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/ADT/ArrayRef.h"
16 #include "llvm/ADT/SmallString.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/StringExtras.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/ADT/Triple.h"
21 #include "llvm/IR/Constant.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/GlobalValue.h"
25 #include "llvm/IR/GlobalVariable.h"
26 #include "llvm/IR/Instruction.h"
27 #include "llvm/IR/LLVMContext.h"
28 #include "llvm/IR/MDBuilder.h"
29 #include "llvm/IR/Metadata.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/IR/Type.h"
32 #include "llvm/ProfileData/InstrProf.h"
33 #include "llvm/Support/Casting.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Compiler.h"
36 #include "llvm/Support/Compression.h"
37 #include "llvm/Support/Endian.h"
38 #include "llvm/Support/Error.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Support/LEB128.h"
41 #include "llvm/Support/ManagedStatic.h"
42 #include "llvm/Support/MathExtras.h"
43 #include "llvm/Support/Path.h"
44 #include "llvm/Support/SwapByteOrder.h"
52 #include <system_error>
58 static cl::opt<bool> StaticFuncFullModulePrefix(
59 "static-func-full-module-prefix", cl::init(true),
60 cl::desc("Use full module build paths in the profile counter names for "
61 "static functions."));
63 // This option is tailored to users that have different top-level directory in
64 // profile-gen and profile-use compilation. Users need to specific the number
65 // of levels to strip. A value larger than the number of directories in the
66 // source file will strip all the directory names and only leave the basename.
68 // Note current ThinLTO module importing for the indirect-calls assumes
69 // the source directory name not being stripped. A non-zero option value here
70 // can potentially prevent some inter-module indirect-call-promotions.
71 static cl::opt<unsigned> StaticFuncStripDirNamePrefix(
72 "static-func-strip-dirname-prefix", cl::init(0),
73 cl::desc("Strip specified level of directory name from source path in "
74 "the profile counter name for static functions."));
76 static std::string getInstrProfErrString(instrprof_error Err) {
78 case instrprof_error::success:
80 case instrprof_error::eof:
82 case instrprof_error::unrecognized_format:
83 return "Unrecognized instrumentation profile encoding format";
84 case instrprof_error::bad_magic:
85 return "Invalid instrumentation profile data (bad magic)";
86 case instrprof_error::bad_header:
87 return "Invalid instrumentation profile data (file header is corrupt)";
88 case instrprof_error::unsupported_version:
89 return "Unsupported instrumentation profile format version";
90 case instrprof_error::unsupported_hash_type:
91 return "Unsupported instrumentation profile hash type";
92 case instrprof_error::too_large:
93 return "Too much profile data";
94 case instrprof_error::truncated:
95 return "Truncated profile data";
96 case instrprof_error::malformed:
97 return "Malformed instrumentation profile data";
98 case instrprof_error::unknown_function:
99 return "No profile data available for function";
100 case instrprof_error::hash_mismatch:
101 return "Function control flow change detected (hash mismatch)";
102 case instrprof_error::count_mismatch:
103 return "Function basic block count change detected (counter mismatch)";
104 case instrprof_error::counter_overflow:
105 return "Counter overflow";
106 case instrprof_error::value_site_count_mismatch:
107 return "Function value site count change detected (counter mismatch)";
108 case instrprof_error::compress_failed:
109 return "Failed to compress data (zlib)";
110 case instrprof_error::uncompress_failed:
111 return "Failed to uncompress data (zlib)";
112 case instrprof_error::empty_raw_profile:
113 return "Empty raw profile file";
115 llvm_unreachable("A value of instrprof_error has no message.");
120 // FIXME: This class is only here to support the transition to llvm::Error. It
121 // will be removed once this transition is complete. Clients should prefer to
122 // deal with the Error value directly, rather than converting to error_code.
