rustc_trait_selection/traits/

coherence.rs

//! See Rustc Dev Guide chapters on [trait-resolution] and [trait-specialization] for more info on
//! how this works.
//!
//! [trait-resolution]: https://rustc-dev-guide.rust-lang.org/traits/resolution.html
//! [trait-specialization]: https://rustc-dev-guide.rust-lang.org/traits/specialization.html

use std::fmt::Debug;

use rustc_data_structures::fx::FxIndexSet;
use rustc_errors::{Diag, EmissionGuarantee};
use rustc_hir::def::DefKind;
use rustc_hir::def_id::DefId;
use rustc_infer::infer::{DefineOpaqueTypes, InferCtxt, TyCtxtInferExt};
use rustc_infer::traits::PredicateObligations;
use rustc_middle::bug;
use rustc_middle::traits::query::NoSolution;
use rustc_middle::traits::solve::{CandidateSource, Certainty, Goal};
use rustc_middle::traits::specialization_graph::OverlapMode;
use rustc_middle::ty::fast_reject::DeepRejectCtxt;
use rustc_middle::ty::visit::{TypeSuperVisitable, TypeVisitable, TypeVisitableExt, TypeVisitor};
use rustc_middle::ty::{self, Ty, TyCtxt, TypingMode};
pub use rustc_next_trait_solver::coherence::*;
use rustc_next_trait_solver::solve::SolverDelegateEvalExt;
use rustc_span::symbol::sym;
use rustc_span::{DUMMY_SP, Span};
use tracing::{debug, instrument, warn};

use super::ObligationCtxt;
use crate::error_reporting::traits::suggest_new_overflow_limit;
use crate::infer::InferOk;
use crate::infer::outlives::env::OutlivesEnvironment;
use crate::solve::inspect::{InspectGoal, ProofTreeInferCtxtExt, ProofTreeVisitor};
use crate::solve::{SolverDelegate, deeply_normalize_for_diagnostics, inspect};
use crate::traits::query::evaluate_obligation::InferCtxtExt;
use crate::traits::select::IntercrateAmbiguityCause;
use crate::traits::{
    FulfillmentErrorCode, NormalizeExt, Obligation, ObligationCause, PredicateObligation,
    SelectionContext, SkipLeakCheck, util,
};

pub struct OverlapResult<'tcx> {
    pub impl_header: ty::ImplHeader<'tcx>,
    pub intercrate_ambiguity_causes: FxIndexSet<IntercrateAmbiguityCause<'tcx>>,

    /// `true` if the overlap might've been permitted before the shift
    /// to universes.
    pub involves_placeholder: bool,

    /// Used in the new solver to suggest increasing the recursion limit.
    pub overflowing_predicates: Vec<ty::Predicate<'tcx>>,
}

pub fn add_placeholder_note<G: EmissionGuarantee>(err: &mut Diag<'_, G>) {
    err.note(
        "this behavior recently changed as a result of a bug fix; \
         see rust-lang/rust#56105 for details",
    );
}

pub(crate) fn suggest_increasing_recursion_limit<'tcx, G: EmissionGuarantee>(
    tcx: TyCtxt<'tcx>,
    err: &mut Diag<'_, G>,
    overflowing_predicates: &[ty::Predicate<'tcx>],
) {
    for pred in overflowing_predicates {
        err.note(format!("overflow evaluating the requirement `{}`", pred));
    }

    suggest_new_overflow_limit(tcx, err);
}

#[derive(Debug, Clone, Copy)]
enum TrackAmbiguityCauses {
    Yes,
    No,
}

impl TrackAmbiguityCauses {
    fn is_yes(self) -> bool {
        match self {
            TrackAmbiguityCauses::Yes => true,
            TrackAmbiguityCauses::No => false,
        }
    }
}

