[llvm-profdata] Add support for weighted merge of profile data
authorNathan Slingerland <slingn@gmail.com>
Fri, 4 Dec 2015 00:00:20 +0000 (00:00 +0000)
committerNathan Slingerland <slingn@gmail.com>
Fri, 4 Dec 2015 00:00:20 +0000 (00:00 +0000)
This change adds support for an optional weight when merging profile data with the llvm-profdata tool.
Weights are specified by adding an option ':<weight>' suffix to the input file names.

Adding support for arbitrary weighting of input profile data allows for relative importance to be placed on the
input data from multiple training runs.

Both sampled and instrumented profiles are supported.

Reviewers: dnovillo, bogner, davidxl

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14547

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@254669 91177308-0d34-0410-b5e6-96231b3b80d8

13 files changed:
docs/CommandGuide/llvm-profdata.rst
include/llvm/ProfileData/InstrProf.h
include/llvm/ProfileData/InstrProfWriter.h
include/llvm/ProfileData/SampleProf.h
lib/ProfileData/InstrProfWriter.cpp
test/tools/llvm-profdata/Inputs/weight-instr-bar.profdata [new file with mode: 0644]
test/tools/llvm-profdata/Inputs/weight-instr-foo.profdata [new file with mode: 0644]
test/tools/llvm-profdata/Inputs/weight-sample-bar.proftext [new file with mode: 0644]
test/tools/llvm-profdata/Inputs/weight-sample-foo.proftext [new file with mode: 0644]
test/tools/llvm-profdata/weight-instr.test [new file with mode: 0644]
test/tools/llvm-profdata/weight-sample.test [new file with mode: 0644]
tools/llvm-profdata/llvm-profdata.cpp
unittests/ProfileData/InstrProfTest.cpp

index 210826a..a4b18f3 100644 (file)
@@ -28,7 +28,7 @@ MERGE
 SYNOPSIS
 ^^^^^^^^
 
-:program:`llvm-profdata merge` [*options*] [*filenames...*]
+:program:`llvm-profdata merge` [*options*] [*filename[:weight]...*]
 
 DESCRIPTION
 ^^^^^^^^^^^
@@ -37,6 +37,10 @@ DESCRIPTION
 generated by PGO instrumentation and merges them together into a single
 indexed profile data file.
 
+The profile counts in each input file can be scaled (multiplied) by specifying
+``<filename>:<weight>``, where `<weight>` is a decimal integer >= 1.
+A default weight of 1 is assumed if only `<filename>` is given.
+
 OPTIONS
 ^^^^^^^
 
index 9564851..e1ed2e9 100644 (file)
@@ -218,7 +218,8 @@ struct InstrProfValueSiteRecord {
   }
 
   /// Merge data from another InstrProfValueSiteRecord
-  void mergeValueData(InstrProfValueSiteRecord &Input) {
+  /// Optionally scale merged counts by \p Weight.
+  void mergeValueData(InstrProfValueSiteRecord &Input, uint64_t Weight = 1) {
     this->sortByTargetValues();
     Input.sortByTargetValues();
     auto I = ValueData.begin();
@@ -228,7 +229,11 @@ struct InstrProfValueSiteRecord {
       while (I != IE && I->Value < J->Value)
         ++I;
       if (I != IE && I->Value == J->Value) {
-        I->Count = SaturatingAdd(I->Count, J->Count);
+        // TODO: Check for counter overflow and return error if it occurs.
+        uint64_t JCount = J->Count;
+        if (Weight > 1)
+          JCount = SaturatingMultiply(JCount, Weight);
+        I->Count = SaturatingAdd(I->Count, JCount);
         ++I;
         continue;
       }
@@ -274,7 +279,8 @@ struct InstrProfRecord {
                            ValueMapType *HashKeys);
 
   /// Merge the counts in \p Other into this one.
-  inline instrprof_error merge(InstrProfRecord &Other);
+  /// Optionally scale merged counts by \p Weight.
+  inline instrprof_error merge(InstrProfRecord &Other, uint64_t Weight = 1);
 
   /// Used by InstrProfWriter: update the value strings to commoned strings in
   /// the writer instance.
@@ -326,7 +332,9 @@ private:
   }
 
