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 "llvm/ProfileData/InstrProf.h"
19 #include <cassert>
20 
21 using namespace llvm;
22 
23 static ErrorOr<std::unique_ptr<MemoryBuffer>>
setupMemoryBuffer(std::string Path)24 setupMemoryBuffer(std::string Path) {
25   ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
26       MemoryBuffer::getFileOrSTDIN(Path);
27   if (std::error_code EC = BufferOrErr.getError())
28     return EC;
29   return std::move(BufferOrErr.get());
30 }
31 
initializeReader(InstrProfReader & Reader)32 static std::error_code initializeReader(InstrProfReader &Reader) {
33   return Reader.readHeader();
34 }
35 
36 ErrorOr<std::unique_ptr<InstrProfReader>>
create(std::string Path)37 InstrProfReader::create(std::string Path) {
38   // Set up the buffer to read.
39   auto BufferOrError = setupMemoryBuffer(Path);
40   if (std::error_code EC = BufferOrError.getError())
41     return EC;
42   return InstrProfReader::create(std::move(BufferOrError.get()));
43 }
44 
45 ErrorOr<std::unique_ptr<InstrProfReader>>
create(std::unique_ptr<MemoryBuffer> Buffer)46 InstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer) {
47   // Sanity check the buffer.
48   if (Buffer->getBufferSize() > std::numeric_limits<unsigned>::max())
49     return instrprof_error::too_large;
50 
51   std::unique_ptr<InstrProfReader> Result;
52   // Create the reader.
53   if (IndexedInstrProfReader::hasFormat(*Buffer))
54     Result.reset(new IndexedInstrProfReader(std::move(Buffer)));
55   else if (RawInstrProfReader64::hasFormat(*Buffer))
56     Result.reset(new RawInstrProfReader64(std::move(Buffer)));
57   else if (RawInstrProfReader32::hasFormat(*Buffer))
58     Result.reset(new RawInstrProfReader32(std::move(Buffer)));
59   else
60     Result.reset(new TextInstrProfReader(std::move(Buffer)));
61 
62   // Initialize the reader and return the result.
63   if (std::error_code EC = initializeReader(*Result))
64     return EC;
65 
66   return std::move(Result);
67 }
68 
69 ErrorOr<std::unique_ptr<IndexedInstrProfReader>>
create(std::string Path)70 IndexedInstrProfReader::create(std::string Path) {
71   // Set up the buffer to read.
72   auto BufferOrError = setupMemoryBuffer(Path);
73   if (std::error_code EC = BufferOrError.getError())
74     return EC;
75   return IndexedInstrProfReader::create(std::move(BufferOrError.get()));
76 }
77 
78 
79 ErrorOr<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 instrprof_error::too_large;
84 
85   // Create the reader.
86   if (!IndexedInstrProfReader::hasFormat(*Buffer))
87     return 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 (std::error_code EC = initializeReader(*Result))
92     return EC;
93 
94   return std::move(Result);
95 }
96 
Increment()97 void InstrProfIterator::Increment() {
98   if (Reader->readNextRecord(Record))
99     *this = InstrProfIterator();
100 }
101 
readNextRecord(InstrProfRecord & Record)102 std::error_code TextInstrProfReader::readNextRecord(InstrProfRecord &Record) {
103   // Skip empty lines and comments.
104   while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
105     ++Line;
106   // If we hit EOF while looking for a name, we're done.
107   if (Line.is_at_end())
108     return error(instrprof_error::eof);
109 
110   // Read the function name.
111   Record.Name = *Line++;
112 
113   // Read the function hash.
114   if (Line.is_at_end())
115     return error(instrprof_error::truncated);
116   if ((Line++)->getAsInteger(0, Record.Hash))
117     return error(instrprof_error::malformed);
118 
119   // Read the number of counters.
