1 //===- InstrProfReader.cpp - Instrumented profiling reader ----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains support for reading profiling data for clang's
10 // instrumentation based PGO and coverage.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/ProfileData/InstrProfReader.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/StringExtras.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/IR/ProfileSummary.h"
21 #include "llvm/ProfileData/InstrProf.h"
22 #include "llvm/ProfileData/ProfileCommon.h"
23 #include "llvm/Support/Endian.h"
24 #include "llvm/Support/Error.h"
25 #include "llvm/Support/ErrorOr.h"
26 #include "llvm/Support/MemoryBuffer.h"
27 #include "llvm/Support/SwapByteOrder.h"
28 #include "llvm/Support/SymbolRemappingReader.h"
29 #include <algorithm>
30 #include <cctype>
31 #include <cstddef>
32 #include <cstdint>
33 #include <limits>
34 #include <memory>
35 #include <system_error>
36 #include <utility>
37 #include <vector>
38 
39 using namespace llvm;
40 
41 static Expected<std::unique_ptr<MemoryBuffer>>
42 setupMemoryBuffer(const Twine &Path) {
43   ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
44       MemoryBuffer::getFileOrSTDIN(Path, /*IsText=*/true);
45   if (std::error_code EC = BufferOrErr.getError())
46     return errorCodeToError(EC);
47   return std::move(BufferOrErr.get());
48 }
49 
50 static Error initializeReader(InstrProfReader &Reader) {
51   return Reader.readHeader();
52 }
53 
54 Expected<std::unique_ptr<InstrProfReader>>
55 InstrProfReader::create(const Twine &Path,
56                         const InstrProfCorrelator *Correlator) {
57   // Set up the buffer to read.
58   auto BufferOrError = setupMemoryBuffer(Path);
59   if (Error E = BufferOrError.takeError())
60     return std::move(E);
61   return InstrProfReader::create(std::move(BufferOrError.get()), Correlator);
62 }
63 
64 Expected<std::unique_ptr<InstrProfReader>>
65 InstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer,
66                         const InstrProfCorrelator *Correlator) {
67   // Sanity check the buffer.
68   if (uint64_t(Buffer->getBufferSize()) > std::numeric_limits<uint64_t>::max())
69     return make_error<InstrProfError>(instrprof_error::too_large);
70 
71   if (Buffer->getBufferSize() == 0)
72     return make_error<InstrProfError>(instrprof_error::empty_raw_profile);
73 
74   std::unique_ptr<InstrProfReader> Result;
75   // Create the reader.
76   if (IndexedInstrProfReader::hasFormat(*Buffer))
77     Result.reset(new IndexedInstrProfReader(std::move(Buffer)));
78   else if (RawInstrProfReader64::hasFormat(*Buffer))
79     Result.reset(new RawInstrProfReader64(std::move(Buffer), Correlator));
80   else if (RawInstrProfReader32::hasFormat(*Buffer))
81     Result.reset(new RawInstrProfReader32(std::move(Buffer), Correlator));
82   else if (TextInstrProfReader::hasFormat(*Buffer))
83     Result.reset(new TextInstrProfReader(std::move(Buffer)));
84   else
85     return make_error<InstrProfError>(instrprof_error::unrecognized_format);
86 
87   // Initialize the reader and return the result.
88   if (Error E = initializeReader(*Result))
89     return std::move(E);
90 
91   return std::move(Result);
92 }
93 
94 Expected<std::unique_ptr<IndexedInstrProfReader>>
95 IndexedInstrProfReader::create(const Twine &Path, const Twine &RemappingPath) {
96   // Set up the buffer to read.
97   auto BufferOrError = setupMemoryBuffer(Path);
98   if (Error E = BufferOrError.takeError())
99     return std::move(E);
100 
101   // Set up the remapping buffer if requested.
102   std::unique_ptr<MemoryBuffer> RemappingBuffer;
103   std::string RemappingPathStr = RemappingPath.str();
104   if (!RemappingPathStr.empty()) {
105     auto RemappingBufferOrError = setupMemoryBuffer(RemappingPathStr);
106     if (Error E = RemappingBufferOrError.takeError())
107       return std::move(E);
108     RemappingBuffer = std::move(RemappingBufferOrError.get());
109   }
110 
111   return IndexedInstrProfReader::create(std::move(BufferOrError.get()),
112                                         std::move(RemappingBuffer));
113 }
114 
115 Expected<std::unique_ptr<IndexedInstrProfReader>>
116 IndexedInstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer,
117                                std::unique_ptr<MemoryBuffer> RemappingBuffer) {
118   if (uint64_t(Buffer->getBufferSize()) > std::numeric_limits<uint64_t>::max())
119     return make_error<InstrProfError>(instrprof_error::too_large);
120 
121   // Create the reader.
122   if (!IndexedInstrProfReader::hasFormat(*Buffer))
123     return make_error<InstrProfError>(instrprof_error::bad_magic);
124   auto Result = std::make_unique<IndexedInstrProfReader>(
125       std::move(Buffer), std::move(RemappingBuffer));
126 
127   // Initialize the reader and return the result.
128   if (Error E = initializeReader(*Result))
129     return std::move(E);
130 
131   return std::move(Result);
132 }
133 
134 void InstrProfIterator::Increment() {
135   if (auto E = Reader->readNextRecord(Record)) {
136     // Handle errors in the reader.
