[llvm] [SampleFDO] Stale profile call-graph matching (PR #95135)

[Lei Wang via llvm-commits]

================
@@ -590,14 +688,237 @@ void SampleProfileMatcher::computeAndReportProfileStaleness() {
   }
 }
 
+void SampleProfileMatcher::findNewIRFunctions() {
+  // TODO: Support MD5 profile.
+  if (FunctionSamples::UseMD5)
+    return;
+  StringSet<> NamesInProfile;
+  if (auto NameTable = Reader.getNameTable()) {
+    for (auto Name : *NameTable)
+      NamesInProfile.insert(Name.stringRef());
+  }
+
+  for (auto &F : M) {
+    // Skip declarations, as even if the function can be matched, we have
+    // nothing to do with it.
+    if (F.isDeclaration())
+      continue;
+
+    StringRef CanonFName = FunctionSamples::getCanonicalFnName(F.getName());
+    const auto *FS = getFlattenedSamplesFor(F);
+    if (FS)
+      continue;
+
+    // For extended binary, functions are fully inlined may not be loaded in the
+    // top-level profile, so check the NameTable which has the all symbol names
+    // in profile.
+    if (NamesInProfile.count(CanonFName))
+      continue;
+
+    // For extended binary, non-profiled function symbols are in the profile
+    // symbol list table.
+    if (PSL && PSL->contains(CanonFName))
+      continue;
+
+    LLVM_DEBUG(dbgs() << "Function " << CanonFName
+                      << " is not in profile or symbol list table.\n");
+    NewIRFunctions[FunctionId(CanonFName)] = &F;
+  }
+}
+
+bool SampleProfileMatcher::functionMatchesProfileHelper(
+    const Function &IRFunc, const FunctionId &ProfFunc) {
+  // The value is in the range [0, 1]. The bigger the value is, the more similar
+  // two sequences are.
+  float Similarity = 0.0;
+
+  const auto *FSFlattened = getFlattenedSamplesFor(ProfFunc);
+  assert(FSFlattened && "Flattened profile sample is null");
+  // Similarity check may not be reliable if the function is tiny, we use the
+  // number of basic block as a proxy for the function complexity and skip the
+  // matching if it's too small.
+  if (IRFunc.size() < MinFuncCountForCGMatching ||
+      FSFlattened->getBodySamples().size() < MinFuncCountForCGMatching)
+    return false;
+
+  // For probe-based function, we first trust the checksum info. If the checksum
+  // doesn't match, we continue checking for similarity.
+  if (FunctionSamples::ProfileIsProbeBased) {
+    const auto *FuncDesc = ProbeManager->getDesc(IRFunc);
+    if (FuncDesc &&
+        !ProbeManager->profileIsHashMismatched(*FuncDesc, *FSFlattened)) {
+      LLVM_DEBUG(dbgs() << "The checksums for " << IRFunc.getName()
+                        << "(IR) and " << ProfFunc << "(Profile) match.\n");
+
+      return true;
+    }
+  }
----------------
wlei-llvm wrote:

this is intentional, the reason is checksum may not be reliable for tiny function too( I saw some cases the block num is only one but the function are different).

https://github.com/llvm/llvm-project/pull/95135