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();
108 std::all_of(buffer.begin(), buffer.begin() + count,
109 [](char c) { return ::isprint(c) || ::isspace(c); });
112 std::error_code TextInstrProfReader::readNextRecord(InstrProfRecord &Record) {
113 // Skip empty lines and comments.
114 while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
116 // If we hit EOF while looking for a name, we're done.
117 if (Line.is_at_end())
118 return error(instrprof_error::eof);
120 // Read the function name.
121 Record.Name = *Line++;
123 // Read the function hash.
124 if (Line.is_at_end())
125 return error(instrprof_error::truncated);
126 if ((Line++)->getAsInteger(0, Record.Hash))
127 return error(instrprof_error::malformed);
129 // Read the number of counters.
130 uint64_t NumCounters;
131 if (Line.is_at_end())
132 return error(instrprof_error::truncated);
133 if ((Line++)->getAsInteger(10, NumCounters))
134 return error(instrprof_error::malformed);
135 if (NumCounters == 0)
136 return error(instrprof_error::malformed);
138 // Read each counter and fill our internal storage with the values.
139 Record.Counts.clear();
140 Record.Counts.reserve(NumCounters);
141 for (uint64_t I = 0; I < NumCounters; ++I) {
142 if (Line.is_at_end())
143 return error(instrprof_error::truncated);
145 if ((Line++)->getAsInteger(10, Count))
146 return error(instrprof_error::malformed);
147 Record.Counts.push_back(Count);
153 template <class IntPtrT>
154 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
155 if (DataBuffer.getBufferSize() < sizeof(uint64_t))
158 *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
159 return RawInstrProf::getMagic<IntPtrT>() == Magic ||
160 sys::getSwappedBytes(RawInstrProf::getMagic<IntPtrT>()) == Magic;
163 template <class IntPtrT>
164 std::error_code RawInstrProfReader<IntPtrT>::readHeader() {
165 if (!hasFormat(*DataBuffer))
166 return error(instrprof_error::bad_magic);
167 if (DataBuffer->getBufferSize() < sizeof(RawInstrProf::Header))
168 return error(instrprof_error::bad_header);
169 auto *Header = reinterpret_cast<const RawInstrProf::Header *>(
170 DataBuffer->getBufferStart());
171 ShouldSwapBytes = Header->Magic != RawInstrProf::getMagic<IntPtrT>();
172 return readHeader(*Header);
175 template <class IntPtrT>
177 RawInstrProfReader<IntPtrT>::readNextHeader(const char *CurrentPos) {
178 const char *End = DataBuffer->getBufferEnd();
179 // Skip zero padding between profiles.
180 while (CurrentPos != End && *CurrentPos == 0)
182 // If there's nothing left, we're done.
183 if (CurrentPos == End)
184 return instrprof_error::eof;
185 // If there isn't enough space for another header, this is probably just
186 // garbage at the end of the file.
187 if (CurrentPos + sizeof(RawInstrProf::Header) > End)
188 return instrprof_error::malformed;
189 // The writer ensures each profile is padded to start at an aligned address.
190 if (reinterpret_cast<size_t>(CurrentPos) % alignOf<uint64_t>())
191 return instrprof_error::malformed;
192 // The magic should have the same byte order as in the previous header.
193 uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos);
194 if (Magic != swap(RawInstrProf::getMagic<IntPtrT>()))
195 return instrprof_error::bad_magic;
197 // There's another profile to read, so we need to process the header.