137     InstrProfError::take(std::move(E));
138     *this = InstrProfIterator();
139   }
140 }
141 
142 bool TextInstrProfReader::hasFormat(const MemoryBuffer &Buffer) {
143   // Verify that this really looks like plain ASCII text by checking a
144   // 'reasonable' number of characters (up to profile magic size).
145   size_t count = std::min(Buffer.getBufferSize(), sizeof(uint64_t));
146   StringRef buffer = Buffer.getBufferStart();
147   return count == 0 ||
148          std::all_of(buffer.begin(), buffer.begin() + count,
149                      [](char c) { return isPrint(c) || isSpace(c); });
150 }
151 
152 // Read the profile variant flag from the header: ":FE" means this is a FE
153 // generated profile. ":IR" means this is an IR level profile. Other strings
154 // with a leading ':' will be reported an error format.
155 Error TextInstrProfReader::readHeader() {
156   Symtab.reset(new InstrProfSymtab());
157 
158   while (Line->startswith(":")) {
159     StringRef Str = Line->substr(1);
160     if (Str.equals_insensitive("ir"))
161       ProfileKind |= InstrProfKind::IR;
162     else if (Str.equals_insensitive("fe"))
163       ProfileKind |= InstrProfKind::FE;
164     else if (Str.equals_insensitive("csir")) {
165       ProfileKind |= InstrProfKind::IR;
166       ProfileKind |= InstrProfKind::CS;
167     } else if (Str.equals_insensitive("entry_first"))
168       ProfileKind |= InstrProfKind::BB;
169     else if (Str.equals_insensitive("not_entry_first"))
170       ProfileKind &= ~InstrProfKind::BB;
171     else
172       return error(instrprof_error::bad_header);
173     ++Line;
174   }
175   return success();
176 }
177 
178 Error
179 TextInstrProfReader::readValueProfileData(InstrProfRecord &Record) {
180 
181 #define CHECK_LINE_END(Line)                                                   \
182   if (Line.is_at_end())                                                        \
183     return error(instrprof_error::truncated);
184 #define READ_NUM(Str, Dst)                                                     \
185   if ((Str).getAsInteger(10, (Dst)))                                           \
186     return error(instrprof_error::malformed);
187 #define VP_READ_ADVANCE(Val)                                                   \
188   CHECK_LINE_END(Line);                                                        \
189   uint32_t Val;                                                                \
190   READ_NUM((*Line), (Val));                                                    \
191   Line++;
192 
193   if (Line.is_at_end())
194     return success();
195 
196   uint32_t NumValueKinds;
197   if (Line->getAsInteger(10, NumValueKinds)) {
198     // No value profile data
199     return success();
200   }
201   if (NumValueKinds == 0 || NumValueKinds > IPVK_Last + 1)
202     return error(instrprof_error::malformed,
203                  "number of value kinds is invalid");
204   Line++;
205 
206   for (uint32_t VK = 0; VK < NumValueKinds; VK++) {
207     VP_READ_ADVANCE(ValueKind);
208     if (ValueKind > IPVK_Last)
209       return error(instrprof_error::malformed, "value kind is invalid");
210     ;
211     VP_READ_ADVANCE(NumValueSites);
212     if (!NumValueSites)
213       continue;
214 
215     Record.reserveSites(VK, NumValueSites);
216     for (uint32_t S = 0; S < NumValueSites; S++) {
217       VP_READ_ADVANCE(NumValueData);
218 
219       std::vector<InstrProfValueData> CurrentValues;
220       for (uint32_t V = 0; V < NumValueData; V++) {
221         CHECK_LINE_END(Line);
222         std::pair<StringRef, StringRef> VD = Line->rsplit(':');
223         uint64_t TakenCount, Value;
224         if (ValueKind == IPVK_IndirectCallTarget) {
225           if (InstrProfSymtab::isExternalSymbol(VD.first)) {
226             Value = 0;
227           } else {
228             if (Error E = Symtab->addFuncName(VD.first))
229               return E;
230             Value = IndexedInstrProf::ComputeHash(VD.first);
231           }
232         } else {
233           READ_NUM(VD.first, Value);
234         }
235         READ_NUM(VD.second, TakenCount);
236         CurrentValues.push_back({Value, TakenCount});
237         Line++;
238       }
239       Record.addValueData(ValueKind, S, CurrentValues.data(), NumValueData,
240                           nullptr);
241     }
242   }
243   return success();
244 
245 #undef CHECK_LINE_END
246 #undef READ_NUM
247 #undef VP_READ_ADVANCE
248 }
249 
250 Error TextInstrProfReader::readNextRecord(NamedInstrProfRecord &Record) {
251   // Skip empty lines and comments.
252   while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
253     ++Line;
254   // If we hit EOF while looking for a name, we're done.
255   if (Line.is_at_end()) {
256     return error(instrprof_error::eof);
257   }
258 
259   // Read the function name.
260   Record.Name = *Line++;
261   if (Error E = Symtab->addFuncName(Record.Name))
262     return error(std::move(E));
263 
264   // Read the function hash.
265   if (Line.is_at_end())
266     return error(instrprof_error::truncated);
267   if ((Line++)->getAsInteger(0, Record.Hash))
268     return error(instrprof_error::malformed,
269                  "function hash is not a valid integer");
270 
271   // Read the number of counters.
