[llvm-branch-commits] [llvm] [BOLT] Match functions with call graph (PR #98125)

[Shaw Young via llvm-branch-commits]

================
@@ -446,6 +503,56 @@ size_t YAMLProfileReader::matchWithLTOCommonName() {
   return MatchedWithLTOCommonName;
 }
 
+size_t YAMLProfileReader::matchWithCallGraph(BinaryContext &BC) {
+  if (!opts::MatchWithCallGraph)
+    return 0;
+
+  size_t MatchedWithCallGraph = 0;
+  CGMatcher.computeBFNeighborHashes(BC);
+  CGMatcher.constructYAMLFCG(YamlBP, IdToYamLBF);
+
+  // Matches YAMLBF to BFs with neighbor hashes.
+  for (yaml::bolt::BinaryFunctionProfile &YamlBF : YamlBP.Functions) {
+    if (YamlBF.Used)
+      continue;
+    auto It = CGMatcher.YamlBFAdjacencyMap.find(&YamlBF);
+    if (It == CGMatcher.YamlBFAdjacencyMap.end())
+      continue;
+    // Computes profiled function's neighbor hash.
+    std::set<yaml::bolt::BinaryFunctionProfile *> &AdjacentFunctions =
+        It->second;
+    std::string AdjacentFunctionHashStr;
+    for (auto &AdjacentFunction : AdjacentFunctions) {
+      AdjacentFunctionHashStr += AdjacentFunction->Name;
+    }
+    uint64_t Hash = std::hash<std::string>{}(AdjacentFunctionHashStr);
+    auto NeighborHashToBFsIt = CGMatcher.NeighborHashToBFs.find(Hash);
+    if (NeighborHashToBFsIt == CGMatcher.NeighborHashToBFs.end())
+      continue;
+    // Finds the binary function with the closest block size to the profiled
----------------
shawbyoung wrote:

1. In a binary with 953488 binary functions and 13376 profiled functions, the largest bucket had 151 functions. This however was an outlier - the median & mean bucket size was 2 and ~5 respectively and more than half of buckets only had one binary function. 
2. I agree - just implemented LCP name matching as opposed to block count.

https://github.com/llvm/llvm-project/pull/98125