[clang] Reduce memory usage in AST parent map generation by lazily checking if nodes have been seen (PR #129934)

[via cfe-commits]

================
@@ -70,16 +93,37 @@ class ParentMapContext::ParentMap {
         push_back(Value);
     }
     bool contains(const DynTypedNode &Value) {
-      return Seen.contains(Value);
+      assert(Value.getMemoizationData());
+      bool found = FragileLazySeenCache.contains(&Value);
+      while (!found && ItemsProcessed < Items.size()) {
+        found |= FragileLazySeenCache.insert(&Items[ItemsProcessed]).second;
+        ++ItemsProcessed;
+      }
+      return found;
     }
     void push_back(const DynTypedNode &Value) {
-      if (!Value.getMemoizationData() || Seen.insert(Value).second)
+      if (!Value.getMemoizationData() || !contains(Value)) {
----------------
higher-performance wrote:

I could definitely optimize these further like in the ways you mention, but those are just (very small) constant-factor savings in time, at a significant readability and fragility cost. I didn't get the feeling it's worth it to be honest -- do you? Note that the performance graphs almost coincide with Clang 17's right now.

https://github.com/llvm/llvm-project/pull/129934