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/ADT/Triple.h"
18 #include "llvm/IR/Constants.h"
19 #include "llvm/IR/Function.h"
20 #include "llvm/IR/GlobalVariable.h"
21 #include "llvm/IR/MDBuilder.h"
22 #include "llvm/IR/Module.h"
23 #include "llvm/Support/Compression.h"
24 #include "llvm/Support/ErrorHandling.h"
25 #include "llvm/Support/LEB128.h"
26 #include "llvm/Support/ManagedStatic.h"
27 #include "llvm/Support/Path.h"
31 static cl::opt<bool> StaticFuncFullModulePrefix(
32 "static-func-full-module-prefix", cl::init(false),
33 cl::desc("Use full module build paths in the profile counter names for "
34 "static functions."));
37 std::string getInstrProfErrString(instrprof_error Err) {
39 case instrprof_error::success:
41 case instrprof_error::eof:
43 case instrprof_error::unrecognized_format:
44 return "Unrecognized instrumentation profile encoding format";
45 case instrprof_error::bad_magic:
46 return "Invalid instrumentation profile data (bad magic)";
47 case instrprof_error::bad_header:
48 return "Invalid instrumentation profile data (file header is corrupt)";
49 case instrprof_error::unsupported_version:
50 return "Unsupported instrumentation profile format version";
51 case instrprof_error::unsupported_hash_type:
52 return "Unsupported instrumentation profile hash type";
53 case instrprof_error::too_large:
54 return "Too much profile data";
55 case instrprof_error::truncated:
56 return "Truncated profile data";
57 case instrprof_error::malformed:
58 return "Malformed instrumentation profile data";
59 case instrprof_error::unknown_function:
60 return "No profile data available for function";
61 case instrprof_error::hash_mismatch:
62 return "Function control flow change detected (hash mismatch)";
63 case instrprof_error::count_mismatch:
64 return "Function basic block count change detected (counter mismatch)";
65 case instrprof_error::counter_overflow:
66 return "Counter overflow";
67 case instrprof_error::value_site_count_mismatch:
68 return "Function value site count change detected (counter mismatch)";
69 case instrprof_error::compress_failed:
70 return "Failed to compress data (zlib)";
71 case instrprof_error::uncompress_failed:
72 return "Failed to uncompress data (zlib)";
73 case instrprof_error::empty_raw_profile:
74 return "Empty raw profile file";
76 llvm_unreachable("A value of instrprof_error has no message.");
79 // FIXME: This class is only here to support the transition to llvm::Error. It
80 // will be removed once this transition is complete. Clients should prefer to
81 // deal with the Error value directly, rather than converting to error_code.
82 class InstrProfErrorCategoryType : public std::error_category {
83 const char *name() const noexcept override { return "llvm.instrprof"; }
84 std::string message(int IE) const override {
85 return getInstrProfErrString(static_cast<instrprof_error>(IE));
88 } // end anonymous namespace
90 static ManagedStatic<InstrProfErrorCategoryType> ErrorCategory;
92 const std::error_category &llvm::instrprof_category() {
93 return *ErrorCategory;
98 void SoftInstrProfErrors::addError(instrprof_error IE) {
99 if (IE == instrprof_error::success)
102 if (FirstError == instrprof_error::success)
106 case instrprof_error::hash_mismatch:
109 case instrprof_error::count_mismatch:
110 ++NumCountMismatches;
112 case instrprof_error::counter_overflow:
113 ++NumCounterOverflows;
115 case instrprof_error::value_site_count_mismatch:
116 ++NumValueSiteCountMismatches;
119 llvm_unreachable("Not a soft error");
123 std::string InstrProfError::message() const {
124 return getInstrProfErrString(Err);
127 char InstrProfError::ID = 0;
129 std::string getPGOFuncName(StringRef RawFuncName,
130 GlobalValue::LinkageTypes Linkage,
132 uint64_t Version LLVM_ATTRIBUTE_UNUSED) {
133 return GlobalValue::getGlobalIdentifier(RawFuncName, Linkage, FileName);
136 // Return the PGOFuncName. This function has some special handling when called
137 // in LTO optimization. The following only applies when calling in LTO passes
138 // (when \c InLTO is true): LTO's internalization privatizes many global linkage
139 // symbols. This happens after value profile annotation, but those internal
140 // linkage functions should not have a source prefix.
