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/ArrayRef.h"
17 #include "llvm/ADT/SmallString.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/ADT/StringRef.h"
21 #include "llvm/ADT/Triple.h"
22 #include "llvm/IR/Constant.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/GlobalValue.h"
26 #include "llvm/IR/GlobalVariable.h"
27 #include "llvm/IR/Instruction.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/MDBuilder.h"
30 #include "llvm/IR/Metadata.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/Type.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), cl::Hidden,
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), cl::Hidden,
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";
114 case instrprof_error::zlib_unavailable:
115 return "Profile uses zlib compression but the profile reader was built without zlib support";
117 llvm_unreachable("A value of instrprof_error has no message.");
122 // FIXME: This class is only here to support the transition to llvm::Error. It
123 // will be removed once this transition is complete. Clients should prefer to
124 // deal with the Error value directly, rather than converting to error_code.
125 class InstrProfErrorCategoryType : public std::error_category {
126 const char *name() const noexcept override { return "llvm.instrprof"; }
128 std::string message(int IE) const override {
129 return getInstrProfErrString(static_cast<instrprof_error>(IE));
133 } // end anonymous namespace
135 static ManagedStatic<InstrProfErrorCategoryType> ErrorCategory;
137 const std::error_category &llvm::instrprof_category() {
138 return *ErrorCategory;
143 const char *InstrProfSectNameCommon[] = {
144 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) \
146 #include "llvm/ProfileData/InstrProfData.inc"
149 const char *InstrProfSectNameCoff[] = {
150 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) \
152 #include "llvm/ProfileData/InstrProfData.inc"
155 const char *InstrProfSectNamePrefix[] = {
156 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) \
158 #include "llvm/ProfileData/InstrProfData.inc"
165 std::string getInstrProfSectionName(InstrProfSectKind IPSK,
166 Triple::ObjectFormatType OF,
167 bool AddSegmentInfo) {
168 std::string SectName;
170 if (OF == Triple::MachO && AddSegmentInfo)
171 SectName = InstrProfSectNamePrefix[IPSK];
173 if (OF == Triple::COFF)
174 SectName += InstrProfSectNameCoff[IPSK];
176 SectName += InstrProfSectNameCommon[IPSK];
178 if (OF == Triple::MachO && IPSK == IPSK_data && AddSegmentInfo)
179 SectName += ",regular,live_support";
184 void SoftInstrProfErrors::addError(instrprof_error IE) {
185 if (IE == instrprof_error::success)
188 if (FirstError == instrprof_error::success)
192 case instrprof_error::hash_mismatch:
195 case instrprof_error::count_mismatch:
196 ++NumCountMismatches;
198 case instrprof_error::counter_overflow:
199 ++NumCounterOverflows;
201 case instrprof_error::value_site_count_mismatch:
202 ++NumValueSiteCountMismatches;
205 llvm_unreachable("Not a soft error");
209 std::string InstrProfError::message() const {
210 return getInstrProfErrString(Err);
213 char InstrProfError::ID = 0;
215 std::string getPGOFuncName(StringRef RawFuncName,
216 GlobalValue::LinkageTypes Linkage,
218 uint64_t Version LLVM_ATTRIBUTE_UNUSED) {
219 return GlobalValue::getGlobalIdentifier(RawFuncName, Linkage, FileName);
222 // Strip NumPrefix level of directory name from PathNameStr. If the number of
223 // directory separators is less than NumPrefix, strip all the directories and
224 // leave base file name only.
225 static StringRef stripDirPrefix(StringRef PathNameStr, uint32_t NumPrefix) {
226 uint32_t Count = NumPrefix;
227 uint32_t Pos = 0, LastPos = 0;
228 for (auto & CI : PathNameStr) {
230 if (llvm::sys::path::is_separator(CI)) {
237 return PathNameStr.substr(LastPos);
240 // Return the PGOFuncName. This function has some special handling when called
241 // in LTO optimization. The following only applies when calling in LTO passes
242 // (when \c InLTO is true): LTO's internalization privatizes many global linkage
243 // symbols. This happens after value profile annotation, but those internal
244 // linkage functions should not have a source prefix.
245 // Additionally, for ThinLTO mode, exported internal functions are promoted
246 // and renamed. We need to ensure that the original internal PGO name is
247 // used when computing the GUID that is compared against the profiled GUIDs.
