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 
20 using namespace llvm;
21 
22 namespace {
23 class InstrProfRecordTrait {
24 public:
25   typedef StringRef key_type;
26   typedef StringRef key_type_ref;
27 
28   typedef const InstrProfWriter::ProfilingData *const data_type;
29   typedef const InstrProfWriter::ProfilingData *const data_type_ref;
30 
31   typedef uint64_t hash_value_type;
32   typedef uint64_t offset_type;
33 
34   static hash_value_type ComputeHash(key_type_ref K) {
35     return IndexedInstrProf::ComputeHash(IndexedInstrProf::HashType, K);
36   }
37 
38   static std::pair<offset_type, offset_type>
39   EmitKeyDataLength(raw_ostream &Out, key_type_ref K, data_type_ref V) {
40     using namespace llvm::support;
41     endian::Writer<little> LE(Out);
42 
43     offset_type N = K.size();
44     LE.write<offset_type>(N);
45 
46     offset_type M = 0;
47     for (const auto &ProfileData : *V) {
48       const InstrProfRecord &ProfRecord = ProfileData.second;
49       M += sizeof(uint64_t); // The function hash
50       M += sizeof(uint64_t); // The size of the Counts vector
51       M += ProfRecord.Counts.size() * sizeof(uint64_t);
52 
53       // Value data
54       M += sizeof(uint64_t); // Number of value kinds with value sites.
55       for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
56         uint32_t NumValueSites = ProfRecord.getNumValueSites(Kind);
57         if (NumValueSites == 0) continue;
58         M += sizeof(uint64_t); // Value kind
59         M += sizeof(uint64_t); // The number of value sites for given value kind
60         for (uint32_t I = 0; I < NumValueSites; I++) {
61           M += sizeof(uint64_t); // Number of value data pairs at a value site
62           uint64_t NumValueDataForSite =
63               ProfRecord.getNumValueDataForSite(Kind, I);
64           M += 2 * sizeof(uint64_t) * NumValueDataForSite;  // Value data pairs
65         }
66       }
67     }
68     LE.write<offset_type>(M);
69 
70     return std::make_pair(N, M);
71   }
72 
73   static void EmitKey(raw_ostream &Out, key_type_ref K, offset_type N){
74     Out.write(K.data(), N);
75   }
76 
77   static void EmitData(raw_ostream &Out, key_type_ref, data_type_ref V,
78                        offset_type) {
79     using namespace llvm::support;
80     endian::Writer<little> LE(Out);
81     for (const auto &ProfileData : *V) {
82       const InstrProfRecord &ProfRecord = ProfileData.second;
83 
84       LE.write<uint64_t>(ProfileData.first); // Function hash
85       LE.write<uint64_t>(ProfRecord.Counts.size());
86       for (uint64_t I : ProfRecord.Counts) LE.write<uint64_t>(I);
87 
88       // Compute the number of value kinds with value sites.
89       uint64_t NumValueKinds = ProfRecord.getNumValueKinds();
90       LE.write<uint64_t>(NumValueKinds);
91 
92       // Write value data
93       for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
94         uint32_t NumValueSites = ProfRecord.getNumValueSites(Kind);
95         if (NumValueSites == 0) continue;
96         LE.write<uint64_t>(Kind); // Write value kind
97         // Write number of value sites for current value kind
98         LE.write<uint64_t>(NumValueSites);
99 
100         for (uint32_t I = 0; I < NumValueSites; I++) {
101           // Write number of value data pairs at this value site
102           uint64_t NumValueDataForSite =
103               ProfRecord.getNumValueDataForSite(Kind, I);
104           LE.write<uint64_t>(NumValueDataForSite);
105           std::unique_ptr<InstrProfValueData[]> VD =
106               ProfRecord.getValueForSite(Kind, I);
107 
108           for (uint32_t V = 0; V < NumValueDataForSite; V++) {
109             if (Kind == IPVK_IndirectCallTarget)
110               LE.write<uint64_t>(ComputeHash((const char *)VD[V].Value));
111             else
112               LE.write<uint64_t>(VD[V].Value);
113             LE.write<uint64_t>(VD[V].Count);
114           }
115         }
116       }
117     }
118   }
119 };
120 }
121 
122 static std::error_code combineInstrProfRecords(InstrProfRecord &Dest,
123                                                InstrProfRecord &Source,
124                                                uint64_t &MaxFunctionCount) {
125   // If the number of counters doesn't match we either have bad data
126   // or a hash collision.