272   uint64_t NumCounters;
273   if (Line.is_at_end())
274     return error(instrprof_error::truncated);
275   if ((Line++)->getAsInteger(10, NumCounters))
276     return error(instrprof_error::malformed,
277                  "number of counters is not a valid integer");
278   if (NumCounters == 0)
279     return error(instrprof_error::malformed, "number of counters is zero");
280 
281   // Read each counter and fill our internal storage with the values.
282   Record.Clear();
283   Record.Counts.reserve(NumCounters);
284   for (uint64_t I = 0; I < NumCounters; ++I) {
285     if (Line.is_at_end())
286       return error(instrprof_error::truncated);
287     uint64_t Count;
288     if ((Line++)->getAsInteger(10, Count))
289       return error(instrprof_error::malformed, "count is invalid");
290     Record.Counts.push_back(Count);
291   }
292 
293   // Check if value profile data exists and read it if so.
294   if (Error E = readValueProfileData(Record))
295     return error(std::move(E));
296 
297   return success();
298 }
299 
300 template <class IntPtrT>
301 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
302   if (DataBuffer.getBufferSize() < sizeof(uint64_t))
303     return false;
304   uint64_t Magic =
305     *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
306   return RawInstrProf::getMagic<IntPtrT>() == Magic ||
307          sys::getSwappedBytes(RawInstrProf::getMagic<IntPtrT>()) == Magic;
308 }
309 
310 template <class IntPtrT>
311 Error RawInstrProfReader<IntPtrT>::readHeader() {
312   if (!hasFormat(*DataBuffer))
313     return error(instrprof_error::bad_magic);
314   if (DataBuffer->getBufferSize() < sizeof(RawInstrProf::Header))
315     return error(instrprof_error::bad_header);
316   auto *Header = reinterpret_cast<const RawInstrProf::Header *>(
317       DataBuffer->getBufferStart());
318   ShouldSwapBytes = Header->Magic != RawInstrProf::getMagic<IntPtrT>();
319   return readHeader(*Header);
320 }
321 
322 template <class IntPtrT>
323 Error RawInstrProfReader<IntPtrT>::readNextHeader(const char *CurrentPos) {
324   const char *End = DataBuffer->getBufferEnd();
325   // Skip zero padding between profiles.
326   while (CurrentPos != End && *CurrentPos == 0)
327     ++CurrentPos;
328   // If there's nothing left, we're done.
329   if (CurrentPos == End)
330     return make_error<InstrProfError>(instrprof_error::eof);
331   // If there isn't enough space for another header, this is probably just
332   // garbage at the end of the file.
333   if (CurrentPos + sizeof(RawInstrProf::Header) > End)
334     return make_error<InstrProfError>(instrprof_error::malformed,
335                                       "not enough space for another header");
336   // The writer ensures each profile is padded to start at an aligned address.
337   if (reinterpret_cast<size_t>(CurrentPos) % alignof(uint64_t))
338     return make_error<InstrProfError>(instrprof_error::malformed,
339                                       "insufficient padding");
340   // The magic should have the same byte order as in the previous header.
341   uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos);
342   if (Magic != swap(RawInstrProf::getMagic<IntPtrT>()))
343     return make_error<InstrProfError>(instrprof_error::bad_magic);
344 
345   // There's another profile to read, so we need to process the header.
346   auto *Header = reinterpret_cast<const RawInstrProf::Header *>(CurrentPos);
347   return readHeader(*Header);
348 }
349 
350 template <class IntPtrT>
351 Error RawInstrProfReader<IntPtrT>::createSymtab(InstrProfSymtab &Symtab) {
352   if (Error E = Symtab.create(StringRef(NamesStart, NamesEnd - NamesStart)))
353     return error(std::move(E));
354   for (const RawInstrProf::ProfileData<IntPtrT> *I = Data; I != DataEnd; ++I) {
355     const IntPtrT FPtr = swap(I->FunctionPointer);
356     if (!FPtr)
357       continue;
358     Symtab.mapAddress(FPtr, I->NameRef);
359   }
360   return success();
361 }
362 
363 template <class IntPtrT>
364 Error RawInstrProfReader<IntPtrT>::readHeader(
365     const RawInstrProf::Header &Header) {
366   Version = swap(Header.Version);
367   if (GET_VERSION(Version) != RawInstrProf::Version)
368     return error(instrprof_error::unsupported_version);
369   if (useDebugInfoCorrelate() && !Correlator)
370     return error(instrprof_error::missing_debug_info_for_correlation);
371   if (!useDebugInfoCorrelate() && Correlator)
372     return error(instrprof_error::unexpected_debug_info_for_correlation);
373 
374   BinaryIdsSize = swap(Header.BinaryIdsSize);
375   if (BinaryIdsSize % sizeof(uint64_t))
376     return error(instrprof_error::bad_header);
377 
378   CountersDelta = swap(Header.CountersDelta);
379   NamesDelta = swap(Header.NamesDelta);
380   auto NumData = swap(Header.DataSize);
381   auto PaddingBytesBeforeCounters = swap(Header.PaddingBytesBeforeCounters);
382   auto CountersSize = swap(Header.CountersSize) * getCounterTypeSize();
383   auto PaddingBytesAfterCounters = swap(Header.PaddingBytesAfterCounters);
384   auto NamesSize = swap(Header.NamesSize);
385   ValueKindLast = swap(Header.ValueKindLast);
386 
387   auto DataSize = NumData * sizeof(RawInstrProf::ProfileData<IntPtrT>);
388   auto PaddingSize = getNumPaddingBytes(NamesSize);
389 
390   // Profile data starts after profile header and binary ids if exist.
