compiler/rustc_next_trait_solver/src/solve/assembly/mod.rs

//! Code shared by trait and projection goals for candidate assembly.

pub(super) mod structural_traits;

use std::cell::Cell;
use std::ops::ControlFlow;

use derive_where::derive_where;
use rustc_type_ir::inherent::*;
use rustc_type_ir::lang_items::SolverTraitLangItem;
use rustc_type_ir::search_graph::CandidateHeadUsages;
use rustc_type_ir::solve::{
    AliasBoundKind, MaybeInfo, NoSolutionOrRerunNonErased, QueryResultOrRerunNonErased,
    RerunNonErased, RerunReason, RerunResultExt, SizedTraitKind, StalledOnCoroutines,
};
use rustc_type_ir::{
    self as ty, AliasTy, Interner, MayBeErased, Region, TypeFlags, TypeFoldable, TypeFolder,
    TypeSuperFoldable, TypeSuperVisitable, TypeVisitable, TypeVisitableExt, TypeVisitor,
    TypingMode, Unnormalized, Upcast, elaborate,
};
use tracing::{debug, instrument};

use super::trait_goals::TraitGoalProvenVia;
use super::{has_only_region_constraints, inspect};
use crate::delegate::SolverDelegate;
use crate::solve::inspect::ProbeKind;
use crate::solve::{
    BuiltinImplSource, CandidateSource, CanonicalResponse, Certainty, EvalCtxt, Goal, GoalSource,
    MaybeCause, NoSolution, OpaqueTypesJank, ParamEnvSource, QueryResult,
    has_no_inference_or_external_constraints,
};

/// A candidate is a possible way to prove a goal.
///
/// It consists of both the `source`, which describes how that goal would be proven,
/// and the `result` when using the given `source`.
#[derive_where(Debug; I: Interner)]
pub(super) struct Candidate<I: Interner> {
    pub(super) source: CandidateSource<I>,
    pub(super) result: CanonicalResponse<I>,
    pub(super) head_usages: CandidateHeadUsages,
}

/// Methods used to assemble candidates for either trait or projection goals.
pub(super) trait GoalKind<D, I = <D as SolverDelegate>::Interner>:
    TypeFoldable<I> + Copy + Eq + std::fmt::Display
where
    D: SolverDelegate<Interner = I>,
    I: Interner,
{
    fn self_ty(self) -> I::Ty;

    fn trait_ref(self, cx: I) -> ty::TraitRef<I>;

    fn with_replaced_self_ty(self, cx: I, self_ty: I::Ty) -> Self;

    fn trait_def_id(self, cx: I) -> I::TraitId;

    /// Consider a clause, which consists of a "assumption" and some "requirements",
    /// to satisfy a goal. If the requirements hold, then attempt to satisfy our
    /// goal by equating it with the assumption.
    fn probe_and_consider_implied_clause(
        ecx: &mut EvalCtxt<'_, D>,
        parent_source: CandidateSource<I>,
        goal: Goal<I, Self>,
        assumption: I::Clause,
        requirements: impl IntoIterator<Item = (GoalSource, Goal<I, I::Predicate>)>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        Self::probe_and_match_goal_against_assumption(ecx, parent_source, goal, assumption, |ecx| {
            for (nested_source, goal) in requirements {
                ecx.add_goal(nested_source, goal)?;
            }
            ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
        })
    }

    /// Consider a clause specifically for a `dyn Trait` self type. This requires
    /// additionally checking all of the supertraits and object bounds to hold,
    /// since they're not implied by the well-formedness of the object type.
    /// `NormalizesTo` overrides this to not check the supertraits for backwards
    /// compatibility with the old solver. cc trait-system-refactor-initiative#245.
    fn probe_and_consider_object_bound_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        source: CandidateSource<I>,
        goal: Goal<I, Self>,
        assumption: I::Clause,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        Self::probe_and_match_goal_against_assumption(ecx, source, goal, assumption, |ecx| {
            let cx = ecx.cx();
            let ty::Dynamic(bounds, _) = goal.predicate.self_ty().kind() else {
                panic!("expected object type in `probe_and_consider_object_bound_candidate`");
            };

            let trait_ref = assumption.kind().map_bound(|clause| match clause {
                ty::ClauseKind::Trait(pred) => pred.trait_ref,
                ty::ClauseKind::Projection(proj) => proj.projection_term.trait_ref(cx),

                ty::ClauseKind::RegionOutlives(..)
                | ty::ClauseKind::TypeOutlives(..)
                | ty::ClauseKind::ConstArgHasType(..)
                | ty::ClauseKind::WellFormed(..)
                | ty::ClauseKind::ConstEvaluatable(..)
                | ty::ClauseKind::HostEffect(..)
                | ty::ClauseKind::UnstableFeature(..) => {
                    unreachable!("expected trait or projection predicate as an assumption")
                }
            });

            match structural_traits::predicates_for_object_candidate(
                ecx,
                goal.param_env,
                trait_ref,
                bounds,
            ) {
                Ok(requirements) => {
                    ecx.add_goals(GoalSource::ImplWhereBound, requirements)?;
                    ecx.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
                }
                Err(_) => {
                    ecx.evaluate_added_goals_and_make_canonical_response(Certainty::AMBIGUOUS)
                }
            }
        })
    }

    /// Assemble additional assumptions for an alias that are not included
    /// in the item bounds of the alias. For now, this is limited to the
    /// `explicit_implied_const_bounds` for an associated type.
    fn consider_additional_alias_assumptions(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        alias_ty: ty::AliasTy<I>,
    ) -> Vec<Candidate<I>>;

    fn probe_and_consider_param_env_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        assumption: I::Clause,
    ) -> Result<Result<Candidate<I>, CandidateHeadUsages>, RerunNonErased> {
        match Self::fast_reject_assumption(ecx, goal, assumption) {
            Ok(()) => {}
            Err(NoSolution) => return Ok(Err(CandidateHeadUsages::default())),
        }

        // Dealing with `ParamEnv` candidates is a bit of a mess as we need to lazily
        // check whether the candidate is global while considering normalization.
        //
        // We need to write into `source` inside of `match_assumption`, but need to access it
        // in `probe` even if the candidate does not apply before we get there. We handle this
        // by using a `Cell` here. We only ever write into it inside of `match_assumption`.
        let source = Cell::new(CandidateSource::ParamEnv(ParamEnvSource::Global));
        let (result, head_usages) = ecx
            .probe(|result: &QueryResult<I>| inspect::ProbeKind::TraitCandidate {
                source: source.get(),
                result: *result,
            })
            .enter_single_candidate(|ecx| {
                Self::match_assumption(
                    ecx,
                    goal,
                    assumption,
                    |ecx| -> Result<_, NoSolutionOrRerunNonErased> {
                        ecx.try_evaluate_added_goals()?;
                        let (src, certainty) =
                            ecx.characterize_param_env_assumption(goal.param_env, assumption)?;
                        source.set(src);
                        ecx.evaluate_added_goals_and_make_canonical_response(certainty)
                    },
                )
                .map_err(Into::into)
            });

