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 SmallVector<StringRef, 0> Names;
249 NameStrings.split(Names, ' ');
250 for (StringRef &Name : Names)
251 Symtab.addFuncName(Name);
253 while (P < EndP && *P == 0)
256 Symtab.finalizeSymtab();
261 InstrProfValueSiteRecord::mergeValueData(InstrProfValueSiteRecord &Input,
263 this->sortByTargetValues();
264 Input.sortByTargetValues();
265 auto I = ValueData.begin();
266 auto IE = ValueData.end();
267 instrprof_error Result = instrprof_error::success;
268 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
270 while (I != IE && I->Value < J->Value)
272 if (I != IE && I->Value == J->Value) {
273 uint64_t JCount = J->Count;
276 JCount = SaturatingMultiply(JCount, Weight, &Overflowed);
278 Result = instrprof_error::counter_overflow;
280 I->Count = SaturatingAdd(I->Count, JCount, &Overflowed);
282 Result = instrprof_error::counter_overflow;
286 ValueData.insert(I, *J);
291 // Merge Value Profile data from Src record to this record for ValueKind.
292 // Scale merged value counts by \p Weight.
293 instrprof_error InstrProfRecord::mergeValueProfData(uint32_t ValueKind,
294 InstrProfRecord &Src,
296 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
297 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
298 if (ThisNumValueSites != OtherNumValueSites)
299 return instrprof_error::value_site_count_mismatch;
300 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
301 getValueSitesForKind(ValueKind);
302 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
303 Src.getValueSitesForKind(ValueKind);
304 instrprof_error Result = instrprof_error::success;
305 for (uint32_t I = 0; I < ThisNumValueSites; I++)
307 ThisSiteRecords[I].mergeValueData(OtherSiteRecords[I], Weight));
311 instrprof_error InstrProfRecord::merge(InstrProfRecord &Other,
313 // If the number of counters doesn't match we either have bad data
314 // or a hash collision.
315 if (Counts.size() != Other.Counts.size())
316 return instrprof_error::count_mismatch;
318 instrprof_error Result = instrprof_error::success;
320 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
322 uint64_t OtherCount = Other.Counts[I];
324 OtherCount = SaturatingMultiply(OtherCount, Weight, &Overflowed);
326 Result = instrprof_error::counter_overflow;
328 Counts[I] = SaturatingAdd(Counts[I], OtherCount, &Overflowed);
330 Result = instrprof_error::counter_overflow;
333 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
334 MergeResult(Result, mergeValueProfData(Kind, Other, Weight));
339 // Map indirect call target name hash to name string.
340 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind,
341 ValueMapType *ValueMap) {
345 case IPVK_IndirectCallTarget: {
347 std::lower_bound(ValueMap->begin(), ValueMap->end(), Value,
348 [](const std::pair<uint64_t, uint64_t> &LHS,
349 uint64_t RHS) { return LHS.first < RHS; });
350 if (Result != ValueMap->end())
351 Value = (uint64_t)Result->second;
358 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
359 InstrProfValueData *VData, uint32_t N,
360 ValueMapType *ValueMap) {
361 for (uint32_t I = 0; I < N; I++) {
362 VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap);
364 std::vector<InstrProfValueSiteRecord> &ValueSites =
365 getValueSitesForKind(ValueKind);
367 ValueSites.push_back(InstrProfValueSiteRecord());
369 ValueSites.emplace_back(VData, VData + N);
372 #define INSTR_PROF_COMMON_API_IMPL
373 #include "llvm/ProfileData/InstrProfData.inc"
376 * \brief ValueProfRecordClosure Interface implementation for InstrProfRecord
377 * class. These C wrappers are used as adaptors so that C++ code can be
378 * invoked as callbacks.
380 uint32_t getNumValueKindsInstrProf(const void *Record) {
381 return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds();
384 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) {
385 return reinterpret_cast<const InstrProfRecord *>(Record)
386 ->getNumValueSites(VKind);
389 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) {
390 return reinterpret_cast<const InstrProfRecord *>(Record)
391 ->getNumValueData(VKind);
394 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK,
396 return reinterpret_cast<const InstrProfRecord *>(R)
397 ->getNumValueDataForSite(VK, S);
400 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst,
401 uint32_t K, uint32_t S,
402 uint64_t (*Mapper)(uint32_t, uint64_t)) {
403 return reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(
407 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) {
409 (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData());
410 memset(VD, 0, TotalSizeInBytes);
414 static ValueProfRecordClosure InstrProfRecordClosure = {
416 getNumValueKindsInstrProf,
417 getNumValueSitesInstrProf,
418 getNumValueDataInstrProf,
419 getNumValueDataForSiteInstrProf,
421 getValueForSiteInstrProf,
422 allocValueProfDataInstrProf};
424 // Wrapper implementation using the closure mechanism.
425 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) {
426 InstrProfRecordClosure.Record = &Record;
427 return getValueProfDataSize(&InstrProfRecordClosure);
430 // Wrapper implementation using the closure mechanism.
