1 //===- InstrProf.h - Instrumented profiling format support ------*- C++ -*-===//
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 // Instrumentation-based profiling data is generated by instrumented
11 // binaries through library functions in compiler-rt, and read by the clang
12 // frontend to feed PGO.
14 //===----------------------------------------------------------------------===//
16 #ifndef LLVM_PROFILEDATA_INSTRPROF_H
17 #define LLVM_PROFILEDATA_INSTRPROF_H
19 #include "llvm/ADT/ArrayRef.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/StringRef.h"
22 #include "llvm/ADT/StringSet.h"
23 #include "llvm/ADT/Triple.h"
24 #include "llvm/IR/GlobalValue.h"
25 #include "llvm/IR/ProfileSummary.h"
26 #include "llvm/ProfileData/InstrProfData.inc"
27 #include "llvm/Support/Compiler.h"
28 #include "llvm/Support/Endian.h"
29 #include "llvm/Support/Error.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/Host.h"
32 #include "llvm/Support/MD5.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
43 #include <system_error>
51 struct InstrProfRecord;
52 class InstrProfSymtab;
57 enum InstrProfSectKind {
58 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) Kind,
59 #include "llvm/ProfileData/InstrProfData.inc"
62 /// Return the name of the profile section corresponding to \p IPSK.
64 /// The name of the section depends on the object format type \p OF. If
65 /// \p AddSegmentInfo is true, a segment prefix and additional linker hints may
66 /// be added to the section name (this is the default).
67 std::string getInstrProfSectionName(InstrProfSectKind IPSK,
68 Triple::ObjectFormatType OF,
69 bool AddSegmentInfo = true);
71 /// Return the name profile runtime entry point to do value profiling
73 inline StringRef getInstrProfValueProfFuncName() {
74 return INSTR_PROF_VALUE_PROF_FUNC_STR;
77 /// Return the name profile runtime entry point to do value range profiling.
78 inline StringRef getInstrProfValueRangeProfFuncName() {
79 return INSTR_PROF_VALUE_RANGE_PROF_FUNC_STR;
82 /// Return the name of the section containing function coverage mapping
84 std::string getInstrProfCoverageSectionName(const Module *M = nullptr);
85 /// Similar to the above, but used by host tool (e.g, coverage) which has
86 /// object format information. The section name returned is not prefixed
87 /// with segment name.
88 std::string getInstrProfCoverageSectionNameInObject(bool isCoff);
90 /// Return the name prefix of variables containing instrumented function names.
91 inline StringRef getInstrProfNameVarPrefix() { return "__profn_"; }
93 /// Return the name prefix of variables containing per-function control data.
94 inline StringRef getInstrProfDataVarPrefix() { return "__profd_"; }
96 /// Return the name prefix of profile counter variables.
97 inline StringRef getInstrProfCountersVarPrefix() { return "__profc_"; }
99 /// Return the name prefix of value profile variables.
100 inline StringRef getInstrProfValuesVarPrefix() { return "__profvp_"; }
102 /// Return the name of value profile node array variables:
103 inline StringRef getInstrProfVNodesVarName() { return "__llvm_prf_vnodes"; }
105 /// Return the name prefix of the COMDAT group for instrumentation variables
106 /// associated with a COMDAT function.
107 inline StringRef getInstrProfComdatPrefix() { return "__profv_"; }
109 /// Return the name of the variable holding the strings (possibly compressed)
110 /// of all function's PGO names.
111 inline StringRef getInstrProfNamesVarName() {
112 return "__llvm_prf_nm";
115 /// Return the name of a covarage mapping variable (internal linkage)
116 /// for each instrumented source module. Such variables are allocated
117 /// in the __llvm_covmap section.
