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 "llvm/ADT/STLExtras.h"
21 static ErrorOr<std::unique_ptr<MemoryBuffer>>
22 setupMemoryBuffer(std::string Path) {
23 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
24 MemoryBuffer::getFileOrSTDIN(Path);
25 if (std::error_code EC = BufferOrErr.getError())
27 return std::move(BufferOrErr.get());
30 static std::error_code initializeReader(InstrProfReader &Reader) {
31 return Reader.readHeader();
34 ErrorOr<std::unique_ptr<InstrProfReader>>
35 InstrProfReader::create(std::string Path) {
36 // Set up the buffer to read.
37 auto BufferOrError = setupMemoryBuffer(Path);
38 if (std::error_code EC = BufferOrError.getError())
40 return InstrProfReader::create(std::move(BufferOrError.get()));
43 ErrorOr<std::unique_ptr<InstrProfReader>>
44 InstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer) {
45 // Sanity check the buffer.
46 if (Buffer->getBufferSize() > std::numeric_limits<unsigned>::max())
47 return instrprof_error::too_large;
49 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)));
57 else if (TextInstrProfReader::hasFormat(*Buffer))
58 Result.reset(new TextInstrProfReader(std::move(Buffer)));
60 return instrprof_error::unrecognized_format;
62 // Initialize the reader and return the result.
63 if (std::error_code EC = initializeReader(*Result))
66 return std::move(Result);
69 ErrorOr<std::unique_ptr<IndexedInstrProfReader>>
70 IndexedInstrProfReader::create(std::string Path) {
71 // Set up the buffer to read.
72 auto BufferOrError = setupMemoryBuffer(Path);
73 if (std::error_code EC = BufferOrError.getError())
75 return IndexedInstrProfReader::create(std::move(BufferOrError.get()));
79 ErrorOr<std::unique_ptr<IndexedInstrProfReader>>
80 IndexedInstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer) {
81 // Sanity check the buffer.
82 if (Buffer->getBufferSize() > std::numeric_limits<unsigned>::max())
83 return instrprof_error::too_large;
86 if (!IndexedInstrProfReader::hasFormat(*Buffer))
87 return instrprof_error::bad_magic;
88 auto Result = llvm::make_unique<IndexedInstrProfReader>(std::move(Buffer));
90 // Initialize the reader and return the result.
91 if (std::error_code EC = initializeReader(*Result))
94 return std::move(Result);
97 void InstrProfIterator::Increment() {
98 if (Reader->readNextRecord(Record))
99 *this = InstrProfIterator();
102 bool TextInstrProfReader::hasFormat(const MemoryBuffer &Buffer) {
103 // Verify that this really looks like plain ASCII text by checking a
104 // 'reasonable' number of characters (up to profile magic size).
105 size_t count = std::min(Buffer.getBufferSize(), sizeof(uint64_t));
106 StringRef buffer = Buffer.getBufferStart();
107 return count == 0 || std::all_of(buffer.begin(), buffer.begin() + count,
108 [](char c) { return ::isprint(c) || ::isspace(c); });
111 std::error_code TextInstrProfReader::readNextRecord(InstrProfRecord &Record) {
112 // Skip empty lines and comments.
113 while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
115 // If we hit EOF while looking for a name, we're done.
116 if (Line.is_at_end())
117 return error(instrprof_error::eof);
119 // Read the function name.
120 Record.Name = *Line++;
122 // Read the function hash.
123 if (Line.is_at_end())
124 return error(instrprof_error::truncated);
125 if ((Line++)->getAsInteger(0, Record.Hash))
126 return error(instrprof_error::malformed);
128 // Read the number of counters.
129 uint64_t NumCounters;
130 if (Line.is_at_end())
131 return error(instrprof_error::truncated);
132 if ((Line++)->getAsInteger(10, NumCounters))
133 return error(instrprof_error::malformed);
134 if (NumCounters == 0)
135 return error(instrprof_error::malformed);
137 // Read each counter and fill our internal storage with the values.