248 // To differentiate compiler generated internal symbols from original ones,
249 // PGOFuncName meta data are created and attached to the original internal
250 // symbols in the value profile annotation step
251 // (PGOUseFunc::annotateIndirectCallSites). If a symbol does not have the meta
252 // data, its original linkage must be non-internal.
253 std::string getPGOFuncName(const Function &F, bool InLTO, uint64_t Version) {
255 StringRef FileName = (StaticFuncFullModulePrefix
256 ? F.getParent()->getName()
257 : sys::path::filename(F.getParent()->getName()));
258 if (StaticFuncFullModulePrefix && StaticFuncStripDirNamePrefix != 0)
259 FileName = stripDirPrefix(FileName, StaticFuncStripDirNamePrefix);
260 return getPGOFuncName(F.getName(), F.getLinkage(), FileName, Version);
263 // In LTO mode (when InLTO is true), first check if there is a meta data.
264 if (MDNode *MD = getPGOFuncNameMetadata(F)) {
265 StringRef S = cast<MDString>(MD->getOperand(0))->getString();
269 // If there is no meta data, the function must be a global before the value
270 // profile annotation pass. Its current linkage may be internal if it is
271 // internalized in LTO mode.
272 return getPGOFuncName(F.getName(), GlobalValue::ExternalLinkage, "");
275 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, StringRef FileName) {
276 if (FileName.empty())
278 // Drop the file name including ':'. See also getPGOFuncName.
279 if (PGOFuncName.startswith(FileName))
280 PGOFuncName = PGOFuncName.drop_front(FileName.size() + 1);
284 // \p FuncName is the string used as profile lookup key for the function. A
285 // symbol is created to hold the name. Return the legalized symbol name.
286 std::string getPGOFuncNameVarName(StringRef FuncName,
287 GlobalValue::LinkageTypes Linkage) {
288 std::string VarName = getInstrProfNameVarPrefix();
291 if (!GlobalValue::isLocalLinkage(Linkage))
294 // Now fix up illegal chars in local VarName that may upset the assembler.
295 const char *InvalidChars = "-:<>/\"'";
296 size_t found = VarName.find_first_of(InvalidChars);
297 while (found != std::string::npos) {
298 VarName[found] = '_';
299 found = VarName.find_first_of(InvalidChars, found + 1);
304 GlobalVariable *createPGOFuncNameVar(Module &M,
305 GlobalValue::LinkageTypes Linkage,
306 StringRef PGOFuncName) {
307 // We generally want to match the function's linkage, but available_externally
308 // and extern_weak both have the wrong semantics, and anything that doesn't
309 // need to link across compilation units doesn't need to be visible at all.
310 if (Linkage == GlobalValue::ExternalWeakLinkage)
311 Linkage = GlobalValue::LinkOnceAnyLinkage;
312 else if (Linkage == GlobalValue::AvailableExternallyLinkage)
313 Linkage = GlobalValue::LinkOnceODRLinkage;
314 else if (Linkage == GlobalValue::InternalLinkage ||
315 Linkage == GlobalValue::ExternalLinkage)
316 Linkage = GlobalValue::PrivateLinkage;
319 ConstantDataArray::getString(M.getContext(), PGOFuncName, false);
321 new GlobalVariable(M, Value->getType(), true, Linkage, Value,
322 getPGOFuncNameVarName(PGOFuncName, Linkage));
324 // Hide the symbol so that we correctly get a copy for each executable.
325 if (!GlobalValue::isLocalLinkage(FuncNameVar->getLinkage()))
326 FuncNameVar->setVisibility(GlobalValue::HiddenVisibility);
331 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName) {
332 return createPGOFuncNameVar(*F.getParent(), F.getLinkage(), PGOFuncName);
335 Error InstrProfSymtab::create(Module &M, bool InLTO) {
336 for (Function &F : M) {
337 // Function may not have a name: like using asm("") to overwrite the name.
338 // Ignore in this case.
341 const std::string &PGOFuncName = getPGOFuncName(F, InLTO);
342 if (Error E = addFuncName(PGOFuncName))
344 MD5FuncMap.emplace_back(Function::getGUID(PGOFuncName), &F);
345 // In ThinLTO, local function may have been promoted to global and have
346 // suffix added to the function name. We need to add the stripped function
347 // name to the symbol table so that we can find a match from profile.
