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 RawInstrProfReader<IntPtrT>::readValueData(
300     InstrProfRecord &Record) {
301 
302   Record.clearValueData();
303   if (!Data->Values || (ValueDataDelta == 0))
304     return success();
305 
306   // Read value data.
307   uint64_t NumVSites = 0;
308   for (uint32_t Kind = IPVK_First; Kind <= ValueKindLast; ++Kind)
309     NumVSites += swap(Data->NumValueSites[Kind]);
310   NumVSites += getNumPaddingBytes(NumVSites);
311 
312   auto VDataCounts = makeArrayRef(getValueDataCounts(Data->Values), NumVSites);
313   // Check bounds.
314   if (VDataCounts.data() < ValueDataStart ||
315       VDataCounts.data() + VDataCounts.size() >
316           reinterpret_cast<const uint8_t *>(ProfileEnd))
317     return error(instrprof_error::malformed);
318 
319   const InstrProfValueData *VDataPtr =
320       getValueData(swap(Data->Values) + NumVSites);
321   for (uint32_t Kind = IPVK_First; Kind <= ValueKindLast; ++Kind) {
322     NumVSites = swap(Data->NumValueSites[Kind]);
323     Record.reserveSites(Kind, NumVSites);
324     for (uint32_t VSite = 0; VSite < NumVSites; ++VSite) {
325 
326       uint32_t VDataCount = VDataCounts[VSite];
327       if ((const char *)(VDataPtr + VDataCount) > ProfileEnd)
328         return error(instrprof_error::malformed);
329 
330       std::vector<InstrProfValueData> CurrentValues;
331       CurrentValues.reserve(VDataCount);
332       for (uint32_t VIndex = 0; VIndex < VDataCount; ++VIndex) {
333         uint64_t TargetValue = swap(VDataPtr->Value);
334         uint64_t Count = swap(VDataPtr->Count);
335         CurrentValues.push_back({TargetValue, Count});
336         ++VDataPtr;
337       }
338       Record.addValueData(Kind, VSite, CurrentValues.data(),
339                           VDataCount, &FunctionPtrToNameMap);
340     }
341   }
342   return success();
343 }
344 
345 template <class IntPtrT>
346 std::error_code RawInstrProfReader<IntPtrT>::readNextRecord(
347     InstrProfRecord &Record) {
348   if (atEnd())
349     if (std::error_code EC = readNextHeader(ProfileEnd))
350       return EC;
351 
352   // Read name ad set it in Record.
353   if (std::error_code EC = readName(Record))
354     return EC;
355 
356   // Read FuncHash and set it in Record.
357   if (std::error_code EC = readFuncHash(Record))
358     return EC;
359 
360   // Read raw counts and set Record.
361   if (std::error_code EC = readRawCounts(Record))
362     return EC;
363 
364   // Read value data and set Record.
365   if (std::error_code EC = readValueData(Record)) return EC;
366 
367   // Iterate.
368   advanceData();
369   return success();
370 }
371 
372 namespace llvm {
373 template class RawInstrProfReader<uint32_t>;
374 template class RawInstrProfReader<uint64_t>;
375 }
376 
377 InstrProfLookupTrait::hash_value_type
378 InstrProfLookupTrait::ComputeHash(StringRef K) {
379   return IndexedInstrProf::ComputeHash(HashType, K);
380 }
381 
382 typedef InstrProfLookupTrait::data_type data_type;
383 typedef InstrProfLookupTrait::offset_type offset_type;
384 
385 bool InstrProfLookupTrait::ReadValueProfilingData(
386     const unsigned char *&D, const unsigned char *const End) {
387   ErrorOr<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
388       ValueProfData::getValueProfData(D, End, ValueProfDataEndianness);
389 
390   if (VDataPtrOrErr.getError())
391     return false;
392 
393   VDataPtrOrErr.get()->deserializeTo(DataBuffer.back(), &HashKeys);
394   D += VDataPtrOrErr.get()->TotalSize;
395 
396   return true;
397 }
398 
399 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
400                                          offset_type N) {
401   // Check if the data is corrupt. If so, don't try to read it.
402   if (N % sizeof(uint64_t))
403     return data_type();
404 
405   DataBuffer.clear();
406   std::vector<uint64_t> CounterBuffer;
407 
408   using namespace support;
409   const unsigned char *End = D + N;
410   while (D < End) {
411     // Read hash.
412     if (D + sizeof(uint64_t) >= End)
413       return data_type();
414     uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
415 
416     // Initialize number of counters for FormatVersion == 1.
417     uint64_t CountsSize = N / sizeof(uint64_t) - 1;
418     // If format version is different then read the number of counters.
419     if (FormatVersion != 1) {
420       if (D + sizeof(uint64_t) > End)
421         return data_type();
422       CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
423     }
424     // Read counter values.
425     if (D + CountsSize * sizeof(uint64_t) > End)
426       return data_type();
427 
428     CounterBuffer.clear();
429     CounterBuffer.reserve(CountsSize);
430     for (uint64_t J = 0; J < CountsSize; ++J)
431       CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
432 
433     DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer));
434 
435     // Read value profiling data.