123 class InstrProfErrorCategoryType : public std::error_category {
124 const char *name() const noexcept override { return "llvm.instrprof"; }
126 std::string message(int IE) const override {
127 return getInstrProfErrString(static_cast<instrprof_error>(IE));
131 } // end anonymous namespace
133 static ManagedStatic<InstrProfErrorCategoryType> ErrorCategory;
135 const std::error_category &llvm::instrprof_category() {
136 return *ErrorCategory;
141 void SoftInstrProfErrors::addError(instrprof_error IE) {
142 if (IE == instrprof_error::success)
145 if (FirstError == instrprof_error::success)
149 case instrprof_error::hash_mismatch:
152 case instrprof_error::count_mismatch:
153 ++NumCountMismatches;
155 case instrprof_error::counter_overflow:
156 ++NumCounterOverflows;
158 case instrprof_error::value_site_count_mismatch:
159 ++NumValueSiteCountMismatches;
162 llvm_unreachable("Not a soft error");
166 std::string InstrProfError::message() const {
167 return getInstrProfErrString(Err);
170 char InstrProfError::ID = 0;
172 std::string getPGOFuncName(StringRef RawFuncName,
173 GlobalValue::LinkageTypes Linkage,
175 uint64_t Version LLVM_ATTRIBUTE_UNUSED) {
176 return GlobalValue::getGlobalIdentifier(RawFuncName, Linkage, FileName);
179 // Strip NumPrefix level of directory name from PathNameStr. If the number of
180 // directory separators is less than NumPrefix, strip all the directories and
181 // leave base file name only.
182 static StringRef stripDirPrefix(StringRef PathNameStr, uint32_t NumPrefix) {
183 uint32_t Count = NumPrefix;
184 uint32_t Pos = 0, LastPos = 0;
185 for (auto & CI : PathNameStr) {
187 if (llvm::sys::path::is_separator(CI)) {
194 return PathNameStr.substr(LastPos);
197 // Return the PGOFuncName. This function has some special handling when called
198 // in LTO optimization. The following only applies when calling in LTO passes
199 // (when \c InLTO is true): LTO's internalization privatizes many global linkage
200 // symbols. This happens after value profile annotation, but those internal
201 // linkage functions should not have a source prefix.
202 // Additionally, for ThinLTO mode, exported internal functions are promoted
203 // and renamed. We need to ensure that the original internal PGO name is
204 // used when computing the GUID that is compared against the profiled GUIDs.
205 // To differentiate compiler generated internal symbols from original ones,
206 // PGOFuncName meta data are created and attached to the original internal
207 // symbols in the value profile annotation step
208 // (PGOUseFunc::annotateIndirectCallSites). If a symbol does not have the meta
209 // data, its original linkage must be non-internal.
210 std::string getPGOFuncName(const Function &F, bool InLTO, uint64_t Version) {
212 StringRef FileName = (StaticFuncFullModulePrefix
213 ? F.getParent()->getName()
214 : sys::path::filename(F.getParent()->getName()));
215 if (StaticFuncFullModulePrefix && StaticFuncStripDirNamePrefix != 0)
216 FileName = stripDirPrefix(FileName, StaticFuncStripDirNamePrefix);
217 return getPGOFuncName(F.getName(), F.getLinkage(), FileName, Version);
220 // In LTO mode (when InLTO is true), first check if there is a meta data.
221 if (MDNode *MD = getPGOFuncNameMetadata(F)) {
222 StringRef S = cast<MDString>(MD->getOperand(0))->getString();
226 // If there is no meta data, the function must be a global before the value
227 // profile annotation pass. Its current linkage may be internal if it is
228 // internalized in LTO mode.
229 return getPGOFuncName(F.getName(), GlobalValue::ExternalLinkage, "");
232 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, StringRef FileName) {
233 if (FileName.empty())
235 // Drop the file name including ':'. See also getPGOFuncName.
236 if (PGOFuncName.startswith(FileName))
237 PGOFuncName = PGOFuncName.drop_front(FileName.size() + 1);
241 // \p FuncName is the string used as profile lookup key for the function. A
242 // symbol is created to hold the name. Return the legalized symbol name.
243 std::string getPGOFuncNameVarName(StringRef FuncName,
244 GlobalValue::LinkageTypes Linkage) {
245 std::string VarName = getInstrProfNameVarPrefix();
248 if (!GlobalValue::isLocalLinkage(Linkage))
251 // Now fix up illegal chars in local VarName that may upset the assembler.
252 const char *InvalidChars = "-:<>/\"'";
253 size_t found = VarName.find_first_of(InvalidChars);
254 while (found != std::string::npos) {
255 VarName[found] = '_';
256 found = VarName.find_first_of(InvalidChars, found + 1);
261 GlobalVariable *createPGOFuncNameVar(Module &M,
262 GlobalValue::LinkageTypes Linkage,
263 StringRef PGOFuncName) {
264 // We generally want to match the function's linkage, but available_externally
265 // and extern_weak both have the wrong semantics, and anything that doesn't
266 // need to link across compilation units doesn't need to be visible at all.
