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 enum InstrProfSectKind {
142 #define INSTR_PROF_SECT_ENTRY(Kind, SectName, SectNameCommon, SectNameCoff, \
145 #include "llvm/ProfileData/InstrProfData.inc"
148 const char *InstrProfSectName[] = {
149 #define INSTR_PROF_SECT_ENTRY(Kind, SectName, SectNameCommon, SectNameCoff, \
152 #include "llvm/ProfileData/InstrProfData.inc"
155 const char *InstrProfSectNameCommon[] = {
156 #define INSTR_PROF_SECT_ENTRY(Kind, SectName, SectNameCommon, SectNameCoff, \
159 #include "llvm/ProfileData/InstrProfData.inc"
162 const char *InstrProfSectNameCoff[] = {
163 #define INSTR_PROF_SECT_ENTRY(Kind, SectName, SectNameCommon, SectNameCoff, \
166 #include "llvm/ProfileData/InstrProfData.inc"
169 const char *InstrProfSectNamePrefix[] = {
170 #define INSTR_PROF_SECT_ENTRY(Kind, SectName, SectNameCommon, SectNameCoff, \
173 #include "llvm/ProfileData/InstrProfData.inc"
176 std::string getInstrProfSectionName(const Module *M, InstrProfSectKind Kind) {
179 return InstrProfSectName[Kind];
181 bool AddSegment = Triple(M->getTargetTriple()).isOSBinFormatMachO();
182 std::string SectName;
183 if (Triple(M->getTargetTriple()).isOSBinFormatCOFF())
184 SectName = InstrProfSectNameCoff[Kind];
186 SectName = InstrProfSectNameCommon[Kind];
189 SectName = InstrProfSectNamePrefix[Kind] + SectName;
190 if (Kind == IPSK_data) {
191 SectName += ",regular,live_support";
201 std::string getInstrProfCountersSectionName(const Module *M) {
202 return getInstrProfSectionName(M, IPSK_cnts);
205 std::string getInstrProfNameSectionName(const Module *M) {
206 return getInstrProfSectionName(M, IPSK_name);
209 std::string getInstrProfDataSectionName(const Module *M) {
210 return getInstrProfSectionName(M, IPSK_data);
213 std::string getInstrProfValuesSectionName(const Module *M) {
214 return getInstrProfSectionName(M, IPSK_vals);
217 std::string getInstrProfVNodesSectionName(const Module *M) {
218 return getInstrProfSectionName(M, IPSK_vnodes);
221 std::string getInstrProfCoverageSectionName(const Module *M) {
222 return getInstrProfSectionName(M, IPSK_covmap);
225 void SoftInstrProfErrors::addError(instrprof_error IE) {
226 if (IE == instrprof_error::success)
229 if (FirstError == instrprof_error::success)
233 case instrprof_error::hash_mismatch:
236 case instrprof_error::count_mismatch:
237 ++NumCountMismatches;
239 case instrprof_error::counter_overflow:
240 ++NumCounterOverflows;
242 case instrprof_error::value_site_count_mismatch:
243 ++NumValueSiteCountMismatches;
246 llvm_unreachable("Not a soft error");
250 std::string InstrProfError::message() const {
251 return getInstrProfErrString(Err);
254 char InstrProfError::ID = 0;
256 std::string getPGOFuncName(StringRef RawFuncName,
257 GlobalValue::LinkageTypes Linkage,
259 uint64_t Version LLVM_ATTRIBUTE_UNUSED) {
260 return GlobalValue::getGlobalIdentifier(RawFuncName, Linkage, FileName);
263 // Strip NumPrefix level of directory name from PathNameStr. If the number of
264 // directory separators is less than NumPrefix, strip all the directories and
265 // leave base file name only.
266 static StringRef stripDirPrefix(StringRef PathNameStr, uint32_t NumPrefix) {
267 uint32_t Count = NumPrefix;
268 uint32_t Pos = 0, LastPos = 0;
269 for (auto & CI : PathNameStr) {
271 if (llvm::sys::path::is_separator(CI)) {
278 return PathNameStr.substr(LastPos);
281 // Return the PGOFuncName. This function has some special handling when called
282 // in LTO optimization. The following only applies when calling in LTO passes
283 // (when \c InLTO is true): LTO's internalization privatizes many global linkage
284 // symbols. This happens after value profile annotation, but those internal
285 // linkage functions should not have a source prefix.
