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/STLExtras.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/IR/ProfileSummary.h"
19 #include "llvm/ProfileData/InstrProf.h"
20 #include "llvm/ProfileData/ProfileCommon.h"
21 #include "llvm/Support/Endian.h"
22 #include "llvm/Support/EndianStream.h"
23 #include "llvm/Support/Error.h"
24 #include "llvm/Support/MemoryBuffer.h"
25 #include "llvm/Support/OnDiskHashTable.h"
26 #include "llvm/Support/raw_ostream.h"
27 #include <algorithm>
28 #include <cstdint>
29 #include <memory>
30 #include <string>
31 #include <tuple>
32 #include <utility>
33 #include <vector>
34 
35 using namespace llvm;
36 
37 // A struct to define how the data stream should be patched. For Indexed
38 // profiling, only uint64_t data type is needed.
39 struct PatchItem {
40   uint64_t Pos; // Where to patch.
41   uint64_t *D;  // Pointer to an array of source data.
42   int N;        // Number of elements in \c D array.
43 };
44 
45 namespace llvm {
46 
47 // A wrapper class to abstract writer stream with support of bytes
48 // back patching.
49 class ProfOStream {
50 public:
51   ProfOStream(raw_fd_ostream &FD) : IsFDOStream(true), OS(FD), LE(FD) {}
52   ProfOStream(raw_string_ostream &STR)
53       : IsFDOStream(false), OS(STR), LE(STR) {}
54 
55   uint64_t tell() { return OS.tell(); }
56   void write(uint64_t V) { LE.write<uint64_t>(V); }
57 
58   // \c patch can only be called when all data is written and flushed.
59   // For raw_string_ostream, the patch is done on the target string
60   // directly and it won't be reflected in the stream's internal buffer.
61   void patch(PatchItem *P, int NItems) {
62     using namespace support;
63 
64     if (IsFDOStream) {
65       raw_fd_ostream &FDOStream = static_cast<raw_fd_ostream &>(OS);
66       for (int K = 0; K < NItems; K++) {
67         FDOStream.seek(P[K].Pos);
68         for (int I = 0; I < P[K].N; I++)
69           write(P[K].D[I]);
70       }
71     } else {
72       raw_string_ostream &SOStream = static_cast<raw_string_ostream &>(OS);
73       std::string &Data = SOStream.str(); // with flush
74       for (int K = 0; K < NItems; K++) {
75         for (int I = 0; I < P[K].N; I++) {
76           uint64_t Bytes = endian::byte_swap<uint64_t, little>(P[K].D[I]);
77           Data.replace(P[K].Pos + I * sizeof(uint64_t), sizeof(uint64_t),
78                        (const char *)&Bytes, sizeof(uint64_t));
79         }
80       }
81     }
82   }
83 
84   // If \c OS is an instance of \c raw_fd_ostream, this field will be
85   // true. Otherwise, \c OS will be an raw_string_ostream.