349 auto pos = PGOFuncName.find('.');
350 if (pos != std::string::npos) {
351 const std::string &OtherFuncName = PGOFuncName.substr(0, pos);
352 if (Error E = addFuncName(OtherFuncName))
354 MD5FuncMap.emplace_back(Function::getGUID(OtherFuncName), &F);
360 return Error::success();
363 Error collectPGOFuncNameStrings(ArrayRef<std::string> NameStrs,
364 bool doCompression, std::string &Result) {
365 assert(!NameStrs.empty() && "No name data to emit");
367 uint8_t Header[16], *P = Header;
368 std::string UncompressedNameStrings =
369 join(NameStrs.begin(), NameStrs.end(), getInstrProfNameSeparator());
371 assert(StringRef(UncompressedNameStrings)
372 .count(getInstrProfNameSeparator()) == (NameStrs.size() - 1) &&
373 "PGO name is invalid (contains separator token)");
375 unsigned EncLen = encodeULEB128(UncompressedNameStrings.length(), P);
378 auto WriteStringToResult = [&](size_t CompressedLen, StringRef InputStr) {
379 EncLen = encodeULEB128(CompressedLen, P);
381 char *HeaderStr = reinterpret_cast<char *>(&Header[0]);
382 unsigned HeaderLen = P - &Header[0];
383 Result.append(HeaderStr, HeaderLen);
385 return Error::success();
388 if (!doCompression) {
389 return WriteStringToResult(0, UncompressedNameStrings);
392 SmallString<128> CompressedNameStrings;
393 Error E = zlib::compress(StringRef(UncompressedNameStrings),
394 CompressedNameStrings, zlib::BestSizeCompression);
396 consumeError(std::move(E));
397 return make_error<InstrProfError>(instrprof_error::compress_failed);
400 return WriteStringToResult(CompressedNameStrings.size(),
401 CompressedNameStrings);
404 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar) {
405 auto *Arr = cast<ConstantDataArray>(NameVar->getInitializer());
407 Arr->isCString() ? Arr->getAsCString() : Arr->getAsString();
411 Error collectPGOFuncNameStrings(ArrayRef<GlobalVariable *> NameVars,
412 std::string &Result, bool doCompression) {
413 std::vector<std::string> NameStrs;
414 for (auto *NameVar : NameVars) {
415 NameStrs.push_back(getPGOFuncNameVarInitializer(NameVar));
417 return collectPGOFuncNameStrings(
418 NameStrs, zlib::isAvailable() && doCompression, Result);
421 Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab) {
422 const uint8_t *P = reinterpret_cast<const uint8_t *>(NameStrings.data());
423 const uint8_t *EndP = reinterpret_cast<const uint8_t *>(NameStrings.data() +
427 uint64_t UncompressedSize = decodeULEB128(P, &N);
429 uint64_t CompressedSize = decodeULEB128(P, &N);
431 bool isCompressed = (CompressedSize != 0);
432 SmallString<128> UncompressedNameStrings;
433 StringRef NameStrings;
435 if (!llvm::zlib::isAvailable())
436 return make_error<InstrProfError>(instrprof_error::zlib_unavailable);
438 StringRef CompressedNameStrings(reinterpret_cast<const char *>(P),
441 zlib::uncompress(CompressedNameStrings, UncompressedNameStrings,
443 consumeError(std::move(E));
444 return make_error<InstrProfError>(instrprof_error::uncompress_failed);
447 NameStrings = StringRef(UncompressedNameStrings.data(),
448 UncompressedNameStrings.size());
451 StringRef(reinterpret_cast<const char *>(P), UncompressedSize);
452 P += UncompressedSize;
454 // Now parse the name strings.