286 // Additionally, for ThinLTO mode, exported internal functions are promoted
287 // and renamed. We need to ensure that the original internal PGO name is
288 // used when computing the GUID that is compared against the profiled GUIDs.
289 // To differentiate compiler generated internal symbols from original ones,
290 // PGOFuncName meta data are created and attached to the original internal
291 // symbols in the value profile annotation step
292 // (PGOUseFunc::annotateIndirectCallSites). If a symbol does not have the meta
293 // data, its original linkage must be non-internal.
294 std::string getPGOFuncName(const Function &F, bool InLTO, uint64_t Version) {
296 StringRef FileName = (StaticFuncFullModulePrefix
297 ? F.getParent()->getName()
298 : sys::path::filename(F.getParent()->getName()));
299 if (StaticFuncFullModulePrefix && StaticFuncStripDirNamePrefix != 0)
300 FileName = stripDirPrefix(FileName, StaticFuncStripDirNamePrefix);
301 return getPGOFuncName(F.getName(), F.getLinkage(), FileName, Version);
304 // In LTO mode (when InLTO is true), first check if there is a meta data.
305 if (MDNode *MD = getPGOFuncNameMetadata(F)) {
306 StringRef S = cast<MDString>(MD->getOperand(0))->getString();
310 // If there is no meta data, the function must be a global before the value
311 // profile annotation pass. Its current linkage may be internal if it is
312 // internalized in LTO mode.
313 return getPGOFuncName(F.getName(), GlobalValue::ExternalLinkage, "");
316 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, StringRef FileName) {
317 if (FileName.empty())
319 // Drop the file name including ':'. See also getPGOFuncName.
320 if (PGOFuncName.startswith(FileName))
321 PGOFuncName = PGOFuncName.drop_front(FileName.size() + 1);
325 // \p FuncName is the string used as profile lookup key for the function. A
326 // symbol is created to hold the name. Return the legalized symbol name.
327 std::string getPGOFuncNameVarName(StringRef FuncName,
328 GlobalValue::LinkageTypes Linkage) {
329 std::string VarName = getInstrProfNameVarPrefix();
332 if (!GlobalValue::isLocalLinkage(Linkage))
335 // Now fix up illegal chars in local VarName that may upset the assembler.
336 const char *InvalidChars = "-:<>/\"'";
337 size_t found = VarName.find_first_of(InvalidChars);
338 while (found != std::string::npos) {
339 VarName[found] = '_';
340 found = VarName.find_first_of(InvalidChars, found + 1);
345 GlobalVariable *createPGOFuncNameVar(Module &M,
346 GlobalValue::LinkageTypes Linkage,
347 StringRef PGOFuncName) {
348 // We generally want to match the function's linkage, but available_externally
349 // and extern_weak both have the wrong semantics, and anything that doesn't
350 // need to link across compilation units doesn't need to be visible at all.
351 if (Linkage == GlobalValue::ExternalWeakLinkage)
352 Linkage = GlobalValue::LinkOnceAnyLinkage;
353 else if (Linkage == GlobalValue::AvailableExternallyLinkage)
354 Linkage = GlobalValue::LinkOnceODRLinkage;
355 else if (Linkage == GlobalValue::InternalLinkage ||
356 Linkage == GlobalValue::ExternalLinkage)
357 Linkage = GlobalValue::PrivateLinkage;
360 ConstantDataArray::getString(M.getContext(), PGOFuncName, false);
362 new GlobalVariable(M, Value->getType(), true, Linkage, Value,
363 getPGOFuncNameVarName(PGOFuncName, Linkage));
365 // Hide the symbol so that we correctly get a copy for each executable.
366 if (!GlobalValue::isLocalLinkage(FuncNameVar->getLinkage()))
367 FuncNameVar->setVisibility(GlobalValue::HiddenVisibility);
372 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName) {
373 return createPGOFuncNameVar(*F.getParent(), F.getLinkage(), PGOFuncName);
376 void InstrProfSymtab::create(Module &M, bool InLTO) {
377 for (Function &F : M) {
378 // Function may not have a name: like using asm("") to overwrite the name.
379 // Ignore in this case.
