factor variances into a proper query

There are now two queries: crate and item. The crate one computes the
variance of all items in the crate; it is sort of an implementation
detail, and not meant to be used. The item one reads from the crate one,
synthesizing correct deps in lieu of the red-green algorithm.

At the same time, remove the `variance_computed` flag, which was a
horrible hack used to force invariance early on (e.g. when type-checking
constants). This is only needed because of trait applications, and
traits are always invariant anyway. Therefore, we now change to take
advantage of the query system:

- When asked to compute variances for a trait, just return a vector
  saying 'all invariant'.
- Remove the corresponding "inferreds" from traits, and tweak the
  constraint generation code to understand that traits are always
  inferred.

Author: nikomatsakis

src/librustc/dep_graph/dep_node.rs

@@ -81,6 +81,7 @@ pub enum DepNode<D: Clone + Debug> {
     TransCrateItem(D),
     TransInlinedItem(D),
     TransWriteMetadata,
+    CrateVariances,
 
     // Nodes representing bits of computed IR in the tcx. Each shared
     // table in the tcx (or elsewhere) maps to one of these
@@ -89,6 +90,8 @@ pub enum DepNode<D: Clone + Debug> {
     // predicates for an item wind up in `ItemSignature`).
     AssociatedItems(D),
     ItemSignature(D),
+    ItemVarianceConstraints(D),
+    ItemVariances(D),
     IsForeignItem(D),
     TypeParamPredicates((D, D)),
     SizedConstraint(D),
@@ -199,6 +202,7 @@ impl<D: Clone + Debug> DepNode<D> {
             MirKrate => Some(MirKrate),
             TypeckBodiesKrate => Some(TypeckBodiesKrate),
             Coherence => Some(Coherence),
+            CrateVariances => Some(CrateVariances),
             Resolve => Some(Resolve),
             Variance => Some(Variance),
             PrivacyAccessLevels(k) => Some(PrivacyAccessLevels(k)),
@@ -230,6 +234,8 @@ impl<D: Clone + Debug> DepNode<D> {
             TransInlinedItem(ref d) => op(d).map(TransInlinedItem),
             AssociatedItems(ref d) => op(d).map(AssociatedItems),
             ItemSignature(ref d) => op(d).map(ItemSignature),
+            ItemVariances(ref d) => op(d).map(ItemVariances),
+            ItemVarianceConstraints(ref d) => op(d).map(ItemVarianceConstraints),
             IsForeignItem(ref d) => op(d).map(IsForeignItem),
             TypeParamPredicates((ref item, ref param)) => {
                 Some(TypeParamPredicates((try_opt!(op(item)), try_opt!(op(param)))))

src/librustc/ty/context.rs

@@ -468,9 +468,6 @@ pub struct GlobalCtxt<'tcx> {
 
     pub lang_items: middle::lang_items::LanguageItems,
 
-    /// True if the variance has been computed yet; false otherwise.
-    pub variance_computed: Cell<bool>,
-
     /// Set of used unsafe nodes (functions or blocks). Unsafe nodes not
     /// present in this set can be warned about.
     pub used_unsafe: RefCell<NodeSet>,
@@ -753,7 +750,6 @@ impl<'a, 'gcx, 'tcx> TyCtxt<'a, 'gcx, 'tcx> {
             dep_graph: dep_graph.clone(),
             types: common_types,
             named_region_map: named_region_map,
-            variance_computed: Cell::new(false),
             trait_map: resolutions.trait_map,
             export_map: resolutions.export_map,
             fulfilled_predicates: RefCell::new(fulfilled_predicates),

src/librustc/ty/maps.rs

@@ -265,6 +265,12 @@ impl<'tcx> QueryDescription for queries::crate_inherent_impls_overlap_check<'tcx
     }
 }
 
+impl<'tcx> QueryDescription for queries::crate_variances<'tcx> {
+    fn describe(_tcx: TyCtxt, _: CrateNum) -> String {
+        format!("computing the variances for items in this crate")
+    }
+}
+
 impl<'tcx> QueryDescription for queries::mir_shims<'tcx> {
     fn describe(tcx: TyCtxt, def: ty::InstanceDef<'tcx>) -> String {
         format!("generating MIR shim for `{}`",
@@ -673,9 +679,13 @@ define_maps! { <'tcx>
     /// True if this is a foreign item (i.e., linked via `extern { ... }`).
     [] is_foreign_item: IsForeignItem(DefId) -> bool,
 
