• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 Expected<std::unique_ptr<MemoryBuffer>>
setupMemoryBuffer(const Twine & Path)22 setupMemoryBuffer(const Twine &Path) {
23   ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
24       MemoryBuffer::getFileOrSTDIN(Path);
25   if (std::error_code EC = BufferOrErr.getError())
26     return errorCodeToError(EC);
27   return std::move(BufferOrErr.get());
28 }
29 
initializeReader(InstrProfReader & Reader)30 static Error initializeReader(InstrProfReader &Reader) {
31   return Reader.readHeader();
32 }
33 
34 Expected<std::unique_ptr<InstrProfReader>>
create(const Twine & Path)35 InstrProfReader::create(const Twine &Path) {
36   // Set up the buffer to read.
37   auto BufferOrError = setupMemoryBuffer(Path);
38   if (Error E = BufferOrError.takeError())
39     return std::move(E);
40   return InstrProfReader::create(std::move(BufferOrError.get()));
41 }
42 
43 Expected<std::unique_ptr<InstrProfReader>>
create(std::unique_ptr<MemoryBuffer> Buffer)44 InstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer) {
45   // Sanity check the buffer.
46   if (Buffer->getBufferSize() > std::numeric_limits<unsigned>::max())
47     return make_error<InstrProfError>(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 make_error<InstrProfError>(instrprof_error::unrecognized_format);
61 
62   // Initialize the reader and return the result.
63   if (Error E = initializeReader(*Result))
64     return std::move(E);
65 
66   return std::move(Result);
67 }
68 
69 Expected<std::unique_ptr<IndexedInstrProfReader>>
create(const Twine & Path)70 IndexedInstrProfReader::create(const Twine &Path) {
71   // Set up the buffer to read.
72   auto BufferOrError = setupMemoryBuffer(Path);
73   if (Error E = BufferOrError.takeError())
74     return std::move(E);
75   return IndexedInstrProfReader::create(std::move(BufferOrError.get()));
76 }
77 
78 
79 Expected<std::unique_ptr<IndexedInstrProfReader>>
create(std::unique_ptr<MemoryBuffer> Buffer)80 IndexedInstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer) {
81   // Sanity check the buffer.
82   if (Buffer->getBufferSize() > std::numeric_limits<unsigned>::max())
83     return make_error<InstrProfError>(instrprof_error::too_large);
84 
85   // Create the reader.
86   if (!IndexedInstrProfReader::hasFormat(*Buffer))
87     return make_error<InstrProfError>(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 (Error E = initializeReader(*Result))
92     return std::move(E);
93 
94   return std::move(Result);
95 }
96 
Increment()97 void InstrProfIterator::Increment() {
98   if (auto E = Reader->readNextRecord(Record)) {
99     // Handle errors in the reader.
100     InstrProfError::take(std::move(E));
101     *this = InstrProfIterator();
102   }
103 }
104 
hasFormat(const MemoryBuffer & Buffer)105 bool TextInstrProfReader::hasFormat(const MemoryBuffer &Buffer) {
106   // Verify that this really looks like plain ASCII text by checking a
107   // 'reasonable' number of characters (up to profile magic size).
108   size_t count = std::min(Buffer.getBufferSize(), sizeof(uint64_t));
109   StringRef buffer = Buffer.getBufferStart();
110   return count == 0 ||
111          std::all_of(buffer.begin(), buffer.begin() + count,
112                      [](char c) { return ::isprint(c) || ::isspace(c); });
113 }
114 
115 // Read the profile variant flag from the header: ":FE" means this is a FE
116 // generated profile. ":IR" means this is an IR level profile. Other strings
117 // with a leading ':' will be reported an error format.
