[llvm] Reland "[MLGO][IR2Vec] Integrating IR2Vec with MLInliner (#143479)" (PR #145664)

[Mircea Trofin via llvm-commits]

================
@@ -433,6 +447,24 @@ std::unique_ptr<InlineAdvice> MLInlineAdvisor::getAdviceImpl(CallBase &CB) {
   *ModelRunner->getTensor<int64_t>(FeatureIndex::is_caller_avail_external) =
       Caller.hasAvailableExternallyLinkage();
 
+  if (UseIR2Vec) {
+    // Python side expects float embeddings. The IR2Vec embeddings are doubles
+    // as of now due to the restriction of fromJSON method used by the
+    // readVocabulary method in ir2vec::Embeddings.
+    auto setEmbedding = [&](const ir2vec::Embedding &Embedding,
+                            FeatureIndex Index) {
+      auto Embedding_float =
----------------
mtrofin wrote:

Also, this results in 2 copies. Maybe we can do in 1:

```
for (auto I = 0, E = Embedding.size(); ++I)
  ModelRunner->getTensor<float>(Index)[I] = Embedding[I];
```

https://github.com/llvm/llvm-project/pull/145664