1 //=-- InstrProfReader.cpp - Instrumented profiling reader -------------------=//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains support for reading profiling data for clang's
11 // instrumentation based PGO and coverage.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/ProfileData/InstrProfReader.h"
16 #include "InstrProfIndexed.h"
17 #include "llvm/ProfileData/InstrProf.h"
22 static ErrorOr<std::unique_ptr<MemoryBuffer>>
23 setupMemoryBuffer(std::string Path) {
24 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
25 MemoryBuffer::getFileOrSTDIN(Path);
26 if (std::error_code EC = BufferOrErr.getError())
28 auto Buffer = std::move(BufferOrErr.get());
30 // Sanity check the file.
31 if (Buffer->getBufferSize() > std::numeric_limits<unsigned>::max())
32 return instrprof_error::too_large;
33 return std::move(Buffer);
36 static std::error_code initializeReader(InstrProfReader &Reader) {
37 return Reader.readHeader();
40 ErrorOr<std::unique_ptr<InstrProfReader>>
41 InstrProfReader::create(std::string Path) {
42 // Set up the buffer to read.
43 auto BufferOrError = setupMemoryBuffer(Path);
44 if (std::error_code EC = BufferOrError.getError())
47 auto Buffer = std::move(BufferOrError.get());
48 std::unique_ptr<InstrProfReader> Result;
51 if (IndexedInstrProfReader::hasFormat(*Buffer))
52 Result.reset(new IndexedInstrProfReader(std::move(Buffer)));
53 else if (RawInstrProfReader64::hasFormat(*Buffer))
54 Result.reset(new RawInstrProfReader64(std::move(Buffer)));
55 else if (RawInstrProfReader32::hasFormat(*Buffer))
56 Result.reset(new RawInstrProfReader32(std::move(Buffer)));
58 Result.reset(new TextInstrProfReader(std::move(Buffer)));
60 // Initialize the reader and return the result.
61 if (std::error_code EC = initializeReader(*Result))
64 return std::move(Result);
67 ErrorOr<std::unique_ptr<IndexedInstrProfReader>>
68 IndexedInstrProfReader::create(std::string Path) {
69 // Set up the buffer to read.
70 auto BufferOrError = setupMemoryBuffer(Path);
71 if (std::error_code EC = BufferOrError.getError())
74 auto Buffer = std::move(BufferOrError.get());
75 std::unique_ptr<IndexedInstrProfReader> Result;
78 if (!IndexedInstrProfReader::hasFormat(*Buffer))
79 return instrprof_error::bad_magic;
80 Result.reset(new IndexedInstrProfReader(std::move(Buffer)));
82 // Initialize the reader and return the result.
83 if (std::error_code EC = initializeReader(*Result))
86 return std::move(Result);
89 void InstrProfIterator::Increment() {
90 if (Reader->readNextRecord(Record))
91 *this = InstrProfIterator();
94 std::error_code TextInstrProfReader::readNextRecord(InstrProfRecord &Record) {
95 // Skip empty lines and comments.
96 while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
98 // If we hit EOF while looking for a name, we're done.
100 return error(instrprof_error::eof);
102 // Read the function name.
103 Record.Name = *Line++;
105 // Read the function hash.
106 if (Line.is_at_end())
107 return error(instrprof_error::truncated);
108 if ((Line++)->getAsInteger(10, Record.Hash))
109 return error(instrprof_error::malformed);
111 // Read the number of counters.
112 uint64_t NumCounters;
113 if (Line.is_at_end())
114 return error(instrprof_error::truncated);
115 if ((Line++)->getAsInteger(10, NumCounters))
116 return error(instrprof_error::malformed);
117 if (NumCounters == 0)
118 return error(instrprof_error::malformed);
120 // Read each counter and fill our internal storage with the values.
122 Counts.reserve(NumCounters);
123 for (uint64_t I = 0; I < NumCounters; ++I) {
124 if (Line.is_at_end())
125 return error(instrprof_error::truncated);
127 if ((Line++)->getAsInteger(10, Count))
128 return error(instrprof_error::malformed);
129 Counts.push_back(Count);
131 // Give the record a reference to our internal counter storage.
