1 //===- SampleProfReader.cpp - Read LLVM sample profile data ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the class that reads LLVM sample profiles. It
11 // supports three file formats: text, binary and gcov.
13 // The textual representation is useful for debugging and testing purposes. The
14 // binary representation is more compact, resulting in smaller file sizes.
16 // The gcov encoding is the one generated by GCC's AutoFDO profile creation
17 // tool (https://github.com/google/autofdo)
19 // All three encodings can be used interchangeably as an input sample profile.
21 //===----------------------------------------------------------------------===//
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/MD5.h"
34 #include "llvm/Support/MemoryBuffer.h"
35 #include "llvm/Support/raw_ostream.h"
41 #include <system_error>
45 using namespace sampleprof;
47 /// Dump the function profile for \p FName.
49 /// \param FName Name of the function to print.
50 /// \param OS Stream to emit the output to.
51 void SampleProfileReader::dumpFunctionProfile(StringRef FName,
53 OS << "Function: " << FName << ": " << Profiles[FName];
56 /// Dump all the function profiles found on stream \p OS.
57 void SampleProfileReader::dump(raw_ostream &OS) {
58 for (const auto &I : Profiles)
59 dumpFunctionProfile(I.getKey(), OS);
62 /// Parse \p Input as function head.
64 /// Parse one line of \p Input, and update function name in \p FName,
65 /// function's total sample count in \p NumSamples, function's entry
66 /// count in \p NumHeadSamples.
68 /// \returns true if parsing is successful.
69 static bool ParseHead(const StringRef &Input, StringRef &FName,
70 uint64_t &NumSamples, uint64_t &NumHeadSamples) {
73 size_t n2 = Input.rfind(':');
74 size_t n1 = Input.rfind(':', n2 - 1);
75 FName = Input.substr(0, n1);
76 if (Input.substr(n1 + 1, n2 - n1 - 1).getAsInteger(10, NumSamples))
78 if (Input.substr(n2 + 1).getAsInteger(10, NumHeadSamples))
83 /// Returns true if line offset \p L is legal (only has 16 bits).
84 static bool isOffsetLegal(unsigned L) { return (L & 0xffff) == L; }
86 /// Parse \p Input as line sample.
88 /// \param Input input line.
89 /// \param IsCallsite true if the line represents an inlined callsite.
90 /// \param Depth the depth of the inline stack.
91 /// \param NumSamples total samples of the line/inlined callsite.
92 /// \param LineOffset line offset to the start of the function.
93 /// \param Discriminator discriminator of the line.
94 /// \param TargetCountMap map from indirect call target to count.
96 /// returns true if parsing is successful.
97 static bool ParseLine(const StringRef &Input, bool &IsCallsite, uint32_t &Depth,
98 uint64_t &NumSamples, uint32_t &LineOffset,
99 uint32_t &Discriminator, StringRef &CalleeName,
100 DenseMap<StringRef, uint64_t> &TargetCountMap) {
101 for (Depth = 0; Input[Depth] == ' '; Depth++)
106 size_t n1 = Input.find(':');
107 StringRef Loc = Input.substr(Depth, n1 - Depth);
108 size_t n2 = Loc.find('.');
109 if (n2 == StringRef::npos) {
110 if (Loc.getAsInteger(10, LineOffset) || !isOffsetLegal(LineOffset))
114 if (Loc.substr(0, n2).getAsInteger(10, LineOffset))
116 if (Loc.substr(n2 + 1).getAsInteger(10, Discriminator))
120 StringRef Rest = Input.substr(n1 + 2);
121 if (Rest[0] >= '0' && Rest[0] <= '9') {
123 size_t n3 = Rest.find(' ');
124 if (n3 == StringRef::npos) {
125 if (Rest.getAsInteger(10, NumSamples))
128 if (Rest.substr(0, n3).getAsInteger(10, NumSamples))
131 // Find call targets and their sample counts.
132 // Note: In some cases, there are symbols in the profile which are not
133 // mangled. To accommodate such cases, use colon + integer pairs as the
136 // _M_construct<char *>:1000 string_view<std::allocator<char> >:437
137 // ":1000" and ":437" are used as anchor points so the string above will
139 // target: _M_construct<char *>
141 // target: string_view<std::allocator<char> >
143 while (n3 != StringRef::npos) {
144 n3 += Rest.substr(n3).find_first_not_of(' ');
145 Rest = Rest.substr(n3);
146 n3 = Rest.find_first_of(':');
147 if (n3 == StringRef::npos || n3 == 0)
153 // Get the segment after the current colon.
