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/SmallVector.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/ErrorOr.h"
28 #include "llvm/Support/LEB128.h"
29 #include "llvm/Support/LineIterator.h"
30 #include "llvm/Support/MemoryBuffer.h"
32 using namespace llvm::sampleprof;
35 /// \brief Dump the function profile for \p FName.
37 /// \param FName Name of the function to print.
38 /// \param OS Stream to emit the output to.
39 void SampleProfileReader::dumpFunctionProfile(StringRef FName,
41 OS << "Function: " << FName << ": ";
42 Profiles[FName].print(OS);
45 /// \brief Dump all the function profiles found on stream \p OS.
46 void SampleProfileReader::dump(raw_ostream &OS) {
47 for (const auto &I : Profiles)
48 dumpFunctionProfile(I.getKey(), OS);
51 /// \brief Parse \p Input as function head.
53 /// Parse one line of \p Input, and update function name in \p FName,
54 /// function's total sample count in \p NumSamples, function's entry
55 /// count in \p NumHeadSamples.
57 /// \returns true if parsing is successful.
58 static bool ParseHead(const StringRef &Input, StringRef &FName,
59 uint64_t &NumSamples, uint64_t &NumHeadSamples) {
62 size_t n2 = Input.rfind(':');
63 size_t n1 = Input.rfind(':', n2 - 1);
64 FName = Input.substr(0, n1);
65 if (Input.substr(n1 + 1, n2 - n1 - 1).getAsInteger(10, NumSamples))
67 if (Input.substr(n2 + 1).getAsInteger(10, NumHeadSamples))
73 /// \brief Returns true if line offset \p L is legal (only has 16 bits).
74 static bool isOffsetLegal(unsigned L) {
75 return (L & 0xffff) == L;
78 /// \brief Parse \p Input as line sample.
80 /// \param Input input line.
81 /// \param IsCallsite true if the line represents an inlined callsite.
82 /// \param Depth the depth of the inline stack.
83 /// \param NumSamples total samples of the line/inlined callsite.
84 /// \param LineOffset line offset to the start of the function.
85 /// \param Discriminator discriminator of the line.
86 /// \param TargetCountMap map from indirect call target to count.
88 /// returns true if parsing is successful.
89 static bool ParseLine(const StringRef &Input, bool &IsCallsite, uint32_t &Depth,
90 uint64_t &NumSamples, uint32_t &LineOffset,
91 uint32_t &Discriminator, StringRef &CalleeName,
92 DenseMap<StringRef, uint64_t> &TargetCountMap) {
93 for (Depth = 0; Input[Depth] == ' '; Depth++)
98 size_t n1 = Input.find(':');
99 StringRef Loc = Input.substr(Depth, n1 - Depth);
100 size_t n2 = Loc.find('.');
101 if (n2 == StringRef::npos) {
102 if (Loc.getAsInteger(10, LineOffset) || !isOffsetLegal(LineOffset))
106 if (Loc.substr(0, n2).getAsInteger(10, LineOffset))
108 if (Loc.substr(n2 + 1).getAsInteger(10, Discriminator))
112 StringRef Rest = Input.substr(n1 + 2);
113 if (Rest[0] >= '0' && Rest[0] <= '9') {
115 size_t n3 = Rest.find(' ');
116 if (n3 == StringRef::npos) {
117 if (Rest.getAsInteger(10, NumSamples))
120 if (Rest.substr(0, n3).getAsInteger(10, NumSamples))
123 while (n3 != StringRef::npos) {
124 n3 += Rest.substr(n3).find_first_not_of(' ');
125 Rest = Rest.substr(n3);
127 StringRef pair = Rest;
128 if (n3 != StringRef::npos) {
129 pair = Rest.substr(0, n3);
131 size_t n4 = pair.find(':');
133 if (pair.substr(n4 + 1).getAsInteger(10, count))
135 TargetCountMap[pair.substr(0, n4)] = count;
139 size_t n3 = Rest.find_last_of(':');
140 CalleeName = Rest.substr(0, n3);
141 if (Rest.substr(n3 + 1).getAsInteger(10, NumSamples))
147 /// \brief Load samples from a text file.
149 /// See the documentation at the top of the file for an explanation of
150 /// the expected format.
152 /// \returns true if the file was loaded successfully, false otherwise.
153 std::error_code SampleProfileReaderText::read() {
154 line_iterator LineIt(*Buffer, /*SkipBlanks=*/true, '#');
156 InlineCallStack InlineStack;
158 for (; !LineIt.is_at_eof(); ++LineIt) {
159 if ((*LineIt)[(*LineIt).find_first_not_of(' ')] == '#')
161 // Read the header of each function.