   // Merge Value Profile data from Src record to this record for ValueKind.
-  instrprof_error mergeValueProfData(uint32_t ValueKind, InstrProfRecord &Src) {
+  // Scale merged value counts by \p Weight.
+  instrprof_error mergeValueProfData(uint32_t ValueKind, InstrProfRecord &Src,
+                                     uint64_t Weight) {
     uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
     uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
     if (ThisNumValueSites != OtherNumValueSites)
@@ -336,7 +344,7 @@ private:
     std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
         Src.getValueSitesForKind(ValueKind);
     for (uint32_t I = 0; I < ThisNumValueSites; I++)
-      ThisSiteRecords[I].mergeValueData(OtherSiteRecords[I]);
+      ThisSiteRecords[I].mergeValueData(OtherSiteRecords[I], Weight);
     return instrprof_error::success;
   }
 };
@@ -422,7 +430,8 @@ void InstrProfRecord::updateStrings(InstrProfStringTable *StrTab) {
       VData.Value = (uint64_t)StrTab->insertString((const char *)VData.Value);
 }
 
-instrprof_error InstrProfRecord::merge(InstrProfRecord &Other) {
+instrprof_error InstrProfRecord::merge(InstrProfRecord &Other,
+                                       uint64_t Weight) {
   // If the number of counters doesn't match we either have bad data
   // or a hash collision.
   if (Counts.size() != Other.Counts.size())
@@ -432,13 +441,19 @@ instrprof_error InstrProfRecord::merge(InstrProfRecord &Other) {
 
   for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
     bool ResultOverflowed;
-    Counts[I] = SaturatingAdd(Counts[I], Other.Counts[I], ResultOverflowed);
+    uint64_t OtherCount = Other.Counts[I];
+    if (Weight > 1) {
+      OtherCount = SaturatingMultiply(OtherCount, Weight, ResultOverflowed);
+      if (ResultOverflowed)
+        Result = instrprof_error::counter_overflow;
+    }
+    Counts[I] = SaturatingAdd(Counts[I], OtherCount, ResultOverflowed);
     if (ResultOverflowed)
       Result = instrprof_error::counter_overflow;
   }
 
   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
-    instrprof_error MergeValueResult = mergeValueProfData(Kind, Other);
+    instrprof_error MergeValueResult = mergeValueProfData(Kind, Other, Weight);
     if (MergeValueResult != instrprof_error::success)
       Result = MergeValueResult;
   }
index d026e08..1958d5f 100644 (file)
@@ -39,8 +39,8 @@ public:
   void updateStringTableReferences(InstrProfRecord &I);
   /// Add function counts for the given function. If there are already counts
   /// for this function and the hash and number of counts match, each counter is
-  /// summed.
-  std::error_code addRecord(InstrProfRecord &&I);
+  /// summed. Optionally scale counts by \p Weight.
+  std::error_code addRecord(InstrProfRecord &&I, uint64_t Weight = 1);
   /// Write the profile to \c OS
   void write(raw_fd_ostream &OS);
   /// Write the profile in text format to \c OS
index a7b22c7..3337f4d 100644 (file)
@@ -173,19 +173,25 @@ public:
   SampleRecord() : NumSamples(0), CallTargets() {}
 
   /// Increment the number of samples for this record by \p S.
+  /// Optionally scale sample count \p S by \p Weight.
   ///
   /// Sample counts accumulate using saturating arithmetic, to avoid wrapping
   /// around unsigned integers.
-  void addSamples(uint64_t S) {
+  void addSamples(uint64_t S, uint64_t Weight = 1) {
+    if (Weight > 1)
+      S = SaturatingMultiply(S, Weight);
     NumSamples = SaturatingAdd(NumSamples, S);
   }
 
   /// Add called function \p F with samples \p S.
+  /// Optionally scale sample count \p S by \p Weight.
   ///
   /// Sample counts accumulate using saturating arithmetic, to avoid wrapping
   /// around unsigned integers.
-  void addCalledTarget(StringRef F, uint64_t S) {
+  void addCalledTarget(StringRef F, uint64_t S, uint64_t Weight = 1) {
     uint64_t &TargetSamples = CallTargets[F];
+    if (Weight > 1)
+      S = SaturatingMultiply(S, Weight);
     TargetSamples = SaturatingAdd(TargetSamples, S);
   }
 
