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 ValueDataDelta = swap(Header.ValueDataDelta);
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() > (char*)ValueDataStart)
259 return error(instrprof_error::malformed);
263 template <class IntPtrT>
264 std::error_code RawInstrProfReader<IntPtrT>::readFuncHash(
265 InstrProfRecord &Record) {
266 Record.Hash = swap(Data->FuncHash);
270 template <class IntPtrT>
271 std::error_code RawInstrProfReader<IntPtrT>::readRawCounts(
272 InstrProfRecord &Record) {
273 uint32_t NumCounters = swap(Data->NumCounters);
274 IntPtrT CounterPtr = Data->CounterPtr;
275 if (NumCounters == 0)
276 return error(instrprof_error::malformed);
278 auto RawCounts = makeArrayRef(getCounter(CounterPtr), NumCounters);
279 auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
282 if (RawCounts.data() < CountersStart ||
283 RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
284 return error(instrprof_error::malformed);
286 if (ShouldSwapBytes) {
287 Record.Counts.clear();
288 Record.Counts.reserve(RawCounts.size());
289 for (uint64_t Count : RawCounts)
290 Record.Counts.push_back(swap(Count));
292 Record.Counts = RawCounts;
297 template <class IntPtrT>
298 std::error_code RawInstrProfReader<IntPtrT>::readValueData(
299 InstrProfRecord &Record) {
301 Record.clearValueData();
302 if (!Data->Values || (ValueDataDelta == 0))
306 uint64_t NumVSites = 0;
307 for (uint32_t Kind = IPVK_First; Kind <= ValueKindLast; ++Kind)
308 NumVSites += swap(Data->NumValueSites[Kind]);
309 NumVSites += getNumPaddingBytes(NumVSites);
311 auto VDataCounts = makeArrayRef(getValueDataCounts(Data->Values), NumVSites);
313 if (VDataCounts.data() < ValueDataStart ||
314 VDataCounts.data() + VDataCounts.size() > (const uint8_t *)ProfileEnd)
315 return error(instrprof_error::malformed);
317 const InstrProfValueData *VDataPtr =
318 getValueData(swap(Data->Values) + NumVSites);
319 for (uint32_t Kind = IPVK_First; Kind <= ValueKindLast; ++Kind) {
320 NumVSites = swap(Data->NumValueSites[Kind]);
321 Record.reserveSites(Kind, NumVSites);
322 for (uint32_t VSite = 0; VSite < NumVSites; ++VSite) {
324 uint32_t VDataCount = VDataCounts[VSite];
325 if ((const char *)(VDataPtr + VDataCount) > ProfileEnd)
326 return error(instrprof_error::malformed);
328 std::vector<InstrProfValueData> CurrentValues;
329 CurrentValues.reserve(VDataCount);
330 for (uint32_t VIndex = 0; VIndex < VDataCount; ++VIndex) {
331 uint64_t TargetValue = swap(VDataPtr->Value);
332 uint64_t Count = swap(VDataPtr->Count);
333 CurrentValues.push_back({TargetValue, Count});
336 Record.addValueData(Kind, VSite, CurrentValues.data(),
337 VDataCount, &FunctionPtrToNameMap);
343 template <class IntPtrT>
344 std::error_code RawInstrProfReader<IntPtrT>::readNextRecord(
345 InstrProfRecord &Record) {
347 if (std::error_code EC = readNextHeader(ProfileEnd))
350 // Read name ad set it in Record.
351 if (std::error_code EC = readName(Record))
354 // Read FuncHash and set it in Record.
355 if (std::error_code EC = readFuncHash(Record))
358 // Read raw counts and set Record.
359 if (std::error_code EC = readRawCounts(Record))
362 // Read value data and set Record.
363 if (std::error_code EC = readValueData(Record)) return EC;
371 template class RawInstrProfReader<uint32_t>;
372 template class RawInstrProfReader<uint64_t>;
375 InstrProfLookupTrait::hash_value_type
376 InstrProfLookupTrait::ComputeHash(StringRef K) {
377 return IndexedInstrProf::ComputeHash(HashType, K);
380 typedef InstrProfLookupTrait::data_type data_type;
381 typedef InstrProfLookupTrait::offset_type offset_type;
383 bool InstrProfLookupTrait::ReadValueProfilingData(
384 const unsigned char *&D, const unsigned char *const End) {
385 ErrorOr<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
386 ValueProfData::getValueProfData(D, End, ValueProfDataEndianness);
388 if (VDataPtrOrErr.getError())
391 VDataPtrOrErr.get()->deserializeTo(DataBuffer.back(), &HashKeys);
392 D += VDataPtrOrErr.get()->TotalSize;
397 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
399 // Check if the data is corrupt. If so, don't try to read it.
400 if (N % sizeof(uint64_t))
404 std::vector<uint64_t> CounterBuffer;
406 using namespace support;
407 const unsigned char *End = D + N;
410 if (D + sizeof(uint64_t) >= End)
412 uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
414 // Initialize number of counters for FormatVersion == 1.
