1 //=-- InstrProf.cpp - Instrumented profiling format support -----------------=//
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 contains support for clang's instrumentation based PGO and
13 //===----------------------------------------------------------------------===//
15 #include "llvm/ProfileData/InstrProf.h"
16 #include "llvm/ADT/StringExtras.h"
17 #include "llvm/IR/Constants.h"
18 #include "llvm/IR/Function.h"
19 #include "llvm/IR/GlobalVariable.h"
20 #include "llvm/IR/Module.h"
21 #include "llvm/Support/Compression.h"
22 #include "llvm/Support/ErrorHandling.h"
23 #include "llvm/Support/LEB128.h"
24 #include "llvm/Support/ManagedStatic.h"
29 class InstrProfErrorCategoryType : public std::error_category {
30 const char *name() const LLVM_NOEXCEPT override { return "llvm.instrprof"; }
31 std::string message(int IE) const override {
32 instrprof_error E = static_cast<instrprof_error>(IE);
34 case instrprof_error::success:
36 case instrprof_error::eof:
38 case instrprof_error::unrecognized_format:
39 return "Unrecognized instrumentation profile encoding format";
40 case instrprof_error::bad_magic:
41 return "Invalid instrumentation profile data (bad magic)";
42 case instrprof_error::bad_header:
43 return "Invalid instrumentation profile data (file header is corrupt)";
44 case instrprof_error::unsupported_version:
45 return "Unsupported instrumentation profile format version";
46 case instrprof_error::unsupported_hash_type:
47 return "Unsupported instrumentation profile hash type";
48 case instrprof_error::too_large:
49 return "Too much profile data";
50 case instrprof_error::truncated:
51 return "Truncated profile data";
52 case instrprof_error::malformed:
53 return "Malformed instrumentation profile data";
54 case instrprof_error::unknown_function:
55 return "No profile data available for function";
56 case instrprof_error::hash_mismatch:
57 return "Function control flow change detected (hash mismatch)";
58 case instrprof_error::count_mismatch:
59 return "Function basic block count change detected (counter mismatch)";
60 case instrprof_error::counter_overflow:
61 return "Counter overflow";
62 case instrprof_error::value_site_count_mismatch:
63 return "Function value site count change detected (counter mismatch)";
65 llvm_unreachable("A value of instrprof_error has no message.");
70 static ManagedStatic<InstrProfErrorCategoryType> ErrorCategory;
72 const std::error_category &llvm::instrprof_category() {
73 return *ErrorCategory;
78 std::string getPGOFuncName(StringRef RawFuncName,
79 GlobalValue::LinkageTypes Linkage,
81 uint64_t Version LLVM_ATTRIBUTE_UNUSED) {
83 // Function names may be prefixed with a binary '1' to indicate
84 // that the backend should not modify the symbols due to any platform
85 // naming convention. Do not include that '1' in the PGO profile name.
86 if (RawFuncName[0] == '\1')
87 RawFuncName = RawFuncName.substr(1);
89 std::string FuncName = RawFuncName;
90 if (llvm::GlobalValue::isLocalLinkage(Linkage)) {
91 // For local symbols, prepend the main file name to distinguish them.
92 // Do not include the full path in the file name since there's no guarantee
93 // that it will stay the same, e.g., if the files are checked out from
94 // version control in different locations.
96 FuncName = FuncName.insert(0, "<unknown>:");
98 FuncName = FuncName.insert(0, FileName.str() + ":");
103 std::string getPGOFuncName(const Function &F, uint64_t Version) {
104 return getPGOFuncName(F.getName(), F.getLinkage(), F.getParent()->getName(),
108 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, StringRef FileName) {
109 if (FileName.empty())
111 // Drop the file name including ':'. See also getPGOFuncName.
112 if (PGOFuncName.startswith(FileName))
113 PGOFuncName = PGOFuncName.drop_front(FileName.size() + 1);
117 // \p FuncName is the string used as profile lookup key for the function. A
118 // symbol is created to hold the name. Return the legalized symbol name.
119 static std::string getPGOFuncNameVarName(StringRef FuncName,
120 GlobalValue::LinkageTypes Linkage) {
121 std::string VarName = getInstrProfNameVarPrefix();
124 if (!GlobalValue::isLocalLinkage(Linkage))
127 // Now fix up illegal chars in local VarName that may upset the assembler.
128 const char *InvalidChars = "-:<>\"'";
129 size_t found = VarName.find_first_of(InvalidChars);
130 while (found != std::string::npos) {
131 VarName[found] = '_';
132 found = VarName.find_first_of(InvalidChars, found + 1);
137 GlobalVariable *createPGOFuncNameVar(Module &M,
138 GlobalValue::LinkageTypes Linkage,
139 StringRef FuncName) {
141 // We generally want to match the function's linkage, but available_externally
142 // and extern_weak both have the wrong semantics, and anything that doesn't
143 // need to link across compilation units doesn't need to be visible at all.