        Ok(match result.map_err_to_rerun()? {
            Ok(result) => Ok(Candidate { source: source.get(), result, head_usages }),
            Err(NoSolution) => Err(head_usages),
        })
    }

    /// Try equating an assumption predicate against a goal's predicate. If it
    /// holds, then execute the `then` callback, which should do any additional
    /// work, then produce a response (typically by executing
    /// [`EvalCtxt::evaluate_added_goals_and_make_canonical_response`]).
    fn probe_and_match_goal_against_assumption(
        ecx: &mut EvalCtxt<'_, D>,
        source: CandidateSource<I>,
        goal: Goal<I, Self>,
        assumption: I::Clause,
        then: impl FnOnce(&mut EvalCtxt<'_, D>) -> QueryResultOrRerunNonErased<I>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        Self::fast_reject_assumption(ecx, goal, assumption)?;

        ecx.probe_trait_candidate(source)
            .enter(|ecx| Self::match_assumption(ecx, goal, assumption, then))
    }

    /// Try to reject the assumption based off of simple heuristics, such as [`ty::ClauseKind`]
    /// and `DefId`.
    fn fast_reject_assumption(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        assumption: I::Clause,
    ) -> Result<(), NoSolution>;

    /// Relate the goal and assumption.
    fn match_assumption(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        assumption: I::Clause,
        then: impl FnOnce(&mut EvalCtxt<'_, D>) -> QueryResultOrRerunNonErased<I>,
    ) -> QueryResultOrRerunNonErased<I>;

    fn consider_impl_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        impl_def_id: I::ImplId,
        then: impl FnOnce(&mut EvalCtxt<'_, D>, Certainty) -> QueryResultOrRerunNonErased<I>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// If the predicate contained an error, we want to avoid emitting unnecessary trait
    /// errors but still want to emit errors for other trait goals. We have some special
    /// handling for this case.
    ///
    /// Trait goals always hold while projection goals never do. This is a bit arbitrary
    /// but prevents incorrect normalization while hiding any trait errors.
    fn consider_error_guaranteed_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        guar: I::ErrorGuaranteed,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A type implements an `auto trait` if its components do as well.
    ///
    /// These components are given by built-in rules from
    /// [`structural_traits::instantiate_constituent_tys_for_auto_trait`].
    fn consider_auto_trait_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A trait alias holds if the RHS traits and `where` clauses hold.
    fn consider_trait_alias_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A type is `Sized` if its tail component is `Sized` and a type is `MetaSized` if its tail
    /// component is `MetaSized`.
    ///
    /// These components are given by built-in rules from
    /// [`structural_traits::instantiate_constituent_tys_for_sizedness_trait`].
    fn consider_builtin_sizedness_candidates(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        sizedness: SizedTraitKind,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A type is `Copy` or `Clone` if its components are `Copy` or `Clone`.
    ///
    /// These components are given by built-in rules from
    /// [`structural_traits::instantiate_constituent_tys_for_copy_clone_trait`].
    fn consider_builtin_copy_clone_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A type is a `FnPtr` if it is of `FnPtr` type.
    fn consider_builtin_fn_ptr_trait_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A callable type (a closure, fn def, or fn ptr) is known to implement the `Fn<A>`
    /// family of traits where `A` is given by the signature of the type.
    fn consider_builtin_fn_trait_candidates(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        kind: ty::ClosureKind,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// An async closure is known to implement the `AsyncFn<A>` family of traits
    /// where `A` is given by the signature of the type.
    fn consider_builtin_async_fn_trait_candidates(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
        kind: ty::ClosureKind,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// Compute the built-in logic of the `AsyncFnKindHelper` helper trait, which
    /// is used internally to delay computation for async closures until after
    /// upvar analysis is performed in HIR typeck.
    fn consider_builtin_async_fn_kind_helper_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// `Tuple` is implemented if the `Self` type is a tuple.
    fn consider_builtin_tuple_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// `Pointee` is always implemented.
    ///
    /// See the projection implementation for the `Metadata` types for all of
    /// the built-in types. For structs, the metadata type is given by the struct
    /// tail.
    fn consider_builtin_pointee_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A coroutine (that comes from an `async` desugaring) is known to implement
    /// `Future<Output = O>`, where `O` is given by the coroutine's return type
    /// that was computed during type-checking.
    fn consider_builtin_future_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A coroutine (that comes from a `gen` desugaring) is known to implement
    /// `Iterator<Item = O>`, where `O` is given by the generator's yield type
    /// that was computed during type-checking.
    fn consider_builtin_iterator_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A coroutine (that comes from a `gen` desugaring) is known to implement
    /// `FusedIterator`
    fn consider_builtin_fused_iterator_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    fn consider_builtin_async_iterator_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// A coroutine (that doesn't come from an `async` or `gen` desugaring) is known to
    /// implement `Coroutine<R, Yield = Y, Return = O>`, given the resume, yield,
    /// and return types of the coroutine computed during type-checking.
    fn consider_builtin_coroutine_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    fn consider_builtin_discriminant_kind_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    fn consider_builtin_destruct_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

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

    fn consider_builtin_bikeshed_guaranteed_no_drop_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    fn consider_builtin_try_as_dyn_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;

    /// Consider (possibly several) candidates to upcast or unsize a type to another
    /// type, excluding the coercion of a sized type into a `dyn Trait`.
    ///
    /// We return the `BuiltinImplSource` for each candidate as it is needed
    /// for unsize coercion in hir typeck and because it is difficult to
    /// otherwise recompute this for codegen. This is a bit of a mess but the
    /// easiest way to maintain the existing behavior for now.
    fn consider_structural_builtin_unsize_candidates(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Vec<Candidate<I>>, RerunNonErased>;

    fn consider_builtin_field_candidate(
        ecx: &mut EvalCtxt<'_, D>,
        goal: Goal<I, Self>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased>;
}

/// Allows callers of `assemble_and_evaluate_candidates` to choose whether to limit
/// candidate assembly to param-env and alias-bound candidates.
///
/// On top of being a micro-optimization, as it avoids doing unnecessary work when
/// a param-env trait bound candidate shadows impls for normalization, this is also
/// required to prevent query cycles due to RPITIT inference. See the issue at:
/// <https://github.com/rust-lang/trait-system-refactor-initiative/issues/173>.
pub(super) enum AssembleCandidatesFrom {
    All,
    /// Only assemble candidates from the environment and alias bounds, ignoring
    /// user-written and built-in impls. We only expect `ParamEnv` and `AliasBound`
    /// candidates to be assembled.
    EnvAndBounds,
}

impl AssembleCandidatesFrom {
    fn should_assemble_impl_candidates(&self) -> bool {
        match self {
            AssembleCandidatesFrom::All => true,
            AssembleCandidatesFrom::EnvAndBounds => false,
        }
    }
}