127   if (Dest.Counts.size() != Source.Counts.size())
128     return instrprof_error::count_mismatch;
129 
130   for (size_t I = 0, E = Source.Counts.size(); I < E; ++I) {
131     if (Dest.Counts[I] + Source.Counts[I] < Dest.Counts[I])
132       return instrprof_error::counter_overflow;
133     Dest.Counts[I] += Source.Counts[I];
134   }
135 
136   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
137     if (std::error_code EC = Dest.mergeValueProfData(Kind, Source)) return EC;
138   }
139 
140   // We keep track of the max function count as we go for simplicity.
141   if (Dest.Counts[0] > MaxFunctionCount)
142     MaxFunctionCount = Dest.Counts[0];
143 
144   return instrprof_error::success;
145 }
146 
147 void InstrProfWriter::updateStringTableReferences(InstrProfRecord &I) {
148   I.updateStrings(&StringTable);
149 }
150 
151 std::error_code InstrProfWriter::addRecord(InstrProfRecord &&I) {
152   updateStringTableReferences(I);
153   auto &ProfileDataMap = FunctionData[I.Name];
154 
155   auto Where = ProfileDataMap.find(I.Hash);
156   if (Where == ProfileDataMap.end()) {
157     // We've never seen a function with this name and hash, add it.
158     ProfileDataMap[I.Hash] = I;
159 
160     // We keep track of the max function count as we go for simplicity.
161     if (I.Counts[0] > MaxFunctionCount)
162       MaxFunctionCount = I.Counts[0];
163     return instrprof_error::success;
164   }
165 
166   // We're updating a function we've seen before.
167   return combineInstrProfRecords(Where->second, I, MaxFunctionCount);
168 }
169 
170 std::pair<uint64_t, uint64_t> InstrProfWriter::writeImpl(raw_ostream &OS) {
171   OnDiskChainedHashTableGenerator<InstrProfRecordTrait> Generator;
172 
173   // Populate the hash table generator.
174   for (const auto &I : FunctionData)
175     Generator.insert(I.getKey(), &I.getValue());
176 
177   using namespace llvm::support;
178   endian::Writer<little> LE(OS);
179 
180   // Write the header.
181   IndexedInstrProf::Header Header;
182   Header.Magic = IndexedInstrProf::Magic;
183   Header.Version = IndexedInstrProf::Version;
184   Header.MaxFunctionCount = MaxFunctionCount;
185   Header.HashType = static_cast<uint64_t>(IndexedInstrProf::HashType);
186   Header.HashOffset = 0;
187   int N = sizeof(IndexedInstrProf::Header) / sizeof(uint64_t);
188 
189   // Only write out all the fields execpt 'HashOffset'. We need
190   // to remember the offset of that field to allow back patching
191   // later.
192   for (int I = 0; I < N - 1; I++)
193     LE.write<uint64_t>(reinterpret_cast<uint64_t *>(&Header)[I]);
194 
195   // Save a space to write the hash table start location.
196   uint64_t HashTableStartLoc = OS.tell();
197   // Reserve the space for HashOffset field.
198   LE.write<uint64_t>(0);
199   // Write the hash table.
200   uint64_t HashTableStart = Generator.Emit(OS);
201 
202   return std::make_pair(HashTableStartLoc, HashTableStart);
203 }
204 
205 void InstrProfWriter::write(raw_fd_ostream &OS) {
206   // Write the hash table.
207   auto TableStart = writeImpl(OS);
208 
209   // Go back and fill in the hash table start.
210   using namespace support;
211   OS.seek(TableStart.first);
212   // Now patch the HashOffset field previously reserved.
213   endian::Writer<little>(OS).write<uint64_t>(TableStart.second);
214 }
215 
216 std::unique_ptr<MemoryBuffer> InstrProfWriter::writeBuffer() {
217   std::string Data;
218   llvm::raw_string_ostream OS(Data);
219   // Write the hash table.
220   auto TableStart = writeImpl(OS);
221   OS.flush();
222 
223   // Go back and fill in the hash table start.
224   using namespace support;
225   uint64_t Bytes = endian::byte_swap<uint64_t, little>(TableStart.second);
226   Data.replace(TableStart.first, sizeof(uint64_t), (const char *)&Bytes,
227                sizeof(uint64_t));
228 
229   // Return this in an aligned memory buffer.
230   return MemoryBuffer::getMemBufferCopy(Data);
231 }
232