120   uint64_t NumCounters;
121   if (Line.is_at_end())
122     return error(instrprof_error::truncated);
123   if ((Line++)->getAsInteger(10, NumCounters))
124     return error(instrprof_error::malformed);
125   if (NumCounters == 0)
126     return error(instrprof_error::malformed);
127 
128   // Read each counter and fill our internal storage with the values.
129   Counts.clear();
130   Counts.reserve(NumCounters);
131   for (uint64_t I = 0; I < NumCounters; ++I) {
132     if (Line.is_at_end())
133       return error(instrprof_error::truncated);
134     uint64_t Count;
135     if ((Line++)->getAsInteger(10, Count))
136       return error(instrprof_error::malformed);
137     Counts.push_back(Count);
138   }
139   // Give the record a reference to our internal counter storage.
140   Record.Counts = Counts;
141 
142   return success();
143 }
144 
145 template <class IntPtrT>
146 static uint64_t getRawMagic();
147 
148 template <>
getRawMagic()149 uint64_t getRawMagic<uint64_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 <>
getRawMagic()162 uint64_t getRawMagic<uint32_t>() {
163   return
164     uint64_t(255) << 56 |
165     uint64_t('l') << 48 |
166     uint64_t('p') << 40 |
167     uint64_t('r') << 32 |
168     uint64_t('o') << 24 |
169     uint64_t('f') << 16 |
170     uint64_t('R') <<  8 |
171     uint64_t(129);
172 }
173 
174 template <class IntPtrT>
hasFormat(const MemoryBuffer & DataBuffer)175 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
176   if (DataBuffer.getBufferSize() < sizeof(uint64_t))
177     return false;
178   uint64_t Magic =
179     *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
180   return getRawMagic<IntPtrT>() == Magic ||
181     sys::getSwappedBytes(getRawMagic<IntPtrT>()) == Magic;
182 }
183 
184 template <class IntPtrT>
readHeader()185 std::error_code RawInstrProfReader<IntPtrT>::readHeader() {
186   if (!hasFormat(*DataBuffer))
187     return error(instrprof_error::bad_magic);
188   if (DataBuffer->getBufferSize() < sizeof(RawHeader))
189     return error(instrprof_error::bad_header);
190   auto *Header =
191     reinterpret_cast<const RawHeader *>(DataBuffer->getBufferStart());
192   ShouldSwapBytes = Header->Magic != getRawMagic<IntPtrT>();
193   return readHeader(*Header);
194 }
195 
196 template <class IntPtrT>
197 std::error_code
readNextHeader(const char * CurrentPos)198 RawInstrProfReader<IntPtrT>::readNextHeader(const char *CurrentPos) {
199   const char *End = DataBuffer->getBufferEnd();
200   // Skip zero padding between profiles.
201   while (CurrentPos != End && *CurrentPos == 0)
202     ++CurrentPos;
203   // If there's nothing left, we're done.
204   if (CurrentPos == End)
205     return instrprof_error::eof;
206   // If there isn't enough space for another header, this is probably just
207   // garbage at the end of the file.
208   if (CurrentPos + sizeof(RawHeader) > End)
209     return instrprof_error::malformed;
210   // The writer ensures each profile is padded to start at an aligned address.
211   if (reinterpret_cast<size_t>(CurrentPos) % alignOf<uint64_t>())
212     return instrprof_error::malformed;
213   // The magic should have the same byte order as in the previous header.
214   uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos);
215   if (Magic != swap(getRawMagic<IntPtrT>()))
216     return instrprof_error::bad_magic;
217 
218   // There's another profile to read, so we need to process the header.
