1 //=-- InstrProfReader.cpp - Instrumented profiling reader -------------------=//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains support for reading profiling data for clang's
11 // instrumentation based PGO and coverage.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/ProfileData/InstrProfReader.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include <cassert>
18 
19 using namespace llvm;
20 
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())
26     return EC;
27   return std::move(BufferOrErr.get());
28 }
29 
30 static std::error_code initializeReader(InstrProfReader &Reader) {
31   return Reader.readHeader();
32 }
33 
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())
39     return EC;
40   return InstrProfReader::create(std::move(BufferOrError.get()));
41 }
42 
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;
48 
49   std::unique_ptr<InstrProfReader> Result;
50   // Create the reader.
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)));
59   else
60     return instrprof_error::unrecognized_format;
61 
62   // Initialize the reader and return the result.
63   if (std::error_code EC = initializeReader(*Result))
64     return EC;
65 
66   return std::move(Result);
67 }
68 
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())
74     return EC;
75   return IndexedInstrProfReader::create(std::move(BufferOrError.get()));
76 }
77 
78 
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;
84 
85   // Create the reader.
86   if (!IndexedInstrProfReader::hasFormat(*Buffer))
87     return instrprof_error::bad_magic;
88   auto Result = llvm::make_unique<IndexedInstrProfReader>(std::move(Buffer));
89 
90   // Initialize the reader and return the result.
91   if (std::error_code EC = initializeReader(*Result))
92     return EC;
93 
94   return std::move(Result);
95 }
96 
97 void InstrProfIterator::Increment() {
98   if (Reader->readNextRecord(Record))
99     *this = InstrProfIterator();
100 }
101 
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); });
109 }
110 
111 std::error_code TextInstrProfReader::readNextRecord(InstrProfRecord &Record) {
112   // Skip empty lines and comments.
113   while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
114     ++Line;
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);
118 
119   // Read the function name.
120   Record.Name = *Line++;
121 
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);
127 
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);
136 
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);
143     uint64_t Count;
144     if ((Line++)->getAsInteger(10, Count))
145       return error(instrprof_error::malformed);
146     Record.Counts.push_back(Count);
147   }
148 
149   return success();
150 }
151 
152 template <class IntPtrT>
153 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
154   if (DataBuffer.getBufferSize() < sizeof(uint64_t))
155     return false;
156   uint64_t Magic =
157     *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
158   return RawInstrProf::getMagic<IntPtrT>() == Magic ||
159          sys::getSwappedBytes(RawInstrProf::getMagic<IntPtrT>()) == Magic;
160 }
161 
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);
172 }
173 
174 template <class IntPtrT>
175 std::error_code
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)
180     ++CurrentPos;
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;
195 
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);
199 }
200 
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);
206 
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);
215 
216   auto DataSizeInBytes = DataSize * sizeof(RawInstrProf::ProfileData<IntPtrT>);
217   auto PaddingSize = getNumPaddingBytes(NamesSize);
218 
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;
224 
225   auto *Start = reinterpret_cast<const char *>(&Header);
226   if (Start + ProfileSize > DataBuffer->getBufferEnd())
227     return error(instrprof_error::bad_header);
228 
229   Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>(
230       Start + DataOffset);
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;
236 
237   FunctionPtrToNameMap.clear();
238   for (const RawInstrProf::ProfileData<IntPtrT> *I = Data; I != DataEnd; ++I) {
239     const IntPtrT FPtr = swap(I->FunctionPointer);
240     if (!FPtr)
241       continue;
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));
246   }
247   std::sort(FunctionPtrToNameMap.begin(), FunctionPtrToNameMap.end(), less_first());
248   FunctionPtrToNameMap.erase(std::unique(FunctionPtrToNameMap.begin(),
249                                          FunctionPtrToNameMap.end()),
250                                          FunctionPtrToNameMap.end());
251   return success();
252 }
253 
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);
261   return success();
262 }
263 
264 template <class IntPtrT>
265 std::error_code RawInstrProfReader<IntPtrT>::readFuncHash(
266     InstrProfRecord &Record) {
267   Record.Hash = swap(Data->FuncHash);
268   return success();
269 }
270 
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);
278 
279   auto RawCounts = makeArrayRef(getCounter(CounterPtr), NumCounters);
280   auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
281 
282   // Check bounds.
