1 //===- SampleProfReader.cpp - Read LLVM sample profile data ---------------===//
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 implements the class that reads LLVM sample profiles. It
11 // supports three file formats: text, binary and gcov.
12 //
13 // The textual representation is useful for debugging and testing purposes. The
14 // binary representation is more compact, resulting in smaller file sizes.
15 //
16 // The gcov encoding is the one generated by GCC's AutoFDO profile creation
17 // tool (https://github.com/google/autofdo)
18 //
19 // All three encodings can be used interchangeably as an input sample profile.
20 //
21 //===----------------------------------------------------------------------===//
22
23 #include "llvm/ProfileData/SampleProfReader.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/IR/ProfileSummary.h"
28 #include "llvm/ProfileData/ProfileCommon.h"
29 #include "llvm/ProfileData/SampleProf.h"
30 #include "llvm/Support/ErrorOr.h"
31 #include "llvm/Support/LEB128.h"
32 #include "llvm/Support/LineIterator.h"
33 #include "llvm/Support/MemoryBuffer.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include <algorithm>
36 #include <cstddef>
37 #include <cstdint>
38 #include <limits>
39 #include <memory>
40 #include <system_error>
41 #include <vector>
42
43 using namespace llvm;
44 using namespace sampleprof;
45
46 /// Dump the function profile for \p FName.
47 ///
48 /// \param FName Name of the function to print.
49 /// \param OS Stream to emit the output to.
dumpFunctionProfile(StringRef FName,raw_ostream & OS)50 void SampleProfileReader::dumpFunctionProfile(StringRef FName,
51 raw_ostream &OS) {
52 OS << "Function: " << FName << ": " << Profiles[FName];
53 }
54
55 /// Dump all the function profiles found on stream \p OS.
dump(raw_ostream & OS)56 void SampleProfileReader::dump(raw_ostream &OS) {
57 for (const auto &I : Profiles)
58 dumpFunctionProfile(I.getKey(), OS);
59 }
60
61 /// Parse \p Input as function head.
62 ///
63 /// Parse one line of \p Input, and update function name in \p FName,
64 /// function's total sample count in \p NumSamples, function's entry
65 /// count in \p NumHeadSamples.
66 ///
67 /// \returns true if parsing is successful.
ParseHead(const StringRef & Input,StringRef & FName,uint64_t & NumSamples,uint64_t & NumHeadSamples)68 static bool ParseHead(const StringRef &Input, StringRef &FName,
69 uint64_t &NumSamples, uint64_t &NumHeadSamples) {
70 if (Input[0] == ' ')
71 return false;
72 size_t n2 = Input.rfind(':');
73 size_t n1 = Input.rfind(':', n2 - 1);
74 FName = Input.substr(0, n1);
75 if (Input.substr(n1 + 1, n2 - n1 - 1).getAsInteger(10, NumSamples))
76 return false;
77 if (Input.substr(n2 + 1).getAsInteger(10, NumHeadSamples))
78 return false;
79 return true;
80 }
81
82 /// Returns true if line offset \p L is legal (only has 16 bits).
isOffsetLegal(unsigned L)83 static bool isOffsetLegal(unsigned L) { return (L & 0xffff) == L; }
84
85 /// Parse \p Input as line sample.
86 ///
87 /// \param Input input line.
88 /// \param IsCallsite true if the line represents an inlined callsite.
89 /// \param Depth the depth of the inline stack.
90 /// \param NumSamples total samples of the line/inlined callsite.
91 /// \param LineOffset line offset to the start of the function.
92 /// \param Discriminator discriminator of the line.
93 /// \param TargetCountMap map from indirect call target to count.
94 ///
95 /// returns true if parsing is successful.
