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 
21 using namespace llvm;
22 
23 // A struct to define how the data stream should be patched. For Indexed
24 // profiling, only uint64_t data type is needed.
25 struct PatchItem {
26   uint64_t Pos; // Where to patch.
27   uint64_t *D;  // Pointer to an array of source data.
28   int N;        // Number of elements in \c D array.
29 };
30 
31 namespace llvm {
32 // A wrapper class to abstract writer stream with support of bytes
33 // back patching.
34 class ProfOStream {
35 
36 public:
37   ProfOStream(llvm::raw_fd_ostream &FD) : IsFDOStream(true), OS(FD), LE(FD) {}
38   ProfOStream(llvm::raw_string_ostream &STR)
39       : IsFDOStream(false), OS(STR), LE(STR) {}
40 
41   uint64_t tell() { return OS.tell(); }
42   void write(uint64_t V) { LE.write<uint64_t>(V); }
43   // \c patch can only be called when all data is written and flushed.
44   // For raw_string_ostream, the patch is done on the target string
45   // directly and it won't be reflected in the stream's internal buffer.
46   void patch(PatchItem *P, int NItems) {
47     using namespace support;
48     if (IsFDOStream) {
49       llvm::raw_fd_ostream &FDOStream = static_cast<llvm::raw_fd_ostream &>(OS);
50       for (int K = 0; K < NItems; K++) {
51         FDOStream.seek(P[K].Pos);
52         for (int I = 0; I < P[K].N; I++)
53           write(P[K].D[I]);
54       }
55     } else {
56       llvm::raw_string_ostream &SOStream =
57           static_cast<llvm::raw_string_ostream &>(OS);
58       std::string &Data = SOStream.str(); // with flush
59       for (int K = 0; K < NItems; K++) {
60         for (int I = 0; I < P[K].N; I++) {
61           uint64_t Bytes = endian::byte_swap<uint64_t, little>(P[K].D[I]);
62           Data.replace(P[K].Pos + I * sizeof(uint64_t), sizeof(uint64_t),
63                        (const char *)&Bytes, sizeof(uint64_t));
64         }
65       }
66     }
67   }
68   // If \c OS is an instance of \c raw_fd_ostream, this field will be
69   // true. Otherwise, \c OS will be an raw_string_ostream.
70   bool IsFDOStream;
71   raw_ostream &OS;
72   support::endian::Writer<support::little> LE;
73 };
74 }
75 
76 namespace {
77 static support::endianness ValueProfDataEndianness = support::little;
78 
79 class InstrProfRecordTrait {
80 public:
81   typedef StringRef key_type;
82   typedef StringRef key_type_ref;
83 
84   typedef const InstrProfWriter::ProfilingData *const data_type;
85   typedef const InstrProfWriter::ProfilingData *const data_type_ref;
86 
87   typedef uint64_t hash_value_type;
88   typedef uint64_t offset_type;
89 
90   static hash_value_type ComputeHash(key_type_ref K) {
91     return IndexedInstrProf::ComputeHash(K);
92   }
93 
94   static std::pair<offset_type, offset_type>
95   EmitKeyDataLength(raw_ostream &Out, key_type_ref K, data_type_ref V) {
96     using namespace llvm::support;
97     endian::Writer<little> LE(Out);
98 
99     offset_type N = K.size();
100     LE.write<offset_type>(N);
101 
102     offset_type M = 0;
103     for (const auto &ProfileData : *V) {
104       const InstrProfRecord &ProfRecord = ProfileData.second;
105       M += sizeof(uint64_t); // The function hash
106       M += sizeof(uint64_t); // The size of the Counts vector
107       M += ProfRecord.Counts.size() * sizeof(uint64_t);
108 
109       // Value data
110       M += ValueProfData::getSize(ProfileData.second);
111     }
112     LE.write<offset_type>(M);
113 
114     return std::make_pair(N, M);
115   }
116 
117   static void EmitKey(raw_ostream &Out, key_type_ref K, offset_type N){
118     Out.write(K.data(), N);
119   }
120 
121   static void EmitData(raw_ostream &Out, key_type_ref, data_type_ref V,
122                        offset_type) {
123     using namespace llvm::support;
124     endian::Writer<little> LE(Out);
125     for (const auto &ProfileData : *V) {
126       const InstrProfRecord &ProfRecord = ProfileData.second;
127 
128       LE.write<uint64_t>(ProfileData.first); // Function hash
129       LE.write<uint64_t>(ProfRecord.Counts.size());
130       for (uint64_t I : ProfRecord.Counts)
131         LE.write<uint64_t>(I);
132 
133       // Write value data
134       std::unique_ptr<ValueProfData> VDataPtr =
135           ValueProfData::serializeFrom(ProfileData.second);
136       uint32_t S = VDataPtr->getSize();
137       VDataPtr->swapBytesFromHost(ValueProfDataEndianness);
138       Out.write((const char *)VDataPtr.get(), S);
139     }
140   }
141 };
142 }
143 
144 // Internal interface for testing purpose only.
145 void InstrProfWriter::setValueProfDataEndianness(
146     support::endianness Endianness) {
147   ValueProfDataEndianness = Endianness;
148 }
149 
150 std::error_code InstrProfWriter::addRecord(InstrProfRecord &&I,
151                                            uint64_t Weight) {
152   auto &ProfileDataMap = FunctionData[I.Name];
153 
154   bool NewFunc;
155   ProfilingData::iterator Where;
156   std::tie(Where, NewFunc) =
157       ProfileDataMap.insert(std::make_pair(I.Hash, InstrProfRecord()));
158   InstrProfRecord &Dest = Where->second;
159 
160   instrprof_error Result = instrprof_error::success;
161   if (NewFunc) {
162     // We've never seen a function with this name and hash, add it.