readHeader()118 Error TextInstrProfReader::readHeader() {
119   Symtab.reset(new InstrProfSymtab());
120   bool IsIRInstr = false;
121   if (!Line->startswith(":")) {
122     IsIRLevelProfile = false;
123     return success();
124   }
125   StringRef Str = (Line)->substr(1);
126   if (Str.equals_lower("ir"))
127     IsIRInstr = true;
128   else if (Str.equals_lower("fe"))
129     IsIRInstr = false;
130   else
131     return error(instrprof_error::bad_header);
132 
133   ++Line;
134   IsIRLevelProfile = IsIRInstr;
135   return success();
136 }
137 
138 Error
readValueProfileData(InstrProfRecord & Record)139 TextInstrProfReader::readValueProfileData(InstrProfRecord &Record) {
140 
141 #define CHECK_LINE_END(Line)                                                   \
142   if (Line.is_at_end())                                                        \
143     return error(instrprof_error::truncated);
144 #define READ_NUM(Str, Dst)                                                     \
145   if ((Str).getAsInteger(10, (Dst)))                                           \
146     return error(instrprof_error::malformed);
147 #define VP_READ_ADVANCE(Val)                                                   \
148   CHECK_LINE_END(Line);                                                        \
149   uint32_t Val;                                                                \
150   READ_NUM((*Line), (Val));                                                    \
151   Line++;
152 
153   if (Line.is_at_end())
154     return success();
155 
156   uint32_t NumValueKinds;
157   if (Line->getAsInteger(10, NumValueKinds)) {
158     // No value profile data
159     return success();
160   }
161   if (NumValueKinds == 0 || NumValueKinds > IPVK_Last + 1)
162     return error(instrprof_error::malformed);
163   Line++;
164 
165   for (uint32_t VK = 0; VK < NumValueKinds; VK++) {
166     VP_READ_ADVANCE(ValueKind);
167     if (ValueKind > IPVK_Last)
168       return error(instrprof_error::malformed);
169     VP_READ_ADVANCE(NumValueSites);
170     if (!NumValueSites)
171       continue;
172 
173     Record.reserveSites(VK, NumValueSites);
174     for (uint32_t S = 0; S < NumValueSites; S++) {
175       VP_READ_ADVANCE(NumValueData);
176 
177       std::vector<InstrProfValueData> CurrentValues;
178       for (uint32_t V = 0; V < NumValueData; V++) {
179         CHECK_LINE_END(Line);
180         std::pair<StringRef, StringRef> VD = Line->rsplit(':');
181         uint64_t TakenCount, Value;
182         if (VK == IPVK_IndirectCallTarget) {
183           Symtab->addFuncName(VD.first);
184           Value = IndexedInstrProf::ComputeHash(VD.first);
185         } else {
186           READ_NUM(VD.first, Value);
187         }
188         READ_NUM(VD.second, TakenCount);
189         CurrentValues.push_back({Value, TakenCount});
190         Line++;
191       }
192       Record.addValueData(VK, S, CurrentValues.data(), NumValueData, nullptr);
193     }
194   }
195   return success();
196 
197 #undef CHECK_LINE_END
198 #undef READ_NUM
199 #undef VP_READ_ADVANCE
200 }
201 
readNextRecord(InstrProfRecord & Record)202 Error TextInstrProfReader::readNextRecord(InstrProfRecord &Record) {
203   // Skip empty lines and comments.
204   while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
205     ++Line;
206   // If we hit EOF while looking for a name, we're done.
207   if (Line.is_at_end()) {
208     Symtab->finalizeSymtab();
209     return error(instrprof_error::eof);
210   }
211 
212   // Read the function name.
213   Record.Name = *Line++;
214   Symtab->addFuncName(Record.Name);
215 
216   // Read the function hash.
217   if (Line.is_at_end())
218     return error(instrprof_error::truncated);
219   if ((Line++)->getAsInteger(0, Record.Hash))
220     return error(instrprof_error::malformed);
221 
222   // Read the number of counters.
223   uint64_t NumCounters;
224   if (Line.is_at_end())
225     return error(instrprof_error::truncated);
226   if ((Line++)->getAsInteger(10, NumCounters))
227     return error(instrprof_error::malformed);
228   if (NumCounters == 0)
229     return error(instrprof_error::malformed);
230 
231   // Read each counter and fill our internal storage with the values.
232   Record.Counts.clear();
233   Record.Counts.reserve(NumCounters);
234   for (uint64_t I = 0; I < NumCounters; ++I) {
235     if (Line.is_at_end())
236       return error(instrprof_error::truncated);
237     uint64_t Count;
238     if ((Line++)->getAsInteger(10, Count))
239       return error(instrprof_error::malformed);
240     Record.Counts.push_back(Count);
241   }
242 
243   // Check if value profile data exists and read it if so.