283   if (RawCounts.data() < CountersStart ||
284       RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
285     return error(instrprof_error::malformed);
286 
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));
292   } else
293     Record.Counts = RawCounts;
294 
295   return success();
296 }
297 
298 template <class IntPtrT>
299 std::error_code
300 RawInstrProfReader<IntPtrT>::readValueProfilingData(InstrProfRecord &Record) {
301 
302   Record.clearValueData();
303   if (!Data->Values || (ValueDataDelta == 0))
304     return success();
305 
306   ErrorOr<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
307       ValueProfData::getValueProfData(getValueDataCounts(Data->Values),
308                                       (const unsigned char *)ProfileEnd,
309                                       getDataEndianness());
310 
311   if (VDataPtrOrErr.getError())
312     return VDataPtrOrErr.getError();
313 
314   VDataPtrOrErr.get()->deserializeTo(Record, &FunctionPtrToNameMap);
315   return success();
316 }
317 
318 template <class IntPtrT>
319 std::error_code
320 RawInstrProfReader<IntPtrT>::readNextRecord(InstrProfRecord &Record) {
321   if (atEnd())
322     if (std::error_code EC = readNextHeader(ProfileEnd))
323       return EC;
324 
325   // Read name ad set it in Record.
326   if (std::error_code EC = readName(Record))
327     return EC;
328 
329   // Read FuncHash and set it in Record.
330   if (std::error_code EC = readFuncHash(Record))
331     return EC;
332 
333   // Read raw counts and set Record.
334   if (std::error_code EC = readRawCounts(Record))
335     return EC;
336 
337   // Read value data and set Record.
338   if (std::error_code EC = readValueProfilingData(Record)) return EC;
339 
340   // Iterate.
341   advanceData();
342   return success();
343 }
344 
345 namespace llvm {
346 template class RawInstrProfReader<uint32_t>;
347 template class RawInstrProfReader<uint64_t>;
348 }
349 
350 InstrProfLookupTrait::hash_value_type
351 InstrProfLookupTrait::ComputeHash(StringRef K) {
352   return IndexedInstrProf::ComputeHash(HashType, K);
353 }
354 
355 typedef InstrProfLookupTrait::data_type data_type;
356 typedef InstrProfLookupTrait::offset_type offset_type;
357 
358 bool InstrProfLookupTrait::readValueProfilingData(
359     const unsigned char *&D, const unsigned char *const End) {
360   ErrorOr<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
361       ValueProfData::getValueProfData(D, End, ValueProfDataEndianness);
362 
363   if (VDataPtrOrErr.getError())
364     return false;
365 
366   VDataPtrOrErr.get()->deserializeTo(DataBuffer.back(), &HashKeys);
367   D += VDataPtrOrErr.get()->TotalSize;
368 
369   return true;
370 }
371 
372 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
373                                          offset_type N) {
374   // Check if the data is corrupt. If so, don't try to read it.
375   if (N % sizeof(uint64_t))
376     return data_type();
377 
378   DataBuffer.clear();
379   std::vector<uint64_t> CounterBuffer;
380 
381   using namespace support;
382   const unsigned char *End = D + N;
383   while (D < End) {
384     // Read hash.
385     if (D + sizeof(uint64_t) >= End)
386       return data_type();
387     uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
388 
389     // Initialize number of counters for FormatVersion == 1.
390     uint64_t CountsSize = N / sizeof(uint64_t) - 1;
391     // If format version is different then read the number of counters.
392     if (FormatVersion != 1) {
393       if (D + sizeof(uint64_t) > End)
394         return data_type();
395       CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
396     }
397     // Read counter values.
398     if (D + CountsSize * sizeof(uint64_t) > End)
399       return data_type();
400 
401     CounterBuffer.clear();
402     CounterBuffer.reserve(CountsSize);
403     for (uint64_t J = 0; J < CountsSize; ++J)
404       CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
405 
406     DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer));
407 
408     // Read value profiling data.
409     if (FormatVersion > 2 && !readValueProfilingData(D, End)) {
410       DataBuffer.clear();
411       return data_type();
412     }
413   }
414   return DataBuffer;
415 }
416 
417 template <typename HashTableImpl>
418 std::error_code InstrProfReaderIndex<HashTableImpl>::getRecords(
419     StringRef FuncName, ArrayRef<InstrProfRecord> &Data) {
420   auto Iter = HashTable->find(FuncName);
421   if (Iter == HashTable->end())
422     return instrprof_error::unknown_function;
423 
424   Data = (*Iter);
425   if (Data.empty())
426     return instrprof_error::malformed;
427 
428   return instrprof_error::success;
429 }
430 
431 template <typename HashTableImpl>
432 std::error_code InstrProfReaderIndex<HashTableImpl>::getRecords(
433     ArrayRef<InstrProfRecord> &Data) {
434   if (atEnd())
435     return instrprof_error::eof;
436 
437   Data = *RecordIterator;
438 
439   if (Data.empty()) return instrprof_error::malformed;
440 
441   return instrprof_error::success;
442 }
443 
444 template <typename HashTableImpl>
445 InstrProfReaderIndex<HashTableImpl>::InstrProfReaderIndex(
446     const unsigned char *Buckets, const unsigned char *const Payload,
447     const unsigned char *const Base, IndexedInstrProf::HashT HashType,
448     uint64_t Version) {
449   FormatVersion = Version;
450   HashTable.reset(HashTableImpl::Create(
451       Buckets, Payload, Base,
452       typename HashTableImpl::InfoType(HashType, Version)));
453   // Form the map of hash values to const char* keys in profiling data.