118 inline StringRef getCoverageMappingVarName() {
119 return "__llvm_coverage_mapping";
122 /// Return the name of the internal variable recording the array
123 /// of PGO name vars referenced by the coverage mapping. The owning
124 /// functions of those names are not emitted by FE (e.g, unused inline
126 inline StringRef getCoverageUnusedNamesVarName() {
127 return "__llvm_coverage_names";
130 /// Return the name of function that registers all the per-function control
131 /// data at program startup time by calling __llvm_register_function. This
132 /// function has internal linkage and is called by __llvm_profile_init
133 /// runtime method. This function is not generated for these platforms:
134 /// Darwin, Linux, and FreeBSD.
135 inline StringRef getInstrProfRegFuncsName() {
136 return "__llvm_profile_register_functions";
139 /// Return the name of the runtime interface that registers per-function control
140 /// data for one instrumented function.
141 inline StringRef getInstrProfRegFuncName() {
142 return "__llvm_profile_register_function";
145 /// Return the name of the runtime interface that registers the PGO name strings.
146 inline StringRef getInstrProfNamesRegFuncName() {
147 return "__llvm_profile_register_names_function";
150 /// Return the name of the runtime initialization method that is generated by
151 /// the compiler. The function calls __llvm_profile_register_functions and
152 /// __llvm_profile_override_default_filename functions if needed. This function
153 /// has internal linkage and invoked at startup time via init_array.
154 inline StringRef getInstrProfInitFuncName() { return "__llvm_profile_init"; }
156 /// Return the name of the hook variable defined in profile runtime library.
157 /// A reference to the variable causes the linker to link in the runtime
158 /// initialization module (which defines the hook variable).
159 inline StringRef getInstrProfRuntimeHookVarName() {
160 return INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_RUNTIME_VAR);
163 /// Return the name of the compiler generated function that references the
164 /// runtime hook variable. The function is a weak global.
165 inline StringRef getInstrProfRuntimeHookVarUseFuncName() {
166 return "__llvm_profile_runtime_user";
169 /// Return the marker used to separate PGO names during serialization.
170 inline StringRef getInstrProfNameSeparator() { return "\01"; }
172 /// Return the modified name for function \c F suitable to be
173 /// used the key for profile lookup. Variable \c InLTO indicates if this
174 /// is called in LTO optimization passes.
175 std::string getPGOFuncName(const Function &F, bool InLTO = false,
176 uint64_t Version = INSTR_PROF_INDEX_VERSION);
178 /// Return the modified name for a function suitable to be
179 /// used the key for profile lookup. The function's original
180 /// name is \c RawFuncName and has linkage of type \c Linkage.
181 /// The function is defined in module \c FileName.
182 std::string getPGOFuncName(StringRef RawFuncName,
183 GlobalValue::LinkageTypes Linkage,
185 uint64_t Version = INSTR_PROF_INDEX_VERSION);
187 /// Return the name of the global variable used to store a function
188 /// name in PGO instrumentation. \c FuncName is the name of the function
189 /// returned by the \c getPGOFuncName call.
190 std::string getPGOFuncNameVarName(StringRef FuncName,
191 GlobalValue::LinkageTypes Linkage);
193 /// Create and return the global variable for function name used in PGO
194 /// instrumentation. \c FuncName is the name of the function returned
195 /// by \c getPGOFuncName call.
196 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName);
198 /// Create and return the global variable for function name used in PGO
199 /// instrumentation. /// \c FuncName is the name of the function
200 /// returned by \c getPGOFuncName call, \c M is the owning module,
201 /// and \c Linkage is the linkage of the instrumented function.
202 GlobalVariable *createPGOFuncNameVar(Module &M,
203 GlobalValue::LinkageTypes Linkage,
204 StringRef PGOFuncName);
206 /// Return the initializer in string of the PGO name var \c NameVar.
207 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar);
209 /// Given a PGO function name, remove the filename prefix and return
210 /// the original (static) function name.