415 uint64_t CountsSize = N / sizeof(uint64_t) - 1;
416 // If format version is different then read the number of counters.
417 if (FormatVersion != 1) {
418 if (D + sizeof(uint64_t) > End)
420 CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
422 // Read counter values.
423 if (D + CountsSize * sizeof(uint64_t) > End)
426 CounterBuffer.clear();
427 CounterBuffer.reserve(CountsSize);
428 for (uint64_t J = 0; J < CountsSize; ++J)
429 CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
431 DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer));
433 // Read value profiling data.
434 if (FormatVersion > 2 && !ReadValueProfilingData(D, End)) {
443 InstrProfReaderIndex::getRecords(StringRef FuncName,
444 ArrayRef<InstrProfRecord> &Data) {
445 auto Iter = Index->find(FuncName);
446 if (Iter == Index->end())
447 return instrprof_error::unknown_function;
451 return instrprof_error::malformed;
453 return instrprof_error::success;
456 std::error_code InstrProfReaderIndex::getRecords(
457 ArrayRef<InstrProfRecord> &Data) {
458 if (atEnd()) return instrprof_error::eof;
460 Data = *RecordIterator;
462 if (Data.empty()) return instrprof_error::malformed;
464 return instrprof_error::success;
467 void InstrProfReaderIndex::Init(const unsigned char *Buckets,
468 const unsigned char *const Payload,
469 const unsigned char *const Base,
470 IndexedInstrProf::HashT HashType,
472 FormatVersion = Version;
473 Index.reset(IndexType::Create(Buckets, Payload, Base,
474 InstrProfLookupTrait(HashType, Version)));
475 // Form the map of hash values to const char* keys in profiling data.
476 std::vector<std::pair<uint64_t, const char *>> HashKeys;
477 for (auto Key : Index->keys()) {
478 const char *KeyTableRef = StringTable.insertString(Key);
479 HashKeys.push_back(std::make_pair(ComputeHash(HashType, Key), KeyTableRef));
481 std::sort(HashKeys.begin(), HashKeys.end(), less_first());
482 HashKeys.erase(std::unique(HashKeys.begin(), HashKeys.end()), HashKeys.end());
483 // Set the hash key map for the InstrLookupTrait
484 Index->getInfoObj().setHashKeys(std::move(HashKeys));
485 RecordIterator = Index->data_begin();
488 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
489 if (DataBuffer.getBufferSize() < 8)
491 using namespace support;
493 endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
494 // Verify that it's magical.
495 return Magic == IndexedInstrProf::Magic;
498 std::error_code IndexedInstrProfReader::readHeader() {
499 const unsigned char *Start =
500 (const unsigned char *)DataBuffer->getBufferStart();
501 const unsigned char *Cur = Start;
502 if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
503 return error(instrprof_error::truncated);
505 using namespace support;
507 auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
508 Cur += sizeof(IndexedInstrProf::Header);
510 // Check the magic number.
511 uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
512 if (Magic != IndexedInstrProf::Magic)
513 return error(instrprof_error::bad_magic);
516 uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
517 if (FormatVersion > IndexedInstrProf::Version)
518 return error(instrprof_error::unsupported_version);
520 // Read the maximal function count.
522 endian::byte_swap<uint64_t, little>(Header->MaxFunctionCount);
524 // Read the hash type and start offset.
525 IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
526 endian::byte_swap<uint64_t, little>(Header->HashType));
527 if (HashType > IndexedInstrProf::HashT::Last)
528 return error(instrprof_error::unsupported_hash_type);
530 uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
532 // The rest of the file is an on disk hash table.
533 Index.Init(Start + HashOffset, Cur, Start, HashType, FormatVersion);
538 ErrorOr<InstrProfRecord>
539 IndexedInstrProfReader::getInstrProfRecord(StringRef FuncName,
541 ArrayRef<InstrProfRecord> Data;
542 std::error_code EC = Index.getRecords(FuncName, Data);
543 if (EC != instrprof_error::success)
545 // Found it. Look for counters with the right hash.
546 for (unsigned I = 0, E = Data.size(); I < E; ++I) {
547 // Check for a match and fill the vector if there is one.
548 if (Data[I].Hash == FuncHash) {
549 return std::move(Data[I]);
552 return error(instrprof_error::hash_mismatch);
556 IndexedInstrProfReader::getFunctionCounts(StringRef FuncName, uint64_t FuncHash,
557 std::vector<uint64_t> &Counts) {
558 ErrorOr<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash);
559 if (std::error_code EC = Record.getError())
562 Counts = Record.get().Counts;
566 std::error_code IndexedInstrProfReader::readNextRecord(
567 InstrProfRecord &Record) {
568 static unsigned RecordIndex = 0;
570 ArrayRef<InstrProfRecord> Data;
572 std::error_code EC = Index.getRecords(Data);
573 if (EC != instrprof_error::success)
576 Record = Data[RecordIndex++];
577 if (RecordIndex >= Data.size()) {
578 Index.advanceToNextKey();