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)));
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())
73 return IndexedInstrProfReader::create(std::move(BufferOrError.get()));
77 ErrorOr<std::unique_ptr<IndexedInstrProfReader>>
78 IndexedInstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer) {
79 // Sanity check the buffer.
80 if (Buffer->getBufferSize() > std::numeric_limits<unsigned>::max())
81 return instrprof_error::too_large;
84 if (!IndexedInstrProfReader::hasFormat(*Buffer))
85 return instrprof_error::bad_magic;
86 auto Result = llvm::make_unique<IndexedInstrProfReader>(std::move(Buffer));
88 // Initialize the reader and return the result.
89 if (std::error_code EC = initializeReader(*Result))
92 return std::move(Result);
95 void InstrProfIterator::Increment() {
96 if (Reader->readNextRecord(Record))
97 *this = InstrProfIterator();
100 std::error_code TextInstrProfReader::readNextRecord(InstrProfRecord &Record) {
101 // Skip empty lines and comments.
102 while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
104 // If we hit EOF while looking for a name, we're done.
105 if (Line.is_at_end())
106 return error(instrprof_error::eof);
108 // Read the function name.
109 Record.Name = *Line++;
111 // Read the function hash.
112 if (Line.is_at_end())
113 return error(instrprof_error::truncated);
114 if ((Line++)->getAsInteger(0, Record.Hash))
115 return error(instrprof_error::malformed);
117 // Read the number of counters.
118 uint64_t NumCounters;
119 if (Line.is_at_end())
120 return error(instrprof_error::truncated);
121 if ((Line++)->getAsInteger(10, NumCounters))
122 return error(instrprof_error::malformed);
123 if (NumCounters == 0)
124 return error(instrprof_error::malformed);
126 // Read each counter and fill our internal storage with the values.
127 Record.Counts.clear();
128 Record.Counts.reserve(NumCounters);
129 for (uint64_t I = 0; I < NumCounters; ++I) {
130 if (Line.is_at_end())
131 return error(instrprof_error::truncated);
133 if ((Line++)->getAsInteger(10, Count))
134 return error(instrprof_error::malformed);
135 Record.Counts.push_back(Count);
141 template <class IntPtrT>
142 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
143 if (DataBuffer.getBufferSize() < sizeof(uint64_t))
146 *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
147 return RawInstrProf::getMagic<IntPtrT>() == Magic ||
148 sys::getSwappedBytes(RawInstrProf::getMagic<IntPtrT>()) == Magic;
151 template <class IntPtrT>
152 std::error_code RawInstrProfReader<IntPtrT>::readHeader() {
153 if (!hasFormat(*DataBuffer))
154 return error(instrprof_error::bad_magic);
155 if (DataBuffer->getBufferSize() < sizeof(RawInstrProf::Header))
156 return error(instrprof_error::bad_header);
157 auto *Header = reinterpret_cast<const RawInstrProf::Header *>(
158 DataBuffer->getBufferStart());
159 ShouldSwapBytes = Header->Magic != RawInstrProf::getMagic<IntPtrT>();
160 return readHeader(*Header);
163 template <class IntPtrT>
165 RawInstrProfReader<IntPtrT>::readNextHeader(const char *CurrentPos) {
166 const char *End = DataBuffer->getBufferEnd();
167 // Skip zero padding between profiles.
168 while (CurrentPos != End && *CurrentPos == 0)
170 // If there's nothing left, we're done.
171 if (CurrentPos == End)
172 return instrprof_error::eof;
173 // If there isn't enough space for another header, this is probably just
174 // garbage at the end of the file.
175 if (CurrentPos + sizeof(RawInstrProf::Header) > End)
176 return instrprof_error::malformed;
177 // The writer ensures each profile is padded to start at an aligned address.
178 if (reinterpret_cast<size_t>(CurrentPos) % alignOf<uint64_t>())
179 return instrprof_error::malformed;
180 // The magic should have the same byte order as in the previous header.