+    /// Get a map with the variance of every item; use `item_variance`
+    /// instead.
+    [] crate_variances: crate_variances(CrateNum) -> Rc<ty::CrateVariancesMap>,
+
     /// Maps from def-id of a type or region parameter to its
     /// (inferred) variance.
-    [pub] variances_of: ItemSignature(DefId) -> Rc<Vec<ty::Variance>>,
+    [pub] variances_of: ItemVariances(DefId) -> Rc<Vec<ty::Variance>>,
 
     /// Maps from an impl/trait def-id to a list of the def-ids of its items
     [] associated_item_def_ids: AssociatedItemDefIds(DefId) -> Rc<Vec<DefId>>,
@@ -810,3 +820,7 @@ fn const_eval_dep_node((def_id, _): (DefId, &Substs)) -> DepNode<DefId> {
 fn mir_keys(_: CrateNum) -> DepNode<DefId> {
     DepNode::MirKeys
 }
+
+fn crate_variances(_: CrateNum) -> DepNode<DefId> {
+    DepNode::CrateVariances
+}

src/librustc/ty/mod.rs

@@ -55,6 +55,7 @@ use rustc_const_math::ConstInt;
 use rustc_data_structures::accumulate_vec::IntoIter as AccIntoIter;
 use rustc_data_structures::stable_hasher::{StableHasher, StableHasherResult,
                                            HashStable};
+use rustc_data_structures::transitive_relation::TransitiveRelation;
 
 use hir;
 use hir::itemlikevisit::ItemLikeVisitor;
@@ -309,6 +310,27 @@ pub enum Variance {
     Bivariant,      // T<A> <: T<B>            -- e.g., unused type parameter
 }
 
+/// The crate variances map is computed during typeck and contains the
+/// variance of every item in the local crate. You should not use it
+/// directly, because to do so will make your pass dependent on the
+/// HIR of every item in the local crate. Instead, use
+/// `tcx.item_variances()` to get the variance for a *particular*
+/// item.
+pub struct CrateVariancesMap {
+    /// This relation tracks the dependencies between the variance of
+    /// various items. In particular, if `a < b`, then the variance of
+    /// `a` depends on the sources of `b`.
+    pub dependencies: TransitiveRelation<DefId>,
+
+    /// For each item with generics, maps to a vector of the variance
+    /// of its generics.  If an item has no generics, it will have no
+    /// entry.
+    pub variances: FxHashMap<DefId, Rc<Vec<ty::Variance>>>,
+
+    /// An empty vector, useful for cloning.
+    pub empty_variance: Rc<Vec<ty::Variance>>,
+}
+
 #[derive(Clone, Copy, Debug, RustcDecodable, RustcEncodable)]
 pub struct MethodCallee<'tcx> {
     /// Impl method ID, for inherent methods, or trait method ID, otherwise.

src/librustc/ty/relate.rs

@@ -124,14 +124,8 @@ fn relate_item_substs<'a, 'gcx, 'tcx, R>(relation: &mut R,
            a_subst,
            b_subst);
 
-    let variances;
-    let opt_variances = if relation.tcx().variance_computed.get() {
-        variances = relation.tcx().variances_of(item_def_id);
-        Some(&*variances)
-    } else {
-        None
-    };
-    relate_substs(relation, opt_variances, a_subst, b_subst)
+    let opt_variances = relation.tcx().variances_of(item_def_id);
+    relate_substs(relation, Some(&opt_variances), a_subst, b_subst)
 }
 
 pub fn relate_substs<'a, 'gcx, 'tcx, R>(relation: &mut R,

src/librustc_data_structures/transitive_relation.rs

@@ -134,6 +134,20 @@ impl<T: Clone + Debug + Eq + Hash + Clone> TransitiveRelation<T> {
         }
     }
 