154 StringRef AfterColon = Rest.substr(n3 + 1);
155 // Get the target symbol before the current colon.
156 Target = Rest.substr(0, n3);
157 // Check if the word after the current colon is an integer.
158 n4 = AfterColon.find_first_of(' ');
159 n4 = (n4 != StringRef::npos) ? n3 + n4 + 1 : Rest.size();
160 StringRef WordAfterColon = Rest.substr(n3 + 1, n4 - n3 - 1);
161 if (!WordAfterColon.getAsInteger(10, count))
164 // Try to find the next colon.
165 uint64_t n5 = AfterColon.find_first_of(':');
166 if (n5 == StringRef::npos)
171 // An anchor point is found. Save the {target, count} pair
172 TargetCountMap[Target] = count;
173 if (n4 == Rest.size())
175 // Change n3 to the next blank space after colon + integer pair.
180 size_t n3 = Rest.find_last_of(':');
181 CalleeName = Rest.substr(0, n3);
182 if (Rest.substr(n3 + 1).getAsInteger(10, NumSamples))
188 /// Load samples from a text file.
190 /// See the documentation at the top of the file for an explanation of
191 /// the expected format.
193 /// \returns true if the file was loaded successfully, false otherwise.
194 std::error_code SampleProfileReaderText::read() {
195 line_iterator LineIt(*Buffer, /*SkipBlanks=*/true, '#');
196 sampleprof_error Result = sampleprof_error::success;
198 InlineCallStack InlineStack;
200 for (; !LineIt.is_at_eof(); ++LineIt) {
201 if ((*LineIt)[(*LineIt).find_first_not_of(' ')] == '#')
203 // Read the header of each function.
205 // Note that for function identifiers we are actually expecting
206 // mangled names, but we may not always get them. This happens when
207 // the compiler decides not to emit the function (e.g., it was inlined
208 // and removed). In this case, the binary will not have the linkage
209 // name for the function, so the profiler will emit the function's
210 // unmangled name, which may contain characters like ':' and '>' in its
211 // name (member functions, templates, etc).
213 // The only requirement we place on the identifier, then, is that it
214 // should not begin with a number.
215 if ((*LineIt)[0] != ' ') {
216 uint64_t NumSamples, NumHeadSamples;
218 if (!ParseHead(*LineIt, FName, NumSamples, NumHeadSamples)) {
219 reportError(LineIt.line_number(),
220 "Expected 'mangled_name:NUM:NUM', found " + *LineIt);
221 return sampleprof_error::malformed;
223 Profiles[FName] = FunctionSamples();
224 FunctionSamples &FProfile = Profiles[FName];
225 FProfile.setName(FName);
226 MergeResult(Result, FProfile.addTotalSamples(NumSamples));
227 MergeResult(Result, FProfile.addHeadSamples(NumHeadSamples));
229 InlineStack.push_back(&FProfile);
233 DenseMap<StringRef, uint64_t> TargetCountMap;
235 uint32_t Depth, LineOffset, Discriminator;
236 if (!ParseLine(*LineIt, IsCallsite, Depth, NumSamples, LineOffset,
237 Discriminator, FName, TargetCountMap)) {
238 reportError(LineIt.line_number(),
239 "Expected 'NUM[.NUM]: NUM[ mangled_name:NUM]*', found " +
241 return sampleprof_error::malformed;
244 while (InlineStack.size() > Depth) {
245 InlineStack.pop_back();
247 FunctionSamples &FSamples = InlineStack.back()->functionSamplesAt(
248 LineLocation(LineOffset, Discriminator))[FName];
249 FSamples.setName(FName);
250 MergeResult(Result, FSamples.addTotalSamples(NumSamples));
251 InlineStack.push_back(&FSamples);
253 while (InlineStack.size() > Depth) {
254 InlineStack.pop_back();
256 FunctionSamples &FProfile = *InlineStack.back();
257 for (const auto &name_count : TargetCountMap) {
258 MergeResult(Result, FProfile.addCalledTargetSamples(
259 LineOffset, Discriminator, name_count.first,
262 MergeResult(Result, FProfile.addBodySamples(LineOffset, Discriminator,
267 if (Result == sampleprof_error::success)
273 bool SampleProfileReaderText::hasFormat(const MemoryBuffer &Buffer) {
276 // Check that the first non-comment line is a valid function header.