/// This is currently used to track the [CandidateHeadUsages] of all failed `ParamEnv`
/// candidates. This is then used to ignore their head usages in case there's another
/// always applicable `ParamEnv` candidate. Look at how `param_env_head_usages` is
/// used in the code for more details.
///
/// We could easily extend this to also ignore head usages of other ignored candidates.
/// However, we currently don't have any tests where this matters and the complexity of
/// doing so does not feel worth it for now.
#[derive(Debug)]
pub(super) struct FailedCandidateInfo {
    pub param_env_head_usages: CandidateHeadUsages,
}

impl<D, I> EvalCtxt<'_, D>
where
    D: SolverDelegate<Interner = I>,
    I: Interner,
{
    // FIXME(#155443): This function should only ever return an error
    // as we want to force a rerun when accessing opaques. We should change
    // this file to revert all the newly added places which return `NoSolution`.
    pub(super) fn assemble_and_evaluate_candidates<G: GoalKind<D>>(
        &mut self,
        goal: Goal<I, G>,
        assemble_from: AssembleCandidatesFrom,
    ) -> Result<(Vec<Candidate<I>>, FailedCandidateInfo), RerunNonErased> {
        let mut candidates = vec![];
        let mut failed_candidate_info =
            FailedCandidateInfo { param_env_head_usages: CandidateHeadUsages::default() };
        let Ok(normalized_self_ty) =
            self.structurally_normalize_ty(goal.param_env, goal.predicate.self_ty())
        else {
            return Ok((candidates, failed_candidate_info));
        };

        let goal: Goal<I, G> = goal
            .with(self.cx(), goal.predicate.with_replaced_self_ty(self.cx(), normalized_self_ty));

        if normalized_self_ty.is_ty_var() {
            debug!("self type has been normalized to infer");
            self.try_assemble_bounds_via_registered_opaques(goal, assemble_from, &mut candidates)?;
            return Ok((candidates, failed_candidate_info));
        }

        // Vars that show up in the rest of the goal substs may have been constrained by
        // normalizing the self type as well, since type variables are not uniquified.
        let goal = self.resolve_vars_if_possible(goal);

        if self.typing_mode().is_coherence()
            && let Ok(candidate) = self.consider_coherence_unknowable_candidate(goal)
        {
            candidates.push(candidate);
            return Ok((candidates, failed_candidate_info));
        }

        self.assemble_alias_bound_candidates(goal, &mut candidates)?;
        self.assemble_param_env_candidates(goal, &mut candidates, &mut failed_candidate_info)?;

        match assemble_from {
            AssembleCandidatesFrom::All => {
                self.assemble_builtin_impl_candidates(goal, &mut candidates)?;
                // For performance we only assemble impls if there are no candidates
                // which would shadow them. This is necessary to avoid hangs in rayon,
                // see trait-system-refactor-initiative#109 for more details.
                //
                // We always assemble builtin impls as trivial builtin impls have a higher
                // priority than where-clauses.
                //
                // We only do this if any such candidate applies without any constraints
                // as we may want to weaken inference guidance in the future and don't want
                // to worry about causing major performance regressions when doing so.
                // See trait-system-refactor-initiative#226 for some ideas here.
                let assemble_impls = match self.typing_mode() {
                    TypingMode::Coherence => true,
                    TypingMode::Typeck { .. }
                    | TypingMode::PostTypeckUntilBorrowck { .. }
                    | TypingMode::Reflection
                    | TypingMode::PostBorrowck { .. }
                    | TypingMode::PostAnalysis
                    | TypingMode::Codegen
                    | TypingMode::ErasedNotCoherence(MayBeErased) => !candidates.iter().any(|c| {
                        matches!(
                            c.source,
                            CandidateSource::ParamEnv(ParamEnvSource::NonGlobal)
                                | CandidateSource::AliasBound(_)
                        ) && has_no_inference_or_external_constraints(c.result)
                    }),
                };
                if assemble_impls {
                    self.assemble_impl_candidates(goal, &mut candidates)?;
                    self.assemble_object_bound_candidates(goal, &mut candidates);
                }
            }
            AssembleCandidatesFrom::EnvAndBounds => {
                // This is somewhat inconsistent and may make #57893 slightly easier to exploit.
                // However, it matches the behavior of the old solver. See
                // `tests/ui/traits/next-solver/normalization-shadowing/use_object_if_empty_env.rs`.
                if matches!(normalized_self_ty.kind(), ty::Dynamic(..))
                    && !candidates.iter().any(|c| matches!(c.source, CandidateSource::ParamEnv(_)))
                {
                    self.assemble_object_bound_candidates(goal, &mut candidates);
                }
            }
        }

        Ok((candidates, failed_candidate_info))
    }

    pub(super) fn forced_ambiguity(
        &mut self,
        maybe: MaybeInfo,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        // This may fail if `try_evaluate_added_goals` overflows because it
        // fails to reach a fixpoint but ends up getting an error after
        // running for some additional step.
        //
        // FIXME(@lcnr): While I believe an error here to be possible, we
        // currently don't have any test which actually triggers it. @lqd
        // created a minimization for an ICE in typenum, but that one no
        // longer fails here. cc trait-system-refactor-initiative#105.
        let source = CandidateSource::BuiltinImpl(BuiltinImplSource::Misc);
        let certainty = Certainty::Maybe(maybe);
        self.probe_trait_candidate(source)
            .enter(|this| this.evaluate_added_goals_and_make_canonical_response(certainty))
    }

    #[instrument(level = "trace", skip_all)]
    fn assemble_impl_candidates<G: GoalKind<D>>(
        &mut self,
        goal: Goal<I, G>,
        candidates: &mut Vec<Candidate<I>>,
    ) -> Result<(), RerunNonErased> {
        let cx = self.cx();
        cx.for_each_relevant_impl(goal.predicate.trait_ref(cx), |impl_def_id| -> Result<_, _> {
            // 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 Ok(());
            }
            match G::consider_impl_candidate(self, goal, impl_def_id, |ecx, certainty| {
                ecx.evaluate_added_goals_and_make_canonical_response(certainty)
            })
            .map_err_to_rerun()?
            {
                Ok(candidate) => candidates.push(candidate),
                Err(NoSolution) => {}
            }

            Ok(())
        })
    }

    #[instrument(level = "trace", skip_all)]
    fn assemble_builtin_impl_candidates<G: GoalKind<D>>(
        &mut self,
        goal: Goal<I, G>,
        candidates: &mut Vec<Candidate<I>>,
    ) -> Result<(), RerunNonErased> {
        let cx = self.cx();
        let trait_def_id = goal.predicate.trait_def_id(cx);

        // Builtin impls regularly are not `is_fully_generic_for_reflection`, so instead
        // of trying to handle these manually, we just reject all builtin impls in reflection
        // mode. We can probably lift this restriction for specific cases, but this is safer.
        // See `try_as_dyn_builtin_impl` for how just allowing all builtin impls is unsound.
        if self.typing_mode().is_reflection() {
            return Ok(());
        }