ParseLine(const StringRef & Input,bool & IsCallsite,uint32_t & Depth,uint64_t & NumSamples,uint32_t & LineOffset,uint32_t & Discriminator,StringRef & CalleeName,DenseMap<StringRef,uint64_t> & TargetCountMap)96 static bool ParseLine(const StringRef &Input, bool &IsCallsite, uint32_t &Depth,
97 uint64_t &NumSamples, uint32_t &LineOffset,
98 uint32_t &Discriminator, StringRef &CalleeName,
99 DenseMap<StringRef, uint64_t> &TargetCountMap) {
100 for (Depth = 0; Input[Depth] == ' '; Depth++)
101 ;
102 if (Depth == 0)
103 return false;
104
105 size_t n1 = Input.find(':');
106 StringRef Loc = Input.substr(Depth, n1 - Depth);
107 size_t n2 = Loc.find('.');
108 if (n2 == StringRef::npos) {
109 if (Loc.getAsInteger(10, LineOffset) || !isOffsetLegal(LineOffset))
110 return false;
111 Discriminator = 0;
112 } else {
113 if (Loc.substr(0, n2).getAsInteger(10, LineOffset))
114 return false;
115 if (Loc.substr(n2 + 1).getAsInteger(10, Discriminator))
116 return false;
117 }
118
119 StringRef Rest = Input.substr(n1 + 2);
120 if (Rest[0] >= '0' && Rest[0] <= '9') {
121 IsCallsite = false;
122 size_t n3 = Rest.find(' ');
123 if (n3 == StringRef::npos) {
124 if (Rest.getAsInteger(10, NumSamples))
125 return false;
126 } else {
127 if (Rest.substr(0, n3).getAsInteger(10, NumSamples))
128 return false;
129 }
130 // Find call targets and their sample counts.
131 // Note: In some cases, there are symbols in the profile which are not
132 // mangled. To accommodate such cases, use colon + integer pairs as the
133 // anchor points.
134 // An example:
135 // _M_construct<char *>:1000 string_view<std::allocator<char> >:437
136 // ":1000" and ":437" are used as anchor points so the string above will
137 // be interpreted as
138 // target: _M_construct<char *>
139 // count: 1000
140 // target: string_view<std::allocator<char> >
141 // count: 437
142 while (n3 != StringRef::npos) {
143 n3 += Rest.substr(n3).find_first_not_of(' ');
144 Rest = Rest.substr(n3);
145 n3 = Rest.find_first_of(':');
146 if (n3 == StringRef::npos || n3 == 0)
147 return false;
148
149 StringRef Target;
150 uint64_t count, n4;
151 while (true) {
152 // Get the segment after the current colon.
153 StringRef AfterColon = Rest.substr(n3 + 1);
154 // Get the target symbol before the current colon.
155 Target = Rest.substr(0, n3);
156 // Check if the word after the current colon is an integer.
157 n4 = AfterColon.find_first_of(' ');
158 n4 = (n4 != StringRef::npos) ? n3 + n4 + 1 : Rest.size();
159 StringRef WordAfterColon = Rest.substr(n3 + 1, n4 - n3 - 1);
160 if (!WordAfterColon.getAsInteger(10, count))
161 break;
162
163 // Try to find the next colon.
164 uint64_t n5 = AfterColon.find_first_of(':');
165 if (n5 == StringRef::npos)
166 return false;
167 n3 += n5 + 1;
168 }
169
170 // An anchor point is found. Save the {target, count} pair
171 TargetCountMap[Target] = count;
172 if (n4 == Rest.size())
173 break;
174 // Change n3 to the next blank space after colon + integer pair.
175 n3 = n4;
176 }
177 } else {
178 IsCallsite = true;
179 size_t n3 = Rest.find_last_of(':');
180 CalleeName = Rest.substr(0, n3);
181 if (Rest.substr(n3 + 1).getAsInteger(10, NumSamples))
182 return false;
183 }
184 return true;
185 }
186
187 /// Load samples from a text file.
188 ///
189 /// See the documentation at the top of the file for an explanation of
190 /// the expected format.
191 ///
192 /// \returns true if the file was loaded successfully, false otherwise.
read()193 std::error_code SampleProfileReaderText::read() {
194 line_iterator LineIt(*Buffer, /*SkipBlanks=*/true, '#');
195 sampleprof_error Result = sampleprof_error::success;
196
197 InlineCallStack InlineStack;
198
199 for (; !LineIt.is_at_eof(); ++LineIt) {
200 if ((*LineIt)[(*LineIt).find_first_not_of(' ')] == '#')
201 continue;
202 // Read the header of each function.
203 //
204 // Note that for function identifiers we are actually expecting
205 // mangled names, but we may not always get them. This happens when
206 // the compiler decides not to emit the function (e.g., it was inlined
207 // and removed). In this case, the binary will not have the linkage
208 // name for the function, so the profiler will emit the function's
209 // unmangled name, which may contain characters like ':' and '>' in its
210 // name (member functions, templates, etc).
211 //
212 // The only requirement we place on the identifier, then, is that it
213 // should not begin with a number.
