[llvm-branch-commits] [llvm] [MLGO][IR2Vec] Integrating IR2Vec with MLInliner (PR #143479)

[Mircea Trofin via llvm-branch-commits]

================
@@ -186,6 +186,20 @@ MLInlineAdvisor::MLInlineAdvisor(
     EdgeCount += getLocalCalls(KVP.first->getFunction());
   }
   NodeCount = AllNodes.size();
+
+  if (auto IR2VecVocabResult = MAM.getCachedResult<IR2VecVocabAnalysis>(M)) {
+    if (!IR2VecVocabResult->isValid()) {
+      M.getContext().emitError("IR2VecVocabAnalysis is not valid");
+      return;
+    }
+    // Add the IR2Vec features to the feature map
+    auto IR2VecDim = IR2VecVocabResult->getDimension();
+    FeatureMap.push_back(
+        TensorSpec::createSpec<float>("callee_embedding", {IR2VecDim}));
+    FeatureMap.push_back(
+        TensorSpec::createSpec<float>("caller_embedding", {IR2VecDim}));
+    UseIR2Vec = true;
----------------
mtrofin wrote:

nit: if it makes sense, you could have `const UseIR2Vec` and init it in the ctor based on `MAM.getCachedResult`. Then here, if the result isn't valid, splat the error.

main value is readability - the bool flag would say it's set once. 

https://github.com/llvm/llvm-project/pull/143479