267 if (Linkage == GlobalValue::ExternalWeakLinkage)
268 Linkage = GlobalValue::LinkOnceAnyLinkage;
269 else if (Linkage == GlobalValue::AvailableExternallyLinkage)
270 Linkage = GlobalValue::LinkOnceODRLinkage;
271 else if (Linkage == GlobalValue::InternalLinkage ||
272 Linkage == GlobalValue::ExternalLinkage)
273 Linkage = GlobalValue::PrivateLinkage;
276 ConstantDataArray::getString(M.getContext(), PGOFuncName, false);
278 new GlobalVariable(M, Value->getType(), true, Linkage, Value,
279 getPGOFuncNameVarName(PGOFuncName, Linkage));
281 // Hide the symbol so that we correctly get a copy for each executable.
282 if (!GlobalValue::isLocalLinkage(FuncNameVar->getLinkage()))
283 FuncNameVar->setVisibility(GlobalValue::HiddenVisibility);
288 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName) {
289 return createPGOFuncNameVar(*F.getParent(), F.getLinkage(), PGOFuncName);
292 void InstrProfSymtab::create(Module &M, bool InLTO) {
293 for (Function &F : M) {
294 // Function may not have a name: like using asm("") to overwrite the name.
295 // Ignore in this case.
298 const std::string &PGOFuncName = getPGOFuncName(F, InLTO);
299 addFuncName(PGOFuncName);
300 MD5FuncMap.emplace_back(Function::getGUID(PGOFuncName), &F);
306 Error collectPGOFuncNameStrings(const std::vector<std::string> &NameStrs,
307 bool doCompression, std::string &Result) {
308 assert(!NameStrs.empty() && "No name data to emit");
310 uint8_t Header[16], *P = Header;
311 std::string UncompressedNameStrings =
312 join(NameStrs.begin(), NameStrs.end(), getInstrProfNameSeparator());
314 assert(StringRef(UncompressedNameStrings)
315 .count(getInstrProfNameSeparator()) == (NameStrs.size() - 1) &&
316 "PGO name is invalid (contains separator token)");
318 unsigned EncLen = encodeULEB128(UncompressedNameStrings.length(), P);
321 auto WriteStringToResult = [&](size_t CompressedLen, StringRef InputStr) {
322 EncLen = encodeULEB128(CompressedLen, P);
324 char *HeaderStr = reinterpret_cast<char *>(&Header[0]);
325 unsigned HeaderLen = P - &Header[0];
326 Result.append(HeaderStr, HeaderLen);
328 return Error::success();
331 if (!doCompression) {
332 return WriteStringToResult(0, UncompressedNameStrings);
335 SmallString<128> CompressedNameStrings;
336 Error E = zlib::compress(StringRef(UncompressedNameStrings),
337 CompressedNameStrings, zlib::BestSizeCompression);
339 consumeError(std::move(E));
340 return make_error<InstrProfError>(instrprof_error::compress_failed);
343 return WriteStringToResult(CompressedNameStrings.size(),
344 CompressedNameStrings);
347 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar) {
348 auto *Arr = cast<ConstantDataArray>(NameVar->getInitializer());
350 Arr->isCString() ? Arr->getAsCString() : Arr->getAsString();
354 Error collectPGOFuncNameStrings(const std::vector<GlobalVariable *> &NameVars,
355 std::string &Result, bool doCompression) {
356 std::vector<std::string> NameStrs;
357 for (auto *NameVar : NameVars) {
358 NameStrs.push_back(getPGOFuncNameVarInitializer(NameVar));
360 return collectPGOFuncNameStrings(
361 NameStrs, zlib::isAvailable() && doCompression, Result);
364 Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab) {
365 const uint8_t *P = reinterpret_cast<const uint8_t *>(NameStrings.data());
366 const uint8_t *EndP = reinterpret_cast<const uint8_t *>(NameStrings.data() +
370 uint64_t UncompressedSize = decodeULEB128(P, &N);
372 uint64_t CompressedSize = decodeULEB128(P, &N);
374 bool isCompressed = (CompressedSize != 0);
375 SmallString<128> UncompressedNameStrings;
376 StringRef NameStrings;
378 StringRef CompressedNameStrings(reinterpret_cast<const char *>(P),
381 zlib::uncompress(CompressedNameStrings, UncompressedNameStrings,
383 consumeError(std::move(E));
384 return make_error<InstrProfError>(instrprof_error::uncompress_failed);
387 NameStrings = StringRef(UncompressedNameStrings.data(),
388 UncompressedNameStrings.size());
391 StringRef(reinterpret_cast<const char *>(P), UncompressedSize);
392 P += UncompressedSize;
394 // Now parse the name strings.
