[llvm-profdata] Implement llvm-profdata overlap for sample profiles

[weihe]

Summary

Implemented the llvm-profdata overlap feature for sample profiles. It reports weighted similarity and unweighted overlap metrics at program and function level for two input profiles. Similarity metrics are symmetric with regards to the order of two input profiles. By default, the tool only reports program-level summary. Users can look into function-level details via additional options --function, --similarity-cutoff, and --value-cutoff.

The similarity metrics are designed as follows:

• Program-level summary
  • Whole program profile similarity is an aggregate over function-level similarity FS: PS = sum(FS(A) * avg_weight(A)) for all function A.
  • Whole program sample overlap: PSO = common_samples / total_samples.
  • Function overlap: FO = #common_function / #total_function.
  • Hot-function overlap: HFO = #common_hot_function / #total_hot_function.
  • Hot-block overlap: HBO = #common_hot_block / #total_hot_block.
• Function-level details
  • Function-level similarity is an aggregate over line/block-level similarities BS of all sample lines/blocks in the function, weighted by the closeness of the function's weights in two profiles: FS = sum(BS(i)) * (1 - weight_distance(A)).
  • Function-level sample overlap: FSO = common_samples / total_samples for samples in the function.

[weihe]

Corrected two typos in comment.

[hoyFB]

llvm/tools/llvm-profdata/llvm-profdata.cpp

858

function-level similarity?

1048

I think you want just const auto Match = .... The return value will be returned by reference if it is large.

1050

May just increase HotFuncOverlap.UnionCount before continue, instead of erasing every matched function from TestHotFunc? like

for (const auto &F : TestHotFunc) {
    if (BaseHotFunc.count(F.first()))
       ++HotFuncOverlap.OverlapCount;
    else
       ++HotFuncOverlap.UnionCount;
}

[weihe]

Refactored computeHotFuncOverlap() and renamed computeSampleFunctionOverlap() to computeSampleFunctionInternalOverlap()

[weihe]

llvm/tools/llvm-profdata/llvm-profdata.cpp

858

Thank you for the comment! This is not exactly the function-level similarity we report in the tool, but an intermediate result towards it. The formula of function-level similarity is given at line 882.

I renamed this function to computeSampleFunctionInternalOverlap() and moved it to a private member to reduce confusions.

1048

Thank you for the suggestion! I've changed the code accordingly.

1050

This suggestion is really good! The refactored code is much cleaner. Thank you very much!

[hoyFB]

@wmi We'd like to hear about your input on this. Thanks!

[wmi]

Thanks for the work. It is a very useful feature.

llvm/tools/llvm-profdata/llvm-profdata.cpp

966–967

Can we make the similarity within range 0~1 to be consistent with Block and profile similarity? It is more natural to reason the similarity with range 0~1.

1068–1074

Seemly a lot of complexity of the function comes from lock step iteration of the maps from two profiles at the same time. Could you extract the lock step iteration logic into a separate class? This way we don't have to deal with the logic multiple times in iterating BodySampleMap, CallsiteSamplesMap and FunctionSamplesMap.

1093–1098

updateForUnmatchedBlock is a special case of the block above. We may be able to share the code.

[weihe]

Extracted lock step iteration logic into class MatchStep and revised code according to other suggestions.

[weihe]

llvm/tools/llvm-profdata/llvm-profdata.cpp

966–967

Thank you for the suggestion! I have changed the function computeSampleFunctionInternalOverlap() to return a double in range 0~1.

1068–1074

I have extracted the logic of lock step iteration to class MatchStep. This is a great suggestion. Thank you very much!

1093–1098

Thank you for the suggestion! I combined this code with updateForUnmatchedBlock() into a new function updateOverlapStatsForFunction().

[wmi]

Thanks. Another suggestion about comment. Other than that, LGTM.

llvm/tools/llvm-profdata/llvm-profdata.cpp

1053–1054

Both weightFuncSimilarity and weightByImportance considers the weight of the function in the profile, so at the beginning I felt confused what are their difference. I find out the difference is weightFuncSimilarity absorbs the weight difference into the similarity so the similarity is still in the range of 0~1, while weightByImportance multiplies the similarity by weight ratio of the function in the profile (the average ratio of the two profiles), so the aggregate similarity of all the functions in the profiles will be in the range of 0~1. Please correct me if I am wrong. But it is better to make the intention of these two functions more clear in the comments.

[weihe]

Grouped previous weightFuncSimilarity() and computeSampleFunctionInternalOverlap() to computeSampleFunctionOverlap() and added comments for readability.

[weihe]

llvm/tools/llvm-profdata/llvm-profdata.cpp

1053–1054

Yes, that's right! In fact, the previous weightFuncSimilarity() is part of the computation of "function-level similarity", whereas weightByImportance() is part of the computation that aggregates "function-level similarity" to "profile-level similarity". So the two functions use weights in slightly different ways.

I also realized these two functions may be confusing, so I grouped the previous weightFuncSimilarity() (renamed as weightForFuncSimilarity()) and computeSampleFunctionInternalOverlap() into one computeSampleFunctionOverlap() function. In addition, I added comments to weightForFuncSimilarity() and weightByImportance() to explain the different purpose of these functions.

Thank you for pointing this out!

[hoyFB]

llvm/tools/llvm-profdata/llvm-profdata.cpp

811–817

Can you please decorate these functions with const?

[weihe]

Added const keyword to member functions of MatchStep.

[weihe]

llvm/tools/llvm-profdata/llvm-profdata.cpp

811–817

Thank you for this suggestion! I have added const keyword to the member functions.

[MaskRay]

llvm/test/tools/llvm-profdata/sample-overlap.test

1

You may consider a new test utility split-file, which can group multiple auxiliary files. D83834

[wenlei]

Thank you for working on this during internship, @weihe! The extra tweaks here on top of our internal version look good to me as well. The test failure doesn't seem related, I will rebase and land this on your behalf.

llvm/test/tools/llvm-profdata/sample-overlap.test

1

Thanks for the pointer - I wasn't aware of the recently added utility. That will work, but in this particular case, I think it's still better to keep profile inputs separate so they're semantically legal afdo profile by themselves.

[aganea]

Hello @wenlei, @weihe,

llvm/tools/llvm-profdata/llvm-profdata.cpp

1399

Match is invalidated after this line, so it cannot be compared with BaseFuncProf.end() afterwards at L1607 and L1609.

In a Debug build on Windows/MSVC this asserts in MS-STL:

The following tests fail because of this:

LLVM :: tools/llvm-profdata/compact-sample.proftext
LLVM :: tools/llvm-profdata/sample-overlap.test

The following patch seems to fi the issue, but I thought I'll let you decide what to do?

diff --git a/llvm/tools/llvm-profdata/llvm-profdata.cpp b/llvm/tools/llvm-profdata/llvm-profdata.cpp
index 488dc8fa4317..38d9cb9461bb 100644
--- a/llvm/tools/llvm-profdata/llvm-profdata.cpp
+++ b/llvm/tools/llvm-profdata/llvm-profdata.cpp
@@ -1633,6 +1633,7 @@ void SampleOverlapAggregator::computeSampleProfileOverlap(raw_fd_ostream &OS) {
            "except inlinees");
     FuncOverlap.TestSample = TestStats[FuncOverlap.TestName].SampleSum;

+    bool Matched = false;
     const auto Match = BaseFuncProf.find(FuncOverlap.TestName);
     if (Match == BaseFuncProf.end()) {
       const FuncSampleStats &FuncStats = TestStats[FuncOverlap.TestName];
@@ -1677,6 +1678,8 @@ void SampleOverlapAggregator::computeSampleProfileOverlap(raw_fd_ostream &OS) {
       // Remove matched base functions for later reporting functions not found
       // in test profile.
       BaseFuncProf.erase(Match);
+
+      Matched = true;
     }

     // Print function-level similarity information if specified by options.
@@ -1684,9 +1687,8 @@ void SampleOverlapAggregator::computeSampleProfileOverlap(raw_fd_ostream &OS) {
            "TestStats should have records for all functions in test profile "
            "except inlinees");
     if (TestStats[FuncOverlap.TestName].MaxSample >= FuncFilter.ValueCutoff ||
-        (Match != BaseFuncProf.end() &&
-         FuncOverlap.Similarity < LowSimilarityThreshold) ||
-        (Match != BaseFuncProf.end() && !FuncFilter.NameFilter.empty() &&
+        (Matched && FuncOverlap.Similarity < LowSimilarityThreshold) ||
+        (Matched && !FuncFilter.NameFilter.empty() &&
          FuncOverlap.BaseName.toString().find(FuncFilter.NameFilter) !=
              std::string::npos)) {
       assert(ProfOverlap.BaseSample > 0 &&

[wenlei]

llvm/tools/llvm-profdata/llvm-profdata.cpp

1399

Good catch, thanks! I will fix it as you suggested.

