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