395 SmallVector<StringRef, 0> Names;
396 NameStrings.split(Names, getInstrProfNameSeparator());
397 for (StringRef &Name : Names)
398 Symtab.addFuncName(Name);
400 while (P < EndP && *P == 0)
403 Symtab.finalizeSymtab();
404 return Error::success();
407 void InstrProfValueSiteRecord::merge(SoftInstrProfErrors &SIPE,
408 InstrProfValueSiteRecord &Input,
410 this->sortByTargetValues();
411 Input.sortByTargetValues();
412 auto I = ValueData.begin();
413 auto IE = ValueData.end();
414 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
416 while (I != IE && I->Value < J->Value)
418 if (I != IE && I->Value == J->Value) {
420 I->Count = SaturatingMultiplyAdd(J->Count, Weight, I->Count, &Overflowed);
422 SIPE.addError(instrprof_error::counter_overflow);
426 ValueData.insert(I, *J);
430 void InstrProfValueSiteRecord::scale(SoftInstrProfErrors &SIPE,
432 for (auto I = ValueData.begin(), IE = ValueData.end(); I != IE; ++I) {
434 I->Count = SaturatingMultiply(I->Count, Weight, &Overflowed);
436 SIPE.addError(instrprof_error::counter_overflow);
440 // Merge Value Profile data from Src record to this record for ValueKind.
441 // Scale merged value counts by \p Weight.
442 void InstrProfRecord::mergeValueProfData(uint32_t ValueKind,
443 InstrProfRecord &Src,
445 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
446 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
447 if (ThisNumValueSites != OtherNumValueSites) {
448 SIPE.addError(instrprof_error::value_site_count_mismatch);
451 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
452 getValueSitesForKind(ValueKind);
453 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
454 Src.getValueSitesForKind(ValueKind);
455 for (uint32_t I = 0; I < ThisNumValueSites; I++)
456 ThisSiteRecords[I].merge(SIPE, OtherSiteRecords[I], Weight);
459 void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight) {
460 // If the number of counters doesn't match we either have bad data
461 // or a hash collision.
462 if (Counts.size() != Other.Counts.size()) {
463 SIPE.addError(instrprof_error::count_mismatch);
467 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
470 SaturatingMultiplyAdd(Other.Counts[I], Weight, Counts[I], &Overflowed);
472 SIPE.addError(instrprof_error::counter_overflow);
475 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
476 mergeValueProfData(Kind, Other, Weight);
479 void InstrProfRecord::scaleValueProfData(uint32_t ValueKind, uint64_t Weight) {
480 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
481 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
482 getValueSitesForKind(ValueKind);
483 for (uint32_t I = 0; I < ThisNumValueSites; I++)
484 ThisSiteRecords[I].scale(SIPE, Weight);
487 void InstrProfRecord::scale(uint64_t Weight) {
488 for (auto &Count : this->Counts) {
490 Count = SaturatingMultiply(Count, Weight, &Overflowed);
492 SIPE.addError(instrprof_error::counter_overflow);
494 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
495 scaleValueProfData(Kind, Weight);
498 // Map indirect call target name hash to name string.
499 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind,
500 ValueMapType *ValueMap) {
504 case IPVK_IndirectCallTarget: {
506 std::lower_bound(ValueMap->begin(), ValueMap->end(), Value,
507 [](const std::pair<uint64_t, uint64_t> &LHS,
508 uint64_t RHS) { return LHS.first < RHS; });
509 // Raw function pointer collected by value profiler may be from
510 // external functions that are not instrumented. They won't have
511 // mapping data to be used by the deserializer. Force the value to
512 // be 0 in this case.