181 uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos);
182 if (Magic != swap(RawInstrProf::getMagic<IntPtrT>()))
183 return instrprof_error::bad_magic;
185 // There's another profile to read, so we need to process the header.
186 auto *Header = reinterpret_cast<const RawInstrProf::Header *>(CurrentPos);
187 return readHeader(*Header);
190 template <class IntPtrT>
191 std::error_code RawInstrProfReader<IntPtrT>::readHeader(
192 const RawInstrProf::Header &Header) {
193 if (swap(Header.Version) != RawInstrProf::Version)
194 return error(instrprof_error::unsupported_version);
196 CountersDelta = swap(Header.CountersDelta);
197 NamesDelta = swap(Header.NamesDelta);
198 auto DataSize = swap(Header.DataSize);
199 auto CountersSize = swap(Header.CountersSize);
200 auto NamesSize = swap(Header.NamesSize);
202 ptrdiff_t DataOffset = sizeof(RawInstrProf::Header);
203 ptrdiff_t CountersOffset =
204 DataOffset + sizeof(RawInstrProf::ProfileData<IntPtrT>) * DataSize;
205 ptrdiff_t NamesOffset = CountersOffset + sizeof(uint64_t) * CountersSize;
206 size_t ProfileSize = NamesOffset + sizeof(char) * NamesSize;
208 auto *Start = reinterpret_cast<const char *>(&Header);
209 if (Start + ProfileSize > DataBuffer->getBufferEnd())
210 return error(instrprof_error::bad_header);
212 Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>(
214 DataEnd = Data + DataSize;
215 CountersStart = reinterpret_cast<const uint64_t *>(Start + CountersOffset);
216 NamesStart = Start + NamesOffset;
217 ProfileEnd = Start + ProfileSize;
222 template <class IntPtrT>
223 std::error_code RawInstrProfReader<IntPtrT>::readName(InstrProfRecord &Record) {
224 Record.Name = StringRef(getName(Data->NamePtr), swap(Data->NameSize));
225 if (Record.Name.data() < NamesStart ||
226 Record.Name.data() + Record.Name.size() > DataBuffer->getBufferEnd())
227 return error(instrprof_error::malformed);
232 template <class IntPtrT>
233 std::error_code RawInstrProfReader<IntPtrT>::readFuncHash(
234 InstrProfRecord &Record) {
235 Record.Hash = swap(Data->FuncHash);
239 template <class IntPtrT>
240 std::error_code RawInstrProfReader<IntPtrT>::readRawCounts(
241 InstrProfRecord &Record) {
242 uint32_t NumCounters = swap(Data->NumCounters);
243 IntPtrT CounterPtr = Data->CounterPtr;
244 if (NumCounters == 0)
245 return error(instrprof_error::malformed);
247 auto RawCounts = makeArrayRef(getCounter(CounterPtr), NumCounters);
248 auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
251 if (RawCounts.data() < CountersStart ||
252 RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
253 return error(instrprof_error::malformed);
255 if (ShouldSwapBytes) {
256 Record.Counts.clear();
257 Record.Counts.reserve(RawCounts.size());
258 for (uint64_t Count : RawCounts)
259 Record.Counts.push_back(swap(Count));
261 Record.Counts = RawCounts;
266 template <class IntPtrT>
267 std::error_code RawInstrProfReader<IntPtrT>::readNextRecord(
268 InstrProfRecord &Record) {
270 if (std::error_code EC = readNextHeader(ProfileEnd)) return EC;
272 // Read name ad set it in Record.
273 if (std::error_code EC = readName(Record)) return EC;
275 // Read FuncHash and set it in Record.
276 if (std::error_code EC = readFuncHash(Record)) return EC;
278 // Read raw counts and set Record.