141 // Additionally, for ThinLTO mode, exported internal functions are promoted
142 // and renamed. We need to ensure that the original internal PGO name is
143 // used when computing the GUID that is compared against the profiled GUIDs.
144 // To differentiate compiler generated internal symbols from original ones,
145 // PGOFuncName meta data are created and attached to the original internal
146 // symbols in the value profile annotation step
147 // (PGOUseFunc::annotateIndirectCallSites). If a symbol does not have the meta
148 // data, its original linkage must be non-internal.
149 std::string getPGOFuncName(const Function &F, bool InLTO, uint64_t Version) {
151 StringRef FileName = (StaticFuncFullModulePrefix
152 ? F.getParent()->getName()
153 : sys::path::filename(F.getParent()->getName()));
154 return getPGOFuncName(F.getName(), F.getLinkage(), FileName, Version);
157 // In LTO mode (when InLTO is true), first check if there is a meta data.
158 if (MDNode *MD = getPGOFuncNameMetadata(F)) {
159 StringRef S = cast<MDString>(MD->getOperand(0))->getString();
163 // If there is no meta data, the function must be a global before the value
164 // profile annotation pass. Its current linkage may be internal if it is
165 // internalized in LTO mode.
166 return getPGOFuncName(F.getName(), GlobalValue::ExternalLinkage, "");
169 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, StringRef FileName) {
170 if (FileName.empty())
172 // Drop the file name including ':'. See also getPGOFuncName.
173 if (PGOFuncName.startswith(FileName))
174 PGOFuncName = PGOFuncName.drop_front(FileName.size() + 1);
178 // \p FuncName is the string used as profile lookup key for the function. A
179 // symbol is created to hold the name. Return the legalized symbol name.
180 std::string getPGOFuncNameVarName(StringRef FuncName,
181 GlobalValue::LinkageTypes Linkage) {
182 std::string VarName = getInstrProfNameVarPrefix();
185 if (!GlobalValue::isLocalLinkage(Linkage))
188 // Now fix up illegal chars in local VarName that may upset the assembler.
189 const char *InvalidChars = "-:<>/\"'";
190 size_t found = VarName.find_first_of(InvalidChars);
191 while (found != std::string::npos) {
192 VarName[found] = '_';
193 found = VarName.find_first_of(InvalidChars, found + 1);
198 GlobalVariable *createPGOFuncNameVar(Module &M,
199 GlobalValue::LinkageTypes Linkage,
200 StringRef PGOFuncName) {
202 // We generally want to match the function's linkage, but available_externally
203 // and extern_weak both have the wrong semantics, and anything that doesn't
204 // need to link across compilation units doesn't need to be visible at all.
205 if (Linkage == GlobalValue::ExternalWeakLinkage)
206 Linkage = GlobalValue::LinkOnceAnyLinkage;
207 else if (Linkage == GlobalValue::AvailableExternallyLinkage)
208 Linkage = GlobalValue::LinkOnceODRLinkage;
209 else if (Linkage == GlobalValue::InternalLinkage ||
210 Linkage == GlobalValue::ExternalLinkage)
211 Linkage = GlobalValue::PrivateLinkage;
214 ConstantDataArray::getString(M.getContext(), PGOFuncName, false);
216 new GlobalVariable(M, Value->getType(), true, Linkage, Value,
217 getPGOFuncNameVarName(PGOFuncName, Linkage));
219 // Hide the symbol so that we correctly get a copy for each executable.
220 if (!GlobalValue::isLocalLinkage(FuncNameVar->getLinkage()))
221 FuncNameVar->setVisibility(GlobalValue::HiddenVisibility);
226 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName) {
227 return createPGOFuncNameVar(*F.getParent(), F.getLinkage(), PGOFuncName);
230 void InstrProfSymtab::create(Module &M, bool InLTO) {
231 for (Function &F : M) {
232 // Function may not have a name: like using asm("") to overwrite the name.
