Only use the new node hashmap for anonymous nodes.

Author: cjgillot

compiler/rustc_codegen_ssa/src/base.rs

@@ -1056,12 +1056,6 @@ pub fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) ->
     // know that later). If we are not doing LTO, there is only one optimized
     // version of each module, so we re-use that.
     let dep_node = cgu.codegen_dep_node(tcx);
-    assert!(
-        !tcx.dep_graph.dep_node_exists(&dep_node),
-        "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
-        cgu.name()
-    );
-
     if tcx.try_mark_green(&dep_node) {
         // We can re-use either the pre- or the post-thinlto state. If no LTO is
         // being performed then we can use post-LTO artifacts, otherwise we must

compiler/rustc_query_system/src/dep_graph/graph.rs

@@ -345,18 +345,6 @@ impl<D: Deps> DepGraphData<D> {
         task: fn(Ctxt, A) -> R,
         hash_result: Option<fn(&mut StableHashingContext<'_>, &R) -> Fingerprint>,
     ) -> (R, DepNodeIndex) {
-        // If the following assertion triggers, it can have two reasons:
-        // 1. Something is wrong with DepNode creation, either here or
-        //    in `DepGraph::try_mark_green()`.
-        // 2. Two distinct query keys get mapped to the same `DepNode`
-        //    (see for example #48923).
-        assert!(
-            !self.dep_node_exists(&key),
-            "forcing query with already existing `DepNode`\n\
-                 - query-key: {arg:?}\n\
-                 - dep-node: {key:?}"
-        );
-
         let with_deps = |task_deps| D::with_deps(task_deps, || task(cx, arg));
         let (result, edges) = if cx.dep_context().is_eval_always(key.kind) {
             (with_deps(TaskDepsRef::EvalAlways), EdgesVec::new())
@@ -443,7 +431,31 @@ impl<D: Deps> DepGraphData<D> {
                     hash: self.current.anon_id_seed.combine(hasher.finish()).into(),
                 };
 
-                self.current.intern_new_node(target_dep_node, task_deps, Fingerprint::ZERO)
+                // The DepNodes generated by the process above are not unique. 2 queries could
+                // have exactly the same dependencies. However, deserialization does not handle
+                // duplicated nodes, so we do the deduplication here directly.
+                //
+                // As anonymous nodes are a small quantity compared to the full dep-graph, the
+                // memory impact of this `anon_node_to_index` map remains tolerable, and helps
+                // us avoid useless growth of the graph with almost-equivalent nodes.
+                match self
+                    .current
+                    .anon_node_to_index
+                    .get_shard_by_value(&target_dep_node)
+                    .lock()
+                    .entry(target_dep_node)
+                {
+                    Entry::Occupied(entry) => *entry.get(),
+                    Entry::Vacant(entry) => {
+                        let dep_node_index = self.current.intern_new_node(
+                            target_dep_node,
+                            task_deps,
+                            Fingerprint::ZERO,
+                        );
+                        entry.insert(dep_node_index);
+                        dep_node_index
+                    }
+                }
             }
         };
 
@@ -607,20 +619,6 @@ impl<D: Deps> DepGraph<D> {
 }
 
 impl<D: Deps> DepGraphData<D> {
-    #[inline]
-    fn dep_node_index_of_opt(&self, dep_node: &DepNode) -> Option<DepNodeIndex> {
-        if let Some(prev_index) = self.previous.node_to_index_opt(dep_node) {
-            self.current.prev_index_to_index.lock()[prev_index]
-        } else {
-            self.current.new_node_to_index.lock_shard_by_value(dep_node).get(dep_node).copied()
-        }
-    }
-
-    #[inline]
-    fn dep_node_exists(&self, dep_node: &DepNode) -> bool {
-        self.dep_node_index_of_opt(dep_node).is_some()
-    }
-
     fn node_color(&self, dep_node: &DepNode) -> Option<DepNodeColor> {
         if let Some(prev_index) = self.previous.node_to_index_opt(dep_node) {
             self.colors.get(prev_index)
@@ -653,11 +651,6 @@ impl<D: Deps> DepGraphData<D> {
 }
 
 impl<D: Deps> DepGraph<D> {
-    #[inline]
-    pub fn dep_node_exists(&self, dep_node: &DepNode) -> bool {
-        self.data.as_ref().is_some_and(|data| data.dep_node_exists(dep_node))
-    }
-
     /// Checks whether a previous work product exists for `v` and, if
     /// so, return the path that leads to it. Used to skip doing work.
     pub fn previous_work_product(&self, v: &WorkProductId) -> Option<WorkProduct> {
@@ -838,10 +831,7 @@ impl<D: Deps> DepGraphData<D> {
         let frame = MarkFrame { index: prev_dep_node_index, parent: frame };
 
         #[cfg(not(parallel_compiler))]
-        {
-            debug_assert!(!self.dep_node_exists(dep_node));
-            debug_assert!(self.colors.get(prev_dep_node_index).is_none());
-        }
+        debug_assert!(self.colors.get(prev_dep_node_index).is_none());
 