279 if (std::error_code EC = readRawCounts(Record)) return EC;
287 template class RawInstrProfReader<uint32_t>;
288 template class RawInstrProfReader<uint64_t>;
291 InstrProfLookupTrait::hash_value_type
292 InstrProfLookupTrait::ComputeHash(StringRef K) {
293 return IndexedInstrProf::ComputeHash(HashType, K);
296 typedef InstrProfLookupTrait::data_type data_type;
297 typedef InstrProfLookupTrait::offset_type offset_type;
299 bool InstrProfLookupTrait::ReadValueProfilingData(
300 const unsigned char *&D, const unsigned char *const End) {
302 using namespace support;
303 // Read number of value kinds with value sites.
304 if (D + sizeof(uint64_t) > End)
306 uint64_t ValueKindCount = endian::readNext<uint64_t, little, unaligned>(D);
308 for (uint32_t Kind = 0; Kind < ValueKindCount; ++Kind) {
310 // Read value kind and number of value sites for kind.
311 if (D + 2 * sizeof(uint64_t) > End)
313 uint64_t ValueKind = endian::readNext<uint64_t, little, unaligned>(D);
314 uint64_t ValueSiteCount = endian::readNext<uint64_t, little, unaligned>(D);
316 std::vector<InstrProfValueSiteRecord> &ValueSites =
317 DataBuffer.back().getValueSitesForKind(ValueKind);
318 ValueSites.reserve(ValueSiteCount);
319 for (uint64_t VSite = 0; VSite < ValueSiteCount; ++VSite) {
320 // Read number of value data pairs at value site.
321 if (D + sizeof(uint64_t) > End)
323 uint64_t ValueDataCount =
324 endian::readNext<uint64_t, little, unaligned>(D);
326 // Check if there are as many ValueDataPairs as ValueDataCount in memory.
327 if (D + (ValueDataCount << 1) * sizeof(uint64_t) > End)
330 InstrProfValueSiteRecord VSiteRecord;
331 for (uint64_t VCount = 0; VCount < ValueDataCount; ++VCount) {
332 uint64_t Value = endian::readNext<uint64_t, little, unaligned>(D);
333 uint64_t NumTaken = endian::readNext<uint64_t, little, unaligned>(D);
335 case IPVK_IndirectCallTarget: {
337 std::lower_bound(HashKeys.begin(), HashKeys.end(), Value,
338 [](const std::pair<uint64_t, const char *> &LHS,
339 uint64_t RHS) { return LHS.first < RHS; });
340 assert(Result != HashKeys.end() &&
341 "Hash does not match any known keys\n");
342 Value = (uint64_t)Result->second;
346 VSiteRecord.ValueData.push_back(std::make_pair(Value, NumTaken));
348 ValueSites.push_back(std::move(VSiteRecord));
354 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
356 // Check if the data is corrupt. If so, don't try to read it.
357 if (N % sizeof(uint64_t))
361 std::vector<uint64_t> CounterBuffer;
363 using namespace support;
364 const unsigned char *End = D + N;
367 if (D + sizeof(uint64_t) >= End)
369 uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
371 // Initialize number of counters for FormatVersion == 1.
372 uint64_t CountsSize = N / sizeof(uint64_t) - 1;
373 // If format version is different then read the number of counters.
374 if (FormatVersion != 1) {
375 if (D + sizeof(uint64_t) > End)
377 CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
379 // Read counter values.
380 if (D + CountsSize * sizeof(uint64_t) > End)
383 CounterBuffer.clear();
384 CounterBuffer.reserve(CountsSize);
385 for (uint64_t J = 0; J < CountsSize; ++J)
386 CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
388 DataBuffer.push_back(InstrProfRecord(K, Hash, std::move(CounterBuffer)));
390 // Read value profiling data.
