[llvm] [profcheck][coro] Adding Branch weights PGO in Coroutines Passes (PR #184466)

[Chuanqi Xu via llvm-commits]

================
@@ -1594,6 +1619,30 @@ struct SwitchCoroutineSplitter {
       ++SuspendIndex;
     }
 
+    // Add the branch weights to the switch instruction.
+    if (!ProfcheckDisableMetadataFixes) {
+      if (Switch->getNumCases() > 0) {
+        SmallVector<uint32_t> Weights;
+        // Add the Unlikely weight for the default case.
+        Weights.push_back(llvm::MDBuilder::kUnlikelyBranchWeight);
+        size_t NumCases = Switch->getNumCases();
+        if (NumCases == 1) {
+          Weights.push_back(llvm::MDBuilder::kUnlikelyBranchWeight);
+        } else {
+          // First case is unlikely.
+          Weights.push_back(llvm::MDBuilder::kUnlikelyBranchWeight);
+          // Cases between first and last are likely.
+          for (size_t i = 1; i < NumCases - 1; ++i)
+            Weights.push_back(llvm::MDBuilder::kLikelyBranchWeight);
+          // The last case is unlikely.
+          Weights.push_back(llvm::MDBuilder::kUnlikelyBranchWeight);
+        }
+        MDBuilder MDB(C);
+        Switch->setMetadata(LLVMContext::MD_prof,
+                            MDB.createBranchWeights(Weights));
+      }
+    }
----------------
ChuanqiXu9 wrote:

I am still not sure if this is correct.

e.g, we can have a pattern that always destroy the coroutine after initialization. in this case, the destroyer has higher occurance than the other resuming point.

And also, I am wondering if, techniquely, can we have the first suspend point in a loop. For example, we can have an initial_suspend which is not always_suspend, then the first co_await appears in a loop, then the first resuming point may have higher probability.

So I think, we can leave a comment here and not landing this. And let's work it in later patches.

https://github.com/llvm/llvm-project/pull/184466