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 class InstrProfRecordWriterTrait {
76 public:
77   typedef StringRef key_type;
78   typedef StringRef key_type_ref;
79 
80   typedef const InstrProfWriter::ProfilingData *const data_type;
81   typedef const InstrProfWriter::ProfilingData *const data_type_ref;
82 
83   typedef uint64_t hash_value_type;
84   typedef uint64_t offset_type;
85 
86   support::endianness ValueProfDataEndianness;
87   ProfileSummary *TheProfileSummary;
88 
89   InstrProfRecordWriterTrait() : ValueProfDataEndianness(support::little) {}
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   void EmitKey(raw_ostream &Out, key_type_ref K, offset_type N) {
118     Out.write(K.data(), N);
119   }
120 
121   void EmitData(raw_ostream &Out, key_type_ref, data_type_ref V, offset_type) {
122     using namespace llvm::support;
123     endian::Writer<little> LE(Out);
124     for (const auto &ProfileData : *V) {
125       const InstrProfRecord &ProfRecord = ProfileData.second;
126       TheProfileSummary->addRecord(ProfRecord);
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 InstrProfWriter::InstrProfWriter(bool Sparse)
145     : Sparse(Sparse), FunctionData(), ProfileKind(PF_Unknown),
146       InfoObj(new InstrProfRecordWriterTrait()) {}
147 
148 InstrProfWriter::~InstrProfWriter() { delete InfoObj; }
149 
150 // Internal interface for testing purpose only.
151 void InstrProfWriter::setValueProfDataEndianness(
152     support::endianness Endianness) {
153   InfoObj->ValueProfDataEndianness = Endianness;
154 }
155 void InstrProfWriter::setOutputSparse(bool Sparse) {
156   this->Sparse = Sparse;
157 }
158 
159 std::error_code InstrProfWriter::addRecord(InstrProfRecord &&I,
160                                            uint64_t Weight) {
161   auto &ProfileDataMap = FunctionData[I.Name];
162 
163   bool NewFunc;
164   ProfilingData::iterator Where;
165   std::tie(Where, NewFunc) =
166       ProfileDataMap.insert(std::make_pair(I.Hash, InstrProfRecord()));
167   InstrProfRecord &Dest = Where->second;
168 
169   instrprof_error Result = instrprof_error::success;
170   if (NewFunc) {
171     // We've never seen a function with this name and hash, add it.
172     Dest = std::move(I);
173     // Fix up the name to avoid dangling reference.
174     Dest.Name = FunctionData.find(Dest.Name)->getKey();
175     if (Weight > 1)
176       Result = Dest.scale(Weight);
177   } else {
178     // We're updating a function we've seen before.
179     Result = Dest.merge(I, Weight);
180   }
181 
182   Dest.sortValueData();
183 
184   return Result;
185 }
186 
187 bool InstrProfWriter::shouldEncodeData(const ProfilingData &PD) {
188   if (!Sparse)
189     return true;
190   for (const auto &Func : PD) {
191     const InstrProfRecord &IPR = Func.second;
192     if (std::any_of(IPR.Counts.begin(), IPR.Counts.end(),
193                     [](uint64_t Count) { return Count > 0; }))
194       return true;
195   }
196   return false;
197 }
198 
199 static void setSummary(IndexedInstrProf::Summary *TheSummary,
200                        ProfileSummary &PS) {
201   using namespace IndexedInstrProf;
202   std::vector<ProfileSummaryEntry> &Res = PS.getDetailedSummary();
203   TheSummary->NumSummaryFields = Summary::NumKinds;
204   TheSummary->NumCutoffEntries = Res.size();
205   TheSummary->set(Summary::MaxFunctionCount, PS.getMaxFunctionCount());
206   TheSummary->set(Summary::MaxBlockCount, PS.getMaxBlockCount());
207   TheSummary->set(Summary::MaxInternalBlockCount,
208                   PS.getMaxInternalBlockCount());
209   TheSummary->set(Summary::TotalBlockCount, PS.getTotalCount());
210   TheSummary->set(Summary::TotalNumBlocks, PS.getNumBlocks());
211   TheSummary->set(Summary::TotalNumFunctions, PS.getNumFunctions());
212   for (unsigned I = 0; I < Res.size(); I++)
213     TheSummary->setEntry(I, Res[I]);
214 }
215 
216 void InstrProfWriter::writeImpl(ProfOStream &OS) {
217   OnDiskChainedHashTableGenerator<InstrProfRecordWriterTrait> Generator;
218 
219   using namespace IndexedInstrProf;
220   std::vector<uint32_t> Cutoffs(&SummaryCutoffs[0],
221                                 &SummaryCutoffs[NumSummaryCutoffs]);
222   ProfileSummary PS(Cutoffs);
223   InfoObj->TheProfileSummary = &PS;
224 
225   // Populate the hash table generator.
226   for (const auto &I : FunctionData)
227     if (shouldEncodeData(I.getValue()))
228       Generator.insert(I.getKey(), &I.getValue());
229   // Write the header.