382 const std::string &PGOFuncName = getPGOFuncName(F, InLTO);
383 addFuncName(PGOFuncName);
384 MD5FuncMap.emplace_back(Function::getGUID(PGOFuncName), &F);
385 // In ThinLTO, local function may have been promoted to global and have
386 // suffix added to the function name. We need to add the stripped function
387 // name to the symbol table so that we can find a match from profile.
389 auto pos = PGOFuncName.find('.');
390 if (pos != std::string::npos) {
391 const std::string &OtherFuncName = PGOFuncName.substr(0, pos);
392 addFuncName(OtherFuncName);
393 MD5FuncMap.emplace_back(Function::getGUID(OtherFuncName), &F);
401 Error collectPGOFuncNameStrings(const std::vector<std::string> &NameStrs,
402 bool doCompression, std::string &Result) {
403 assert(!NameStrs.empty() && "No name data to emit");
405 uint8_t Header[16], *P = Header;
406 std::string UncompressedNameStrings =
407 join(NameStrs.begin(), NameStrs.end(), getInstrProfNameSeparator());
409 assert(StringRef(UncompressedNameStrings)
410 .count(getInstrProfNameSeparator()) == (NameStrs.size() - 1) &&
411 "PGO name is invalid (contains separator token)");
413 unsigned EncLen = encodeULEB128(UncompressedNameStrings.length(), P);
416 auto WriteStringToResult = [&](size_t CompressedLen, StringRef InputStr) {
417 EncLen = encodeULEB128(CompressedLen, P);
419 char *HeaderStr = reinterpret_cast<char *>(&Header[0]);
420 unsigned HeaderLen = P - &Header[0];
421 Result.append(HeaderStr, HeaderLen);
423 return Error::success();
426 if (!doCompression) {
427 return WriteStringToResult(0, UncompressedNameStrings);
430 SmallString<128> CompressedNameStrings;
431 Error E = zlib::compress(StringRef(UncompressedNameStrings),
432 CompressedNameStrings, zlib::BestSizeCompression);
434 consumeError(std::move(E));
435 return make_error<InstrProfError>(instrprof_error::compress_failed);
438 return WriteStringToResult(CompressedNameStrings.size(),
439 CompressedNameStrings);
442 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar) {
443 auto *Arr = cast<ConstantDataArray>(NameVar->getInitializer());
445 Arr->isCString() ? Arr->getAsCString() : Arr->getAsString();
449 Error collectPGOFuncNameStrings(const std::vector<GlobalVariable *> &NameVars,
450 std::string &Result, bool doCompression) {
451 std::vector<std::string> NameStrs;
452 for (auto *NameVar : NameVars) {
453 NameStrs.push_back(getPGOFuncNameVarInitializer(NameVar));
455 return collectPGOFuncNameStrings(
456 NameStrs, zlib::isAvailable() && doCompression, Result);
459 Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab) {
460 const uint8_t *P = reinterpret_cast<const uint8_t *>(NameStrings.data());
461 const uint8_t *EndP = reinterpret_cast<const uint8_t *>(NameStrings.data() +
465 uint64_t UncompressedSize = decodeULEB128(P, &N);
467 uint64_t CompressedSize = decodeULEB128(P, &N);
469 bool isCompressed = (CompressedSize != 0);
470 SmallString<128> UncompressedNameStrings;
471 StringRef NameStrings;
473 StringRef CompressedNameStrings(reinterpret_cast<const char *>(P),
476 zlib::uncompress(CompressedNameStrings, UncompressedNameStrings,
478 consumeError(std::move(E));
479 return make_error<InstrProfError>(instrprof_error::uncompress_failed);
482 NameStrings = StringRef(UncompressedNameStrings.data(),
483 UncompressedNameStrings.size());
486 StringRef(reinterpret_cast<const char *>(P), UncompressedSize);
487 P += UncompressedSize;
489 // Now parse the name strings.