163     Dest = std::move(I);
164     // Fix up the name to avoid dangling reference.
165     Dest.Name = FunctionData.find(Dest.Name)->getKey();
166     if (Weight > 1)
167       Result = Dest.scale(Weight);
168   } else {
169     // We're updating a function we've seen before.
170     Result = Dest.merge(I, Weight);
171   }
172 
173   Dest.sortValueData();
174 
175   // We keep track of the max function count as we go for simplicity.
176   // Update this statistic no matter the result of the merge.
177   if (Dest.Counts[0] > MaxFunctionCount)
178     MaxFunctionCount = Dest.Counts[0];
179 
180   return Result;
181 }
182 
183 void InstrProfWriter::writeImpl(ProfOStream &OS) {
184   OnDiskChainedHashTableGenerator<InstrProfRecordTrait> Generator;
185   // Populate the hash table generator.
186   for (const auto &I : FunctionData)
187     Generator.insert(I.getKey(), &I.getValue());
188   // Write the header.
189   IndexedInstrProf::Header Header;
190   Header.Magic = IndexedInstrProf::Magic;
191   Header.Version = IndexedInstrProf::ProfVersion::CurrentVersion;
192   Header.MaxFunctionCount = MaxFunctionCount;
193   Header.HashType = static_cast<uint64_t>(IndexedInstrProf::HashType);
194   Header.HashOffset = 0;
195   int N = sizeof(IndexedInstrProf::Header) / sizeof(uint64_t);
196 
197   // Only write out all the fields execpt 'HashOffset'. We need
198   // to remember the offset of that field to allow back patching
199   // later.
200   for (int I = 0; I < N - 1; I++)
201     OS.write(reinterpret_cast<uint64_t *>(&Header)[I]);
202 
203   // Save a space to write the hash table start location.
204   uint64_t HashTableStartLoc = OS.tell();
205   // Reserve the space for HashOffset field.
206   OS.write(0);
207   // Write the hash table.
208   uint64_t HashTableStart = Generator.Emit(OS.OS);
209 
210   // Now do the final patch:
211   PatchItem PatchItems[1] = {{HashTableStartLoc, &HashTableStart, 1}};
212   OS.patch(PatchItems, sizeof(PatchItems) / sizeof(*PatchItems));
213 }
214 
215 void InstrProfWriter::write(raw_fd_ostream &OS) {
216   // Write the hash table.
217   ProfOStream POS(OS);
218   writeImpl(POS);
219 }
220 
221 std::unique_ptr<MemoryBuffer> InstrProfWriter::writeBuffer() {
222   std::string Data;
223   llvm::raw_string_ostream OS(Data);
224   ProfOStream POS(OS);
225   // Write the hash table.
226   writeImpl(POS);
227   // Return this in an aligned memory buffer.
228   return MemoryBuffer::getMemBufferCopy(Data);
229 }
230 
231 static const char *ValueProfKindStr[] = {
232 #define VALUE_PROF_KIND(Enumerator, Value) #Enumerator,
233 #include "llvm/ProfileData/InstrProfData.inc"
234 };
235 
236 void InstrProfWriter::writeRecordInText(const InstrProfRecord &Func,
237                                         InstrProfSymtab &Symtab,
238                                         raw_fd_ostream &OS) {
239   OS << Func.Name << "\n";
240   OS << "# Func Hash:\n" << Func.Hash << "\n";
241   OS << "# Num Counters:\n" << Func.Counts.size() << "\n";
242   OS << "# Counter Values:\n";
243   for (uint64_t Count : Func.Counts)
244     OS << Count << "\n";
245 
246   uint32_t NumValueKinds = Func.getNumValueKinds();
247   if (!NumValueKinds) {
248     OS << "\n";
249     return;
250   }
251 
252   OS << "# Num Value Kinds:\n" << Func.getNumValueKinds() << "\n";
253   for (uint32_t VK = 0; VK < IPVK_Last + 1; VK++) {
254     uint32_t NS = Func.getNumValueSites(VK);
255     if (!NS)
256       continue;
257     OS << "# ValueKind = " << ValueProfKindStr[VK] << ":\n" << VK << "\n";
258     OS << "# NumValueSites:\n" << NS << "\n";
259     for (uint32_t S = 0; S < NS; S++) {
260       uint32_t ND = Func.getNumValueDataForSite(VK, S);
261       OS << ND << "\n";
262       std::unique_ptr<InstrProfValueData[]> VD = Func.getValueForSite(VK, S);
263       for (uint32_t I = 0; I < ND; I++) {
264         if (VK == IPVK_IndirectCallTarget)
265           OS << Symtab.getFuncName(VD[I].Value) << ":" << VD[I].Count << "\n";
266         else
267           OS << VD[I].Value << ":" << VD[I].Count << "\n";
268       }
269     }
270   }
271 
272   OS << "\n";
273 }
274 
275 void InstrProfWriter::writeText(raw_fd_ostream &OS) {
276   InstrProfSymtab Symtab;
277   for (const auto &I : FunctionData)
278     Symtab.addFuncName(I.getKey());
279   Symtab.finalizeSymtab();
280 
281   for (const auto &I : FunctionData)
282     for (const auto &Func : I.getValue())
283       writeRecordInText(Func.second, Symtab, OS);
284 }
285