277 line_iterator LineIt(Buffer, /*SkipBlanks=*/true, '#');
278 if (!LineIt.is_at_eof()) {
279 if ((*LineIt)[0] != ' ') {
280 uint64_t NumSamples, NumHeadSamples;
282 result = ParseHead(*LineIt, FName, NumSamples, NumHeadSamples);
289 template <typename T> ErrorOr<T> SampleProfileReaderBinary::readNumber() {
290 unsigned NumBytesRead = 0;
292 uint64_t Val = decodeULEB128(Data, &NumBytesRead);
294 if (Val > std::numeric_limits<T>::max())
295 EC = sampleprof_error::malformed;
296 else if (Data + NumBytesRead > End)
297 EC = sampleprof_error::truncated;
299 EC = sampleprof_error::success;
302 reportError(0, EC.message());
306 Data += NumBytesRead;
307 return static_cast<T>(Val);
310 ErrorOr<StringRef> SampleProfileReaderBinary::readString() {
312 StringRef Str(reinterpret_cast<const char *>(Data));
313 if (Data + Str.size() + 1 > End) {
314 EC = sampleprof_error::truncated;
315 reportError(0, EC.message());
319 Data += Str.size() + 1;
323 template <typename T>
324 ErrorOr<T> SampleProfileReaderBinary::readUnencodedNumber() {
327 if (Data + sizeof(T) > End) {
328 EC = sampleprof_error::truncated;
329 reportError(0, EC.message());
333 using namespace support;
334 T Val = endian::readNext<T, little, unaligned>(Data);
338 template <typename T>
339 inline ErrorOr<uint32_t> SampleProfileReaderBinary::readStringIndex(T &Table) {
341 auto Idx = readNumber<uint32_t>();
342 if (std::error_code EC = Idx.getError())
344 if (*Idx >= Table.size())
345 return sampleprof_error::truncated_name_table;
349 ErrorOr<StringRef> SampleProfileReaderRawBinary::readStringFromTable() {
350 auto Idx = readStringIndex(NameTable);
351 if (std::error_code EC = Idx.getError())
354 return NameTable[*Idx];
357 ErrorOr<StringRef> SampleProfileReaderCompactBinary::readStringFromTable() {
358 auto Idx = readStringIndex(NameTable);
359 if (std::error_code EC = Idx.getError())
362 return StringRef(NameTable[*Idx]);
366 SampleProfileReaderBinary::readProfile(FunctionSamples &FProfile) {
367 auto NumSamples = readNumber<uint64_t>();
368 if (std::error_code EC = NumSamples.getError())
370 FProfile.addTotalSamples(*NumSamples);
372 // Read the samples in the body.
373 auto NumRecords = readNumber<uint32_t>();
374 if (std::error_code EC = NumRecords.getError())
377 for (uint32_t I = 0; I < *NumRecords; ++I) {
378 auto LineOffset = readNumber<uint64_t>();
379 if (std::error_code EC = LineOffset.getError())
382 if (!isOffsetLegal(*LineOffset)) {
383 return std::error_code();
386 auto Discriminator = readNumber<uint64_t>();
387 if (std::error_code EC = Discriminator.getError())
390 auto NumSamples = readNumber<uint64_t>();
391 if (std::error_code EC = NumSamples.getError())
394 auto NumCalls = readNumber<uint32_t>();
395 if (std::error_code EC = NumCalls.getError())
398 for (uint32_t J = 0; J < *NumCalls; ++J) {
399 auto CalledFunction(readStringFromTable());
400 if (std::error_code EC = CalledFunction.getError())
403 auto CalledFunctionSamples = readNumber<uint64_t>();
404 if (std::error_code EC = CalledFunctionSamples.getError())
407 FProfile.addCalledTargetSamples(*LineOffset, *Discriminator,
408 *CalledFunction, *CalledFunctionSamples);
411 FProfile.addBodySamples(*LineOffset, *Discriminator, *NumSamples);
414 // Read all the samples for inlined function calls.