        // N.B. When assembling built-in candidates for lang items that are also
        // `auto` traits, then the auto trait candidate that is assembled in
        // `consider_auto_trait_candidate` MUST be disqualified to remain sound.
        //
        // Instead of adding the logic here, it's a better idea to add it in
        // `EvalCtxt::disqualify_auto_trait_candidate_due_to_possible_impl` in
        // `solve::trait_goals` instead.
        let result = if let ty::Error(guar) = goal.predicate.self_ty().kind() {
            G::consider_error_guaranteed_candidate(self, goal, guar)
        } else if cx.trait_is_auto(trait_def_id) {
            G::consider_auto_trait_candidate(self, goal)
        } else if cx.trait_is_alias(trait_def_id) {
            G::consider_trait_alias_candidate(self, goal)
        } else {
            match cx.as_trait_lang_item(trait_def_id) {
                Some(SolverTraitLangItem::Sized) => {
                    G::consider_builtin_sizedness_candidates(self, goal, SizedTraitKind::Sized)
                }
                Some(SolverTraitLangItem::MetaSized) => {
                    G::consider_builtin_sizedness_candidates(self, goal, SizedTraitKind::MetaSized)
                }
                Some(SolverTraitLangItem::PointeeSized) => {
                    unreachable!("`PointeeSized` is removed during lowering");
                }
                Some(
                    SolverTraitLangItem::Copy
                    | SolverTraitLangItem::Clone
                    | SolverTraitLangItem::TrivialClone,
                ) => G::consider_builtin_copy_clone_candidate(self, goal),
                Some(SolverTraitLangItem::Fn) => {
                    G::consider_builtin_fn_trait_candidates(self, goal, ty::ClosureKind::Fn)
                }
                Some(SolverTraitLangItem::FnMut) => {
                    G::consider_builtin_fn_trait_candidates(self, goal, ty::ClosureKind::FnMut)
                }
                Some(SolverTraitLangItem::FnOnce) => {
                    G::consider_builtin_fn_trait_candidates(self, goal, ty::ClosureKind::FnOnce)
                }
                Some(SolverTraitLangItem::AsyncFn) => {
                    G::consider_builtin_async_fn_trait_candidates(self, goal, ty::ClosureKind::Fn)
                }
                Some(SolverTraitLangItem::AsyncFnMut) => {
                    G::consider_builtin_async_fn_trait_candidates(
                        self,
                        goal,
                        ty::ClosureKind::FnMut,
                    )
                }
                Some(SolverTraitLangItem::AsyncFnOnce) => {
                    G::consider_builtin_async_fn_trait_candidates(
                        self,
                        goal,
                        ty::ClosureKind::FnOnce,
                    )
                }
                Some(SolverTraitLangItem::FnPtrTrait) => {
                    G::consider_builtin_fn_ptr_trait_candidate(self, goal)
                }
                Some(SolverTraitLangItem::AsyncFnKindHelper) => {
                    G::consider_builtin_async_fn_kind_helper_candidate(self, goal)
                }
                Some(SolverTraitLangItem::Tuple) => G::consider_builtin_tuple_candidate(self, goal),
                Some(SolverTraitLangItem::PointeeTrait) => {
                    G::consider_builtin_pointee_candidate(self, goal)
                }
                Some(SolverTraitLangItem::Future) => {
                    G::consider_builtin_future_candidate(self, goal)
                }
                Some(SolverTraitLangItem::Iterator) => {
                    G::consider_builtin_iterator_candidate(self, goal)
                }
                Some(SolverTraitLangItem::FusedIterator) => {
                    G::consider_builtin_fused_iterator_candidate(self, goal)
                }
                Some(SolverTraitLangItem::AsyncIterator) => {
                    G::consider_builtin_async_iterator_candidate(self, goal)
                }
                Some(SolverTraitLangItem::Coroutine) => {
                    G::consider_builtin_coroutine_candidate(self, goal)
                }
                Some(SolverTraitLangItem::DiscriminantKind) => {
                    G::consider_builtin_discriminant_kind_candidate(self, goal)
                }
                Some(SolverTraitLangItem::Destruct) => {
                    G::consider_builtin_destruct_candidate(self, goal)
                }
                Some(SolverTraitLangItem::TransmuteTrait) => {
                    G::consider_builtin_transmute_candidate(self, goal)
                }
                Some(SolverTraitLangItem::BikeshedGuaranteedNoDrop) => {
                    G::consider_builtin_bikeshed_guaranteed_no_drop_candidate(self, goal)
                }
                Some(SolverTraitLangItem::TryAsDyn) => {
                    G::consider_builtin_try_as_dyn_candidate(self, goal)
                }
                Some(SolverTraitLangItem::Field) => G::consider_builtin_field_candidate(self, goal),
                _ => Err(NoSolution.into()),
            }
        };

        candidates.extend(result);

        // There may be multiple unsize candidates for a trait with several supertraits:
        // `trait Foo: Bar<A> + Bar<B>` and `dyn Foo: Unsize<dyn Bar<_>>`
        if cx.is_trait_lang_item(trait_def_id, SolverTraitLangItem::Unsize) {
            candidates.extend(G::consider_structural_builtin_unsize_candidates(self, goal)?);
        }

        Ok(())
    }

    #[instrument(level = "trace", skip_all)]
    fn assemble_param_env_candidates<G: GoalKind<D>>(
        &mut self,
        goal: Goal<I, G>,
        candidates: &mut Vec<Candidate<I>>,
        failed_candidate_info: &mut FailedCandidateInfo,
    ) -> Result<(), RerunNonErased> {
        for assumption in goal.param_env.caller_bounds().iter() {
            match G::probe_and_consider_param_env_candidate(self, goal, assumption)? {
                Ok(candidate) => candidates.push(candidate),
                Err(head_usages) => {
                    failed_candidate_info.param_env_head_usages.merge_usages(head_usages)
                }
            }
        }

        Ok(())
    }

    #[instrument(level = "trace", skip_all)]
    fn assemble_alias_bound_candidates<G: GoalKind<D>>(
        &mut self,
        goal: Goal<I, G>,
        candidates: &mut Vec<Candidate<I>>,
    ) -> Result<(), RerunNonErased> {
        let res = self.probe(|_| ProbeKind::NormalizedSelfTyAssembly).enter(|ecx| {
            ecx.assemble_alias_bound_candidates_recur(
                goal.predicate.self_ty(),
                goal,
                candidates,
                AliasBoundKind::SelfBounds,
            )?;
            Ok(())
        });

        // always returns Ok
        match res {
            Ok(_) => Ok(()),
            Err(NoSolutionOrRerunNonErased::RerunNonErased(e)) => Err(e),
            Err(NoSolutionOrRerunNonErased::NoSolution(NoSolution)) => {
                unreachable!()
            }
        }
    }