198 auto *Header = reinterpret_cast<const RawInstrProf::Header *>(CurrentPos);
199 return readHeader(*Header);
202 template <class IntPtrT>
203 std::error_code RawInstrProfReader<IntPtrT>::readHeader(
204 const RawInstrProf::Header &Header) {
205 if (swap(Header.Version) != RawInstrProf::Version)
206 return error(instrprof_error::unsupported_version);
208 CountersDelta = swap(Header.CountersDelta);
209 NamesDelta = swap(Header.NamesDelta);
210 ValueDataDelta = swap(Header.ValueDataDelta);
211 auto DataSize = swap(Header.DataSize);
212 auto CountersSize = swap(Header.CountersSize);
213 auto NamesSize = swap(Header.NamesSize);
214 auto ValueDataSize = swap(Header.ValueDataSize);
215 ValueKindLast = swap(Header.ValueKindLast);
217 auto DataSizeInBytes = DataSize * sizeof(RawInstrProf::ProfileData<IntPtrT>);
218 auto PaddingSize = getNumPaddingBytes(NamesSize);
220 ptrdiff_t DataOffset = sizeof(RawInstrProf::Header);
221 ptrdiff_t CountersOffset = DataOffset + DataSizeInBytes;
222 ptrdiff_t NamesOffset = CountersOffset + sizeof(uint64_t) * CountersSize;
223 ptrdiff_t ValueDataOffset = NamesOffset + NamesSize + PaddingSize;
224 size_t ProfileSize = ValueDataOffset + ValueDataSize;
226 auto *Start = reinterpret_cast<const char *>(&Header);
227 if (Start + ProfileSize > DataBuffer->getBufferEnd())
228 return error(instrprof_error::bad_header);
230 Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>(
232 DataEnd = Data + DataSize;
233 CountersStart = reinterpret_cast<const uint64_t *>(Start + CountersOffset);
234 NamesStart = Start + NamesOffset;
235 ValueDataStart = reinterpret_cast<const uint8_t*>(Start + ValueDataOffset);
236 ProfileEnd = Start + ProfileSize;
238 FunctionPtrToNameMap.clear();
239 for (const RawInstrProf::ProfileData<IntPtrT> *I = Data; I != DataEnd; ++I) {
240 const IntPtrT FPtr = swap(I->FunctionPointer);
243 StringRef FunctionName(getName(I->NamePtr), swap(I->NameSize));
244 const char* NameEntryPtr = StringTable.insertString(FunctionName);
245 FunctionPtrToNameMap.push_back(std::pair<const IntPtrT, const char*>
246 (FPtr, NameEntryPtr));
248 std::sort(FunctionPtrToNameMap.begin(), FunctionPtrToNameMap.end(), less_first());
249 FunctionPtrToNameMap.erase(std::unique(FunctionPtrToNameMap.begin(),
250 FunctionPtrToNameMap.end()),
251 FunctionPtrToNameMap.end());
255 template <class IntPtrT>
256 std::error_code RawInstrProfReader<IntPtrT>::readName(InstrProfRecord &Record) {
257 Record.Name = StringRef(getName(Data->NamePtr), swap(Data->NameSize));
258 if (Record.Name.data() < NamesStart ||
259 Record.Name.data() + Record.Name.size() >
260 reinterpret_cast<const char *>(ValueDataStart))
261 return error(instrprof_error::malformed);
265 template <class IntPtrT>
266 std::error_code RawInstrProfReader<IntPtrT>::readFuncHash(
267 InstrProfRecord &Record) {
268 Record.Hash = swap(Data->FuncHash);
272 template <class IntPtrT>
273 std::error_code RawInstrProfReader<IntPtrT>::readRawCounts(
274 InstrProfRecord &Record) {
275 uint32_t NumCounters = swap(Data->NumCounters);
276 IntPtrT CounterPtr = Data->CounterPtr;
277 if (NumCounters == 0)
278 return error(instrprof_error::malformed);
280 auto RawCounts = makeArrayRef(getCounter(CounterPtr), NumCounters);
281 auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
284 if (RawCounts.data() < CountersStart ||
285 RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
286 return error(instrprof_error::malformed);
288 if (ShouldSwapBytes) {
289 Record.Counts.clear();
290 Record.Counts.reserve(RawCounts.size());
291 for (uint64_t Count : RawCounts)
292 Record.Counts.push_back(swap(Count));
294 Record.Counts = RawCounts;
299 template <class IntPtrT>
301 RawInstrProfReader<IntPtrT>::readValueProfilingData(InstrProfRecord &Record) {
303 Record.clearValueData();
304 if (!Data->Values || (ValueDataDelta == 0))
307 ErrorOr<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
308 ValueProfData::getValueProfData(getValueDataCounts(Data->Values),
309 (const unsigned char *)ProfileEnd,
310 getDataEndianness());
312 if (VDataPtrOrErr.getError())
313 return VDataPtrOrErr.getError();
315 VDataPtrOrErr.get()->deserializeTo(Record, &FunctionPtrToNameMap);
319 template <class IntPtrT>
321 RawInstrProfReader<IntPtrT>::readNextRecord(InstrProfRecord &Record) {
323 if (std::error_code EC = readNextHeader(ProfileEnd))
326 // Read name ad set it in Record.
