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 "InstrProfIndexed.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include <cassert>
19 
20 using namespace llvm;
21 
22 static ErrorOr<std::unique_ptr<MemoryBuffer>>
23 setupMemoryBuffer(std::string Path) {
24   ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
25       MemoryBuffer::getFileOrSTDIN(Path);
26   if (std::error_code EC = BufferOrErr.getError())
27     return EC;
28   return std::move(BufferOrErr.get());
29 }
30 
31 static std::error_code initializeReader(InstrProfReader &Reader) {
32   return Reader.readHeader();
33 }
34 
35 ErrorOr<std::unique_ptr<InstrProfReader>>
36 InstrProfReader::create(std::string Path) {
37   // Set up the buffer to read.
38   auto BufferOrError = setupMemoryBuffer(Path);
39   if (std::error_code EC = BufferOrError.getError())
40     return EC;
41   return InstrProfReader::create(std::move(BufferOrError.get()));
42 }
43 
44 ErrorOr<std::unique_ptr<InstrProfReader>>
45 InstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer) {
46   // Sanity check the buffer.
47   if (Buffer->getBufferSize() > std::numeric_limits<unsigned>::max())
48     return instrprof_error::too_large;
49 
50   std::unique_ptr<InstrProfReader> Result;
51   // Create the reader.
52   if (IndexedInstrProfReader::hasFormat(*Buffer))
53     Result.reset(new IndexedInstrProfReader(std::move(Buffer)));
54   else if (RawInstrProfReader64::hasFormat(*Buffer))
55     Result.reset(new RawInstrProfReader64(std::move(Buffer)));
56   else if (RawInstrProfReader32::hasFormat(*Buffer))
57     Result.reset(new RawInstrProfReader32(std::move(Buffer)));
58   else
59     Result.reset(new TextInstrProfReader(std::move(Buffer)));
60 
61   // Initialize the reader and return the result.
62   if (std::error_code EC = initializeReader(*Result))
63     return EC;
64 
65   return std::move(Result);
66 }
67 
68 ErrorOr<std::unique_ptr<IndexedInstrProfReader>>
69 IndexedInstrProfReader::create(std::string Path) {
70   // Set up the buffer to read.
71   auto BufferOrError = setupMemoryBuffer(Path);
72   if (std::error_code EC = BufferOrError.getError())
73     return EC;
74   return IndexedInstrProfReader::create(std::move(BufferOrError.get()));
75 }
76 
77 
78 ErrorOr<std::unique_ptr<IndexedInstrProfReader>>
79 IndexedInstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer) {
80   // Sanity check the buffer.
81   if (Buffer->getBufferSize() > std::numeric_limits<unsigned>::max())
82     return instrprof_error::too_large;
83 
84   // Create the reader.
85   if (!IndexedInstrProfReader::hasFormat(*Buffer))
86     return instrprof_error::bad_magic;
87   auto Result = llvm::make_unique<IndexedInstrProfReader>(std::move(Buffer));
88 
89   // Initialize the reader and return the result.
90   if (std::error_code EC = initializeReader(*Result))
91     return EC;
92 
93   return std::move(Result);
94 }
95 
96 void InstrProfIterator::Increment() {
97   if (Reader->readNextRecord(Record))
98     *this = InstrProfIterator();
99 }
100 
101 std::error_code TextInstrProfReader::readNextRecord(InstrProfRecord &Record) {
102   // Skip empty lines and comments.
103   while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
104     ++Line;
105   // If we hit EOF while looking for a name, we're done.
106   if (Line.is_at_end())
107     return error(instrprof_error::eof);
108 
109   // Read the function name.
110   Record.Name = *Line++;
111 
112   // Read the function hash.
113   if (Line.is_at_end())
114     return error(instrprof_error::truncated);
115   if ((Line++)->getAsInteger(0, Record.Hash))
116     return error(instrprof_error::malformed);
117 
118   // Read the number of counters.