454   std::vector<std::pair<uint64_t, const char *>> HashKeys;
455   for (auto Key : HashTable->keys()) {
456     const char *KeyTableRef = StringTable.insertString(Key);
457     HashKeys.push_back(std::make_pair(ComputeHash(HashType, Key), KeyTableRef));
458   }
459   std::sort(HashKeys.begin(), HashKeys.end(), less_first());
460   HashKeys.erase(std::unique(HashKeys.begin(), HashKeys.end()), HashKeys.end());
461   // Set the hash key map for the InstrLookupTrait
462   HashTable->getInfoObj().setHashKeys(std::move(HashKeys));
463   RecordIterator = HashTable->data_begin();
464 }
465 
466 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
467   if (DataBuffer.getBufferSize() < 8)
468     return false;
469   using namespace support;
470   uint64_t Magic =
471       endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
472   // Verify that it's magical.
473   return Magic == IndexedInstrProf::Magic;
474 }
475 
476 std::error_code IndexedInstrProfReader::readHeader() {
477   const unsigned char *Start =
478       (const unsigned char *)DataBuffer->getBufferStart();
479   const unsigned char *Cur = Start;
480   if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
481     return error(instrprof_error::truncated);
482 
483   using namespace support;
484 
485   auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
486   Cur += sizeof(IndexedInstrProf::Header);
487 
488   // Check the magic number.
489   uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
490   if (Magic != IndexedInstrProf::Magic)
491     return error(instrprof_error::bad_magic);
492 
493   // Read the version.
494   uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
495   if (FormatVersion > IndexedInstrProf::Version)
496     return error(instrprof_error::unsupported_version);
497 
498   // Read the maximal function count.
499   MaxFunctionCount =
500       endian::byte_swap<uint64_t, little>(Header->MaxFunctionCount);
501 
502   // Read the hash type and start offset.
503   IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
504       endian::byte_swap<uint64_t, little>(Header->HashType));
505   if (HashType > IndexedInstrProf::HashT::Last)
506     return error(instrprof_error::unsupported_hash_type);
507 
508   uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
509 
510   // The rest of the file is an on disk hash table.
511   InstrProfReaderIndexBase *IndexPtr = nullptr;
512   IndexPtr = new InstrProfReaderIndex<OnDiskHashTableImplV3>(
513       Start + HashOffset, Cur, Start, HashType, FormatVersion);
514   Index.reset(IndexPtr);
515   return success();
516 }
517 
518 ErrorOr<InstrProfRecord>
519 IndexedInstrProfReader::getInstrProfRecord(StringRef FuncName,
520                                            uint64_t FuncHash) {
521   ArrayRef<InstrProfRecord> Data;
522   std::error_code EC = Index->getRecords(FuncName, Data);
523   if (EC != instrprof_error::success)
524     return EC;
525   // Found it. Look for counters with the right hash.
526   for (unsigned I = 0, E = Data.size(); I < E; ++I) {
527     // Check for a match and fill the vector if there is one.
528     if (Data[I].Hash == FuncHash) {
529       return std::move(Data[I]);
530     }
531   }
532   return error(instrprof_error::hash_mismatch);
533 }
534 
535 std::error_code
536 IndexedInstrProfReader::getFunctionCounts(StringRef FuncName, uint64_t FuncHash,
537                                           std::vector<uint64_t> &Counts) {
538   ErrorOr<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash);
539   if (std::error_code EC = Record.getError())
540     return EC;
541 
542   Counts = Record.get().Counts;
543   return success();
544 }
545 
546 std::error_code IndexedInstrProfReader::readNextRecord(
547     InstrProfRecord &Record) {
548   static unsigned RecordIndex = 0;
549 
550   ArrayRef<InstrProfRecord> Data;
551 
552   std::error_code EC = Index->getRecords(Data);
553   if (EC != instrprof_error::success)
554     return error(EC);
555 
556   Record = Data[RecordIndex++];
557   if (RecordIndex >= Data.size()) {
558     Index->advanceToNextKey();
559     RecordIndex = 0;
560   }
561   return success();
562 }
563