233 // Ignore in this case.
236 const std::string &PGOFuncName = getPGOFuncName(F, InLTO);
237 addFuncName(PGOFuncName);
238 MD5FuncMap.emplace_back(Function::getGUID(PGOFuncName), &F);
244 Error collectPGOFuncNameStrings(const std::vector<std::string> &NameStrs,
245 bool doCompression, std::string &Result) {
246 assert(NameStrs.size() && "No name data to emit");
248 uint8_t Header[16], *P = Header;
249 std::string UncompressedNameStrings =
250 join(NameStrs.begin(), NameStrs.end(), getInstrProfNameSeparator());
252 assert(StringRef(UncompressedNameStrings)
253 .count(getInstrProfNameSeparator()) == (NameStrs.size() - 1) &&
254 "PGO name is invalid (contains separator token)");
256 unsigned EncLen = encodeULEB128(UncompressedNameStrings.length(), P);
259 auto WriteStringToResult = [&](size_t CompressedLen, StringRef InputStr) {
260 EncLen = encodeULEB128(CompressedLen, P);
262 char *HeaderStr = reinterpret_cast<char *>(&Header[0]);
263 unsigned HeaderLen = P - &Header[0];
264 Result.append(HeaderStr, HeaderLen);
266 return Error::success();
269 if (!doCompression) {
270 return WriteStringToResult(0, UncompressedNameStrings);
273 SmallString<128> CompressedNameStrings;
274 zlib::Status Success =
275 zlib::compress(StringRef(UncompressedNameStrings), CompressedNameStrings,
276 zlib::BestSizeCompression);
278 if (Success != zlib::StatusOK)
279 return make_error<InstrProfError>(instrprof_error::compress_failed);
281 return WriteStringToResult(CompressedNameStrings.size(),
282 CompressedNameStrings);
285 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar) {
286 auto *Arr = cast<ConstantDataArray>(NameVar->getInitializer());
288 Arr->isCString() ? Arr->getAsCString() : Arr->getAsString();
292 Error collectPGOFuncNameStrings(const std::vector<GlobalVariable *> &NameVars,
293 std::string &Result, bool doCompression) {
294 std::vector<std::string> NameStrs;
295 for (auto *NameVar : NameVars) {
296 NameStrs.push_back(getPGOFuncNameVarInitializer(NameVar));
298 return collectPGOFuncNameStrings(
299 NameStrs, zlib::isAvailable() && doCompression, Result);
302 Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab) {
303 const uint8_t *P = reinterpret_cast<const uint8_t *>(NameStrings.data());
304 const uint8_t *EndP = reinterpret_cast<const uint8_t *>(NameStrings.data() +
308 uint64_t UncompressedSize = decodeULEB128(P, &N);
310 uint64_t CompressedSize = decodeULEB128(P, &N);
312 bool isCompressed = (CompressedSize != 0);
313 SmallString<128> UncompressedNameStrings;
314 StringRef NameStrings;
316 StringRef CompressedNameStrings(reinterpret_cast<const char *>(P),
318 if (zlib::uncompress(CompressedNameStrings, UncompressedNameStrings,
319 UncompressedSize) != zlib::StatusOK)
320 return make_error<InstrProfError>(instrprof_error::uncompress_failed);
322 NameStrings = StringRef(UncompressedNameStrings.data(),
323 UncompressedNameStrings.size());
326 StringRef(reinterpret_cast<const char *>(P), UncompressedSize);
327 P += UncompressedSize;
329 // Now parse the name strings.