132 Record.Counts = Counts;
137 template <class IntPtrT>
138 static uint64_t getRawMagic();
141 uint64_t getRawMagic<uint64_t>() {
143 uint64_t(255) << 56 |
144 uint64_t('l') << 48 |
145 uint64_t('p') << 40 |
146 uint64_t('r') << 32 |
147 uint64_t('o') << 24 |
148 uint64_t('f') << 16 |
154 uint64_t getRawMagic<uint32_t>() {
156 uint64_t(255) << 56 |
157 uint64_t('l') << 48 |
158 uint64_t('p') << 40 |
159 uint64_t('r') << 32 |
160 uint64_t('o') << 24 |
161 uint64_t('f') << 16 |
166 template <class IntPtrT>
167 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
168 if (DataBuffer.getBufferSize() < sizeof(uint64_t))
171 *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
172 return getRawMagic<IntPtrT>() == Magic ||
173 sys::getSwappedBytes(getRawMagic<IntPtrT>()) == Magic;
176 template <class IntPtrT>
177 std::error_code RawInstrProfReader<IntPtrT>::readHeader() {
178 if (!hasFormat(*DataBuffer))
179 return error(instrprof_error::bad_magic);
180 if (DataBuffer->getBufferSize() < sizeof(RawHeader))
181 return error(instrprof_error::bad_header);
183 reinterpret_cast<const RawHeader *>(DataBuffer->getBufferStart());
184 ShouldSwapBytes = Header->Magic != getRawMagic<IntPtrT>();
185 return readHeader(*Header);
188 template <class IntPtrT>
190 RawInstrProfReader<IntPtrT>::readNextHeader(const char *CurrentPos) {
191 const char *End = DataBuffer->getBufferEnd();
192 // Skip zero padding between profiles.
193 while (CurrentPos != End && *CurrentPos == 0)
195 // If there's nothing left, we're done.
196 if (CurrentPos == End)
197 return instrprof_error::eof;
198 // If there isn't enough space for another header, this is probably just
199 // garbage at the end of the file.
200 if (CurrentPos + sizeof(RawHeader) > End)
201 return instrprof_error::malformed;
202 // The writer ensures each profile is padded to start at an aligned address.
203 if (reinterpret_cast<size_t>(CurrentPos) % alignOf<uint64_t>())
204 return instrprof_error::malformed;
205 // The magic should have the same byte order as in the previous header.
206 uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos);
207 if (Magic != swap(getRawMagic<IntPtrT>()))
208 return instrprof_error::bad_magic;
210 // There's another profile to read, so we need to process the header.
211 auto *Header = reinterpret_cast<const RawHeader *>(CurrentPos);
212 return readHeader(*Header);
215 static uint64_t getRawVersion() {
219 template <class IntPtrT>
221 RawInstrProfReader<IntPtrT>::readHeader(const RawHeader &Header) {
222 if (swap(Header.Version) != getRawVersion())
223 return error(instrprof_error::unsupported_version);
225 CountersDelta = swap(Header.CountersDelta);
226 NamesDelta = swap(Header.NamesDelta);
227 auto DataSize = swap(Header.DataSize);
228 auto CountersSize = swap(Header.CountersSize);
229 auto NamesSize = swap(Header.NamesSize);
231 ptrdiff_t DataOffset = sizeof(RawHeader);
232 ptrdiff_t CountersOffset = DataOffset + sizeof(ProfileData) * DataSize;
233 ptrdiff_t NamesOffset = CountersOffset + sizeof(uint64_t) * CountersSize;
234 size_t ProfileSize = NamesOffset + sizeof(char) * NamesSize;
236 auto *Start = reinterpret_cast<const char *>(&Header);
237 if (Start + ProfileSize > DataBuffer->getBufferEnd())
238 return error(instrprof_error::bad_header);
240 Data = reinterpret_cast<const ProfileData *>(Start + DataOffset);
241 DataEnd = Data + DataSize;
242 CountersStart = reinterpret_cast<const uint64_t *>(Start + CountersOffset);
243 NamesStart = Start + NamesOffset;
244 ProfileEnd = Start + ProfileSize;
249 template <class IntPtrT>
251 RawInstrProfReader<IntPtrT>::readNextRecord(InstrProfRecord &Record) {
253 if (std::error_code EC = readNextHeader(ProfileEnd))
257 StringRef RawName(getName(Data->NamePtr), swap(Data->NameSize));
258 uint32_t NumCounters = swap(Data->NumCounters);
259 if (NumCounters == 0)
260 return error(instrprof_error::malformed);
261 auto RawCounts = makeArrayRef(getCounter(Data->CounterPtr), NumCounters);
264 auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
265 if (RawName.data() < NamesStart ||
266 RawName.data() + RawName.size() > DataBuffer->getBufferEnd() ||
267 RawCounts.data() < CountersStart ||
268 RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
269 return error(instrprof_error::malformed);
271 // Store the data in Record, byte-swapping as necessary.