    /// For some deeply nested `<T>::A::B::C::D` rigid associated type,
    /// we should explore the item bounds for all levels, since the
    /// `associated_type_bounds` feature means that a parent associated
    /// type may carry bounds for a nested associated type.
    ///
    /// If we have a projection, check that its self type is a rigid projection.
    /// If so, continue searching by recursively calling after normalization.
    // FIXME: This may recurse infinitely, but I can't seem to trigger it without
    // hitting another overflow error something. Add a depth parameter needed later.
    fn assemble_alias_bound_candidates_recur<G: GoalKind<D>>(
        &mut self,
        self_ty: I::Ty,
        goal: Goal<I, G>,
        candidates: &mut Vec<Candidate<I>>,
        consider_self_bounds: AliasBoundKind,
    ) -> Result<(), RerunNonErased> {
        let (alias_ty, def_id) = match self_ty.kind() {
            ty::Bool
            | ty::Char
            | ty::Int(_)
            | ty::Uint(_)
            | ty::Float(_)
            | ty::Adt(_, _)
            | ty::Foreign(_)
            | ty::Str
            | ty::Array(_, _)
            | ty::Pat(_, _)
            | ty::Slice(_)
            | ty::RawPtr(_, _)
            | ty::Ref(_, _, _)
            | ty::FnDef(_, _)
            | ty::FnPtr(..)
            | ty::UnsafeBinder(_)
            | ty::Dynamic(..)
            | ty::Closure(..)
            | ty::CoroutineClosure(..)
            | ty::Coroutine(..)
            | ty::CoroutineWitness(..)
            | ty::Never
            | ty::Tuple(_)
            | ty::Param(_)
            | ty::Placeholder(..)
            | ty::Infer(ty::IntVar(_) | ty::FloatVar(_))
            | ty::Error(_) => return Ok(()),
            ty::Infer(ty::FreshTy(_) | ty::FreshIntTy(_) | ty::FreshFloatTy(_)) | ty::Bound(..) => {
                panic!("unexpected self type for `{goal:?}`")
            }

            ty::Infer(ty::TyVar(_)) => {
                // If we hit infer when normalizing the self type of an alias,
                // then bail with ambiguity. We should never encounter this on
                // the *first* iteration of this recursive function.
                if let Ok(result) =
                    self.evaluate_added_goals_and_make_canonical_response(Certainty::AMBIGUOUS)
                {
                    candidates.push(Candidate {
                        source: CandidateSource::AliasBound(consider_self_bounds),
                        result,
                        head_usages: CandidateHeadUsages::default(),
                    });
                }
                return Ok(());
            }

            ty::Alias(
                ty::IsRigid::Yes,
                alias_ty @ AliasTy { kind: ty::Projection { def_id }, .. },
            ) => (alias_ty, def_id.into()),

            ty::Alias(ty::IsRigid::Yes, alias_ty @ AliasTy { kind: ty::Opaque { def_id }, .. }) => {
                (alias_ty, def_id.into())
            }

            ty::Alias(ty::IsRigid::No, _) => unreachable!("non-rigid self type: {self_ty:?}"),

            ty::Alias(
                ty::IsRigid::Yes,
                AliasTy { kind: ty::Inherent { .. } | ty::Free { .. }, .. },
            ) => {
                self.cx().delay_bug(format!("could not normalize {self_ty:?}, it is not WF"));
                return Ok(());
            }
        };

        match consider_self_bounds {
            AliasBoundKind::SelfBounds => {
                for assumption in self
                    .cx()
                    .item_self_bounds(def_id)
                    .iter_instantiated(self.cx(), alias_ty.args)
                    .map(Unnormalized::skip_norm_wip)
                {
                    candidates.extend(G::probe_and_consider_implied_clause(
                        self,
                        CandidateSource::AliasBound(consider_self_bounds),
                        goal,
                        assumption,
                        [],
                    ));
                }
            }
            AliasBoundKind::NonSelfBounds => {
                for assumption in self
                    .cx()
                    .item_non_self_bounds(def_id)
                    .iter_instantiated(self.cx(), alias_ty.args)
                    .map(Unnormalized::skip_norm_wip)
                {
                    candidates.extend(G::probe_and_consider_implied_clause(
                        self,
                        CandidateSource::AliasBound(consider_self_bounds),
                        goal,
                        assumption,
                        [],
                    ));
                }
            }
        }

        candidates.extend(G::consider_additional_alias_assumptions(self, goal, alias_ty));

        let Some(projection_ty) = alias_ty.try_to_projection() else {
            return Ok(());
        };

        // Recurse on the self type of the projection.
        match self.structurally_normalize_ty(goal.param_env, projection_ty.projection_self_ty()) {
            Ok(next_self_ty) => self.assemble_alias_bound_candidates_recur(
                next_self_ty,
                goal,
                candidates,
                AliasBoundKind::NonSelfBounds,
            ),
            Err(NoSolutionOrRerunNonErased::NoSolution(NoSolution)) => Ok(()),
            Err(NoSolutionOrRerunNonErased::RerunNonErased(e)) => Err(e),
        }
    }

    #[instrument(level = "trace", skip_all)]
    fn assemble_object_bound_candidates<G: GoalKind<D>>(
        &mut self,
        goal: Goal<I, G>,
        candidates: &mut Vec<Candidate<I>>,
    ) {
        let cx = self.cx();
        if cx.is_sizedness_trait(goal.predicate.trait_def_id(cx)) {
            // `dyn MetaSized` is valid, but should get its `MetaSized` impl from
            // being `dyn` (SizedCandidate), not from the object candidate.
            return;
        }

        // Builtin impls regularly are not `is_fully_generic_for_reflection`, so instead
        // of trying to handle these manually, we just reject all builtin impls in reflection
        // mode. We can probably lift this restriction for specific cases, but this is safer.
        // See `try_as_dyn_builtin_impl` for how just allowing all builtin impls is unsound.
        if self.typing_mode().is_reflection() {
            return;
        }

        let self_ty = goal.predicate.self_ty();
        let bounds = match self_ty.kind() {
            ty::Bool
            | ty::Char
            | ty::Int(_)
            | ty::Uint(_)
            | ty::Float(_)
            | ty::Adt(_, _)
            | ty::Foreign(_)
            | ty::Str
            | ty::Array(_, _)
            | ty::Pat(_, _)
            | ty::Slice(_)
            | ty::RawPtr(_, _)
            | ty::Ref(_, _, _)
            | ty::FnDef(_, _)
            | ty::FnPtr(..)
            | ty::UnsafeBinder(_)
            | ty::Alias(..)
            | ty::Closure(..)
            | ty::CoroutineClosure(..)
            | ty::Coroutine(..)
            | ty::CoroutineWitness(..)
            | ty::Never
            | ty::Tuple(_)
            | ty::Param(_)
            | ty::Placeholder(..)
            | ty::Infer(ty::IntVar(_) | ty::FloatVar(_))
            | ty::Error(_) => return,
            ty::Infer(ty::TyVar(_) | ty::FreshTy(_) | ty::FreshIntTy(_) | ty::FreshFloatTy(_))
            | ty::Bound(..) => panic!("unexpected self type for `{goal:?}`"),
            ty::Dynamic(bounds, ..) => bounds,
        };

        // Do not consider built-in object impls for dyn-incompatible types.
        if bounds.principal_def_id().is_some_and(|def_id| !cx.trait_is_dyn_compatible(def_id)) {
            return;
        }