244   if (Error E = readValueProfileData(Record))
245     return E;
246 
247   // This is needed to avoid two pass parsing because llvm-profdata
248   // does dumping while reading.
249   Symtab->finalizeSymtab();
250   return success();
251 }
252 
253 template <class IntPtrT>
hasFormat(const MemoryBuffer & DataBuffer)254 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
255   if (DataBuffer.getBufferSize() < sizeof(uint64_t))
256     return false;
257   uint64_t Magic =
258     *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
259   return RawInstrProf::getMagic<IntPtrT>() == Magic ||
260          sys::getSwappedBytes(RawInstrProf::getMagic<IntPtrT>()) == Magic;
261 }
262 
263 template <class IntPtrT>
readHeader()264 Error RawInstrProfReader<IntPtrT>::readHeader() {
265   if (!hasFormat(*DataBuffer))
266     return error(instrprof_error::bad_magic);
267   if (DataBuffer->getBufferSize() < sizeof(RawInstrProf::Header))
268     return error(instrprof_error::bad_header);
269   auto *Header = reinterpret_cast<const RawInstrProf::Header *>(
270       DataBuffer->getBufferStart());
271   ShouldSwapBytes = Header->Magic != RawInstrProf::getMagic<IntPtrT>();
272   return readHeader(*Header);
273 }
274 
275 template <class IntPtrT>
readNextHeader(const char * CurrentPos)276 Error RawInstrProfReader<IntPtrT>::readNextHeader(const char *CurrentPos) {
277   const char *End = DataBuffer->getBufferEnd();
278   // Skip zero padding between profiles.
279   while (CurrentPos != End && *CurrentPos == 0)
280     ++CurrentPos;
281   // If there's nothing left, we're done.
282   if (CurrentPos == End)
283     return make_error<InstrProfError>(instrprof_error::eof);
284   // If there isn't enough space for another header, this is probably just
285   // garbage at the end of the file.
286   if (CurrentPos + sizeof(RawInstrProf::Header) > End)
287     return make_error<InstrProfError>(instrprof_error::malformed);
288   // The writer ensures each profile is padded to start at an aligned address.
289   if (reinterpret_cast<size_t>(CurrentPos) % alignOf<uint64_t>())
290     return make_error<InstrProfError>(instrprof_error::malformed);
291   // The magic should have the same byte order as in the previous header.
292   uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos);
293   if (Magic != swap(RawInstrProf::getMagic<IntPtrT>()))
294     return make_error<InstrProfError>(instrprof_error::bad_magic);
295 
296   // There's another profile to read, so we need to process the header.
297   auto *Header = reinterpret_cast<const RawInstrProf::Header *>(CurrentPos);
298   return readHeader(*Header);
299 }
300 
301 template <class IntPtrT>
createSymtab(InstrProfSymtab & Symtab)302 Error RawInstrProfReader<IntPtrT>::createSymtab(InstrProfSymtab &Symtab) {
303   if (Error E = Symtab.create(StringRef(NamesStart, NamesSize)))
304     return error(std::move(E));
305   for (const RawInstrProf::ProfileData<IntPtrT> *I = Data; I != DataEnd; ++I) {
306     const IntPtrT FPtr = swap(I->FunctionPointer);
307     if (!FPtr)
308       continue;
309     Symtab.mapAddress(FPtr, I->NameRef);
310   }
311   Symtab.finalizeSymtab();
312   return success();
313 }
314 
315 template <class IntPtrT>
readHeader(const RawInstrProf::Header & Header)316 Error RawInstrProfReader<IntPtrT>::readHeader(
317     const RawInstrProf::Header &Header) {
318   Version = swap(Header.Version);
319   if (GET_VERSION(Version) != RawInstrProf::Version)
320     return error(instrprof_error::unsupported_version);
321 
322   CountersDelta = swap(Header.CountersDelta);
323   NamesDelta = swap(Header.NamesDelta);
324   auto DataSize = swap(Header.DataSize);
325   auto CountersSize = swap(Header.CountersSize);