330 SmallVector<StringRef, 0> Names;
331 NameStrings.split(Names, getInstrProfNameSeparator());
332 for (StringRef &Name : Names)
333 Symtab.addFuncName(Name);
335 while (P < EndP && *P == 0)
338 Symtab.finalizeSymtab();
339 return Error::success();
342 void InstrProfValueSiteRecord::merge(SoftInstrProfErrors &SIPE,
343 InstrProfValueSiteRecord &Input,
345 this->sortByTargetValues();
346 Input.sortByTargetValues();
347 auto I = ValueData.begin();
348 auto IE = ValueData.end();
349 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
351 while (I != IE && I->Value < J->Value)
353 if (I != IE && I->Value == J->Value) {
355 I->Count = SaturatingMultiplyAdd(J->Count, Weight, I->Count, &Overflowed);
357 SIPE.addError(instrprof_error::counter_overflow);
361 ValueData.insert(I, *J);
365 void InstrProfValueSiteRecord::scale(SoftInstrProfErrors &SIPE,
367 for (auto I = ValueData.begin(), IE = ValueData.end(); I != IE; ++I) {
369 I->Count = SaturatingMultiply(I->Count, Weight, &Overflowed);
371 SIPE.addError(instrprof_error::counter_overflow);
375 // Merge Value Profile data from Src record to this record for ValueKind.
376 // Scale merged value counts by \p Weight.
377 void InstrProfRecord::mergeValueProfData(uint32_t ValueKind,
378 InstrProfRecord &Src,
380 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
381 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
382 if (ThisNumValueSites != OtherNumValueSites) {
383 SIPE.addError(instrprof_error::value_site_count_mismatch);
386 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
387 getValueSitesForKind(ValueKind);
388 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
389 Src.getValueSitesForKind(ValueKind);
390 for (uint32_t I = 0; I < ThisNumValueSites; I++)
391 ThisSiteRecords[I].merge(SIPE, OtherSiteRecords[I], Weight);
394 void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight) {
395 // If the number of counters doesn't match we either have bad data
396 // or a hash collision.
397 if (Counts.size() != Other.Counts.size()) {
398 SIPE.addError(instrprof_error::count_mismatch);
402 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
405 SaturatingMultiplyAdd(Other.Counts[I], Weight, Counts[I], &Overflowed);
407 SIPE.addError(instrprof_error::counter_overflow);
410 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
411 mergeValueProfData(Kind, Other, Weight);
414 void InstrProfRecord::scaleValueProfData(uint32_t ValueKind, uint64_t Weight) {
415 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
416 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
417 getValueSitesForKind(ValueKind);
418 for (uint32_t I = 0; I < ThisNumValueSites; I++)
419 ThisSiteRecords[I].scale(SIPE, Weight);
422 void InstrProfRecord::scale(uint64_t Weight) {
423 for (auto &Count : this->Counts) {
425 Count = SaturatingMultiply(Count, Weight, &Overflowed);
427 SIPE.addError(instrprof_error::counter_overflow);
429 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
430 scaleValueProfData(Kind, Weight);
433 // Map indirect call target name hash to name string.
434 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind,
435 ValueMapType *ValueMap) {
439 case IPVK_IndirectCallTarget: {
441 std::lower_bound(ValueMap->begin(), ValueMap->end(), Value,
442 [](const std::pair<uint64_t, uint64_t> &LHS,
443 uint64_t RHS) { return LHS.first < RHS; });
444 // Raw function pointer collected by value profiler may be from
445 // external functions that are not instrumented. They won't have
446 // mapping data to be used by the deserializer. Force the value to
447 // be 0 in this case.
448 if (Result != ValueMap->end() && Result->first == Value)
449 Value = (uint64_t)Result->second;
458 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
459 InstrProfValueData *VData, uint32_t N,
460 ValueMapType *ValueMap) {
461 for (uint32_t I = 0; I < N; I++) {
462 VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap);
464 std::vector<InstrProfValueSiteRecord> &ValueSites =
465 getValueSitesForKind(ValueKind);
467 ValueSites.emplace_back();
469 ValueSites.emplace_back(VData, VData + N);
472 #define INSTR_PROF_COMMON_API_IMPL
473 #include "llvm/ProfileData/InstrProfData.inc"
476 * \brief ValueProfRecordClosure Interface implementation for InstrProfRecord
477 * class. These C wrappers are used as adaptors so that C++ code can be
478 * invoked as callbacks.