/// If there are types that satisfy both impls, returns `Some`
/// with a suitably-freshened `ImplHeader` with those types
/// instantiated. Otherwise, returns `None`.
#[instrument(skip(tcx, skip_leak_check), level = "debug")]
pub fn overlapping_impls(
    tcx: TyCtxt<'_>,
    impl1_def_id: DefId,
    impl2_def_id: DefId,
    skip_leak_check: SkipLeakCheck,
    overlap_mode: OverlapMode,
) -> Option<OverlapResult<'_>> {
    // Before doing expensive operations like entering an inference context, do
    // a quick check via fast_reject to tell if the impl headers could possibly
    // unify.
    let drcx = DeepRejectCtxt::relate_infer_infer(tcx);
    let impl1_ref = tcx.impl_trait_ref(impl1_def_id);
    let impl2_ref = tcx.impl_trait_ref(impl2_def_id);
    let may_overlap = match (impl1_ref, impl2_ref) {
        (Some(a), Some(b)) => drcx.args_may_unify(a.skip_binder().args, b.skip_binder().args),
        (None, None) => {
            let self_ty1 = tcx.type_of(impl1_def_id).skip_binder();
            let self_ty2 = tcx.type_of(impl2_def_id).skip_binder();
            drcx.types_may_unify(self_ty1, self_ty2)
        }
        _ => bug!("unexpected impls: {impl1_def_id:?} {impl2_def_id:?}"),
    };

    if !may_overlap {
        // Some types involved are definitely different, so the impls couldn't possibly overlap.
        debug!("overlapping_impls: fast_reject early-exit");
        return None;
    }

    let _overlap_with_bad_diagnostics = overlap(
        tcx,
        TrackAmbiguityCauses::No,
        skip_leak_check,
        impl1_def_id,
        impl2_def_id,
        overlap_mode,
    )?;

    // In the case where we detect an error, run the check again, but
    // this time tracking intercrate ambiguity causes for better
    // diagnostics. (These take time and can lead to false errors.)
    let overlap = overlap(
        tcx,
        TrackAmbiguityCauses::Yes,
        skip_leak_check,
        impl1_def_id,
        impl2_def_id,
        overlap_mode,
    )
    .unwrap();
    Some(overlap)
}

fn fresh_impl_header<'tcx>(infcx: &InferCtxt<'tcx>, impl_def_id: DefId) -> ty::ImplHeader<'tcx> {
    let tcx = infcx.tcx;
    let impl_args = infcx.fresh_args_for_item(DUMMY_SP, impl_def_id);

    ty::ImplHeader {
        impl_def_id,
        impl_args,
        self_ty: tcx.type_of(impl_def_id).instantiate(tcx, impl_args),
        trait_ref: tcx.impl_trait_ref(impl_def_id).map(|i| i.instantiate(tcx, impl_args)),
        predicates: tcx
            .predicates_of(impl_def_id)
            .instantiate(tcx, impl_args)
            .iter()
            .map(|(c, _)| c.as_predicate())
            .collect(),
    }
}

fn fresh_impl_header_normalized<'tcx>(
    infcx: &InferCtxt<'tcx>,
    param_env: ty::ParamEnv<'tcx>,
    impl_def_id: DefId,
) -> ty::ImplHeader<'tcx> {
    let header = fresh_impl_header(infcx, impl_def_id);

    let InferOk { value: mut header, obligations } =
        infcx.at(&ObligationCause::dummy(), param_env).normalize(header);

    header.predicates.extend(obligations.into_iter().map(|o| o.predicate));
    header
}

/// Can both impl `a` and impl `b` be satisfied by a common type (including
/// where-clauses)? If so, returns an `ImplHeader` that unifies the two impls.
#[instrument(level = "debug", skip(tcx))]
fn overlap<'tcx>(
    tcx: TyCtxt<'tcx>,
    track_ambiguity_causes: TrackAmbiguityCauses,
    skip_leak_check: SkipLeakCheck,
    impl1_def_id: DefId,
    impl2_def_id: DefId,
    overlap_mode: OverlapMode,
) -> Option<OverlapResult<'tcx>> {
    if overlap_mode.use_negative_impl() {
        if impl_intersection_has_negative_obligation(tcx, impl1_def_id, impl2_def_id)
            || impl_intersection_has_negative_obligation(tcx, impl2_def_id, impl1_def_id)
        {
            return None;
        }
    }

    let infcx = tcx
        .infer_ctxt()
        .skip_leak_check(skip_leak_check.is_yes())
        .with_next_trait_solver(tcx.next_trait_solver_in_coherence())
        .build(TypingMode::Coherence);
    let selcx = &mut SelectionContext::new(&infcx);
    if track_ambiguity_causes.is_yes() {
        selcx.enable_tracking_intercrate_ambiguity_causes();
    }