@@ -196,10 +202,11 @@ public:
   const CallTargetMap &getCallTargets() const { return CallTargets; }
 
   /// Merge the samples in \p Other into this record.
-  void merge(const SampleRecord &Other) {
-    addSamples(Other.getSamples());
+  /// Optionally scale sample counts by \p Weight.
+  void merge(const SampleRecord &Other, uint64_t Weight = 1) {
+    addSamples(Other.getSamples(), Weight);
     for (const auto &I : Other.getCallTargets())
-      addCalledTarget(I.first(), I.second);
+      addCalledTarget(I.first(), I.second, Weight);
   }
 
   void print(raw_ostream &OS, unsigned Indent) const;
@@ -226,16 +233,26 @@ public:
   FunctionSamples() : TotalSamples(0), TotalHeadSamples(0) {}
   void print(raw_ostream &OS = dbgs(), unsigned Indent = 0) const;
   void dump() const;
-  void addTotalSamples(uint64_t Num) { TotalSamples += Num; }
-  void addHeadSamples(uint64_t Num) { TotalHeadSamples += Num; }
-  void addBodySamples(uint32_t LineOffset, uint32_t Discriminator,
-                      uint64_t Num) {
-    BodySamples[LineLocation(LineOffset, Discriminator)].addSamples(Num);
+  void addTotalSamples(uint64_t Num, uint64_t Weight = 1) {
+    if (Weight > 1)
+      Num = SaturatingMultiply(Num, Weight);
+    TotalSamples += Num;
+  }
+  void addHeadSamples(uint64_t Num, uint64_t Weight = 1) {
+    if (Weight > 1)
+      Num = SaturatingMultiply(Num, Weight);
+    TotalHeadSamples += Num;
+  }
+  void addBodySamples(uint32_t LineOffset, uint32_t Discriminator, uint64_t Num,
+                      uint64_t Weight = 1) {
+    BodySamples[LineLocation(LineOffset, Discriminator)].addSamples(Num,
+                                                                    Weight);
   }
   void addCalledTargetSamples(uint32_t LineOffset, uint32_t Discriminator,
-                              std::string FName, uint64_t Num) {
-    BodySamples[LineLocation(LineOffset, Discriminator)].addCalledTarget(FName,
-                                                                         Num);
+                              std::string FName, uint64_t Num,
+                              uint64_t Weight = 1) {
+    BodySamples[LineLocation(LineOffset, Discriminator)].addCalledTarget(
+        FName, Num, Weight);
   }
 
   /// Return the number of samples collected at the given location.
@@ -284,18 +301,19 @@ public:
   }
 
   /// Merge the samples in \p Other into this one.
-  void merge(const FunctionSamples &Other) {
-    addTotalSamples(Other.getTotalSamples());
-    addHeadSamples(Other.getHeadSamples());
+  /// Optionally scale samples by \p Weight.
+  void merge(const FunctionSamples &Other, uint64_t Weight = 1) {
+    addTotalSamples(Other.getTotalSamples(), Weight);
+    addHeadSamples(Other.getHeadSamples(), Weight);
     for (const auto &I : Other.getBodySamples()) {
       const LineLocation &Loc = I.first;
       const SampleRecord &Rec = I.second;
-      BodySamples[Loc].merge(Rec);
+      BodySamples[Loc].merge(Rec, Weight);
     }
     for (const auto &I : Other.getCallsiteSamples()) {
       const CallsiteLocation &Loc = I.first;
       const FunctionSamples &Rec = I.second;
-      functionSamplesAt(Loc).merge(Rec);
+      functionSamplesAt(Loc).merge(Rec, Weight);
     }
   }
 
index 78bec01..2261c92 100644 (file)
@@ -98,7 +98,8 @@ void InstrProfWriter::updateStringTableReferences(InstrProfRecord &I) {
   I.updateStrings(&StringTable);
 }
 
-std::error_code InstrProfWriter::addRecord(InstrProfRecord &&I) {
+std::error_code InstrProfWriter::addRecord(InstrProfRecord &&I,
+                                           uint64_t Weight) {
   updateStringTableReferences(I);
   auto &ProfileDataMap = FunctionData[I.Name];
 
