
            | ty::RawPtr(_, _)
            | ty::Ref(_, _, _)
            | ty::FnDef(_, _)
            | ty::FnPtr(..)
            | ty::Closure(..)
            | ty::CoroutineClosure(..)
            | ty::Coroutine(_, _)
            | ty::CoroutineWitness(..)
            | ty::Never
            | ty::Tuple(_)
            | ty::Adt(_, _)
            | ty::UnsafeBinder(_) => check_impls(),
            ty::Error(_) => None,

            ty::Infer(_) | ty::Alias(ty::IsRigid::No, _) | ty::Bound(_, _) => {
                panic!("unexpected type `{self_ty:?}`")
            }
        }
    }

    /// Convenience function for traits that are structural, i.e. that only
    /// have nested subgoals that only change the self type. Unlike other
    /// evaluate-like helpers, this does a probe, so it doesn't need to be
    /// wrapped in one.
    fn probe_and_evaluate_goal_for_constituent_tys(
        &mut self,
        source: CandidateSource<I>,
        goal: Goal<I, TraitClause<I>>,
        constituent_tys: impl Fn(
            &EvalCtxt<'_, D>,
            I::Ty,
        ) -> Result<ty::Binder<I, Vec<I::Ty>>, NoSolution>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        self.probe_trait_candidate(source).enter(|ecx| {
            let goals = ecx.enter_forall_with_assumptions(
                constituent_tys(ecx, goal.predicate.self_ty())?,
                goal.param_env,
                |ecx, tys| {
                    tys.into_iter()
                        .map(|ty| {
                            goal.with(ecx.cx(), goal.predicate.with_replaced_self_ty(ecx.cx(), ty))
                        })
                        .collect::<Vec<_>>()
                },
            );
            ecx.add_goals(GoalSource::ImplWhereBound, goals)?;
            ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
        })
    }
}

/// How we've proven this trait goal.
///
/// This is used by `NormalizesTo` goals to only normalize
/// by using the same 'kind of candidate' we've used to prove
/// its corresponding trait goal. Most notably, we do not
/// normalize by using an impl if the trait goal has been
/// proven via a `ParamEnv` candidate.
///
/// This is necessary to avoid unnecessary region constraints,
/// see trait-system-refactor-initiative#125 for more details.
#[derive(Debug, Clone, Copy)]
pub(super) enum TraitGoalProvenVia {
    /// We've proven the trait goal by something which is
    /// is not a non-global where-bound or an alias-bound.
    ///
    /// This means we don't disable any candidates during
    /// normalization.
    Misc,
    ParamEnv,
    AliasBound,
}

impl<D, I> EvalCtxt<'_, D>
where
    D: SolverDelegate<Interner = I>,
    I: Interner,
{
    /// FIXME(#57893): For backwards compatibility with the old trait solver implementation,
    /// we need to handle overlap between builtin and user-written impls for trait objects.
    ///
    /// This overlap is unsound in general and something which we intend to fix separately.
    /// To avoid blocking the stabilization of the trait solver, we add this hack to avoid
    /// breakage in cases which are *mostly fine*™. Importantly, this preference is strictly
    /// weaker than the old behavior.
    ///
    /// We only prefer builtin over user-written impls if there are no inference constraints.
    /// Importantly, we also only prefer the builtin impls for trait goals, and not during
    /// normalization. This means the only case where this special-case results in exploitable
    /// unsoundness should be lifetime dependent user-written impls.
    pub(super) fn unsound_prefer_builtin_dyn_impl(&mut self, candidates: &mut Vec<Candidate<I>>) {
        if self.typing_mode().is_coherence() {
            return;
        }

        if candidates
            .iter()
            .find(|c| {
                matches!(c.source, CandidateSource::BuiltinImpl(BuiltinImplSource::Object(_)))
            })
            .is_some_and(|c| has_only_region_constraints(c.result))
        {
            candidates.retain(|c| {
                if matches!(c.source, CandidateSource::Impl(_)) {
                    debug!(?c, "unsoundly dropping impl in favor of builtin dyn-candidate");
                    false
                } else {
                    true
                }
            });
        }
    }