211 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName,
212 StringRef FileName = "<unknown>");
214 /// Given a vector of strings (function PGO names) \c NameStrs, the
215 /// method generates a combined string \c Result thatis ready to be
216 /// serialized. The \c Result string is comprised of three fields:
217 /// The first field is the legnth of the uncompressed strings, and the
218 /// the second field is the length of the zlib-compressed string.
219 /// Both fields are encoded in ULEB128. If \c doCompress is false, the
220 /// third field is the uncompressed strings; otherwise it is the
221 /// compressed string. When the string compression is off, the
222 /// second field will have value zero.
223 Error collectPGOFuncNameStrings(const std::vector<std::string> &NameStrs,
224 bool doCompression, std::string &Result);
226 /// Produce \c Result string with the same format described above. The input
227 /// is vector of PGO function name variables that are referenced.
228 Error collectPGOFuncNameStrings(const std::vector<GlobalVariable *> &NameVars,
229 std::string &Result, bool doCompression = true);
231 /// \c NameStrings is a string composed of one of more sub-strings encoded in
232 /// the format described above. The substrings are separated by 0 or more zero
233 /// bytes. This method decodes the string and populates the \c Symtab.
234 Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab);
236 /// Check if INSTR_PROF_RAW_VERSION_VAR is defined. This global is only being
237 /// set in IR PGO compilation.
238 bool isIRPGOFlagSet(const Module *M);
240 /// Check if we can safely rename this Comdat function. Instances of the same
241 /// comdat function may have different control flows thus can not share the
242 /// same counter variable.
243 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken = false);
245 enum InstrProfValueKind : uint32_t {
246 #define VALUE_PROF_KIND(Enumerator, Value) Enumerator = Value,
247 #include "llvm/ProfileData/InstrProfData.inc"
250 /// Get the value profile data for value site \p SiteIdx from \p InstrProfR
251 /// and annotate the instruction \p Inst with the value profile meta data.
252 /// Annotate up to \p MaxMDCount (default 3) number of records per value site.
253 void annotateValueSite(Module &M, Instruction &Inst,
254 const InstrProfRecord &InstrProfR,
255 InstrProfValueKind ValueKind, uint32_t SiteIndx,
256 uint32_t MaxMDCount = 3);
258 /// Same as the above interface but using an ArrayRef, as well as \p Sum.
259 void annotateValueSite(Module &M, Instruction &Inst,
260 ArrayRef<InstrProfValueData> VDs,
261 uint64_t Sum, InstrProfValueKind ValueKind,
262 uint32_t MaxMDCount);
264 /// Extract the value profile data from \p Inst which is annotated with
265 /// value profile meta data. Return false if there is no value data annotated,
266 /// otherwise return true.
267 bool getValueProfDataFromInst(const Instruction &Inst,
268 InstrProfValueKind ValueKind,
269 uint32_t MaxNumValueData,
270 InstrProfValueData ValueData[],
271 uint32_t &ActualNumValueData, uint64_t &TotalC);
273 inline StringRef getPGOFuncNameMetadataName() { return "PGOFuncName"; }
275 /// Return the PGOFuncName meta data associated with a function.
276 MDNode *getPGOFuncNameMetadata(const Function &F);
278 /// Create the PGOFuncName meta data if PGOFuncName is different from
279 /// function's raw name. This should only apply to internal linkage functions
280 /// declared by users only.
281 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName);
283 /// Check if we can use Comdat for profile variables. This will eliminate
284 /// the duplicated profile variables for Comdat functions.
285 bool needsComdatForCounter(const Function &F, const Module &M);
287 const std::error_category &instrprof_category();
289 enum class instrprof_error {
296 unsupported_hash_type,
304 value_site_count_mismatch,
310 inline std::error_code make_error_code(instrprof_error E) {
311 return std::error_code(static_cast<int>(E), instrprof_category());
314 class InstrProfError : public ErrorInfo<InstrProfError> {
316 InstrProfError(instrprof_error Err) : Err(Err) {
317 assert(Err != instrprof_error::success && "Not an error");
320 std::string message() const override;
322 void log(raw_ostream &OS) const override { OS << message(); }
324 std::error_code convertToErrorCode() const override {
325 return make_error_code(Err);
328 instrprof_error get() const { return Err; }
330 /// Consume an Error and return the raw enum value contained within it. The
331 /// Error must either be a success value, or contain a single InstrProfError.