326   NamesSize = swap(Header.NamesSize);
327   ValueKindLast = swap(Header.ValueKindLast);
328 
329   auto DataSizeInBytes = DataSize * sizeof(RawInstrProf::ProfileData<IntPtrT>);
330   auto PaddingSize = getNumPaddingBytes(NamesSize);
331 
332   ptrdiff_t DataOffset = sizeof(RawInstrProf::Header);
333   ptrdiff_t CountersOffset = DataOffset + DataSizeInBytes;
334   ptrdiff_t NamesOffset = CountersOffset + sizeof(uint64_t) * CountersSize;
335   ptrdiff_t ValueDataOffset = NamesOffset + NamesSize + PaddingSize;
336 
337   auto *Start = reinterpret_cast<const char *>(&Header);
338   if (Start + ValueDataOffset > DataBuffer->getBufferEnd())
339     return error(instrprof_error::bad_header);
340 
341   Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>(
342       Start + DataOffset);
343   DataEnd = Data + DataSize;
344   CountersStart = reinterpret_cast<const uint64_t *>(Start + CountersOffset);
345   NamesStart = Start + NamesOffset;
346   ValueDataStart = reinterpret_cast<const uint8_t *>(Start + ValueDataOffset);
347 
348   std::unique_ptr<InstrProfSymtab> NewSymtab = make_unique<InstrProfSymtab>();
349   if (Error E = createSymtab(*NewSymtab.get()))
350     return E;
351 
352   Symtab = std::move(NewSymtab);
353   return success();
354 }
355 
356 template <class IntPtrT>
readName(InstrProfRecord & Record)357 Error RawInstrProfReader<IntPtrT>::readName(InstrProfRecord &Record) {
358   Record.Name = getName(Data->NameRef);
359   return success();
360 }
361 
362 template <class IntPtrT>
readFuncHash(InstrProfRecord & Record)363 Error RawInstrProfReader<IntPtrT>::readFuncHash(InstrProfRecord &Record) {
364   Record.Hash = swap(Data->FuncHash);
365   return success();
366 }
367 
368 template <class IntPtrT>
readRawCounts(InstrProfRecord & Record)369 Error RawInstrProfReader<IntPtrT>::readRawCounts(
370     InstrProfRecord &Record) {
371   uint32_t NumCounters = swap(Data->NumCounters);
372   IntPtrT CounterPtr = Data->CounterPtr;
373   if (NumCounters == 0)
374     return error(instrprof_error::malformed);
375 
376   auto RawCounts = makeArrayRef(getCounter(CounterPtr), NumCounters);
377   auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
378 
379   // Check bounds.
380   if (RawCounts.data() < CountersStart ||
381       RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
382     return error(instrprof_error::malformed);
383 
384   if (ShouldSwapBytes) {
385     Record.Counts.clear();
386     Record.Counts.reserve(RawCounts.size());
387     for (uint64_t Count : RawCounts)
388       Record.Counts.push_back(swap(Count));
389   } else
390     Record.Counts = RawCounts;
391 
392   return success();
393 }
394 
395 template <class IntPtrT>
readValueProfilingData(InstrProfRecord & Record)396 Error RawInstrProfReader<IntPtrT>::readValueProfilingData(
397     InstrProfRecord &Record) {
398 
399   Record.clearValueData();
400   CurValueDataSize = 0;
401   // Need to match the logic in value profile dumper code in compiler-rt:
402   uint32_t NumValueKinds = 0;
403   for (uint32_t I = 0; I < IPVK_Last + 1; I++)
404     NumValueKinds += (Data->NumValueSites[I] != 0);
405 
406   if (!NumValueKinds)
407     return success();
408 
409   Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
410       ValueProfData::getValueProfData(
411           ValueDataStart, (const unsigned char *)DataBuffer->getBufferEnd(),
412           getDataEndianness());
413 
414   if (Error E = VDataPtrOrErr.takeError())
415     return E;
416 
417   // Note that besides deserialization, this also performs the conversion for
418   // indirect call targets.  The function pointers from the raw profile are
419   // remapped into function name hashes.