@@ -113,9 +114,18 @@ std::error_code InstrProfWriter::addRecord(InstrProfRecord &&I) {
     // We've never seen a function with this name and hash, add it.
     Dest = std::move(I);
     Result = instrprof_error::success;
+    if (Weight > 1) {
+      for (auto &Count : Dest.Counts) {
+        bool Overflowed;
+        Count = SaturatingMultiply(Count, Weight, Overflowed);
+        if (Overflowed && Result == instrprof_error::success) {
+          Result = instrprof_error::counter_overflow;
+        }
+      }
+    }
   } else {
     // We're updating a function we've seen before.
-    Result = Dest.merge(I);
+    Result = Dest.merge(I, Weight);
   }
 
   // We keep track of the max function count as we go for simplicity.
diff --git a/test/tools/llvm-profdata/Inputs/weight-instr-bar.profdata b/test/tools/llvm-profdata/Inputs/weight-instr-bar.profdata
new file mode 100644 (file)
index 0000000..4ed0766
Binary files /dev/null and b/test/tools/llvm-profdata/Inputs/weight-instr-bar.profdata differ
diff --git a/test/tools/llvm-profdata/Inputs/weight-instr-foo.profdata b/test/tools/llvm-profdata/Inputs/weight-instr-foo.profdata
new file mode 100644 (file)
index 0000000..581ef39
Binary files /dev/null and b/test/tools/llvm-profdata/Inputs/weight-instr-foo.profdata differ
diff --git a/test/tools/llvm-profdata/Inputs/weight-sample-bar.proftext b/test/tools/llvm-profdata/Inputs/weight-sample-bar.proftext
new file mode 100644 (file)
index 0000000..a910f74
--- /dev/null
@@ -0,0 +1,8 @@
+bar:1772037:35370
+ 17: 35370
+ 18: 35370
+ 19: 7005
+ 20: 29407
+ 21: 12170
+ 23: 18150 bar:19829
+ 25: 36666
diff --git a/test/tools/llvm-profdata/Inputs/weight-sample-foo.proftext b/test/tools/llvm-profdata/Inputs/weight-sample-foo.proftext
new file mode 100644 (file)
index 0000000..155ec5d
--- /dev/null
@@ -0,0 +1,8 @@
+foo:1763288:35327
+ 7: 35327
+ 8: 35327
+ 9: 6930
+ 10: 29341
+ 11: 11906
+ 13: 18185 foo:19531
+ 15: 36458
diff --git a/test/tools/llvm-profdata/weight-instr.test b/test/tools/llvm-profdata/weight-instr.test
new file mode 100644 (file)
index 0000000..bc0b506
--- /dev/null
@@ -0,0 +1,55 @@
+Tests for weighted merge of instrumented profiles.
+
+1- Merge the foo and bar profiles with unity weight and verify the combined output
+RUN: llvm-profdata merge --instr %p/Inputs/weight-instr-bar.profdata:1 %p/Inputs/weight-instr-foo.profdata:1 -o %t
+RUN: llvm-profdata show --instr -all-functions %t | FileCheck %s --check-prefix=WEIGHT1
+WEIGHT1: Counters:
+WEIGHT1:   usage:
+WEIGHT1:     Hash: 0x0000000000000000
+WEIGHT1:     Counters: 1
+WEIGHT1:     Function count: 0
+WEIGHT1:   foo:
+WEIGHT1:     Hash: 0x000000000000028a
+WEIGHT1:     Counters: 3
+WEIGHT1:     Function count: 866988873
+WEIGHT1:   bar:
+WEIGHT1:     Hash: 0x000000000000028a
+WEIGHT1:     Counters: 3
+WEIGHT1:     Function count: 866988873
+WEIGHT1:   main:
+WEIGHT1:     Hash: 0x7d31c47ea98f8248
+WEIGHT1:     Counters: 60
+WEIGHT1:     Function count: 2
+WEIGHT1: Functions shown: 4
+WEIGHT1: Total functions: 4
+WEIGHT1: Maximum function count: 866988873
+WEIGHT1: Maximum internal block count: 267914296
+
+2- Merge the foo and bar profiles with weight 3x and 5x respectively and verify the combined output
