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);
176 if (!doCompression) {
177 EncLen = encodeULEB128(0, P);
179 Result.append(reinterpret_cast<char *>(&Header[0]), P - &Header[0]);
180 Result += UncompressedNameStrings;
183 SmallVector<char, 128> CompressedNameStrings;
184 zlib::Status Success =
185 zlib::compress(StringRef(UncompressedNameStrings), CompressedNameStrings,
186 zlib::BestSizeCompression);
187 assert(Success == zlib::StatusOK);
188 if (Success != zlib::StatusOK)
190 EncLen = encodeULEB128(CompressedNameStrings.size(), P);
192 Result.append(reinterpret_cast<char *>(&Header[0]), P - &Header[0]);
194 std::string(CompressedNameStrings.data(), CompressedNameStrings.size());
198 StringRef getPGOFuncNameInitializer(GlobalVariable *NameVar) {
199 auto *Arr = cast<ConstantDataArray>(NameVar->getInitializer());
201 Arr->isCString() ? Arr->getAsCString() : Arr->getAsString();
205 int collectPGOFuncNameStrings(const std::vector<GlobalVariable *> &NameVars,
206 std::string &Result) {
207 std::vector<std::string> NameStrs;
208 for (auto *NameVar : NameVars) {
209 NameStrs.push_back(getPGOFuncNameInitializer(NameVar));
211 return collectPGOFuncNameStrings(NameStrs, zlib::isAvailable(), Result);
214 int readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab) {
215 const uint8_t *P = reinterpret_cast<const uint8_t *>(NameStrings.data());
216 const uint8_t *EndP = reinterpret_cast<const uint8_t *>(NameStrings.data() +
220 uint64_t UncompressedSize = decodeULEB128(P, &N);
222 uint64_t CompressedSize = decodeULEB128(P, &N);
224 bool isCompressed = (CompressedSize != 0);
225 SmallString<128> UncompressedNameStrings;
226 StringRef NameStrings;
228 StringRef CompressedNameStrings(reinterpret_cast<const char *>(P),
230 if (zlib::uncompress(CompressedNameStrings, UncompressedNameStrings,
231 UncompressedSize) != zlib::StatusOK)
234 NameStrings = StringRef(UncompressedNameStrings.data(),
235 UncompressedNameStrings.size());
238 StringRef(reinterpret_cast<const char *>(P), UncompressedSize);
239 P += UncompressedSize;
241 // Now parse the name strings.
242 size_t NameStart = 0;
245 size_t NameStop = NameStrings.find(' ', NameStart);
246 if (NameStop == StringRef::npos)
247 NameStop = NameStrings.size();
248 if (NameStop >= NameStrings.size() - 1)
250 StringRef Name = NameStrings.substr(NameStart, NameStop - NameStart);
251 Symtab.addFuncName(Name);
254 NameStart = NameStop + 1;
257 while (P < EndP && *P == 0)
260 Symtab.finalizeSymtab();
265 InstrProfValueSiteRecord::mergeValueData(InstrProfValueSiteRecord &Input,
267 this->sortByTargetValues();
268 Input.sortByTargetValues();
269 auto I = ValueData.begin();
270 auto IE = ValueData.end();
271 instrprof_error Result = instrprof_error::success;
272 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
274 while (I != IE && I->Value < J->Value)
276 if (I != IE && I->Value == J->Value) {
277 uint64_t JCount = J->Count;
280 JCount = SaturatingMultiply(JCount, Weight, &Overflowed);
282 Result = instrprof_error::counter_overflow;
284 I->Count = SaturatingAdd(I->Count, JCount, &Overflowed);
286 Result = instrprof_error::counter_overflow;
290 ValueData.insert(I, *J);
295 // Merge Value Profile data from Src record to this record for ValueKind.
296 // Scale merged value counts by \p Weight.
297 instrprof_error InstrProfRecord::mergeValueProfData(uint32_t ValueKind,
298 InstrProfRecord &Src,
300 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
301 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
302 if (ThisNumValueSites != OtherNumValueSites)
303 return instrprof_error::value_site_count_mismatch;
304 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
305 getValueSitesForKind(ValueKind);
306 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
307 Src.getValueSitesForKind(ValueKind);
308 instrprof_error Result = instrprof_error::success;
309 for (uint32_t I = 0; I < ThisNumValueSites; I++)
311 ThisSiteRecords[I].mergeValueData(OtherSiteRecords[I], Weight));
315 instrprof_error InstrProfRecord::merge(InstrProfRecord &Other,
317 // If the number of counters doesn't match we either have bad data
318 // or a hash collision.
