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 auto DataSize = swap(Header.DataSize);
211 auto CountersSize = swap(Header.CountersSize);
212 auto NamesSize = swap(Header.NamesSize);
213 auto ValueDataSize = swap(Header.ValueDataSize);
214 ValueKindLast = swap(Header.ValueKindLast);
216 auto DataSizeInBytes = DataSize * sizeof(RawInstrProf::ProfileData<IntPtrT>);
217 auto PaddingSize = getNumPaddingBytes(NamesSize);
219 ptrdiff_t DataOffset = sizeof(RawInstrProf::Header);
220 ptrdiff_t CountersOffset = DataOffset + DataSizeInBytes;
221 ptrdiff_t NamesOffset = CountersOffset + sizeof(uint64_t) * CountersSize;
222 ptrdiff_t ValueDataOffset = NamesOffset + NamesSize + PaddingSize;
223 size_t ProfileSize = ValueDataOffset + ValueDataSize;
225 auto *Start = reinterpret_cast<const char *>(&Header);
226 if (Start + ProfileSize > DataBuffer->getBufferEnd())
227 return error(instrprof_error::bad_header);
229 Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>(
231 DataEnd = Data + DataSize;
232 CountersStart = reinterpret_cast<const uint64_t *>(Start + CountersOffset);
233 NamesStart = Start + NamesOffset;
234 ValueDataStart = reinterpret_cast<const uint8_t*>(Start + ValueDataOffset);
235 ProfileEnd = Start + ProfileSize;
237 FunctionPtrToNameMap.clear();
238 for (const RawInstrProf::ProfileData<IntPtrT> *I = Data; I != DataEnd; ++I) {
239 const IntPtrT FPtr = swap(I->FunctionPointer);
242 StringRef FunctionName(getName(I->NamePtr), swap(I->NameSize));
243 const char* NameEntryPtr = StringTable.insertString(FunctionName);
244 FunctionPtrToNameMap.push_back(std::pair<const IntPtrT, const char*>
245 (FPtr, NameEntryPtr));
247 std::sort(FunctionPtrToNameMap.begin(), FunctionPtrToNameMap.end(), less_first());
248 FunctionPtrToNameMap.erase(std::unique(FunctionPtrToNameMap.begin(),
249 FunctionPtrToNameMap.end()),
250 FunctionPtrToNameMap.end());
254 template <class IntPtrT>
255 std::error_code RawInstrProfReader<IntPtrT>::readName(InstrProfRecord &Record) {
256 Record.Name = StringRef(getName(Data->NamePtr), swap(Data->NameSize));
257 if (Record.Name.data() < NamesStart ||
258 Record.Name.data() + Record.Name.size() >
259 reinterpret_cast<const char *>(ValueDataStart))
260 return error(instrprof_error::malformed);
264 template <class IntPtrT>
265 std::error_code RawInstrProfReader<IntPtrT>::readFuncHash(
266 InstrProfRecord &Record) {
267 Record.Hash = swap(Data->FuncHash);
271 template <class IntPtrT>
272 std::error_code RawInstrProfReader<IntPtrT>::readRawCounts(
273 InstrProfRecord &Record) {
274 uint32_t NumCounters = swap(Data->NumCounters);
275 IntPtrT CounterPtr = Data->CounterPtr;
276 if (NumCounters == 0)
277 return error(instrprof_error::malformed);
279 auto RawCounts = makeArrayRef(getCounter(CounterPtr), NumCounters);
280 auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
283 if (RawCounts.data() < CountersStart ||
284 RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
285 return error(instrprof_error::malformed);
287 if (ShouldSwapBytes) {
288 Record.Counts.clear();
289 Record.Counts.reserve(RawCounts.size());
290 for (uint64_t Count : RawCounts)
291 Record.Counts.push_back(swap(Count));
293 Record.Counts = RawCounts;
298 template <class IntPtrT>
300 RawInstrProfReader<IntPtrT>::readValueProfilingData(InstrProfRecord &Record) {
302 Record.clearValueData();
303 CurValueDataSize = 0;
304 // Need to match the logic in value profile dumper code in compiler-rt:
305 uint32_t NumValueKinds = 0;
306 for (uint32_t I = 0; I < IPVK_Last + 1; I++)
307 NumValueKinds += (Data->NumValueSites[I] != 0);
312 ErrorOr<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
313 ValueProfData::getValueProfData(ValueDataStart,
314 (const unsigned char *)ProfileEnd,
315 getDataEndianness());
317 if (VDataPtrOrErr.getError())
318 return VDataPtrOrErr.getError();
320 VDataPtrOrErr.get()->deserializeTo(Record, &FunctionPtrToNameMap);
321 CurValueDataSize = VDataPtrOrErr.get()->getSize();
325 template <class IntPtrT>
327 RawInstrProfReader<IntPtrT>::readNextRecord(InstrProfRecord &Record) {
329 if (std::error_code EC = readNextHeader(ProfileEnd))
332 // Read name ad set it in Record.