230   IndexedInstrProf::Header Header;
231   Header.Magic = IndexedInstrProf::Magic;
232   Header.Version = IndexedInstrProf::ProfVersion::CurrentVersion;
233   if (ProfileKind == PF_IRLevel)
234     Header.Version |= VARIANT_MASK_IR_PROF;
235   Header.Unused = 0;
236   Header.HashType = static_cast<uint64_t>(IndexedInstrProf::HashType);
237   Header.HashOffset = 0;
238   int N = sizeof(IndexedInstrProf::Header) / sizeof(uint64_t);
239 
240   // Only write out all the fields except 'HashOffset'. We need
241   // to remember the offset of that field to allow back patching
242   // later.
243   for (int I = 0; I < N - 1; I++)
244     OS.write(reinterpret_cast<uint64_t *>(&Header)[I]);
245 
246   // Save the location of Header.HashOffset field in \c OS.
247   uint64_t HashTableStartFieldOffset = OS.tell();
248   // Reserve the space for HashOffset field.
249   OS.write(0);
250 
251   // Reserve space to write profile summary data.
252   uint32_t NumEntries = Cutoffs.size();
253   uint32_t SummarySize = Summary::getSize(Summary::NumKinds, NumEntries);
254   // Remember the summary offset.
255   uint64_t SummaryOffset = OS.tell();
256   for (unsigned I = 0; I < SummarySize / sizeof(uint64_t); I++)
257     OS.write(0);
258 
259   // Write the hash table.
260   uint64_t HashTableStart = Generator.Emit(OS.OS, *InfoObj);
261 
262   // Allocate space for data to be serialized out.
263   std::unique_ptr<IndexedInstrProf::Summary> TheSummary =
264       IndexedInstrProf::allocSummary(SummarySize);
265   // Compute the Summary and copy the data to the data
266   // structure to be serialized out (to disk or buffer).
267   setSummary(TheSummary.get(), PS);
268   InfoObj->TheProfileSummary = 0;
269 
270   // Now do the final patch:
271   PatchItem PatchItems[] = {
272       // Patch the Header.HashOffset field.
273       {HashTableStartFieldOffset, &HashTableStart, 1},
274       // Patch the summary data.
275       {SummaryOffset, reinterpret_cast<uint64_t *>(TheSummary.get()),
276        (int)(SummarySize / sizeof(uint64_t))}};
277   OS.patch(PatchItems, sizeof(PatchItems) / sizeof(*PatchItems));
278 }
279 
280 void InstrProfWriter::write(raw_fd_ostream &OS) {
281   // Write the hash table.
282   ProfOStream POS(OS);
283   writeImpl(POS);
284 }
285 
286 std::unique_ptr<MemoryBuffer> InstrProfWriter::writeBuffer() {
287   std::string Data;
288   llvm::raw_string_ostream OS(Data);
289   ProfOStream POS(OS);
290   // Write the hash table.
291   writeImpl(POS);
292   // Return this in an aligned memory buffer.
293   return MemoryBuffer::getMemBufferCopy(Data);
294 }
295 
296 static const char *ValueProfKindStr[] = {
297 #define VALUE_PROF_KIND(Enumerator, Value) #Enumerator,
298 #include "llvm/ProfileData/InstrProfData.inc"
299 };
300 
301 void InstrProfWriter::writeRecordInText(const InstrProfRecord &Func,
302                                         InstrProfSymtab &Symtab,
303                                         raw_fd_ostream &OS) {
304   OS << Func.Name << "\n";
305   OS << "# Func Hash:\n" << Func.Hash << "\n";
306   OS << "# Num Counters:\n" << Func.Counts.size() << "\n";
307   OS << "# Counter Values:\n";
308   for (uint64_t Count : Func.Counts)
309     OS << Count << "\n";
310 
311   uint32_t NumValueKinds = Func.getNumValueKinds();
312   if (!NumValueKinds) {
313     OS << "\n";
314     return;
315   }
316 
317   OS << "# Num Value Kinds:\n" << Func.getNumValueKinds() << "\n";
318   for (uint32_t VK = 0; VK < IPVK_Last + 1; VK++) {
319     uint32_t NS = Func.getNumValueSites(VK);
320     if (!NS)
321       continue;
322     OS << "# ValueKind = " << ValueProfKindStr[VK] << ":\n" << VK << "\n";
323     OS << "# NumValueSites:\n" << NS << "\n";
324     for (uint32_t S = 0; S < NS; S++) {
325       uint32_t ND = Func.getNumValueDataForSite(VK, S);
326       OS << ND << "\n";
327       std::unique_ptr<InstrProfValueData[]> VD = Func.getValueForSite(VK, S);
328       for (uint32_t I = 0; I < ND; I++) {
329         if (VK == IPVK_IndirectCallTarget)
330           OS << Symtab.getFuncName(VD[I].Value) << ":" << VD[I].Count << "\n";
331         else
332           OS << VD[I].Value << ":" << VD[I].Count << "\n";
333       }
334     }
335   }
336 
337   OS << "\n";
338 }
339 
340 void InstrProfWriter::writeText(raw_fd_ostream &OS) {
341   if (ProfileKind == PF_IRLevel)
342     OS << "# IR level Instrumentation Flag\n:ir\n";
343   InstrProfSymtab Symtab;
344   for (const auto &I : FunctionData)
345     if (shouldEncodeData(I.getValue()))
346       Symtab.addFuncName(I.getKey());
347   Symtab.finalizeSymtab();
348 
349   for (const auto &I : FunctionData)
350     if (shouldEncodeData(I.getValue()))
351       for (const auto &Func : I.getValue())
352         writeRecordInText(Func.second, Symtab, OS);
353 }
354