[llvm] [SampleFDO] Stale profile call-graph matching (PR #92151)

[Lei Wang via llvm-commits]

================
@@ -590,14 +617,318 @@ void SampleProfileMatcher::computeAndReportProfileStaleness() {
   }
 }
 
-void SampleProfileMatcher::runOnModule() {
-  ProfileConverter::flattenProfile(Reader.getProfiles(), FlattenedProfiles,
-                                   FunctionSamples::ProfileIsCS);
+void SampleProfileMatcher::findNewIRFunctions(
+    StringMap<Function *> &NewIRFunctions) {
+  // 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) {
-    if (skipProfileForFunction(F))
+    // Skip declarations, as even if the function can be recognized renamed, we
+    // have nothing to do with it.
+    if (F.isDeclaration())
       continue;
-    runOnFunction(F);
+
+    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[CanonFName] = &F;
   }
+}
+
+void SampleProfileMatcher::findNewIRCallees(
+    Function &Caller, const StringMap<Function *> &NewIRFunctions,
+    std::vector<Function *> &NewIRCallees) {
+  for (auto &BB : Caller) {
----------------
wlei-llvm wrote:

I just realized that this function may be run redundantly, we traverse the nested tree for all callers, but it's possible that the sample function appears multiple times in different inline context and will be run repeatedly.  Then the solution could be use a cache or run on a flattened profile(but that needs a new function to update the original profile)



 

https://github.com/llvm/llvm-project/pull/92151