214 if ((*LineIt)[0] != ' ') {
215 uint64_t NumSamples, NumHeadSamples;
216 StringRef FName;
217 if (!ParseHead(*LineIt, FName, NumSamples, NumHeadSamples)) {
218 reportError(LineIt.line_number(),
219 "Expected 'mangled_name:NUM:NUM', found " + *LineIt);
220 return sampleprof_error::malformed;
221 }
222 Profiles[FName] = FunctionSamples();
223 FunctionSamples &FProfile = Profiles[FName];
224 FProfile.setName(FName);
225 MergeResult(Result, FProfile.addTotalSamples(NumSamples));
226 MergeResult(Result, FProfile.addHeadSamples(NumHeadSamples));
227 InlineStack.clear();
228 InlineStack.push_back(&FProfile);
229 } else {
230 uint64_t NumSamples;
231 StringRef FName;
232 DenseMap<StringRef, uint64_t> TargetCountMap;
233 bool IsCallsite;
234 uint32_t Depth, LineOffset, Discriminator;
235 if (!ParseLine(*LineIt, IsCallsite, Depth, NumSamples, LineOffset,
236 Discriminator, FName, TargetCountMap)) {
237 reportError(LineIt.line_number(),
238 "Expected 'NUM[.NUM]: NUM[ mangled_name:NUM]*', found " +
239 *LineIt);
240 return sampleprof_error::malformed;
241 }
242 if (IsCallsite) {
243 while (InlineStack.size() > Depth) {
244 InlineStack.pop_back();
245 }
246 FunctionSamples &FSamples = InlineStack.back()->functionSamplesAt(
247 LineLocation(LineOffset, Discriminator))[FName];
248 FSamples.setName(FName);
249 MergeResult(Result, FSamples.addTotalSamples(NumSamples));
250 InlineStack.push_back(&FSamples);
251 } else {
252 while (InlineStack.size() > Depth) {
253 InlineStack.pop_back();
254 }
255 FunctionSamples &FProfile = *InlineStack.back();
256 for (const auto &name_count : TargetCountMap) {
257 MergeResult(Result, FProfile.addCalledTargetSamples(
258 LineOffset, Discriminator, name_count.first,
259 name_count.second));
260 }
261 MergeResult(Result, FProfile.addBodySamples(LineOffset, Discriminator,
262 NumSamples));
263 }
264 }
265 }
266 if (Result == sampleprof_error::success)
267 computeSummary();
268
269 return Result;
270 }
271
hasFormat(const MemoryBuffer & Buffer)272 bool SampleProfileReaderText::hasFormat(const MemoryBuffer &Buffer) {
273 bool result = false;
274
275 // Check that the first non-comment line is a valid function header.
276 line_iterator LineIt(Buffer, /*SkipBlanks=*/true, '#');
277 if (!LineIt.is_at_eof()) {
278 if ((*LineIt)[0] != ' ') {
279 uint64_t NumSamples, NumHeadSamples;
280 StringRef FName;
281 result = ParseHead(*LineIt, FName, NumSamples, NumHeadSamples);
282 }
283 }
284
285 return result;
286 }
287
readNumber()288 template <typename T> ErrorOr<T> SampleProfileReaderBinary::readNumber() {
289 unsigned NumBytesRead = 0;
290 std::error_code EC;
291 uint64_t Val = decodeULEB128(Data, &NumBytesRead);
292
293 if (Val > std::numeric_limits<T>::max())
294 EC = sampleprof_error::malformed;
295 else if (Data + NumBytesRead > End)
296 EC = sampleprof_error::truncated;
297 else
298 EC = sampleprof_error::success;
299
300 if (EC) {
301 reportError(0, EC.message());
302 return EC;
303 }
304
305 Data += NumBytesRead;
306 return static_cast<T>(Val);
307 }
308
readString()309 ErrorOr<StringRef> SampleProfileReaderBinary::readString() {
310 std::error_code EC;
311 StringRef Str(reinterpret_cast<const char *>(Data));
312 if (Data + Str.size() + 1 > End) {
313 EC = sampleprof_error::truncated;
314 reportError(0, EC.message());
315 return EC;
316 }
317
318 Data += Str.size() + 1;
319 return Str;
320 }
321
322 template <typename T>
readStringIndex(T & Table)323 inline ErrorOr<uint32_t> SampleProfileReaderBinary::readStringIndex(T &Table) {
324 std::error_code EC;
325 auto Idx = readNumber<uint32_t>();
326 if (std::error_code EC = Idx.getError())
327 return EC;
328 if (*Idx >= Table.size())
329 return sampleprof_error::truncated_name_table;
330 return *Idx;
331 }
332
readStringFromTable()333 ErrorOr<StringRef> SampleProfileReaderRawBinary::readStringFromTable() {
334 auto Idx = readStringIndex(NameTable);
335 if (std::error_code EC = Idx.getError())
336 return EC;
337
338 return NameTable[*Idx];
339 }
340
readStringFromTable()341 ErrorOr<StringRef> SampleProfileReaderCompactBinary::readStringFromTable() {
342 auto Idx = readStringIndex(NameTable);
343 if (std::error_code EC = Idx.getError())
344 return EC;
345
346 return StringRef(NameTable[*Idx]);
347 }
348
349 std::error_code
readProfile(FunctionSamples & FProfile)350 SampleProfileReaderBinary::readProfile(FunctionSamples &FProfile) {
351 auto NumSamples = readNumber<uint64_t>();
352 if (std::error_code EC = NumSamples.getError())
353 return EC;
354 FProfile.addTotalSamples(*NumSamples);
355
356 // Read the samples in the body.