    // For the purposes of this check, we don't bring any placeholder
    // types into scope; instead, we replace the generic types with
    // fresh type variables, and hence we do our evaluations in an
    // empty environment.
    let param_env = ty::ParamEnv::empty();

    let impl1_header = fresh_impl_header_normalized(selcx.infcx, param_env, impl1_def_id);
    let impl2_header = fresh_impl_header_normalized(selcx.infcx, param_env, impl2_def_id);

    // Equate the headers to find their intersection (the general type, with infer vars,
    // that may apply both impls).
    let mut obligations =
        equate_impl_headers(selcx.infcx, param_env, &impl1_header, &impl2_header)?;
    debug!("overlap: unification check succeeded");

    obligations.extend(
        [&impl1_header.predicates, &impl2_header.predicates].into_iter().flatten().map(
            |&predicate| Obligation::new(infcx.tcx, ObligationCause::dummy(), param_env, predicate),
        ),
    );

    let mut overflowing_predicates = Vec::new();
    if overlap_mode.use_implicit_negative() {
        match impl_intersection_has_impossible_obligation(selcx, &obligations) {
            IntersectionHasImpossibleObligations::Yes => return None,
            IntersectionHasImpossibleObligations::No { overflowing_predicates: p } => {
                overflowing_predicates = p
            }
        }
    }

    // We toggle the `leak_check` by using `skip_leak_check` when constructing the
    // inference context, so this may be a noop.
    if infcx.leak_check(ty::UniverseIndex::ROOT, None).is_err() {
        debug!("overlap: leak check failed");
        return None;
    }

    let intercrate_ambiguity_causes = if !overlap_mode.use_implicit_negative() {
        Default::default()
    } else if infcx.next_trait_solver() {
        compute_intercrate_ambiguity_causes(&infcx, &obligations)
    } else {
        selcx.take_intercrate_ambiguity_causes()
    };

    debug!("overlap: intercrate_ambiguity_causes={:#?}", intercrate_ambiguity_causes);
    let involves_placeholder = infcx
        .inner
        .borrow_mut()
        .unwrap_region_constraints()
        .data()
        .constraints
        .iter()
        .any(|c| c.0.involves_placeholders());

    let mut impl_header = infcx.resolve_vars_if_possible(impl1_header);

    // Deeply normalize the impl header for diagnostics, ignoring any errors if this fails.
    if infcx.next_trait_solver() {
        impl_header = deeply_normalize_for_diagnostics(&infcx, param_env, impl_header);
    }

    Some(OverlapResult {
        impl_header,
        intercrate_ambiguity_causes,
        involves_placeholder,
        overflowing_predicates,
    })
}

#[instrument(level = "debug", skip(infcx), ret)]
fn equate_impl_headers<'tcx>(
    infcx: &InferCtxt<'tcx>,
    param_env: ty::ParamEnv<'tcx>,
    impl1: &ty::ImplHeader<'tcx>,
    impl2: &ty::ImplHeader<'tcx>,
) -> Option<PredicateObligations<'tcx>> {
    let result =
        match (impl1.trait_ref, impl2.trait_ref) {
            (Some(impl1_ref), Some(impl2_ref)) => infcx
                .at(&ObligationCause::dummy(), param_env)
                .eq(DefineOpaqueTypes::Yes, impl1_ref, impl2_ref),
            (None, None) => infcx.at(&ObligationCause::dummy(), param_env).eq(
                DefineOpaqueTypes::Yes,
                impl1.self_ty,
                impl2.self_ty,
            ),
            _ => bug!("equate_impl_headers given mismatched impl kinds"),
        };

    result.map(|infer_ok| infer_ok.obligations).ok()
}

/// The result of [fn impl_intersection_has_impossible_obligation].
#[derive(Debug)]
enum IntersectionHasImpossibleObligations<'tcx> {
    Yes,
    No {
        /// With `-Znext-solver=coherence`, some obligations may
        /// fail if only the user increased the recursion limit.
        ///
        /// We return those obligations here and mention them in the
        /// error message.
        overflowing_predicates: Vec<ty::Predicate<'tcx>>,
    },
}