513 if (Result != ValueMap->end() && Result->first == Value)
514 Value = (uint64_t)Result->second;
523 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
524 InstrProfValueData *VData, uint32_t N,
525 ValueMapType *ValueMap) {
526 for (uint32_t I = 0; I < N; I++) {
527 VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap);
529 std::vector<InstrProfValueSiteRecord> &ValueSites =
530 getValueSitesForKind(ValueKind);
532 ValueSites.emplace_back();
534 ValueSites.emplace_back(VData, VData + N);
537 #define INSTR_PROF_COMMON_API_IMPL
538 #include "llvm/ProfileData/InstrProfData.inc"
541 * \brief ValueProfRecordClosure Interface implementation for InstrProfRecord
542 * class. These C wrappers are used as adaptors so that C++ code can be
543 * invoked as callbacks.
545 uint32_t getNumValueKindsInstrProf(const void *Record) {
546 return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds();
549 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) {
550 return reinterpret_cast<const InstrProfRecord *>(Record)
551 ->getNumValueSites(VKind);
554 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) {
555 return reinterpret_cast<const InstrProfRecord *>(Record)
556 ->getNumValueData(VKind);
559 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK,
561 return reinterpret_cast<const InstrProfRecord *>(R)
562 ->getNumValueDataForSite(VK, S);
565 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst,
566 uint32_t K, uint32_t S) {
567 reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(Dst, K, S);
570 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) {
572 (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData());
573 memset(VD, 0, TotalSizeInBytes);
577 static ValueProfRecordClosure InstrProfRecordClosure = {
579 getNumValueKindsInstrProf,
580 getNumValueSitesInstrProf,
581 getNumValueDataInstrProf,
582 getNumValueDataForSiteInstrProf,
584 getValueForSiteInstrProf,
585 allocValueProfDataInstrProf};
587 // Wrapper implementation using the closure mechanism.
588 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) {
589 InstrProfRecordClosure.Record = &Record;
590 return getValueProfDataSize(&InstrProfRecordClosure);
593 // Wrapper implementation using the closure mechanism.
594 std::unique_ptr<ValueProfData>
595 ValueProfData::serializeFrom(const InstrProfRecord &Record) {
596 InstrProfRecordClosure.Record = &Record;
598 std::unique_ptr<ValueProfData> VPD(
599 serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr));
603 void ValueProfRecord::deserializeTo(InstrProfRecord &Record,
604 InstrProfRecord::ValueMapType *VMap) {
605 Record.reserveSites(Kind, NumValueSites);
607 InstrProfValueData *ValueData = getValueProfRecordValueData(this);
608 for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) {
609 uint8_t ValueDataCount = this->SiteCountArray[VSite];
610 Record.addValueData(Kind, VSite, ValueData, ValueDataCount, VMap);
611 ValueData += ValueDataCount;
615 // For writing/serializing, Old is the host endianness, and New is
616 // byte order intended on disk. For Reading/deserialization, Old
617 // is the on-disk source endianness, and New is the host endianness.
618 void ValueProfRecord::swapBytes(support::endianness Old,
619 support::endianness New) {
620 using namespace support;
625 if (getHostEndianness() != Old) {
626 sys::swapByteOrder<uint32_t>(NumValueSites);
627 sys::swapByteOrder<uint32_t>(Kind);
629 uint32_t ND = getValueProfRecordNumValueData(this);
630 InstrProfValueData *VD = getValueProfRecordValueData(this);
632 // No need to swap byte array: SiteCountArrray.
633 for (uint32_t I = 0; I < ND; I++) {
634 sys::swapByteOrder<uint64_t>(VD[I].Value);
635 sys::swapByteOrder<uint64_t>(VD[I].Count);
637 if (getHostEndianness() == Old) {
638 sys::swapByteOrder<uint32_t>(NumValueSites);
639 sys::swapByteOrder<uint32_t>(Kind);
643 void ValueProfData::deserializeTo(InstrProfRecord &Record,
644 InstrProfRecord::ValueMapType *VMap) {
645 if (NumValueKinds == 0)
648 ValueProfRecord *VR = getFirstValueProfRecord(this);
649 for (uint32_t K = 0; K < NumValueKinds; K++) {
650 VR->deserializeTo(Record, VMap);
651 VR = getValueProfRecordNext(VR);
656 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) {
657 using namespace support;
660 return endian::readNext<T, little, unaligned>(D);
662 return endian::readNext<T, big, unaligned>(D);
665 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) {
666 return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize))
670 Error ValueProfData::checkIntegrity() {
671 if (NumValueKinds > IPVK_Last + 1)
672 return make_error<InstrProfError>(instrprof_error::malformed);
673 // Total size needs to be mulltiple of quadword size.