490 SmallVector<StringRef, 0> Names;
491 NameStrings.split(Names, getInstrProfNameSeparator());
492 for (StringRef &Name : Names)
493 Symtab.addFuncName(Name);
495 while (P < EndP && *P == 0)
498 Symtab.finalizeSymtab();
499 return Error::success();
502 void InstrProfValueSiteRecord::merge(SoftInstrProfErrors &SIPE,
503 InstrProfValueSiteRecord &Input,
505 this->sortByTargetValues();
506 Input.sortByTargetValues();
507 auto I = ValueData.begin();
508 auto IE = ValueData.end();
509 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
511 while (I != IE && I->Value < J->Value)
513 if (I != IE && I->Value == J->Value) {
515 I->Count = SaturatingMultiplyAdd(J->Count, Weight, I->Count, &Overflowed);
517 SIPE.addError(instrprof_error::counter_overflow);
521 ValueData.insert(I, *J);
525 void InstrProfValueSiteRecord::scale(SoftInstrProfErrors &SIPE,
527 for (auto I = ValueData.begin(), IE = ValueData.end(); I != IE; ++I) {
529 I->Count = SaturatingMultiply(I->Count, Weight, &Overflowed);
531 SIPE.addError(instrprof_error::counter_overflow);
535 // Merge Value Profile data from Src record to this record for ValueKind.
536 // Scale merged value counts by \p Weight.
537 void InstrProfRecord::mergeValueProfData(uint32_t ValueKind,
538 InstrProfRecord &Src,
540 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
541 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
542 if (ThisNumValueSites != OtherNumValueSites) {
543 SIPE.addError(instrprof_error::value_site_count_mismatch);
546 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
547 getValueSitesForKind(ValueKind);
548 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
549 Src.getValueSitesForKind(ValueKind);
550 for (uint32_t I = 0; I < ThisNumValueSites; I++)
551 ThisSiteRecords[I].merge(SIPE, OtherSiteRecords[I], Weight);
554 void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight) {
555 // If the number of counters doesn't match we either have bad data
556 // or a hash collision.
557 if (Counts.size() != Other.Counts.size()) {
558 SIPE.addError(instrprof_error::count_mismatch);
562 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
565 SaturatingMultiplyAdd(Other.Counts[I], Weight, Counts[I], &Overflowed);
567 SIPE.addError(instrprof_error::counter_overflow);
570 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
571 mergeValueProfData(Kind, Other, Weight);
574 void InstrProfRecord::scaleValueProfData(uint32_t ValueKind, uint64_t Weight) {
575 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
576 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
577 getValueSitesForKind(ValueKind);
578 for (uint32_t I = 0; I < ThisNumValueSites; I++)
579 ThisSiteRecords[I].scale(SIPE, Weight);
582 void InstrProfRecord::scale(uint64_t Weight) {
583 for (auto &Count : this->Counts) {
585 Count = SaturatingMultiply(Count, Weight, &Overflowed);
587 SIPE.addError(instrprof_error::counter_overflow);
589 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
590 scaleValueProfData(Kind, Weight);
593 // Map indirect call target name hash to name string.
594 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind,
595 ValueMapType *ValueMap) {
599 case IPVK_IndirectCallTarget: {
601 std::lower_bound(ValueMap->begin(), ValueMap->end(), Value,
602 [](const std::pair<uint64_t, uint64_t> &LHS,
603 uint64_t RHS) { return LHS.first < RHS; });
604 // Raw function pointer collected by value profiler may be from
605 // external functions that are not instrumented. They won't have
606 // mapping data to be used by the deserializer. Force the value to
607 // be 0 in this case.
608 if (Result != ValueMap->end() && Result->first == Value)
609 Value = (uint64_t)Result->second;
618 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
619 InstrProfValueData *VData, uint32_t N,
620 ValueMapType *ValueMap) {
621 for (uint32_t I = 0; I < N; I++) {
622 VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap);
624 std::vector<InstrProfValueSiteRecord> &ValueSites =
625 getValueSitesForKind(ValueKind);
627 ValueSites.emplace_back();
629 ValueSites.emplace_back(VData, VData + N);
632 #define INSTR_PROF_COMMON_API_IMPL
633 #include "llvm/ProfileData/InstrProfData.inc"
636 * \brief ValueProfRecordClosure Interface implementation for InstrProfRecord
637 * class. These C wrappers are used as adaptors so that C++ code can be
638 * invoked as callbacks.