144 if (Linkage == GlobalValue::ExternalWeakLinkage)
145 Linkage = GlobalValue::LinkOnceAnyLinkage;
146 else if (Linkage == GlobalValue::AvailableExternallyLinkage)
147 Linkage = GlobalValue::LinkOnceODRLinkage;
148 else if (Linkage == GlobalValue::InternalLinkage ||
149 Linkage == GlobalValue::ExternalLinkage)
150 Linkage = GlobalValue::PrivateLinkage;
152 auto *Value = ConstantDataArray::getString(M.getContext(), FuncName, false);
154 new GlobalVariable(M, Value->getType(), true, Linkage, Value,
155 getPGOFuncNameVarName(FuncName, Linkage));
157 // Hide the symbol so that we correctly get a copy for each executable.
158 if (!GlobalValue::isLocalLinkage(FuncNameVar->getLinkage()))
159 FuncNameVar->setVisibility(GlobalValue::HiddenVisibility);
164 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef FuncName) {
165 return createPGOFuncNameVar(*F.getParent(), F.getLinkage(), FuncName);
168 int collectPGOFuncNameStrings(const std::vector<std::string> &NameStrs,
169 bool doCompression, std::string &Result) {
170 uint8_t Header[16], *P = Header;
171 std::string UncompressedNameStrings =
172 join(NameStrs.begin(), NameStrs.end(), StringRef(" "));
174 unsigned EncLen = encodeULEB128(UncompressedNameStrings.length(), P);
177 auto WriteStringToResult = [&](size_t CompressedLen,
178 const std::string &InputStr) {
179 EncLen = encodeULEB128(CompressedLen, P);
181 char *HeaderStr = reinterpret_cast<char *>(&Header[0]);
182 unsigned HeaderLen = P - &Header[0];
183 Result.append(HeaderStr, HeaderLen);
189 return WriteStringToResult(0, UncompressedNameStrings);
191 SmallVector<char, 128> CompressedNameStrings;
192 zlib::Status Success =
193 zlib::compress(StringRef(UncompressedNameStrings), CompressedNameStrings,
194 zlib::BestSizeCompression);
196 if (Success != zlib::StatusOK)
199 return WriteStringToResult(
200 CompressedNameStrings.size(),
201 std::string(CompressedNameStrings.data(), CompressedNameStrings.size()));
204 StringRef getPGOFuncNameInitializer(GlobalVariable *NameVar) {
205 auto *Arr = cast<ConstantDataArray>(NameVar->getInitializer());
207 Arr->isCString() ? Arr->getAsCString() : Arr->getAsString();
211 int collectPGOFuncNameStrings(const std::vector<GlobalVariable *> &NameVars,
212 std::string &Result) {
213 std::vector<std::string> NameStrs;
214 for (auto *NameVar : NameVars) {
215 NameStrs.push_back(getPGOFuncNameInitializer(NameVar));
217 return collectPGOFuncNameStrings(NameStrs, zlib::isAvailable(), Result);
220 int readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab) {
221 const uint8_t *P = reinterpret_cast<const uint8_t *>(NameStrings.data());
222 const uint8_t *EndP = reinterpret_cast<const uint8_t *>(NameStrings.data() +
226 uint64_t UncompressedSize = decodeULEB128(P, &N);
228 uint64_t CompressedSize = decodeULEB128(P, &N);
230 bool isCompressed = (CompressedSize != 0);
231 SmallString<128> UncompressedNameStrings;
232 StringRef NameStrings;
234 StringRef CompressedNameStrings(reinterpret_cast<const char *>(P),
236 if (zlib::uncompress(CompressedNameStrings, UncompressedNameStrings,
237 UncompressedSize) != zlib::StatusOK)
240 NameStrings = StringRef(UncompressedNameStrings.data(),
241 UncompressedNameStrings.size());
244 StringRef(reinterpret_cast<const char *>(P), UncompressedSize);
245 P += UncompressedSize;
247 // Now parse the name strings.
248 size_t NameStart = 0;
251 size_t NameStop = NameStrings.find(' ', NameStart);
252 if (NameStop == StringRef::npos)
253 NameStop = NameStrings.size();
254 if (NameStop >= NameStrings.size() - 1)
256 StringRef Name = NameStrings.substr(NameStart, NameStop - NameStart);
257 Symtab.addFuncName(Name);
260 NameStart = NameStop + 1;
263 while (P < EndP && *P == 0)
266 Symtab.finalizeSymtab();
271 InstrProfValueSiteRecord::mergeValueData(InstrProfValueSiteRecord &Input,
273 this->sortByTargetValues();
274 Input.sortByTargetValues();
275 auto I = ValueData.begin();
276 auto IE = ValueData.end();
277 instrprof_error Result = instrprof_error::success;
278 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
280 while (I != IE && I->Value < J->Value)
282 if (I != IE && I->Value == J->Value) {
283 uint64_t JCount = J->Count;
286 JCount = SaturatingMultiply(JCount, Weight, &Overflowed);
288 Result = instrprof_error::counter_overflow;
290 I->Count = SaturatingAdd(I->Count, JCount, &Overflowed);
292 Result = instrprof_error::counter_overflow;
296 ValueData.insert(I, *J);
301 // Merge Value Profile data from Src record to this record for ValueKind.
