1//=-- InstrProfWriter.cpp - Instrumented profiling writer -------------------=//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file contains support for writing profiling data for clang's
11// instrumentation based PGO and coverage.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/ProfileData/InstrProfWriter.h"
16#include "llvm/ADT/StringExtras.h"
17#include "llvm/Support/EndianStream.h"
18#include "llvm/Support/OnDiskHashTable.h"
19#include <tuple>
20
21using namespace llvm;
22
23namespace {
24static support::endianness ValueProfDataEndianness = support::little;
25
26class InstrProfRecordTrait {
27public:
28  typedef StringRef key_type;
29  typedef StringRef key_type_ref;
30
31  typedef const InstrProfWriter::ProfilingData *const data_type;
32  typedef const InstrProfWriter::ProfilingData *const data_type_ref;
33
34  typedef uint64_t hash_value_type;
35  typedef uint64_t offset_type;
36
37  static hash_value_type ComputeHash(key_type_ref K) {
38    return IndexedInstrProf::ComputeHash(K);
39  }
40
41  static std::pair<offset_type, offset_type>
42  EmitKeyDataLength(raw_ostream &Out, key_type_ref K, data_type_ref V) {
43    using namespace llvm::support;
44    endian::Writer<little> LE(Out);
45
46    offset_type N = K.size();
47    LE.write<offset_type>(N);
48
49    offset_type M = 0;
50    for (const auto &ProfileData : *V) {
51      const InstrProfRecord &ProfRecord = ProfileData.second;
52      M += sizeof(uint64_t); // The function hash
53      M += sizeof(uint64_t); // The size of the Counts vector
54      M += ProfRecord.Counts.size() * sizeof(uint64_t);
55
56      // Value data
57      M += ValueProfData::getSize(ProfileData.second);
58    }
59    LE.write<offset_type>(M);
60
61    return std::make_pair(N, M);
62  }
63
64  static void EmitKey(raw_ostream &Out, key_type_ref K, offset_type N){
65    Out.write(K.data(), N);
66  }
67
68  static void EmitData(raw_ostream &Out, key_type_ref, data_type_ref V,
69                       offset_type) {
70    using namespace llvm::support;
71    endian::Writer<little> LE(Out);
72    for (const auto &ProfileData : *V) {
73      const InstrProfRecord &ProfRecord = ProfileData.second;
74
75      LE.write<uint64_t>(ProfileData.first); // Function hash
76      LE.write<uint64_t>(ProfRecord.Counts.size());
77      for (uint64_t I : ProfRecord.Counts)
78        LE.write<uint64_t>(I);
79
80      // Write value data
81      std::unique_ptr<ValueProfData> VDataPtr =
82          ValueProfData::serializeFrom(ProfileData.second);
83      uint32_t S = VDataPtr->getSize();
84      VDataPtr->swapBytesFromHost(ValueProfDataEndianness);
85      Out.write((const char *)VDataPtr.get(), S);
86    }
87  }
88};
89}
90
91// Internal interface for testing purpose only.
92void InstrProfWriter::setValueProfDataEndianness(
93    support::endianness Endianness) {
94  ValueProfDataEndianness = Endianness;
95}
96
97std::error_code InstrProfWriter::addRecord(InstrProfRecord &&I,
98                                           uint64_t Weight) {
99  auto &ProfileDataMap = FunctionData[I.Name];
100
101  bool NewFunc;
102  ProfilingData::iterator Where;
103  std::tie(Where, NewFunc) =
104      ProfileDataMap.insert(std::make_pair(I.Hash, InstrProfRecord()));
105  InstrProfRecord &Dest = Where->second;
106
107  instrprof_error Result = instrprof_error::success;
108  if (NewFunc) {
109    // We've never seen a function with this name and hash, add it.
110    Dest = std::move(I);
111    // Fix up the name to avoid dangling reference.
112    Dest.Name = FunctionData.find(Dest.Name)->getKey();
113    if (Weight > 1)
114      Result = Dest.scale(Weight);
115  } else {
116    // We're updating a function we've seen before.
117    Result = Dest.merge(I, Weight);
118  }
119
120  Dest.sortValueData();
121
122  // We keep track of the max function count as we go for simplicity.
123  // Update this statistic no matter the result of the merge.
124  if (Dest.Counts[0] > MaxFunctionCount)
125    MaxFunctionCount = Dest.Counts[0];
126
127  return Result;
128}
129
130std::pair<uint64_t, uint64_t> InstrProfWriter::writeImpl(raw_ostream &OS) {
131  OnDiskChainedHashTableGenerator<InstrProfRecordTrait> Generator;
132
133  // Populate the hash table generator.
