[llvm] [MemProf] Support cloning through recursive cycles (PR #127429)

[Snehasish Kumar via llvm-commits]

================
@@ -293,37 +298,35 @@ class CallsiteContextGraph {
     // TODO: Should this be a map (from Caller node) for more efficient lookup?
     std::vector<std::shared_ptr<ContextEdge>> CallerEdges;
 
-    // Get the list of edges from which we can compute allocation information
-    // such as the context ids and allocation type of this node.
-    const std::vector<std::shared_ptr<ContextEdge>> *
-    getEdgesWithAllocInfo() const {
-      // If node has any callees, compute from those, otherwise compute from
-      // callers (i.e. if this is the leaf allocation node).
-      if (!CalleeEdges.empty())
-        return &CalleeEdges;
+    // Returns true if we need to look at the callee edges for determining the
+    // node context ids and allocation type.
+    bool useCallerEdgesForContextInfo() const {
       // Typically if the callee edges are empty either the caller edges are
       // also empty, or this is an allocation (leaf node). However, if we are
       // allowing recursive callsites and contexts this will be violated for
       // incompletely cloned recursive cycles.
-      assert(CallerEdges.empty() || IsAllocation ||
+      assert(!CalleeEdges.empty() || CallerEdges.empty() || IsAllocation ||
              (AllowRecursiveCallsites && AllowRecursiveContexts));
-      if (!CallerEdges.empty() && IsAllocation)
-        return &CallerEdges;
-      return nullptr;
+      // When cloning for a recursive context, during cloning we might be in the
+      // midst of cloning for a recurrence and have moved context ids off of a
+      // caller edge onto the clone but not yet off of the incoming caller
+      // (back) edge. If we don't look at those we miss the fact that this node
+      // still has context ids of interest.
+      return IsAllocation || CloneRecursiveContexts;
     }
 
     // Compute the context ids for this node from the union of its edge context
     // ids.
     DenseSet<uint32_t> getContextIds() const {
-      DenseSet<uint32_t> ContextIds;
-      auto *Edges = getEdgesWithAllocInfo();
-      if (!Edges)
-        return {};
       unsigned Count = 0;
-      for (auto &Edge : *Edges)
+      for (auto &Edge : CalleeEdges.empty() ? CallerEdges : CalleeEdges)
         Count += Edge->getContextIds().size();
+      DenseSet<uint32_t> ContextIds;
       ContextIds.reserve(Count);
-      for (auto &Edge : *Edges)
+      std::vector<std::shared_ptr<ContextEdge>> Empty;
----------------
snehasish wrote:

I wonder if we can use a default constructed RangeTs instead of constructing an empty vector.

https://github.com/llvm/llvm-project/pull/127429