332 static instrprof_error take(Error E) {
333 auto Err = instrprof_error::success;
334 handleAllErrors(std::move(E), [&Err](const InstrProfError &IPE) {
335 assert(Err == instrprof_error::success && "Multiple errors encountered");
347 class SoftInstrProfErrors {
348 /// Count the number of soft instrprof_errors encountered and keep track of
349 /// the first such error for reporting purposes.
351 /// The first soft error encountered.
352 instrprof_error FirstError = instrprof_error::success;
354 /// The number of hash mismatches.
355 unsigned NumHashMismatches = 0;
357 /// The number of count mismatches.
358 unsigned NumCountMismatches = 0;
360 /// The number of counter overflows.
361 unsigned NumCounterOverflows = 0;
363 /// The number of value site count mismatches.
364 unsigned NumValueSiteCountMismatches = 0;
367 SoftInstrProfErrors() = default;
369 ~SoftInstrProfErrors() {
370 assert(FirstError == instrprof_error::success &&
371 "Unchecked soft error encountered");
374 /// Track a soft error (\p IE) and increment its associated counter.
375 void addError(instrprof_error IE);
377 /// Get the number of hash mismatches.
378 unsigned getNumHashMismatches() const { return NumHashMismatches; }
380 /// Get the number of count mismatches.
381 unsigned getNumCountMismatches() const { return NumCountMismatches; }
383 /// Get the number of counter overflows.
384 unsigned getNumCounterOverflows() const { return NumCounterOverflows; }
386 /// Get the number of value site count mismatches.
387 unsigned getNumValueSiteCountMismatches() const {
388 return NumValueSiteCountMismatches;
391 /// Return the first encountered error and reset FirstError to a success
394 if (FirstError == instrprof_error::success)
395 return Error::success();
396 auto E = make_error<InstrProfError>(FirstError);
397 FirstError = instrprof_error::success;
406 } // end namespace object
408 namespace IndexedInstrProf {
410 uint64_t ComputeHash(StringRef K);
412 } // end namespace IndexedInstrProf
414 /// A symbol table used for function PGO name look-up with keys
415 /// (such as pointers, md5hash values) to the function. A function's
416 /// PGO name or name's md5hash are used in retrieving the profile
417 /// data of the function. See \c getPGOFuncName() method for details
418 /// on how PGO name is formed.
419 class InstrProfSymtab {
421 typedef std::vector<std::pair<uint64_t, uint64_t>> AddrHashMap;
425 uint64_t Address = 0;
426 // Unique name strings.
428 // A map from MD5 keys to function name strings.
429 std::vector<std::pair<uint64_t, StringRef>> MD5NameMap;
430 // A map from MD5 keys to function define. We only populate this map
431 // when build the Symtab from a Module.
432 std::vector<std::pair<uint64_t, Function *>> MD5FuncMap;
433 // A map from function runtime address to function name MD5 hash.
434 // This map is only populated and used by raw instr profile reader.
435 AddrHashMap AddrToMD5Map;
438 InstrProfSymtab() = default;
440 /// Create InstrProfSymtab from an object file section which
441 /// contains function PGO names. When section may contain raw
442 /// string data or string data in compressed form. This method
443 /// only initialize the symtab with reference to the data and
444 /// the section base address. The decompression will be delayed
445 /// until before it is used. See also \c create(StringRef) method.
446 Error create(object::SectionRef &Section);
448 /// This interface is used by reader of CoverageMapping test
450 inline Error create(StringRef D, uint64_t BaseAddr);
452 /// \c NameStrings is a string composed of one of more sub-strings
453 /// encoded in the format described in \c collectPGOFuncNameStrings.