/// Check if both impls can be satisfied by a common type by considering whether
/// any of either impl's obligations is not known to hold.
///
/// For example, given these two impls:
///     `impl From<MyLocalType> for Box<dyn Error>` (in my crate)
///     `impl<E> From<E> for Box<dyn Error> where E: Error` (in libstd)
///
/// After replacing both impl headers with inference vars (which happens before
/// this function is called), we get:
///     `Box<dyn Error>: From<MyLocalType>`
///     `Box<dyn Error>: From<?E>`
///
/// This gives us `?E = MyLocalType`. We then certainly know that `MyLocalType: Error`
/// never holds in intercrate mode since a local impl does not exist, and a
/// downstream impl cannot be added -- therefore can consider the intersection
/// of the two impls above to be empty.
///
/// Importantly, this works even if there isn't a `impl !Error for MyLocalType`.
#[instrument(level = "debug", skip(selcx), ret)]
fn impl_intersection_has_impossible_obligation<'a, 'cx, 'tcx>(
    selcx: &mut SelectionContext<'cx, 'tcx>,
    obligations: &'a [PredicateObligation<'tcx>],
) -> IntersectionHasImpossibleObligations<'tcx> {
    let infcx = selcx.infcx;

    if infcx.next_trait_solver() {
        // A fast path optimization, try evaluating all goals with
        // a very low recursion depth and bail if any of them don't
        // hold.
        if !obligations.iter().all(|o| {
            <&SolverDelegate<'tcx>>::from(infcx)
                .root_goal_may_hold_with_depth(8, Goal::new(infcx.tcx, o.param_env, o.predicate))
        }) {
            return IntersectionHasImpossibleObligations::Yes;
        }

        let ocx = ObligationCtxt::new_with_diagnostics(infcx);
        ocx.register_obligations(obligations.iter().cloned());
        let errors_and_ambiguities = ocx.select_all_or_error();
        // We only care about the obligations that are *definitely* true errors.
        // Ambiguities do not prove the disjointness of two impls.
        let (errors, ambiguities): (Vec<_>, Vec<_>) =
            errors_and_ambiguities.into_iter().partition(|error| error.is_true_error());

        if errors.is_empty() {
            IntersectionHasImpossibleObligations::No {
                overflowing_predicates: ambiguities
                    .into_iter()
                    .filter(|error| {
                        matches!(error.code, FulfillmentErrorCode::Ambiguity {
                            overflow: Some(true)
                        })
                    })
                    .map(|e| infcx.resolve_vars_if_possible(e.obligation.predicate))
                    .collect(),
            }
        } else {
            IntersectionHasImpossibleObligations::Yes
        }
    } else {
        for obligation in obligations {
            // We use `evaluate_root_obligation` to correctly track intercrate
            // ambiguity clauses.
            let evaluation_result = selcx.evaluate_root_obligation(obligation);

            match evaluation_result {
                Ok(result) => {
                    if !result.may_apply() {
                        return IntersectionHasImpossibleObligations::Yes;
                    }
                }
                // If overflow occurs, we need to conservatively treat the goal as possibly holding,
                // since there can be instantiations of this goal that don't overflow and result in
                // success. While this isn't much of a problem in the old solver, since we treat overflow
                // fatally, this still can be encountered: <https://github.com/rust-lang/rust/issues/105231>.
                Err(_overflow) => {}
            }
        }

        IntersectionHasImpossibleObligations::No { overflowing_predicates: Vec::new() }
    }
}

/// Check if both impls can be satisfied by a common type by considering whether
/// any of first impl's obligations is known not to hold *via a negative predicate*.
///
/// For example, given these two impls:
///     `struct MyCustomBox<T: ?Sized>(Box<T>);`
///     `impl From<&str> for MyCustomBox<dyn Error>` (in my crate)
///     `impl<E> From<E> for MyCustomBox<dyn Error> where E: Error` (in my crate)
///
/// After replacing the second impl's header with inference vars, we get:
///     `MyCustomBox<dyn Error>: From<&str>`
///     `MyCustomBox<dyn Error>: From<?E>`
///
/// This gives us `?E = &str`. We then try to prove the first impl's predicates
/// after negating, giving us `&str: !Error`. This is a negative impl provided by
/// libstd, and therefore we can guarantee for certain that libstd will never add
/// a positive impl for `&str: Error` (without it being a breaking change).
fn impl_intersection_has_negative_obligation(
    tcx: TyCtxt<'_>,
    impl1_def_id: DefId,
    impl2_def_id: DefId,
) -> bool {
    debug!("negative_impl(impl1_def_id={:?}, impl2_def_id={:?})", impl1_def_id, impl2_def_id);