455 SmallVector<StringRef, 0> Names;
456 NameStrings.split(Names, getInstrProfNameSeparator());
457 for (StringRef &Name : Names)
458 if (Error E = Symtab.addFuncName(Name))
461 while (P < EndP && *P == 0)
464 Symtab.finalizeSymtab();
465 return Error::success();
468 void InstrProfValueSiteRecord::merge(InstrProfValueSiteRecord &Input,
470 function_ref<void(instrprof_error)> Warn) {
471 this->sortByTargetValues();
472 Input.sortByTargetValues();
473 auto I = ValueData.begin();
474 auto IE = ValueData.end();
475 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
477 while (I != IE && I->Value < J->Value)
479 if (I != IE && I->Value == J->Value) {
481 I->Count = SaturatingMultiplyAdd(J->Count, Weight, I->Count, &Overflowed);
483 Warn(instrprof_error::counter_overflow);
487 ValueData.insert(I, *J);
491 void InstrProfValueSiteRecord::scale(uint64_t Weight,
492 function_ref<void(instrprof_error)> Warn) {
493 for (auto I = ValueData.begin(), IE = ValueData.end(); I != IE; ++I) {
495 I->Count = SaturatingMultiply(I->Count, Weight, &Overflowed);
497 Warn(instrprof_error::counter_overflow);
501 // Merge Value Profile data from Src record to this record for ValueKind.
502 // Scale merged value counts by \p Weight.
503 void InstrProfRecord::mergeValueProfData(
504 uint32_t ValueKind, InstrProfRecord &Src, uint64_t Weight,
505 function_ref<void(instrprof_error)> Warn) {
506 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
507 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
508 if (ThisNumValueSites != OtherNumValueSites) {
509 Warn(instrprof_error::value_site_count_mismatch);
512 if (!ThisNumValueSites)
514 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
515 getOrCreateValueSitesForKind(ValueKind);
516 MutableArrayRef<InstrProfValueSiteRecord> OtherSiteRecords =
517 Src.getValueSitesForKind(ValueKind);
518 for (uint32_t I = 0; I < ThisNumValueSites; I++)
519 ThisSiteRecords[I].merge(OtherSiteRecords[I], Weight, Warn);
522 void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight,
523 function_ref<void(instrprof_error)> Warn) {
524 // If the number of counters doesn't match we either have bad data
525 // or a hash collision.
526 if (Counts.size() != Other.Counts.size()) {
527 Warn(instrprof_error::count_mismatch);
531 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
534 SaturatingMultiplyAdd(Other.Counts[I], Weight, Counts[I], &Overflowed);
536 Warn(instrprof_error::counter_overflow);
539 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
540 mergeValueProfData(Kind, Other, Weight, Warn);
543 void InstrProfRecord::scaleValueProfData(
544 uint32_t ValueKind, uint64_t Weight,
545 function_ref<void(instrprof_error)> Warn) {
546 for (auto &R : getValueSitesForKind(ValueKind))
547 R.scale(Weight, Warn);
550 void InstrProfRecord::scale(uint64_t Weight,
551 function_ref<void(instrprof_error)> Warn) {
552 for (auto &Count : this->Counts) {
554 Count = SaturatingMultiply(Count, Weight, &Overflowed);
556 Warn(instrprof_error::counter_overflow);
558 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
559 scaleValueProfData(Kind, Weight, Warn);
562 // Map indirect call target name hash to name string.
563 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind,
564 ValueMapType *ValueMap) {
568 case IPVK_IndirectCallTarget: {
570 std::lower_bound(ValueMap->begin(), ValueMap->end(), Value,
571 [](const std::pair<uint64_t, uint64_t> &LHS,
572 uint64_t RHS) { return LHS.first < RHS; });
573 // Raw function pointer collected by value profiler may be from
574 // external functions that are not instrumented. They won't have
575 // mapping data to be used by the deserializer. Force the value to
576 // be 0 in this case.
577 if (Result != ValueMap->end() && Result->first == Value)
578 Value = (uint64_t)Result->second;
587 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
588 InstrProfValueData *VData, uint32_t N,
589 ValueMapType *ValueMap) {
590 for (uint32_t I = 0; I < N; I++) {
591 VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap);
593 std::vector<InstrProfValueSiteRecord> &ValueSites =
594 getOrCreateValueSitesForKind(ValueKind);
596 ValueSites.emplace_back();
598 ValueSites.emplace_back(VData, VData + N);
601 #define INSTR_PROF_COMMON_API_IMPL
602 #include "llvm/ProfileData/InstrProfData.inc"
605 * \brief ValueProfRecordClosure Interface implementation for InstrProfRecord
606 * class. These C wrappers are used as adaptors so that C++ code can be
607 * invoked as callbacks.