        // Consider all of the auto-trait and projection bounds, which don't
        // need to be recorded as a `BuiltinImplSource::Object` since they don't
        // really have a vtable base...
        for bound in bounds.iter() {
            match bound.skip_binder() {
                ty::ExistentialPredicate::Trait(_) => {
                    // Skip principal
                }
                ty::ExistentialPredicate::Projection(_)
                | ty::ExistentialPredicate::AutoTrait(_) => {
                    candidates.extend(G::probe_and_consider_object_bound_candidate(
                        self,
                        CandidateSource::BuiltinImpl(BuiltinImplSource::Misc),
                        goal,
                        bound.with_self_ty(cx, self_ty),
                    ));
                }
            }
        }

        // FIXME: We only need to do *any* of this if we're considering a trait goal,
        // since we don't need to look at any supertrait or anything if we are doing
        // a projection goal.
        if let Some(principal) = bounds.principal() {
            let principal_trait_ref = principal.with_self_ty(cx, self_ty);
            for (idx, assumption) in elaborate::supertraits(cx, principal_trait_ref).enumerate() {
                candidates.extend(G::probe_and_consider_object_bound_candidate(
                    self,
                    CandidateSource::BuiltinImpl(BuiltinImplSource::Object(idx)),
                    goal,
                    assumption.upcast(cx),
                ));
            }
        }
    }

    /// In coherence we have to not only care about all impls we know about, but
    /// also consider impls which may get added in a downstream or sibling crate
    /// or which an upstream impl may add in a minor release.
    ///
    /// To do so we return a single ambiguous candidate in case such an unknown
    /// impl could apply to the current goal.
    #[instrument(level = "trace", skip_all)]
    fn consider_coherence_unknowable_candidate<G: GoalKind<D>>(
        &mut self,
        goal: Goal<I, G>,
    ) -> Result<Candidate<I>, NoSolutionOrRerunNonErased> {
        self.probe_trait_candidate(CandidateSource::CoherenceUnknowable).enter(|ecx| {
            let cx = ecx.cx();
            let trait_ref = goal.predicate.trait_ref(cx);
            if ecx.trait_ref_is_knowable(goal.param_env, trait_ref)? {
                Err(NoSolution.into())
            } else {
                // While the trait bound itself may be unknowable, we may be able to
                // prove that a super trait is not implemented. For this, we recursively
                // prove the super trait bounds of the current goal.
                //
                // We skip the goal itself as that one would cycle.
                let predicate: I::Predicate = trait_ref.upcast(cx);
                ecx.add_goals(
                    GoalSource::Misc,
                    elaborate::elaborate(cx, [predicate])
                        .skip(1)
                        .map(|predicate| goal.with(cx, predicate)),
                )?;
                ecx.evaluate_added_goals_and_make_canonical_response(Certainty::AMBIGUOUS)
            }
        })
    }
}

pub(super) enum AllowInferenceConstraints {
    Yes,
    No,
}

impl<D, I> EvalCtxt<'_, D>
where
    D: SolverDelegate<Interner = I>,
    I: Interner,
{
    /// Check whether we can ignore impl candidates due to specialization.
    ///
    /// This is only necessary for `feature(specialization)` and seems quite ugly.
    pub(super) fn filter_specialized_impls(
        &mut self,
        allow_inference_constraints: AllowInferenceConstraints,
        candidates: &mut Vec<Candidate<I>>,
    ) {
        if self.typing_mode().is_coherence() {
            return;
        }

        let mut i = 0;
        'outer: while i < candidates.len() {
            let CandidateSource::Impl(victim_def_id) = candidates[i].source else {
                i += 1;
                continue;
            };

            for (j, c) in candidates.iter().enumerate() {
                if i == j {
                    continue;
                }

                let CandidateSource::Impl(other_def_id) = c.source else {
                    continue;
                };

                // See if we can toss out `victim` based on specialization.
                //
                // While this requires us to know *for sure* that the `lhs` impl applies
                // we still use modulo regions here. This is fine as specialization currently
                // assumes that specializing impls have to be always applicable, meaning that
                // the only allowed region constraints may be constraints also present on the default impl.
                if matches!(allow_inference_constraints, AllowInferenceConstraints::Yes)
                    || has_only_region_constraints(c.result)
                {
                    if self.cx().impl_specializes(other_def_id, victim_def_id) {
                        candidates.remove(i);
                        continue 'outer;
                    }
                }
            }

            i += 1;
        }
    }

    /// If the self type is the hidden type of an opaque, try to assemble
    /// candidates for it by consider its item bounds and by using blanket
    /// impls. This is used to incompletely guide type inference when handling
    /// non-defining uses in the defining scope.
    ///
    /// We otherwise just fail fail with ambiguity. Even if we're using an
    /// opaque type item bound or a blank impls, we still force its certainty
    /// to be `Maybe` so that we properly prove this goal later.
    ///
    /// See <https://github.com/rust-lang/trait-system-refactor-initiative/issues/182>
    /// for why this is necessary.
    #[tracing::instrument(skip(self, assemble_from))]
    fn try_assemble_bounds_via_registered_opaques<G: GoalKind<D>>(
        &mut self,
        goal: Goal<I, G>,
        assemble_from: AssembleCandidatesFrom,
        candidates: &mut Vec<Candidate<I>>,
    ) -> Result<(), RerunNonErased> {
        let self_ty = goal.predicate.self_ty();
        // We only use this hack during HIR typeck.
        let opaque_types = match self.typing_mode() {
            TypingMode::Typeck { .. } => self.opaques_with_sub_unified_hidden_type(self_ty),
            TypingMode::Coherence
            | TypingMode::PostTypeckUntilBorrowck { .. }
            | TypingMode::PostBorrowck { .. }
            | TypingMode::PostAnalysis
            | TypingMode::Reflection
            | TypingMode::Codegen => vec![],
            TypingMode::ErasedNotCoherence(MayBeErased) => {
                self.opaque_accesses
                    .rerun_if_any_opaque_has_infer_as_hidden_type(RerunReason::SelfTyInfer)?;
                Vec::new()
            }
        };

        if opaque_types.is_empty() {
            candidates.extend(self.forced_ambiguity(MaybeInfo::AMBIGUOUS));
            return Ok(());
        }

        for &opaque_ty in &opaque_types {
            debug!("self ty is sub unified with {opaque_ty:?}");

            struct ReplaceOpaque<I: Interner> {
                cx: I,
                opaque_ty: ty::OpaqueAliasTy<I>,
                self_ty: I::Ty,
            }
            impl<I: Interner> TypeFolder<I> for ReplaceOpaque<I> {
                fn cx(&self) -> I {
                    self.cx
                }
                fn fold_ty(&mut self, ty: I::Ty) -> I::Ty {
                    if let ty::Alias(is_rigid, alias_ty) = ty.kind()
                        && let Some(opaque_ty) = alias_ty.try_to_opaque()
                    {
                        if opaque_ty == self.opaque_ty {
                            debug_assert_eq!(is_rigid, ty::IsRigid::No);
                            return self.self_ty;
                        }
                    }
                    ty.super_fold_with(self)
                }
            }