302 // Scale merged value counts by \p Weight.
303 instrprof_error InstrProfRecord::mergeValueProfData(uint32_t ValueKind,
304 InstrProfRecord &Src,
306 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
307 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
308 if (ThisNumValueSites != OtherNumValueSites)
309 return instrprof_error::value_site_count_mismatch;
310 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
311 getValueSitesForKind(ValueKind);
312 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
313 Src.getValueSitesForKind(ValueKind);
314 instrprof_error Result = instrprof_error::success;
315 for (uint32_t I = 0; I < ThisNumValueSites; I++)
317 ThisSiteRecords[I].mergeValueData(OtherSiteRecords[I], Weight));
321 instrprof_error InstrProfRecord::merge(InstrProfRecord &Other,
323 // If the number of counters doesn't match we either have bad data
324 // or a hash collision.
325 if (Counts.size() != Other.Counts.size())
326 return instrprof_error::count_mismatch;
328 instrprof_error Result = instrprof_error::success;
330 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
332 uint64_t OtherCount = Other.Counts[I];
334 OtherCount = SaturatingMultiply(OtherCount, Weight, &Overflowed);
336 Result = instrprof_error::counter_overflow;
338 Counts[I] = SaturatingAdd(Counts[I], OtherCount, &Overflowed);
340 Result = instrprof_error::counter_overflow;
343 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
344 MergeResult(Result, mergeValueProfData(Kind, Other, Weight));
349 // Map indirect call target name hash to name string.
350 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind,
351 ValueMapType *ValueMap) {
355 case IPVK_IndirectCallTarget: {
357 std::lower_bound(ValueMap->begin(), ValueMap->end(), Value,
358 [](const std::pair<uint64_t, uint64_t> &LHS,
359 uint64_t RHS) { return LHS.first < RHS; });
360 if (Result != ValueMap->end())
361 Value = (uint64_t)Result->second;
368 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
369 InstrProfValueData *VData, uint32_t N,
370 ValueMapType *ValueMap) {
371 for (uint32_t I = 0; I < N; I++) {
372 VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap);
374 std::vector<InstrProfValueSiteRecord> &ValueSites =
375 getValueSitesForKind(ValueKind);
377 ValueSites.push_back(InstrProfValueSiteRecord());
379 ValueSites.emplace_back(VData, VData + N);
382 #define INSTR_PROF_COMMON_API_IMPL
383 #include "llvm/ProfileData/InstrProfData.inc"
386 * \brief ValueProfRecordClosure Interface implementation for InstrProfRecord
387 * class. These C wrappers are used as adaptors so that C++ code can be
388 * invoked as callbacks.
390 uint32_t getNumValueKindsInstrProf(const void *Record) {
391 return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds();
394 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) {
395 return reinterpret_cast<const InstrProfRecord *>(Record)
396 ->getNumValueSites(VKind);
399 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) {
400 return reinterpret_cast<const InstrProfRecord *>(Record)
401 ->getNumValueData(VKind);
404 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK,
406 return reinterpret_cast<const InstrProfRecord *>(R)
407 ->getNumValueDataForSite(VK, S);
410 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst,
411 uint32_t K, uint32_t S,
412 uint64_t (*Mapper)(uint32_t, uint64_t)) {
413 return reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(
417 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) {
419 (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData());
420 memset(VD, 0, TotalSizeInBytes);
424 static ValueProfRecordClosure InstrProfRecordClosure = {
426 getNumValueKindsInstrProf,
427 getNumValueSitesInstrProf,
428 getNumValueDataInstrProf,
429 getNumValueDataForSiteInstrProf,
431 getValueForSiteInstrProf,
432 allocValueProfDataInstrProf};
434 // Wrapper implementation using the closure mechanism.
435 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) {
436 InstrProfRecordClosure.Record = &Record;
437 return getValueProfDataSize(&InstrProfRecordClosure);
440 // Wrapper implementation using the closure mechanism.