    // N.B. We need to unify impl headers *with* intercrate mode, even if proving negative predicates
    // do not need intercrate mode enabled.
    let ref infcx = tcx.infer_ctxt().with_next_trait_solver(true).build(TypingMode::Coherence);
    let root_universe = infcx.universe();
    assert_eq!(root_universe, ty::UniverseIndex::ROOT);

    let impl1_header = fresh_impl_header(infcx, impl1_def_id);
    let param_env =
        ty::EarlyBinder::bind(tcx.param_env(impl1_def_id)).instantiate(tcx, impl1_header.impl_args);

    let impl2_header = fresh_impl_header(infcx, impl2_def_id);

    // Equate the headers to find their intersection (the general type, with infer vars,
    // that may apply both impls).
    let Some(equate_obligations) =
        equate_impl_headers(infcx, param_env, &impl1_header, &impl2_header)
    else {
        return false;
    };

    // FIXME(with_negative_coherence): the infcx has constraints from equating
    // the impl headers. We should use these constraints as assumptions, not as
    // requirements, when proving the negated where clauses below.
    drop(equate_obligations);
    drop(infcx.take_registered_region_obligations());
    drop(infcx.take_and_reset_region_constraints());

    plug_infer_with_placeholders(
        infcx,
        root_universe,
        (impl1_header.impl_args, impl2_header.impl_args),
    );
    let param_env = infcx.resolve_vars_if_possible(param_env);

    util::elaborate(tcx, tcx.predicates_of(impl2_def_id).instantiate(tcx, impl2_header.impl_args))
        .any(|(clause, _)| try_prove_negated_where_clause(infcx, clause, param_env))
}

fn plug_infer_with_placeholders<'tcx>(
    infcx: &InferCtxt<'tcx>,
    universe: ty::UniverseIndex,
    value: impl TypeVisitable<TyCtxt<'tcx>>,
) {
    struct PlugInferWithPlaceholder<'a, 'tcx> {
        infcx: &'a InferCtxt<'tcx>,
        universe: ty::UniverseIndex,
        var: ty::BoundVar,
    }

    impl<'tcx> PlugInferWithPlaceholder<'_, 'tcx> {
        fn next_var(&mut self) -> ty::BoundVar {
            let var = self.var;
            self.var = self.var + 1;
            var
        }
    }

    impl<'tcx> TypeVisitor<TyCtxt<'tcx>> for PlugInferWithPlaceholder<'_, 'tcx> {
        fn visit_ty(&mut self, ty: Ty<'tcx>) {
            let ty = self.infcx.shallow_resolve(ty);
            if ty.is_ty_var() {
                let Ok(InferOk { value: (), obligations }) =
                    self.infcx.at(&ObligationCause::dummy(), ty::ParamEnv::empty()).eq(
                        // Comparing against a type variable never registers hidden types anyway
                        DefineOpaqueTypes::Yes,
                        ty,
                        Ty::new_placeholder(self.infcx.tcx, ty::Placeholder {
                            universe: self.universe,
                            bound: ty::BoundTy {
                                var: self.next_var(),
                                kind: ty::BoundTyKind::Anon,
                            },
                        }),
                    )
                else {
                    bug!("we always expect to be able to plug an infer var with placeholder")
                };
                assert_eq!(obligations.len(), 0);
            } else {
                ty.super_visit_with(self);
            }
        }

        fn visit_const(&mut self, ct: ty::Const<'tcx>) {
            let ct = self.infcx.shallow_resolve_const(ct);
            if ct.is_ct_infer() {
                let Ok(InferOk { value: (), obligations }) =
                    self.infcx.at(&ObligationCause::dummy(), ty::ParamEnv::empty()).eq(
                        // The types of the constants are the same, so there is no hidden type
                        // registration happening anyway.
                        DefineOpaqueTypes::Yes,
                        ct,
                        ty::Const::new_placeholder(self.infcx.tcx, ty::Placeholder {
                            universe: self.universe,
                            bound: self.next_var(),
                        }),
                    )
                else {
                    bug!("we always expect to be able to plug an infer var with placeholder")
                };
                assert_eq!(obligations.len(), 0);
            } else {
                ct.super_visit_with(self);
            }
        }

