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(InstrProfRecord &&I, uint64_t Weight) {
180   auto &ProfileDataMap = FunctionData[I.Name];
181 
182   bool NewFunc;
183   ProfilingData::iterator Where;
184   std::tie(Where, NewFunc) =
185       ProfileDataMap.insert(std::make_pair(I.Hash, InstrProfRecord()));
186   InstrProfRecord &Dest = Where->second;
187 
188   if (NewFunc) {
189     // We've never seen a function with this name and hash, add it.
190     Dest = std::move(I);
191     // Fix up the name to avoid dangling reference.
192     Dest.Name = FunctionData.find(Dest.Name)->getKey();
193     if (Weight > 1)
194       Dest.scale(Weight);
195   } else {
196     // We're updating a function we've seen before.
197     Dest.merge(I, Weight);
198   }
199 
200   Dest.sortValueData();
201 
202   return Dest.takeError();
203 }
204 
205 Error InstrProfWriter::mergeRecordsFromWriter(InstrProfWriter &&IPW) {
206   for (auto &I : IPW.FunctionData)
207     for (auto &Func : I.getValue())
208       if (Error E = addRecord(std::move(Func.second), 1))
209         return E;
210   return Error::success();
211 }
212 
213 bool InstrProfWriter::shouldEncodeData(const ProfilingData &PD) {
214   if (!Sparse)
215     return true;
216   for (const auto &Func : PD) {
217     const InstrProfRecord &IPR = Func.second;
218     if (llvm::any_of(IPR.Counts, [](uint64_t Count) { return Count > 0; }))
219       return true;
220   }
221   return false;
222 }
223 
224 static void setSummary(IndexedInstrProf::Summary *TheSummary,
225                        ProfileSummary &PS) {
226   using namespace IndexedInstrProf;
227 
228   std::vector<ProfileSummaryEntry> &Res = PS.getDetailedSummary();
229   TheSummary->NumSummaryFields = Summary::NumKinds;
230   TheSummary->NumCutoffEntries = Res.size();
231   TheSummary->set(Summary::MaxFunctionCount, PS.getMaxFunctionCount());
232   TheSummary->set(Summary::MaxBlockCount, PS.getMaxCount());
233   TheSummary->set(Summary::MaxInternalBlockCount, PS.getMaxInternalCount());
234   TheSummary->set(Summary::TotalBlockCount, PS.getTotalCount());
235   TheSummary->set(Summary::TotalNumBlocks, PS.getNumCounts());
236   TheSummary->set(Summary::TotalNumFunctions, PS.getNumFunctions());
237   for (unsigned I = 0; I < Res.size(); I++)
238     TheSummary->setEntry(I, Res[I]);
239 }
240 
241 void InstrProfWriter::writeImpl(ProfOStream &OS) {
242   using namespace IndexedInstrProf;
243 
244   OnDiskChainedHashTableGenerator<InstrProfRecordWriterTrait> Generator;
245 
246   InstrProfSummaryBuilder ISB(ProfileSummaryBuilder::DefaultCutoffs);
247   InfoObj->SummaryBuilder = &ISB;
248 
249   // Populate the hash table generator.
250   for (const auto &I : FunctionData)
251     if (shouldEncodeData(I.getValue()))
252       Generator.insert(I.getKey(), &I.getValue());
253   // Write the header.
254   IndexedInstrProf::Header Header;
255   Header.Magic = IndexedInstrProf::Magic;
256   Header.Version = IndexedInstrProf::ProfVersion::CurrentVersion;
257   if (ProfileKind == PF_IRLevel)
258     Header.Version |= VARIANT_MASK_IR_PROF;
259   Header.Unused = 0;
260   Header.HashType = static_cast<uint64_t>(IndexedInstrProf::HashType);
261   Header.HashOffset = 0;
262   int N = sizeof(IndexedInstrProf::Header) / sizeof(uint64_t);
263 
264   // Only write out all the fields except 'HashOffset'. We need
265   // to remember the offset of that field to allow back patching
266   // later.
267   for (int I = 0; I < N - 1; I++)
268     OS.write(reinterpret_cast<uint64_t *>(&Header)[I]);
269 
270   // Save the location of Header.HashOffset field in \c OS.
271   uint64_t HashTableStartFieldOffset = OS.tell();
272   // Reserve the space for HashOffset field.
273   OS.write(0);
274 
275   // Reserve space to write profile summary data.
276   uint32_t NumEntries = ProfileSummaryBuilder::DefaultCutoffs.size();
277   uint32_t SummarySize = Summary::getSize(Summary::NumKinds, NumEntries);
278   // Remember the summary offset.