llvm/test/tools/llvm-profdata/Inputs/sample-overlap-0.proftext

+_Z3bari:20301:1437
+ 1: 1437
+_Z3fooi:7711:610
+ 1: 610
+main:184019:0
+ 4: 534
+ 4.2: 534
+ 5: 1075
+ 5.1: 1075
+ 6: 2080
+ 7: 534
+ 9: 2064 _Z3bari:1471 _Z3fooi:631
+ 10: inline1:1000
+  1: 1000
+ 10: inline2:2000
+  1: 2000
+_Z3bazi:20301:1000
+ 1: 1000

llvm/test/tools/llvm-profdata/Inputs/sample-overlap-1.proftext

+_Z3bari:203010:14370
+ 1: 14370
+_Z3fooi:77110:6100
+ 1: 6100
+main:1840190:0
+ 4: 5340
+ 4.2: 5340
+ 5: 10750
+ 5.1: 10750
+ 6: 20800
+ 7: 5340
+ 9: 20640 _Z3bari:14710 _Z3fooi:6310
+ 10: inline1:10000
+  1: 10000
+ 10: inline2:20000
+  1: 20000
+_Z3bazi:203010:10000
+ 1: 10000

llvm/test/tools/llvm-profdata/Inputs/sample-overlap-2.proftext

+_Z3bari:20301:1437
+ 1: 1437
+_Z3fooi:7711:610
+ 1: 610
+main:18401:0
+ 4: 53
+ 4.2: 53
+ 5: 107
+ 5.1: 107
+ 6: 208
+ 7: 53
+ 9: 206 _Z3bari:1471 _Z3fooi:631
+ 10: inline1:100
+  1: 100
+ 10: inline2:200
+  1: 200
+_Z3bazi:20301:1000
+ 1: 1000

llvm/test/tools/llvm-profdata/Inputs/sample-overlap-3.proftext

+_Z3bari:20301:1437
+ 1: 1437
+_Z3fooi2:7711:610
+ 1: 610
+main2:184019:0
+ 4: 534
+ 4.2: 534
+ 5: 1075
+ 5.1: 1075
+ 6: 2080
+ 7: 534
+ 9: 2064 _Z3bari:1471 _Z3fooi:631
+ 10: inline1:1000
+  1: 1000
+ 10: inline2:2000
+  1: 2000
+_Z3bazi:20301:1000
+ 1: 100

llvm/test/tools/llvm-profdata/Inputs/sample-overlap-4.proftext

+_Z3bari:20301:1437
+ 2: 1437
+_Z3fooi:7711:610
+ 2: 610
+main:184019:0
+ 5: 534
+ 5.2: 534
+ 6: 1075
+ 6.1: 1075
+ 7: 208
+ 8: 534
+ 10: 206 _Z3bari:1471 _Z3fooi:631
+ 11: inline1:1000
+  1: 1000
+ 11: inline2:2000
+  1: 2000
+_Z3bazi:20301:1000
+ 2: 1000

llvm/test/tools/llvm-profdata/Inputs/sample-overlap-5.proftext

+_Z3bari:0:0
+ 1: 0
+_Z3fooi:0:0
+ 1: 0
+main:0:0
+ 4: 0
+ 4.2: 0
+ 5: 0
+ 5.1: 0
+ 6: 0
+ 7: 0
+ 9: 0
+ 10: inline1:0
+  1: 0
+ 10: inline2:0
+  1: 0
+_Z3bazi:0:0
+ 1: 0

llvm/test/tools/llvm-profdata/sample-overlap.test

+; RUN: llvm-profdata overlap --sample %S/Inputs/sample-overlap-0.proftext %S/Inputs/sample-overlap-0.proftext | FileCheck %s --check-prefix=OVERLAP0  --match-full-lines --strict-whitespace
+; OVERLAP0:Program level:
+; OVERLAP0:  Whole program profile similarity: 100.000%
+; OVERLAP0:  Whole program sample overlap: 100.000%
+; OVERLAP0:    percentage of samples unique in base profile: 0.000%
+; OVERLAP0:    percentage of samples unique in test profile: 0.000%
+; OVERLAP0:    total samples in base profile: 13943
+; OVERLAP0:    total samples in test profile: 13943
+; OVERLAP0:  Function overlap: 100.000%
+; OVERLAP0:    overlap functions: 4
+; OVERLAP0:    functions unique in base profile: 0
+; OVERLAP0:    functions unique in test profile: 0
+; OVERLAP0:  Hot-function overlap: 100.000%
+; OVERLAP0:    overlap hot functions: 4
+; OVERLAP0:    hot functions unique in base profile: 0
+; OVERLAP0:    hot functions unique in test profile: 0
+; OVERLAP0:  Hot-block overlap: 100.000%
+; OVERLAP0:    overlap hot blocks: 12
+; OVERLAP0:    hot blocks unique in base profile: 0
+; OVERLAP0:    hot blocks unique in test profile: 0
+
+; RUN: llvm-profdata overlap --sample %S/Inputs/sample-overlap-0.proftext %S/Inputs/sample-overlap-1.proftext | FileCheck %s --check-prefix=OVERLAP1  --match-full-lines --strict-whitespace
+; OVERLAP1:Program level:
+; OVERLAP1:  Whole program profile similarity: 100.000%
+; OVERLAP1:  Whole program sample overlap: 10.000%
+; OVERLAP1:    percentage of samples unique in base profile: 0.000%
+; OVERLAP1:    percentage of samples unique in test profile: 0.000%
+; OVERLAP1:    total samples in base profile: 13943
+; OVERLAP1:    total samples in test profile: 139430
+; OVERLAP1:  Function overlap: 100.000%
+; OVERLAP1:    overlap functions: 4
+; OVERLAP1:    functions unique in base profile: 0
+; OVERLAP1:    functions unique in test profile: 0
+; OVERLAP1:  Hot-function overlap: 100.000%
+; OVERLAP1:    overlap hot functions: 4
+; OVERLAP1:    hot functions unique in base profile: 0
+; OVERLAP1:    hot functions unique in test profile: 0
+; OVERLAP1:  Hot-block overlap: 100.000%
+; OVERLAP1:    overlap hot blocks: 12
+; OVERLAP1:    hot blocks unique in base profile: 0
+; OVERLAP1:    hot blocks unique in test profile: 0
+
+; RUN: llvm-profdata overlap --sample --similarity-cutoff=800000 %S/Inputs/sample-overlap-0.proftext %S/Inputs/sample-overlap-2.proftext | FileCheck %s --check-prefix=OVERLAP2  --match-full-lines --strict-whitespace
+; OVERLAP2:Program level:
+; OVERLAP2:  Whole program profile similarity: 63.720%
+; OVERLAP2:  Whole program sample overlap: 29.649%
+; OVERLAP2:    percentage of samples unique in base profile: 0.000%
+; OVERLAP2:    percentage of samples unique in test profile: 0.000%
+; OVERLAP2:    total samples in base profile: 13943
+; OVERLAP2:    total samples in test profile: 4134
+; OVERLAP2:  Function overlap: 100.000%
+; OVERLAP2:    overlap functions: 4
+; OVERLAP2:    functions unique in base profile: 0
+; OVERLAP2:    functions unique in test profile: 0
+; OVERLAP2:  Hot-function overlap: 100.000%
+; OVERLAP2:    overlap hot functions: 4
+; OVERLAP2:    hot functions unique in base profile: 0
+; OVERLAP2:    hot functions unique in test profile: 0
+; OVERLAP2:  Hot-block overlap: 100.000%
+; OVERLAP2:    overlap hot blocks: 12
+; OVERLAP2:    hot blocks unique in base profile: 0
+; OVERLAP2:    hot blocks unique in test profile: 0
+; OVERLAP2:Function-level details:
+; OVERLAP2:Base weight    Test weight    Similarity   Overlap   Base unique   Test unique   Base samples   Test samples   Function name
+; OVERLAP2:78.15%         26.29%         48.09%       9.98%     0.00%         0.00%         10896          1087           main
+; OVERLAP2:10.31%         34.76%         75.55%       100.00%   0.00%         0.00%         1437           1437           _Z3bari
+
+; RUN: llvm-profdata overlap --sample --value-cutoff=1000 %S/Inputs/sample-overlap-0.proftext %S/Inputs/sample-overlap-3.proftext | FileCheck %s --check-prefix=OVERLAP3  --match-full-lines --strict-whitespace
+; OVERLAP3:Program level:
+; OVERLAP3:  Whole program profile similarity: 14.301%
+; OVERLAP3:  Whole program sample overlap: 6.040%
+; OVERLAP3:    percentage of samples unique in base profile: 82.522%
+; OVERLAP3:    percentage of samples unique in test profile: 88.216%
+; OVERLAP3:    total samples in base profile: 13943
+; OVERLAP3:    total samples in test profile: 13043
+; OVERLAP3:  Function overlap: 33.333%
+; OVERLAP3:    overlap functions: 2
+; OVERLAP3:    functions unique in base profile: 2
+; OVERLAP3:    functions unique in test profile: 2
+; OVERLAP3:  Hot-function overlap: 16.667%
+; OVERLAP3:    overlap hot functions: 1
+; OVERLAP3:    hot functions unique in base profile: 3
+; OVERLAP3:    hot functions unique in test profile: 2
+; OVERLAP3:  Hot-block overlap: 4.545%
+; OVERLAP3:    overlap hot blocks: 1
+; OVERLAP3:    hot blocks unique in base profile: 11
+; OVERLAP3:    hot blocks unique in test profile: 10
+; OVERLAP3:Function-level details:
+; OVERLAP3:Base weight    Test weight    Similarity   Overlap   Base unique   Test unique   Base samples   Test samples   Function name
+; OVERLAP3:10.31%         11.02%         99.29%       100.00%   0.00%         0.00%         1437           1437           _Z3bari
+; OVERLAP3:0.00%          83.54%         0.00%        0.00%     0.00%         100.00%       0              10896          main2
+
+; RUN: llvm-profdata overlap --sample --function=main %S/Inputs/sample-overlap-0.proftext %S/Inputs/sample-overlap-4.proftext | FileCheck %s --check-prefix=OVERLAP4  --match-full-lines --strict-whitespace
+; OVERLAP4:Program level:
+; OVERLAP4:  Whole program profile similarity: 17.302%
+; OVERLAP4:  Whole program sample overlap: 8.134%
+; OVERLAP4:    percentage of samples unique in base profile: 73.542%
+; OVERLAP4:    percentage of samples unique in test profile: 82.209%
+; OVERLAP4:    total samples in base profile: 13943
+; OVERLAP4:    total samples in test profile: 10213
+; OVERLAP4:  Function overlap: 100.000%
+; OVERLAP4:    overlap functions: 4
+; OVERLAP4:    functions unique in base profile: 0
+; OVERLAP4:    functions unique in test profile: 0
+; OVERLAP4:  Hot-function overlap: 100.000%
+; OVERLAP4:    overlap hot functions: 4
+; OVERLAP4:    hot functions unique in base profile: 0
+; OVERLAP4:    hot functions unique in test profile: 0
+; OVERLAP4:  Hot-block overlap: 14.286%
+; OVERLAP4:    overlap hot blocks: 3
+; OVERLAP4:    hot blocks unique in base profile: 9
+; OVERLAP4:    hot blocks unique in test profile: 9
+; OVERLAP4:Function-level details:
+; OVERLAP4:Base weight    Test weight    Similarity   Overlap   Base unique   Test unique   Base samples   Test samples   Function name
+; OVERLAP4:78.15%         70.17%         23.33%       11.18%    66.14%        74.64%        10896          7166           main
+
+; RUN: llvm-profdata overlap --sample %S/Inputs/sample-overlap-0.proftext %S/Inputs/sample-overlap-5.proftext | FileCheck %s --check-prefix=OVERLAP5  --match-full-lines --strict-whitespace
+; OVERLAP5:Sum of sample counts for profile {{.*}}/Inputs/sample-overlap-5.proftext is 0.

