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)
58           continue;
59         M += sizeof(uint64_t); // Value kind
60         M += sizeof(uint64_t); // The number of value sites for given value kind
61         for (uint32_t I = 0; I < NumValueSites; I++) {
62           M += sizeof(uint64_t); // Number of value data pairs at a value site
63           uint64_t NumValueDataForSite =
64               ProfRecord.getNumValueDataForSite(Kind, I);
65           M += 2 * sizeof(uint64_t) * NumValueDataForSite; // Value data pairs
66         }
67       }
68     }
69     LE.write<offset_type>(M);
70 
71     return std::make_pair(N, M);
72   }
73 
74   static void EmitKey(raw_ostream &Out, key_type_ref K, offset_type N){
75     Out.write(K.data(), N);
76   }
77 
78   static void EmitData(raw_ostream &Out, key_type_ref, data_type_ref V,
79                        offset_type) {
80     using namespace llvm::support;
81     endian::Writer<little> LE(Out);
82     for (const auto &ProfileData : *V) {
83       const InstrProfRecord &ProfRecord = ProfileData.second;
84 
85       LE.write<uint64_t>(ProfileData.first); // Function hash
86       LE.write<uint64_t>(ProfRecord.Counts.size());
87       for (uint64_t I : ProfRecord.Counts)
88         LE.write<uint64_t>(I);
89 
90       // Compute the number of value kinds with value sites.
91       uint64_t NumValueKinds = ProfRecord.getNumValueKinds();
92       LE.write<uint64_t>(NumValueKinds);
93 
94       // Write value data
95       for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
96         uint32_t NumValueSites = ProfRecord.getNumValueSites(Kind);
97         if (NumValueSites == 0)
98           continue;
99         LE.write<uint64_t>(Kind); // Write value kind
100         // Write number of value sites for current value kind
101         LE.write<uint64_t>(NumValueSites);
102 
103         for (uint32_t I = 0; I < NumValueSites; I++) {
104           // Write number of value data pairs at this value site
105           uint64_t NumValueDataForSite =
106               ProfRecord.getNumValueDataForSite(Kind, I);
107           LE.write<uint64_t>(NumValueDataForSite);
108           std::unique_ptr<InstrProfValueData[]> VD =
109               ProfRecord.getValueForSite(Kind, I);
110 
111           for (uint32_t V = 0; V < NumValueDataForSite; V++) {
112             if (Kind == IPVK_IndirectCallTarget)
113               LE.write<uint64_t>(ComputeHash((const char *)VD[V].Value));
114             else
115               LE.write<uint64_t>(VD[V].Value);
116             LE.write<uint64_t>(VD[V].Count);
117           }
118         }
119       }
120     }
121   }
122 };
123 }
124 
125 static std::error_code combineInstrProfRecords(InstrProfRecord &Dest,
126                                                InstrProfRecord &Source,
127                                                uint64_t &MaxFunctionCount) {
128   // If the number of counters doesn't match we either have bad data
129   // or a hash collision.
130   if (Dest.Counts.size() != Source.Counts.size())
131     return instrprof_error::count_mismatch;
132 
133   for (size_t I = 0, E = Source.Counts.size(); I < E; ++I) {
134     if (Dest.Counts[I] + Source.Counts[I] < Dest.Counts[I])
135       return instrprof_error::counter_overflow;
136     Dest.Counts[I] += Source.Counts[I];
137   }
138 
139   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
140     if (std::error_code EC = Dest.mergeValueProfData(Kind, Source))
141       return EC;
142   }
143 
144   // We keep track of the max function count as we go for simplicity.
145   if (Dest.Counts[0] > MaxFunctionCount)
146     MaxFunctionCount = Dest.Counts[0];
147 
148   return instrprof_error::success;
149 }
150 
151 void InstrProfWriter::updateStringTableReferences(InstrProfRecord &I) {
152   I.updateStrings(&StringTable);
153 }
154 
155 std::error_code InstrProfWriter::addRecord(InstrProfRecord &&I) {
156   updateStringTableReferences(I);
157   auto &ProfileDataMap = FunctionData[I.Name];
158 
159   auto Where = ProfileDataMap.find(I.Hash);
160   if (Where == ProfileDataMap.end()) {
161     // We've never seen a function with this name and hash, add it.
162     ProfileDataMap[I.Hash] = I;
163 
164     // We keep track of the max function count as we go for simplicity.
165     if (I.Counts[0] > MaxFunctionCount)
166       MaxFunctionCount = I.Counts[0];
167     return instrprof_error::success;
168   }
169 
170   // We're updating a function we've seen before.
171   return combineInstrProfRecords(Where->second, I, MaxFunctionCount);
172 }
173 
174 std::pair<uint64_t, uint64_t> InstrProfWriter::writeImpl(raw_ostream &OS) {
175   OnDiskChainedHashTableGenerator<InstrProfRecordTrait> Generator;
176 
177   // Populate the hash table generator.
178   for (const auto &I : FunctionData)
179     Generator.insert(I.getKey(), &I.getValue());
180 
181   using namespace llvm::support;
182   endian::Writer<little> LE(OS);
183 
184   // Write the header.
185   IndexedInstrProf::Header Header;
186   Header.Magic = IndexedInstrProf::Magic;
187   Header.Version = IndexedInstrProf::Version;
188   Header.MaxFunctionCount = MaxFunctionCount;
189   Header.HashType = static_cast<uint64_t>(IndexedInstrProf::HashType);
190   Header.HashOffset = 0;
191   int N = sizeof(IndexedInstrProf::Header) / sizeof(uint64_t);
192 
193   // Only write out all the fields execpt 'HashOffset'. We need
194   // to remember the offset of that field to allow back patching
195   // later.
196   for (int I = 0; I < N - 1; I++)
197     LE.write<uint64_t>(reinterpret_cast<uint64_t *>(&Header)[I]);
198 
199   // Save a space to write the hash table start location.
200   uint64_t HashTableStartLoc = OS.tell();
201   // Reserve the space for HashOffset field.
202   LE.write<uint64_t>(0);
203   // Write the hash table.
204   uint64_t HashTableStart = Generator.Emit(OS);
205 
206   return std::make_pair(HashTableStartLoc, HashTableStart);
207 }
208 
209 void InstrProfWriter::write(raw_fd_ostream &OS) {
210   // Write the hash table.
211   auto TableStart = writeImpl(OS);
212 
213   // Go back and fill in the hash table start.
214   using namespace support;
215   OS.seek(TableStart.first);
216   // Now patch the HashOffset field previously reserved.
217   endian::Writer<little>(OS).write<uint64_t>(TableStart.second);
218 }
219 
220 std::unique_ptr<MemoryBuffer> InstrProfWriter::writeBuffer() {
221   std::string Data;
222   llvm::raw_string_ostream OS(Data);
223   // Write the hash table.
224   auto TableStart = writeImpl(OS);
225   OS.flush();
226 
227   // Go back and fill in the hash table start.
228   using namespace support;
229   uint64_t Bytes = endian::byte_swap<uint64_t, little>(TableStart.second);
230   Data.replace(TableStart.first, sizeof(uint64_t), (const char *)&Bytes,
231                sizeof(uint64_t));
232 
233   // Return this in an aligned memory buffer.
234   return MemoryBuffer::getMemBufferCopy(Data);
235 }
236