333 if (std::error_code EC = readName(Record))
336 // Read FuncHash and set it in Record.
337 if (std::error_code EC = readFuncHash(Record))
340 // Read raw counts and set Record.
341 if (std::error_code EC = readRawCounts(Record))
344 // Read value data and set Record.
345 if (std::error_code EC = readValueProfilingData(Record))
354 template class RawInstrProfReader<uint32_t>;
355 template class RawInstrProfReader<uint64_t>;
358 InstrProfLookupTrait::hash_value_type
359 InstrProfLookupTrait::ComputeHash(StringRef K) {
360 return IndexedInstrProf::ComputeHash(HashType, K);
363 typedef InstrProfLookupTrait::data_type data_type;
364 typedef InstrProfLookupTrait::offset_type offset_type;
366 bool InstrProfLookupTrait::readValueProfilingData(
367 const unsigned char *&D, const unsigned char *const End) {
368 ErrorOr<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
369 ValueProfData::getValueProfData(D, End, ValueProfDataEndianness);
371 if (VDataPtrOrErr.getError())
374 VDataPtrOrErr.get()->deserializeTo(DataBuffer.back(), &HashKeys);
375 D += VDataPtrOrErr.get()->TotalSize;
380 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
382 // Check if the data is corrupt. If so, don't try to read it.
383 if (N % sizeof(uint64_t))
387 std::vector<uint64_t> CounterBuffer;
389 using namespace support;
390 const unsigned char *End = D + N;
393 if (D + sizeof(uint64_t) >= End)
395 uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
397 // Initialize number of counters for FormatVersion == 1.
398 uint64_t CountsSize = N / sizeof(uint64_t) - 1;
399 // If format version is different then read the number of counters.
400 if (FormatVersion != 1) {
401 if (D + sizeof(uint64_t) > End)
403 CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
405 // Read counter values.
406 if (D + CountsSize * sizeof(uint64_t) > End)
409 CounterBuffer.clear();
410 CounterBuffer.reserve(CountsSize);
411 for (uint64_t J = 0; J < CountsSize; ++J)
412 CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
414 DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer));
416 // Read value profiling data.
417 if (FormatVersion > 2 && !readValueProfilingData(D, End)) {
425 template <typename HashTableImpl>
426 std::error_code InstrProfReaderIndex<HashTableImpl>::getRecords(
427 StringRef FuncName, ArrayRef<InstrProfRecord> &Data) {
428 auto Iter = HashTable->find(FuncName);
429 if (Iter == HashTable->end())
430 return instrprof_error::unknown_function;
434 return instrprof_error::malformed;
436 return instrprof_error::success;
439 template <typename HashTableImpl>
440 std::error_code InstrProfReaderIndex<HashTableImpl>::getRecords(
441 ArrayRef<InstrProfRecord> &Data) {
443 return instrprof_error::eof;
445 Data = *RecordIterator;
448 return instrprof_error::malformed;
450 return instrprof_error::success;
453 template <typename HashTableImpl>
454 InstrProfReaderIndex<HashTableImpl>::InstrProfReaderIndex(
455 const unsigned char *Buckets, const unsigned char *const Payload,
456 const unsigned char *const Base, IndexedInstrProf::HashT HashType,
458 FormatVersion = Version;
459 HashTable.reset(HashTableImpl::Create(
460 Buckets, Payload, Base,
461 typename HashTableImpl::InfoType(HashType, Version)));
462 // Form the map of hash values to const char* keys in profiling data.