llvm/tools/llvm-profdata/llvm-profdata.cpp

if (Overlap.Test.CountSum < 1.0f) {
     exit(0);
   }
   loadInput(WeightedInput, nullptr, &Context);
   overlapInput(BaseFilename, TestFilename, &Context, Overlap, FuncFilter, OS,
                IsCS);
   Overlap.dump(OS);
 }
 
+namespace {
+struct SampleOverlapStats {
+  StringRef BaseName;
+  StringRef TestName;
+  // Number of overlap units
+  uint64_t OverlapCount;
+  // Total samples of overlap units
+  uint64_t OverlapSample;
+  // Number of and total samples of units that only present in base or test
+  // profile
+  uint64_t BaseUniqueCount;
+  uint64_t BaseUniqueSample;
+  uint64_t TestUniqueCount;
+  uint64_t TestUniqueSample;
+  // Number of units and total samples in base or test profile
+  uint64_t BaseCount;
+  uint64_t BaseSample;
+  uint64_t TestCount;
+  uint64_t TestSample;
+  // Number of and total samples of units that present in at least one profile
+  uint64_t UnionCount;
+  uint64_t UnionSample;
+  // Weighted similarity
+  double Similarity;
+  // For SampleOverlapStats instances representing functions, weights of the
+  // function in base and test profiles
+  double BaseWeight;
+  double TestWeight;
+
+  SampleOverlapStats()
+      : OverlapCount(0), OverlapSample(0), BaseUniqueCount(0),
+        BaseUniqueSample(0), TestUniqueCount(0), TestUniqueSample(0),
+        BaseCount(0), BaseSample(0), TestCount(0), TestSample(0), UnionCount(0),
+        UnionSample(0), Similarity(0.0), BaseWeight(0.0), TestWeight(0.0) {}
+};
+} // end anonymous namespace
+
+namespace {
+struct FuncSampleStats {
+  uint64_t SampleSum;
+  uint64_t MaxSample;
+  uint64_t HotBlockCount;
+  FuncSampleStats() : SampleSum(0), MaxSample(0), HotBlockCount(0) {}
+  FuncSampleStats(uint64_t SampleSum, uint64_t MaxSample,
+                  uint64_t HotBlockCount)
+      : SampleSum(SampleSum), MaxSample(MaxSample),
+        HotBlockCount(HotBlockCount) {}
+};
+} // end anonymous namespace
+
+namespace {
+enum MatchStatus { MS_Match, MS_FirstUnique, MS_SecondUnique, MS_None };
+
+// Class for updating merging steps for two sorted maps. The class should be
+// instantiated with a map iterator type.
+template <class T> class MatchStep {
+public:
+  MatchStep() = delete;
+
+  MatchStep(T FirstIter, T FirstEnd, T SecondIter, T SecondEnd)
+      : FirstIter(FirstIter), FirstEnd(FirstEnd), SecondIter(SecondIter),
+        SecondEnd(SecondEnd), Status(MS_None) {}
+
+  bool areBothFinished() const {
+    return (FirstIter == FirstEnd && SecondIter == SecondEnd);
+  }
+
+  bool isFirstFinished() const { return FirstIter == FirstEnd; }
+
+  bool isSecondFinished() const { return SecondIter == SecondEnd; }
+
+  /// Advance one step based on the previous match status unless the previous
+  /// status is MS_None. Then update Status based on the comparison between two
+  /// container iterators at the current step. If the previous status is
+  /// MS_None, it means two iterators are at the beginning and no comparison has
+  /// been made, so we simply update Status without advancing the iterators.
+  void updateOneStep();
+
+  T getFirstIter() const { return FirstIter; }
+
+  T getSecondIter() const { return SecondIter; }
+
+  MatchStatus getMatchStatus() const { return Status; }
+
+private:
+  // Current iterator and end iterator of the first container.
+  T FirstIter;
+  T FirstEnd;
+  // Current iterator and end iterator of the second container.
+  T SecondIter;
+  T SecondEnd;
+  // Match status of the current step.
+  MatchStatus Status;
+};
+} // end anonymous namespace
+
+template <class T> void MatchStep<T>::updateOneStep() {
+  switch (Status) {
+  case MS_Match:
+    ++FirstIter;
+    ++SecondIter;
+    break;
+  case MS_FirstUnique:
+    ++FirstIter;
+    break;
+  case MS_SecondUnique:
+    ++SecondIter;
+    break;
+  case MS_None:
+    break;
+  }
+
+  // Update Status according to iterators at the current step.
+  if (areBothFinished())
+    return;
+  if (FirstIter != FirstEnd &&
+      (SecondIter == SecondEnd || FirstIter->first < SecondIter->first))
+    Status = MS_FirstUnique;
+  else if (SecondIter != SecondEnd &&
+           (FirstIter == FirstEnd || SecondIter->first < FirstIter->first))
+    Status = MS_SecondUnique;
+  else
+    Status = MS_Match;
+}
+
+// Return the sum of line/block samples, the max line/block sample, and the
+// number of line/block samples above the given threshold in a function
+// including its inlinees.
+static void getFuncSampleStats(const sampleprof::FunctionSamples &Func,
+                               FuncSampleStats &FuncStats,
+                               uint64_t HotThreshold) {
+  for (const auto &L : Func.getBodySamples()) {
+    uint64_t Sample = L.second.getSamples();
+    FuncStats.SampleSum += Sample;
+    FuncStats.MaxSample = std::max(FuncStats.MaxSample, Sample);
+    if (Sample >= HotThreshold)
+      ++FuncStats.HotBlockCount;
+  }
+
+  for (const auto &C : Func.getCallsiteSamples()) {
+    for (const auto &F : C.second)
+      getFuncSampleStats(F.second, FuncStats, HotThreshold);
+  }
+}
+
+/// Predicate that determines if a function is hot with a given threshold. We
+/// keep it separate from its callsites for possible extension in the future.
+static bool isFunctionHot(const FuncSampleStats &FuncStats,
+                          uint64_t HotThreshold) {
+  // We intentionally compare the maximum sample count in a function with the
+  // HotThreshold to get an approximate determination on hot functions.
+  return (FuncStats.MaxSample >= HotThreshold);
+}
+
+namespace {
+class SampleOverlapAggregator {
+public:
+  SampleOverlapAggregator(const std::string &BaseFilename,
+                          const std::string &TestFilename,
+                          double LowSimilarityThreshold, double Epsilon,
+                          const OverlapFuncFilters &FuncFilter)
+      : BaseFilename(BaseFilename), TestFilename(TestFilename),
+        LowSimilarityThreshold(LowSimilarityThreshold), Epsilon(Epsilon),
+        FuncFilter(FuncFilter) {}
+
+  /// Detect 0-sample input profile and report to output stream. This interface
+  /// should be called after loadProfiles().
+  bool detectZeroSampleProfile(raw_fd_ostream &OS) const;
+
+  /// Write out function-level similarity statistics for functions specified by
+  /// options --function, --value-cutoff, and --similarity-cutoff.
+  void dumpFuncSimilarity(raw_fd_ostream &OS) const;
+
+  /// Write out program-level similarity and overlap statistics.
+  void dumpProgramSummary(raw_fd_ostream &OS) const;
+
+  /// Write out hot-function and hot-block statistics for base_profile,
+  /// test_profile, and their overlap. For both cases, the overlap HO is
+  /// calculated as follows:
+  ///    Given the number of functions (or blocks) that are hot in both profiles
+  ///    HCommon and the number of functions (or blocks) that are hot in at
+  ///    least one profile HUnion, HO = HCommon / HUnion.
+  void dumpHotFuncAndBlockOverlap(raw_fd_ostream &OS) const;
+
+  /// This function tries matching functions in base and test profiles. For each
+  /// pair of matched functions, it aggregates the function-level
+  /// similarity into a profile-level similarity. It also dump function-level
+  /// similarity information of functions specified by --function,
+  /// --value-cutoff, and --similarity-cutoff options. The program-level
+  /// similarity PS is computed as follows:
+  ///     Given function-level similarity FS(A) for all function A, the
+  ///     weight of function A in base profile WB(A), and the weight of function
+  ///     A in test profile WT(A), compute PS(base_profile, test_profile) =
+  ///     sum_A(FS(A) * avg(WB(A), WT(A))) ranging in [0.0f to 1.0f] with 0.0
+  ///     meaning no-overlap.
+  void computeSampleProfileOverlap(raw_fd_ostream &OS);
+
+  /// Initialize ProfOverlap with the sum of samples in base and test
+  /// profiles. This function also computes and keeps the sum of samples and
+  /// max sample counts of each function in BaseStats and TestStats for later
+  /// use to avoid re-computations.
+  void initializeSampleProfileOverlap();
+
+  /// Load profiles specified by BaseFilename and TestFilename.
+  std::error_code loadProfiles();
+
+private:
+  SampleOverlapStats ProfOverlap;