219   auto *Header = reinterpret_cast<const RawHeader *>(CurrentPos);
220   return readHeader(*Header);
221 }
222 
getRawVersion()223 static uint64_t getRawVersion() {
224   return 1;
225 }
226 
227 template <class IntPtrT>
228 std::error_code
readHeader(const RawHeader & Header)229 RawInstrProfReader<IntPtrT>::readHeader(const RawHeader &Header) {
230   if (swap(Header.Version) != getRawVersion())
231     return error(instrprof_error::unsupported_version);
232 
233   CountersDelta = swap(Header.CountersDelta);
234   NamesDelta = swap(Header.NamesDelta);
235   auto DataSize = swap(Header.DataSize);
236   auto CountersSize = swap(Header.CountersSize);
237   auto NamesSize = swap(Header.NamesSize);
238 
239   ptrdiff_t DataOffset = sizeof(RawHeader);
240   ptrdiff_t CountersOffset = DataOffset + sizeof(ProfileData) * DataSize;
241   ptrdiff_t NamesOffset = CountersOffset + sizeof(uint64_t) * CountersSize;
242   size_t ProfileSize = NamesOffset + sizeof(char) * NamesSize;
243 
244   auto *Start = reinterpret_cast<const char *>(&Header);
245   if (Start + ProfileSize > DataBuffer->getBufferEnd())
246     return error(instrprof_error::bad_header);
247 
248   Data = reinterpret_cast<const ProfileData *>(Start + DataOffset);
249   DataEnd = Data + DataSize;
250   CountersStart = reinterpret_cast<const uint64_t *>(Start + CountersOffset);
251   NamesStart = Start + NamesOffset;
252   ProfileEnd = Start + ProfileSize;
253 
254   return success();
255 }
256 
257 template <class IntPtrT>
258 std::error_code
readNextRecord(InstrProfRecord & Record)259 RawInstrProfReader<IntPtrT>::readNextRecord(InstrProfRecord &Record) {
260   if (Data == DataEnd)
261     if (std::error_code EC = readNextHeader(ProfileEnd))
262       return EC;
263 
264   // Get the raw data.
265   StringRef RawName(getName(Data->NamePtr), swap(Data->NameSize));
266   uint32_t NumCounters = swap(Data->NumCounters);
267   if (NumCounters == 0)
268     return error(instrprof_error::malformed);
269   auto RawCounts = makeArrayRef(getCounter(Data->CounterPtr), NumCounters);
270 
271   // Check bounds.
272   auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
273   if (RawName.data() < NamesStart ||
274       RawName.data() + RawName.size() > DataBuffer->getBufferEnd() ||
275       RawCounts.data() < CountersStart ||
276       RawCounts.data() + RawCounts.size() > NamesStartAsCounter)
277     return error(instrprof_error::malformed);
278 
279   // Store the data in Record, byte-swapping as necessary.
280   Record.Hash = swap(Data->FuncHash);
281   Record.Name = RawName;
282   if (ShouldSwapBytes) {
283     Counts.clear();
284     Counts.reserve(RawCounts.size());
285     for (uint64_t Count : RawCounts)
286       Counts.push_back(swap(Count));
287     Record.Counts = Counts;
288   } else
289     Record.Counts = RawCounts;
290 
291   // Iterate.
292   ++Data;
293   return success();
294 }
295 
296 namespace llvm {
297 template class RawInstrProfReader<uint32_t>;
298 template class RawInstrProfReader<uint64_t>;
299 }
300 
301 InstrProfLookupTrait::hash_value_type
ComputeHash(StringRef K)302 InstrProfLookupTrait::ComputeHash(StringRef K) {
303   return IndexedInstrProf::ComputeHash(HashType, K);
304 }
305 
hasFormat(const MemoryBuffer & DataBuffer)306 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
307   if (DataBuffer.getBufferSize() < 8)
308     return false;
309   using namespace support;
310   uint64_t Magic =
311       endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
312   return Magic == IndexedInstrProf::Magic;
313 }
314 
readHeader()315 std::error_code IndexedInstrProfReader::readHeader() {
316   const unsigned char *Start =
317       (const unsigned char *)DataBuffer->getBufferStart();
318   const unsigned char *Cur = Start;
319   if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
320     return error(instrprof_error::truncated);
321 
322   using namespace support;
323 
324   // Check the magic number.