357 auto NumRecords = readNumber<uint32_t>();
358 if (std::error_code EC = NumRecords.getError())
359 return EC;
360
361 for (uint32_t I = 0; I < *NumRecords; ++I) {
362 auto LineOffset = readNumber<uint64_t>();
363 if (std::error_code EC = LineOffset.getError())
364 return EC;
365
366 if (!isOffsetLegal(*LineOffset)) {
367 return std::error_code();
368 }
369
370 auto Discriminator = readNumber<uint64_t>();
371 if (std::error_code EC = Discriminator.getError())
372 return EC;
373
374 auto NumSamples = readNumber<uint64_t>();
375 if (std::error_code EC = NumSamples.getError())
376 return EC;
377
378 auto NumCalls = readNumber<uint32_t>();
379 if (std::error_code EC = NumCalls.getError())
380 return EC;
381
382 for (uint32_t J = 0; J < *NumCalls; ++J) {
383 auto CalledFunction(readStringFromTable());
384 if (std::error_code EC = CalledFunction.getError())
385 return EC;
386
387 auto CalledFunctionSamples = readNumber<uint64_t>();
388 if (std::error_code EC = CalledFunctionSamples.getError())
389 return EC;
390
391 FProfile.addCalledTargetSamples(*LineOffset, *Discriminator,
392 *CalledFunction, *CalledFunctionSamples);
393 }
394
395 FProfile.addBodySamples(*LineOffset, *Discriminator, *NumSamples);
396 }
397
398 // Read all the samples for inlined function calls.
399 auto NumCallsites = readNumber<uint32_t>();
400 if (std::error_code EC = NumCallsites.getError())
401 return EC;
402
403 for (uint32_t J = 0; J < *NumCallsites; ++J) {
404 auto LineOffset = readNumber<uint64_t>();
405 if (std::error_code EC = LineOffset.getError())
406 return EC;
407
408 auto Discriminator = readNumber<uint64_t>();
409 if (std::error_code EC = Discriminator.getError())
410 return EC;
411
412 auto FName(readStringFromTable());
413 if (std::error_code EC = FName.getError())
414 return EC;
415
416 FunctionSamples &CalleeProfile = FProfile.functionSamplesAt(
417 LineLocation(*LineOffset, *Discriminator))[*FName];
418 CalleeProfile.setName(*FName);
419 if (std::error_code EC = readProfile(CalleeProfile))
420 return EC;
421 }
422
423 return sampleprof_error::success;
424 }
425
read()426 std::error_code SampleProfileReaderBinary::read() {
427 while (!at_eof()) {
428 auto NumHeadSamples = readNumber<uint64_t>();
429 if (std::error_code EC = NumHeadSamples.getError())
430 return EC;
431
432 auto FName(readStringFromTable());
433 if (std::error_code EC = FName.getError())
434 return EC;
435
436 Profiles[*FName] = FunctionSamples();
437 FunctionSamples &FProfile = Profiles[*FName];
438 FProfile.setName(*FName);
439
440 FProfile.addHeadSamples(*NumHeadSamples);
441
442 if (std::error_code EC = readProfile(FProfile))
443 return EC;
444 }
445
446 return sampleprof_error::success;
447 }
448
verifySPMagic(uint64_t Magic)449 std::error_code SampleProfileReaderRawBinary::verifySPMagic(uint64_t Magic) {
450 if (Magic == SPMagic())
451 return sampleprof_error::success;
452 return sampleprof_error::bad_magic;
453 }
454
455 std::error_code
verifySPMagic(uint64_t Magic)456 SampleProfileReaderCompactBinary::verifySPMagic(uint64_t Magic) {
457 if (Magic == SPMagic(SPF_Compact_Binary))
458 return sampleprof_error::success;
459 return sampleprof_error::bad_magic;
460 }
461
readNameTable()462 std::error_code SampleProfileReaderRawBinary::readNameTable() {
463 auto Size = readNumber<uint32_t>();
464 if (std::error_code EC = Size.getError())
465 return EC;
466 NameTable.reserve(*Size);
467 for (uint32_t I = 0; I < *Size; ++I) {
468 auto Name(readString());
469 if (std::error_code EC = Name.getError())
470 return EC;
471 NameTable.push_back(*Name);
472 }
473
474 return sampleprof_error::success;
475 }
476
readNameTable()477 std::error_code SampleProfileReaderCompactBinary::readNameTable() {
478 auto Size = readNumber<uint64_t>();
479 if (std::error_code EC = Size.getError())
480 return EC;
481 NameTable.reserve(*Size);
482 for (uint32_t I = 0; I < *Size; ++I) {