+  SampleOverlapStats HotFuncOverlap;
+  SampleOverlapStats HotBlockOverlap;
+  std::string BaseFilename;
+  std::string TestFilename;
+  std::unique_ptr<sampleprof::SampleProfileReader> BaseReader;
+  std::unique_ptr<sampleprof::SampleProfileReader> TestReader;
+  // BaseStats and TestStats hold FuncSampleStats for each function, with
+  // function name as the key.
+  StringMap<FuncSampleStats> BaseStats;
+  StringMap<FuncSampleStats> TestStats;
+  // Low similarity threshold in floating point number
+  double LowSimilarityThreshold;
+  // Block samples above BaseHotThreshold or TestHotThreshold are considered hot
+  // for tracking hot blocks.
+  uint64_t BaseHotThreshold;
+  uint64_t TestHotThreshold;
+  // A small threshold used to round the results of floating point accumulations
+  // to resolve imprecision.
+  const double Epsilon;
+  std::multimap<double, SampleOverlapStats, std::greater<double>>
+      FuncSimilarityDump;
+  // FuncFilter carries specifications in options --value-cutoff and
+  // --function.
+  OverlapFuncFilters FuncFilter;
+  // Column offsets for printing the function-level details table.
+  static const unsigned int TestWeightCol = 15;
+  static const unsigned int SimilarityCol = 30;
+  static const unsigned int OverlapCol = 43;
+  static const unsigned int BaseUniqueCol = 53;
+  static const unsigned int TestUniqueCol = 67;
+  static const unsigned int BaseSampleCol = 81;
+  static const unsigned int TestSampleCol = 96;
+  static const unsigned int FuncNameCol = 111;
+
+  /// Return a similarity of two line/block sample counters in the same
+  /// function in base and test profiles. The line/block-similarity BS(i) is
+  /// computed as follows:
+  ///    For an offsets i, given the sample count at i in base profile BB(i),
+  ///    the sample count at i in test profile BT(i), the sum of sample counts
+  ///    in this function in base profile SB, and the sum of sample counts in
+  ///    this function in test profile ST, compute BS(i) = 1.0 - fabs(BB(i)/SB -
+  ///    BT(i)/ST), ranging in [0.0f to 1.0f] with 0.0 meaning no-overlap.
+  double computeBlockSimilarity(uint64_t BaseSample, uint64_t TestSample,
+                                const SampleOverlapStats &FuncOverlap) const;
+
+  void updateHotBlockOverlap(uint64_t BaseSample, uint64_t TestSample,
+                             uint64_t HotBlockCount);
+
+  void getHotFunctions(const StringMap<FuncSampleStats> &ProfStats,
+                       StringMap<FuncSampleStats> &HotFunc,
+                       uint64_t HotThreshold) const;
+
+  void computeHotFuncOverlap();
+
+  /// This function updates statistics in FuncOverlap, HotBlockOverlap, and
+  /// Difference for two sample units in a matched function according to the
+  /// given match status.
+  void updateOverlapStatsForFunction(uint64_t BaseSample, uint64_t TestSample,
+                                     uint64_t HotBlockCount,
+                                     SampleOverlapStats &FuncOverlap,
+                                     double &Difference, MatchStatus Status);
+
+  /// This function updates statistics in FuncOverlap, HotBlockOverlap, and
+  /// Difference for unmatched callees that only present in one profile in a
+  /// matched caller function.
+  void updateForUnmatchedCallee(const sampleprof::FunctionSamples &Func,
+                                SampleOverlapStats &FuncOverlap,
+                                double &Difference, MatchStatus Status);
+
+  /// This function updates sample overlap statistics of an overlap function in
+  /// base and test profile. It also calculates a function-internal similarity
+  /// FIS as follows:
+  ///    For offsets i that have samples in at least one profile in this
+  ///    function A, given BS(i) returned by computeBlockSimilarity(), compute
+  ///    FIS(A) = (2.0 - sum_i(1.0 - BS(i))) / 2, ranging in [0.0f to 1.0f] with
+  ///    0.0 meaning no overlap.
+  double computeSampleFunctionInternalOverlap(
+      const sampleprof::FunctionSamples &BaseFunc,
+      const sampleprof::FunctionSamples &TestFunc,
+      SampleOverlapStats &FuncOverlap);
+
+  /// Function-level similarity (FS) is a weighted value over function internal
+  /// similarity (FIS). This function computes a function's FS from its FIS by
+  /// applying the weight.
+  double weightForFuncSimilarity(double FuncSimilarity, uint64_t BaseFuncSample,
+                                 uint64_t TestFuncSample) const;
+
+  /// The function-level similarity FS(A) for a function A is computed as
+  /// follows:
+  ///     Compute a function-internal similarity FIS(A) by
+  ///     computeSampleFunctionInternalOverlap(). Then, with the weight of
+  ///     function A in base profile WB(A), and the weight of function A in test
+  ///     profile WT(A), compute FS(A) = FIS(A) * (1.0 - fabs(WB(A) - WT(A)))
+  ///     ranging in [0.0f to 1.0f] with 0.0 meaning no overlap.
+  double
+  computeSampleFunctionOverlap(const sampleprof::FunctionSamples *BaseFunc,
+                               const sampleprof::FunctionSamples *TestFunc,
+                               SampleOverlapStats *FuncOverlap,
+                               uint64_t BaseFuncSample,
+                               uint64_t TestFuncSample);
+
+  /// Profile-level similarity (PS) is a weighted aggregate over function-level
+  /// similarities (FS). This method weights the FS value by the function
+  /// weights in the base and test profiles for the aggregation.
+  double weightByImportance(double FuncSimilarity, uint64_t BaseFuncSample,
+                            uint64_t TestFuncSample) const;
+};
+} // end anonymous namespace
+
+bool SampleOverlapAggregator::detectZeroSampleProfile(
+    raw_fd_ostream &OS) const {
+  bool HaveZeroSample = false;
+  if (ProfOverlap.BaseSample == 0) {
+    OS << "Sum of sample counts for profile " << BaseFilename << " is 0.\n";
+    HaveZeroSample = true;
+  }
+  if (ProfOverlap.TestSample == 0) {
+    OS << "Sum of sample counts for profile " << TestFilename << " is 0.\n";
+    HaveZeroSample = true;
+  }
+  return HaveZeroSample;
+}
+
+double SampleOverlapAggregator::computeBlockSimilarity(
+    uint64_t BaseSample, uint64_t TestSample,
+    const SampleOverlapStats &FuncOverlap) const {
+  double BaseFrac = 0.0;
+  double TestFrac = 0.0;
+  if (FuncOverlap.BaseSample > 0)
+    BaseFrac = static_cast<double>(BaseSample) / FuncOverlap.BaseSample;
+  if (FuncOverlap.TestSample > 0)
+    TestFrac = static_cast<double>(TestSample) / FuncOverlap.TestSample;
+  return 1.0 - std::fabs(BaseFrac - TestFrac);
+}
+
+void SampleOverlapAggregator::updateHotBlockOverlap(uint64_t BaseSample,
+                                                    uint64_t TestSample,
+                                                    uint64_t HotBlockCount) {
+  bool IsBaseHot = (BaseSample >= BaseHotThreshold);
+  bool IsTestHot = (TestSample >= TestHotThreshold);
+  if (!IsBaseHot && !IsTestHot)
+    return;
+
+  HotBlockOverlap.UnionCount += HotBlockCount;
+  if (IsBaseHot)
+    HotBlockOverlap.BaseCount += HotBlockCount;
+  if (IsTestHot)
+    HotBlockOverlap.TestCount += HotBlockCount;
+  if (IsBaseHot && IsTestHot)
+    HotBlockOverlap.OverlapCount += HotBlockCount;
+}
+
+void SampleOverlapAggregator::getHotFunctions(
+    const StringMap<FuncSampleStats> &ProfStats,
+    StringMap<FuncSampleStats> &HotFunc, uint64_t HotThreshold) const {
+  for (const auto &F : ProfStats) {
+    if (isFunctionHot(F.second, HotThreshold))
+      HotFunc.try_emplace(F.first(), F.second);
+  }
+}
+
+void SampleOverlapAggregator::computeHotFuncOverlap() {
+  StringMap<FuncSampleStats> BaseHotFunc;
+  getHotFunctions(BaseStats, BaseHotFunc, BaseHotThreshold);
+  HotFuncOverlap.BaseCount = BaseHotFunc.size();
+
+  StringMap<FuncSampleStats> TestHotFunc;
+  getHotFunctions(TestStats, TestHotFunc, TestHotThreshold);
+  HotFuncOverlap.TestCount = TestHotFunc.size();
+  HotFuncOverlap.UnionCount = HotFuncOverlap.TestCount;
+
+  for (const auto &F : BaseHotFunc) {
+    if (TestHotFunc.count(F.first()))
+      ++HotFuncOverlap.OverlapCount;
+    else
+      ++HotFuncOverlap.UnionCount;
+  }
+}
+
+void SampleOverlapAggregator::updateOverlapStatsForFunction(
+    uint64_t BaseSample, uint64_t TestSample, uint64_t HotBlockCount,