         // We never try to mark eval_always nodes as green
         debug_assert!(!qcx.dep_context().is_eval_always(dep_node.kind));
@@ -1038,24 +1028,24 @@ rustc_index::newtype_index! {
 /// largest in the compiler.
 ///
 /// For this reason, we avoid storing `DepNode`s more than once as map
-/// keys. The `new_node_to_index` map only contains nodes not in the previous
+/// keys. The `anon_node_to_index` map only contains nodes of anonymous queries not in the previous
 /// graph, and we map nodes in the previous graph to indices via a two-step
 /// mapping. `SerializedDepGraph` maps from `DepNode` to `SerializedDepNodeIndex`,
 /// and the `prev_index_to_index` vector (which is more compact and faster than
 /// using a map) maps from `SerializedDepNodeIndex` to `DepNodeIndex`.
 ///
-/// This struct uses three locks internally. The `data`, `new_node_to_index`,
+/// This struct uses three locks internally. The `data`, `anon_node_to_index`,
 /// and `prev_index_to_index` fields are locked separately. Operations that take
 /// a `DepNodeIndex` typically just access the `data` field.
 ///
 /// We only need to manipulate at most two locks simultaneously:
-/// `new_node_to_index` and `data`, or `prev_index_to_index` and `data`. When
-/// manipulating both, we acquire `new_node_to_index` or `prev_index_to_index`
+/// `anon_node_to_index` and `data`, or `prev_index_to_index` and `data`. When
+/// manipulating both, we acquire `anon_node_to_index` or `prev_index_to_index`
 /// first, and `data` second.
 pub(super) struct CurrentDepGraph<D: Deps> {
     encoder: GraphEncoder<D>,
-    new_node_to_index: Sharded<FxHashMap<DepNode, DepNodeIndex>>,
     prev_index_to_index: Lock<IndexVec<SerializedDepNodeIndex, Option<DepNodeIndex>>>,
+    anon_node_to_index: Sharded<FxHashMap<DepNode, DepNodeIndex>>,
 
     /// This is used to verify that fingerprints do not change between the creation of a node
     /// and its recomputation.
@@ -1123,7 +1113,7 @@ impl<D: Deps> CurrentDepGraph<D> {
                 profiler,
                 previous,
             ),
-            new_node_to_index: Sharded::new(|| {
+            anon_node_to_index: Sharded::new(|| {
                 FxHashMap::with_capacity_and_hasher(
                     new_node_count_estimate / sharded::shards(),
                     Default::default(),
@@ -1158,14 +1148,7 @@ impl<D: Deps> CurrentDepGraph<D> {
         edges: EdgesVec,
         current_fingerprint: Fingerprint,
     ) -> DepNodeIndex {
-        let dep_node_index = match self.new_node_to_index.lock_shard_by_value(&key).entry(key) {
-            Entry::Occupied(entry) => *entry.get(),
-            Entry::Vacant(entry) => {
-                let dep_node_index = self.encoder.send(key, current_fingerprint, edges);
-                entry.insert(dep_node_index);
-                dep_node_index
-            }
-        };
+        let dep_node_index = self.encoder.send(key, current_fingerprint, edges);
 
         #[cfg(debug_assertions)]
         self.record_edge(dep_node_index, key, current_fingerprint);
@@ -1235,8 +1218,6 @@ impl<D: Deps> CurrentDepGraph<D> {
         prev_graph: &SerializedDepGraph,
         prev_index: SerializedDepNodeIndex,
     ) -> DepNodeIndex {
-        self.debug_assert_not_in_new_nodes(prev_graph, prev_index);
-
         let mut prev_index_to_index = self.prev_index_to_index.lock();
 
         match prev_index_to_index[prev_index] {
@@ -1254,19 +1235,6 @@ impl<D: Deps> CurrentDepGraph<D> {
             }
         }
     }
-
-    #[inline]
-    fn debug_assert_not_in_new_nodes(
-        &self,
-        prev_graph: &SerializedDepGraph,
-        prev_index: SerializedDepNodeIndex,
-    ) {
-        let node = &prev_graph.index_to_node(prev_index);
-        debug_assert!(
-            !self.new_node_to_index.lock_shard_by_value(node).contains_key(node),
-            "node from previous graph present in new node collection"
-        );
-    }
 }
 
 #[derive(Debug, Clone, Copy)]
@@ -1388,7 +1356,7 @@ fn panic_on_forbidden_read<D: Deps>(data: &DepGraphData<D>, dep_node_index: DepN
 
     if dep_node.is_none() {
         // Try to find it among the new nodes
-        for shard in data.current.new_node_to_index.lock_shards() {
+        for shard in data.current.anon_node_to_index.lock_shards() {
             if let Some((node, _)) = shard.iter().find(|(_, index)| **index == dep_node_index) {
                 dep_node = Some(*node);
                 break;