640 uint32_t getNumValueKindsInstrProf(const void *Record) {
641 return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds();
644 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) {
645 return reinterpret_cast<const InstrProfRecord *>(Record)
646 ->getNumValueSites(VKind);
649 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) {
650 return reinterpret_cast<const InstrProfRecord *>(Record)
651 ->getNumValueData(VKind);
654 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK,
656 return reinterpret_cast<const InstrProfRecord *>(R)
657 ->getNumValueDataForSite(VK, S);
660 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst,
661 uint32_t K, uint32_t S) {
662 reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(Dst, K, S);
665 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) {
667 (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData());
668 memset(VD, 0, TotalSizeInBytes);
672 static ValueProfRecordClosure InstrProfRecordClosure = {
674 getNumValueKindsInstrProf,
675 getNumValueSitesInstrProf,
676 getNumValueDataInstrProf,
677 getNumValueDataForSiteInstrProf,
679 getValueForSiteInstrProf,
680 allocValueProfDataInstrProf};
682 // Wrapper implementation using the closure mechanism.
683 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) {
684 InstrProfRecordClosure.Record = &Record;
685 return getValueProfDataSize(&InstrProfRecordClosure);
688 // Wrapper implementation using the closure mechanism.
689 std::unique_ptr<ValueProfData>
690 ValueProfData::serializeFrom(const InstrProfRecord &Record) {
691 InstrProfRecordClosure.Record = &Record;
693 std::unique_ptr<ValueProfData> VPD(
694 serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr));
698 void ValueProfRecord::deserializeTo(InstrProfRecord &Record,
699 InstrProfRecord::ValueMapType *VMap) {
700 Record.reserveSites(Kind, NumValueSites);
702 InstrProfValueData *ValueData = getValueProfRecordValueData(this);
703 for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) {
704 uint8_t ValueDataCount = this->SiteCountArray[VSite];
705 Record.addValueData(Kind, VSite, ValueData, ValueDataCount, VMap);
706 ValueData += ValueDataCount;
710 // For writing/serializing, Old is the host endianness, and New is
711 // byte order intended on disk. For Reading/deserialization, Old
712 // is the on-disk source endianness, and New is the host endianness.
713 void ValueProfRecord::swapBytes(support::endianness Old,
714 support::endianness New) {
715 using namespace support;
720 if (getHostEndianness() != Old) {
721 sys::swapByteOrder<uint32_t>(NumValueSites);
722 sys::swapByteOrder<uint32_t>(Kind);
724 uint32_t ND = getValueProfRecordNumValueData(this);
725 InstrProfValueData *VD = getValueProfRecordValueData(this);
727 // No need to swap byte array: SiteCountArrray.
728 for (uint32_t I = 0; I < ND; I++) {
729 sys::swapByteOrder<uint64_t>(VD[I].Value);
730 sys::swapByteOrder<uint64_t>(VD[I].Count);
732 if (getHostEndianness() == Old) {
733 sys::swapByteOrder<uint32_t>(NumValueSites);
734 sys::swapByteOrder<uint32_t>(Kind);
738 void ValueProfData::deserializeTo(InstrProfRecord &Record,
739 InstrProfRecord::ValueMapType *VMap) {
740 if (NumValueKinds == 0)
743 ValueProfRecord *VR = getFirstValueProfRecord(this);
744 for (uint32_t K = 0; K < NumValueKinds; K++) {
745 VR->deserializeTo(Record, VMap);
746 VR = getValueProfRecordNext(VR);
751 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) {
752 using namespace support;
755 return endian::readNext<T, little, unaligned>(D);
757 return endian::readNext<T, big, unaligned>(D);
760 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) {
761 return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize))
765 Error ValueProfData::checkIntegrity() {
766 if (NumValueKinds > IPVK_Last + 1)
767 return make_error<InstrProfError>(instrprof_error::malformed);
768 // Total size needs to be mulltiple of quadword size.