483 auto FID = readNumber<uint64_t>();
484 if (std::error_code EC = FID.getError())
485 return EC;
486 NameTable.push_back(std::to_string(*FID));
487 }
488 return sampleprof_error::success;
489 }
490
readHeader()491 std::error_code SampleProfileReaderBinary::readHeader() {
492 Data = reinterpret_cast<const uint8_t *>(Buffer->getBufferStart());
493 End = Data + Buffer->getBufferSize();
494
495 // Read and check the magic identifier.
496 auto Magic = readNumber<uint64_t>();
497 if (std::error_code EC = Magic.getError())
498 return EC;
499 else if (std::error_code EC = verifySPMagic(*Magic))
500 return EC;
501
502 // Read the version number.
503 auto Version = readNumber<uint64_t>();
504 if (std::error_code EC = Version.getError())
505 return EC;
506 else if (*Version != SPVersion())
507 return sampleprof_error::unsupported_version;
508
509 if (std::error_code EC = readSummary())
510 return EC;
511
512 if (std::error_code EC = readNameTable())
513 return EC;
514 return sampleprof_error::success;
515 }
516
readSummaryEntry(std::vector<ProfileSummaryEntry> & Entries)517 std::error_code SampleProfileReaderBinary::readSummaryEntry(
518 std::vector<ProfileSummaryEntry> &Entries) {
519 auto Cutoff = readNumber<uint64_t>();
520 if (std::error_code EC = Cutoff.getError())
521 return EC;
522
523 auto MinBlockCount = readNumber<uint64_t>();
524 if (std::error_code EC = MinBlockCount.getError())
525 return EC;
526
527 auto NumBlocks = readNumber<uint64_t>();
528 if (std::error_code EC = NumBlocks.getError())
529 return EC;
530
531 Entries.emplace_back(*Cutoff, *MinBlockCount, *NumBlocks);
532 return sampleprof_error::success;
533 }
534
readSummary()535 std::error_code SampleProfileReaderBinary::readSummary() {
536 auto TotalCount = readNumber<uint64_t>();
537 if (std::error_code EC = TotalCount.getError())
538 return EC;
539
540 auto MaxBlockCount = readNumber<uint64_t>();
541 if (std::error_code EC = MaxBlockCount.getError())
542 return EC;
543
544 auto MaxFunctionCount = readNumber<uint64_t>();
545 if (std::error_code EC = MaxFunctionCount.getError())
546 return EC;
547
548 auto NumBlocks = readNumber<uint64_t>();
549 if (std::error_code EC = NumBlocks.getError())
550 return EC;
551
552 auto NumFunctions = readNumber<uint64_t>();
553 if (std::error_code EC = NumFunctions.getError())
554 return EC;
555
556 auto NumSummaryEntries = readNumber<uint64_t>();
557 if (std::error_code EC = NumSummaryEntries.getError())
558 return EC;
559
560 std::vector<ProfileSummaryEntry> Entries;
561 for (unsigned i = 0; i < *NumSummaryEntries; i++) {
562 std::error_code EC = readSummaryEntry(Entries);
563 if (EC != sampleprof_error::success)
564 return EC;
565 }
566 Summary = llvm::make_unique<ProfileSummary>(
567 ProfileSummary::PSK_Sample, Entries, *TotalCount, *MaxBlockCount, 0,
568 *MaxFunctionCount, *NumBlocks, *NumFunctions);
569
570 return sampleprof_error::success;
571 }
572
hasFormat(const MemoryBuffer & Buffer)573 bool SampleProfileReaderRawBinary::hasFormat(const MemoryBuffer &Buffer) {
574 const uint8_t *Data =
575 reinterpret_cast<const uint8_t *>(Buffer.getBufferStart());
576 uint64_t Magic = decodeULEB128(Data);
577 return Magic == SPMagic();
578 }
579
hasFormat(const MemoryBuffer & Buffer)580 bool SampleProfileReaderCompactBinary::hasFormat(const MemoryBuffer &Buffer) {
581 const uint8_t *Data =
582 reinterpret_cast<const uint8_t *>(Buffer.getBufferStart());
583 uint64_t Magic = decodeULEB128(Data);
584 return Magic == SPMagic(SPF_Compact_Binary);
585 }
586
skipNextWord()587 std::error_code SampleProfileReaderGCC::skipNextWord() {
588 uint32_t dummy;
589 if (!GcovBuffer.readInt(dummy))
590 return sampleprof_error::truncated;
591 return sampleprof_error::success;
592 }
593
readNumber()594 template <typename T> ErrorOr<T> SampleProfileReaderGCC::readNumber() {
595 if (sizeof(T) <= sizeof(uint32_t)) {
596 uint32_t Val;