119   uint64_t NumCounters;
120   if (Line.is_at_end())
121     return error(instrprof_error::truncated);
122   if ((Line++)->getAsInteger(10, NumCounters))
123     return error(instrprof_error::malformed);
124   if (NumCounters == 0)
125     return error(instrprof_error::malformed);
126 
127   // Read each counter and fill our internal storage with the values.
128   Record.Counts.clear();
129   Record.Counts.reserve(NumCounters);
130   for (uint64_t I = 0; I < NumCounters; ++I) {
131     if (Line.is_at_end())
132       return error(instrprof_error::truncated);
133     uint64_t Count;
134     if ((Line++)->getAsInteger(10, Count))
135       return error(instrprof_error::malformed);
136     Record.Counts.push_back(Count);
137   }
138 
139   return success();
140 }
141 
142 template <class IntPtrT>
143 static uint64_t getRawMagic();
144 
145 template <>
146 uint64_t getRawMagic<uint64_t>() {
147   return
148     uint64_t(255) << 56 |
149     uint64_t('l') << 48 |
150     uint64_t('p') << 40 |
151     uint64_t('r') << 32 |
152     uint64_t('o') << 24 |
153     uint64_t('f') << 16 |
154     uint64_t('r') <<  8 |
155     uint64_t(129);
156 }
157 
158 template <>
159 uint64_t getRawMagic<uint32_t>() {
160   return
161     uint64_t(255) << 56 |
162     uint64_t('l') << 48 |
163     uint64_t('p') << 40 |
164     uint64_t('r') << 32 |
165     uint64_t('o') << 24 |
166     uint64_t('f') << 16 |
167     uint64_t('R') <<  8 |
168     uint64_t(129);
169 }
170 
171 template <class IntPtrT>
172 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
173   if (DataBuffer.getBufferSize() < sizeof(uint64_t))
174     return false;
175   uint64_t Magic =
176     *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
177   return getRawMagic<IntPtrT>() == Magic ||
178     sys::getSwappedBytes(getRawMagic<IntPtrT>()) == Magic;
179 }
180 
181 template <class IntPtrT>
182 std::error_code RawInstrProfReader<IntPtrT>::readHeader() {
183   if (!hasFormat(*DataBuffer))
184     return error(instrprof_error::bad_magic);
185   if (DataBuffer->getBufferSize() < sizeof(RawHeader))
186     return error(instrprof_error::bad_header);
187   auto *Header =
188     reinterpret_cast<const RawHeader *>(DataBuffer->getBufferStart());
189   ShouldSwapBytes = Header->Magic != getRawMagic<IntPtrT>();
190   return readHeader(*Header);
191 }
192 
193 template <class IntPtrT>
194 std::error_code
195 RawInstrProfReader<IntPtrT>::readNextHeader(const char *CurrentPos) {
196   const char *End = DataBuffer->getBufferEnd();
197   // Skip zero padding between profiles.
198   while (CurrentPos != End && *CurrentPos == 0)
199     ++CurrentPos;
200   // If there's nothing left, we're done.
201   if (CurrentPos == End)
202     return instrprof_error::eof;
203   // If there isn't enough space for another header, this is probably just
204   // garbage at the end of the file.
205   if (CurrentPos + sizeof(RawHeader) > End)
206     return instrprof_error::malformed;
207   // The writer ensures each profile is padded to start at an aligned address.
208   if (reinterpret_cast<size_t>(CurrentPos) % alignOf<uint64_t>())
209     return instrprof_error::malformed;
210   // The magic should have the same byte order as in the previous header.
211   uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos);
212   if (Magic != swap(getRawMagic<IntPtrT>()))
213     return instrprof_error::bad_magic;
214 
215   // There's another profile to read, so we need to process the header.