420   VDataPtrOrErr.get()->deserializeTo(Record, &Symtab->getAddrHashMap());
421   CurValueDataSize = VDataPtrOrErr.get()->getSize();
422   return success();
423 }
424 
425 template <class IntPtrT>
readNextRecord(InstrProfRecord & Record)426 Error RawInstrProfReader<IntPtrT>::readNextRecord(InstrProfRecord &Record) {
427   if (atEnd())
428     // At this point, ValueDataStart field points to the next header.
429     if (Error E = readNextHeader(getNextHeaderPos()))
430       return E;
431 
432   // Read name ad set it in Record.
433   if (Error E = readName(Record))
434     return E;
435 
436   // Read FuncHash and set it in Record.
437   if (Error E = readFuncHash(Record))
438     return E;
439 
440   // Read raw counts and set Record.
441   if (Error E = readRawCounts(Record))
442     return E;
443 
444   // Read value data and set Record.
445   if (Error E = readValueProfilingData(Record))
446     return E;
447 
448   // Iterate.
449   advanceData();
450   return success();
451 }
452 
453 namespace llvm {
454 template class RawInstrProfReader<uint32_t>;
455 template class RawInstrProfReader<uint64_t>;
456 }
457 
458 InstrProfLookupTrait::hash_value_type
ComputeHash(StringRef K)459 InstrProfLookupTrait::ComputeHash(StringRef K) {
460   return IndexedInstrProf::ComputeHash(HashType, K);
461 }
462 
463 typedef InstrProfLookupTrait::data_type data_type;
464 typedef InstrProfLookupTrait::offset_type offset_type;
465 
readValueProfilingData(const unsigned char * & D,const unsigned char * const End)466 bool InstrProfLookupTrait::readValueProfilingData(
467     const unsigned char *&D, const unsigned char *const End) {
468   Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
469       ValueProfData::getValueProfData(D, End, ValueProfDataEndianness);
470 
471   if (VDataPtrOrErr.takeError())
472     return false;
473 
474   VDataPtrOrErr.get()->deserializeTo(DataBuffer.back(), nullptr);
475   D += VDataPtrOrErr.get()->TotalSize;
476 
477   return true;
478 }
479 
ReadData(StringRef K,const unsigned char * D,offset_type N)480 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
481                                          offset_type N) {
482   // Check if the data is corrupt. If so, don't try to read it.
483   if (N % sizeof(uint64_t))
484     return data_type();
485 
486   DataBuffer.clear();
487   std::vector<uint64_t> CounterBuffer;
488 
489   using namespace support;
490   const unsigned char *End = D + N;
491   while (D < End) {
492     // Read hash.
493     if (D + sizeof(uint64_t) >= End)
494       return data_type();
495     uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
496 
497     // Initialize number of counters for GET_VERSION(FormatVersion) == 1.
498     uint64_t CountsSize = N / sizeof(uint64_t) - 1;
499     // If format version is different then read the number of counters.
500     if (GET_VERSION(FormatVersion) != IndexedInstrProf::ProfVersion::Version1) {
501       if (D + sizeof(uint64_t) > End)
502         return data_type();
503       CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
504     }
505     // Read counter values.
506     if (D + CountsSize * sizeof(uint64_t) > End)
507       return data_type();
508 
509     CounterBuffer.clear();
510     CounterBuffer.reserve(CountsSize);
511     for (uint64_t J = 0; J < CountsSize; ++J)
512       CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
513 
514     DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer));
515 
516     // Read value profiling data.