138 Record.Counts.clear();
139 Record.Counts.reserve(NumCounters);
140 for (uint64_t I = 0; I < NumCounters; ++I) {
141 if (Line.is_at_end())
142 return error(instrprof_error::truncated);
144 if ((Line++)->getAsInteger(10, Count))
145 return error(instrprof_error::malformed);
146 Record.Counts.push_back(Count);
152 template <class IntPtrT>
153 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
154 if (DataBuffer.getBufferSize() < sizeof(uint64_t))
157 *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
158 return RawInstrProf::getMagic<IntPtrT>() == Magic ||
159 sys::getSwappedBytes(RawInstrProf::getMagic<IntPtrT>()) == Magic;
162 template <class IntPtrT>
163 std::error_code RawInstrProfReader<IntPtrT>::readHeader() {
164 if (!hasFormat(*DataBuffer))
165 return error(instrprof_error::bad_magic);
166 if (DataBuffer->getBufferSize() < sizeof(RawInstrProf::Header))
167 return error(instrprof_error::bad_header);
168 auto *Header = reinterpret_cast<const RawInstrProf::Header *>(
169 DataBuffer->getBufferStart());
170 ShouldSwapBytes = Header->Magic != RawInstrProf::getMagic<IntPtrT>();
171 return readHeader(*Header);
174 template <class IntPtrT>
176 RawInstrProfReader<IntPtrT>::readNextHeader(const char *CurrentPos) {
177 const char *End = DataBuffer->getBufferEnd();
178 // Skip zero padding between profiles.
179 while (CurrentPos != End && *CurrentPos == 0)
181 // If there's nothing left, we're done.
182 if (CurrentPos == End)
183 return instrprof_error::eof;
184 // If there isn't enough space for another header, this is probably just
185 // garbage at the end of the file.
186 if (CurrentPos + sizeof(RawInstrProf::Header) > End)
187 return instrprof_error::malformed;
188 // The writer ensures each profile is padded to start at an aligned address.
189 if (reinterpret_cast<size_t>(CurrentPos) % alignOf<uint64_t>())
190 return instrprof_error::malformed;
191 // The magic should have the same byte order as in the previous header.
192 uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos);
193 if (Magic != swap(RawInstrProf::getMagic<IntPtrT>()))
194 return instrprof_error::bad_magic;
196 // There's another profile to read, so we need to process the header.
197 auto *Header = reinterpret_cast<const RawInstrProf::Header *>(CurrentPos);
198 return readHeader(*Header);
201 template <class IntPtrT>
202 std::error_code RawInstrProfReader<IntPtrT>::readHeader(
203 const RawInstrProf::Header &Header) {
204 if (swap(Header.Version) != RawInstrProf::Version)
205 return error(instrprof_error::unsupported_version);
207 CountersDelta = swap(Header.CountersDelta);
208 NamesDelta = swap(Header.NamesDelta);
209 auto DataSize = swap(Header.DataSize);
210 auto CountersSize = swap(Header.CountersSize);
211 auto NamesSize = swap(Header.NamesSize);
213 ptrdiff_t DataOffset = sizeof(RawInstrProf::Header);
214 ptrdiff_t CountersOffset =
215 DataOffset + sizeof(RawInstrProf::ProfileData<IntPtrT>) * DataSize;
216 ptrdiff_t NamesOffset = CountersOffset + sizeof(uint64_t) * CountersSize;
217 size_t ProfileSize = NamesOffset + sizeof(char) * NamesSize;
219 auto *Start = reinterpret_cast<const char *>(&Header);
220 if (Start + ProfileSize > DataBuffer->getBufferEnd())
221 return error(instrprof_error::bad_header);
223 Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>(
225 DataEnd = Data + DataSize;
226 CountersStart = reinterpret_cast<const uint64_t *>(Start + CountersOffset);
227 NamesStart = Start + NamesOffset;
228 ProfileEnd = Start + ProfileSize;
233 template <class IntPtrT>