            // We look at all item-bounds of the opaque, replacing the
            // opaque with the current self type before considering
            // them as a candidate. Imagine we've got `?x: Trait<?y>`
            // and `?x` has been sub-unified with the hidden type of
            // `impl Trait<u32>`, We take the item bound `opaque: Trait<u32>`
            // and replace all occurrences of `opaque` with `?x`. This results
            // in a `?x: Trait<u32>` alias-bound candidate.
            for item_bound in self
                .cx()
                .item_self_bounds(opaque_ty.kind.into())
                .iter_instantiated(self.cx(), opaque_ty.args)
                .map(Unnormalized::skip_norm_wip)
            {
                let assumption =
                    item_bound.fold_with(&mut ReplaceOpaque { cx: self.cx(), opaque_ty, self_ty });
                candidates.extend(G::probe_and_match_goal_against_assumption(
                    self,
                    CandidateSource::AliasBound(AliasBoundKind::SelfBounds),
                    goal,
                    assumption,
                    |ecx| {
                        // We want to reprove this goal once we've inferred the
                        // hidden type, so we force the certainty to `Maybe`.
                        ecx.evaluate_added_goals_and_make_canonical_response(Certainty::AMBIGUOUS)
                    },
                ));
            }
        }

        // If the self type is sub unified with any opaque type, we also look at blanket
        // impls for it.
        //
        // See tests/ui/impl-trait/non-defining-uses/use-blanket-impl.rs for an example.
        if assemble_from.should_assemble_impl_candidates() {
            let cx = self.cx();
            cx.for_each_blanket_impl(goal.predicate.trait_def_id(cx), |impl_def_id| {
                // 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 Ok(());
                }

                match G::consider_impl_candidate(self, goal, impl_def_id, |ecx, certainty| {
                    if ecx.shallow_resolve(self_ty).is_ty_var() {
                        // We force the certainty of impl candidates to be `Maybe`.
                        let certainty = certainty.and(Certainty::AMBIGUOUS);
                        ecx.evaluate_added_goals_and_make_canonical_response(certainty)
                    } else {
                        // We don't want to use impls if they constrain the opaque.
                        //
                        // FIXME(trait-system-refactor-initiative#229): This isn't
                        // perfect yet as it still allows us to incorrectly constrain
                        // other inference variables.
                        Err(NoSolution.into())
                    }
                })
                .map_err_to_rerun()?
                {
                    Ok(candidate) => candidates.push(candidate),
                    Err(NoSolution) => {}
                }

                Ok(())
            })?;
        }

        if candidates.is_empty() {
            let source = CandidateSource::BuiltinImpl(BuiltinImplSource::Misc);
            let certainty = Certainty::Maybe(MaybeInfo {
                cause: MaybeCause::Ambiguity,
                opaque_types_jank: OpaqueTypesJank::ErrorIfRigidSelfTy,
                stalled_on_coroutines: StalledOnCoroutines::No,
            });
            candidates
                .extend(self.probe_trait_candidate(source).enter(|this| {
                    this.evaluate_added_goals_and_make_canonical_response(certainty)
                }));
        }

        Ok(())
    }

    /// Assemble and merge candidates for goals which are related to an underlying trait
    /// goal. Right now, this is normalizes-to and host effect goals.
    ///
    /// We sadly can't simply take all possible candidates for normalization goals
    /// and check whether they result in the same constraints. We want to make sure
    /// that trying to normalize an alias doesn't result in constraints which aren't
    /// otherwise required.
    ///
    /// Most notably, when proving a trait goal by via a where-bound, we should not
    /// normalize via impls which have stricter region constraints than the where-bound:
    ///
    /// ```rust
    /// trait Trait<'a> {
    ///     type Assoc;
    /// }
    ///
    /// impl<'a, T: 'a> Trait<'a> for T {
    ///     type Assoc = u32;
    /// }
    ///
    /// fn with_bound<'a, T: Trait<'a>>(_value: T::Assoc) {}
    /// ```
    ///
    /// The where-bound of `with_bound` doesn't specify the associated type, so we would
    /// only be able to normalize `<T as Trait<'a>>::Assoc` by using the impl. This impl
    /// adds a `T: 'a` bound however, which would result in a region error. Given that the
    /// user explicitly wrote that `T: Trait<'a>` holds, this is undesirable and we instead
    /// treat the alias as rigid.
    ///
    /// See trait-system-refactor-initiative#124 for more details.
    #[instrument(level = "debug", skip_all, fields(proven_via, goal), ret)]
    pub(super) fn assemble_and_merge_candidates<G: GoalKind<D>>(
        &mut self,
        proven_via: Option<TraitGoalProvenVia>,
        goal: Goal<I, G>,
        inject_forced_ambiguity_candidate: impl FnOnce(
            &mut EvalCtxt<'_, D>,
        ) -> Option<
            Result<CanonicalResponse<I>, NoSolutionOrRerunNonErased>,
        >,
        inject_normalize_to_rigid_candidate: impl FnOnce(
            &mut EvalCtxt<'_, D>,
        ) -> Result<
            CanonicalResponse<I>,
            NoSolutionOrRerunNonErased,
        >,
    ) -> QueryResultOrRerunNonErased<I> {
        let Some(proven_via) = proven_via else {
            // We don't care about overflow. If proving the trait goal overflowed, then
            // it's enough to report an overflow error for that, we don't also have to
            // overflow during normalization.
            //
            // We use `forced_ambiguity` here over `make_ambiguous_response_no_constraints`
            // because the former will also record a built-in candidate in the inspector.
            return self.forced_ambiguity(MaybeInfo::AMBIGUOUS).map(|cand| cand.result);
        };

        match proven_via {
            TraitGoalProvenVia::ParamEnv | TraitGoalProvenVia::AliasBound => {
                // Even when a trait bound has been proven using a where-bound, we
                // still need to consider alias-bounds for normalization, see
                // `tests/ui/next-solver/alias-bound-shadowed-by-env.rs`.
                let (mut candidates, _) = self
                    .assemble_and_evaluate_candidates(goal, AssembleCandidatesFrom::EnvAndBounds)?;
                debug!(?candidates);

                // If the trait goal has been proven by using the environment, we want to treat
                // aliases as rigid if there are no applicable projection bounds in the environment.
                if candidates.is_empty() {
                    return inject_normalize_to_rigid_candidate(self);
                }

                // If we're normalizing an GAT, we bail if using a where-bound would constrain
                // its generic arguments.
                if let Some(result) = inject_forced_ambiguity_candidate(self) {
                    return result;
                }

                // We still need to prefer where-bounds over alias-bounds however.
                // See `tests/ui/winnowing/norm-where-bound-gt-alias-bound.rs`.
                if candidates.iter().any(|c| matches!(c.source, CandidateSource::ParamEnv(_))) {
                    candidates.retain(|c| matches!(c.source, CandidateSource::ParamEnv(_)));
                }

                if let Some((response, _)) = self.try_merge_candidates(&candidates) {
                    Ok(response)
                } else {
                    self.flounder(&candidates).map_err(Into::into)
                }
            }
            TraitGoalProvenVia::Misc => {
                let (mut candidates, _) =
                    self.assemble_and_evaluate_candidates(goal, AssembleCandidatesFrom::All)?;