+    /// Returns a vector of all things less than `a`.
+    ///
+    /// Really this probably ought to be `impl Iterator<Item=&T>`, but
+    /// I'm too lazy to make that work, and -- given the caching
+    /// strategy -- it'd be a touch tricky anyhow.
+    pub fn less_than(&self, a: &T) -> Vec<&T> {
+        match self.index(a) {
+            Some(a) => self.with_closure(|closure| {
+                closure.iter(a.0).map(|i| &self.elements[i]).collect()
+            }),
+            None => vec![],
+        }
+    }
+
     /// Picks what I am referring to as the "postdominating"
     /// upper-bound for `a` and `b`. This is usually the least upper
     /// bound, but in cases where there is no single least upper

src/librustc_typeck/lib.rs

@@ -293,6 +293,7 @@ pub fn provide(providers: &mut Providers) {
     collect::provide(providers);
     coherence::provide(providers);
     check::provide(providers);
+    variance::provide(providers);
 }
 
 pub fn check_crate<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>)
@@ -307,9 +308,6 @@ pub fn check_crate<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>)
 
     })?;
 
-    time(time_passes, "variance inference", ||
-         variance::infer_variance(tcx));
-
     tcx.sess.track_errors(|| {
         time(time_passes, "impl wf inference", ||
              impl_wf_check::impl_wf_check(tcx));

src/librustc_typeck/variance/README.md

@@ -97,51 +97,29 @@ types involved before considering variance.
 
 #### Dependency graph management
 
-Because variance works in two phases, if we are not careful, we wind
-up with a muddled mess of a dep-graph. Basically, when gathering up
-the constraints, things are fairly well-structured, but then we do a
-fixed-point iteration and write the results back where they
-belong. You can't give this fixed-point iteration a single task
-because it reads from (and writes to) the variance of all types in the
-crate. In principle, we *could* switch the "current task" in a very
-fine-grained way while propagating constraints in the fixed-point
-iteration and everything would be automatically tracked, but that
-would add some overhead and isn't really necessary anyway.
-
-Instead what we do is to add edges into the dependency graph as we
-construct the constraint set: so, if computing the constraints for
-node `X` requires loading the inference variables from node `Y`, then
-we can add an edge `Y -> X`, since the variance we ultimately infer
-for `Y` will affect the variance we ultimately infer for `X`.
-
-At this point, we've basically mirrored the inference graph in the
-dependency graph. This means we can just completely ignore the
-fixed-point iteration, since it is just shuffling values along this
-graph. In other words, if we added the fine-grained switching of tasks
-I described earlier, all it would show is that we repeatedly read the
-values described by the constraints, but those edges were already
-added when building the constraints in the first place.
-
-Here is how this is implemented (at least as of the time of this
-writing). The associated `DepNode` for the variance map is (at least
-presently) `Signature(DefId)`. This means that, in `constraints.rs`,
-when we visit an item to load up its constraints, we set
-`Signature(DefId)` as the current task (the "memoization" pattern
-described in the `dep-graph` README). Then whenever we find an
-embedded type or trait, we add a synthetic read of `Signature(DefId)`,
-which covers the variances we will compute for all of its
-parameters. This read is synthetic (i.e., we call
-`variance_map.read()`) because, in fact, the final variance is not yet
-computed -- the read *will* occur (repeatedly) during the fixed-point
-iteration phase.
-
-In fact, we don't really *need* this synthetic read. That's because we
-do wind up looking up the `TypeScheme` or `TraitDef` for all
-references types/traits, and those reads add an edge from
-`Signature(DefId)` (that is, they share the same dep node as
-variance). However, I've kept the synthetic reads in place anyway,
-just for future-proofing (in case we change the dep-nodes in the
-future), and because it makes the intention a bit clearer I think.
+Because variance is a whole-crate inference, its dependency graph
+can become quite muddled if we are not careful. To resolve this, we refactor
+into two queries:
+
+- `crate_variances` computes the variance for all items in the current crate.
+- `item_variances` accesses the variance for an individual reading; it
+  works by requesting `crate_variances` and extracting the relevant data.
+  
+If you limit yourself to reading `item_variances`, your code will only
+depend then on the inference inferred for that particular item.
+
+Eventually, the goal is to rely on the red-green dependency management
+algorithm. At the moment, however, we rely instead on a hack, where
+`item_variances` ignores the dependencies of accessing
+`crate_variances` and instead computes the *correct* dependencies
+itself. To this end, when we build up the constraints in the system,
+we also built up a transitive `dependencies` relation as part of the
+crate map. A `(X, Y)` pair is added to the map each time we have a
+constraint that the variance of some inferred for the item `X` depends
+on the variance of some element of `Y`. This is to some extent a
+mirroring of the inference graph in the dependency graph. This means
+we can just completely ignore the fixed-point iteration, since it is
+just shuffling values along this graph.
 
 ### Addendum: Variance on traits