517     if (GET_VERSION(FormatVersion) > IndexedInstrProf::ProfVersion::Version2 &&
518         !readValueProfilingData(D, End)) {
519       DataBuffer.clear();
520       return data_type();
521     }
522   }
523   return DataBuffer;
524 }
525 
526 template <typename HashTableImpl>
getRecords(StringRef FuncName,ArrayRef<InstrProfRecord> & Data)527 Error InstrProfReaderIndex<HashTableImpl>::getRecords(
528     StringRef FuncName, ArrayRef<InstrProfRecord> &Data) {
529   auto Iter = HashTable->find(FuncName);
530   if (Iter == HashTable->end())
531     return make_error<InstrProfError>(instrprof_error::unknown_function);
532 
533   Data = (*Iter);
534   if (Data.empty())
535     return make_error<InstrProfError>(instrprof_error::malformed);
536 
537   return Error::success();
538 }
539 
540 template <typename HashTableImpl>
getRecords(ArrayRef<InstrProfRecord> & Data)541 Error InstrProfReaderIndex<HashTableImpl>::getRecords(
542     ArrayRef<InstrProfRecord> &Data) {
543   if (atEnd())
544     return make_error<InstrProfError>(instrprof_error::eof);
545 
546   Data = *RecordIterator;
547 
548   if (Data.empty())
549     return make_error<InstrProfError>(instrprof_error::malformed);
550 
551   return Error::success();
552 }
553 
554 template <typename HashTableImpl>
InstrProfReaderIndex(const unsigned char * Buckets,const unsigned char * const Payload,const unsigned char * const Base,IndexedInstrProf::HashT HashType,uint64_t Version)555 InstrProfReaderIndex<HashTableImpl>::InstrProfReaderIndex(
556     const unsigned char *Buckets, const unsigned char *const Payload,
557     const unsigned char *const Base, IndexedInstrProf::HashT HashType,
558     uint64_t Version) {
559   FormatVersion = Version;
560   HashTable.reset(HashTableImpl::Create(
561       Buckets, Payload, Base,
562       typename HashTableImpl::InfoType(HashType, Version)));
563   RecordIterator = HashTable->data_begin();
564 }
565 
hasFormat(const MemoryBuffer & DataBuffer)566 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
567   if (DataBuffer.getBufferSize() < 8)
568     return false;
569   using namespace support;
570   uint64_t Magic =
571       endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
572   // Verify that it's magical.
573   return Magic == IndexedInstrProf::Magic;
574 }
575 
576 const unsigned char *
readSummary(IndexedInstrProf::ProfVersion Version,const unsigned char * Cur)577 IndexedInstrProfReader::readSummary(IndexedInstrProf::ProfVersion Version,
578                                     const unsigned char *Cur) {
579   using namespace IndexedInstrProf;
580   using namespace support;
581   if (Version >= IndexedInstrProf::Version4) {
582     const IndexedInstrProf::Summary *SummaryInLE =
583         reinterpret_cast<const IndexedInstrProf::Summary *>(Cur);
584     uint64_t NFields =
585         endian::byte_swap<uint64_t, little>(SummaryInLE->NumSummaryFields);
586     uint64_t NEntries =
587         endian::byte_swap<uint64_t, little>(SummaryInLE->NumCutoffEntries);
588     uint32_t SummarySize =
589         IndexedInstrProf::Summary::getSize(NFields, NEntries);
590     std::unique_ptr<IndexedInstrProf::Summary> SummaryData =
591         IndexedInstrProf::allocSummary(SummarySize);
592 
593     const uint64_t *Src = reinterpret_cast<const uint64_t *>(SummaryInLE);
594     uint64_t *Dst = reinterpret_cast<uint64_t *>(SummaryData.get());
595     for (unsigned I = 0; I < SummarySize / sizeof(uint64_t); I++)
596       Dst[I] = endian::byte_swap<uint64_t, little>(Src[I]);
597 
598     llvm::SummaryEntryVector DetailedSummary;
599     for (unsigned I = 0; I < SummaryData->NumCutoffEntries; I++) {
600       const IndexedInstrProf::Summary::Entry &Ent = SummaryData->getEntry(I);
601       DetailedSummary.emplace_back((uint32_t)Ent.Cutoff, Ent.MinBlockCount,
602                                    Ent.NumBlocks);
603     }
604     // initialize InstrProfSummary using the SummaryData from disk.
605     this->Summary = llvm::make_unique<ProfileSummary>(
606         ProfileSummary::PSK_Instr, DetailedSummary,
607         SummaryData->get(Summary::TotalBlockCount),
608         SummaryData->get(Summary::MaxBlockCount),
609         SummaryData->get(Summary::MaxInternalBlockCount),
610         SummaryData->get(Summary::MaxFunctionCount),
611         SummaryData->get(Summary::TotalNumBlocks),
612         SummaryData->get(Summary::TotalNumFunctions));
613     return Cur + SummarySize;
614   } else {
615     // For older version of profile data, we need to compute on the fly:
616     using namespace IndexedInstrProf;
617     InstrProfSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs);
618     // FIXME: This only computes an empty summary. Need to call addRecord for
619     // all InstrProfRecords to get the correct summary.