597 if (GcovBuffer.readInt(Val) && Val <= std::numeric_limits<T>::max())
598 return static_cast<T>(Val);
599 } else if (sizeof(T) <= sizeof(uint64_t)) {
600 uint64_t Val;
601 if (GcovBuffer.readInt64(Val) && Val <= std::numeric_limits<T>::max())
602 return static_cast<T>(Val);
603 }
604
605 std::error_code EC = sampleprof_error::malformed;
606 reportError(0, EC.message());
607 return EC;
608 }
609
readString()610 ErrorOr<StringRef> SampleProfileReaderGCC::readString() {
611 StringRef Str;
612 if (!GcovBuffer.readString(Str))
613 return sampleprof_error::truncated;
614 return Str;
615 }
616
readHeader()617 std::error_code SampleProfileReaderGCC::readHeader() {
618 // Read the magic identifier.
619 if (!GcovBuffer.readGCDAFormat())
620 return sampleprof_error::unrecognized_format;
621
622 // Read the version number. Note - the GCC reader does not validate this
623 // version, but the profile creator generates v704.
624 GCOV::GCOVVersion version;
625 if (!GcovBuffer.readGCOVVersion(version))
626 return sampleprof_error::unrecognized_format;
627
628 if (version != GCOV::V704)
629 return sampleprof_error::unsupported_version;
630
631 // Skip the empty integer.
632 if (std::error_code EC = skipNextWord())
633 return EC;
634
635 return sampleprof_error::success;
636 }
637
readSectionTag(uint32_t Expected)638 std::error_code SampleProfileReaderGCC::readSectionTag(uint32_t Expected) {
639 uint32_t Tag;
640 if (!GcovBuffer.readInt(Tag))
641 return sampleprof_error::truncated;
642
643 if (Tag != Expected)
644 return sampleprof_error::malformed;
645
646 if (std::error_code EC = skipNextWord())
647 return EC;
648
649 return sampleprof_error::success;
650 }
651
readNameTable()652 std::error_code SampleProfileReaderGCC::readNameTable() {
653 if (std::error_code EC = readSectionTag(GCOVTagAFDOFileNames))
654 return EC;
655
656 uint32_t Size;
657 if (!GcovBuffer.readInt(Size))
658 return sampleprof_error::truncated;
659
660 for (uint32_t I = 0; I < Size; ++I) {
661 StringRef Str;
662 if (!GcovBuffer.readString(Str))
663 return sampleprof_error::truncated;
664 Names.push_back(Str);
665 }
666
667 return sampleprof_error::success;
668 }
669
readFunctionProfiles()670 std::error_code SampleProfileReaderGCC::readFunctionProfiles() {
671 if (std::error_code EC = readSectionTag(GCOVTagAFDOFunction))
672 return EC;
673
674 uint32_t NumFunctions;
675 if (!GcovBuffer.readInt(NumFunctions))
676 return sampleprof_error::truncated;
677
678 InlineCallStack Stack;
679 for (uint32_t I = 0; I < NumFunctions; ++I)
680 if (std::error_code EC = readOneFunctionProfile(Stack, true, 0))
681 return EC;
682
683 computeSummary();
684 return sampleprof_error::success;
685 }
686
readOneFunctionProfile(const InlineCallStack & InlineStack,bool Update,uint32_t Offset)687 std::error_code SampleProfileReaderGCC::readOneFunctionProfile(
688 const InlineCallStack &InlineStack, bool Update, uint32_t Offset) {
689 uint64_t HeadCount = 0;
690 if (InlineStack.size() == 0)
691 if (!GcovBuffer.readInt64(HeadCount))
692 return sampleprof_error::truncated;
693
694 uint32_t NameIdx;
695 if (!GcovBuffer.readInt(NameIdx))
696 return sampleprof_error::truncated;
697
698 StringRef Name(Names[NameIdx]);
699
700 uint32_t NumPosCounts;
701 if (!GcovBuffer.readInt(NumPosCounts))
702 return sampleprof_error::truncated;
703
704 uint32_t NumCallsites;
705 if (!GcovBuffer.readInt(NumCallsites))
706 return sampleprof_error::truncated;
707
708 FunctionSamples *FProfile = nullptr;
709 if (InlineStack.size() == 0) {
710 // If this is a top function that we have already processed, do not
711 // update its profile again. This happens in the presence of
712 // function aliases. Since these aliases share the same function
713 // body, there will be identical replicated profiles for the
714 // original function. In this case, we simply not bother updating
715 // the profile of the original function.