+    SampleOverlapStats &FuncOverlap, double &Difference, MatchStatus Status) {
+  assert(Status != MS_None &&
+         "Match status should be updated before updating overlap statistics");
+  if (Status == MS_FirstUnique) {
+    TestSample = 0;
+    FuncOverlap.BaseUniqueSample += BaseSample;
+  } else if (Status == MS_SecondUnique) {
+    BaseSample = 0;
+    FuncOverlap.TestUniqueSample += TestSample;
+  } else {
+    ++FuncOverlap.OverlapCount;
+  }
+
+  FuncOverlap.UnionSample += std::max(BaseSample, TestSample);
+  FuncOverlap.OverlapSample += std::min(BaseSample, TestSample);
+  Difference +=
+      1.0 - computeBlockSimilarity(BaseSample, TestSample, FuncOverlap);
+  updateHotBlockOverlap(BaseSample, TestSample, HotBlockCount);
+}
+
+void SampleOverlapAggregator::updateForUnmatchedCallee(
+    const sampleprof::FunctionSamples &Func, SampleOverlapStats &FuncOverlap,
+    double &Difference, MatchStatus Status) {
+  assert((Status == MS_FirstUnique || Status == MS_SecondUnique) &&
+         "Status must be either of the two unmatched cases");
+  FuncSampleStats FuncStats;
+  if (Status == MS_FirstUnique) {
+    getFuncSampleStats(Func, FuncStats, BaseHotThreshold);
+    updateOverlapStatsForFunction(FuncStats.SampleSum, 0,
+                                  FuncStats.HotBlockCount, FuncOverlap,
+                                  Difference, Status);
+  } else {
+    getFuncSampleStats(Func, FuncStats, TestHotThreshold);
+    updateOverlapStatsForFunction(0, FuncStats.SampleSum,
+                                  FuncStats.HotBlockCount, FuncOverlap,
+                                  Difference, Status);
+  }
+}
+
+double SampleOverlapAggregator::computeSampleFunctionInternalOverlap(
+    const sampleprof::FunctionSamples &BaseFunc,
+    const sampleprof::FunctionSamples &TestFunc,
+    SampleOverlapStats &FuncOverlap) {
+
+  using namespace sampleprof;
+
+  double Difference = 0;
+
+  // Accumulate Difference for regular line/block samples in the function.
+  // We match them through sort-merge join algorithm because
+  // FunctionSamples::getBodySamples() returns a map of sample counters ordered
+  // by their offsets.
+  MatchStep<BodySampleMap::const_iterator> BlockIterStep(
+      BaseFunc.getBodySamples().cbegin(), BaseFunc.getBodySamples().cend(),
+      TestFunc.getBodySamples().cbegin(), TestFunc.getBodySamples().cend());
+  BlockIterStep.updateOneStep();
+  while (!BlockIterStep.areBothFinished()) {
+    uint64_t BaseSample =
+        BlockIterStep.isFirstFinished()
+            ? 0
+            : BlockIterStep.getFirstIter()->second.getSamples();
+    uint64_t TestSample =
+        BlockIterStep.isSecondFinished()
+            ? 0
+            : BlockIterStep.getSecondIter()->second.getSamples();
+    updateOverlapStatsForFunction(BaseSample, TestSample, 1, FuncOverlap,
+                                  Difference, BlockIterStep.getMatchStatus());
+
+    BlockIterStep.updateOneStep();
+  }
+
+  // Accumulate Difference for callsite lines in the function. We match
+  // them through sort-merge algorithm because
+  // FunctionSamples::getCallsiteSamples() returns a map of callsite records
+  // ordered by their offsets.
+  MatchStep<CallsiteSampleMap::const_iterator> CallsiteIterStep(
+      BaseFunc.getCallsiteSamples().cbegin(),
+      BaseFunc.getCallsiteSamples().cend(),
+      TestFunc.getCallsiteSamples().cbegin(),
+      TestFunc.getCallsiteSamples().cend());
+  CallsiteIterStep.updateOneStep();
+  while (!CallsiteIterStep.areBothFinished()) {
+    MatchStatus CallsiteStepStatus = CallsiteIterStep.getMatchStatus();
+    assert(CallsiteStepStatus != MS_None &&
+           "Match status should be updated before entering loop body");
+
+    if (CallsiteStepStatus != MS_Match) {
+      auto Callsite = (CallsiteStepStatus == MS_FirstUnique)
+                          ? CallsiteIterStep.getFirstIter()
+                          : CallsiteIterStep.getSecondIter();
+      for (const auto &F : Callsite->second)
+        updateForUnmatchedCallee(F.second, FuncOverlap, Difference,
+                                 CallsiteStepStatus);
+    } else {
+      // There may be multiple inlinees at the same offset, so we need to try
+      // matching all of them. This match is implemented through sort-merge
+      // algorithm because callsite records at the same offset are ordered by
+      // function names.
+      MatchStep<FunctionSamplesMap::const_iterator> CalleeIterStep(
+          CallsiteIterStep.getFirstIter()->second.cbegin(),
+          CallsiteIterStep.getFirstIter()->second.cend(),
+          CallsiteIterStep.getSecondIter()->second.cbegin(),
+          CallsiteIterStep.getSecondIter()->second.cend());
+      CalleeIterStep.updateOneStep();
+      while (!CalleeIterStep.areBothFinished()) {
+        MatchStatus CalleeStepStatus = CalleeIterStep.getMatchStatus();
+        if (CalleeStepStatus != MS_Match) {
+          auto Callee = (CalleeStepStatus == MS_FirstUnique)
+                            ? CalleeIterStep.getFirstIter()
+                            : CalleeIterStep.getSecondIter();
+          updateForUnmatchedCallee(Callee->second, FuncOverlap, Difference,
+                                   CalleeStepStatus);
+        } else {
+          // An inlined function can contain other inlinees inside, so compute
+          // the Difference recursively.
+          Difference += 2.0 - 2 * computeSampleFunctionInternalOverlap(
+                                      CalleeIterStep.getFirstIter()->second,
+                                      CalleeIterStep.getSecondIter()->second,
+                                      FuncOverlap);
+        }
+        CalleeIterStep.updateOneStep();
+      }
+    }
+    CallsiteIterStep.updateOneStep();
+  }
+
+  // Difference reflects the total differences of line/block samples in this
+  // function and ranges in [0.0f to 2.0f]. Take (2.0 - Difference) / 2 to
+  // reflect the similarity between function profiles in [0.0f to 1.0f].
+  return (2.0 - Difference) / 2;
+}
+
+double SampleOverlapAggregator::weightForFuncSimilarity(
+    double FuncInternalSimilarity, uint64_t BaseFuncSample,
+    uint64_t TestFuncSample) const {
+  // Compute the weight as the distance between the function weights in two
+  // profiles.
+  double BaseFrac = 0.0;
+  double TestFrac = 0.0;
+  assert(ProfOverlap.BaseSample > 0 &&
+         "Total samples in base profile should be greater than 0");
+  BaseFrac = static_cast<double>(BaseFuncSample) / ProfOverlap.BaseSample;
+  assert(ProfOverlap.TestSample > 0 &&
+         "Total samples in test profile should be greater than 0");
+  TestFrac = static_cast<double>(TestFuncSample) / ProfOverlap.TestSample;
+  double WeightDistance = std::fabs(BaseFrac - TestFrac);
+
+  // Take WeightDistance into the similarity.
+  return FuncInternalSimilarity * (1 - WeightDistance);
+}
+
+double
+SampleOverlapAggregator::weightByImportance(double FuncSimilarity,
+                                            uint64_t BaseFuncSample,
+                                            uint64_t TestFuncSample) const {
+
+  double BaseFrac = 0.0;
+  double TestFrac = 0.0;
+  assert(ProfOverlap.BaseSample > 0 &&
+         "Total samples in base profile should be greater than 0");
+  BaseFrac = static_cast<double>(BaseFuncSample) / ProfOverlap.BaseSample / 2.0;
+  assert(ProfOverlap.TestSample > 0 &&
+         "Total samples in test profile should be greater than 0");
+  TestFrac = static_cast<double>(TestFuncSample) / ProfOverlap.TestSample / 2.0;
+  return FuncSimilarity * (BaseFrac + TestFrac);
+}
+
+double SampleOverlapAggregator::computeSampleFunctionOverlap(
+    const sampleprof::FunctionSamples *BaseFunc,
+    const sampleprof::FunctionSamples *TestFunc,
+    SampleOverlapStats *FuncOverlap, uint64_t BaseFuncSample,
+    uint64_t TestFuncSample) {
+  // Default function internal similarity before weighted, meaning two functions
+  // has no overlap.
+  const double DefaultFuncInternalSimilarity = 0;
+  double FuncSimilarity;
+  double FuncInternalSimilarity;
+
+  // If BaseFunc or TestFunc is nullptr, it means the functions do not overlap.
+  // In this case, we use DefaultFuncInternalSimilarity as the function internal
+  // similarity.
