[llvm-profgen] Filter out ambiguous profiles during profile generation

Author: wlei-llvm

llvm/test/tools/llvm-profgen/Inputs/filter-ambiguous-profile.prof

@@ -0,0 +1,16 @@
+foo:12345:1000
+ 1: 1000
+ 4: bar:1000
+  1: 1000
+  2: __tls_init.llvm.123:1
+   1: 1
+  3: goo:300
+   1: 300
+ 8: __cxx_global_var_init.4:4
+  1: 1
+  2: goo:3
+   1: 3
+__cxx_global_var_init.1:1:1
+ 1: 1
+__tls_init.llvm.345:1:1
+ 1: 1

llvm/test/tools/llvm-profgen/filter-ambiguous-profile.test

@@ -0,0 +1,8 @@
+; RUN: llvm-profgen --format=text --llvm-sample-profile=%S/Inputs/filter-ambiguous-profile.prof --binary=%S/Inputs/inline-cs-noprobe.perfbin --csspgo-preinliner=0 --output=%t1 || FileCheck %s --input-file %t1
+
+;CHECK:    foo:12345:1000
+;CHECK-NEXT  1: 1000
+;CHECK-NEXT  4: bar:1000
+;CHECK-NEXT   1: 1000
+;CHECK-NEXT   3: goo:300
+;CHECK-NEXT    1: 300

llvm/tools/llvm-profgen/ProfileGenerator.cpp

@@ -196,6 +196,33 @@ void ProfileGeneratorBase::showDensitySuggestion(double Density) {
            << "% total samples: " << format("%.1f", Density) << "\n";
 }
 
+bool ProfileGeneratorBase::filterAmbiguousProfile(FunctionSamples &FS) {
+  for (const auto &Prefix : FuncPrefixsToFilter) {
+    if (FS.getFuncName().starts_with(Prefix))
+      return true;
+  }
+
+  // Filter the function profiles for the inlinees.
+  for (auto &Callees :
+       const_cast<CallsiteSampleMap &>(FS.getCallsiteSamples())) {
+    auto &CalleesMap = Callees.second;
+    for (auto I = CalleesMap.begin(); I != CalleesMap.end();) {
+      auto Tmp = I++;
+      if (filterAmbiguousProfile(Tmp->second))
+        CalleesMap.erase(Tmp);
+    }
+  }
+  return false;
+}
+
+void ProfileGeneratorBase::filterAmbiguousProfile(SampleProfileMap &Profiles) {
+  for (auto I = ProfileMap.begin(); I != ProfileMap.end();) {
+    auto Tmp = I++;
+    if (filterAmbiguousProfile(Tmp->second))
+      ProfileMap.erase(Tmp);
+  }
+}
+
 double ProfileGeneratorBase::calculateDensity(const SampleProfileMap &Profiles,
                                               uint64_t HotCntThreshold) {
   double Density = DBL_MAX;
@@ -491,6 +518,7 @@ void ProfileGenerator::generateProfile() {
 void ProfileGenerator::postProcessProfiles() {
   computeSummaryAndThreshold(ProfileMap);
   trimColdProfiles(ProfileMap, ColdCountThreshold);
+  filterAmbiguousProfile(ProfileMap);
   calculateAndShowDensity(ProfileMap);
 }
 
@@ -1024,6 +1052,7 @@ void CSProfileGenerator::postProcessProfiles() {
     CSConverter.convertCSProfiles();
     FunctionSamples::ProfileIsCS = false;
   }
+  filterAmbiguousProfile(ProfileMap);
 }
 
 void ProfileGeneratorBase::computeSummaryAndThreshold(

llvm/tools/llvm-profgen/ProfileGenerator.h

@@ -108,6 +108,10 @@ class ProfileGeneratorBase {
 
   void updateCallsiteSamples();
 
+  void filterAmbiguousProfile(SampleProfileMap &Profiles);
+
+  bool filterAmbiguousProfile(FunctionSamples &FS);
+
   StringRef getCalleeNameForAddress(uint64_t TargetAddress);
 
   void computeSummaryAndThreshold(SampleProfileMap &ProfileMap);
@@ -128,6 +132,13 @@ class ProfileGeneratorBase {
   virtual bool collectFunctionsFromLLVMProfile(
       std::unordered_set<const BinaryFunction *> &ProfiledFunctions) = 0;
 
+  // Those functions are usually cold but could have multiple versions, the
+  // profile are either from one version or the merged version, which is
+  // ambiguous. We can't attribute the profile for each version accurately, so
+  // filter them out from the profile map.
+  static constexpr const char *FuncPrefixsToFilter[] = {"__cxx_global_var_init",
+                                                        "__tls_init"};
+
   // Thresholds from profile summary to answer isHotCount/isColdCount queries.
   uint64_t HotCountThreshold;