234 std::error_code RawInstrProfReader<IntPtrT>::readName(InstrProfRecord &Record) {
235 Record.Name = StringRef(getName(Data->NamePtr), swap(Data->NameSize));
236 if (Record.Name.data() < NamesStart ||
237 Record.Name.data() + Record.Name.size() > DataBuffer->getBufferEnd())
238 return error(instrprof_error::malformed);
243 template <class IntPtrT>
244 std::error_code RawInstrProfReader<IntPtrT>::readFuncHash(
245 InstrProfRecord &Record) {
246 Record.Hash = swap(Data->FuncHash);
250 template <class IntPtrT>
251 std::error_code RawInstrProfReader<IntPtrT>::readRawCounts(
252 InstrProfRecord &Record) {
253 uint32_t NumCounters = swap(Data->NumCounters);
254 IntPtrT CounterPtr = Data->CounterPtr;
255 if (NumCounters == 0)
256 return error(instrprof_error::malformed);
258 auto RawCounts = makeArrayRef(getCounter(CounterPtr), NumCounters);
259 auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
262 if (RawCounts.data() < CountersStart ||
263 RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
264 return error(instrprof_error::malformed);
266 if (ShouldSwapBytes) {
267 Record.Counts.clear();
268 Record.Counts.reserve(RawCounts.size());
269 for (uint64_t Count : RawCounts)
270 Record.Counts.push_back(swap(Count));
272 Record.Counts = RawCounts;
277 template <class IntPtrT>
279 RawInstrProfReader<IntPtrT>::readNextRecord(InstrProfRecord &Record) {
281 if (std::error_code EC = readNextHeader(ProfileEnd))
284 // Read name ad set it in Record.
285 if (std::error_code EC = readName(Record))
288 // Read FuncHash and set it in Record.
289 if (std::error_code EC = readFuncHash(Record))
292 // Read raw counts and set Record.
293 if (std::error_code EC = readRawCounts(Record))
302 template class RawInstrProfReader<uint32_t>;
303 template class RawInstrProfReader<uint64_t>;
306 InstrProfLookupTrait::hash_value_type
307 InstrProfLookupTrait::ComputeHash(StringRef K) {
308 return IndexedInstrProf::ComputeHash(HashType, K);
311 typedef InstrProfLookupTrait::data_type data_type;
312 typedef InstrProfLookupTrait::offset_type offset_type;
314 bool InstrProfLookupTrait::ReadValueProfilingData(
315 const unsigned char *&D, const unsigned char *const End) {
316 ErrorOr<std::unique_ptr<IndexedInstrProf::ValueProfData>> VDataPtrOrErr =
317 IndexedInstrProf::ValueProfData::getValueProfData(
318 D, End, ValueProfDataEndianness);
320 if (VDataPtrOrErr.getError())
323 VDataPtrOrErr.get()->deserializeTo(DataBuffer.back(), &HashKeys);
324 D += VDataPtrOrErr.get()->TotalSize;
329 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
331 // Check if the data is corrupt. If so, don't try to read it.
332 if (N % sizeof(uint64_t))
336 std::vector<uint64_t> CounterBuffer;
338 using namespace support;
339 const unsigned char *End = D + N;
342 if (D + sizeof(uint64_t) >= End)
344 uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
346 // Initialize number of counters for FormatVersion == 1.
347 uint64_t CountsSize = N / sizeof(uint64_t) - 1;
348 // If format version is different then read the number of counters.
349 if (FormatVersion != 1) {
350 if (D + sizeof(uint64_t) > End)
352 CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
354 // Read counter values.
355 if (D + CountsSize * sizeof(uint64_t) > End)
358 CounterBuffer.clear();
359 CounterBuffer.reserve(CountsSize);
360 for (uint64_t J = 0; J < CountsSize; ++J)
361 CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
363 DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer));
365 // Read value profiling data.