436     if (FormatVersion > 2 && !ReadValueProfilingData(D, End)) {
437       DataBuffer.clear();
438       return data_type();
439     }
440   }
441   return DataBuffer;
442 }
443 
444 std::error_code
445 InstrProfReaderIndex::getRecords(StringRef FuncName,
446                                  ArrayRef<InstrProfRecord> &Data) {
447   auto Iter = Index->find(FuncName);
448   if (Iter == Index->end())
449     return instrprof_error::unknown_function;
450 
451   Data = (*Iter);
452   if (Data.empty())
453     return instrprof_error::malformed;
454 
455   return instrprof_error::success;
456 }
457 
458 std::error_code InstrProfReaderIndex::getRecords(
459     ArrayRef<InstrProfRecord> &Data) {
460   if (atEnd()) return instrprof_error::eof;
461 
462   Data = *RecordIterator;
463 
464   if (Data.empty()) return instrprof_error::malformed;
465 
466   return instrprof_error::success;
467 }
468 
469 void InstrProfReaderIndex::Init(const unsigned char *Buckets,
470                                 const unsigned char *const Payload,
471                                 const unsigned char *const Base,
472                                 IndexedInstrProf::HashT HashType,
473                                 uint64_t Version) {
474   FormatVersion = Version;
475   Index.reset(IndexType::Create(Buckets, Payload, Base,
476                                 InstrProfLookupTrait(HashType, Version)));
477   // Form the map of hash values to const char* keys in profiling data.
478   std::vector<std::pair<uint64_t, const char *>> HashKeys;
479   for (auto Key : Index->keys()) {
480     const char *KeyTableRef = StringTable.insertString(Key);
481     HashKeys.push_back(std::make_pair(ComputeHash(HashType, Key), KeyTableRef));
482   }
483   std::sort(HashKeys.begin(), HashKeys.end(), less_first());
484   HashKeys.erase(std::unique(HashKeys.begin(), HashKeys.end()), HashKeys.end());
485   // Set the hash key map for the InstrLookupTrait
486   Index->getInfoObj().setHashKeys(std::move(HashKeys));
487   RecordIterator = Index->data_begin();
488 }
489 
490 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
491   if (DataBuffer.getBufferSize() < 8)
492     return false;
493   using namespace support;
494   uint64_t Magic =
495       endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
496   // Verify that it's magical.
497   return Magic == IndexedInstrProf::Magic;
498 }
499 
500 std::error_code IndexedInstrProfReader::readHeader() {
501   const unsigned char *Start =
502       (const unsigned char *)DataBuffer->getBufferStart();
503   const unsigned char *Cur = Start;
504   if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
505     return error(instrprof_error::truncated);
506 
507   using namespace support;
508 
509   auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
510   Cur += sizeof(IndexedInstrProf::Header);
511 
512   // Check the magic number.
513   uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
514   if (Magic != IndexedInstrProf::Magic)
515     return error(instrprof_error::bad_magic);
516 
517   // Read the version.
518   uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
519   if (FormatVersion > IndexedInstrProf::Version)
520     return error(instrprof_error::unsupported_version);
521 
522   // Read the maximal function count.
523   MaxFunctionCount =
524       endian::byte_swap<uint64_t, little>(Header->MaxFunctionCount);
525 
526   // Read the hash type and start offset.
527   IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
528       endian::byte_swap<uint64_t, little>(Header->HashType));
529   if (HashType > IndexedInstrProf::HashT::Last)
530     return error(instrprof_error::unsupported_hash_type);
531 
532   uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
533 
534   // The rest of the file is an on disk hash table.
535   Index.Init(Start + HashOffset, Cur, Start, HashType, FormatVersion);
536 
537   return success();
538 }
539 
540 ErrorOr<InstrProfRecord>
541 IndexedInstrProfReader::getInstrProfRecord(StringRef FuncName,
542                                            uint64_t FuncHash) {
543   ArrayRef<InstrProfRecord> Data;
544   std::error_code EC = Index.getRecords(FuncName, Data);
545   if (EC != instrprof_error::success)
546     return EC;
547   // Found it. Look for counters with the right hash.
548   for (unsigned I = 0, E = Data.size(); I < E; ++I) {
549     // Check for a match and fill the vector if there is one.
550     if (Data[I].Hash == FuncHash) {
551       return std::move(Data[I]);
552     }
553   }
554   return error(instrprof_error::hash_mismatch);
555 }
556 
557 std::error_code
558 IndexedInstrProfReader::getFunctionCounts(StringRef FuncName, uint64_t FuncHash,
559                                           std::vector<uint64_t> &Counts) {
560   ErrorOr<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash);
561   if (std::error_code EC = Record.getError())
562     return EC;
563 
564   Counts = Record.get().Counts;
565   return success();
566 }
567 
568 std::error_code IndexedInstrProfReader::readNextRecord(
569     InstrProfRecord &Record) {
570   static unsigned RecordIndex = 0;
571 
572   ArrayRef<InstrProfRecord> Data;
573 
574   std::error_code EC = Index.getRecords(Data);
575   if (EC != instrprof_error::success)
576     return error(EC);
577 
578   Record = Data[RecordIndex++];
579   if (RecordIndex >= Data.size()) {
580     Index.advanceToNextKey();
581     RecordIndex = 0;
582   }
583   return success();
584 }
585