609 uint32_t getNumValueKindsInstrProf(const void *Record) {
610 return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds();
613 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) {
614 return reinterpret_cast<const InstrProfRecord *>(Record)
615 ->getNumValueSites(VKind);
618 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) {
619 return reinterpret_cast<const InstrProfRecord *>(Record)
620 ->getNumValueData(VKind);
623 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK,
625 return reinterpret_cast<const InstrProfRecord *>(R)
626 ->getNumValueDataForSite(VK, S);
629 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst,
630 uint32_t K, uint32_t S) {
631 reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(Dst, K, S);
634 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) {
636 (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData());
637 memset(VD, 0, TotalSizeInBytes);
641 static ValueProfRecordClosure InstrProfRecordClosure = {
643 getNumValueKindsInstrProf,
644 getNumValueSitesInstrProf,
645 getNumValueDataInstrProf,
646 getNumValueDataForSiteInstrProf,
648 getValueForSiteInstrProf,
649 allocValueProfDataInstrProf};
651 // Wrapper implementation using the closure mechanism.
652 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) {
653 auto Closure = InstrProfRecordClosure;
654 Closure.Record = &Record;
655 return getValueProfDataSize(&Closure);
658 // Wrapper implementation using the closure mechanism.
659 std::unique_ptr<ValueProfData>
660 ValueProfData::serializeFrom(const InstrProfRecord &Record) {
661 InstrProfRecordClosure.Record = &Record;
663 std::unique_ptr<ValueProfData> VPD(
664 serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr));
668 void ValueProfRecord::deserializeTo(InstrProfRecord &Record,
669 InstrProfRecord::ValueMapType *VMap) {
670 Record.reserveSites(Kind, NumValueSites);
672 InstrProfValueData *ValueData = getValueProfRecordValueData(this);
673 for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) {
674 uint8_t ValueDataCount = this->SiteCountArray[VSite];
675 Record.addValueData(Kind, VSite, ValueData, ValueDataCount, VMap);
676 ValueData += ValueDataCount;
680 // For writing/serializing, Old is the host endianness, and New is
681 // byte order intended on disk. For Reading/deserialization, Old
682 // is the on-disk source endianness, and New is the host endianness.
683 void ValueProfRecord::swapBytes(support::endianness Old,
684 support::endianness New) {
685 using namespace support;
690 if (getHostEndianness() != Old) {
691 sys::swapByteOrder<uint32_t>(NumValueSites);
692 sys::swapByteOrder<uint32_t>(Kind);
694 uint32_t ND = getValueProfRecordNumValueData(this);
695 InstrProfValueData *VD = getValueProfRecordValueData(this);
697 // No need to swap byte array: SiteCountArrray.
698 for (uint32_t I = 0; I < ND; I++) {
699 sys::swapByteOrder<uint64_t>(VD[I].Value);
700 sys::swapByteOrder<uint64_t>(VD[I].Count);
702 if (getHostEndianness() == Old) {
703 sys::swapByteOrder<uint32_t>(NumValueSites);
704 sys::swapByteOrder<uint32_t>(Kind);
708 void ValueProfData::deserializeTo(InstrProfRecord &Record,
709 InstrProfRecord::ValueMapType *VMap) {
710 if (NumValueKinds == 0)
713 ValueProfRecord *VR = getFirstValueProfRecord(this);
714 for (uint32_t K = 0; K < NumValueKinds; K++) {
715 VR->deserializeTo(Record, VMap);
716 VR = getValueProfRecordNext(VR);
721 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) {
722 using namespace support;
725 return endian::readNext<T, little, unaligned>(D);
727 return endian::readNext<T, big, unaligned>(D);
730 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) {
731 return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize))
735 Error ValueProfData::checkIntegrity() {
736 if (NumValueKinds > IPVK_Last + 1)
737 return make_error<InstrProfError>(instrprof_error::malformed);
738 // Total size needs to be mulltiple of quadword size.