391   ptrdiff_t DataOffset = sizeof(RawInstrProf::Header) + BinaryIdsSize;
392   ptrdiff_t CountersOffset = DataOffset + DataSize + PaddingBytesBeforeCounters;
393   ptrdiff_t NamesOffset =
394       CountersOffset + CountersSize + PaddingBytesAfterCounters;
395   ptrdiff_t ValueDataOffset = NamesOffset + NamesSize + PaddingSize;
396 
397   auto *Start = reinterpret_cast<const char *>(&Header);
398   if (Start + ValueDataOffset > DataBuffer->getBufferEnd())
399     return error(instrprof_error::bad_header);
400 
401   if (Correlator) {
402     // These sizes in the raw file are zero because we constructed them in the
403     // Correlator.
404     assert(DataSize == 0 && NamesSize == 0);
405     assert(CountersDelta == 0 && NamesDelta == 0);
406     Data = Correlator->getDataPointer();
407     DataEnd = Data + Correlator->getDataSize();
408     NamesStart = Correlator->getNamesPointer();
409     NamesEnd = NamesStart + Correlator->getNamesSize();
410   } else {
411     Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>(
412         Start + DataOffset);
413     DataEnd = Data + NumData;
414     NamesStart = Start + NamesOffset;
415     NamesEnd = NamesStart + NamesSize;
416   }
417 
418   // Binary ids start just after the header.
419   BinaryIdsStart =
420       reinterpret_cast<const uint8_t *>(&Header) + sizeof(RawInstrProf::Header);
421   CountersStart = Start + CountersOffset;
422   CountersEnd = CountersStart + CountersSize;
423   ValueDataStart = reinterpret_cast<const uint8_t *>(Start + ValueDataOffset);
424 
425   const uint8_t *BufferEnd = (const uint8_t *)DataBuffer->getBufferEnd();
426   if (BinaryIdsStart + BinaryIdsSize > BufferEnd)
427     return error(instrprof_error::bad_header);
428 
429   std::unique_ptr<InstrProfSymtab> NewSymtab = std::make_unique<InstrProfSymtab>();
430   if (Error E = createSymtab(*NewSymtab.get()))
431     return E;
432 
433   Symtab = std::move(NewSymtab);
434   return success();
435 }
436 
437 template <class IntPtrT>
438 Error RawInstrProfReader<IntPtrT>::readName(NamedInstrProfRecord &Record) {
439   Record.Name = getName(Data->NameRef);
440   return success();
441 }
442 
443 template <class IntPtrT>
444 Error RawInstrProfReader<IntPtrT>::readFuncHash(NamedInstrProfRecord &Record) {
445   Record.Hash = swap(Data->FuncHash);
446   return success();
447 }
448 
449 template <class IntPtrT>
450 Error RawInstrProfReader<IntPtrT>::readRawCounts(
451     InstrProfRecord &Record) {
452   uint32_t NumCounters = swap(Data->NumCounters);
453   if (NumCounters == 0)
454     return error(instrprof_error::malformed, "number of counters is zero");
455 
456   ptrdiff_t CounterBaseOffset = swap(Data->CounterPtr) - CountersDelta;
457   if (CounterBaseOffset < 0)
458     return error(
459         instrprof_error::malformed,
460         ("counter offset " + Twine(CounterBaseOffset) + " is negative").str());
461 
462   if (CounterBaseOffset >= CountersEnd - CountersStart)
463     return error(instrprof_error::malformed,
464                  ("counter offset " + Twine(CounterBaseOffset) +
465                   " is greater than the maximum counter offset " +
466                   Twine(CountersEnd - CountersStart - 1))
467                      .str());
468 
469   uint64_t MaxNumCounters =
470       (CountersEnd - (CountersStart + CounterBaseOffset)) /
471       getCounterTypeSize();
472   if (NumCounters > MaxNumCounters)
473     return error(instrprof_error::malformed,
474                  ("number of counters " + Twine(NumCounters) +
475                   " is greater than the maximum number of counters " +
476                   Twine(MaxNumCounters))
477                      .str());
478 
479   Record.Counts.clear();
480   Record.Counts.reserve(NumCounters);
481   for (uint32_t I = 0; I < NumCounters; I++) {
482     const char *Ptr =
483         CountersStart + CounterBaseOffset + I * getCounterTypeSize();
484     if (hasSingleByteCoverage()) {
485       // A value of zero signifies the block is covered.