366 if (FormatVersion > 2 && !ReadValueProfilingData(D, End)) {
375 InstrProfReaderIndex::getRecords(StringRef FuncName,
376 ArrayRef<InstrProfRecord> &Data) {
377 auto Iter = Index->find(FuncName);
378 if (Iter == Index->end())
379 return instrprof_error::unknown_function;
383 return instrprof_error::malformed;
385 return instrprof_error::success;
388 std::error_code InstrProfReaderIndex::getRecords(
389 ArrayRef<InstrProfRecord> &Data) {
390 if (atEnd()) return instrprof_error::eof;
392 Data = *RecordIterator;
394 if (Data.empty()) return instrprof_error::malformed;
396 return instrprof_error::success;
399 void InstrProfReaderIndex::Init(const unsigned char *Buckets,
400 const unsigned char *const Payload,
401 const unsigned char *const Base,
402 IndexedInstrProf::HashT HashType,
404 FormatVersion = Version;
405 Index.reset(IndexType::Create(Buckets, Payload, Base,
406 InstrProfLookupTrait(HashType, Version)));
407 // Form the map of hash values to const char* keys in profiling data.
408 std::vector<std::pair<uint64_t, const char *>> HashKeys;
409 for (auto Key : Index->keys()) {
410 const char *KeyTableRef = StringTable.insertString(Key);
411 HashKeys.push_back(std::make_pair(ComputeHash(HashType, Key), KeyTableRef));
413 std::sort(HashKeys.begin(), HashKeys.end(), less_first());
414 HashKeys.erase(std::unique(HashKeys.begin(), HashKeys.end()), HashKeys.end());
415 // Set the hash key map for the InstrLookupTrait
416 Index->getInfoObj().setHashKeys(std::move(HashKeys));
417 RecordIterator = Index->data_begin();
420 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
421 if (DataBuffer.getBufferSize() < 8)
423 using namespace support;
425 endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
426 // Verify that it's magical.
427 return Magic == IndexedInstrProf::Magic;
430 std::error_code IndexedInstrProfReader::readHeader() {
431 const unsigned char *Start =
432 (const unsigned char *)DataBuffer->getBufferStart();
433 const unsigned char *Cur = Start;
434 if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
435 return error(instrprof_error::truncated);
437 using namespace support;
439 auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
440 Cur += sizeof(IndexedInstrProf::Header);
442 // Check the magic number.
443 uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
444 if (Magic != IndexedInstrProf::Magic)
445 return error(instrprof_error::bad_magic);
448 uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
449 if (FormatVersion > IndexedInstrProf::Version)
450 return error(instrprof_error::unsupported_version);
452 // Read the maximal function count.
454 endian::byte_swap<uint64_t, little>(Header->MaxFunctionCount);
456 // Read the hash type and start offset.
457 IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
458 endian::byte_swap<uint64_t, little>(Header->HashType));
459 if (HashType > IndexedInstrProf::HashT::Last)
460 return error(instrprof_error::unsupported_hash_type);
462 uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
464 // The rest of the file is an on disk hash table.
465 Index.Init(Start + HashOffset, Cur, Start, HashType, FormatVersion);
470 ErrorOr<InstrProfRecord>
471 IndexedInstrProfReader::getInstrProfRecord(StringRef FuncName,
473 ArrayRef<InstrProfRecord> Data;
474 std::error_code EC = Index.getRecords(FuncName, Data);
475 if (EC != instrprof_error::success)
477 // Found it. Look for counters with the right hash.
478 for (unsigned I = 0, E = Data.size(); I < E; ++I) {
479 // Check for a match and fill the vector if there is one.
480 if (Data[I].Hash == FuncHash) {
481 return std::move(Data[I]);
484 return error(instrprof_error::hash_mismatch);
488 IndexedInstrProfReader::getFunctionCounts(StringRef FuncName, uint64_t FuncHash,
489 std::vector<uint64_t> &Counts) {
490 ErrorOr<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash);
491 if (std::error_code EC = Record.getError())
494 Counts = Record.get().Counts;
498 std::error_code IndexedInstrProfReader::readNextRecord(
499 InstrProfRecord &Record) {
500 static unsigned RecordIndex = 0;
502 ArrayRef<InstrProfRecord> Data;
504 std::error_code EC = Index.getRecords(Data);
505 if (EC != instrprof_error::success)
508 Record = Data[RecordIndex++];
509 if (RecordIndex >= Data.size()) {
510 Index.advanceToNextKey();