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

Author: wlei-llvm

llvm/include/llvm/ProfileData/SampleProf.h

@@ -466,7 +466,7 @@ struct SampleContextFrame {
   LineLocation Location;
 
   SampleContextFrame() : Location(0, 0) {}
-  
+
   SampleContextFrame(FunctionId Func, LineLocation Location)
       : Func(Func), Location(Location) {}
 
@@ -527,7 +527,7 @@ class SampleContext {
       : Func(Name), State(UnknownContext), Attributes(ContextNone) {
         assert(!Name.empty() && "Name is empty");
       }
-  
+
   SampleContext(FunctionId Func)
       : Func(Func), State(UnknownContext), Attributes(ContextNone) {}
 
@@ -975,6 +975,8 @@ class FunctionSamples {
     return CallsiteSamples;
   }
 
+  CallsiteSampleMap &getCallsiteSamples() { return CallsiteSamples; }
+
   /// Return the maximum of sample counts in a function body. When SkipCallSite
   /// is false, which is the default, the return count includes samples in the
   /// inlined functions. When SkipCallSite is true, the return count only

llvm/lib/Transforms/IPO/SampleProfile.cpp

@@ -1129,7 +1129,7 @@ void SampleProfileLoader::findExternalInlineCandidate(
         CalleeSample->getContext().hasAttribute(ContextShouldBeInlined);
     if (!PreInline && CalleeSample->getHeadSamplesEstimate() < Threshold)
       continue;
-    
+
     Function *Func = SymbolMap.lookup(CalleeSample->getFunction());
     // Add to the import list only when it's defined out of module.
     if (!Func || Func->isDeclaration())

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 +196,39 @@ void ProfileGeneratorBase::showDensitySuggestion(double Density) {
            << "% total samples: " << format("%.1f", Density) << "\n";
 }
 
+// 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"};
+
+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 : 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;
@@ -492,6 +525,7 @@ void ProfileGenerator::postProcessProfiles() {
   computeSummaryAndThreshold(ProfileMap);
   trimColdProfiles(ProfileMap, ColdCountThreshold);
   calculateAndShowDensity(ProfileMap);
+  filterAmbiguousProfile(ProfileMap);
 }
 
 void ProfileGenerator::trimColdProfiles(const SampleProfileMap &Profiles,
@@ -1024,6 +1058,7 @@ 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);