480 uint32_t getNumValueKindsInstrProf(const void *Record) {
481 return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds();
484 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) {
485 return reinterpret_cast<const InstrProfRecord *>(Record)
486 ->getNumValueSites(VKind);
489 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) {
490 return reinterpret_cast<const InstrProfRecord *>(Record)
491 ->getNumValueData(VKind);
494 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK,
496 return reinterpret_cast<const InstrProfRecord *>(R)
497 ->getNumValueDataForSite(VK, S);
500 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst,
501 uint32_t K, uint32_t S) {
502 reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(Dst, K, S);
505 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) {
507 (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData());
508 memset(VD, 0, TotalSizeInBytes);
512 static ValueProfRecordClosure InstrProfRecordClosure = {
514 getNumValueKindsInstrProf,
515 getNumValueSitesInstrProf,
516 getNumValueDataInstrProf,
517 getNumValueDataForSiteInstrProf,
519 getValueForSiteInstrProf,
520 allocValueProfDataInstrProf};
522 // Wrapper implementation using the closure mechanism.
523 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) {
524 InstrProfRecordClosure.Record = &Record;
525 return getValueProfDataSize(&InstrProfRecordClosure);
528 // Wrapper implementation using the closure mechanism.
529 std::unique_ptr<ValueProfData>
530 ValueProfData::serializeFrom(const InstrProfRecord &Record) {
531 InstrProfRecordClosure.Record = &Record;
533 std::unique_ptr<ValueProfData> VPD(
534 serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr));
538 void ValueProfRecord::deserializeTo(InstrProfRecord &Record,
539 InstrProfRecord::ValueMapType *VMap) {
540 Record.reserveSites(Kind, NumValueSites);
542 InstrProfValueData *ValueData = getValueProfRecordValueData(this);
543 for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) {
544 uint8_t ValueDataCount = this->SiteCountArray[VSite];
545 Record.addValueData(Kind, VSite, ValueData, ValueDataCount, VMap);
546 ValueData += ValueDataCount;
550 // For writing/serializing, Old is the host endianness, and New is
551 // byte order intended on disk. For Reading/deserialization, Old
552 // is the on-disk source endianness, and New is the host endianness.
553 void ValueProfRecord::swapBytes(support::endianness Old,
554 support::endianness New) {
555 using namespace support;
559 if (getHostEndianness() != Old) {
560 sys::swapByteOrder<uint32_t>(NumValueSites);
561 sys::swapByteOrder<uint32_t>(Kind);
563 uint32_t ND = getValueProfRecordNumValueData(this);
564 InstrProfValueData *VD = getValueProfRecordValueData(this);
566 // No need to swap byte array: SiteCountArrray.
567 for (uint32_t I = 0; I < ND; I++) {
568 sys::swapByteOrder<uint64_t>(VD[I].Value);
569 sys::swapByteOrder<uint64_t>(VD[I].Count);
571 if (getHostEndianness() == Old) {
572 sys::swapByteOrder<uint32_t>(NumValueSites);
573 sys::swapByteOrder<uint32_t>(Kind);
577 void ValueProfData::deserializeTo(InstrProfRecord &Record,
578 InstrProfRecord::ValueMapType *VMap) {
579 if (NumValueKinds == 0)
582 ValueProfRecord *VR = getFirstValueProfRecord(this);
583 for (uint32_t K = 0; K < NumValueKinds; K++) {
584 VR->deserializeTo(Record, VMap);
585 VR = getValueProfRecordNext(VR);
590 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) {
591 using namespace support;
593 return endian::readNext<T, little, unaligned>(D);
595 return endian::readNext<T, big, unaligned>(D);
598 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) {
599 return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize))
603 Error ValueProfData::checkIntegrity() {
604 if (NumValueKinds > IPVK_Last + 1)
605 return make_error<InstrProfError>(instrprof_error::malformed);
606 // Total size needs to be mulltiple of quadword size.