163 // Note that for function identifiers we are actually expecting
164 // mangled names, but we may not always get them. This happens when
165 // the compiler decides not to emit the function (e.g., it was inlined
166 // and removed). In this case, the binary will not have the linkage
167 // name for the function, so the profiler will emit the function's
168 // unmangled name, which may contain characters like ':' and '>' in its
169 // name (member functions, templates, etc).
171 // The only requirement we place on the identifier, then, is that it
172 // should not begin with a number.
173 if ((*LineIt)[0] != ' ') {
174 uint64_t NumSamples, NumHeadSamples;
176 if (!ParseHead(*LineIt, FName, NumSamples, NumHeadSamples)) {
177 reportError(LineIt.line_number(),
178 "Expected 'mangled_name:NUM:NUM', found " + *LineIt);
179 return sampleprof_error::malformed;
181 Profiles[FName] = FunctionSamples();
182 FunctionSamples &FProfile = Profiles[FName];
183 FProfile.addTotalSamples(NumSamples);
184 FProfile.addHeadSamples(NumHeadSamples);
186 InlineStack.push_back(&FProfile);
190 DenseMap<StringRef, uint64_t> TargetCountMap;
192 uint32_t Depth, LineOffset, Discriminator;
193 if (!ParseLine(*LineIt, IsCallsite, Depth, NumSamples, LineOffset,
194 Discriminator, FName, TargetCountMap)) {
195 reportError(LineIt.line_number(),
196 "Expected 'NUM[.NUM]: NUM[ mangled_name:NUM]*', found " +
198 return sampleprof_error::malformed;
201 while (InlineStack.size() > Depth) {
202 InlineStack.pop_back();
204 FunctionSamples &FSamples = InlineStack.back()->functionSamplesAt(
205 CallsiteLocation(LineOffset, Discriminator, FName));
206 FSamples.addTotalSamples(NumSamples);
207 InlineStack.push_back(&FSamples);
209 while (InlineStack.size() > Depth) {
210 InlineStack.pop_back();
212 FunctionSamples &FProfile = *InlineStack.back();
213 for (const auto &name_count : TargetCountMap) {
214 FProfile.addCalledTargetSamples(LineOffset, Discriminator,
215 name_count.first, name_count.second);
217 FProfile.addBodySamples(LineOffset, Discriminator, NumSamples);
222 return sampleprof_error::success;
225 template <typename T> ErrorOr<T> SampleProfileReaderBinary::readNumber() {
226 unsigned NumBytesRead = 0;
228 uint64_t Val = decodeULEB128(Data, &NumBytesRead);
230 if (Val > std::numeric_limits<T>::max())
231 EC = sampleprof_error::malformed;
232 else if (Data + NumBytesRead > End)
233 EC = sampleprof_error::truncated;
235 EC = sampleprof_error::success;
238 reportError(0, EC.message());
242 Data += NumBytesRead;
243 return static_cast<T>(Val);
246 ErrorOr<StringRef> SampleProfileReaderBinary::readString() {
248 StringRef Str(reinterpret_cast<const char *>(Data));
249 if (Data + Str.size() + 1 > End) {
250 EC = sampleprof_error::truncated;
251 reportError(0, EC.message());
255 Data += Str.size() + 1;
259 ErrorOr<StringRef> SampleProfileReaderBinary::readStringFromTable() {
261 auto Idx = readNumber<uint32_t>();
262 if (std::error_code EC = Idx.getError())
264 if (*Idx >= NameTable.size())
265 return sampleprof_error::truncated_name_table;
266 return NameTable[*Idx];
270 SampleProfileReaderBinary::readProfile(FunctionSamples &FProfile) {
271 auto NumSamples = readNumber<uint64_t>();
272 if (std::error_code EC = NumSamples.getError())
274 FProfile.addTotalSamples(*NumSamples);
276 // Read the samples in the body.