415 auto NumCallsites = readNumber<uint32_t>();
416 if (std::error_code EC = NumCallsites.getError())
419 for (uint32_t J = 0; J < *NumCallsites; ++J) {
420 auto LineOffset = readNumber<uint64_t>();
421 if (std::error_code EC = LineOffset.getError())
424 auto Discriminator = readNumber<uint64_t>();
425 if (std::error_code EC = Discriminator.getError())
428 auto FName(readStringFromTable());
429 if (std::error_code EC = FName.getError())
432 FunctionSamples &CalleeProfile = FProfile.functionSamplesAt(
433 LineLocation(*LineOffset, *Discriminator))[*FName];
434 CalleeProfile.setName(*FName);
435 if (std::error_code EC = readProfile(CalleeProfile))
439 return sampleprof_error::success;
442 std::error_code SampleProfileReaderBinary::readFuncProfile() {
443 auto NumHeadSamples = readNumber<uint64_t>();
444 if (std::error_code EC = NumHeadSamples.getError())
447 auto FName(readStringFromTable());
448 if (std::error_code EC = FName.getError())
451 Profiles[*FName] = FunctionSamples();
452 FunctionSamples &FProfile = Profiles[*FName];
453 FProfile.setName(*FName);
455 FProfile.addHeadSamples(*NumHeadSamples);
457 if (std::error_code EC = readProfile(FProfile))
459 return sampleprof_error::success;
462 std::error_code SampleProfileReaderBinary::read() {
464 if (std::error_code EC = readFuncProfile())
468 return sampleprof_error::success;
471 std::error_code SampleProfileReaderCompactBinary::read() {
472 for (auto Name : FuncsToUse) {
473 auto GUID = std::to_string(MD5Hash(Name));
474 auto iter = FuncOffsetTable.find(StringRef(GUID));
475 if (iter == FuncOffsetTable.end())
477 const uint8_t *SavedData = Data;
478 Data = reinterpret_cast<const uint8_t *>(Buffer->getBufferStart()) +
480 if (std::error_code EC = readFuncProfile())
484 return sampleprof_error::success;
487 std::error_code SampleProfileReaderRawBinary::verifySPMagic(uint64_t Magic) {
488 if (Magic == SPMagic())
489 return sampleprof_error::success;
490 return sampleprof_error::bad_magic;
494 SampleProfileReaderCompactBinary::verifySPMagic(uint64_t Magic) {
495 if (Magic == SPMagic(SPF_Compact_Binary))
496 return sampleprof_error::success;
497 return sampleprof_error::bad_magic;
500 std::error_code SampleProfileReaderRawBinary::readNameTable() {
501 auto Size = readNumber<uint32_t>();
502 if (std::error_code EC = Size.getError())
504 NameTable.reserve(*Size);
505 for (uint32_t I = 0; I < *Size; ++I) {
506 auto Name(readString());
507 if (std::error_code EC = Name.getError())
509 NameTable.push_back(*Name);
512 return sampleprof_error::success;
515 std::error_code SampleProfileReaderCompactBinary::readNameTable() {
516 auto Size = readNumber<uint64_t>();
517 if (std::error_code EC = Size.getError())
519 NameTable.reserve(*Size);
520 for (uint32_t I = 0; I < *Size; ++I) {
521 auto FID = readNumber<uint64_t>();
522 if (std::error_code EC = FID.getError())
524 NameTable.push_back(std::to_string(*FID));
526 return sampleprof_error::success;
529 std::error_code SampleProfileReaderBinary::readHeader() {
530 Data = reinterpret_cast<const uint8_t *>(Buffer->getBufferStart());
531 End = Data + Buffer->getBufferSize();
533 // Read and check the magic identifier.
534 auto Magic = readNumber<uint64_t>();
535 if (std::error_code EC = Magic.getError())
537 else if (std::error_code EC = verifySPMagic(*Magic))
540 // Read the version number.