674 if (TotalSize % sizeof(uint64_t))
675 return make_error<InstrProfError>(instrprof_error::malformed);
677 ValueProfRecord *VR = getFirstValueProfRecord(this);
678 for (uint32_t K = 0; K < this->NumValueKinds; K++) {
679 if (VR->Kind > IPVK_Last)
680 return make_error<InstrProfError>(instrprof_error::malformed);
681 VR = getValueProfRecordNext(VR);
682 if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize)
683 return make_error<InstrProfError>(instrprof_error::malformed);
685 return Error::success();
688 Expected<std::unique_ptr<ValueProfData>>
689 ValueProfData::getValueProfData(const unsigned char *D,
690 const unsigned char *const BufferEnd,
691 support::endianness Endianness) {
692 using namespace support;
694 if (D + sizeof(ValueProfData) > BufferEnd)
695 return make_error<InstrProfError>(instrprof_error::truncated);
697 const unsigned char *Header = D;
698 uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness);
699 if (D + TotalSize > BufferEnd)
700 return make_error<InstrProfError>(instrprof_error::too_large);
702 std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize);
703 memcpy(VPD.get(), D, TotalSize);
705 VPD->swapBytesToHost(Endianness);
707 Error E = VPD->checkIntegrity();
711 return std::move(VPD);
714 void ValueProfData::swapBytesToHost(support::endianness Endianness) {
715 using namespace support;
717 if (Endianness == getHostEndianness())
720 sys::swapByteOrder<uint32_t>(TotalSize);
721 sys::swapByteOrder<uint32_t>(NumValueKinds);
723 ValueProfRecord *VR = getFirstValueProfRecord(this);
724 for (uint32_t K = 0; K < NumValueKinds; K++) {
725 VR->swapBytes(Endianness, getHostEndianness());
726 VR = getValueProfRecordNext(VR);
730 void ValueProfData::swapBytesFromHost(support::endianness Endianness) {
731 using namespace support;
733 if (Endianness == getHostEndianness())
736 ValueProfRecord *VR = getFirstValueProfRecord(this);
737 for (uint32_t K = 0; K < NumValueKinds; K++) {
738 ValueProfRecord *NVR = getValueProfRecordNext(VR);
739 VR->swapBytes(getHostEndianness(), Endianness);
742 sys::swapByteOrder<uint32_t>(TotalSize);
743 sys::swapByteOrder<uint32_t>(NumValueKinds);
746 void annotateValueSite(Module &M, Instruction &Inst,
747 const InstrProfRecord &InstrProfR,
748 InstrProfValueKind ValueKind, uint32_t SiteIdx,
749 uint32_t MaxMDCount) {
750 uint32_t NV = InstrProfR.getNumValueDataForSite(ValueKind, SiteIdx);
755 std::unique_ptr<InstrProfValueData[]> VD =
756 InstrProfR.getValueForSite(ValueKind, SiteIdx, &Sum);
758 ArrayRef<InstrProfValueData> VDs(VD.get(), NV);
759 annotateValueSite(M, Inst, VDs, Sum, ValueKind, MaxMDCount);
762 void annotateValueSite(Module &M, Instruction &Inst,
763 ArrayRef<InstrProfValueData> VDs,
764 uint64_t Sum, InstrProfValueKind ValueKind,
765 uint32_t MaxMDCount) {
766 LLVMContext &Ctx = M.getContext();
767 MDBuilder MDHelper(Ctx);
768 SmallVector<Metadata *, 3> Vals;
770 Vals.push_back(MDHelper.createString("VP"));
772 Vals.push_back(MDHelper.createConstant(
773 ConstantInt::get(Type::getInt32Ty(Ctx), ValueKind)));
776 MDHelper.createConstant(ConstantInt::get(Type::getInt64Ty(Ctx), Sum)));
778 // Value Profile Data
779 uint32_t MDCount = MaxMDCount;
780 for (auto &VD : VDs) {
781 Vals.push_back(MDHelper.createConstant(
782 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Value)));
783 Vals.push_back(MDHelper.createConstant(
784 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Count)));
788 Inst.setMetadata(LLVMContext::MD_prof, MDNode::get(Ctx, Vals));
791 bool getValueProfDataFromInst(const Instruction &Inst,
792 InstrProfValueKind ValueKind,
793 uint32_t MaxNumValueData,
794 InstrProfValueData ValueData[],
795 uint32_t &ActualNumValueData, uint64_t &TotalC) {