739 if (TotalSize % sizeof(uint64_t))
740 return make_error<InstrProfError>(instrprof_error::malformed);
742 ValueProfRecord *VR = getFirstValueProfRecord(this);
743 for (uint32_t K = 0; K < this->NumValueKinds; K++) {
744 if (VR->Kind > IPVK_Last)
745 return make_error<InstrProfError>(instrprof_error::malformed);
746 VR = getValueProfRecordNext(VR);
747 if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize)
748 return make_error<InstrProfError>(instrprof_error::malformed);
750 return Error::success();
753 Expected<std::unique_ptr<ValueProfData>>
754 ValueProfData::getValueProfData(const unsigned char *D,
755 const unsigned char *const BufferEnd,
756 support::endianness Endianness) {
757 using namespace support;
759 if (D + sizeof(ValueProfData) > BufferEnd)
760 return make_error<InstrProfError>(instrprof_error::truncated);
762 const unsigned char *Header = D;
763 uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness);
764 if (D + TotalSize > BufferEnd)
765 return make_error<InstrProfError>(instrprof_error::too_large);
767 std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize);
768 memcpy(VPD.get(), D, TotalSize);
770 VPD->swapBytesToHost(Endianness);
772 Error E = VPD->checkIntegrity();
776 return std::move(VPD);
779 void ValueProfData::swapBytesToHost(support::endianness Endianness) {
780 using namespace support;
782 if (Endianness == getHostEndianness())
785 sys::swapByteOrder<uint32_t>(TotalSize);
786 sys::swapByteOrder<uint32_t>(NumValueKinds);
788 ValueProfRecord *VR = getFirstValueProfRecord(this);
789 for (uint32_t K = 0; K < NumValueKinds; K++) {
790 VR->swapBytes(Endianness, getHostEndianness());
791 VR = getValueProfRecordNext(VR);
795 void ValueProfData::swapBytesFromHost(support::endianness Endianness) {
796 using namespace support;
798 if (Endianness == getHostEndianness())
801 ValueProfRecord *VR = getFirstValueProfRecord(this);
802 for (uint32_t K = 0; K < NumValueKinds; K++) {
803 ValueProfRecord *NVR = getValueProfRecordNext(VR);
804 VR->swapBytes(getHostEndianness(), Endianness);
807 sys::swapByteOrder<uint32_t>(TotalSize);
808 sys::swapByteOrder<uint32_t>(NumValueKinds);
811 void annotateValueSite(Module &M, Instruction &Inst,
812 const InstrProfRecord &InstrProfR,
813 InstrProfValueKind ValueKind, uint32_t SiteIdx,
814 uint32_t MaxMDCount) {
815 uint32_t NV = InstrProfR.getNumValueDataForSite(ValueKind, SiteIdx);
820 std::unique_ptr<InstrProfValueData[]> VD =
821 InstrProfR.getValueForSite(ValueKind, SiteIdx, &Sum);
823 ArrayRef<InstrProfValueData> VDs(VD.get(), NV);
824 annotateValueSite(M, Inst, VDs, Sum, ValueKind, MaxMDCount);
827 void annotateValueSite(Module &M, Instruction &Inst,
828 ArrayRef<InstrProfValueData> VDs,
829 uint64_t Sum, InstrProfValueKind ValueKind,
830 uint32_t MaxMDCount) {
831 LLVMContext &Ctx = M.getContext();
832 MDBuilder MDHelper(Ctx);
833 SmallVector<Metadata *, 3> Vals;
835 Vals.push_back(MDHelper.createString("VP"));
837 Vals.push_back(MDHelper.createConstant(
838 ConstantInt::get(Type::getInt32Ty(Ctx), ValueKind)));
841 MDHelper.createConstant(ConstantInt::get(Type::getInt64Ty(Ctx), Sum)));
843 // Value Profile Data
844 uint32_t MDCount = MaxMDCount;
845 for (auto &VD : VDs) {
846 Vals.push_back(MDHelper.createConstant(
847 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Value)));
848 Vals.push_back(MDHelper.createConstant(
849 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Count)));
853 Inst.setMetadata(LLVMContext::MD_prof, MDNode::get(Ctx, Vals));
856 bool getValueProfDataFromInst(const Instruction &Inst,
857 InstrProfValueKind ValueKind,
858 uint32_t MaxNumValueData,
859 InstrProfValueData ValueData[],
860 uint32_t &ActualNumValueData, uint64_t &TotalC) {
861 MDNode *MD = Inst.getMetadata(LLVMContext::MD_prof);
865 unsigned NOps = MD->getNumOperands();
870 // Operand 0 is a string tag "VP":
871 MDString *Tag = cast<MDString>(MD->getOperand(0));
875 if (!Tag->getString().equals("VP"))
879 ConstantInt *KindInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
882 if (KindInt->getZExtValue() != ValueKind)
886 ConstantInt *TotalCInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
889 TotalC = TotalCInt->getZExtValue();
891 ActualNumValueData = 0;
893 for (unsigned I = 3; I < NOps; I += 2) {
894 if (ActualNumValueData >= MaxNumValueData)
896 ConstantInt *Value = mdconst::dyn_extract<ConstantInt>(MD->getOperand(I));
898 mdconst::dyn_extract<ConstantInt>(MD->getOperand(I + 1));
899 if (!Value || !Count)
901 ValueData[ActualNumValueData].Value = Value->getZExtValue();
902 ValueData[ActualNumValueData].Count = Count->getZExtValue();
903 ActualNumValueData++;
908 MDNode *getPGOFuncNameMetadata(const Function &F) {
909 return F.getMetadata(getPGOFuncNameMetadataName());
912 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName) {
913 // Only for internal linkage functions.