607 if (TotalSize % sizeof(uint64_t))
608 return make_error<InstrProfError>(instrprof_error::malformed);
610 ValueProfRecord *VR = getFirstValueProfRecord(this);
611 for (uint32_t K = 0; K < this->NumValueKinds; K++) {
612 if (VR->Kind > IPVK_Last)
613 return make_error<InstrProfError>(instrprof_error::malformed);
614 VR = getValueProfRecordNext(VR);
615 if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize)
616 return make_error<InstrProfError>(instrprof_error::malformed);
618 return Error::success();
621 Expected<std::unique_ptr<ValueProfData>>
622 ValueProfData::getValueProfData(const unsigned char *D,
623 const unsigned char *const BufferEnd,
624 support::endianness Endianness) {
625 using namespace support;
626 if (D + sizeof(ValueProfData) > BufferEnd)
627 return make_error<InstrProfError>(instrprof_error::truncated);
629 const unsigned char *Header = D;
630 uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness);
631 if (D + TotalSize > BufferEnd)
632 return make_error<InstrProfError>(instrprof_error::too_large);
634 std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize);
635 memcpy(VPD.get(), D, TotalSize);
637 VPD->swapBytesToHost(Endianness);
639 Error E = VPD->checkIntegrity();
643 return std::move(VPD);
646 void ValueProfData::swapBytesToHost(support::endianness Endianness) {
647 using namespace support;
648 if (Endianness == getHostEndianness())
651 sys::swapByteOrder<uint32_t>(TotalSize);
652 sys::swapByteOrder<uint32_t>(NumValueKinds);
654 ValueProfRecord *VR = getFirstValueProfRecord(this);
655 for (uint32_t K = 0; K < NumValueKinds; K++) {
656 VR->swapBytes(Endianness, getHostEndianness());
657 VR = getValueProfRecordNext(VR);
661 void ValueProfData::swapBytesFromHost(support::endianness Endianness) {
662 using namespace support;
663 if (Endianness == getHostEndianness())
666 ValueProfRecord *VR = getFirstValueProfRecord(this);
667 for (uint32_t K = 0; K < NumValueKinds; K++) {
668 ValueProfRecord *NVR = getValueProfRecordNext(VR);
669 VR->swapBytes(getHostEndianness(), Endianness);
672 sys::swapByteOrder<uint32_t>(TotalSize);
673 sys::swapByteOrder<uint32_t>(NumValueKinds);
676 void annotateValueSite(Module &M, Instruction &Inst,
677 const InstrProfRecord &InstrProfR,
678 InstrProfValueKind ValueKind, uint32_t SiteIdx,
679 uint32_t MaxMDCount) {
680 uint32_t NV = InstrProfR.getNumValueDataForSite(ValueKind, SiteIdx);
685 std::unique_ptr<InstrProfValueData[]> VD =
686 InstrProfR.getValueForSite(ValueKind, SiteIdx, &Sum);
688 ArrayRef<InstrProfValueData> VDs(VD.get(), NV);
689 annotateValueSite(M, Inst, VDs, Sum, ValueKind, MaxMDCount);
692 void annotateValueSite(Module &M, Instruction &Inst,
693 ArrayRef<InstrProfValueData> VDs,
694 uint64_t Sum, InstrProfValueKind ValueKind,
695 uint32_t MaxMDCount) {
696 LLVMContext &Ctx = M.getContext();
697 MDBuilder MDHelper(Ctx);
698 SmallVector<Metadata *, 3> Vals;
700 Vals.push_back(MDHelper.createString("VP"));
702 Vals.push_back(MDHelper.createConstant(
703 ConstantInt::get(Type::getInt32Ty(Ctx), ValueKind)));
706 MDHelper.createConstant(ConstantInt::get(Type::getInt64Ty(Ctx), Sum)));
708 // Value Profile Data
709 uint32_t MDCount = MaxMDCount;
710 for (auto &VD : VDs) {
711 Vals.push_back(MDHelper.createConstant(
712 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Value)));
713 Vals.push_back(MDHelper.createConstant(
714 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Count)));
718 Inst.setMetadata(LLVMContext::MD_prof, MDNode::get(Ctx, Vals));
721 bool getValueProfDataFromInst(const Instruction &Inst,
722 InstrProfValueKind ValueKind,
723 uint32_t MaxNumValueData,
724 InstrProfValueData ValueData[],
725 uint32_t &ActualNumValueData, uint64_t &TotalC) {
