outlives.rs

// Copyright 2012 The Rust Project Developers. See the COPYRIGHT
// file at the top-level directory of this distribution and at
// http://rust-lang.org/COPYRIGHT.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.

// The outlines relation `T: 'a` or `'a: 'b`. This code frequently
// refers to rules defined in RFC 1214 (`OutlivesFooBar`), so see that
// RFC for reference.

use ty::{self, Ty, TyCtxt, TypeFoldable};

#[derive(Debug)]
pub enum Component<'tcx> {
    Region(&'tcx ty::Region),
    Param(ty::ParamTy),
    UnresolvedInferenceVariable(ty::InferTy),

    // Projections like `T::Foo` are tricky because a constraint like
    // `T::Foo: 'a` can be satisfied in so many ways. There may be a
    // where-clause that says `T::Foo: 'a`, or the defining trait may
    // include a bound like `type Foo: 'static`, or -- in the most
    // conservative way -- we can prove that `T: 'a` (more generally,
    // that all components in the projection outlive `'a`). This code
    // is not in a position to judge which is the best technique, so
    // we just product the projection as a component and leave it to
    // the consumer to decide (but see `EscapingProjection` below).
    Projection(ty::ProjectionTy<'tcx>),

    // In the case where a projection has escaping regions -- meaning
    // regions bound within the type itself -- we always use
    // the most conservative rule, which requires that all components
    // outlive the bound. So for example if we had a type like this:
    //
    //     for<'a> Trait1<  <T as Trait2<'a,'b>>::Foo  >
    //                      ~~~~~~~~~~~~~~~~~~~~~~~~~
    //
    // then the inner projection (underlined) has an escaping region
    // `'a`. We consider that outer trait `'c` to meet a bound if `'b`
    // outlives `'b: 'c`, and we don't consider whether the trait
    // declares that `Foo: 'static` etc. Therefore, we just return the
    // free components of such a projection (in this case, `'b`).
    //
    // However, in the future, we may want to get smarter, and
    // actually return a "higher-ranked projection" here. Therefore,
    // we mark that these components are part of an escaping
    // projection, so that implied bounds code can avoid relying on
    // them. This gives us room to improve the regionck reasoning in
    // the future without breaking backwards compat.
    EscapingProjection(Vec<Component<'tcx>>),
}

impl<'a, 'gcx, 'tcx> TyCtxt<'a, 'gcx, 'tcx> {
    /// Returns all the things that must outlive `'a` for the condition
    /// `ty0: 'a` to hold. Note that `ty0` must be a **fully resolved type**.
    pub fn outlives_components(&self, ty0: Ty<'tcx>)
                               -> Vec<Component<'tcx>> {
        let mut components = vec![];
        self.compute_components(ty0, &mut components);
        debug!("components({:?}) = {:?}", ty0, components);
        components
    }

    fn compute_components(&self, ty: Ty<'tcx>, out: &mut Vec<Component<'tcx>>) {
        // Descend through the types, looking for the various "base"
        // components and collecting them into `out`. This is not written
        // with `collect()` because of the need to sometimes skip subtrees
        // in the `subtys` iterator (e.g., when encountering a
        // projection).
        match ty.sty {
            ty::TyClosure(def_id, ref substs) => {
                // FIXME(#27086). We do not accumulate from substs, since they
                // don't represent reachable data. This means that, in
                // practice, some of the lifetime parameters might not
                // be in scope when the body runs, so long as there is
                // no reachable data with that lifetime. For better or
                // worse, this is consistent with fn types, however,
                // which can also encapsulate data in this fashion
                // (though it's somewhat harder, and typically
                // requires virtual dispatch).
                //
                // Note that changing this (in a naive way, at least)
                // causes regressions for what appears to be perfectly
                // reasonable code like this:
                //
                // ```
                // fn foo<'a>(p: &Data<'a>) {
                //    bar(|q: &mut Parser| q.read_addr())
                // }
                // fn bar(p: Box<FnMut(&mut Parser)+'static>) {
                // }
                // ```
                //
                // Note that `p` (and `'a`) are not used in the
                // closure at all, but to meet the requirement that
                // the closure type `C: 'static` (so it can be coerced
                // to the object type), we get the requirement that
                // `'a: 'static` since `'a` appears in the closure
                // type `C`.
                //
                // A smarter fix might "prune" unused `func_substs` --
                // this would avoid breaking simple examples like
                // this, but would still break others (which might
                // indeed be invalid, depending on your POV). Pruning
                // would be a subtle process, since we have to see
                // what func/type parameters are used and unused,
                // taking into consideration UFCS and so forth.

