[llvm] [Coroutines][LazyCallGraph] resumes are not really SCC (PR #116285)

[Tyler Nowicki via llvm-commits]

================
@@ -1082,11 +1082,14 @@ class LazyCallGraph {
   /// Add new ref-recursive functions split/outlined from an existing function.
   ///
   /// The new functions may only reference other functions that the original
-  /// function did. The new functions may reference (not call) the original
-  /// function.
+  /// function did. The new functions may reference the original function. New
+  /// functions must not call other new functions or the original function.
   ///
-  /// The original function must reference (not call) all new functions.
-  /// All new functions must reference (not call) each other.
+  /// Marks the original function as referencing all new functions.
+  ///
+  /// The CG must be updated following the use of this helper, for example with
+  /// updateCGAndAnalysisManagerForCGSCCPass(), to ensure the RefSCCs and SCCs
+  /// are correctly identified.
----------------
TylerNowicki wrote:

> the point of updateCGAndAnalysisManagerForCGSCCPass() is to do the call graph update for you when there are potential edge additions/removals that can change the call graph and requeue any modified SCCs to be visited again

This point is new to me actually. I only previously understood that we wanted the call-graph's (Ref)SCCs to match the IR. But it sounds like what you are saying is that it will recompute the list of SCCs that need to be visited. In our case the original and new functions just need to be requeued.

**I have a follow-up question on this**: after we call addSplitRefRecursiveFunctions we immediately call updateCGAndAnalysisManagerForCGSCCPass, followed by simplifying the original function, then a call to updateCGAndAnalysisManagerForFunctionPass. You can see this in CoroSplit.cpp::updateCallGraphAfterCoroutineSplit(...). When we simplify the original function some of the references to the new functions may be removed. I am starting to suspect the call to updateCGAndAnalysisManagerForCGSCCPass is not required, does that match your interpretation as well?

> another clarifying question, does the original function still always reference all newly split functions?

To clarify your question below as it relates to this. At the time when addSplitRefRecursiveFunctions and updateCGAndAnalysisManagerForCGSCCPass are called the original function still always references all the newly split functions.

https://github.com/llvm/llvm-project/pull/116285