726 MDNode *MD = Inst.getMetadata(LLVMContext::MD_prof);
730 unsigned NOps = MD->getNumOperands();
735 // Operand 0 is a string tag "VP":
736 MDString *Tag = cast<MDString>(MD->getOperand(0));
740 if (!Tag->getString().equals("VP"))
744 ConstantInt *KindInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
747 if (KindInt->getZExtValue() != ValueKind)
751 ConstantInt *TotalCInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
754 TotalC = TotalCInt->getZExtValue();
756 ActualNumValueData = 0;
758 for (unsigned I = 3; I < NOps; I += 2) {
759 if (ActualNumValueData >= MaxNumValueData)
761 ConstantInt *Value = mdconst::dyn_extract<ConstantInt>(MD->getOperand(I));
763 mdconst::dyn_extract<ConstantInt>(MD->getOperand(I + 1));
764 if (!Value || !Count)
766 ValueData[ActualNumValueData].Value = Value->getZExtValue();
767 ValueData[ActualNumValueData].Count = Count->getZExtValue();
768 ActualNumValueData++;
773 MDNode *getPGOFuncNameMetadata(const Function &F) {
774 return F.getMetadata(getPGOFuncNameMetadataName());
777 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName) {
778 // Only for internal linkage functions.
779 if (PGOFuncName == F.getName())
781 // Don't create duplicated meta-data.
782 if (getPGOFuncNameMetadata(F))
784 LLVMContext &C = F.getContext();
785 MDNode *N = MDNode::get(C, MDString::get(C, PGOFuncName));
786 F.setMetadata(getPGOFuncNameMetadataName(), N);
789 bool needsComdatForCounter(const Function &F, const Module &M) {
793 Triple TT(M.getTargetTriple());
794 if (!TT.isOSBinFormatELF())
797 // See createPGOFuncNameVar for more details. To avoid link errors, profile
798 // counters for function with available_externally linkage needs to be changed
799 // to linkonce linkage. On ELF based systems, this leads to weak symbols to be
800 // created. Without using comdat, duplicate entries won't be removed by the
801 // linker leading to increased data segement size and raw profile size. Even
802 // worse, since the referenced counter from profile per-function data object
803 // will be resolved to the common strong definition, the profile counts for
804 // available_externally functions will end up being duplicated in raw profile
805 // data. This can result in distorted profile as the counts of those dups
806 // will be accumulated by the profile merger.
807 GlobalValue::LinkageTypes Linkage = F.getLinkage();
808 if (Linkage != GlobalValue::ExternalWeakLinkage &&
809 Linkage != GlobalValue::AvailableExternallyLinkage)
815 // Check if INSTR_PROF_RAW_VERSION_VAR is defined.
816 bool isIRPGOFlagSet(const Module *M) {
818 M->getNamedGlobal(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
819 if (!IRInstrVar || IRInstrVar->isDeclaration() ||
820 IRInstrVar->hasLocalLinkage())
823 // Check if the flag is set.
824 if (!IRInstrVar->hasInitializer())
827 const Constant *InitVal = IRInstrVar->getInitializer();
831 return (dyn_cast<ConstantInt>(InitVal)->getZExtValue() &
832 VARIANT_MASK_IR_PROF) != 0;
835 // Check if we can safely rename this Comdat function.
836 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken) {
837 if (F.getName().empty())
839 if (!needsComdatForCounter(F, *(F.getParent())))
841 // Unsafe to rename the address-taken function (which can be used in
842 // function comparison).
843 if (CheckAddressTaken && F.hasAddressTaken())
845 // Only safe to do if this function may be discarded if it is not used
846 // in the compilation unit.
847 if (!GlobalValue::isDiscardableIfUnused(F.getLinkage()))
850 // For AvailableExternallyLinkage functions.
851 if (!F.hasComdat()) {
852 assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage);
857 } // end namespace llvm