716 FProfile = &Profiles[Name];
717 FProfile->addHeadSamples(HeadCount);
718 if (FProfile->getTotalSamples() > 0)
719 Update = false;
720 } else {
721 // Otherwise, we are reading an inlined instance. The top of the
722 // inline stack contains the profile of the caller. Insert this
723 // callee in the caller's CallsiteMap.
724 FunctionSamples *CallerProfile = InlineStack.front();
725 uint32_t LineOffset = Offset >> 16;
726 uint32_t Discriminator = Offset & 0xffff;
727 FProfile = &CallerProfile->functionSamplesAt(
728 LineLocation(LineOffset, Discriminator))[Name];
729 }
730 FProfile->setName(Name);
731
732 for (uint32_t I = 0; I < NumPosCounts; ++I) {
733 uint32_t Offset;
734 if (!GcovBuffer.readInt(Offset))
735 return sampleprof_error::truncated;
736
737 uint32_t NumTargets;
738 if (!GcovBuffer.readInt(NumTargets))
739 return sampleprof_error::truncated;
740
741 uint64_t Count;
742 if (!GcovBuffer.readInt64(Count))
743 return sampleprof_error::truncated;
744
745 // The line location is encoded in the offset as:
746 // high 16 bits: line offset to the start of the function.
747 // low 16 bits: discriminator.
748 uint32_t LineOffset = Offset >> 16;
749 uint32_t Discriminator = Offset & 0xffff;
750
751 InlineCallStack NewStack;
752 NewStack.push_back(FProfile);
753 NewStack.insert(NewStack.end(), InlineStack.begin(), InlineStack.end());
754 if (Update) {
755 // Walk up the inline stack, adding the samples on this line to
756 // the total sample count of the callers in the chain.
757 for (auto CallerProfile : NewStack)
758 CallerProfile->addTotalSamples(Count);
759
760 // Update the body samples for the current profile.
761 FProfile->addBodySamples(LineOffset, Discriminator, Count);
762 }
763
764 // Process the list of functions called at an indirect call site.
765 // These are all the targets that a function pointer (or virtual
766 // function) resolved at runtime.
767 for (uint32_t J = 0; J < NumTargets; J++) {
768 uint32_t HistVal;
769 if (!GcovBuffer.readInt(HistVal))
770 return sampleprof_error::truncated;
771
772 if (HistVal != HIST_TYPE_INDIR_CALL_TOPN)
773 return sampleprof_error::malformed;
774
775 uint64_t TargetIdx;
776 if (!GcovBuffer.readInt64(TargetIdx))
777 return sampleprof_error::truncated;
778 StringRef TargetName(Names[TargetIdx]);
779
780 uint64_t TargetCount;
781 if (!GcovBuffer.readInt64(TargetCount))
782 return sampleprof_error::truncated;
783
784 if (Update)
785 FProfile->addCalledTargetSamples(LineOffset, Discriminator,
786 TargetName, TargetCount);
787 }
788 }
789
790 // Process all the inlined callers into the current function. These
791 // are all the callsites that were inlined into this function.
792 for (uint32_t I = 0; I < NumCallsites; I++) {
793 // The offset is encoded as:
794 // high 16 bits: line offset to the start of the function.
795 // low 16 bits: discriminator.