454 /// This method is a wrapper to \c readPGOFuncNameStrings method.
455 inline Error create(StringRef NameStrings);
457 /// A wrapper interface to populate the PGO symtab with functions
458 /// decls from module \c M. This interface is used by transformation
459 /// passes such as indirect function call promotion. Variable \c InLTO
460 /// indicates if this is called from LTO optimization passes.
461 void create(Module &M, bool InLTO = false);
463 /// Create InstrProfSymtab from a set of names iteratable from
464 /// \p IterRange. This interface is used by IndexedProfReader.
465 template <typename NameIterRange> void create(const NameIterRange &IterRange);
467 // If the symtab is created by a series of calls to \c addFuncName, \c
468 // finalizeSymtab needs to be called before looking up function names.
469 // This is required because the underlying map is a vector (for space
470 // efficiency) which needs to be sorted.
471 inline void finalizeSymtab();
473 /// Update the symtab by adding \p FuncName to the table. This interface
474 /// is used by the raw and text profile readers.
475 void addFuncName(StringRef FuncName) {
476 auto Ins = NameTab.insert(FuncName);
478 MD5NameMap.push_back(std::make_pair(
479 IndexedInstrProf::ComputeHash(FuncName), Ins.first->getKey()));
482 /// Map a function address to its name's MD5 hash. This interface
483 /// is only used by the raw profiler reader.
484 void mapAddress(uint64_t Addr, uint64_t MD5Val) {
485 AddrToMD5Map.push_back(std::make_pair(Addr, MD5Val));
488 AddrHashMap &getAddrHashMap() { return AddrToMD5Map; }
490 /// Return function's PGO name from the function name's symbol
491 /// address in the object file. If an error occurs, return
493 StringRef getFuncName(uint64_t FuncNameAddress, size_t NameSize);
495 /// Return function's PGO name from the name's md5 hash value.
496 /// If not found, return an empty string.
497 inline StringRef getFuncName(uint64_t FuncMD5Hash);
499 /// Return function from the name's md5 hash. Return nullptr if not found.
500 inline Function *getFunction(uint64_t FuncMD5Hash);
502 /// Return the function's original assembly name by stripping off
503 /// the prefix attached (to symbols with priviate linkage). For
504 /// global functions, it returns the same string as getFuncName.
505 inline StringRef getOrigFuncName(uint64_t FuncMD5Hash);
507 /// Return the name section data.
508 inline StringRef getNameData() const { return Data; }
511 Error InstrProfSymtab::create(StringRef D, uint64_t BaseAddr) {
514 return Error::success();
517 Error InstrProfSymtab::create(StringRef NameStrings) {
518 return readPGOFuncNameStrings(NameStrings, *this);
521 template <typename NameIterRange>
522 void InstrProfSymtab::create(const NameIterRange &IterRange) {
523 for (auto Name : IterRange)
529 void InstrProfSymtab::finalizeSymtab() {
530 std::sort(MD5NameMap.begin(), MD5NameMap.end(), less_first());
531 std::sort(MD5FuncMap.begin(), MD5FuncMap.end(), less_first());
532 std::sort(AddrToMD5Map.begin(), AddrToMD5Map.end(), less_first());
533 AddrToMD5Map.erase(std::unique(AddrToMD5Map.begin(), AddrToMD5Map.end()),
537 StringRef InstrProfSymtab::getFuncName(uint64_t FuncMD5Hash) {
539 std::lower_bound(MD5NameMap.begin(), MD5NameMap.end(), FuncMD5Hash,
540 [](const std::pair<uint64_t, std::string> &LHS,
541 uint64_t RHS) { return LHS.first < RHS; });
542 if (Result != MD5NameMap.end() && Result->first == FuncMD5Hash)
543 return Result->second;
547 Function* InstrProfSymtab::getFunction(uint64_t FuncMD5Hash) {
549 std::lower_bound(MD5FuncMap.begin(), MD5FuncMap.end(), FuncMD5Hash,
550 [](const std::pair<uint64_t, Function*> &LHS,
551 uint64_t RHS) { return LHS.first < RHS; });
552 if (Result != MD5FuncMap.end() && Result->first == FuncMD5Hash)
553 return Result->second;
557 // See also getPGOFuncName implementation. These two need to be
559 StringRef InstrProfSymtab::getOrigFuncName(uint64_t FuncMD5Hash) {
560 StringRef PGOName = getFuncName(FuncMD5Hash);
561 size_t S = PGOName.find_first_of(':');
562 if (S == StringRef::npos)
564 return PGOName.drop_front(S + 1);
567 struct InstrProfValueSiteRecord {
568 /// Value profiling data pairs at a given value site.