541 auto Version = readNumber<uint64_t>();
542 if (std::error_code EC = Version.getError())
544 else if (*Version != SPVersion())
545 return sampleprof_error::unsupported_version;
547 if (std::error_code EC = readSummary())
550 if (std::error_code EC = readNameTable())
552 return sampleprof_error::success;
555 std::error_code SampleProfileReaderCompactBinary::readHeader() {
556 SampleProfileReaderBinary::readHeader();
557 if (std::error_code EC = readFuncOffsetTable())
559 return sampleprof_error::success;
562 std::error_code SampleProfileReaderCompactBinary::readFuncOffsetTable() {
563 auto TableOffset = readUnencodedNumber<uint64_t>();
564 if (std::error_code EC = TableOffset.getError())
567 const uint8_t *SavedData = Data;
568 const uint8_t *TableStart =
569 reinterpret_cast<const uint8_t *>(Buffer->getBufferStart()) +
573 auto Size = readNumber<uint64_t>();
574 if (std::error_code EC = Size.getError())
577 FuncOffsetTable.reserve(*Size);
578 for (uint32_t I = 0; I < *Size; ++I) {
579 auto FName(readStringFromTable());
580 if (std::error_code EC = FName.getError())
583 auto Offset = readNumber<uint64_t>();
584 if (std::error_code EC = Offset.getError())
587 FuncOffsetTable[*FName] = *Offset;
591 return sampleprof_error::success;
594 void SampleProfileReaderCompactBinary::collectFuncsToUse(const Module &M) {
597 StringRef Fname = F.getName().split('.').first;
598 FuncsToUse.insert(Fname);
602 std::error_code SampleProfileReaderBinary::readSummaryEntry(
603 std::vector<ProfileSummaryEntry> &Entries) {
604 auto Cutoff = readNumber<uint64_t>();
605 if (std::error_code EC = Cutoff.getError())
608 auto MinBlockCount = readNumber<uint64_t>();
609 if (std::error_code EC = MinBlockCount.getError())
612 auto NumBlocks = readNumber<uint64_t>();
613 if (std::error_code EC = NumBlocks.getError())
616 Entries.emplace_back(*Cutoff, *MinBlockCount, *NumBlocks);
617 return sampleprof_error::success;
620 std::error_code SampleProfileReaderBinary::readSummary() {
621 auto TotalCount = readNumber<uint64_t>();
622 if (std::error_code EC = TotalCount.getError())
625 auto MaxBlockCount = readNumber<uint64_t>();
626 if (std::error_code EC = MaxBlockCount.getError())
629 auto MaxFunctionCount = readNumber<uint64_t>();
630 if (std::error_code EC = MaxFunctionCount.getError())
633 auto NumBlocks = readNumber<uint64_t>();
634 if (std::error_code EC = NumBlocks.getError())
637 auto NumFunctions = readNumber<uint64_t>();
638 if (std::error_code EC = NumFunctions.getError())
641 auto NumSummaryEntries = readNumber<uint64_t>();
642 if (std::error_code EC = NumSummaryEntries.getError())
645 std::vector<ProfileSummaryEntry> Entries;
646 for (unsigned i = 0; i < *NumSummaryEntries; i++) {
647 std::error_code EC = readSummaryEntry(Entries);
648 if (EC != sampleprof_error::success)
651 Summary = llvm::make_unique<ProfileSummary>(
652 ProfileSummary::PSK_Sample, Entries, *TotalCount, *MaxBlockCount, 0,
653 *MaxFunctionCount, *NumBlocks, *NumFunctions);
655 return sampleprof_error::success;
658 bool SampleProfileReaderRawBinary::hasFormat(const MemoryBuffer &Buffer) {
659 const uint8_t *Data =
660 reinterpret_cast<const uint8_t *>(Buffer.getBufferStart());
661 uint64_t Magic = decodeULEB128(Data);
662 return Magic == SPMagic();
665 bool SampleProfileReaderCompactBinary::hasFormat(const MemoryBuffer &Buffer) {
666 const uint8_t *Data =
667 reinterpret_cast<const uint8_t *>(Buffer.getBufferStart());
668 uint64_t Magic = decodeULEB128(Data);
669 return Magic == SPMagic(SPF_Compact_Binary);
672 std::error_code SampleProfileReaderGCC::skipNextWord() {
674 if (!GcovBuffer.readInt(dummy))
675 return sampleprof_error::truncated;
676 return sampleprof_error::success;
679 template <typename T> ErrorOr<T> SampleProfileReaderGCC::readNumber() {
680 if (sizeof(T) <= sizeof(uint32_t)) {
682 if (GcovBuffer.readInt(Val) && Val <= std::numeric_limits<T>::max())
683 return static_cast<T>(Val);
684 } else if (sizeof(T) <= sizeof(uint64_t)) {
686 if (GcovBuffer.readInt64(Val) && Val <= std::numeric_limits<T>::max())
687 return static_cast<T>(Val);
690 std::error_code EC = sampleprof_error::malformed;
691 reportError(0, EC.message());
695 ErrorOr<StringRef> SampleProfileReaderGCC::readString() {
697 if (!GcovBuffer.readString(Str))
698 return sampleprof_error::truncated;
702 std::error_code SampleProfileReaderGCC::readHeader() {
703 // Read the magic identifier.