914 if (PGOFuncName == F.getName())
916 // Don't create duplicated meta-data.
917 if (getPGOFuncNameMetadata(F))
919 LLVMContext &C = F.getContext();
920 MDNode *N = MDNode::get(C, MDString::get(C, PGOFuncName));
921 F.setMetadata(getPGOFuncNameMetadataName(), N);
924 bool needsComdatForCounter(const Function &F, const Module &M) {
928 Triple TT(M.getTargetTriple());
929 if (!TT.isOSBinFormatELF() && !TT.isOSBinFormatWasm())
932 // See createPGOFuncNameVar for more details. To avoid link errors, profile
933 // counters for function with available_externally linkage needs to be changed
934 // to linkonce linkage. On ELF based systems, this leads to weak symbols to be
935 // created. Without using comdat, duplicate entries won't be removed by the
936 // linker leading to increased data segement size and raw profile size. Even
937 // worse, since the referenced counter from profile per-function data object
938 // will be resolved to the common strong definition, the profile counts for
939 // available_externally functions will end up being duplicated in raw profile
940 // data. This can result in distorted profile as the counts of those dups
941 // will be accumulated by the profile merger.
942 GlobalValue::LinkageTypes Linkage = F.getLinkage();
943 if (Linkage != GlobalValue::ExternalWeakLinkage &&
944 Linkage != GlobalValue::AvailableExternallyLinkage)
950 // Check if INSTR_PROF_RAW_VERSION_VAR is defined.
951 bool isIRPGOFlagSet(const Module *M) {
953 M->getNamedGlobal(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
954 if (!IRInstrVar || IRInstrVar->isDeclaration() ||
955 IRInstrVar->hasLocalLinkage())
958 // Check if the flag is set.
959 if (!IRInstrVar->hasInitializer())
962 const Constant *InitVal = IRInstrVar->getInitializer();
966 return (dyn_cast<ConstantInt>(InitVal)->getZExtValue() &
967 VARIANT_MASK_IR_PROF) != 0;
970 // Check if we can safely rename this Comdat function.
971 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken) {
972 if (F.getName().empty())
974 if (!needsComdatForCounter(F, *(F.getParent())))
976 // Unsafe to rename the address-taken function (which can be used in
977 // function comparison).
978 if (CheckAddressTaken && F.hasAddressTaken())
980 // Only safe to do if this function may be discarded if it is not used
981 // in the compilation unit.
982 if (!GlobalValue::isDiscardableIfUnused(F.getLinkage()))
985 // For AvailableExternallyLinkage functions.
986 if (!F.hasComdat()) {
987 assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage);
993 // Parse the value profile options.
994 void getMemOPSizeRangeFromOption(StringRef MemOPSizeRange, int64_t &RangeStart,
995 int64_t &RangeLast) {
996 static const int64_t DefaultMemOPSizeRangeStart = 0;
997 static const int64_t DefaultMemOPSizeRangeLast = 8;
998 RangeStart = DefaultMemOPSizeRangeStart;
999 RangeLast = DefaultMemOPSizeRangeLast;
1001 if (!MemOPSizeRange.empty()) {
1002 auto Pos = MemOPSizeRange.find(':');
1003 if (Pos != std::string::npos) {
1005 MemOPSizeRange.substr(0, Pos).getAsInteger(10, RangeStart);
1006 if (Pos < MemOPSizeRange.size() - 1)
1007 MemOPSizeRange.substr(Pos + 1).getAsInteger(10, RangeLast);
1009 MemOPSizeRange.getAsInteger(10, RangeLast);
1011 assert(RangeLast >= RangeStart);
1014 } // end namespace llvm