620     this->Summary = Builder.getSummary();
621     return Cur;
622   }
623 }
624 
readHeader()625 Error IndexedInstrProfReader::readHeader() {
626   const unsigned char *Start =
627       (const unsigned char *)DataBuffer->getBufferStart();
628   const unsigned char *Cur = Start;
629   if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
630     return error(instrprof_error::truncated);
631 
632   using namespace support;
633 
634   auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
635   Cur += sizeof(IndexedInstrProf::Header);
636 
637   // Check the magic number.
638   uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
639   if (Magic != IndexedInstrProf::Magic)
640     return error(instrprof_error::bad_magic);
641 
642   // Read the version.
643   uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
644   if (GET_VERSION(FormatVersion) >
645       IndexedInstrProf::ProfVersion::CurrentVersion)
646     return error(instrprof_error::unsupported_version);
647 
648   Cur = readSummary((IndexedInstrProf::ProfVersion)FormatVersion, Cur);
649 
650   // Read the hash type and start offset.
651   IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
652       endian::byte_swap<uint64_t, little>(Header->HashType));
653   if (HashType > IndexedInstrProf::HashT::Last)
654     return error(instrprof_error::unsupported_hash_type);
655 
656   uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
657 
658   // The rest of the file is an on disk hash table.
659   InstrProfReaderIndexBase *IndexPtr = nullptr;
660   IndexPtr = new InstrProfReaderIndex<OnDiskHashTableImplV3>(
661       Start + HashOffset, Cur, Start, HashType, FormatVersion);
662   Index.reset(IndexPtr);
663   return success();
664 }
665 
getSymtab()666 InstrProfSymtab &IndexedInstrProfReader::getSymtab() {
667   if (Symtab.get())
668     return *Symtab.get();
669 
670   std::unique_ptr<InstrProfSymtab> NewSymtab = make_unique<InstrProfSymtab>();
671   Index->populateSymtab(*NewSymtab.get());
672 
673   Symtab = std::move(NewSymtab);
674   return *Symtab.get();
675 }
676 
677 Expected<InstrProfRecord>
getInstrProfRecord(StringRef FuncName,uint64_t FuncHash)678 IndexedInstrProfReader::getInstrProfRecord(StringRef FuncName,
679                                            uint64_t FuncHash) {
680   ArrayRef<InstrProfRecord> Data;
681   Error Err = Index->getRecords(FuncName, Data);
682   if (Err)
683     return std::move(Err);
684   // Found it. Look for counters with the right hash.
685   for (unsigned I = 0, E = Data.size(); I < E; ++I) {
686     // Check for a match and fill the vector if there is one.
687     if (Data[I].Hash == FuncHash) {
688       return std::move(Data[I]);
689     }
690   }
691   return error(instrprof_error::hash_mismatch);
692 }
693 
getFunctionCounts(StringRef FuncName,uint64_t FuncHash,std::vector<uint64_t> & Counts)694 Error IndexedInstrProfReader::getFunctionCounts(StringRef FuncName,
695                                                 uint64_t FuncHash,
696                                                 std::vector<uint64_t> &Counts) {
697   Expected<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash);
698   if (Error E = Record.takeError())
699     return error(std::move(E));
700 
701   Counts = Record.get().Counts;
702   return success();
703 }
704 
readNextRecord(InstrProfRecord & Record)705 Error IndexedInstrProfReader::readNextRecord(InstrProfRecord &Record) {
706   static unsigned RecordIndex = 0;
707 
708   ArrayRef<InstrProfRecord> Data;
709 
710   Error E = Index->getRecords(Data);
711   if (E)
712     return error(std::move(E));
713 
714   Record = Data[RecordIndex++];
715   if (RecordIndex >= Data.size()) {
716     Index->advanceToNextKey();
717     RecordIndex = 0;
718   }
719   return success();
720 }
721