704 if (!GcovBuffer.readGCDAFormat())
705 return sampleprof_error::unrecognized_format;
707 // Read the version number. Note - the GCC reader does not validate this
708 // version, but the profile creator generates v704.
709 GCOV::GCOVVersion version;
710 if (!GcovBuffer.readGCOVVersion(version))
711 return sampleprof_error::unrecognized_format;
713 if (version != GCOV::V704)
714 return sampleprof_error::unsupported_version;
716 // Skip the empty integer.
717 if (std::error_code EC = skipNextWord())
720 return sampleprof_error::success;
723 std::error_code SampleProfileReaderGCC::readSectionTag(uint32_t Expected) {
725 if (!GcovBuffer.readInt(Tag))
726 return sampleprof_error::truncated;
729 return sampleprof_error::malformed;
731 if (std::error_code EC = skipNextWord())
734 return sampleprof_error::success;
737 std::error_code SampleProfileReaderGCC::readNameTable() {
738 if (std::error_code EC = readSectionTag(GCOVTagAFDOFileNames))
742 if (!GcovBuffer.readInt(Size))
743 return sampleprof_error::truncated;
745 for (uint32_t I = 0; I < Size; ++I) {
747 if (!GcovBuffer.readString(Str))
748 return sampleprof_error::truncated;
749 Names.push_back(Str);
752 return sampleprof_error::success;
755 std::error_code SampleProfileReaderGCC::readFunctionProfiles() {
756 if (std::error_code EC = readSectionTag(GCOVTagAFDOFunction))
759 uint32_t NumFunctions;
760 if (!GcovBuffer.readInt(NumFunctions))
761 return sampleprof_error::truncated;
763 InlineCallStack Stack;
764 for (uint32_t I = 0; I < NumFunctions; ++I)
765 if (std::error_code EC = readOneFunctionProfile(Stack, true, 0))
769 return sampleprof_error::success;
772 std::error_code SampleProfileReaderGCC::readOneFunctionProfile(
773 const InlineCallStack &InlineStack, bool Update, uint32_t Offset) {
774 uint64_t HeadCount = 0;
775 if (InlineStack.size() == 0)
776 if (!GcovBuffer.readInt64(HeadCount))
777 return sampleprof_error::truncated;
780 if (!GcovBuffer.readInt(NameIdx))
781 return sampleprof_error::truncated;
783 StringRef Name(Names[NameIdx]);
785 uint32_t NumPosCounts;
786 if (!GcovBuffer.readInt(NumPosCounts))
787 return sampleprof_error::truncated;
789 uint32_t NumCallsites;
790 if (!GcovBuffer.readInt(NumCallsites))
791 return sampleprof_error::truncated;
793 FunctionSamples *FProfile = nullptr;
794 if (InlineStack.size() == 0) {
795 // If this is a top function that we have already processed, do not
796 // update its profile again. This happens in the presence of
797 // function aliases. Since these aliases share the same function
798 // body, there will be identical replicated profiles for the
799 // original function. In this case, we simply not bother updating
800 // the profile of the original function.
801 FProfile = &Profiles[Name];
802 FProfile->addHeadSamples(HeadCount);
803 if (FProfile->getTotalSamples() > 0)
806 // Otherwise, we are reading an inlined instance. The top of the
807 // inline stack contains the profile of the caller. Insert this
808 // callee in the caller's CallsiteMap.