277 auto NumRecords = readNumber<uint32_t>();
278 if (std::error_code EC = NumRecords.getError())
281 for (uint32_t I = 0; I < *NumRecords; ++I) {
282 auto LineOffset = readNumber<uint64_t>();
283 if (std::error_code EC = LineOffset.getError())
286 if (!isOffsetLegal(*LineOffset)) {
287 return std::error_code();
290 auto Discriminator = readNumber<uint64_t>();
291 if (std::error_code EC = Discriminator.getError())
294 auto NumSamples = readNumber<uint64_t>();
295 if (std::error_code EC = NumSamples.getError())
298 auto NumCalls = readNumber<uint32_t>();
299 if (std::error_code EC = NumCalls.getError())
302 for (uint32_t J = 0; J < *NumCalls; ++J) {
303 auto CalledFunction(readStringFromTable());
304 if (std::error_code EC = CalledFunction.getError())
307 auto CalledFunctionSamples = readNumber<uint64_t>();
308 if (std::error_code EC = CalledFunctionSamples.getError())
311 FProfile.addCalledTargetSamples(*LineOffset, *Discriminator,
312 *CalledFunction, *CalledFunctionSamples);
315 FProfile.addBodySamples(*LineOffset, *Discriminator, *NumSamples);
318 // Read all the samples for inlined function calls.
319 auto NumCallsites = readNumber<uint32_t>();
320 if (std::error_code EC = NumCallsites.getError())
323 for (uint32_t J = 0; J < *NumCallsites; ++J) {
324 auto LineOffset = readNumber<uint64_t>();
325 if (std::error_code EC = LineOffset.getError())
328 auto Discriminator = readNumber<uint64_t>();
329 if (std::error_code EC = Discriminator.getError())
332 auto FName(readStringFromTable());
333 if (std::error_code EC = FName.getError())
336 FunctionSamples &CalleeProfile = FProfile.functionSamplesAt(
337 CallsiteLocation(*LineOffset, *Discriminator, *FName));
338 if (std::error_code EC = readProfile(CalleeProfile))
342 return sampleprof_error::success;
345 std::error_code SampleProfileReaderBinary::read() {
347 auto NumHeadSamples = readNumber<uint64_t>();
348 if (std::error_code EC = NumHeadSamples.getError())
351 auto FName(readStringFromTable());
352 if (std::error_code EC = FName.getError())
355 Profiles[*FName] = FunctionSamples();
356 FunctionSamples &FProfile = Profiles[*FName];
358 FProfile.addHeadSamples(*NumHeadSamples);
360 if (std::error_code EC = readProfile(FProfile))
364 return sampleprof_error::success;
367 std::error_code SampleProfileReaderBinary::readHeader() {
368 Data = reinterpret_cast<const uint8_t *>(Buffer->getBufferStart());
369 End = Data + Buffer->getBufferSize();
371 // Read and check the magic identifier.
372 auto Magic = readNumber<uint64_t>();
373 if (std::error_code EC = Magic.getError())
375 else if (*Magic != SPMagic())
376 return sampleprof_error::bad_magic;
378 // Read the version number.
379 auto Version = readNumber<uint64_t>();
380 if (std::error_code EC = Version.getError())
382 else if (*Version != SPVersion())
383 return sampleprof_error::unsupported_version;
385 // Read the name table.
386 auto Size = readNumber<uint32_t>();
387 if (std::error_code EC = Size.getError())
389 NameTable.reserve(*Size);
390 for (uint32_t I = 0; I < *Size; ++I) {
391 auto Name(readString());
392 if (std::error_code EC = Name.getError())
394 NameTable.push_back(*Name);
397 return sampleprof_error::success;
400 bool SampleProfileReaderBinary::hasFormat(const MemoryBuffer &Buffer) {
401 const uint8_t *Data =
402 reinterpret_cast<const uint8_t *>(Buffer.getBufferStart());
403 uint64_t Magic = decodeULEB128(Data);
404 return Magic == SPMagic();
407 std::error_code SampleProfileReaderGCC::skipNextWord() {
409 if (!GcovBuffer.readInt(dummy))
410 return sampleprof_error::truncated;
411 return sampleprof_error::success;
414 template <typename T> ErrorOr<T> SampleProfileReaderGCC::readNumber() {
415 if (sizeof(T) <= sizeof(uint32_t)) {
417 if (GcovBuffer.readInt(Val) && Val <= std::numeric_limits<T>::max())
418 return static_cast<T>(Val);
419 } else if (sizeof(T) <= sizeof(uint64_t)) {
421 if (GcovBuffer.readInt64(Val) && Val <= std::numeric_limits<T>::max())
422 return static_cast<T>(Val);
425 std::error_code EC = sampleprof_error::malformed;
426 reportError(0, EC.message());
430 ErrorOr<StringRef> SampleProfileReaderGCC::readString() {
432 if (!GcovBuffer.readString(Str))
433 return sampleprof_error::truncated;
437 std::error_code SampleProfileReaderGCC::readHeader() {
438 // Read the magic identifier.