                for upvar_ty in substs.upvar_tys(def_id, *self) {
                    self.compute_components(upvar_ty, out);
                }
            }

            // OutlivesTypeParameterEnv -- the actual checking that `X:'a`
            // is implied by the environment is done in regionck.
            ty::TyParam(p) => {
                out.push(Component::Param(p));
            }

            // For projections, we prefer to generate an obligation like
            // `<P0 as Trait<P1...Pn>>::Foo: 'a`, because this gives the
            // regionck more ways to prove that it holds. However,
            // regionck is not (at least currently) prepared to deal with
            // higher-ranked regions that may appear in the
            // trait-ref. Therefore, if we see any higher-ranke regions,
            // we simply fallback to the most restrictive rule, which
            // requires that `Pi: 'a` for all `i`.
            ty::TyProjection(ref data) => {
                if !data.has_escaping_regions() {
                    // best case: no escaping regions, so push the
                    // projection and skip the subtree (thus generating no
                    // constraints for Pi). This defers the choice between
                    // the rules OutlivesProjectionEnv,
                    // OutlivesProjectionTraitDef, and
                    // OutlivesProjectionComponents to regionck.
                    out.push(Component::Projection(*data));
                } else {
                    // fallback case: hard code
                    // OutlivesProjectionComponents.  Continue walking
                    // through and constrain Pi.
                    let subcomponents = self.capture_components(ty);
                    out.push(Component::EscapingProjection(subcomponents));
                }
            }

            // We assume that inference variables are fully resolved.
            // So, if we encounter an inference variable, just record
            // the unresolved variable as a component.
            ty::TyInfer(infer_ty) => {
                out.push(Component::UnresolvedInferenceVariable(infer_ty));
            }

            // Most types do not introduce any region binders, nor
            // involve any other subtle cases, and so the WF relation
            // simply constraints any regions referenced directly by
            // the type and then visits the types that are lexically
            // contained within. (The comments refer to relevant rules
            // from RFC1214.)
            ty::TyBool |            // OutlivesScalar
            ty::TyChar |            // OutlivesScalar
            ty::TyInt(..) |         // OutlivesScalar
            ty::TyUint(..) |        // OutlivesScalar
            ty::TyFloat(..) |       // OutlivesScalar
            ty::TyNever |           // ...
            ty::TyAdt(..) |         // OutlivesNominalType
            ty::TyAnon(..) |        // OutlivesNominalType (ish)
            ty::TyStr |             // OutlivesScalar (ish)
            ty::TyArray(..) |       // ...
            ty::TySlice(..) |       // ...
            ty::TyRawPtr(..) |      // ...
            ty::TyRef(..) |         // OutlivesReference
            ty::TyTuple(..) |       // ...
            ty::TyFnDef(..) |       // OutlivesFunction (*)
            ty::TyFnPtr(_) |        // OutlivesFunction (*)
            ty::TyDynamic(..) |       // OutlivesObject, OutlivesFragment (*)
            ty::TyError => {
                // (*) Bare functions and traits are both binders. In the
                // RFC, this means we would add the bound regions to the
                // "bound regions list".  In our representation, no such
                // list is maintained explicitly, because bound regions
                // themselves can be readily identified.

                push_region_constraints(out, ty.regions());
                for subty in ty.walk_shallow() {
                    self.compute_components(subty, out);
                }
            }
        }
    }

    fn capture_components(&self, ty: Ty<'tcx>) -> Vec<Component<'tcx>> {
        let mut temp = vec![];
        push_region_constraints(&mut temp, ty.regions());
        for subty in ty.walk_shallow() {
            self.compute_components(subty, &mut temp);
        }
        temp
    }
}

fn push_region_constraints<'tcx>(out: &mut Vec<Component<'tcx>>, regions: Vec<&'tcx ty::Region>) {
    for r in regions {
        if !r.is_bound() {
            out.push(Component::Region(r));
        }
    }
}