809 FunctionSamples *CallerProfile = InlineStack.front();
810 uint32_t LineOffset = Offset >> 16;
811 uint32_t Discriminator = Offset & 0xffff;
812 FProfile = &CallerProfile->functionSamplesAt(
813 LineLocation(LineOffset, Discriminator))[Name];
815 FProfile->setName(Name);
817 for (uint32_t I = 0; I < NumPosCounts; ++I) {
819 if (!GcovBuffer.readInt(Offset))
820 return sampleprof_error::truncated;
823 if (!GcovBuffer.readInt(NumTargets))
824 return sampleprof_error::truncated;
827 if (!GcovBuffer.readInt64(Count))
828 return sampleprof_error::truncated;
830 // The line location is encoded in the offset as:
831 // high 16 bits: line offset to the start of the function.
832 // low 16 bits: discriminator.
833 uint32_t LineOffset = Offset >> 16;
834 uint32_t Discriminator = Offset & 0xffff;
836 InlineCallStack NewStack;
837 NewStack.push_back(FProfile);
838 NewStack.insert(NewStack.end(), InlineStack.begin(), InlineStack.end());
840 // Walk up the inline stack, adding the samples on this line to
841 // the total sample count of the callers in the chain.
842 for (auto CallerProfile : NewStack)
843 CallerProfile->addTotalSamples(Count);
845 // Update the body samples for the current profile.
846 FProfile->addBodySamples(LineOffset, Discriminator, Count);
849 // Process the list of functions called at an indirect call site.
850 // These are all the targets that a function pointer (or virtual
851 // function) resolved at runtime.
852 for (uint32_t J = 0; J < NumTargets; J++) {
854 if (!GcovBuffer.readInt(HistVal))
855 return sampleprof_error::truncated;
857 if (HistVal != HIST_TYPE_INDIR_CALL_TOPN)
858 return sampleprof_error::malformed;
861 if (!GcovBuffer.readInt64(TargetIdx))
862 return sampleprof_error::truncated;
863 StringRef TargetName(Names[TargetIdx]);
865 uint64_t TargetCount;
866 if (!GcovBuffer.readInt64(TargetCount))
867 return sampleprof_error::truncated;
870 FProfile->addCalledTargetSamples(LineOffset, Discriminator,
871 TargetName, TargetCount);
875 // Process all the inlined callers into the current function. These
876 // are all the callsites that were inlined into this function.
877 for (uint32_t I = 0; I < NumCallsites; I++) {
878 // The offset is encoded as:
879 // high 16 bits: line offset to the start of the function.
880 // low 16 bits: discriminator.
882 if (!GcovBuffer.readInt(Offset))
883 return sampleprof_error::truncated;
884 InlineCallStack NewStack;
885 NewStack.push_back(FProfile);
886 NewStack.insert(NewStack.end(), InlineStack.begin(), InlineStack.end());
887 if (std::error_code EC = readOneFunctionProfile(NewStack, Update, Offset))
891 return sampleprof_error::success;
894 /// Read a GCC AutoFDO profile.
896 /// This format is generated by the Linux Perf conversion tool at
897 /// https://github.com/google/autofdo.
898 std::error_code SampleProfileReaderGCC::read() {
899 // Read the string table.
900 if (std::error_code EC = readNameTable())
903 // Read the source profile.
904 if (std::error_code EC = readFunctionProfiles())
907 return sampleprof_error::success;
910 bool SampleProfileReaderGCC::hasFormat(const MemoryBuffer &Buffer) {
911 StringRef Magic(reinterpret_cast<const char *>(Buffer.getBufferStart()));
912 return Magic == "adcg*704";
915 std::error_code SampleProfileReaderItaniumRemapper::read() {
916 // If the underlying data is in compact format, we can't remap it because
917 // we don't know what the original function names were.
918 if (getFormat() == SPF_Compact_Binary) {
919 Ctx.diagnose(DiagnosticInfoSampleProfile(
920 Buffer->getBufferIdentifier(),
921 "Profile data remapping cannot be applied to profile data "
922 "in compact format (original mangled names are not available).",
924 return sampleprof_error::success;
927 if (Error E = Remappings.read(*Buffer)) {
929 std::move(E), [&](const SymbolRemappingParseError &ParseError) {
930 reportError(ParseError.getLineNum(), ParseError.getMessage());
932 return sampleprof_error::malformed;
935 for (auto &Sample : getProfiles())
936 if (auto Key = Remappings.insert(Sample.first()))
937 SampleMap.insert({Key, &Sample.second});
939 return sampleprof_error::success;
943 SampleProfileReaderItaniumRemapper::getSamplesFor(StringRef Fname) {
944 if (auto Key = Remappings.lookup(Fname))
945 return SampleMap.lookup(Key);
946 return SampleProfileReader::getSamplesFor(Fname);
949 /// Prepare a memory buffer for the contents of \p Filename.