325   uint64_t Magic = endian::readNext<uint64_t, little, unaligned>(Cur);
326   if (Magic != IndexedInstrProf::Magic)
327     return error(instrprof_error::bad_magic);
328 
329   // Read the version.
330   FormatVersion = endian::readNext<uint64_t, little, unaligned>(Cur);
331   if (FormatVersion > IndexedInstrProf::Version)
332     return error(instrprof_error::unsupported_version);
333 
334   // Read the maximal function count.
335   MaxFunctionCount = endian::readNext<uint64_t, little, unaligned>(Cur);
336 
337   // Read the hash type and start offset.
338   IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
339       endian::readNext<uint64_t, little, unaligned>(Cur));
340   if (HashType > IndexedInstrProf::HashT::Last)
341     return error(instrprof_error::unsupported_hash_type);
342   uint64_t HashOffset = endian::readNext<uint64_t, little, unaligned>(Cur);
343 
344   // The rest of the file is an on disk hash table.
345   Index.reset(InstrProfReaderIndex::Create(Start + HashOffset, Cur, Start,
346                                            InstrProfLookupTrait(HashType)));
347   // Set up our iterator for readNextRecord.
348   RecordIterator = Index->data_begin();
349 
350   return success();
351 }
352 
getFunctionCounts(StringRef FuncName,uint64_t FuncHash,std::vector<uint64_t> & Counts)353 std::error_code IndexedInstrProfReader::getFunctionCounts(
354     StringRef FuncName, uint64_t FuncHash, std::vector<uint64_t> &Counts) {
355   auto Iter = Index->find(FuncName);
356   if (Iter == Index->end())
357     return error(instrprof_error::unknown_function);
358 
359   // Found it. Look for counters with the right hash.
360   ArrayRef<uint64_t> Data = (*Iter).Data;
361   uint64_t NumCounts;
362   for (uint64_t I = 0, E = Data.size(); I != E; I += NumCounts) {
363     // The function hash comes first.
364     uint64_t FoundHash = Data[I++];
365     // In v1, we have at least one count. Later, we have the number of counts.
366     if (I == E)
367       return error(instrprof_error::malformed);
368     NumCounts = FormatVersion == 1 ? E - I : Data[I++];
369     // If we have more counts than data, this is bogus.
370     if (I + NumCounts > E)
371       return error(instrprof_error::malformed);
372     // Check for a match and fill the vector if there is one.
373     if (FoundHash == FuncHash) {
374       Counts = Data.slice(I, NumCounts);
375       return success();
376     }
377   }
378   return error(instrprof_error::hash_mismatch);
379 }
380 
381 std::error_code
readNextRecord(InstrProfRecord & Record)382 IndexedInstrProfReader::readNextRecord(InstrProfRecord &Record) {
383   // Are we out of records?
384   if (RecordIterator == Index->data_end())
385     return error(instrprof_error::eof);
386 
387   // Record the current function name.
388   Record.Name = (*RecordIterator).Name;
389 
390   ArrayRef<uint64_t> Data = (*RecordIterator).Data;
391   // Valid data starts with a hash and either a count or the number of counts.
392   if (CurrentOffset + 1 > Data.size())
393     return error(instrprof_error::malformed);
394   // First we have a function hash.
395   Record.Hash = Data[CurrentOffset++];
396   // In version 1 we knew the number of counters implicitly, but in newer
397   // versions we store the number of counters next.
398   uint64_t NumCounts =
399       FormatVersion == 1 ? Data.size() - CurrentOffset : Data[CurrentOffset++];
400   if (CurrentOffset + NumCounts > Data.size())
401     return error(instrprof_error::malformed);
402   // And finally the counts themselves.
403   Record.Counts = Data.slice(CurrentOffset, NumCounts);
404 
405   // If we've exhausted this function's data, increment the record.
406   CurrentOffset += NumCounts;
407   if (CurrentOffset == Data.size()) {
408     ++RecordIterator;
409     CurrentOffset = 0;
410   }
411 
412   return success();
413 }
414