769 if (TotalSize % sizeof(uint64_t))
770 return make_error<InstrProfError>(instrprof_error::malformed);
772 ValueProfRecord *VR = getFirstValueProfRecord(this);
773 for (uint32_t K = 0; K < this->NumValueKinds; K++) {
774 if (VR->Kind > IPVK_Last)
775 return make_error<InstrProfError>(instrprof_error::malformed);
776 VR = getValueProfRecordNext(VR);
777 if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize)
778 return make_error<InstrProfError>(instrprof_error::malformed);
780 return Error::success();
783 Expected<std::unique_ptr<ValueProfData>>
784 ValueProfData::getValueProfData(const unsigned char *D,
785 const unsigned char *const BufferEnd,
786 support::endianness Endianness) {
787 using namespace support;
789 if (D + sizeof(ValueProfData) > BufferEnd)
790 return make_error<InstrProfError>(instrprof_error::truncated);
792 const unsigned char *Header = D;
793 uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness);
794 if (D + TotalSize > BufferEnd)
795 return make_error<InstrProfError>(instrprof_error::too_large);
797 std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize);
798 memcpy(VPD.get(), D, TotalSize);
800 VPD->swapBytesToHost(Endianness);
802 Error E = VPD->checkIntegrity();
806 return std::move(VPD);
809 void ValueProfData::swapBytesToHost(support::endianness Endianness) {
810 using namespace support;
812 if (Endianness == getHostEndianness())
815 sys::swapByteOrder<uint32_t>(TotalSize);
816 sys::swapByteOrder<uint32_t>(NumValueKinds);
818 ValueProfRecord *VR = getFirstValueProfRecord(this);
819 for (uint32_t K = 0; K < NumValueKinds; K++) {
820 VR->swapBytes(Endianness, getHostEndianness());
821 VR = getValueProfRecordNext(VR);
825 void ValueProfData::swapBytesFromHost(support::endianness Endianness) {
826 using namespace support;
828 if (Endianness == getHostEndianness())
831 ValueProfRecord *VR = getFirstValueProfRecord(this);
832 for (uint32_t K = 0; K < NumValueKinds; K++) {
833 ValueProfRecord *NVR = getValueProfRecordNext(VR);
834 VR->swapBytes(getHostEndianness(), Endianness);
837 sys::swapByteOrder<uint32_t>(TotalSize);
838 sys::swapByteOrder<uint32_t>(NumValueKinds);
841 void annotateValueSite(Module &M, Instruction &Inst,
842 const InstrProfRecord &InstrProfR,
843 InstrProfValueKind ValueKind, uint32_t SiteIdx,
844 uint32_t MaxMDCount) {
845 uint32_t NV = InstrProfR.getNumValueDataForSite(ValueKind, SiteIdx);
850 std::unique_ptr<InstrProfValueData[]> VD =
851 InstrProfR.getValueForSite(ValueKind, SiteIdx, &Sum);
853 ArrayRef<InstrProfValueData> VDs(VD.get(), NV);
854 annotateValueSite(M, Inst, VDs, Sum, ValueKind, MaxMDCount);
857 void annotateValueSite(Module &M, Instruction &Inst,
858 ArrayRef<InstrProfValueData> VDs,
859 uint64_t Sum, InstrProfValueKind ValueKind,
860 uint32_t MaxMDCount) {
861 LLVMContext &Ctx = M.getContext();
862 MDBuilder MDHelper(Ctx);
863 SmallVector<Metadata *, 3> Vals;
865 Vals.push_back(MDHelper.createString("VP"));
867 Vals.push_back(MDHelper.createConstant(
868 ConstantInt::get(Type::getInt32Ty(Ctx), ValueKind)));
871 MDHelper.createConstant(ConstantInt::get(Type::getInt64Ty(Ctx), Sum)));
873 // Value Profile Data
874 uint32_t MDCount = MaxMDCount;
875 for (auto &VD : VDs) {
876 Vals.push_back(MDHelper.createConstant(
877 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Value)));
878 Vals.push_back(MDHelper.createConstant(
879 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Count)));
883 Inst.setMetadata(LLVMContext::MD_prof, MDNode::get(Ctx, Vals));
886 bool getValueProfDataFromInst(const Instruction &Inst,
887 InstrProfValueKind ValueKind,
888 uint32_t MaxNumValueData,
889 InstrProfValueData ValueData[],
890 uint32_t &ActualNumValueData, uint64_t &TotalC) {
891 MDNode *MD = Inst.getMetadata(LLVMContext::MD_prof);
895 unsigned NOps = MD->getNumOperands();
900 // Operand 0 is a string tag "VP":
901 MDString *Tag = cast<MDString>(MD->getOperand(0));
905 if (!Tag->getString().equals("VP"))
909 ConstantInt *KindInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
912 if (KindInt->getZExtValue() != ValueKind)
916 ConstantInt *TotalCInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
919 TotalC = TotalCInt->getZExtValue();
921 ActualNumValueData = 0;
923 for (unsigned I = 3; I < NOps; I += 2) {
924 if (ActualNumValueData >= MaxNumValueData)
926 ConstantInt *Value = mdconst::dyn_extract<ConstantInt>(MD->getOperand(I));
928 mdconst::dyn_extract<ConstantInt>(MD->getOperand(I + 1));
929 if (!Value || !Count)
931 ValueData[ActualNumValueData].Value = Value->getZExtValue();
932 ValueData[ActualNumValueData].Count = Count->getZExtValue();
933 ActualNumValueData++;
938 MDNode *getPGOFuncNameMetadata(const Function &F) {
939 return F.getMetadata(getPGOFuncNameMetadataName());
942 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName) {
943 // Only for internal linkage functions.