216   auto *Header = reinterpret_cast<const RawHeader *>(CurrentPos);
217   return readHeader(*Header);
218 }
219 
220 static uint64_t getRawVersion() {
221   return 1;
222 }
223 
224 template <class IntPtrT>
225 std::error_code
226 RawInstrProfReader<IntPtrT>::readHeader(const RawHeader &Header) {
227   if (swap(Header.Version) != getRawVersion())
228     return error(instrprof_error::unsupported_version);
229 
230   CountersDelta = swap(Header.CountersDelta);
231   NamesDelta = swap(Header.NamesDelta);
232   auto DataSize = swap(Header.DataSize);
233   auto CountersSize = swap(Header.CountersSize);
234   auto NamesSize = swap(Header.NamesSize);
235 
236   ptrdiff_t DataOffset = sizeof(RawHeader);
237   ptrdiff_t CountersOffset = DataOffset + sizeof(ProfileData) * DataSize;
238   ptrdiff_t NamesOffset = CountersOffset + sizeof(uint64_t) * CountersSize;
239   size_t ProfileSize = NamesOffset + sizeof(char) * NamesSize;
240 
241   auto *Start = reinterpret_cast<const char *>(&Header);
242   if (Start + ProfileSize > DataBuffer->getBufferEnd())
243     return error(instrprof_error::bad_header);
244 
245   Data = reinterpret_cast<const ProfileData *>(Start + DataOffset);
246   DataEnd = Data + DataSize;
247   CountersStart = reinterpret_cast<const uint64_t *>(Start + CountersOffset);
248   NamesStart = Start + NamesOffset;
249   ProfileEnd = Start + ProfileSize;
250 
251   return success();
252 }
253 
254 template <class IntPtrT>
255 std::error_code
256 RawInstrProfReader<IntPtrT>::readNextRecord(InstrProfRecord &Record) {
257   if (Data == DataEnd)
258     if (std::error_code EC = readNextHeader(ProfileEnd))
259       return EC;
260 
261   // Get the raw data.
262   StringRef RawName(getName(Data->NamePtr), swap(Data->NameSize));
263   uint32_t NumCounters = swap(Data->NumCounters);
264   if (NumCounters == 0)
265     return error(instrprof_error::malformed);
266   auto RawCounts = makeArrayRef(getCounter(Data->CounterPtr), NumCounters);
267 
268   // Check bounds.
269   auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
270   if (RawName.data() < NamesStart ||
271       RawName.data() + RawName.size() > DataBuffer->getBufferEnd() ||
272       RawCounts.data() < CountersStart ||
273       RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
274     return error(instrprof_error::malformed);
275 
276   // Store the data in Record, byte-swapping as necessary.
277   Record.Hash = swap(Data->FuncHash);
278   Record.Name = RawName;
279   if (ShouldSwapBytes) {
280     Record.Counts.clear();
281     Record.Counts.reserve(RawCounts.size());
282     for (uint64_t Count : RawCounts)
283       Record.Counts.push_back(swap(Count));
284   } else
285     Record.Counts = RawCounts;
286 
287   // Iterate.
288   ++Data;
289   return success();
290 }
291 
292 namespace llvm {
293 template class RawInstrProfReader<uint32_t>;
294 template class RawInstrProfReader<uint64_t>;
295 }
296 
297 InstrProfLookupTrait::hash_value_type
298 InstrProfLookupTrait::ComputeHash(StringRef K) {
299   return IndexedInstrProf::ComputeHash(HashType, K);
300 }
301 
302 typedef InstrProfLookupTrait::data_type data_type;
303 typedef InstrProfLookupTrait::offset_type offset_type;
304 
305 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
306                                          offset_type N) {
307 
308   // Check if the data is corrupt. If so, don't try to read it.
309   if (N % sizeof(uint64_t))
310     return data_type();
311 
312   DataBuffer.clear();
313   uint64_t NumCounts;
314   uint64_t NumEntries = N / sizeof(uint64_t);
315   std::vector<uint64_t> CounterBuffer;
316   for (uint64_t I = 0; I < NumEntries; I += NumCounts) {
317     using namespace support;
318     // The function hash comes first.
319     uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
320 
321     if (++I >= NumEntries)
322       return data_type();
323 
324     // In v1, we have at least one count.