86   bool IsFDOStream;
87   raw_ostream &OS;
88   support::endian::Writer<support::little> LE;
89 };
90 
91 class InstrProfRecordWriterTrait {
92 public:
93   using key_type = StringRef;
94   using key_type_ref = StringRef;
95 
96   using data_type = const InstrProfWriter::ProfilingData *const;
97   using data_type_ref = const InstrProfWriter::ProfilingData *const;
98 
99   using hash_value_type = uint64_t;
100   using offset_type = uint64_t;
101 
102   support::endianness ValueProfDataEndianness = support::little;
103   InstrProfSummaryBuilder *SummaryBuilder;
104 
105   InstrProfRecordWriterTrait() = default;
106 
107   static hash_value_type ComputeHash(key_type_ref K) {
108     return IndexedInstrProf::ComputeHash(K);
109   }
110 
111   static std::pair<offset_type, offset_type>
112   EmitKeyDataLength(raw_ostream &Out, key_type_ref K, data_type_ref V) {
113     using namespace support;
114 
115     endian::Writer<little> LE(Out);
116 
117     offset_type N = K.size();
118     LE.write<offset_type>(N);
119 
120     offset_type M = 0;
121     for (const auto &ProfileData : *V) {
122       const InstrProfRecord &ProfRecord = ProfileData.second;
123       M += sizeof(uint64_t); // The function hash
124       M += sizeof(uint64_t); // The size of the Counts vector
125       M += ProfRecord.Counts.size() * sizeof(uint64_t);
126 
127       // Value data
128       M += ValueProfData::getSize(ProfileData.second);
129     }
130     LE.write<offset_type>(M);
131 
132     return std::make_pair(N, M);
133   }
134 
135   void EmitKey(raw_ostream &Out, key_type_ref K, offset_type N) {
136     Out.write(K.data(), N);
137   }
138 
139   void EmitData(raw_ostream &Out, key_type_ref, data_type_ref V, offset_type) {
140     using namespace support;
141 
142     endian::Writer<little> LE(Out);
143     for (const auto &ProfileData : *V) {
144       const InstrProfRecord &ProfRecord = ProfileData.second;
145       SummaryBuilder->addRecord(ProfRecord);
146 
147       LE.write<uint64_t>(ProfileData.first); // Function hash
148       LE.write<uint64_t>(ProfRecord.Counts.size());
149       for (uint64_t I : ProfRecord.Counts)
150         LE.write<uint64_t>(I);
151 
152       // Write value data
153       std::unique_ptr<ValueProfData> VDataPtr =
154           ValueProfData::serializeFrom(ProfileData.second);
155       uint32_t S = VDataPtr->getSize();
156       VDataPtr->swapBytesFromHost(ValueProfDataEndianness);
157       Out.write((const char *)VDataPtr.get(), S);
158     }
159   }
160 };
161 
162 } // end namespace llvm
163 
164 InstrProfWriter::InstrProfWriter(bool Sparse)
165     : Sparse(Sparse), InfoObj(new InstrProfRecordWriterTrait()) {}
166 
167 InstrProfWriter::~InstrProfWriter() { delete InfoObj; }
168 
169 // Internal interface for testing purpose only.
170 void InstrProfWriter::setValueProfDataEndianness(
171     support::endianness Endianness) {
172   InfoObj->ValueProfDataEndianness = Endianness;
173 }
174 
175 void InstrProfWriter::setOutputSparse(bool Sparse) {
176   this->Sparse = Sparse;
177 }
178 
179 Error InstrProfWriter::addRecord(NamedInstrProfRecord &&I, uint64_t Weight) {
180   auto Name = I.Name;
181   auto Hash = I.Hash;
182   return addRecord(Name, Hash, std::move(I), Weight);
183 }
184 
185 Error InstrProfWriter::addRecord(StringRef Name, uint64_t Hash,
186                                  InstrProfRecord &&I, uint64_t Weight) {
187   auto &ProfileDataMap = FunctionData[Name];
188 
189   bool NewFunc;
190   ProfilingData::iterator Where;
191   std::tie(Where, NewFunc) =
192       ProfileDataMap.insert(std::make_pair(Hash, InstrProfRecord()));
193   InstrProfRecord &Dest = Where->second;
194 
195   if (NewFunc) {
196     // We've never seen a function with this name and hash, add it.
197     Dest = std::move(I);
198     if (Weight > 1)
199       Dest.scale(Weight);
200   } else {
201     // We're updating a function we've seen before.