463 std::vector<std::pair<uint64_t, const char *>> HashKeys;
464 for (auto Key : HashTable->keys()) {
465 const char *KeyTableRef = StringTable.insertString(Key);
466 HashKeys.push_back(std::make_pair(ComputeHash(HashType, Key), KeyTableRef));
468 std::sort(HashKeys.begin(), HashKeys.end(), less_first());
469 HashKeys.erase(std::unique(HashKeys.begin(), HashKeys.end()), HashKeys.end());
470 // Set the hash key map for the InstrLookupTrait
471 HashTable->getInfoObj().setHashKeys(std::move(HashKeys));
472 RecordIterator = HashTable->data_begin();
475 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
476 if (DataBuffer.getBufferSize() < 8)
478 using namespace support;
480 endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
481 // Verify that it's magical.
482 return Magic == IndexedInstrProf::Magic;
485 std::error_code IndexedInstrProfReader::readHeader() {
486 const unsigned char *Start =
487 (const unsigned char *)DataBuffer->getBufferStart();
488 const unsigned char *Cur = Start;
489 if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
490 return error(instrprof_error::truncated);
492 using namespace support;
494 auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
495 Cur += sizeof(IndexedInstrProf::Header);
497 // Check the magic number.
498 uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
499 if (Magic != IndexedInstrProf::Magic)
500 return error(instrprof_error::bad_magic);
503 uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
504 if (FormatVersion > IndexedInstrProf::Version)
505 return error(instrprof_error::unsupported_version);
507 // Read the maximal function count.
509 endian::byte_swap<uint64_t, little>(Header->MaxFunctionCount);
511 // Read the hash type and start offset.
512 IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
513 endian::byte_swap<uint64_t, little>(Header->HashType));
514 if (HashType > IndexedInstrProf::HashT::Last)
515 return error(instrprof_error::unsupported_hash_type);
517 uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
519 // The rest of the file is an on disk hash table.
520 InstrProfReaderIndexBase *IndexPtr = nullptr;
521 IndexPtr = new InstrProfReaderIndex<OnDiskHashTableImplV3>(
522 Start + HashOffset, Cur, Start, HashType, FormatVersion);
523 Index.reset(IndexPtr);
527 ErrorOr<InstrProfRecord>
528 IndexedInstrProfReader::getInstrProfRecord(StringRef FuncName,
530 ArrayRef<InstrProfRecord> Data;
531 std::error_code EC = Index->getRecords(FuncName, Data);
532 if (EC != instrprof_error::success)
534 // Found it. Look for counters with the right hash.
535 for (unsigned I = 0, E = Data.size(); I < E; ++I) {
536 // Check for a match and fill the vector if there is one.
537 if (Data[I].Hash == FuncHash) {
538 return std::move(Data[I]);
541 return error(instrprof_error::hash_mismatch);
545 IndexedInstrProfReader::getFunctionCounts(StringRef FuncName, uint64_t FuncHash,
546 std::vector<uint64_t> &Counts) {
547 ErrorOr<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash);
548 if (std::error_code EC = Record.getError())
551 Counts = Record.get().Counts;
555 std::error_code IndexedInstrProfReader::readNextRecord(
556 InstrProfRecord &Record) {
557 static unsigned RecordIndex = 0;
559 ArrayRef<InstrProfRecord> Data;
561 std::error_code EC = Index->getRecords(Data);
562 if (EC != instrprof_error::success)
565 Record = Data[RecordIndex++];
566 if (RecordIndex >= Data.size()) {
567 Index->advanceToNextKey();