325     // Later, we have the number of counts.
326     NumCounts = (1 == FormatVersion)
327                     ? NumEntries - I
328                     : endian::readNext<uint64_t, little, unaligned>(D);
329     if (1 != FormatVersion)
330       ++I;
331 
332     // If we have more counts than data, this is bogus.
333     if (I + NumCounts > NumEntries)
334       return data_type();
335 
336     CounterBuffer.clear();
337     for (unsigned J = 0; J < NumCounts; ++J)
338       CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
339 
340     DataBuffer.push_back(InstrProfRecord(K, Hash, std::move(CounterBuffer)));
341   }
342   return DataBuffer;
343 }
344 
345 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
346   if (DataBuffer.getBufferSize() < 8)
347     return false;
348   using namespace support;
349   uint64_t Magic =
350       endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
351   return Magic == IndexedInstrProf::Magic;
352 }
353 
354 std::error_code IndexedInstrProfReader::readHeader() {
355   const unsigned char *Start =
356       (const unsigned char *)DataBuffer->getBufferStart();
357   const unsigned char *Cur = Start;
358   if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
359     return error(instrprof_error::truncated);
360 
361   using namespace support;
362 
363   // Check the magic number.
364   uint64_t Magic = endian::readNext<uint64_t, little, unaligned>(Cur);
365   if (Magic != IndexedInstrProf::Magic)
366     return error(instrprof_error::bad_magic);
367 
368   // Read the version.
369   FormatVersion = endian::readNext<uint64_t, little, unaligned>(Cur);
370   if (FormatVersion > IndexedInstrProf::Version)
371     return error(instrprof_error::unsupported_version);
372 
373   // Read the maximal function count.
374   MaxFunctionCount = endian::readNext<uint64_t, little, unaligned>(Cur);
375 
376   // Read the hash type and start offset.
377   IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
378       endian::readNext<uint64_t, little, unaligned>(Cur));
379   if (HashType > IndexedInstrProf::HashT::Last)
380     return error(instrprof_error::unsupported_hash_type);
381   uint64_t HashOffset = endian::readNext<uint64_t, little, unaligned>(Cur);
382 
383   // The rest of the file is an on disk hash table.
384   Index.reset(InstrProfReaderIndex::Create(
385       Start + HashOffset, Cur, Start,
386       InstrProfLookupTrait(HashType, FormatVersion)));
387   // Set up our iterator for readNextRecord.
388   RecordIterator = Index->data_begin();
389 
390   return success();
391 }
392 
393 std::error_code IndexedInstrProfReader::getFunctionCounts(
394     StringRef FuncName, uint64_t FuncHash, std::vector<uint64_t> &Counts) {
395   auto Iter = Index->find(FuncName);
396   if (Iter == Index->end())
397     return error(instrprof_error::unknown_function);
398 
399   // Found it. Look for counters with the right hash.
400   ArrayRef<InstrProfRecord> Data = (*Iter);
401   if (Data.empty())
402     return error(instrprof_error::malformed);
403 
404   for (unsigned I = 0, E = Data.size(); I < E; ++I) {
405     // Check for a match and fill the vector if there is one.
406     if (Data[I].Hash == FuncHash) {
407       Counts = Data[I].Counts;
408       return success();
409     }
410   }
411   return error(instrprof_error::hash_mismatch);
412 }
413 
414 std::error_code
415 IndexedInstrProfReader::readNextRecord(InstrProfRecord &Record) {
416   // Are we out of records?
417   if (RecordIterator == Index->data_end())
418     return error(instrprof_error::eof);
419 
420   if ((*RecordIterator).empty())
421     return error(instrprof_error::malformed);
422 
423   static unsigned RecordIndex = 0;
424   ArrayRef<InstrProfRecord> Data = (*RecordIterator);
425   Record = Data[RecordIndex++];
426   if (RecordIndex >= Data.size()) {
427     ++RecordIterator;
428     RecordIndex = 0;
429   }
430   return success();
431 }
432