486       Record.Counts.push_back(*Ptr == 0 ? 1 : 0);
487     } else {
488       const auto *CounterValue = reinterpret_cast<const uint64_t *>(Ptr);
489       Record.Counts.push_back(swap(*CounterValue));
490     }
491   }
492 
493   return success();
494 }
495 
496 template <class IntPtrT>
497 Error RawInstrProfReader<IntPtrT>::readValueProfilingData(
498     InstrProfRecord &Record) {
499   Record.clearValueData();
500   CurValueDataSize = 0;
501   // Need to match the logic in value profile dumper code in compiler-rt:
502   uint32_t NumValueKinds = 0;
503   for (uint32_t I = 0; I < IPVK_Last + 1; I++)
504     NumValueKinds += (Data->NumValueSites[I] != 0);
505 
506   if (!NumValueKinds)
507     return success();
508 
509   Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
510       ValueProfData::getValueProfData(
511           ValueDataStart, (const unsigned char *)DataBuffer->getBufferEnd(),
512           getDataEndianness());
513 
514   if (Error E = VDataPtrOrErr.takeError())
515     return E;
516 
517   // Note that besides deserialization, this also performs the conversion for
518   // indirect call targets.  The function pointers from the raw profile are
519   // remapped into function name hashes.
520   VDataPtrOrErr.get()->deserializeTo(Record, Symtab.get());
521   CurValueDataSize = VDataPtrOrErr.get()->getSize();
522   return success();
523 }
524 
525 template <class IntPtrT>
526 Error RawInstrProfReader<IntPtrT>::readNextRecord(NamedInstrProfRecord &Record) {
527   if (atEnd())
528     // At this point, ValueDataStart field points to the next header.
529     if (Error E = readNextHeader(getNextHeaderPos()))
530       return error(std::move(E));
531 
532   // Read name ad set it in Record.
533   if (Error E = readName(Record))
534     return error(std::move(E));
535 
536   // Read FuncHash and set it in Record.
537   if (Error E = readFuncHash(Record))
538     return error(std::move(E));
539 
540   // Read raw counts and set Record.
541   if (Error E = readRawCounts(Record))
542     return error(std::move(E));
543 
544   // Read value data and set Record.
545   if (Error E = readValueProfilingData(Record))
546     return error(std::move(E));
547 
548   // Iterate.
549   advanceData();
550   return success();
551 }
552 
553 static size_t RoundUp(size_t size, size_t align) {
554   return (size + align - 1) & ~(align - 1);
555 }
556 
557 template <class IntPtrT>
558 Error RawInstrProfReader<IntPtrT>::printBinaryIds(raw_ostream &OS) {
559   if (BinaryIdsSize == 0)
560     return success();
561 
562   OS << "Binary IDs: \n";
563   const uint8_t *BI = BinaryIdsStart;
564   const uint8_t *BIEnd = BinaryIdsStart + BinaryIdsSize;
565   while (BI < BIEnd) {
566     size_t Remaining = BIEnd - BI;
567 
568     // There should be enough left to read the binary ID size field.
569     if (Remaining < sizeof(uint64_t))
570       return make_error<InstrProfError>(
571           instrprof_error::malformed,
572           "not enough data to read binary id length");
573 
574     uint64_t BinaryIdLen = swap(*reinterpret_cast<const uint64_t *>(BI));
575 
576     // There should be enough left to read the binary ID size field, and the
577     // binary ID.
578     if (Remaining < sizeof(BinaryIdLen) + BinaryIdLen)
579       return make_error<InstrProfError>(
580           instrprof_error::malformed, "not enough data to read binary id data");
581 
582     // Increment by binary id length data type size.
583     BI += sizeof(BinaryIdLen);
584     if (BI > (const uint8_t *)DataBuffer->getBufferEnd())
585       return make_error<InstrProfError>(
586           instrprof_error::malformed,
587           "binary id that is read is bigger than buffer size");
588 
589     for (uint64_t I = 0; I < BinaryIdLen; I++)
590       OS << format("%02x", BI[I]);
591     OS << "\n";
592 
593     // Increment by binary id data length, rounded to the next 8 bytes. This
594     // accounts for the zero-padding after each build ID.
595     BI += RoundUp(BinaryIdLen, sizeof(uint64_t));
596     if (BI > (const uint8_t *)DataBuffer->getBufferEnd())
597       return make_error<InstrProfError>(instrprof_error::malformed);
598   }
599 
600   return success();
601 }
602 
603 namespace llvm {
604 
605 template class RawInstrProfReader<uint32_t>;
606 template class RawInstrProfReader<uint64_t>;
607 
608 } // end namespace llvm
609 
610 InstrProfLookupTrait::hash_value_type
611 InstrProfLookupTrait::ComputeHash(StringRef K) {
612   return IndexedInstrProf::ComputeHash(HashType, K);
613 }
614 
615 using data_type = InstrProfLookupTrait::data_type;
616 using offset_type = InstrProfLookupTrait::offset_type;
617 
618 bool InstrProfLookupTrait::readValueProfilingData(
619     const unsigned char *&D, const unsigned char *const End) {
620   Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
621       ValueProfData::getValueProfData(D, End, ValueProfDataEndianness);
622 
623   if (VDataPtrOrErr.takeError())
624     return false;
625 
626   VDataPtrOrErr.get()->deserializeTo(DataBuffer.back(), nullptr);
627   D += VDataPtrOrErr.get()->TotalSize;
628 
629   return true;
630 }
631 
632 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
633                                          offset_type N) {
634   using namespace support;
635 
636   // Check if the data is corrupt. If so, don't try to read it.