    /// Like `try_merge_candidates`, but returns `None` if there are multiple user-written
    /// impls, or a user-written impl and a builtin impl, in which case we bail with ambiguity.
    ///
    /// Candidates with equal responses can still come from conflicting impls, ex - two
    /// overlapping impls which both apply due to impossible where-clauses. We don't want
    /// to merge these. Multiple builtin impls are still merged as usual. This is not
    /// done for marker traits, as their impls are allowed to overlap.
    #[instrument(level = "trace", skip(self), ret)]
    fn try_merge_impl_candidates(
        &mut self,
        trait_def_id: I::TraitId,
        candidates: &[Candidate<I>],
    ) -> Option<CanonicalResponse<I>> {
        let is_marker = self.cx().trait_is_marker(trait_def_id);
        let all_builtin = candidates
            .iter()
            .all(|candidate| matches!(candidate.source, CandidateSource::BuiltinImpl(_)));

        if is_marker || all_builtin {
            self.try_merge_candidates(candidates).map(|(response, _)| response)
        } else if candidates.len() > 1 {
            None
        } else {
            candidates.first().map(|candidate| candidate.result)
        }
    }

    fn merge_candidates_or_bail_with_ambiguity(
        &mut self,
        candidates: &[Candidate<I>],
        proven_via: TraitGoalProvenVia,
    ) -> (CanonicalResponse<I>, Option<TraitGoalProvenVia>) {
        if let Some((response, _)) = self.try_merge_candidates(candidates) {
            (response, Some(proven_via))
        } else {
            (self.bail_with_ambiguity(candidates), None)
        }
    }

    fn merge_impl_candidates_or_bail_with_ambiguity(
        &mut self,
        trait_def_id: I::TraitId,
        candidates: &[Candidate<I>],
    ) -> (CanonicalResponse<I>, Option<TraitGoalProvenVia>) {
        debug_assert!(candidates.iter().all(|c| matches!(
            c.source,
            CandidateSource::Impl(_) | CandidateSource::BuiltinImpl(_)
        )));
        match self.try_merge_impl_candidates(trait_def_id, candidates) {
            Some(response) => (response, Some(TraitGoalProvenVia::Misc)),
            None => (self.bail_with_ambiguity(candidates), None),
        }
    }

    #[instrument(level = "debug", skip(self), ret)]
    pub(super) fn merge_trait_candidates(
        &mut self,
        trait_def_id: I::TraitId,
        mut candidates: Vec<Candidate<I>>,
        failed_candidate_info: FailedCandidateInfo,
    ) -> Result<(CanonicalResponse<I>, Option<TraitGoalProvenVia>), NoSolution> {
        if candidates.is_empty() {
            return Err(NoSolution);
        }

        let candidate_preference_mode = CandidatePreferenceMode::compute(self.cx(), trait_def_id);
        if self.typing_mode().is_coherence() {
            return Ok(match self.try_merge_impl_candidates(trait_def_id, &candidates) {
                Some(response) => (response, Some(TraitGoalProvenVia::Misc)),
                None => (self.bail_with_ambiguity(&candidates), None),
            });
        }

        // We prefer trivial builtin candidates, i.e. builtin impls without any
        // nested requirements, over all others. This is a fix for #53123 and
        // prevents where-bounds from accidentally extending the lifetime of a
        // variable.
        let mut trivial_builtin_impls = candidates.iter().filter(|c| {
            matches!(c.source, CandidateSource::BuiltinImpl(BuiltinImplSource::Trivial))
        });
        if let Some(candidate) = trivial_builtin_impls.next() {
            // There should only ever be a single trivial builtin candidate
            // as they would otherwise overlap.
            assert!(trivial_builtin_impls.next().is_none());
            return Ok((candidate.result, Some(TraitGoalProvenVia::Misc)));
        }

        // Extract non-nested alias bound candidates, will be preferred over where bounds if
        // we're proving an auto-trait, sizedness trait or default trait.
        if matches!(candidate_preference_mode, CandidatePreferenceMode::Marker)
            && candidates.iter().any(|c| {
                matches!(c.source, CandidateSource::AliasBound(AliasBoundKind::SelfBounds))
            })
        {
            let alias_bounds: Vec<_> = candidates
                .extract_if(.., |c| matches!(c.source, CandidateSource::AliasBound(..)))
                .collect();
            return Ok(self.merge_candidates_or_bail_with_ambiguity(
                &alias_bounds,
                TraitGoalProvenVia::AliasBound,
            ));
        }