279   uint64_t SummaryOffset = OS.tell();
280   for (unsigned I = 0; I < SummarySize / sizeof(uint64_t); I++)
281     OS.write(0);
282 
283   // Write the hash table.
284   uint64_t HashTableStart = Generator.Emit(OS.OS, *InfoObj);
285 
286   // Allocate space for data to be serialized out.
287   std::unique_ptr<IndexedInstrProf::Summary> TheSummary =
288       IndexedInstrProf::allocSummary(SummarySize);
289   // Compute the Summary and copy the data to the data
290   // structure to be serialized out (to disk or buffer).
291   std::unique_ptr<ProfileSummary> PS = ISB.getSummary();
292   setSummary(TheSummary.get(), *PS);
293   InfoObj->SummaryBuilder = nullptr;
294 
295   // Now do the final patch:
296   PatchItem PatchItems[] = {
297       // Patch the Header.HashOffset field.
298       {HashTableStartFieldOffset, &HashTableStart, 1},
299       // Patch the summary data.
300       {SummaryOffset, reinterpret_cast<uint64_t *>(TheSummary.get()),
301        (int)(SummarySize / sizeof(uint64_t))}};
302   OS.patch(PatchItems, sizeof(PatchItems) / sizeof(*PatchItems));
303 }
304 
305 void InstrProfWriter::write(raw_fd_ostream &OS) {
306   // Write the hash table.
307   ProfOStream POS(OS);
308   writeImpl(POS);
309 }
310 
311 std::unique_ptr<MemoryBuffer> InstrProfWriter::writeBuffer() {
312   std::string Data;
313   raw_string_ostream OS(Data);
314   ProfOStream POS(OS);
315   // Write the hash table.
316   writeImpl(POS);
317   // Return this in an aligned memory buffer.
318   return MemoryBuffer::getMemBufferCopy(Data);
319 }
320 
321 static const char *ValueProfKindStr[] = {
322 #define VALUE_PROF_KIND(Enumerator, Value) #Enumerator,
323 #include "llvm/ProfileData/InstrProfData.inc"
324 };
325 
326 void InstrProfWriter::writeRecordInText(const InstrProfRecord &Func,
327                                         InstrProfSymtab &Symtab,
328                                         raw_fd_ostream &OS) {
329   OS << Func.Name << "\n";
330   OS << "# Func Hash:\n" << Func.Hash << "\n";
331   OS << "# Num Counters:\n" << Func.Counts.size() << "\n";
332   OS << "# Counter Values:\n";
333   for (uint64_t Count : Func.Counts)
334     OS << Count << "\n";
335 
336   uint32_t NumValueKinds = Func.getNumValueKinds();
337   if (!NumValueKinds) {
338     OS << "\n";
339     return;
340   }
341 
342   OS << "# Num Value Kinds:\n" << Func.getNumValueKinds() << "\n";
343   for (uint32_t VK = 0; VK < IPVK_Last + 1; VK++) {
344     uint32_t NS = Func.getNumValueSites(VK);
345     if (!NS)
346       continue;
347     OS << "# ValueKind = " << ValueProfKindStr[VK] << ":\n" << VK << "\n";
348     OS << "# NumValueSites:\n" << NS << "\n";
349     for (uint32_t S = 0; S < NS; S++) {
350       uint32_t ND = Func.getNumValueDataForSite(VK, S);
351       OS << ND << "\n";
352       std::unique_ptr<InstrProfValueData[]> VD = Func.getValueForSite(VK, S);
353       for (uint32_t I = 0; I < ND; I++) {
354         if (VK == IPVK_IndirectCallTarget)
355           OS << Symtab.getFuncName(VD[I].Value) << ":" << VD[I].Count << "\n";
356         else
357           OS << VD[I].Value << ":" << VD[I].Count << "\n";
358       }
359     }
360   }
361 
362   OS << "\n";
363 }
364 
365 Error InstrProfWriter::writeText(raw_fd_ostream &OS) {
366   if (ProfileKind == PF_IRLevel)
367     OS << "# IR level Instrumentation Flag\n:ir\n";
368   InstrProfSymtab Symtab;
369   for (const auto &I : FunctionData)
370     if (shouldEncodeData(I.getValue()))
371       if (Error E = Symtab.addFuncName(I.getKey()))
372         return E;
373   Symtab.finalizeSymtab();
374 
375   for (const auto &I : FunctionData)
376     if (shouldEncodeData(I.getValue()))
377       for (const auto &Func : I.getValue())
378         writeRecordInText(Func.second, Symtab, OS);
379   return Error::success();
380 }
381