327 if (std::error_code EC = readName(Record))
330 // Read FuncHash and set it in Record.
331 if (std::error_code EC = readFuncHash(Record))
334 // Read raw counts and set Record.
335 if (std::error_code EC = readRawCounts(Record))
338 // Read value data and set Record.
339 if (std::error_code EC = readValueProfilingData(Record))
348 template class RawInstrProfReader<uint32_t>;
349 template class RawInstrProfReader<uint64_t>;
352 InstrProfLookupTrait::hash_value_type
353 InstrProfLookupTrait::ComputeHash(StringRef K) {
354 return IndexedInstrProf::ComputeHash(HashType, K);
357 typedef InstrProfLookupTrait::data_type data_type;
358 typedef InstrProfLookupTrait::offset_type offset_type;
360 bool InstrProfLookupTrait::readValueProfilingData(
361 const unsigned char *&D, const unsigned char *const End) {
362 ErrorOr<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
363 ValueProfData::getValueProfData(D, End, ValueProfDataEndianness);
365 if (VDataPtrOrErr.getError())
368 VDataPtrOrErr.get()->deserializeTo(DataBuffer.back(), &HashKeys);
369 D += VDataPtrOrErr.get()->TotalSize;
374 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
376 // Check if the data is corrupt. If so, don't try to read it.
377 if (N % sizeof(uint64_t))
381 std::vector<uint64_t> CounterBuffer;
383 using namespace support;
384 const unsigned char *End = D + N;
387 if (D + sizeof(uint64_t) >= End)
389 uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
391 // Initialize number of counters for FormatVersion == 1.
392 uint64_t CountsSize = N / sizeof(uint64_t) - 1;
393 // If format version is different then read the number of counters.
394 if (FormatVersion != 1) {
395 if (D + sizeof(uint64_t) > End)
397 CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
399 // Read counter values.
400 if (D + CountsSize * sizeof(uint64_t) > End)
403 CounterBuffer.clear();
404 CounterBuffer.reserve(CountsSize);
405 for (uint64_t J = 0; J < CountsSize; ++J)
406 CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
408 DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer));
410 // Read value profiling data.
411 if (FormatVersion > 2 && !readValueProfilingData(D, End)) {
419 template <typename HashTableImpl>
420 std::error_code InstrProfReaderIndex<HashTableImpl>::getRecords(
421 StringRef FuncName, ArrayRef<InstrProfRecord> &Data) {
422 auto Iter = HashTable->find(FuncName);
423 if (Iter == HashTable->end())
424 return instrprof_error::unknown_function;
428 return instrprof_error::malformed;
430 return instrprof_error::success;
433 template <typename HashTableImpl>
434 std::error_code InstrProfReaderIndex<HashTableImpl>::getRecords(
435 ArrayRef<InstrProfRecord> &Data) {
437 return instrprof_error::eof;
439 Data = *RecordIterator;
442 return instrprof_error::malformed;
444 return instrprof_error::success;
447 template <typename HashTableImpl>
448 InstrProfReaderIndex<HashTableImpl>::InstrProfReaderIndex(
449 const unsigned char *Buckets, const unsigned char *const Payload,
450 const unsigned char *const Base, IndexedInstrProf::HashT HashType,
452 FormatVersion = Version;
453 HashTable.reset(HashTableImpl::Create(
454 Buckets, Payload, Base,
455 typename HashTableImpl::InfoType(HashType, Version)));
456 // Form the map of hash values to const char* keys in profiling data.