134  for (const auto &I : FunctionData)
135    Generator.insert(I.getKey(), &I.getValue());
136
137  using namespace llvm::support;
138  endian::Writer<little> LE(OS);
139
140  // Write the header.
141  IndexedInstrProf::Header Header;
142  Header.Magic = IndexedInstrProf::Magic;
143  Header.Version = IndexedInstrProf::Version;
144  Header.MaxFunctionCount = MaxFunctionCount;
145  Header.HashType = static_cast<uint64_t>(IndexedInstrProf::HashType);
146  Header.HashOffset = 0;
147  int N = sizeof(IndexedInstrProf::Header) / sizeof(uint64_t);
148
149  // Only write out all the fields execpt 'HashOffset'. We need
150  // to remember the offset of that field to allow back patching
151  // later.
152  for (int I = 0; I < N - 1; I++)
153    LE.write<uint64_t>(reinterpret_cast<uint64_t *>(&Header)[I]);
154
155  // Save a space to write the hash table start location.
156  uint64_t HashTableStartLoc = OS.tell();
157  // Reserve the space for HashOffset field.
158  LE.write<uint64_t>(0);
159  // Write the hash table.
160  uint64_t HashTableStart = Generator.Emit(OS);
161
162  return std::make_pair(HashTableStartLoc, HashTableStart);
163}
164
165void InstrProfWriter::write(raw_fd_ostream &OS) {
166  // Write the hash table.
167  auto TableStart = writeImpl(OS);
168
169  // Go back and fill in the hash table start.
170  using namespace support;
171  OS.seek(TableStart.first);
172  // Now patch the HashOffset field previously reserved.
173  endian::Writer<little>(OS).write<uint64_t>(TableStart.second);
174}
175
176static const char *ValueProfKindStr[] = {
177#define VALUE_PROF_KIND(Enumerator, Value) #Enumerator,
178#include "llvm/ProfileData/InstrProfData.inc"
179};
180
181void InstrProfWriter::writeRecordInText(const InstrProfRecord &Func,
182                                        InstrProfSymtab &Symtab,
183                                        raw_fd_ostream &OS) {
184  OS << Func.Name << "\n";
185  OS << "# Func Hash:\n" << Func.Hash << "\n";
186  OS << "# Num Counters:\n" << Func.Counts.size() << "\n";
187  OS << "# Counter Values:\n";
188  for (uint64_t Count : Func.Counts)
189    OS << Count << "\n";
190
191  uint32_t NumValueKinds = Func.getNumValueKinds();
192  if (!NumValueKinds) {
193    OS << "\n";
194    return;
195  }
196
197  OS << "# Num Value Kinds:\n" << Func.getNumValueKinds() << "\n";
198  for (uint32_t VK = 0; VK < IPVK_Last + 1; VK++) {
199    uint32_t NS = Func.getNumValueSites(VK);
200    if (!NS)
201      continue;
202    OS << "# ValueKind = " << ValueProfKindStr[VK] << ":\n" << VK << "\n";
203    OS << "# NumValueSites:\n" << NS << "\n";
204    for (uint32_t S = 0; S < NS; S++) {
205      uint32_t ND = Func.getNumValueDataForSite(VK, S);
206      OS << ND << "\n";
207      std::unique_ptr<InstrProfValueData[]> VD = Func.getValueForSite(VK, S);
208      for (uint32_t I = 0; I < ND; I++) {
209        if (VK == IPVK_IndirectCallTarget)
210          OS << Symtab.getFuncName(VD[I].Value) << ":" << VD[I].Count << "\n";
211        else
212          OS << VD[I].Value << ":" << VD[I].Count << "\n";
213      }
214    }
215  }
216
217  OS << "\n";
218}
219
220void InstrProfWriter::writeText(raw_fd_ostream &OS) {
221  InstrProfSymtab Symtab;
222  for (const auto &I : FunctionData)
223    Symtab.addFuncName(I.getKey());
224  Symtab.finalizeSymtab();
225
226  for (const auto &I : FunctionData)
227    for (const auto &Func : I.getValue())
228      writeRecordInText(Func.second, Symtab, OS);
229}
230
231std::unique_ptr<MemoryBuffer> InstrProfWriter::writeBuffer() {
232  std::string Data;
233  llvm::raw_string_ostream OS(Data);
234  // Write the hash table.
235  auto TableStart = writeImpl(OS);
236  OS.flush();
237
238  // Go back and fill in the hash table start.
239  using namespace support;
240  uint64_t Bytes = endian::byte_swap<uint64_t, little>(TableStart.second);
241  Data.replace(TableStart.first, sizeof(uint64_t), (const char *)&Bytes,
242               sizeof(uint64_t));
243
244  // Return this in an aligned memory buffer.
245  return MemoryBuffer::getMemBufferCopy(Data);
246}
247