+  if (!BaseFunc || !TestFunc) {
+    FuncInternalSimilarity = DefaultFuncInternalSimilarity;
+  } else {
+    assert(FuncOverlap != nullptr &&
+           "FuncOverlap should be provided in this case");
+    FuncInternalSimilarity = computeSampleFunctionInternalOverlap(
+        *BaseFunc, *TestFunc, *FuncOverlap);
+    // Now, FuncInternalSimilarity may be a little less than 0 due to
+    // imprecision of floating point accumulations. Make it zero if the
+    // difference is below Epsilon.
+    FuncInternalSimilarity = (std::fabs(FuncInternalSimilarity - 0) < Epsilon)
+                                 ? 0
+                                 : FuncInternalSimilarity;
+  }
+  FuncSimilarity = weightForFuncSimilarity(FuncInternalSimilarity,
+                                           BaseFuncSample, TestFuncSample);
+  return FuncSimilarity;
+}
+
+void SampleOverlapAggregator::computeSampleProfileOverlap(raw_fd_ostream &OS) {
+  using namespace sampleprof;
+
+  StringMap<const FunctionSamples *> BaseFuncProf;
+  const auto &BaseProfiles = BaseReader->getProfiles();
+  for (const auto &BaseFunc : BaseProfiles) {
+    BaseFuncProf.try_emplace(BaseFunc.second.getFuncName(), &(BaseFunc.second));
+  }
+  ProfOverlap.UnionCount = BaseFuncProf.size();
+
+  const auto &TestProfiles = TestReader->getProfiles();
+  for (const auto &TestFunc : TestProfiles) {
+    SampleOverlapStats FuncOverlap;
+    FuncOverlap.TestName = TestFunc.second.getFuncName();
+    assert(TestStats.count(FuncOverlap.TestName) &&
+           "TestStats should have records for all functions in test profile "
+           "except inlinees");
+    FuncOverlap.TestSample = TestStats[FuncOverlap.TestName].SampleSum;
+
+    const auto Match = BaseFuncProf.find(FuncOverlap.TestName);
+    if (Match == BaseFuncProf.end()) {
+      const FuncSampleStats &FuncStats = TestStats[FuncOverlap.TestName];
+      ++ProfOverlap.TestUniqueCount;
+      ProfOverlap.TestUniqueSample += FuncStats.SampleSum;
+      FuncOverlap.TestUniqueSample = FuncStats.SampleSum;
+
+      updateHotBlockOverlap(0, FuncStats.SampleSum, FuncStats.HotBlockCount);
+
+      double FuncSimilarity = computeSampleFunctionOverlap(
+          nullptr, nullptr, nullptr, 0, FuncStats.SampleSum);
+      ProfOverlap.Similarity +=
+          weightByImportance(FuncSimilarity, 0, FuncStats.SampleSum);
+
+      ++ProfOverlap.UnionCount;
+      ProfOverlap.UnionSample += FuncStats.SampleSum;
+    } else {
+      ++ProfOverlap.OverlapCount;
+
+      // Two functions match with each other. Compute function-level overlap and
+      // aggregate them into profile-level overlap.
+      FuncOverlap.BaseName = Match->second->getFuncName();
+      assert(BaseStats.count(FuncOverlap.BaseName) &&
+             "BaseStats should have records for all functions in base profile "
+             "except inlinees");
+      FuncOverlap.BaseSample = BaseStats[FuncOverlap.BaseName].SampleSum;
+
+      FuncOverlap.Similarity = computeSampleFunctionOverlap(
+          Match->second, &TestFunc.second, &FuncOverlap, FuncOverlap.BaseSample,
+          FuncOverlap.TestSample);
+      ProfOverlap.Similarity +=
+          weightByImportance(FuncOverlap.Similarity, FuncOverlap.BaseSample,
+                             FuncOverlap.TestSample);
+      ProfOverlap.OverlapSample += FuncOverlap.OverlapSample;
+      ProfOverlap.UnionSample += FuncOverlap.UnionSample;
+
+      // Accumulate the percentage of base unique and test unique samples into
+      // ProfOverlap.
+      ProfOverlap.BaseUniqueSample += FuncOverlap.BaseUniqueSample;
+      ProfOverlap.TestUniqueSample += FuncOverlap.TestUniqueSample;
+
+      // Remove matched base functions for later reporting functions not found
+      // in test profile.
+      BaseFuncProf.erase(Match);
+    }
+
+    // Print function-level similarity information if specified by options.
+    assert(TestStats.count(FuncOverlap.TestName) &&
+           "TestStats should have records for all functions in test profile "
+           "except inlinees");
+    if (TestStats[FuncOverlap.TestName].MaxSample >= FuncFilter.ValueCutoff ||
+        (Match != BaseFuncProf.end() &&
+         FuncOverlap.Similarity < LowSimilarityThreshold) ||
+        (Match != BaseFuncProf.end() && !FuncFilter.NameFilter.empty() &&
+         FuncOverlap.BaseName.find(FuncFilter.NameFilter) !=
+             FuncOverlap.BaseName.npos)) {
+      assert(ProfOverlap.BaseSample > 0 &&
+             "Total samples in base profile should be greater than 0");
+      FuncOverlap.BaseWeight =
+          static_cast<double>(FuncOverlap.BaseSample) / ProfOverlap.BaseSample;
+      assert(ProfOverlap.TestSample > 0 &&
+             "Total samples in test profile should be greater than 0");
+      FuncOverlap.TestWeight =
+          static_cast<double>(FuncOverlap.TestSample) / ProfOverlap.TestSample;
+      FuncSimilarityDump.emplace(FuncOverlap.BaseWeight, FuncOverlap);
+    }
+  }
+
+  // Traverse through functions in base profile but not in test profile.
+  for (const auto &F : BaseFuncProf) {
+    assert(BaseStats.count(F.second->getFuncName()) &&
+           "BaseStats should have records for all functions in base profile "
+           "except inlinees");
+    const FuncSampleStats &FuncStats = BaseStats[F.second->getFuncName()];
+    ++ProfOverlap.BaseUniqueCount;
+    ProfOverlap.BaseUniqueSample += FuncStats.SampleSum;
+
+    updateHotBlockOverlap(FuncStats.SampleSum, 0, FuncStats.HotBlockCount);
+
+    double FuncSimilarity = computeSampleFunctionOverlap(
+        nullptr, nullptr, nullptr, FuncStats.SampleSum, 0);
+    ProfOverlap.Similarity +=
+        weightByImportance(FuncSimilarity, FuncStats.SampleSum, 0);
+
+    ProfOverlap.UnionSample += FuncStats.SampleSum;
+  }
+
+  // Now, ProfSimilarity may be a little greater than 1 due to imprecision
+  // of floating point accumulations. Make it 1.0 if the difference is below
+  // Epsilon.
+  ProfOverlap.Similarity = (std::fabs(ProfOverlap.Similarity - 1) < Epsilon)
+                               ? 1
+                               : ProfOverlap.Similarity;
+
+  computeHotFuncOverlap();
+}
+
+void SampleOverlapAggregator::initializeSampleProfileOverlap() {
+  const auto &BaseProf = BaseReader->getProfiles();
+  for (const auto &I : BaseProf) {
+    ++ProfOverlap.BaseCount;
+    FuncSampleStats FuncStats;
+    getFuncSampleStats(I.second, FuncStats, BaseHotThreshold);
+    ProfOverlap.BaseSample += FuncStats.SampleSum;
+    BaseStats.try_emplace(I.second.getFuncName(), FuncStats);
+  }
+
+  const auto &TestProf = TestReader->getProfiles();
+  for (const auto &I : TestProf) {
+    ++ProfOverlap.TestCount;
+    FuncSampleStats FuncStats;
+    getFuncSampleStats(I.second, FuncStats, TestHotThreshold);
+    ProfOverlap.TestSample += FuncStats.SampleSum;
+    TestStats.try_emplace(I.second.getFuncName(), FuncStats);
+  }
+
+  ProfOverlap.BaseName = StringRef(BaseFilename);
+  ProfOverlap.TestName = StringRef(TestFilename);
+}
+
+void SampleOverlapAggregator::dumpFuncSimilarity(raw_fd_ostream &OS) const {
+  using namespace sampleprof;
+
+  if (FuncSimilarityDump.empty())
+    return;
+
+  formatted_raw_ostream FOS(OS);
+  FOS << "Function-level details:\n";
+  FOS << "Base weight";
+  FOS.PadToColumn(TestWeightCol);
+  FOS << "Test weight";
+  FOS.PadToColumn(SimilarityCol);
+  FOS << "Similarity";
+  FOS.PadToColumn(OverlapCol);
+  FOS << "Overlap";
+  FOS.PadToColumn(BaseUniqueCol);
+  FOS << "Base unique";
+  FOS.PadToColumn(TestUniqueCol);
+  FOS << "Test unique";
+  FOS.PadToColumn(BaseSampleCol);
+  FOS << "Base samples";
+  FOS.PadToColumn(TestSampleCol);
+  FOS << "Test samples";
+  FOS.PadToColumn(FuncNameCol);
+  FOS << "Function name\n";
+  for (const auto &F : FuncSimilarityDump) {
+    double OverlapPercent =
+        F.second.UnionSample > 0
+            ? static_cast<double>(F.second.OverlapSample) / F.second.UnionSample
+            : 0;
+    double BaseUniquePercent =
+        F.second.BaseSample > 0
+            ? static_cast<double>(F.second.BaseUniqueSample) /
+                  F.second.BaseSample
+            : 0;
+    double TestUniquePercent =
+        F.second.TestSample > 0
+            ? static_cast<double>(F.second.TestUniqueSample) /
+                  F.second.TestSample
+            : 0;
+