637   if (N % sizeof(uint64_t))
638     return data_type();
639 
640   DataBuffer.clear();
641   std::vector<uint64_t> CounterBuffer;
642 
643   const unsigned char *End = D + N;
644   while (D < End) {
645     // Read hash.
646     if (D + sizeof(uint64_t) >= End)
647       return data_type();
648     uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
649 
650     // Initialize number of counters for GET_VERSION(FormatVersion) == 1.
651     uint64_t CountsSize = N / sizeof(uint64_t) - 1;
652     // If format version is different then read the number of counters.
653     if (GET_VERSION(FormatVersion) != IndexedInstrProf::ProfVersion::Version1) {
654       if (D + sizeof(uint64_t) > End)
655         return data_type();
656       CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
657     }
658     // Read counter values.
659     if (D + CountsSize * sizeof(uint64_t) > End)
660       return data_type();
661 
662     CounterBuffer.clear();
663     CounterBuffer.reserve(CountsSize);
664     for (uint64_t J = 0; J < CountsSize; ++J)
665       CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
666 
667     DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer));
668 
669     // Read value profiling data.
670     if (GET_VERSION(FormatVersion) > IndexedInstrProf::ProfVersion::Version2 &&
671         !readValueProfilingData(D, End)) {
672       DataBuffer.clear();
673       return data_type();
674     }
675   }
676   return DataBuffer;
677 }
678 
679 template <typename HashTableImpl>
680 Error InstrProfReaderIndex<HashTableImpl>::getRecords(
681     StringRef FuncName, ArrayRef<NamedInstrProfRecord> &Data) {
682   auto Iter = HashTable->find(FuncName);
683   if (Iter == HashTable->end())
684     return make_error<InstrProfError>(instrprof_error::unknown_function);
685 
686   Data = (*Iter);
687   if (Data.empty())
688     return make_error<InstrProfError>(instrprof_error::malformed,
689                                       "profile data is empty");
690 
691   return Error::success();
692 }
693 
694 template <typename HashTableImpl>
695 Error InstrProfReaderIndex<HashTableImpl>::getRecords(
696     ArrayRef<NamedInstrProfRecord> &Data) {
697   if (atEnd())
698     return make_error<InstrProfError>(instrprof_error::eof);
699 
700   Data = *RecordIterator;
701 
702   if (Data.empty())
703     return make_error<InstrProfError>(instrprof_error::malformed,
704                                       "profile data is empty");
705 
706   return Error::success();
707 }
708 
709 template <typename HashTableImpl>
710 InstrProfReaderIndex<HashTableImpl>::InstrProfReaderIndex(
711     const unsigned char *Buckets, const unsigned char *const Payload,
712     const unsigned char *const Base, IndexedInstrProf::HashT HashType,
713     uint64_t Version) {
714   FormatVersion = Version;
715   HashTable.reset(HashTableImpl::Create(
716       Buckets, Payload, Base,
717       typename HashTableImpl::InfoType(HashType, Version)));
718   RecordIterator = HashTable->data_begin();
719 }
720 
721 namespace {
722 /// A remapper that does not apply any remappings.
723 class InstrProfReaderNullRemapper : public InstrProfReaderRemapper {
724   InstrProfReaderIndexBase &Underlying;
725 
726 public:
727   InstrProfReaderNullRemapper(InstrProfReaderIndexBase &Underlying)
728       : Underlying(Underlying) {}
729 
730   Error getRecords(StringRef FuncName,
731                    ArrayRef<NamedInstrProfRecord> &Data) override {
732     return Underlying.getRecords(FuncName, Data);
733   }
734 };
735 } // namespace
736 
737 /// A remapper that applies remappings based on a symbol remapping file.
738 template <typename HashTableImpl>
739 class llvm::InstrProfReaderItaniumRemapper
740     : public InstrProfReaderRemapper {
741 public:
742   InstrProfReaderItaniumRemapper(
743       std::unique_ptr<MemoryBuffer> RemapBuffer,
744       InstrProfReaderIndex<HashTableImpl> &Underlying)
745       : RemapBuffer(std::move(RemapBuffer)), Underlying(Underlying) {
746   }
747 
748   /// Extract the original function name from a PGO function name.
749   static StringRef extractName(StringRef Name) {
750     // We can have multiple :-separated pieces; there can be pieces both
751     // before and after the mangled name. Find the first part that starts
752     // with '_Z'; we'll assume that's the mangled name we want.
753     std::pair<StringRef, StringRef> Parts = {StringRef(), Name};
754     while (true) {
755       Parts = Parts.second.split(':');
756       if (Parts.first.startswith("_Z"))
757         return Parts.first;
758       if (Parts.second.empty())
759         return Name;
760     }
761   }
762 
763   /// Given a mangled name extracted from a PGO function name, and a new
764   /// form for that mangled name, reconstitute the name.
765   static void reconstituteName(StringRef OrigName, StringRef ExtractedName,
766                                StringRef Replacement,
767                                SmallVectorImpl<char> &Out) {
768     Out.reserve(OrigName.size() + Replacement.size() - ExtractedName.size());
769     Out.insert(Out.end(), OrigName.begin(), ExtractedName.begin());
770     Out.insert(Out.end(), Replacement.begin(), Replacement.end());
771     Out.insert(Out.end(), ExtractedName.end(), OrigName.end());
772   }
773 
774   Error populateRemappings() override {
775     if (Error E = Remappings.read(*RemapBuffer))
776       return E;
777     for (StringRef Name : Underlying.HashTable->keys()) {
778       StringRef RealName = extractName(Name);
779       if (auto Key = Remappings.insert(RealName)) {
780         // FIXME: We could theoretically map the same equivalence class to
781         // multiple names in the profile data. If that happens, we should
782         // return NamedInstrProfRecords from all of them.