319 if (Counts.size() != Other.Counts.size())
320 return instrprof_error::count_mismatch;
322 instrprof_error Result = instrprof_error::success;
324 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
326 uint64_t OtherCount = Other.Counts[I];
328 OtherCount = SaturatingMultiply(OtherCount, Weight, &Overflowed);
330 Result = instrprof_error::counter_overflow;
332 Counts[I] = SaturatingAdd(Counts[I], OtherCount, &Overflowed);
334 Result = instrprof_error::counter_overflow;
337 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
338 MergeResult(Result, mergeValueProfData(Kind, Other, Weight));
343 // Map indirect call target name hash to name string.
344 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind,
345 ValueMapType *ValueMap) {
349 case IPVK_IndirectCallTarget: {
351 std::lower_bound(ValueMap->begin(), ValueMap->end(), Value,
352 [](const std::pair<uint64_t, uint64_t> &LHS,
353 uint64_t RHS) { return LHS.first < RHS; });
354 if (Result != ValueMap->end())
355 Value = (uint64_t)Result->second;
362 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
363 InstrProfValueData *VData, uint32_t N,
364 ValueMapType *ValueMap) {
365 for (uint32_t I = 0; I < N; I++) {
366 VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap);
368 std::vector<InstrProfValueSiteRecord> &ValueSites =
369 getValueSitesForKind(ValueKind);
371 ValueSites.push_back(InstrProfValueSiteRecord());
373 ValueSites.emplace_back(VData, VData + N);
376 #define INSTR_PROF_COMMON_API_IMPL
377 #include "llvm/ProfileData/InstrProfData.inc"
380 * \brief ValueProfRecordClosure Interface implementation for InstrProfRecord
381 * class. These C wrappers are used as adaptors so that C++ code can be
382 * invoked as callbacks.
384 uint32_t getNumValueKindsInstrProf(const void *Record) {
385 return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds();
388 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) {
389 return reinterpret_cast<const InstrProfRecord *>(Record)
390 ->getNumValueSites(VKind);
393 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) {
394 return reinterpret_cast<const InstrProfRecord *>(Record)
395 ->getNumValueData(VKind);
398 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK,
400 return reinterpret_cast<const InstrProfRecord *>(R)
401 ->getNumValueDataForSite(VK, S);
404 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst,
405 uint32_t K, uint32_t S,
406 uint64_t (*Mapper)(uint32_t, uint64_t)) {
407 return reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(
411 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) {
413 (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData());
414 memset(VD, 0, TotalSizeInBytes);
418 static ValueProfRecordClosure InstrProfRecordClosure = {
420 getNumValueKindsInstrProf,
421 getNumValueSitesInstrProf,
422 getNumValueDataInstrProf,
423 getNumValueDataForSiteInstrProf,
425 getValueForSiteInstrProf,
426 allocValueProfDataInstrProf};
428 // Wrapper implementation using the closure mechanism.
429 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) {
430 InstrProfRecordClosure.Record = &Record;
431 return getValueProfDataSize(&InstrProfRecordClosure);
434 // Wrapper implementation using the closure mechanism.
435 std::unique_ptr<ValueProfData>
436 ValueProfData::serializeFrom(const InstrProfRecord &Record) {
437 InstrProfRecordClosure.Record = &Record;
439 std::unique_ptr<ValueProfData> VPD(
440 serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr));
444 void ValueProfRecord::deserializeTo(InstrProfRecord &Record,
445 InstrProfRecord::ValueMapType *VMap) {
446 Record.reserveSites(Kind, NumValueSites);
448 InstrProfValueData *ValueData = getValueProfRecordValueData(this);
449 for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) {
450 uint8_t ValueDataCount = this->SiteCountArray[VSite];
451 Record.addValueData(Kind, VSite, ValueData, ValueDataCount, VMap);
452 ValueData += ValueDataCount;
456 // For writing/serializing, Old is the host endianness, and New is
457 // byte order intended on disk. For Reading/deserialization, Old
458 // is the on-disk source endianness, and New is the host endianness.