439 if (!GcovBuffer.readGCDAFormat())
440 return sampleprof_error::unrecognized_format;
442 // Read the version number. Note - the GCC reader does not validate this
443 // version, but the profile creator generates v704.
444 GCOV::GCOVVersion version;
445 if (!GcovBuffer.readGCOVVersion(version))
446 return sampleprof_error::unrecognized_format;
448 if (version != GCOV::V704)
449 return sampleprof_error::unsupported_version;
451 // Skip the empty integer.
452 if (std::error_code EC = skipNextWord())
455 return sampleprof_error::success;
458 std::error_code SampleProfileReaderGCC::readSectionTag(uint32_t Expected) {
460 if (!GcovBuffer.readInt(Tag))
461 return sampleprof_error::truncated;
464 return sampleprof_error::malformed;
466 if (std::error_code EC = skipNextWord())
469 return sampleprof_error::success;
472 std::error_code SampleProfileReaderGCC::readNameTable() {
473 if (std::error_code EC = readSectionTag(GCOVTagAFDOFileNames))
477 if (!GcovBuffer.readInt(Size))
478 return sampleprof_error::truncated;
480 for (uint32_t I = 0; I < Size; ++I) {
482 if (!GcovBuffer.readString(Str))
483 return sampleprof_error::truncated;
484 Names.push_back(Str);
487 return sampleprof_error::success;
490 std::error_code SampleProfileReaderGCC::readFunctionProfiles() {
491 if (std::error_code EC = readSectionTag(GCOVTagAFDOFunction))
494 uint32_t NumFunctions;
495 if (!GcovBuffer.readInt(NumFunctions))
496 return sampleprof_error::truncated;
498 InlineCallStack Stack;
499 for (uint32_t I = 0; I < NumFunctions; ++I)
500 if (std::error_code EC = readOneFunctionProfile(Stack, true, 0))
503 return sampleprof_error::success;
506 std::error_code SampleProfileReaderGCC::readOneFunctionProfile(
507 const InlineCallStack &InlineStack, bool Update, uint32_t Offset) {
508 uint64_t HeadCount = 0;
509 if (InlineStack.size() == 0)
510 if (!GcovBuffer.readInt64(HeadCount))
511 return sampleprof_error::truncated;
514 if (!GcovBuffer.readInt(NameIdx))
515 return sampleprof_error::truncated;
517 StringRef Name(Names[NameIdx]);
519 uint32_t NumPosCounts;
520 if (!GcovBuffer.readInt(NumPosCounts))
521 return sampleprof_error::truncated;
523 uint32_t NumCallsites;
524 if (!GcovBuffer.readInt(NumCallsites))
525 return sampleprof_error::truncated;
527 FunctionSamples *FProfile = nullptr;
528 if (InlineStack.size() == 0) {
529 // If this is a top function that we have already processed, do not
530 // update its profile again. This happens in the presence of
531 // function aliases. Since these aliases share the same function
532 // body, there will be identical replicated profiles for the
533 // original function. In this case, we simply not bother updating
534 // the profile of the original function.
535 FProfile = &Profiles[Name];
536 FProfile->addHeadSamples(HeadCount);
537 if (FProfile->getTotalSamples() > 0)
540 // Otherwise, we are reading an inlined instance. The top of the
541 // inline stack contains the profile of the caller. Insert this
542 // callee in the caller's CallsiteMap.
543 FunctionSamples *CallerProfile = InlineStack.front();
544 uint32_t LineOffset = Offset >> 16;
545 uint32_t Discriminator = Offset & 0xffff;
546 FProfile = &CallerProfile->functionSamplesAt(
547 CallsiteLocation(LineOffset, Discriminator, Name));
550 for (uint32_t I = 0; I < NumPosCounts; ++I) {
552 if (!GcovBuffer.readInt(Offset))
553 return sampleprof_error::truncated;
556 if (!GcovBuffer.readInt(NumTargets))
557 return sampleprof_error::truncated;
560 if (!GcovBuffer.readInt64(Count))
561 return sampleprof_error::truncated;
563 // The line location is encoded in the offset as:
564 // high 16 bits: line offset to the start of the function.
565 // low 16 bits: discriminator.
566 uint32_t LineOffset = Offset >> 16;
567 uint32_t Discriminator = Offset & 0xffff;
569 InlineCallStack NewStack;
570 NewStack.push_back(FProfile);
571 NewStack.insert(NewStack.end(), InlineStack.begin(), InlineStack.end());
573 // Walk up the inline stack, adding the samples on this line to
574 // the total sample count of the callers in the chain.