        // If there are non-global where-bounds, prefer where-bounds
        // (including global ones) over everything else.
        let has_non_global_where_bounds = candidates
            .iter()
            .any(|c| matches!(c.source, CandidateSource::ParamEnv(ParamEnvSource::NonGlobal)));
        if has_non_global_where_bounds {
            let where_bounds: Vec<_> = candidates
                .extract_if(.., |c| matches!(c.source, CandidateSource::ParamEnv(_)))
                .collect();
            let Some((response, info)) = self.try_merge_candidates(&where_bounds) else {
                return Ok((self.bail_with_ambiguity(&where_bounds), None));
            };
            match info {
                // If there's an always applicable candidate, the result of all
                // other candidates does not matter. This means we can ignore
                // them when checking whether we've reached a fixpoint.
                //
                // We always prefer the first always applicable candidate, even if a
                // later candidate is also always applicable and would result in fewer
                // reruns. We could slightly improve this by e.g. searching for another
                // always applicable candidate which doesn't depend on any cycle heads.
                //
                // NOTE: This is optimization is observable in case there is an always
                // applicable global candidate and another non-global candidate which only
                // applies because of a provisional result. I can't even think of a test
                // case where this would occur and even then, this would not be unsound.
                // Supporting this makes the code more involved, so I am just going to
                // ignore this for now.
                MergeCandidateInfo::AlwaysApplicable(i) => {
                    for (j, c) in where_bounds.into_iter().enumerate() {
                        if i != j {
                            self.ignore_candidate_head_usages(c.head_usages)
                        }
                    }
                    // If a where-bound does not apply, we don't actually get a
                    // candidate for it. We manually track the head usages
                    // of all failed `ParamEnv` candidates instead.
                    self.ignore_candidate_head_usages(failed_candidate_info.param_env_head_usages);
                }
                MergeCandidateInfo::EqualResponse => {}
            }
            return Ok((response, Some(TraitGoalProvenVia::ParamEnv)));
        }

        // Next, prefer any alias bound (nested or otherwise).
        if candidates.iter().any(|c| matches!(c.source, CandidateSource::AliasBound(_))) {
            let alias_bounds: Vec<_> = candidates
                .extract_if(.., |c| matches!(c.source, CandidateSource::AliasBound(_)))
                .collect();
            return Ok(self.merge_candidates_or_bail_with_ambiguity(
                &alias_bounds,
                TraitGoalProvenVia::AliasBound,
            ));
        }

        self.filter_specialized_impls(AllowInferenceConstraints::No, &mut candidates);
        self.unsound_prefer_builtin_dyn_impl(&mut candidates);

        // If there are *only* global where bounds, then make sure to return that this
        // is still reported as being proven-via the param-env so that rigid projections
        // operate correctly. Otherwise, drop all global where-bounds before merging the
        // remaining candidates.
        let only_global_where_bounds = candidates
            .iter()
            .all(|c| matches!(c.source, CandidateSource::ParamEnv(ParamEnvSource::Global)));
        if only_global_where_bounds {
            Ok(self
                .merge_candidates_or_bail_with_ambiguity(&candidates, TraitGoalProvenVia::ParamEnv))
        } else {
            candidates
                .retain(|c| !matches!(c.source, CandidateSource::ParamEnv(ParamEnvSource::Global)));
            Ok(self.merge_impl_candidates_or_bail_with_ambiguity(trait_def_id, &candidates))
        }
    }

    #[instrument(level = "trace", skip(self))]
    pub(super) fn compute_trait_goal(
        &mut self,
        goal: Goal<I, TraitClause<I>>,
    ) -> Result<(CanonicalResponse<I>, Option<TraitGoalProvenVia>), NoSolutionOrRerunNonErased>
    {
        let (candidates, failed_candidate_info) =
            self.assemble_and_evaluate_candidates(goal, AssembleCandidatesFrom::All)?;
        self.merge_trait_candidates(goal.predicate.def_id(), candidates, failed_candidate_info)
            .map_err(Into::into)
    }

    fn try_stall_coroutine(
        &mut self,
        self_ty: I::Ty,
    ) -> Option<Result<Candidate<I>, NoSolutionOrRerunNonErased>> {
        if let ty::Coroutine(def_id, _) = self_ty.kind() {
            match self.typing_mode() {
                TypingMode::Typeck { defining_opaque_types_and_generators: stalled_generators } => {
                    if def_id.as_local().is_some_and(|def_id| stalled_generators.contains(&def_id))
                    {
                        return Some(self.forced_ambiguity(MaybeInfo {
                            cause: MaybeCause::Ambiguity,
                            opaque_types_jank: OpaqueTypesJank::AllGood,
                            stalled_on_coroutines: StalledOnCoroutines::Yes,
                        }));
                    }
                }
                TypingMode::ErasedNotCoherence(MayBeErased) => {
                    // Trying to continue here isn't worth it.
                    return Some(
                        match self.opaque_accesses.rerun_always(RerunReason::TryStallCoroutine) {
                            Err(e) => Err(e.into()),
                        },
                    );
                }
                TypingMode::Coherence
                | TypingMode::PostAnalysis
                | TypingMode::Reflection
                | TypingMode::Codegen
                | TypingMode::PostTypeckUntilBorrowck { defining_opaque_types: _ }
                | TypingMode::PostBorrowck { defined_opaque_types: _ } => {}
            }
        }

        None
    }
}