+RUN: llvm-profdata merge --instr %p/Inputs/weight-instr-bar.profdata:3 %p/Inputs/weight-instr-foo.profdata:5 -o %t
+RUN: llvm-profdata show --instr -all-functions %t | FileCheck %s --check-prefix=WEIGHT2
+WEIGHT2: Counters:
+WEIGHT2:   usage:
+WEIGHT2:     Hash: 0x0000000000000000
+WEIGHT2:     Counters: 1
+WEIGHT2:     Function count: 0
+WEIGHT2:   foo:
+WEIGHT2:     Hash: 0x000000000000028a
+WEIGHT2:     Counters: 3
+WEIGHT2:     Function count: 4334944365
+WEIGHT2:   bar:
+WEIGHT2:     Hash: 0x000000000000028a
+WEIGHT2:     Counters: 3
+WEIGHT2:     Function count: 2600966619
+WEIGHT2:   main:
+WEIGHT2:     Hash: 0x7d31c47ea98f8248
+WEIGHT2:     Counters: 60
+WEIGHT2:     Function count: 8
+WEIGHT2: Functions shown: 4
+WEIGHT2: Total functions: 4
+WEIGHT2: Maximum function count: 4334944365
+WEIGHT2: Maximum internal block count: 1339571480
+
+3- Bad merge: foo and bar profiles with invalid weights
+RUN: not llvm-profdata merge --instr %p/Inputs/weight-instr-bar.profdata:3 %p/Inputs/weight-instr-foo.profdata:-5 -o %t.out 2>&1 | FileCheck %s --check-prefix=ERROR3
+ERROR3: error: Input weight must be a positive integer.
diff --git a/test/tools/llvm-profdata/weight-sample.test b/test/tools/llvm-profdata/weight-sample.test
new file mode 100644 (file)
index 0000000..a1fe1df
--- /dev/null
@@ -0,0 +1,43 @@
+Tests for weighted merge of sample profiles.
+
+1- Merge the foo and bar profiles with unity weight and verify the combined output
+RUN: llvm-profdata merge --sample --text %p/Inputs/weight-sample-bar.proftext:1 %p/Inputs/weight-sample-foo.proftext:1 -o - | FileCheck %s --check-prefix=WEIGHT1
+WEIGHT1: foo:1763288:35327
+WEIGHT1:  7: 35327
+WEIGHT1:  8: 35327
+WEIGHT1:  9: 6930
+WEIGHT1:  10: 29341
+WEIGHT1:  11: 11906
+WEIGHT1:  13: 18185 foo:19531
+WEIGHT1:  15: 36458
+WEIGHT1: bar:1772037:35370
+WEIGHT1:  17: 35370
+WEIGHT1:  18: 35370
+WEIGHT1:  19: 7005
+WEIGHT1:  20: 29407
+WEIGHT1:  21: 12170
+WEIGHT1:  23: 18150 bar:19829
+WEIGHT1:  25: 36666
+
+2- Merge the foo and bar profiles with weight 3x and 5x respectively and verify the combined output
+RUN: llvm-profdata merge --sample --text %p/Inputs/weight-sample-bar.proftext:3 %p/Inputs/weight-sample-foo.proftext:5 -o - | FileCheck %s --check-prefix=WEIGHT2
+WEIGHT2: foo:8816440:176635
+WEIGHT2:  7: 176635
+WEIGHT2:  8: 176635
+WEIGHT2:  9: 34650
+WEIGHT2:  10: 146705
+WEIGHT2:  11: 59530
+WEIGHT2:  13: 90925 foo:97655
+WEIGHT2:  15: 182290
+WEIGHT2: bar:5316111:106110
+WEIGHT2:  17: 106110
+WEIGHT2:  18: 106110
+WEIGHT2:  19: 21015
+WEIGHT2:  20: 88221
+WEIGHT2:  21: 36510
+WEIGHT2:  23: 54450 bar:59487
+WEIGHT2:  25: 109998
+
+3- Bad merge: foo and bar profiles with invalid weights
+RUN: not llvm-profdata merge --sample --text %p/Inputs/weight-sample-bar.proftext:3 %p/Inputs/weight-sample-foo.proftext:-5 -o %t.out 2>&1 | FileCheck %s --check-prefix=ERROR3
+ERROR3: error: Input weight must be a positive integer.
index 10b6855..56c80f5 100644 (file)
@@ -12,6 +12,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/ProfileData/InstrProfReader.h"
@@ -27,6 +28,7 @@
 #include "llvm/Support/PrettyStackTrace.h"
 #include "llvm/Support/Signals.h"
 #include "llvm/Support/raw_ostream.h"
+#include <tuple>
 
 using namespace llvm;
 