569 std::list<InstrProfValueData> ValueData;
571 InstrProfValueSiteRecord() { ValueData.clear(); }
572 template <class InputIterator>
573 InstrProfValueSiteRecord(InputIterator F, InputIterator L)
576 /// Sort ValueData ascending by Value
577 void sortByTargetValues() {
579 [](const InstrProfValueData &left, const InstrProfValueData &right) {
580 return left.Value < right.Value;
583 /// Sort ValueData Descending by Count
584 inline void sortByCount();
586 /// Merge data from another InstrProfValueSiteRecord
587 /// Optionally scale merged counts by \p Weight.
588 void merge(SoftInstrProfErrors &SIPE, InstrProfValueSiteRecord &Input,
589 uint64_t Weight = 1);
590 /// Scale up value profile data counts.
591 void scale(SoftInstrProfErrors &SIPE, uint64_t Weight);
594 /// Profiling information for a single function.
595 struct InstrProfRecord {
598 std::vector<uint64_t> Counts;
599 SoftInstrProfErrors SIPE;
601 InstrProfRecord() = default;
602 InstrProfRecord(StringRef Name, uint64_t Hash, std::vector<uint64_t> Counts)
603 : Name(Name), Hash(Hash), Counts(std::move(Counts)) {}
605 typedef std::vector<std::pair<uint64_t, uint64_t>> ValueMapType;
607 /// Return the number of value profile kinds with non-zero number
608 /// of profile sites.
609 inline uint32_t getNumValueKinds() const;
611 /// Return the number of instrumented sites for ValueKind.
612 inline uint32_t getNumValueSites(uint32_t ValueKind) const;
614 /// Return the total number of ValueData for ValueKind.
615 inline uint32_t getNumValueData(uint32_t ValueKind) const;
617 /// Return the number of value data collected for ValueKind at profiling
619 inline uint32_t getNumValueDataForSite(uint32_t ValueKind,
620 uint32_t Site) const;
622 /// Return the array of profiled values at \p Site. If \p TotalC
623 /// is not null, the total count of all target values at this site
624 /// will be stored in \c *TotalC.
625 inline std::unique_ptr<InstrProfValueData[]>
626 getValueForSite(uint32_t ValueKind, uint32_t Site,
627 uint64_t *TotalC = nullptr) const;
629 /// Get the target value/counts of kind \p ValueKind collected at site
630 /// \p Site and store the result in array \p Dest. Return the total
631 /// counts of all target values at this site.
632 inline uint64_t getValueForSite(InstrProfValueData Dest[], uint32_t ValueKind,
633 uint32_t Site) const;
635 /// Reserve space for NumValueSites sites.
636 inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites);
638 /// Add ValueData for ValueKind at value Site.
639 void addValueData(uint32_t ValueKind, uint32_t Site,
640 InstrProfValueData *VData, uint32_t N,
641 ValueMapType *ValueMap);
643 /// Merge the counts in \p Other into this one.
644 /// Optionally scale merged counts by \p Weight.