441 std::unique_ptr<ValueProfData>
442 ValueProfData::serializeFrom(const InstrProfRecord &Record) {
443 InstrProfRecordClosure.Record = &Record;
445 std::unique_ptr<ValueProfData> VPD(
446 serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr));
450 void ValueProfRecord::deserializeTo(InstrProfRecord &Record,
451 InstrProfRecord::ValueMapType *VMap) {
452 Record.reserveSites(Kind, NumValueSites);
454 InstrProfValueData *ValueData = getValueProfRecordValueData(this);
455 for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) {
456 uint8_t ValueDataCount = this->SiteCountArray[VSite];
457 Record.addValueData(Kind, VSite, ValueData, ValueDataCount, VMap);
458 ValueData += ValueDataCount;
462 // For writing/serializing, Old is the host endianness, and New is
463 // byte order intended on disk. For Reading/deserialization, Old
464 // is the on-disk source endianness, and New is the host endianness.
465 void ValueProfRecord::swapBytes(support::endianness Old,
466 support::endianness New) {
467 using namespace support;
471 if (getHostEndianness() != Old) {
472 sys::swapByteOrder<uint32_t>(NumValueSites);
473 sys::swapByteOrder<uint32_t>(Kind);
475 uint32_t ND = getValueProfRecordNumValueData(this);
476 InstrProfValueData *VD = getValueProfRecordValueData(this);
478 // No need to swap byte array: SiteCountArrray.
479 for (uint32_t I = 0; I < ND; I++) {
480 sys::swapByteOrder<uint64_t>(VD[I].Value);
481 sys::swapByteOrder<uint64_t>(VD[I].Count);
483 if (getHostEndianness() == Old) {
484 sys::swapByteOrder<uint32_t>(NumValueSites);
485 sys::swapByteOrder<uint32_t>(Kind);
489 void ValueProfData::deserializeTo(InstrProfRecord &Record,
490 InstrProfRecord::ValueMapType *VMap) {
491 if (NumValueKinds == 0)
494 ValueProfRecord *VR = getFirstValueProfRecord(this);
495 for (uint32_t K = 0; K < NumValueKinds; K++) {
496 VR->deserializeTo(Record, VMap);
497 VR = getValueProfRecordNext(VR);
502 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) {
503 using namespace support;
505 return endian::readNext<T, little, unaligned>(D);
507 return endian::readNext<T, big, unaligned>(D);
510 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) {
511 return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize))
515 instrprof_error ValueProfData::checkIntegrity() {
516 if (NumValueKinds > IPVK_Last + 1)
517 return instrprof_error::malformed;
518 // Total size needs to be mulltiple of quadword size.
519 if (TotalSize % sizeof(uint64_t))
520 return instrprof_error::malformed;
522 ValueProfRecord *VR = getFirstValueProfRecord(this);
523 for (uint32_t K = 0; K < this->NumValueKinds; K++) {
524 if (VR->Kind > IPVK_Last)
525 return instrprof_error::malformed;
526 VR = getValueProfRecordNext(VR);
527 if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize)
528 return instrprof_error::malformed;
530 return instrprof_error::success;
533 ErrorOr<std::unique_ptr<ValueProfData>>
534 ValueProfData::getValueProfData(const unsigned char *D,
535 const unsigned char *const BufferEnd,
536 support::endianness Endianness) {
537 using namespace support;
538 if (D + sizeof(ValueProfData) > BufferEnd)
539 return instrprof_error::truncated;
541 const unsigned char *Header = D;
542 uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness);
543 if (D + TotalSize > BufferEnd)
544 return instrprof_error::too_large;
546 std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize);
547 memcpy(VPD.get(), D, TotalSize);
549 VPD->swapBytesToHost(Endianness);
551 instrprof_error EC = VPD->checkIntegrity();
552 if (EC != instrprof_error::success)
555 return std::move(VPD);
558 void ValueProfData::swapBytesToHost(support::endianness Endianness) {
559 using namespace support;
560 if (Endianness == getHostEndianness())
563 sys::swapByteOrder<uint32_t>(TotalSize);
564 sys::swapByteOrder<uint32_t>(NumValueKinds);
566 ValueProfRecord *VR = getFirstValueProfRecord(this);
567 for (uint32_t K = 0; K < NumValueKinds; K++) {
568 VR->swapBytes(Endianness, getHostEndianness());
569 VR = getValueProfRecordNext(VR);
573 void ValueProfData::swapBytesFromHost(support::endianness Endianness) {
574 using namespace support;
575 if (Endianness == getHostEndianness())
578 ValueProfRecord *VR = getFirstValueProfRecord(this);
579 for (uint32_t K = 0; K < NumValueKinds; K++) {
580 ValueProfRecord *NVR = getValueProfRecordNext(VR);
581 VR->swapBytes(getHostEndianness(), Endianness);
584 sys::swapByteOrder<uint32_t>(TotalSize);
585 sys::swapByteOrder<uint32_t>(NumValueKinds);