796 uint32_t Offset;
797 if (!GcovBuffer.readInt(Offset))
798 return sampleprof_error::truncated;
799 InlineCallStack NewStack;
800 NewStack.push_back(FProfile);
801 NewStack.insert(NewStack.end(), InlineStack.begin(), InlineStack.end());
802 if (std::error_code EC = readOneFunctionProfile(NewStack, Update, Offset))
803 return EC;
804 }
805
806 return sampleprof_error::success;
807 }
808
809 /// Read a GCC AutoFDO profile.
810 ///
811 /// This format is generated by the Linux Perf conversion tool at
812 /// https://github.com/google/autofdo.
read()813 std::error_code SampleProfileReaderGCC::read() {
814 // Read the string table.
815 if (std::error_code EC = readNameTable())
816 return EC;
817
818 // Read the source profile.
819 if (std::error_code EC = readFunctionProfiles())
820 return EC;
821
822 return sampleprof_error::success;
823 }
824
hasFormat(const MemoryBuffer & Buffer)825 bool SampleProfileReaderGCC::hasFormat(const MemoryBuffer &Buffer) {
826 StringRef Magic(reinterpret_cast<const char *>(Buffer.getBufferStart()));
827 return Magic == "adcg*704";
828 }
829
830 /// Prepare a memory buffer for the contents of \p Filename.
831 ///
832 /// \returns an error code indicating the status of the buffer.
833 static ErrorOr<std::unique_ptr<MemoryBuffer>>
setupMemoryBuffer(const Twine & Filename)834 setupMemoryBuffer(const Twine &Filename) {
835 auto BufferOrErr = MemoryBuffer::getFileOrSTDIN(Filename);
836 if (std::error_code EC = BufferOrErr.getError())
837 return EC;
838 auto Buffer = std::move(BufferOrErr.get());
839
840 // Sanity check the file.
841 if (uint64_t(Buffer->getBufferSize()) > std::numeric_limits<uint32_t>::max())
842 return sampleprof_error::too_large;
843
844 return std::move(Buffer);
845 }
846
847 /// Create a sample profile reader based on the format of the input file.
848 ///
849 /// \param Filename The file to open.
850 ///
851 /// \param C The LLVM context to use to emit diagnostics.
852 ///
853 /// \returns an error code indicating the status of the created reader.
854 ErrorOr<std::unique_ptr<SampleProfileReader>>
create(const Twine & Filename,LLVMContext & C)855 SampleProfileReader::create(const Twine &Filename, LLVMContext &C) {
856 auto BufferOrError = setupMemoryBuffer(Filename);
857 if (std::error_code EC = BufferOrError.getError())
858 return EC;
859 return create(BufferOrError.get(), C);
860 }
861
862 /// Create a sample profile reader based on the format of the input data.
863 ///
864 /// \param B The memory buffer to create the reader from (assumes ownership).
865 ///
866 /// \param C The LLVM context to use to emit diagnostics.
867 ///
868 /// \returns an error code indicating the status of the created reader.
869 ErrorOr<std::unique_ptr<SampleProfileReader>>
create(std::unique_ptr<MemoryBuffer> & B,LLVMContext & C)870 SampleProfileReader::create(std::unique_ptr<MemoryBuffer> &B, LLVMContext &C) {
871 std::unique_ptr<SampleProfileReader> Reader;
872 if (SampleProfileReaderRawBinary::hasFormat(*B))
873 Reader.reset(new SampleProfileReaderRawBinary(std::move(B), C));
874 else if (SampleProfileReaderCompactBinary::hasFormat(*B))
875 Reader.reset(new SampleProfileReaderCompactBinary(std::move(B), C));
876 else if (SampleProfileReaderGCC::hasFormat(*B))
877 Reader.reset(new SampleProfileReaderGCC(std::move(B), C));
878 else if (SampleProfileReaderText::hasFormat(*B))
879 Reader.reset(new SampleProfileReaderText(std::move(B), C));
880 else
881 return sampleprof_error::unrecognized_format;
882
883 if (std::error_code EC = Reader->readHeader())
884 return EC;
885
886 return std::move(Reader);
887 }
888
889 // For text and GCC file formats, we compute the summary after reading the
890 // profile. Binary format has the profile summary in its header.
computeSummary()891 void SampleProfileReader::computeSummary() {
892 SampleProfileSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs);
893 for (const auto &I : Profiles) {
894 const FunctionSamples &Profile = I.second;
895 Builder.addRecord(Profile);
896 }
897 Summary = Builder.getSummary();
898 }
899