783         MappedNames.insert({Key, RealName});
784       }
785     }
786     return Error::success();
787   }
788 
789   Error getRecords(StringRef FuncName,
790                    ArrayRef<NamedInstrProfRecord> &Data) override {
791     StringRef RealName = extractName(FuncName);
792     if (auto Key = Remappings.lookup(RealName)) {
793       StringRef Remapped = MappedNames.lookup(Key);
794       if (!Remapped.empty()) {
795         if (RealName.begin() == FuncName.begin() &&
796             RealName.end() == FuncName.end())
797           FuncName = Remapped;
798         else {
799           // Try rebuilding the name from the given remapping.
800           SmallString<256> Reconstituted;
801           reconstituteName(FuncName, RealName, Remapped, Reconstituted);
802           Error E = Underlying.getRecords(Reconstituted, Data);
803           if (!E)
804             return E;
805 
806           // If we failed because the name doesn't exist, fall back to asking
807           // about the original name.
808           if (Error Unhandled = handleErrors(
809                   std::move(E), [](std::unique_ptr<InstrProfError> Err) {
810                     return Err->get() == instrprof_error::unknown_function
811                                ? Error::success()
812                                : Error(std::move(Err));
813                   }))
814             return Unhandled;
815         }
816       }
817     }
818     return Underlying.getRecords(FuncName, Data);
819   }
820 
821 private:
822   /// The memory buffer containing the remapping configuration. Remappings
823   /// holds pointers into this buffer.
824   std::unique_ptr<MemoryBuffer> RemapBuffer;
825 
826   /// The mangling remapper.
827   SymbolRemappingReader Remappings;
828 
829   /// Mapping from mangled name keys to the name used for the key in the
830   /// profile data.
831   /// FIXME: Can we store a location within the on-disk hash table instead of
832   /// redoing lookup?
833   DenseMap<SymbolRemappingReader::Key, StringRef> MappedNames;
834 
835   /// The real profile data reader.
836   InstrProfReaderIndex<HashTableImpl> &Underlying;
837 };
838 
839 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
840   using namespace support;
841 
842   if (DataBuffer.getBufferSize() < 8)
843     return false;
844   uint64_t Magic =
845       endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
846   // Verify that it's magical.
847   return Magic == IndexedInstrProf::Magic;
848 }
849 
850 const unsigned char *
851 IndexedInstrProfReader::readSummary(IndexedInstrProf::ProfVersion Version,
852                                     const unsigned char *Cur, bool UseCS) {
853   using namespace IndexedInstrProf;
854   using namespace support;
855 
856   if (Version >= IndexedInstrProf::Version4) {
857     const IndexedInstrProf::Summary *SummaryInLE =
858         reinterpret_cast<const IndexedInstrProf::Summary *>(Cur);
859     uint64_t NFields =
860         endian::byte_swap<uint64_t, little>(SummaryInLE->NumSummaryFields);
861     uint64_t NEntries =
862         endian::byte_swap<uint64_t, little>(SummaryInLE->NumCutoffEntries);
863     uint32_t SummarySize =
864         IndexedInstrProf::Summary::getSize(NFields, NEntries);
865     std::unique_ptr<IndexedInstrProf::Summary> SummaryData =
866         IndexedInstrProf::allocSummary(SummarySize);
867 
868     const uint64_t *Src = reinterpret_cast<const uint64_t *>(SummaryInLE);
869     uint64_t *Dst = reinterpret_cast<uint64_t *>(SummaryData.get());
870     for (unsigned I = 0; I < SummarySize / sizeof(uint64_t); I++)
871       Dst[I] = endian::byte_swap<uint64_t, little>(Src[I]);
872 
873     SummaryEntryVector DetailedSummary;
874     for (unsigned I = 0; I < SummaryData->NumCutoffEntries; I++) {
875       const IndexedInstrProf::Summary::Entry &Ent = SummaryData->getEntry(I);
876       DetailedSummary.emplace_back((uint32_t)Ent.Cutoff, Ent.MinBlockCount,
877                                    Ent.NumBlocks);
878     }
879     std::unique_ptr<llvm::ProfileSummary> &Summary =
880         UseCS ? this->CS_Summary : this->Summary;
881 
882     // initialize InstrProfSummary using the SummaryData from disk.
883     Summary = std::make_unique<ProfileSummary>(
884         UseCS ? ProfileSummary::PSK_CSInstr : ProfileSummary::PSK_Instr,
885         DetailedSummary, SummaryData->get(Summary::TotalBlockCount),
886         SummaryData->get(Summary::MaxBlockCount),
887         SummaryData->get(Summary::MaxInternalBlockCount),
888         SummaryData->get(Summary::MaxFunctionCount),
889         SummaryData->get(Summary::TotalNumBlocks),
890         SummaryData->get(Summary::TotalNumFunctions));
891     return Cur + SummarySize;
892   } else {
893     // The older versions do not support a profile summary. This just computes
894     // an empty summary, which will not result in accurate hot/cold detection.