@@ -93,7 +95,17 @@ static void handleMergeWriterError(std::error_code &Error,
   }
 }
 
-static void mergeInstrProfile(const cl::list<std::string> &Inputs,
+struct WeightedFile {
+  StringRef Filename;
+  uint64_t Weight;
+
+  WeightedFile() {}
+
+  WeightedFile(StringRef F, uint64_t W) : Filename{F}, Weight{W} {}
+};
+typedef SmallVector<WeightedFile, 5> WeightedFileVector;
+
+static void mergeInstrProfile(const WeightedFileVector &Inputs,
                               StringRef OutputFilename,
                               ProfileFormat OutputFormat) {
   if (OutputFilename.compare("-") == 0)
@@ -109,21 +121,21 @@ static void mergeInstrProfile(const cl::list<std::string> &Inputs,
 
   InstrProfWriter Writer;
   SmallSet<std::error_code, 4> WriterErrorCodes;
-  for (const auto &Filename : Inputs) {
-    auto ReaderOrErr = InstrProfReader::create(Filename);
+  for (const auto &Input : Inputs) {
+    auto ReaderOrErr = InstrProfReader::create(Input.Filename);
     if (std::error_code ec = ReaderOrErr.getError())
-      exitWithErrorCode(ec, Filename);
+      exitWithErrorCode(ec, Input.Filename);
 
     auto Reader = std::move(ReaderOrErr.get());
     for (auto &I : *Reader) {
-      if (std::error_code EC = Writer.addRecord(std::move(I))) {
+      if (std::error_code EC = Writer.addRecord(std::move(I), Input.Weight)) {
         // Only show hint the first time an error occurs.
         bool firstTime = WriterErrorCodes.insert(EC).second;
-        handleMergeWriterError(EC, Filename, I.Name, firstTime);
+        handleMergeWriterError(EC, Input.Filename, I.Name, firstTime);
       }
     }
     if (Reader->hasError())
-      exitWithErrorCode(Reader->getError(), Filename);
+      exitWithErrorCode(Reader->getError(), Input.Filename);
   }
   if (OutputFormat == PF_Text)
     Writer.writeText(Output);
@@ -135,7 +147,7 @@ static sampleprof::SampleProfileFormat FormatMap[] = {
     sampleprof::SPF_None, sampleprof::SPF_Text, sampleprof::SPF_Binary,
     sampleprof::SPF_GCC};
 
-static void mergeSampleProfile(const cl::list<std::string> &Inputs,
+static void mergeSampleProfile(const WeightedFileVector &Inputs,
                                StringRef OutputFilename,
                                ProfileFormat OutputFormat) {
   using namespace sampleprof;
@@ -147,11 +159,11 @@ static void mergeSampleProfile(const cl::list<std::string> &Inputs,
   auto Writer = std::move(WriterOrErr.get());
   StringMap<FunctionSamples> ProfileMap;
   SmallVector<std::unique_ptr<sampleprof::SampleProfileReader>, 5> Readers;
-  for (const auto &Filename : Inputs) {
+  for (const auto &Input : Inputs) {
     auto ReaderOrErr =
-        SampleProfileReader::create(Filename, getGlobalContext());
+        SampleProfileReader::create(Input.Filename, getGlobalContext());
     if (std::error_code EC = ReaderOrErr.getError())
-      exitWithErrorCode(EC, Filename);
+      exitWithErrorCode(EC, Input.Filename);
 