431 std::unique_ptr<ValueProfData>
432 ValueProfData::serializeFrom(const InstrProfRecord &Record) {
433 InstrProfRecordClosure.Record = &Record;
435 std::unique_ptr<ValueProfData> VPD(
436 serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr));
440 void ValueProfRecord::deserializeTo(InstrProfRecord &Record,
441 InstrProfRecord::ValueMapType *VMap) {
442 Record.reserveSites(Kind, NumValueSites);
444 InstrProfValueData *ValueData = getValueProfRecordValueData(this);
445 for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) {
446 uint8_t ValueDataCount = this->SiteCountArray[VSite];
447 Record.addValueData(Kind, VSite, ValueData, ValueDataCount, VMap);
448 ValueData += ValueDataCount;
452 // For writing/serializing, Old is the host endianness, and New is
453 // byte order intended on disk. For Reading/deserialization, Old
454 // is the on-disk source endianness, and New is the host endianness.
455 void ValueProfRecord::swapBytes(support::endianness Old,
456 support::endianness New) {
457 using namespace support;
461 if (getHostEndianness() != Old) {
462 sys::swapByteOrder<uint32_t>(NumValueSites);
463 sys::swapByteOrder<uint32_t>(Kind);
465 uint32_t ND = getValueProfRecordNumValueData(this);
466 InstrProfValueData *VD = getValueProfRecordValueData(this);
468 // No need to swap byte array: SiteCountArrray.
469 for (uint32_t I = 0; I < ND; I++) {
470 sys::swapByteOrder<uint64_t>(VD[I].Value);
471 sys::swapByteOrder<uint64_t>(VD[I].Count);
473 if (getHostEndianness() == Old) {
474 sys::swapByteOrder<uint32_t>(NumValueSites);
475 sys::swapByteOrder<uint32_t>(Kind);
479 void ValueProfData::deserializeTo(InstrProfRecord &Record,
480 InstrProfRecord::ValueMapType *VMap) {
481 if (NumValueKinds == 0)
484 ValueProfRecord *VR = getFirstValueProfRecord(this);
485 for (uint32_t K = 0; K < NumValueKinds; K++) {
486 VR->deserializeTo(Record, VMap);
487 VR = getValueProfRecordNext(VR);
492 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) {
493 using namespace support;
495 return endian::readNext<T, little, unaligned>(D);
497 return endian::readNext<T, big, unaligned>(D);
500 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) {
501 return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize))
505 instrprof_error ValueProfData::checkIntegrity() {
506 if (NumValueKinds > IPVK_Last + 1)
507 return instrprof_error::malformed;
508 // Total size needs to be mulltiple of quadword size.
509 if (TotalSize % sizeof(uint64_t))
510 return instrprof_error::malformed;
512 ValueProfRecord *VR = getFirstValueProfRecord(this);
513 for (uint32_t K = 0; K < this->NumValueKinds; K++) {
514 if (VR->Kind > IPVK_Last)
515 return instrprof_error::malformed;
516 VR = getValueProfRecordNext(VR);
517 if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize)
518 return instrprof_error::malformed;
520 return instrprof_error::success;
523 ErrorOr<std::unique_ptr<ValueProfData>>
524 ValueProfData::getValueProfData(const unsigned char *D,
525 const unsigned char *const BufferEnd,
526 support::endianness Endianness) {
527 using namespace support;
528 if (D + sizeof(ValueProfData) > BufferEnd)
529 return instrprof_error::truncated;
531 const unsigned char *Header = D;
532 uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness);
533 if (D + TotalSize > BufferEnd)
534 return instrprof_error::too_large;
536 std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize);
537 memcpy(VPD.get(), D, TotalSize);
539 VPD->swapBytesToHost(Endianness);
541 instrprof_error EC = VPD->checkIntegrity();
542 if (EC != instrprof_error::success)
545 return std::move(VPD);
548 void ValueProfData::swapBytesToHost(support::endianness Endianness) {
549 using namespace support;
550 if (Endianness == getHostEndianness())
553 sys::swapByteOrder<uint32_t>(TotalSize);
554 sys::swapByteOrder<uint32_t>(NumValueKinds);
556 ValueProfRecord *VR = getFirstValueProfRecord(this);
557 for (uint32_t K = 0; K < NumValueKinds; K++) {
558 VR->swapBytes(Endianness, getHostEndianness());
559 VR = getValueProfRecordNext(VR);
563 void ValueProfData::swapBytesFromHost(support::endianness Endianness) {
564 using namespace support;
565 if (Endianness == getHostEndianness())
568 ValueProfRecord *VR = getFirstValueProfRecord(this);
569 for (uint32_t K = 0; K < NumValueKinds; K++) {
570 ValueProfRecord *NVR = getValueProfRecordNext(VR);
571 VR->swapBytes(getHostEndianness(), Endianness);
574 sys::swapByteOrder<uint32_t>(TotalSize);
575 sys::swapByteOrder<uint32_t>(NumValueKinds);