575 for (auto CallerProfile : NewStack)
576 CallerProfile->addTotalSamples(Count);
578 // Update the body samples for the current profile.
579 FProfile->addBodySamples(LineOffset, Discriminator, Count);
582 // Process the list of functions called at an indirect call site.
583 // These are all the targets that a function pointer (or virtual
584 // function) resolved at runtime.
585 for (uint32_t J = 0; J < NumTargets; J++) {
587 if (!GcovBuffer.readInt(HistVal))
588 return sampleprof_error::truncated;
590 if (HistVal != HIST_TYPE_INDIR_CALL_TOPN)
591 return sampleprof_error::malformed;
594 if (!GcovBuffer.readInt64(TargetIdx))
595 return sampleprof_error::truncated;
596 StringRef TargetName(Names[TargetIdx]);
598 uint64_t TargetCount;
599 if (!GcovBuffer.readInt64(TargetCount))
600 return sampleprof_error::truncated;
603 FunctionSamples &TargetProfile = Profiles[TargetName];
604 TargetProfile.addCalledTargetSamples(LineOffset, Discriminator,
605 TargetName, TargetCount);
610 // Process all the inlined callers into the current function. These
611 // are all the callsites that were inlined into this function.
612 for (uint32_t I = 0; I < NumCallsites; I++) {
613 // The offset is encoded as:
614 // high 16 bits: line offset to the start of the function.
615 // low 16 bits: discriminator.
617 if (!GcovBuffer.readInt(Offset))
618 return sampleprof_error::truncated;
619 InlineCallStack NewStack;
620 NewStack.push_back(FProfile);
621 NewStack.insert(NewStack.end(), InlineStack.begin(), InlineStack.end());
622 if (std::error_code EC = readOneFunctionProfile(NewStack, Update, Offset))
626 return sampleprof_error::success;
629 /// \brief Read a GCC AutoFDO profile.
631 /// This format is generated by the Linux Perf conversion tool at
632 /// https://github.com/google/autofdo.
633 std::error_code SampleProfileReaderGCC::read() {
634 // Read the string table.
635 if (std::error_code EC = readNameTable())
638 // Read the source profile.
639 if (std::error_code EC = readFunctionProfiles())
642 return sampleprof_error::success;
645 bool SampleProfileReaderGCC::hasFormat(const MemoryBuffer &Buffer) {
646 StringRef Magic(reinterpret_cast<const char *>(Buffer.getBufferStart()));
647 return Magic == "adcg*704";
650 /// \brief Prepare a memory buffer for the contents of \p Filename.
652 /// \returns an error code indicating the status of the buffer.
653 static ErrorOr<std::unique_ptr<MemoryBuffer>>
654 setupMemoryBuffer(std::string Filename) {
655 auto BufferOrErr = MemoryBuffer::getFileOrSTDIN(Filename);
656 if (std::error_code EC = BufferOrErr.getError())
658 auto Buffer = std::move(BufferOrErr.get());
660 // Sanity check the file.
661 if (Buffer->getBufferSize() > std::numeric_limits<uint32_t>::max())
662 return sampleprof_error::too_large;
664 return std::move(Buffer);
667 /// \brief Create a sample profile reader based on the format of the input file.
669 /// \param Filename The file to open.
671 /// \param Reader The reader to instantiate according to \p Filename's format.
673 /// \param C The LLVM context to use to emit diagnostics.
675 /// \returns an error code indicating the status of the created reader.
676 ErrorOr<std::unique_ptr<SampleProfileReader>>
677 SampleProfileReader::create(StringRef Filename, LLVMContext &C) {
678 auto BufferOrError = setupMemoryBuffer(Filename);
679 if (std::error_code EC = BufferOrError.getError())
682 auto Buffer = std::move(BufferOrError.get());
683 std::unique_ptr<SampleProfileReader> Reader;
684 if (SampleProfileReaderBinary::hasFormat(*Buffer))
685 Reader.reset(new SampleProfileReaderBinary(std::move(Buffer), C));
686 else if (SampleProfileReaderGCC::hasFormat(*Buffer))
687 Reader.reset(new SampleProfileReaderGCC(std::move(Buffer), C));
689 Reader.reset(new SampleProfileReaderText(std::move(Buffer), C));
691 if (std::error_code EC = Reader->readHeader())
694 return std::move(Reader);