        fn visit_region(&mut self, r: ty::Region<'tcx>) {
            if let ty::ReVar(vid) = *r {
                let r = self
                    .infcx
                    .inner
                    .borrow_mut()
                    .unwrap_region_constraints()
                    .opportunistic_resolve_var(self.infcx.tcx, vid);
                if r.is_var() {
                    let Ok(InferOk { value: (), obligations }) =
                        self.infcx.at(&ObligationCause::dummy(), ty::ParamEnv::empty()).eq(
                            // Lifetimes don't contain opaque types (or any types for that matter).
                            DefineOpaqueTypes::Yes,
                            r,
                            ty::Region::new_placeholder(self.infcx.tcx, ty::Placeholder {
                                universe: self.universe,
                                bound: ty::BoundRegion {
                                    var: self.next_var(),
                                    kind: ty::BoundRegionKind::Anon,
                                },
                            }),
                        )
                    else {
                        bug!("we always expect to be able to plug an infer var with placeholder")
                    };
                    assert_eq!(obligations.len(), 0);
                }
            }
        }
    }

    value.visit_with(&mut PlugInferWithPlaceholder { infcx, universe, var: ty::BoundVar::ZERO });
}

fn try_prove_negated_where_clause<'tcx>(
    root_infcx: &InferCtxt<'tcx>,
    clause: ty::Clause<'tcx>,
    param_env: ty::ParamEnv<'tcx>,
) -> bool {
    let Some(negative_predicate) = clause.as_predicate().flip_polarity(root_infcx.tcx) else {
        return false;
    };

    // N.B. We don't need to use intercrate mode here because we're trying to prove
    // the *existence* of a negative goal, not the non-existence of a positive goal.
    // Without this, we over-eagerly register coherence ambiguity candidates when
    // impl candidates do exist.
    // FIXME(#132279): `TypingMode::non_body_analysis` is a bit questionable here as it
    // would cause us to reveal opaque types to leak their auto traits.
    let ref infcx = root_infcx.fork_with_typing_mode(TypingMode::non_body_analysis());
    let ocx = ObligationCtxt::new(infcx);
    ocx.register_obligation(Obligation::new(
        infcx.tcx,
        ObligationCause::dummy(),
        param_env,
        negative_predicate,
    ));
    if !ocx.select_all_or_error().is_empty() {
        return false;
    }

    // FIXME: We could use the assumed_wf_types from both impls, I think,
    // if that wasn't implemented just for LocalDefId, and we'd need to do
    // the normalization ourselves since this is totally fallible...
    let outlives_env = OutlivesEnvironment::new(param_env);
    let errors = ocx.resolve_regions(&outlives_env);
    if !errors.is_empty() {
        return false;
    }

    true
}

/// Compute the `intercrate_ambiguity_causes` for the new solver using
/// "proof trees".
///
/// This is a bit scuffed but seems to be good enough, at least
/// when looking at UI tests. Given that it is only used to improve
/// diagnostics this is good enough. We can always improve it once there
/// are test cases where it is currently not enough.
fn compute_intercrate_ambiguity_causes<'tcx>(
    infcx: &InferCtxt<'tcx>,
    obligations: &[PredicateObligation<'tcx>],
) -> FxIndexSet<IntercrateAmbiguityCause<'tcx>> {
    let mut causes: FxIndexSet<IntercrateAmbiguityCause<'tcx>> = Default::default();

    for obligation in obligations {
        search_ambiguity_causes(infcx, obligation.clone().into(), &mut causes);
    }

    causes
}

struct AmbiguityCausesVisitor<'a, 'tcx> {
    causes: &'a mut FxIndexSet<IntercrateAmbiguityCause<'tcx>>,
}

impl<'a, 'tcx> ProofTreeVisitor<'tcx> for AmbiguityCausesVisitor<'a, 'tcx> {
    fn span(&self) -> Span {
        DUMMY_SP
    }

    fn visit_goal(&mut self, goal: &InspectGoal<'_, 'tcx>) {
        let infcx = goal.infcx();
        for cand in goal.candidates() {
            cand.visit_nested_in_probe(self);
        }
        // When searching for intercrate ambiguity causes, we only need to look
        // at ambiguous goals, as for others the coherence unknowable candidate
        // was irrelevant.
        match goal.result() {
            Ok(Certainty::Yes) | Err(NoSolution) => return,
            Ok(Certainty::Maybe(_)) => {}
        }