202     Dest.merge(I, Weight);
203   }
204 
205   Dest.sortValueData();
206 
207   return Dest.takeError();
208 }
209 
210 Error InstrProfWriter::mergeRecordsFromWriter(InstrProfWriter &&IPW) {
211   for (auto &I : IPW.FunctionData)
212     for (auto &Func : I.getValue())
213       if (Error E = addRecord(I.getKey(), Func.first, std::move(Func.second)))
214         return E;
215   return Error::success();
216 }
217 
218 bool InstrProfWriter::shouldEncodeData(const ProfilingData &PD) {
219   if (!Sparse)
220     return true;
221   for (const auto &Func : PD) {
222     const InstrProfRecord &IPR = Func.second;
223     if (llvm::any_of(IPR.Counts, [](uint64_t Count) { return Count > 0; }))
224       return true;
225   }
226   return false;
227 }
228 
229 static void setSummary(IndexedInstrProf::Summary *TheSummary,
230                        ProfileSummary &PS) {
231   using namespace IndexedInstrProf;
232 
233   std::vector<ProfileSummaryEntry> &Res = PS.getDetailedSummary();
234   TheSummary->NumSummaryFields = Summary::NumKinds;
235   TheSummary->NumCutoffEntries = Res.size();
236   TheSummary->set(Summary::MaxFunctionCount, PS.getMaxFunctionCount());
237   TheSummary->set(Summary::MaxBlockCount, PS.getMaxCount());
238   TheSummary->set(Summary::MaxInternalBlockCount, PS.getMaxInternalCount());
239   TheSummary->set(Summary::TotalBlockCount, PS.getTotalCount());
240   TheSummary->set(Summary::TotalNumBlocks, PS.getNumCounts());
241   TheSummary->set(Summary::TotalNumFunctions, PS.getNumFunctions());
242   for (unsigned I = 0; I < Res.size(); I++)
243     TheSummary->setEntry(I, Res[I]);
244 }
245 
246 void InstrProfWriter::writeImpl(ProfOStream &OS) {
247   using namespace IndexedInstrProf;
248 
249   OnDiskChainedHashTableGenerator<InstrProfRecordWriterTrait> Generator;
250 
251   InstrProfSummaryBuilder ISB(ProfileSummaryBuilder::DefaultCutoffs);
252   InfoObj->SummaryBuilder = &ISB;
253 
254   // Populate the hash table generator.
255   for (const auto &I : FunctionData)
256     if (shouldEncodeData(I.getValue()))
257       Generator.insert(I.getKey(), &I.getValue());
258   // Write the header.
259   IndexedInstrProf::Header Header;
260   Header.Magic = IndexedInstrProf::Magic;
261   Header.Version = IndexedInstrProf::ProfVersion::CurrentVersion;
262   if (ProfileKind == PF_IRLevel)
263     Header.Version |= VARIANT_MASK_IR_PROF;
264   Header.Unused = 0;
265   Header.HashType = static_cast<uint64_t>(IndexedInstrProf::HashType);
266   Header.HashOffset = 0;
267   int N = sizeof(IndexedInstrProf::Header) / sizeof(uint64_t);
268 
269   // Only write out all the fields except 'HashOffset'. We need
270   // to remember the offset of that field to allow back patching
271   // later.
272   for (int I = 0; I < N - 1; I++)
273     OS.write(reinterpret_cast<uint64_t *>(&Header)[I]);
274 
275   // Save the location of Header.HashOffset field in \c OS.
276   uint64_t HashTableStartFieldOffset = OS.tell();
277   // Reserve the space for HashOffset field.
278   OS.write(0);
279 
280   // Reserve space to write profile summary data.
281   uint32_t NumEntries = ProfileSummaryBuilder::DefaultCutoffs.size();
282   uint32_t SummarySize = Summary::getSize(Summary::NumKinds, NumEntries);
283   // Remember the summary offset.
284   uint64_t SummaryOffset = OS.tell();
285   for (unsigned I = 0; I < SummarySize / sizeof(uint64_t); I++)
286     OS.write(0);
287 
288   // Write the hash table.
289   uint64_t HashTableStart = Generator.Emit(OS.OS, *InfoObj);
290 
291   // Allocate space for data to be serialized out.
292   std::unique_ptr<IndexedInstrProf::Summary> TheSummary =
293       IndexedInstrProf::allocSummary(SummarySize);
294   // Compute the Summary and copy the data to the data
295   // structure to be serialized out (to disk or buffer).