+    FOS << format("%.2f%%", F.second.BaseWeight * 100);
+    FOS.PadToColumn(TestWeightCol);
+    FOS << format("%.2f%%", F.second.TestWeight * 100);
+    FOS.PadToColumn(SimilarityCol);
+    FOS << format("%.2f%%", F.second.Similarity * 100);
+    FOS.PadToColumn(OverlapCol);
+    FOS << format("%.2f%%", OverlapPercent * 100);
+    FOS.PadToColumn(BaseUniqueCol);
+    FOS << format("%.2f%%", BaseUniquePercent * 100);
+    FOS.PadToColumn(TestUniqueCol);
+    FOS << format("%.2f%%", TestUniquePercent * 100);
+    FOS.PadToColumn(BaseSampleCol);
+    FOS << F.second.BaseSample;
+    FOS.PadToColumn(TestSampleCol);
+    FOS << F.second.TestSample;
+    FOS.PadToColumn(FuncNameCol);
+    FOS << F.second.TestName << "\n";
+  }
+}
+
+void SampleOverlapAggregator::dumpProgramSummary(raw_fd_ostream &OS) const {
+  OS << "Profile overlap infomation for base_profile: " << ProfOverlap.BaseName
+     << " and test_profile: " << ProfOverlap.TestName << "\nProgram level:\n";
+
+  OS << "  Whole program profile similarity: "
+     << format("%.3f%%", ProfOverlap.Similarity * 100) << "\n";
+
+  assert(ProfOverlap.UnionSample > 0 &&
+         "Total samples in two profile should be greater than 0");
+  double OverlapPercent =
+      static_cast<double>(ProfOverlap.OverlapSample) / ProfOverlap.UnionSample;
+  assert(ProfOverlap.BaseSample > 0 &&
+         "Total samples in base profile should be greater than 0");
+  double BaseUniquePercent = static_cast<double>(ProfOverlap.BaseUniqueSample) /
+                             ProfOverlap.BaseSample;
+  assert(ProfOverlap.TestSample > 0 &&
+         "Total samples in test profile should be greater than 0");
+  double TestUniquePercent = static_cast<double>(ProfOverlap.TestUniqueSample) /
+                             ProfOverlap.TestSample;
+
+  OS << "  Whole program sample overlap: "
+     << format("%.3f%%", OverlapPercent * 100) << "\n";
+  OS << "    percentage of samples unique in base profile: "
+     << format("%.3f%%", BaseUniquePercent * 100) << "\n";
+  OS << "    percentage of samples unique in test profile: "
+     << format("%.3f%%", TestUniquePercent * 100) << "\n";
+  OS << "    total samples in base profile: " << ProfOverlap.BaseSample << "\n"
+     << "    total samples in test profile: " << ProfOverlap.TestSample << "\n";
+
+  assert(ProfOverlap.UnionCount > 0 &&
+         "There should be at least one function in two input profiles");
+  double FuncOverlapPercent =
+      static_cast<double>(ProfOverlap.OverlapCount) / ProfOverlap.UnionCount;
+  OS << "  Function overlap: " << format("%.3f%%", FuncOverlapPercent * 100)
+     << "\n";
+  OS << "    overlap functions: " << ProfOverlap.OverlapCount << "\n";
+  OS << "    functions unique in base profile: " << ProfOverlap.BaseUniqueCount
+     << "\n";
+  OS << "    functions unique in test profile: " << ProfOverlap.TestUniqueCount
+     << "\n";
+}
+
+void SampleOverlapAggregator::dumpHotFuncAndBlockOverlap(
+    raw_fd_ostream &OS) const {
+  assert(HotFuncOverlap.UnionCount > 0 &&
+         "There should be at least one hot function in two input profiles");
+  OS << "  Hot-function overlap: "
+     << format("%.3f%%", static_cast<double>(HotFuncOverlap.OverlapCount) /
+                             HotFuncOverlap.UnionCount * 100)
+     << "\n";
+  OS << "    overlap hot functions: " << HotFuncOverlap.OverlapCount << "\n";
+  OS << "    hot functions unique in base profile: "
+     << HotFuncOverlap.BaseCount - HotFuncOverlap.OverlapCount << "\n";
+  OS << "    hot functions unique in test profile: "
+     << HotFuncOverlap.TestCount - HotFuncOverlap.OverlapCount << "\n";
+
+  assert(HotBlockOverlap.UnionCount > 0 &&
+         "There should be at least one hot block in two input profiles");
+  OS << "  Hot-block overlap: "
+     << format("%.3f%%", static_cast<double>(HotBlockOverlap.OverlapCount) /
+                             HotBlockOverlap.UnionCount * 100)
+     << "\n";
+  OS << "    overlap hot blocks: " << HotBlockOverlap.OverlapCount << "\n";
+  OS << "    hot blocks unique in base profile: "
+     << HotBlockOverlap.BaseCount - HotBlockOverlap.OverlapCount << "\n";
+  OS << "    hot blocks unique in test profile: "
+     << HotBlockOverlap.TestCount - HotBlockOverlap.OverlapCount << "\n";
+}
+
+std::error_code SampleOverlapAggregator::loadProfiles() {
+  using namespace sampleprof;
+
+  LLVMContext Context;
+  auto BaseReaderOrErr = SampleProfileReader::create(BaseFilename, Context);
+  if (std::error_code EC = BaseReaderOrErr.getError())
+    exitWithErrorCode(EC, BaseFilename);
+
+  auto TestReaderOrErr = SampleProfileReader::create(TestFilename, Context);
+  if (std::error_code EC = TestReaderOrErr.getError())
+    exitWithErrorCode(EC, TestFilename);
+
+  BaseReader = std::move(BaseReaderOrErr.get());
+  TestReader = std::move(TestReaderOrErr.get());
+
+  if (std::error_code EC = BaseReader->read())
+    exitWithErrorCode(EC, BaseFilename);
+  if (std::error_code EC = TestReader->read())
+    exitWithErrorCode(EC, TestFilename);
+
+  // Load BaseHotThreshold and TestHotThreshold as 99-percentile threshold in
+  // profile summary.
+  const uint64_t HotCutoff = 990000;
+  ProfileSummary &BasePS = BaseReader->getSummary();
+  for (const auto &SummaryEntry : BasePS.getDetailedSummary()) {
+    if (SummaryEntry.Cutoff == HotCutoff) {
+      BaseHotThreshold = SummaryEntry.MinCount;
+      break;
+    }
+  }
+
+  ProfileSummary &TestPS = TestReader->getSummary();
+  for (const auto &SummaryEntry : TestPS.getDetailedSummary()) {
+    if (SummaryEntry.Cutoff == HotCutoff) {
+      TestHotThreshold = SummaryEntry.MinCount;
+      break;
+    }
+  }
+  return std::error_code();
+}
+
+void overlapSampleProfile(const std::string &BaseFilename,
+                          const std::string &TestFilename,
+                          const OverlapFuncFilters &FuncFilter,
+                          uint64_t SimilarityCutoff, raw_fd_ostream &OS) {
+  using namespace sampleprof;
+
+  // We use 0.000005 to initialize OverlapAggr.Epsilon because the final metrics
+  // report 2--3 places after decimal point in percentage numbers.
+  SampleOverlapAggregator OverlapAggr(
+      BaseFilename, TestFilename,
+      static_cast<double>(SimilarityCutoff) / 1000000, 0.000005, FuncFilter);
+  if (std::error_code EC = OverlapAggr.loadProfiles())
+    exitWithErrorCode(EC);
+
+  OverlapAggr.initializeSampleProfileOverlap();
+  if (OverlapAggr.detectZeroSampleProfile(OS))
+    return;
+
+  OverlapAggr.computeSampleProfileOverlap(OS);
+
+  OverlapAggr.dumpProgramSummary(OS);
+  OverlapAggr.dumpHotFuncAndBlockOverlap(OS);
+  OverlapAggr.dumpFuncSimilarity(OS);
+}
+
 static int overlap_main(int argc, const char *argv[]) {
   cl::opt<std::string> BaseFilename(cl::Positional, cl::Required,
                                     cl::desc("<base profile file>"));
   cl::opt<std::string> TestFilename(cl::Positional, cl::Required,
                                     cl::desc("<test profile file>"));
   cl::opt<std::string> Output("output", cl::value_desc("output"), cl::init("-"),
                               cl::desc("Output file"));
   cl::alias OutputA("o", cl::desc("Alias for --output"), cl::aliasopt(Output));
   cl::opt<bool> IsCS("cs", cl::init(false),
                      cl::desc("For context sensitive counts"));
   cl::opt<unsigned long long> ValueCutoff(
       "value-cutoff", cl::init(-1),
       cl::desc(
           "Function level overlap information for every function in test "
           "profile with max count value greater then the parameter value"));
   cl::opt<std::string> FuncNameFilter(
       "function",
       cl::desc("Function level overlap information for matching functions"));
+  cl::opt<unsigned long long> SimilarityCutoff(
+      "similarity-cutoff", cl::init(0),
+      cl::desc(
+          "For sample profiles, list function names for overlapped functions "
+          "with similarities below the cutoff (percentage times 10000)."));
+  cl::opt<ProfileKinds> ProfileKind(
+      cl::desc("Profile kind:"), cl::init(instr),
+      cl::values(clEnumVal(instr, "Instrumentation profile (default)"),
+                 clEnumVal(sample, "Sample profile")));
   cl::ParseCommandLineOptions(argc, argv, "LLVM profile data overlap tool\n");
 