272 Record.Hash = swap(Data->FuncHash);
273 Record.Name = RawName;
274 if (ShouldSwapBytes) {
276 Counts.reserve(RawCounts.size());
277 for (uint64_t Count : RawCounts)
278 Counts.push_back(swap(Count));
279 Record.Counts = Counts;
281 Record.Counts = RawCounts;
289 template class RawInstrProfReader<uint32_t>;
290 template class RawInstrProfReader<uint64_t>;
293 InstrProfLookupTrait::hash_value_type
294 InstrProfLookupTrait::ComputeHash(StringRef K) {
295 return IndexedInstrProf::ComputeHash(HashType, K);
298 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
299 if (DataBuffer.getBufferSize() < 8)
301 using namespace support;
303 endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
304 return Magic == IndexedInstrProf::Magic;
307 std::error_code IndexedInstrProfReader::readHeader() {
308 const unsigned char *Start =
309 (const unsigned char *)DataBuffer->getBufferStart();
310 const unsigned char *Cur = Start;
311 if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
312 return error(instrprof_error::truncated);
314 using namespace support;
316 // Check the magic number.
317 uint64_t Magic = endian::readNext<uint64_t, little, unaligned>(Cur);
318 if (Magic != IndexedInstrProf::Magic)
319 return error(instrprof_error::bad_magic);
322 FormatVersion = endian::readNext<uint64_t, little, unaligned>(Cur);
323 if (FormatVersion > IndexedInstrProf::Version)
324 return error(instrprof_error::unsupported_version);
326 // Read the maximal function count.
327 MaxFunctionCount = endian::readNext<uint64_t, little, unaligned>(Cur);
329 // Read the hash type and start offset.
330 IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
331 endian::readNext<uint64_t, little, unaligned>(Cur));
332 if (HashType > IndexedInstrProf::HashT::Last)
333 return error(instrprof_error::unsupported_hash_type);
334 uint64_t HashOffset = endian::readNext<uint64_t, little, unaligned>(Cur);
336 // The rest of the file is an on disk hash table.
337 Index.reset(InstrProfReaderIndex::Create(Start + HashOffset, Cur, Start,
338 InstrProfLookupTrait(HashType)));
339 // Set up our iterator for readNextRecord.
340 RecordIterator = Index->data_begin();
345 std::error_code IndexedInstrProfReader::getFunctionCounts(
346 StringRef FuncName, uint64_t FuncHash, std::vector<uint64_t> &Counts) {
347 auto Iter = Index->find(FuncName);
348 if (Iter == Index->end())
349 return error(instrprof_error::unknown_function);
351 // Found it. Look for counters with the right hash.
352 ArrayRef<uint64_t> Data = (*Iter).Data;
354 for (uint64_t I = 0, E = Data.size(); I != E; I += NumCounts) {
355 // The function hash comes first.
356 uint64_t FoundHash = Data[I++];
357 // In v1, we have at least one count. Later, we have the number of counts.
359 return error(instrprof_error::malformed);
360 NumCounts = FormatVersion == 1 ? E - I : Data[I++];
361 // If we have more counts than data, this is bogus.
362 if (I + NumCounts > E)
363 return error(instrprof_error::malformed);
364 // Check for a match and fill the vector if there is one.
365 if (FoundHash == FuncHash) {
366 Counts = Data.slice(I, NumCounts);
370 return error(instrprof_error::hash_mismatch);
374 IndexedInstrProfReader::readNextRecord(InstrProfRecord &Record) {
375 // Are we out of records?
376 if (RecordIterator == Index->data_end())
377 return error(instrprof_error::eof);
379 // Record the current function name.
380 Record.Name = (*RecordIterator).Name;
382 ArrayRef<uint64_t> Data = (*RecordIterator).Data;
383 // Valid data starts with a hash and either a count or the number of counts.
384 if (CurrentOffset + 1 > Data.size())
385 return error(instrprof_error::malformed);
386 // First we have a function hash.
387 Record.Hash = Data[CurrentOffset++];
388 // In version 1 we knew the number of counters implicitly, but in newer
389 // versions we store the number of counters next.
391 FormatVersion == 1 ? Data.size() - CurrentOffset : Data[CurrentOffset++];
392 if (CurrentOffset + NumCounts > Data.size())
393 return error(instrprof_error::malformed);
394 // And finally the counts themselves.
395 Record.Counts = Data.slice(CurrentOffset, NumCounts);
397 // If we've exhausted this function's data, increment the record.
398 CurrentOffset += NumCounts;
399 if (CurrentOffset == Data.size()) {