                // Prefer "orphaned" param-env normalization predicates, which are used
                // (for example, and ideally only) when proving item bounds for an impl.
                if candidates.iter().any(|c| matches!(c.source, CandidateSource::ParamEnv(_))) {
                    candidates.retain(|c| matches!(c.source, CandidateSource::ParamEnv(_)));
                }

                // We drop specialized impls to allow normalization via a final impl here. In case
                // the specializing impl has different inference constraints from the specialized
                // impl, proving the trait goal is already ambiguous, so we never get here. This
                // means we can just ignore inference constraints and don't have to special-case
                // constraining the normalized-to `term`.
                self.filter_specialized_impls(AllowInferenceConstraints::Yes, &mut candidates);
                if let Some((response, _)) = self.try_merge_candidates(&candidates) {
                    Ok(response)
                } else {
                    self.flounder(&candidates).map_err(Into::into)
                }
            }
        }
    }

    /// Compute whether a param-env assumption is global or non-global after normalizing it.
    ///
    /// This is necessary because, for example, given:
    ///
    /// ```ignore,rust
    /// where
    ///     T: Trait<Assoc = u32>,
    ///     i32: From<T::Assoc>,
    /// ```
    ///
    /// The `i32: From<T::Assoc>` bound is non-global before normalization, but is global after.
    /// Since the old trait solver normalized param-envs eagerly, we want to emulate this
    /// behavior lazily.
    fn characterize_param_env_assumption(
        &mut self,
        param_env: I::ParamEnv,
        assumption: I::Clause,
    ) -> Result<(CandidateSource<I>, Certainty), NoSolution> {
        // FIXME: This should be fixed, but it also requires changing the behavior
        // in the old solver which is currently relied on.
        if assumption.has_bound_vars() {
            return Ok((CandidateSource::ParamEnv(ParamEnvSource::NonGlobal), Certainty::Yes));
        }

        match assumption.visit_with(&mut FindParamInClause {
            ecx: self,
            param_env,
            universes: vec![],
            recursion_depth: 0,
        }) {
            ControlFlow::Break(Err(NoSolution)) => Err(NoSolution),
            ControlFlow::Break(Ok(certainty)) => {
                Ok((CandidateSource::ParamEnv(ParamEnvSource::NonGlobal), certainty))
            }
            ControlFlow::Continue(()) => {
                Ok((CandidateSource::ParamEnv(ParamEnvSource::Global), Certainty::Yes))
            }
        }
    }
}

struct FindParamInClause<'a, 'b, D: SolverDelegate<Interner = I>, I: Interner> {
    ecx: &'a mut EvalCtxt<'b, D>,
    param_env: I::ParamEnv,
    universes: Vec<Option<ty::UniverseIndex>>,
    recursion_depth: usize,
}

impl<D, I> TypeVisitor<I> for FindParamInClause<'_, '_, D, I>
where
    D: SolverDelegate<Interner = I>,
    I: Interner,
{
    // - `Continue(())`: no generic parameter was found, the type is global
    // - `Break(Ok(Certainty::Yes))`: a generic parameter was found, the type is non-global
    // - `Break(Ok(Certainty::Maybe(_)))`: the recursion limit reached, assume that the type is non-global
    // - `Break(Err(NoSolution))`: normalization failed
    type Result = ControlFlow<Result<Certainty, NoSolution>>;

    fn visit_binder<T: TypeVisitable<I>>(&mut self, t: &ty::Binder<I, T>) -> Self::Result {
        self.universes.push(None);
        t.super_visit_with(self)?;
        self.universes.pop();
        ControlFlow::Continue(())
    }

    fn visit_ty(&mut self, ty: I::Ty) -> Self::Result {
        let ty = self.ecx.replace_bound_vars(ty, &mut self.universes);
        let Ok(ty) = self.ecx.structurally_normalize_ty(self.param_env, ty) else {
            return ControlFlow::Break(Err(NoSolution));
        };

        match ty.kind() {
            ty::Placeholder(p) => {
                if p.universe() == ty::UniverseIndex::ROOT {
                    ControlFlow::Break(Ok(Certainty::Yes))
                } else {
                    ControlFlow::Continue(())
                }
            }
            ty::Infer(_) => ControlFlow::Break(Ok(Certainty::AMBIGUOUS)),
            _ if ty.has_type_flags(
                TypeFlags::HAS_PLACEHOLDER | TypeFlags::HAS_INFER | TypeFlags::HAS_ALIAS,
            ) =>
            {
                self.recursion_depth += 1;
                if self.recursion_depth > self.ecx.cx().recursion_limit() {
                    return ControlFlow::Break(Ok(Certainty::Maybe(MaybeInfo {
                        cause: MaybeCause::Overflow {
                            suggest_increasing_limit: true,
                            keep_constraints: false,
                        },
                        opaque_types_jank: OpaqueTypesJank::AllGood,
                        stalled_on_coroutines: StalledOnCoroutines::No,
                    })));
                }
                let result = ty.super_visit_with(self);
                self.recursion_depth -= 1;
                result
            }
            _ => ControlFlow::Continue(()),
        }
    }

    fn visit_const(&mut self, ct: I::Const) -> Self::Result {
        let ct = self.ecx.replace_bound_vars(ct, &mut self.universes);
        let Ok(ct) = self.ecx.structurally_normalize_const(self.param_env, ct) else {
            return ControlFlow::Break(Err(NoSolution));
        };

        match ct.kind() {
            ty::ConstKind::Placeholder(p) => {
                if p.universe() == ty::UniverseIndex::ROOT {
                    ControlFlow::Break(Ok(Certainty::Yes))
                } else {
                    ControlFlow::Continue(())
                }
            }
            ty::ConstKind::Infer(_) => ControlFlow::Break(Ok(Certainty::AMBIGUOUS)),
            _ if ct.has_type_flags(
                TypeFlags::HAS_PLACEHOLDER | TypeFlags::HAS_INFER | TypeFlags::HAS_ALIAS,
            ) =>
            {
                // FIXME(mgca): we should also check the recursion limit here
                ct.super_visit_with(self)
            }
            _ => ControlFlow::Continue(()),
        }
    }

    fn visit_region(&mut self, r: Region<I>) -> Self::Result {
        match self.ecx.eager_resolve_region(r).kind() {
            ty::ReStatic | ty::ReError(_) | ty::ReBound(..) => ControlFlow::Continue(()),
            ty::RePlaceholder(p) => {
                if p.universe() == ty::UniverseIndex::ROOT {
                    ControlFlow::Break(Ok(Certainty::Yes))
                } else {
                    ControlFlow::Continue(())
                }
            }
            ty::ReVar(_) => ControlFlow::Break(Ok(Certainty::Yes)),
            ty::ReErased | ty::ReEarlyParam(_) | ty::ReLateParam(_) => {
                unreachable!("unexpected region in param-env clause")
            }
        }
    }
}