895     // We would need to call addRecord for all NamedInstrProfRecords to get the
896     // correct summary. However, this version is old (prior to early 2016) and
897     // has not been supporting an accurate summary for several years.
898     InstrProfSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs);
899     Summary = Builder.getSummary();
900     return Cur;
901   }
902 }
903 
904 Error IndexedInstrProfReader::readHeader() {
905   using namespace support;
906 
907   const unsigned char *Start =
908       (const unsigned char *)DataBuffer->getBufferStart();
909   const unsigned char *Cur = Start;
910   if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
911     return error(instrprof_error::truncated);
912 
913   auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
914   Cur += sizeof(IndexedInstrProf::Header);
915 
916   // Check the magic number.
917   uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
918   if (Magic != IndexedInstrProf::Magic)
919     return error(instrprof_error::bad_magic);
920 
921   // Read the version.
922   uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
923   if (GET_VERSION(FormatVersion) >
924       IndexedInstrProf::ProfVersion::CurrentVersion)
925     return error(instrprof_error::unsupported_version);
926 
927   Cur = readSummary((IndexedInstrProf::ProfVersion)FormatVersion, Cur,
928                     /* UseCS */ false);
929   if (FormatVersion & VARIANT_MASK_CSIR_PROF)
930     Cur = readSummary((IndexedInstrProf::ProfVersion)FormatVersion, Cur,
931                       /* UseCS */ true);
932 
933   // Read the hash type and start offset.
934   IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
935       endian::byte_swap<uint64_t, little>(Header->HashType));
936   if (HashType > IndexedInstrProf::HashT::Last)
937     return error(instrprof_error::unsupported_hash_type);
938 
939   uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
940 
941   // The rest of the file is an on disk hash table.
942   auto IndexPtr =
943       std::make_unique<InstrProfReaderIndex<OnDiskHashTableImplV3>>(
944           Start + HashOffset, Cur, Start, HashType, FormatVersion);
945 
946   // Load the remapping table now if requested.
947   if (RemappingBuffer) {
948     Remapper = std::make_unique<
949         InstrProfReaderItaniumRemapper<OnDiskHashTableImplV3>>(
950         std::move(RemappingBuffer), *IndexPtr);
951     if (Error E = Remapper->populateRemappings())
952       return E;
953   } else {
954     Remapper = std::make_unique<InstrProfReaderNullRemapper>(*IndexPtr);
955   }
956   Index = std::move(IndexPtr);
957 
958   return success();
959 }
960 
961 InstrProfSymtab &IndexedInstrProfReader::getSymtab() {
962   if (Symtab.get())
963     return *Symtab.get();
964 
965   std::unique_ptr<InstrProfSymtab> NewSymtab = std::make_unique<InstrProfSymtab>();
966   if (Error E = Index->populateSymtab(*NewSymtab.get())) {
967     consumeError(error(InstrProfError::take(std::move(E))));
968   }
969 
970   Symtab = std::move(NewSymtab);
971   return *Symtab.get();
972 }
973 
974 Expected<InstrProfRecord>
975 IndexedInstrProfReader::getInstrProfRecord(StringRef FuncName,
976                                            uint64_t FuncHash) {
977   ArrayRef<NamedInstrProfRecord> Data;
978   Error Err = Remapper->getRecords(FuncName, Data);
979   if (Err)
980     return std::move(Err);
981   // Found it. Look for counters with the right hash.
982   for (const NamedInstrProfRecord &I : Data) {
983     // Check for a match and fill the vector if there is one.
984     if (I.Hash == FuncHash)
985       return std::move(I);
986   }
987   return error(instrprof_error::hash_mismatch);
988 }
989 
990 Error IndexedInstrProfReader::getFunctionCounts(StringRef FuncName,
991                                                 uint64_t FuncHash,
992                                                 std::vector<uint64_t> &Counts) {
993   Expected<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash);
994   if (Error E = Record.takeError())
995     return error(std::move(E));
996 
997   Counts = Record.get().Counts;
998   return success();
999 }
1000 
1001 Error IndexedInstrProfReader::readNextRecord(NamedInstrProfRecord &Record) {
1002   ArrayRef<NamedInstrProfRecord> Data;
1003 
1004   Error E = Index->getRecords(Data);
1005   if (E)
1006     return error(std::move(E));
1007 
1008   Record = Data[RecordIndex++];
1009   if (RecordIndex >= Data.size()) {
1010     Index->advanceToNextKey();
1011     RecordIndex = 0;
1012   }
1013   return success();
1014 }
1015 
1016 void InstrProfReader::accumulateCounts(CountSumOrPercent &Sum, bool IsCS) {
1017   uint64_t NumFuncs = 0;
1018   for (const auto &Func : *this) {
1019     if (isIRLevelProfile()) {
1020       bool FuncIsCS = NamedInstrProfRecord::hasCSFlagInHash(Func.Hash);
1021       if (FuncIsCS != IsCS)
1022         continue;
1023     }
1024     Func.accumulateCounts(Sum);
1025     ++NumFuncs;
1026   }
1027   Sum.NumEntries = NumFuncs;
1028 }
1029