Auto merge of #39912 - nikomatsakis:incr-comp-trait-select-no-vec, r=eddyb

rework `TraitSelect` to avoid a vec and just use two def-ids

r? @eddyb

src/librustc/dep_graph/dep_node.rs

@@ -131,7 +131,37 @@ pub enum DepNode<D: Clone + Debug> {
     // which would yield an overly conservative dep-graph.
     TraitItems(D),
     ReprHints(D),
-    TraitSelect(Vec<D>),
+
+    // Trait selection cache is a little funny. Given a trait
+    // reference like `Foo: SomeTrait<Bar>`, there could be
+    // arbitrarily many def-ids to map on in there (e.g., `Foo`,
+    // `SomeTrait`, `Bar`). We could have a vector of them, but it
+    // requires heap-allocation, and trait sel in general can be a
+    // surprisingly hot path. So instead we pick two def-ids: the
+    // trait def-id, and the first def-id in the input types. If there
+    // is no def-id in the input types, then we use the trait def-id
+    // again. So for example:
+    //
+    // - `i32: Clone` -> `TraitSelect { trait_def_id: Clone, self_def_id: Clone }`
+    // - `u32: Clone` -> `TraitSelect { trait_def_id: Clone, self_def_id: Clone }`
+    // - `Clone: Clone` -> `TraitSelect { trait_def_id: Clone, self_def_id: Clone }`
+    // - `Vec<i32>: Clone` -> `TraitSelect { trait_def_id: Clone, self_def_id: Vec }`
+    // - `String: Clone` -> `TraitSelect { trait_def_id: Clone, self_def_id: String }`
+    // - `Foo: Trait<Bar>` -> `TraitSelect { trait_def_id: Trait, self_def_id: Foo }`
+    // - `Foo: Trait<i32>` -> `TraitSelect { trait_def_id: Trait, self_def_id: Foo }`
+    // - `(Foo, Bar): Trait` -> `TraitSelect { trait_def_id: Trait, self_def_id: Foo }`
+    // - `i32: Trait<Foo>` -> `TraitSelect { trait_def_id: Trait, self_def_id: Foo }`
+    //
+    // You can see that we map many trait refs to the same
+    // trait-select node.  This is not a problem, it just means
+    // imprecision in our dep-graph tracking.  The important thing is
+    // that for any given trait-ref, we always map to the **same**
+    // trait-select node.
+    TraitSelect { trait_def_id: D, input_def_id: D },
+
+    // For proj. cache, we just keep a list of all def-ids, since it is
+    // not a hotspot.
+    ProjectionCache { def_ids: Vec<D> },
 }
 
 impl<D: Clone + Debug> DepNode<D> {
@@ -236,9 +266,17 @@ pub fn map_def<E, OP>(&self, mut op: OP) -> Option<DepNode<E>>
             TraitImpls(ref d) => op(d).map(TraitImpls),
             TraitItems(ref d) => op(d).map(TraitItems),
             ReprHints(ref d) => op(d).map(ReprHints),
-            TraitSelect(ref type_ds) => {
+            TraitSelect { ref trait_def_id, ref input_def_id } => {
-                let type_ds = try_opt!(type_ds.iter().map(|d| op(d)).collect());
+                op(trait_def_id).and_then(|trait_def_id| {
-                Some(TraitSelect(type_ds))
+                    op(input_def_id).and_then(|input_def_id| {
+                        Some(TraitSelect { trait_def_id: trait_def_id,
+                                           input_def_id: input_def_id })
+                    })
+                })
+            }
+            ProjectionCache { ref def_ids } => {
+                let def_ids: Option<Vec<E>> = def_ids.iter().map(op).collect();
+                def_ids.map(|d| ProjectionCache { def_ids: d })
             }
         }
     }

src/librustc/ty/mod.rs

@@ -34,7 +34,6 @@
 use std::borrow::Cow;
 use std::cell::{Cell, RefCell, Ref};
 use std::hash::{Hash, Hasher};
-use std::iter;
 use std::ops::Deref;
 use std::rc::Rc;
 use std::slice;
@@ -843,27 +842,22 @@ pub fn def_id(&self) -> DefId {
 
     /// Creates the dep-node for selecting/evaluating this trait reference.
     fn dep_node(&self) -> DepNode<DefId> {
-        // Ideally, the dep-node would just have all the input types
+        // Extact the trait-def and first def-id from inputs.  See the
-        // in it.  But they are limited to including def-ids. So as an
+        // docs for `DepNode::TraitSelect` for more information.
-        // approximation we include the def-ids for all nominal types
+        let trait_def_id = self.def_id();
-        // found somewhere. This means that we will e.g. conflate the
+        let input_def_id =
-        // dep-nodes for `u32: SomeTrait` and `u64: SomeTrait`, but we
-        // would have distinct dep-nodes for `Vec<u32>: SomeTrait`,
-        // `Rc<u32>: SomeTrait`, and `(Vec<u32>, Rc<u32>): SomeTrait`.
-        // Note that it's always sound to conflate dep-nodes, it just
-        // leads to more recompilation.
-        let def_ids: Vec<_> =
             self.input_types()
                 .flat_map(|t| t.walk())
                 .filter_map(|t| match t.sty {
-                    ty::TyAdt(adt_def, _) =>
+                    ty::TyAdt(adt_def, _) => Some(adt_def.did),
-                        Some(adt_def.did),
+                    _ => None
-                    _ =>
-                        None
                 })
-                .chain(iter::once(self.def_id()))
+                .next()
-                .collect();
+                .unwrap_or(trait_def_id);
-        DepNode::TraitSelect(def_ids)
+        DepNode::TraitSelect {
+            trait_def_id: trait_def_id,
+            input_def_id: input_def_id
+        }
     }
 
     pub fn input_types<'a>(&'a self) -> impl DoubleEndedIterator<Item=Ty<'tcx>> + 'a {

src/librustc_trans/context.rs

@@ -208,7 +208,8 @@ fn to_dep_node(key: &Self::Key) -> DepNode<DefId> {
                    _ => None,
                })
                .collect();
-        DepNode::TraitSelect(def_ids)
+
+        DepNode::ProjectionCache { def_ids: def_ids }
     }
 }