796 MDNode *MD = Inst.getMetadata(LLVMContext::MD_prof);
800 unsigned NOps = MD->getNumOperands();
805 // Operand 0 is a string tag "VP":
806 MDString *Tag = cast<MDString>(MD->getOperand(0));
810 if (!Tag->getString().equals("VP"))
814 ConstantInt *KindInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
817 if (KindInt->getZExtValue() != ValueKind)
821 ConstantInt *TotalCInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
824 TotalC = TotalCInt->getZExtValue();
826 ActualNumValueData = 0;
828 for (unsigned I = 3; I < NOps; I += 2) {
829 if (ActualNumValueData >= MaxNumValueData)
831 ConstantInt *Value = mdconst::dyn_extract<ConstantInt>(MD->getOperand(I));
833 mdconst::dyn_extract<ConstantInt>(MD->getOperand(I + 1));
834 if (!Value || !Count)
836 ValueData[ActualNumValueData].Value = Value->getZExtValue();
837 ValueData[ActualNumValueData].Count = Count->getZExtValue();
838 ActualNumValueData++;
843 MDNode *getPGOFuncNameMetadata(const Function &F) {
844 return F.getMetadata(getPGOFuncNameMetadataName());
847 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName) {
848 // Only for internal linkage functions.
849 if (PGOFuncName == F.getName())
851 // Don't create duplicated meta-data.
852 if (getPGOFuncNameMetadata(F))
854 LLVMContext &C = F.getContext();
855 MDNode *N = MDNode::get(C, MDString::get(C, PGOFuncName));
856 F.setMetadata(getPGOFuncNameMetadataName(), N);
859 bool needsComdatForCounter(const Function &F, const Module &M) {
863 Triple TT(M.getTargetTriple());
864 if (!TT.isOSBinFormatELF() && !TT.isOSBinFormatWasm())
867 // See createPGOFuncNameVar for more details. To avoid link errors, profile
868 // counters for function with available_externally linkage needs to be changed
869 // to linkonce linkage. On ELF based systems, this leads to weak symbols to be
870 // created. Without using comdat, duplicate entries won't be removed by the
871 // linker leading to increased data segement size and raw profile size. Even
872 // worse, since the referenced counter from profile per-function data object
873 // will be resolved to the common strong definition, the profile counts for
874 // available_externally functions will end up being duplicated in raw profile
875 // data. This can result in distorted profile as the counts of those dups
876 // will be accumulated by the profile merger.
877 GlobalValue::LinkageTypes Linkage = F.getLinkage();
878 if (Linkage != GlobalValue::ExternalWeakLinkage &&
879 Linkage != GlobalValue::AvailableExternallyLinkage)
885 // Check if INSTR_PROF_RAW_VERSION_VAR is defined.
886 bool isIRPGOFlagSet(const Module *M) {
888 M->getNamedGlobal(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
889 if (!IRInstrVar || IRInstrVar->isDeclaration() ||
890 IRInstrVar->hasLocalLinkage())
893 // Check if the flag is set.
894 if (!IRInstrVar->hasInitializer())
897 const Constant *InitVal = IRInstrVar->getInitializer();
901 return (dyn_cast<ConstantInt>(InitVal)->getZExtValue() &
902 VARIANT_MASK_IR_PROF) != 0;
905 // Check if we can safely rename this Comdat function.
906 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken) {
907 if (F.getName().empty())
909 if (!needsComdatForCounter(F, *(F.getParent())))
911 // Unsafe to rename the address-taken function (which can be used in
912 // function comparison).
913 if (CheckAddressTaken && F.hasAddressTaken())
915 // Only safe to do if this function may be discarded if it is not used
916 // in the compilation unit.
917 if (!GlobalValue::isDiscardableIfUnused(F.getLinkage()))
920 // For AvailableExternallyLinkage functions.
921 if (!F.hasComdat()) {
922 assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage);
928 } // end namespace llvm