459 void ValueProfRecord::swapBytes(support::endianness Old,
460 support::endianness New) {
461 using namespace support;
465 if (getHostEndianness() != Old) {
466 sys::swapByteOrder<uint32_t>(NumValueSites);
467 sys::swapByteOrder<uint32_t>(Kind);
469 uint32_t ND = getValueProfRecordNumValueData(this);
470 InstrProfValueData *VD = getValueProfRecordValueData(this);
472 // No need to swap byte array: SiteCountArrray.
473 for (uint32_t I = 0; I < ND; I++) {
474 sys::swapByteOrder<uint64_t>(VD[I].Value);
475 sys::swapByteOrder<uint64_t>(VD[I].Count);
477 if (getHostEndianness() == Old) {
478 sys::swapByteOrder<uint32_t>(NumValueSites);
479 sys::swapByteOrder<uint32_t>(Kind);
483 void ValueProfData::deserializeTo(InstrProfRecord &Record,
484 InstrProfRecord::ValueMapType *VMap) {
485 if (NumValueKinds == 0)
488 ValueProfRecord *VR = getFirstValueProfRecord(this);
489 for (uint32_t K = 0; K < NumValueKinds; K++) {
490 VR->deserializeTo(Record, VMap);
491 VR = getValueProfRecordNext(VR);
496 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) {
497 using namespace support;
499 return endian::readNext<T, little, unaligned>(D);
501 return endian::readNext<T, big, unaligned>(D);
504 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) {
505 return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize))
509 instrprof_error ValueProfData::checkIntegrity() {
510 if (NumValueKinds > IPVK_Last + 1)
511 return instrprof_error::malformed;
512 // Total size needs to be mulltiple of quadword size.
513 if (TotalSize % sizeof(uint64_t))
514 return instrprof_error::malformed;
516 ValueProfRecord *VR = getFirstValueProfRecord(this);
517 for (uint32_t K = 0; K < this->NumValueKinds; K++) {
518 if (VR->Kind > IPVK_Last)
519 return instrprof_error::malformed;
520 VR = getValueProfRecordNext(VR);
521 if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize)
522 return instrprof_error::malformed;
524 return instrprof_error::success;
527 ErrorOr<std::unique_ptr<ValueProfData>>
528 ValueProfData::getValueProfData(const unsigned char *D,
529 const unsigned char *const BufferEnd,
530 support::endianness Endianness) {
531 using namespace support;
532 if (D + sizeof(ValueProfData) > BufferEnd)
533 return instrprof_error::truncated;
535 const unsigned char *Header = D;
536 uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness);
537 if (D + TotalSize > BufferEnd)
538 return instrprof_error::too_large;
540 std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize);
541 memcpy(VPD.get(), D, TotalSize);
543 VPD->swapBytesToHost(Endianness);
545 instrprof_error EC = VPD->checkIntegrity();
546 if (EC != instrprof_error::success)
549 return std::move(VPD);
552 void ValueProfData::swapBytesToHost(support::endianness Endianness) {
553 using namespace support;
554 if (Endianness == getHostEndianness())
557 sys::swapByteOrder<uint32_t>(TotalSize);
558 sys::swapByteOrder<uint32_t>(NumValueKinds);
560 ValueProfRecord *VR = getFirstValueProfRecord(this);
561 for (uint32_t K = 0; K < NumValueKinds; K++) {
562 VR->swapBytes(Endianness, getHostEndianness());
563 VR = getValueProfRecordNext(VR);
567 void ValueProfData::swapBytesFromHost(support::endianness Endianness) {
568 using namespace support;
569 if (Endianness == getHostEndianness())
572 ValueProfRecord *VR = getFirstValueProfRecord(this);
573 for (uint32_t K = 0; K < NumValueKinds; K++) {
574 ValueProfRecord *NVR = getValueProfRecordNext(VR);
575 VR->swapBytes(getHostEndianness(), Endianness);
578 sys::swapByteOrder<uint32_t>(TotalSize);
579 sys::swapByteOrder<uint32_t>(NumValueKinds);