457 std::vector<std::pair<uint64_t, const char *>> HashKeys;
458 for (auto Key : HashTable->keys()) {
459 const char *KeyTableRef = StringTable.insertString(Key);
460 HashKeys.push_back(std::make_pair(ComputeHash(HashType, Key), KeyTableRef));
462 std::sort(HashKeys.begin(), HashKeys.end(), less_first());
463 HashKeys.erase(std::unique(HashKeys.begin(), HashKeys.end()), HashKeys.end());
464 // Set the hash key map for the InstrLookupTrait
465 HashTable->getInfoObj().setHashKeys(std::move(HashKeys));
466 RecordIterator = HashTable->data_begin();
469 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
470 if (DataBuffer.getBufferSize() < 8)
472 using namespace support;
474 endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
475 // Verify that it's magical.
476 return Magic == IndexedInstrProf::Magic;
479 std::error_code IndexedInstrProfReader::readHeader() {
480 const unsigned char *Start =
481 (const unsigned char *)DataBuffer->getBufferStart();
482 const unsigned char *Cur = Start;
483 if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
484 return error(instrprof_error::truncated);
486 using namespace support;
488 auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
489 Cur += sizeof(IndexedInstrProf::Header);
491 // Check the magic number.
492 uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
493 if (Magic != IndexedInstrProf::Magic)
494 return error(instrprof_error::bad_magic);
497 uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
498 if (FormatVersion > IndexedInstrProf::Version)
499 return error(instrprof_error::unsupported_version);
501 // Read the maximal function count.
503 endian::byte_swap<uint64_t, little>(Header->MaxFunctionCount);
505 // Read the hash type and start offset.
506 IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
507 endian::byte_swap<uint64_t, little>(Header->HashType));
508 if (HashType > IndexedInstrProf::HashT::Last)
509 return error(instrprof_error::unsupported_hash_type);
511 uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
513 // The rest of the file is an on disk hash table.
514 InstrProfReaderIndexBase *IndexPtr = nullptr;
515 IndexPtr = new InstrProfReaderIndex<OnDiskHashTableImplV3>(
516 Start + HashOffset, Cur, Start, HashType, FormatVersion);
517 Index.reset(IndexPtr);
521 ErrorOr<InstrProfRecord>
522 IndexedInstrProfReader::getInstrProfRecord(StringRef FuncName,
524 ArrayRef<InstrProfRecord> Data;
525 std::error_code EC = Index->getRecords(FuncName, Data);
526 if (EC != instrprof_error::success)
528 // Found it. Look for counters with the right hash.
529 for (unsigned I = 0, E = Data.size(); I < E; ++I) {
530 // Check for a match and fill the vector if there is one.
531 if (Data[I].Hash == FuncHash) {
532 return std::move(Data[I]);
535 return error(instrprof_error::hash_mismatch);
539 IndexedInstrProfReader::getFunctionCounts(StringRef FuncName, uint64_t FuncHash,
540 std::vector<uint64_t> &Counts) {
541 ErrorOr<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash);
542 if (std::error_code EC = Record.getError())
545 Counts = Record.get().Counts;
549 std::error_code IndexedInstrProfReader::readNextRecord(
550 InstrProfRecord &Record) {
551 static unsigned RecordIndex = 0;
553 ArrayRef<InstrProfRecord> Data;
555 std::error_code EC = Index->getRecords(Data);
556 if (EC != instrprof_error::success)
559 Record = Data[RecordIndex++];
560 if (RecordIndex >= Data.size()) {
561 Index->advanceToNextKey();