     // We need to keep the readers around until after all the files are
     // read so that we do not lose the function names stored in each
@@ -160,7 +172,7 @@ static void mergeSampleProfile(const cl::list<std::string> &Inputs,
     Readers.push_back(std::move(ReaderOrErr.get()));
     const auto Reader = Readers.back().get();
     if (std::error_code EC = Reader->read())
-      exitWithErrorCode(EC, Filename);
+      exitWithErrorCode(EC, Input.Filename);
 
     StringMap<FunctionSamples> &Profiles = Reader->getProfiles();
     for (StringMap<FunctionSamples>::iterator I = Profiles.begin(),
@@ -168,15 +180,38 @@ static void mergeSampleProfile(const cl::list<std::string> &Inputs,
          I != E; ++I) {
       StringRef FName = I->first();
       FunctionSamples &Samples = I->second;
-      ProfileMap[FName].merge(Samples);
+      ProfileMap[FName].merge(Samples, Input.Weight);
     }
   }
   Writer->write(ProfileMap);
 }
 
+static void parseInputFiles(const cl::list<std::string> &Inputs,
+                            WeightedFileVector &WeightedInputs) {
+  WeightedInputs.reserve(Inputs.size());
+
+  for (StringRef Input : Inputs) {
+    StringRef FileName;
+    StringRef WeightStr;
+    std::tie(FileName, WeightStr) = Input.rsplit(':');
+    if (WeightStr.empty() || sys::fs::exists(Input)) {
+      // No weight specified or valid path containing delimiter.
+      WeightedInputs.push_back(WeightedFile(Input, 1));
+    } else {
+      // Input weight specified.
+      uint64_t Weight;
+      if (WeightStr.getAsInteger(10, Weight) || Weight < 1) {
+        // Invalid input weight.
+        exitWithError("Input weight must be a positive integer.");
+      }
+      WeightedInputs.push_back(WeightedFile(FileName, Weight));
+    }
+  }
+}
+
 static int merge_main(int argc, const char *argv[]) {
   cl::list<std::string> Inputs(cl::Positional, cl::Required, cl::OneOrMore,
-                               cl::desc("<filenames...>"));
+                               cl::desc("<filename[:weight]...>"));
 
   cl::opt<std::string> OutputFilename("output", cl::value_desc("output"),
                                       cl::init("-"), cl::Required,
@@ -198,10 +233,13 @@ static int merge_main(int argc, const char *argv[]) {
 
   cl::ParseCommandLineOptions(argc, argv, "LLVM profile data merger\n");
 
+  WeightedFileVector WeightedInputs;
+  parseInputFiles(Inputs, WeightedInputs);
+
   if (ProfileKind == instr)
-    mergeInstrProfile(Inputs, OutputFilename, OutputFormat);
+    mergeInstrProfile(WeightedInputs, OutputFilename, OutputFormat);
   else
-    mergeSampleProfile(Inputs, OutputFilename, OutputFormat);
+    mergeSampleProfile(WeightedInputs, OutputFilename, OutputFormat);
 
   return 0;
 }
index 635a543..946afda 100644 (file)
@@ -490,4 +490,24 @@ TEST_F(InstrProfTest, get_max_function_count) {
   ASSERT_EQ(1ULL << 63, Reader->getMaximumFunctionCount());
 }
 
+TEST_F(InstrProfTest, get_weighted_function_counts) {
+  InstrProfRecord Record1("foo", 0x1234, {1, 2});
+  InstrProfRecord Record2("foo", 0x1235, {3, 4});
+  Writer.addRecord(std::move(Record1), 3);
+  Writer.addRecord(std::move(Record2), 5);
+  auto Profile = Writer.writeBuffer();
+  readProfile(std::move(Profile));
+
+  std::vector<uint64_t> Counts;
+  ASSERT_TRUE(NoError(Reader->getFunctionCounts("foo", 0x1234, Counts)));
+  ASSERT_EQ(2U, Counts.size());
+  ASSERT_EQ(3U, Counts[0]);
+  ASSERT_EQ(6U, Counts[1]);
+
+  ASSERT_TRUE(NoError(Reader->getFunctionCounts("foo", 0x1235, Counts)));
+  ASSERT_EQ(2U, Counts.size());
+  ASSERT_EQ(15U, Counts[0]);
+  ASSERT_EQ(20U, Counts[1]);
+}
+
 } // end anonymous namespace