        // For bound predicates we simply call `infcx.enter_forall`
        // and then prove the resulting predicate as a nested goal.
        let Goal { param_env, predicate } = goal.goal();
        let trait_ref = match predicate.kind().no_bound_vars() {
            Some(ty::PredicateKind::Clause(ty::ClauseKind::Trait(tr))) => tr.trait_ref,
            Some(ty::PredicateKind::Clause(ty::ClauseKind::Projection(proj)))
                if matches!(
                    infcx.tcx.def_kind(proj.projection_term.def_id),
                    DefKind::AssocTy | DefKind::AssocConst
                ) =>
            {
                proj.projection_term.trait_ref(infcx.tcx)
            }
            _ => return,
        };

        if trait_ref.references_error() {
            return;
        }

        let mut candidates = goal.candidates();
        for cand in goal.candidates() {
            if let inspect::ProbeKind::TraitCandidate {
                source: CandidateSource::Impl(def_id),
                result: Ok(_),
            } = cand.kind()
            {
                if let ty::ImplPolarity::Reservation = infcx.tcx.impl_polarity(def_id) {
                    let message = infcx
                        .tcx
                        .get_attr(def_id, sym::rustc_reservation_impl)
                        .and_then(|a| a.value_str());
                    if let Some(message) = message {
                        self.causes.insert(IntercrateAmbiguityCause::ReservationImpl { message });
                    }
                }
            }
        }

        // We also look for unknowable candidates. In case a goal is unknowable, there's
        // always exactly 1 candidate.
        let Some(cand) = candidates.pop() else {
            return;
        };

        let inspect::ProbeKind::TraitCandidate {
            source: CandidateSource::CoherenceUnknowable,
            result: Ok(_),
        } = cand.kind()
        else {
            return;
        };

        let lazily_normalize_ty = |mut ty: Ty<'tcx>| {
            if matches!(ty.kind(), ty::Alias(..)) {
                let ocx = ObligationCtxt::new(infcx);
                ty = ocx
                    .structurally_normalize(&ObligationCause::dummy(), param_env, ty)
                    .map_err(|_| ())?;
                if !ocx.select_where_possible().is_empty() {
                    return Err(());
                }
            }
            Ok(ty)
        };

        infcx.probe(|_| {
            let conflict = match trait_ref_is_knowable(infcx, trait_ref, lazily_normalize_ty) {
                Err(()) => return,
                Ok(Ok(())) => {
                    warn!("expected an unknowable trait ref: {trait_ref:?}");
                    return;
                }
                Ok(Err(conflict)) => conflict,
            };

            // It is only relevant that a goal is unknowable if it would have otherwise
            // failed.
            // FIXME(#132279): Forking with `TypingMode::non_body_analysis` is a bit questionable
            // as it would allow us to reveal opaque types, potentially causing unexpected
            // cycles.
            let non_intercrate_infcx = infcx.fork_with_typing_mode(TypingMode::non_body_analysis());
            if non_intercrate_infcx.predicate_may_hold(&Obligation::new(
                infcx.tcx,
                ObligationCause::dummy(),
                param_env,
                predicate,
            )) {
                return;
            }

            // Normalize the trait ref for diagnostics, ignoring any errors if this fails.
            let trait_ref = deeply_normalize_for_diagnostics(infcx, param_env, trait_ref);
            let self_ty = trait_ref.self_ty();
            let self_ty = self_ty.has_concrete_skeleton().then(|| self_ty);
            self.causes.insert(match conflict {
                Conflict::Upstream => {
                    IntercrateAmbiguityCause::UpstreamCrateUpdate { trait_ref, self_ty }
                }
                Conflict::Downstream => {
                    IntercrateAmbiguityCause::DownstreamCrate { trait_ref, self_ty }
                }
            });
        });
    }
}

fn search_ambiguity_causes<'tcx>(
    infcx: &InferCtxt<'tcx>,
    goal: Goal<'tcx, ty::Predicate<'tcx>>,
    causes: &mut FxIndexSet<IntercrateAmbiguityCause<'tcx>>,
) {
    infcx.probe(|_| infcx.visit_proof_tree(goal, &mut AmbiguityCausesVisitor { causes }));
}