951 /// \returns an error code indicating the status of the buffer.
952 static ErrorOr<std::unique_ptr<MemoryBuffer>>
953 setupMemoryBuffer(const Twine &Filename) {
954 auto BufferOrErr = MemoryBuffer::getFileOrSTDIN(Filename);
955 if (std::error_code EC = BufferOrErr.getError())
957 auto Buffer = std::move(BufferOrErr.get());
959 // Sanity check the file.
960 if (uint64_t(Buffer->getBufferSize()) > std::numeric_limits<uint32_t>::max())
961 return sampleprof_error::too_large;
963 return std::move(Buffer);
966 /// Create a sample profile reader based on the format of the input file.
968 /// \param Filename The file to open.
970 /// \param C The LLVM context to use to emit diagnostics.
972 /// \returns an error code indicating the status of the created reader.
973 ErrorOr<std::unique_ptr<SampleProfileReader>>
974 SampleProfileReader::create(const Twine &Filename, LLVMContext &C) {
975 auto BufferOrError = setupMemoryBuffer(Filename);
976 if (std::error_code EC = BufferOrError.getError())
978 return create(BufferOrError.get(), C);
981 /// Create a sample profile remapper from the given input, to remap the
982 /// function names in the given profile data.
984 /// \param Filename The file to open.
986 /// \param C The LLVM context to use to emit diagnostics.
988 /// \param Underlying The underlying profile data reader to remap.
990 /// \returns an error code indicating the status of the created reader.
991 ErrorOr<std::unique_ptr<SampleProfileReader>>
992 SampleProfileReaderItaniumRemapper::create(
993 const Twine &Filename, LLVMContext &C,
994 std::unique_ptr<SampleProfileReader> Underlying) {
995 auto BufferOrError = setupMemoryBuffer(Filename);
996 if (std::error_code EC = BufferOrError.getError())
998 return llvm::make_unique<SampleProfileReaderItaniumRemapper>(
999 std::move(BufferOrError.get()), C, std::move(Underlying));
1002 /// Create a sample profile reader based on the format of the input data.
1004 /// \param B The memory buffer to create the reader from (assumes ownership).
1006 /// \param C The LLVM context to use to emit diagnostics.
1008 /// \returns an error code indicating the status of the created reader.
1009 ErrorOr<std::unique_ptr<SampleProfileReader>>
1010 SampleProfileReader::create(std::unique_ptr<MemoryBuffer> &B, LLVMContext &C) {
1011 std::unique_ptr<SampleProfileReader> Reader;
1012 if (SampleProfileReaderRawBinary::hasFormat(*B))
1013 Reader.reset(new SampleProfileReaderRawBinary(std::move(B), C));
1014 else if (SampleProfileReaderCompactBinary::hasFormat(*B))
1015 Reader.reset(new SampleProfileReaderCompactBinary(std::move(B), C));
1016 else if (SampleProfileReaderGCC::hasFormat(*B))
1017 Reader.reset(new SampleProfileReaderGCC(std::move(B), C));
1018 else if (SampleProfileReaderText::hasFormat(*B))
1019 Reader.reset(new SampleProfileReaderText(std::move(B), C));
1021 return sampleprof_error::unrecognized_format;
1023 FunctionSamples::Format = Reader->getFormat();
1024 if (std::error_code EC = Reader->readHeader())
1027 return std::move(Reader);
1030 // For text and GCC file formats, we compute the summary after reading the
1031 // profile. Binary format has the profile summary in its header.
1032 void SampleProfileReader::computeSummary() {
1033 SampleProfileSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs);
1034 for (const auto &I : Profiles) {
1035 const FunctionSamples &Profile = I.second;
1036 Builder.addRecord(Profile);
1038 Summary = Builder.getSummary();