296   std::unique_ptr<ProfileSummary> PS = ISB.getSummary();
297   setSummary(TheSummary.get(), *PS);
298   InfoObj->SummaryBuilder = nullptr;
299 
300   // Now do the final patch:
301   PatchItem PatchItems[] = {
302       // Patch the Header.HashOffset field.
303       {HashTableStartFieldOffset, &HashTableStart, 1},
304       // Patch the summary data.
305       {SummaryOffset, reinterpret_cast<uint64_t *>(TheSummary.get()),
306        (int)(SummarySize / sizeof(uint64_t))}};
307   OS.patch(PatchItems, sizeof(PatchItems) / sizeof(*PatchItems));
308 }
309 
310 void InstrProfWriter::write(raw_fd_ostream &OS) {
311   // Write the hash table.
312   ProfOStream POS(OS);
313   writeImpl(POS);
314 }
315 
316 std::unique_ptr<MemoryBuffer> InstrProfWriter::writeBuffer() {
317   std::string Data;
318   raw_string_ostream OS(Data);
319   ProfOStream POS(OS);
320   // Write the hash table.
321   writeImpl(POS);
322   // Return this in an aligned memory buffer.
323   return MemoryBuffer::getMemBufferCopy(Data);
324 }
325 
326 static const char *ValueProfKindStr[] = {
327 #define VALUE_PROF_KIND(Enumerator, Value) #Enumerator,
328 #include "llvm/ProfileData/InstrProfData.inc"
329 };
330 
331 void InstrProfWriter::writeRecordInText(StringRef Name, uint64_t Hash,
332                                         const InstrProfRecord &Func,
333                                         InstrProfSymtab &Symtab,
334                                         raw_fd_ostream &OS) {
335   OS << Name << "\n";
336   OS << "# Func Hash:\n" << Hash << "\n";
337   OS << "# Num Counters:\n" << Func.Counts.size() << "\n";
338   OS << "# Counter Values:\n";
339   for (uint64_t Count : Func.Counts)
340     OS << Count << "\n";
341 
342   uint32_t NumValueKinds = Func.getNumValueKinds();
343   if (!NumValueKinds) {
344     OS << "\n";
345     return;
346   }
347 
348   OS << "# Num Value Kinds:\n" << Func.getNumValueKinds() << "\n";
349   for (uint32_t VK = 0; VK < IPVK_Last + 1; VK++) {
350     uint32_t NS = Func.getNumValueSites(VK);
351     if (!NS)
352       continue;
353     OS << "# ValueKind = " << ValueProfKindStr[VK] << ":\n" << VK << "\n";
354     OS << "# NumValueSites:\n" << NS << "\n";
355     for (uint32_t S = 0; S < NS; S++) {
356       uint32_t ND = Func.getNumValueDataForSite(VK, S);
357       OS << ND << "\n";
358       std::unique_ptr<InstrProfValueData[]> VD = Func.getValueForSite(VK, S);
359       for (uint32_t I = 0; I < ND; I++) {
360         if (VK == IPVK_IndirectCallTarget)
361           OS << Symtab.getFuncName(VD[I].Value) << ":" << VD[I].Count << "\n";
362         else
363           OS << VD[I].Value << ":" << VD[I].Count << "\n";
364       }
365     }
366   }
367 
368   OS << "\n";
369 }
370 
371 Error InstrProfWriter::writeText(raw_fd_ostream &OS) {
372   if (ProfileKind == PF_IRLevel)
373     OS << "# IR level Instrumentation Flag\n:ir\n";
374   InstrProfSymtab Symtab;
375   for (const auto &I : FunctionData)
376     if (shouldEncodeData(I.getValue()))
377       if (Error E = Symtab.addFuncName(I.getKey()))
378         return E;
379   Symtab.finalizeSymtab();
380 
381   for (const auto &I : FunctionData)
382     if (shouldEncodeData(I.getValue()))
383       for (const auto &Func : I.getValue())
384         writeRecordInText(I.getKey(), Func.first, Func.second, Symtab, OS);
385   return Error::success();
386 }
387