391 if (FormatVersion > 2 && !ReadValueProfilingData(D, End)) {
399 std::error_code InstrProfReaderIndex::getRecords(
400 StringRef FuncName, ArrayRef<InstrProfRecord> &Data) {
401 auto Iter = Index->find(FuncName);
402 if (Iter == Index->end()) return instrprof_error::unknown_function;
405 if (Data.empty()) return instrprof_error::malformed;
407 return instrprof_error::success;
410 std::error_code InstrProfReaderIndex::getRecords(
411 ArrayRef<InstrProfRecord> &Data) {
412 if (atEnd()) return instrprof_error::eof;
414 Data = *RecordIterator;
416 if (Data.empty()) return instrprof_error::malformed;
418 return instrprof_error::success;
421 void InstrProfReaderIndex::Init(const unsigned char *Buckets,
422 const unsigned char *const Payload,
423 const unsigned char *const Base,
424 IndexedInstrProf::HashT HashType,
426 FormatVersion = Version;
427 Index.reset(IndexType::Create(Buckets, Payload, Base,
428 InstrProfLookupTrait(HashType, Version)));
429 // Form the map of hash values to const char* keys in profiling data.
430 std::vector<std::pair<uint64_t, const char *>> HashKeys;
431 for (auto Key : Index->keys()) {
432 const char *KeyTableRef = StringTable.insertString(Key);
433 HashKeys.push_back(std::make_pair(ComputeHash(HashType, Key), KeyTableRef));
435 std::sort(HashKeys.begin(), HashKeys.end(), less_first());
436 HashKeys.erase(std::unique(HashKeys.begin(), HashKeys.end()), HashKeys.end());
437 // Set the hash key map for the InstrLookupTrait
438 Index->getInfoObj().setHashKeys(std::move(HashKeys));
439 RecordIterator = Index->data_begin();
442 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
443 if (DataBuffer.getBufferSize() < 8) return false;
444 using namespace support;
446 endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
447 // Verify that it's magical.
448 return Magic == IndexedInstrProf::Magic;
451 std::error_code IndexedInstrProfReader::readHeader() {
452 const unsigned char *Start =
453 (const unsigned char *)DataBuffer->getBufferStart();
454 const unsigned char *Cur = Start;
455 if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
456 return error(instrprof_error::truncated);
458 using namespace support;
460 auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
461 Cur += sizeof(IndexedInstrProf::Header);
463 // Check the magic number.
464 uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
465 if (Magic != IndexedInstrProf::Magic)
466 return error(instrprof_error::bad_magic);
469 uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
470 if (FormatVersion > IndexedInstrProf::Version)
471 return error(instrprof_error::unsupported_version);
473 // Read the maximal function count.
475 endian::byte_swap<uint64_t, little>(Header->MaxFunctionCount);
477 // Read the hash type and start offset.
478 IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
479 endian::byte_swap<uint64_t, little>(Header->HashType));
480 if (HashType > IndexedInstrProf::HashT::Last)
481 return error(instrprof_error::unsupported_hash_type);
483 uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
485 // The rest of the file is an on disk hash table.
486 Index.Init(Start + HashOffset, Cur, Start, HashType, FormatVersion);
491 std::error_code IndexedInstrProfReader::getFunctionCounts(
492 StringRef FuncName, uint64_t FuncHash, std::vector<uint64_t> &Counts) {
493 ArrayRef<InstrProfRecord> Data;
495 std::error_code EC = Index.getRecords(FuncName, Data);
496 if (EC != instrprof_error::success) return EC;
498 // Found it. Look for counters with the right hash.
499 for (unsigned I = 0, E = Data.size(); I < E; ++I) {
500 // Check for a match and fill the vector if there is one.
501 if (Data[I].Hash == FuncHash) {
502 Counts = Data[I].Counts;
506 return error(instrprof_error::hash_mismatch);
509 std::error_code IndexedInstrProfReader::readNextRecord(
510 InstrProfRecord &Record) {
511 static unsigned RecordIndex = 0;
513 ArrayRef<InstrProfRecord> Data;
515 std::error_code EC = Index.getRecords(Data);
516 if (EC != instrprof_error::success) return error(EC);
518 Record = Data[RecordIndex++];
519 if (RecordIndex >= Data.size()) {
520 Index.advanceToNextKey();