944 if (PGOFuncName == F.getName())
946 // Don't create duplicated meta-data.
947 if (getPGOFuncNameMetadata(F))
949 LLVMContext &C = F.getContext();
950 MDNode *N = MDNode::get(C, MDString::get(C, PGOFuncName));
951 F.setMetadata(getPGOFuncNameMetadataName(), N);
954 bool needsComdatForCounter(const Function &F, const Module &M) {
958 Triple TT(M.getTargetTriple());
959 if (!TT.isOSBinFormatELF() && !TT.isOSBinFormatWasm())
962 // See createPGOFuncNameVar for more details. To avoid link errors, profile
963 // counters for function with available_externally linkage needs to be changed
964 // to linkonce linkage. On ELF based systems, this leads to weak symbols to be
965 // created. Without using comdat, duplicate entries won't be removed by the
966 // linker leading to increased data segement size and raw profile size. Even
967 // worse, since the referenced counter from profile per-function data object
968 // will be resolved to the common strong definition, the profile counts for
969 // available_externally functions will end up being duplicated in raw profile
970 // data. This can result in distorted profile as the counts of those dups
971 // will be accumulated by the profile merger.
972 GlobalValue::LinkageTypes Linkage = F.getLinkage();
973 if (Linkage != GlobalValue::ExternalWeakLinkage &&
974 Linkage != GlobalValue::AvailableExternallyLinkage)
980 // Check if INSTR_PROF_RAW_VERSION_VAR is defined.
981 bool isIRPGOFlagSet(const Module *M) {
983 M->getNamedGlobal(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
984 if (!IRInstrVar || IRInstrVar->isDeclaration() ||
985 IRInstrVar->hasLocalLinkage())
988 // Check if the flag is set.
989 if (!IRInstrVar->hasInitializer())
992 const Constant *InitVal = IRInstrVar->getInitializer();
996 return (dyn_cast<ConstantInt>(InitVal)->getZExtValue() &
997 VARIANT_MASK_IR_PROF) != 0;
1000 // Check if we can safely rename this Comdat function.
1001 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken) {
1002 if (F.getName().empty())
1004 if (!needsComdatForCounter(F, *(F.getParent())))
1006 // Unsafe to rename the address-taken function (which can be used in
1007 // function comparison).
1008 if (CheckAddressTaken && F.hasAddressTaken())
1010 // Only safe to do if this function may be discarded if it is not used
1011 // in the compilation unit.
1012 if (!GlobalValue::isDiscardableIfUnused(F.getLinkage()))
1015 // For AvailableExternallyLinkage functions.
1016 if (!F.hasComdat()) {
1017 assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage);
1023 // Parse the value profile options.
1024 void getMemOPSizeRangeFromOption(std::string MemOPSizeRange,
1025 int64_t &RangeStart, int64_t &RangeLast) {
1026 static const int64_t DefaultMemOPSizeRangeStart = 0;
1027 static const int64_t DefaultMemOPSizeRangeLast = 8;
1028 RangeStart = DefaultMemOPSizeRangeStart;
1029 RangeLast = DefaultMemOPSizeRangeLast;
1031 if (!MemOPSizeRange.empty()) {
1032 auto Pos = MemOPSizeRange.find(":");
1033 if (Pos != std::string::npos) {
1035 RangeStart = atoi(MemOPSizeRange.substr(0, Pos).c_str());
1036 if (Pos < MemOPSizeRange.size() - 1)
1037 RangeLast = atoi(MemOPSizeRange.substr(Pos + 1).c_str());
1039 RangeLast = atoi(MemOPSizeRange.c_str());
1041 assert(RangeLast >= RangeStart);
1044 } // end namespace llvm