   std::error_code EC;
   raw_fd_ostream OS(Output.data(), EC, sys::fs::OF_Text);
   if (EC)
     exitWithErrorCode(EC, Output);
 
+  if (ProfileKind == instr)
     overlapInstrProfile(BaseFilename, TestFilename,
                         OverlapFuncFilters{ValueCutoff, FuncNameFilter}, OS,
                         IsCS);
+  else
+    overlapSampleProfile(BaseFilename, TestFilename,
+                         OverlapFuncFilters{ValueCutoff, FuncNameFilter},
+                         SimilarityCutoff, OS);
 
   return 0;
 }
 
 typedef struct ValueSitesStats {
   ValueSitesStats()
       : TotalNumValueSites(0), TotalNumValueSitesWithValueProfile(0),
         TotalNumValues(0) {}

 // print out or let it be an empty string.
 static void dumpHotFunctionList(const std::vector<std::string> &ColumnTitle,
                                 const std::vector<int> &ColumnOffset,
                                 const std::vector<HotFuncInfo> &PrintValues,
                                 uint64_t HotFuncCount, uint64_t TotalFuncCount,
                                 uint64_t HotProfCount, uint64_t TotalProfCount,
                                 const std::string &HotFuncMetric,
                                 raw_fd_ostream &OS) {
-  assert(ColumnOffset.size() == ColumnTitle.size());
-  assert(ColumnTitle.size() >= 4);
-  assert(TotalFuncCount > 0);
+  assert(ColumnOffset.size() == ColumnTitle.size() &&
+         "ColumnOffset and ColumnTitle should have the same size");
+  assert(ColumnTitle.size() >= 4 &&
+         "ColumnTitle should have at least 4 elements");
+  assert(TotalFuncCount > 0 &&
+         "There should be at least one function in the profile");
   double TotalProfPercent = 0;
   if (TotalProfCount > 0)
-    TotalProfPercent = ((double)HotProfCount) / TotalProfCount * 100;
+    TotalProfPercent = static_cast<double>(HotProfCount) / TotalProfCount * 100;
 
   formatted_raw_ostream FOS(OS);
   FOS << HotFuncCount << " out of " << TotalFuncCount
       << " functions with profile ("
-      << format("%.2f%%", (((double)HotFuncCount) / TotalFuncCount * 100))
+      << format("%.2f%%",
+                (static_cast<double>(HotFuncCount) / TotalFuncCount * 100))
       << ") are considered hot functions";
   if (!HotFuncMetric.empty())
     FOS << " (" << HotFuncMetric << ")";
   FOS << ".\n";
   FOS << HotProfCount << " out of " << TotalProfCount << " profile counts ("
       << format("%.2f%%", TotalProfPercent) << ") are from hot functions.\n";
 
   for (size_t I = 0; I < ColumnTitle.size(); ++I) {
showHotFunctionList(const StringMap<sampleprof::FunctionSamples> &Profiles,
   auto &SummaryVector = PS.getDetailedSummary();
   uint64_t MinCountThreshold = 0;
   for (const ProfileSummaryEntry &SummaryEntry : SummaryVector) {
     if (SummaryEntry.Cutoff == HotFuncCutoff) {
       MinCountThreshold = SummaryEntry.MinCount;
       break;
     }
   }
-  assert(MinCountThreshold != 0);
 
   // Traverse all functions in the profile and keep only hot functions.
   // The following loop also calculates the sum of total samples of all
   // functions.
   std::multimap<uint64_t, std::pair<const FunctionSamples *, const uint64_t>,
                 std::greater<uint64_t>>
       HotFunc;
   uint64_t ProfileTotalSample = 0;
   uint64_t HotFuncSample = 0;
   uint64_t HotFuncCount = 0;
-  uint64_t MaxCount = 0;
+
   for (const auto &I : Profiles) {
+    FuncSampleStats FuncStats;
     const FunctionSamples &FuncProf = I.second;
     ProfileTotalSample += FuncProf.getTotalSamples();
-    MaxCount = FuncProf.getMaxCountInside();
+    getFuncSampleStats(FuncProf, FuncStats, MinCountThreshold);
 
-    // MinCountThreshold is a block/line threshold computed for a given cutoff.
-    // We intentionally compare the maximum sample count in a function with this
-    // threshold to get an approximate threshold for hot functions.
-    if (MaxCount >= MinCountThreshold) {
+    if (isFunctionHot(FuncStats, MinCountThreshold)) {
       HotFunc.emplace(FuncProf.getTotalSamples(),
-                      std::make_pair(&(I.second), MaxCount));
+                      std::make_pair(&(I.second), FuncStats.MaxSample));
       HotFuncSample += FuncProf.getTotalSamples();
       ++HotFuncCount;
     }
   }
 
   std::vector<std::string> ColumnTitle{"Total sample (%)", "Max sample",
                                        "Entry sample", "Function name"};
   std::vector<int> ColumnOffset{0, 24, 42, 58};