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