645 void merge(InstrProfRecord &Other, uint64_t Weight = 1);
647 /// Scale up profile counts (including value profile data) by
649 void scale(uint64_t Weight);
651 /// Sort value profile data (per site) by count.
652 void sortValueData() {
653 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
654 std::vector<InstrProfValueSiteRecord> &SiteRecords =
655 getValueSitesForKind(Kind);
656 for (auto &SR : SiteRecords)
661 /// Clear value data entries and edge counters.
667 /// Clear value data entries
668 void clearValueData() {
669 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
670 getValueSitesForKind(Kind).clear();
673 /// Get the error contained within the record's soft error counter.
674 Error takeError() { return SIPE.takeError(); }
677 std::vector<InstrProfValueSiteRecord> IndirectCallSites;
678 std::vector<InstrProfValueSiteRecord> MemOPSizes;
679 const std::vector<InstrProfValueSiteRecord> &
681 getValueSitesForKind(uint32_t ValueKind) const {
683 case IPVK_IndirectCallTarget:
684 return IndirectCallSites;
688 llvm_unreachable("Unknown value kind!");
690 return IndirectCallSites;
693 std::vector<InstrProfValueSiteRecord> &
694 getValueSitesForKind(uint32_t ValueKind) {
695 return const_cast<std::vector<InstrProfValueSiteRecord> &>(
696 const_cast<const InstrProfRecord *>(this)
697 ->getValueSitesForKind(ValueKind));
700 // Map indirect call target name hash to name string.
701 uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
702 ValueMapType *HashKeys);
704 // Merge Value Profile data from Src record to this record for ValueKind.
705 // Scale merged value counts by \p Weight.
706 void mergeValueProfData(uint32_t ValueKind, InstrProfRecord &Src,
709 // Scale up value profile data count.
710 void scaleValueProfData(uint32_t ValueKind, uint64_t Weight);
713 uint32_t InstrProfRecord::getNumValueKinds() const {
714 uint32_t NumValueKinds = 0;
715 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
716 NumValueKinds += !(getValueSitesForKind(Kind).empty());
717 return NumValueKinds;
720 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
722 const std::vector<InstrProfValueSiteRecord> &SiteRecords =
723 getValueSitesForKind(ValueKind);
724 for (auto &SR : SiteRecords) {
725 N += SR.ValueData.size();
730 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
731 return getValueSitesForKind(ValueKind).size();
734 uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind,
735 uint32_t Site) const {
736 return getValueSitesForKind(ValueKind)[Site].ValueData.size();
739 std::unique_ptr<InstrProfValueData[]>
740 InstrProfRecord::getValueForSite(uint32_t ValueKind, uint32_t Site,
741 uint64_t *TotalC) const {
743 uint64_t &TotalCount = (TotalC == nullptr ? Dummy : *TotalC);
744 uint32_t N = getNumValueDataForSite(ValueKind, Site);
747 return std::unique_ptr<InstrProfValueData[]>(nullptr);
750 auto VD = llvm::make_unique<InstrProfValueData[]>(N);
751 TotalCount = getValueForSite(VD.get(), ValueKind, Site);
756 uint64_t InstrProfRecord::getValueForSite(InstrProfValueData Dest[],
758 uint32_t Site) const {
760 uint64_t TotalCount = 0;
761 for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) {
762 Dest[I].Value = V.Value;
763 Dest[I].Count = V.Count;
764 TotalCount = SaturatingAdd(TotalCount, V.Count);
770 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
771 std::vector<InstrProfValueSiteRecord> &ValueSites =
772 getValueSitesForKind(ValueKind);
773 ValueSites.reserve(NumValueSites);
776 inline support::endianness getHostEndianness() {
777 return sys::IsLittleEndianHost ? support::little : support::big;
780 // Include definitions for value profile data
781 #define INSTR_PROF_VALUE_PROF_DATA
782 #include "llvm/ProfileData/InstrProfData.inc"
784 void InstrProfValueSiteRecord::sortByCount() {
786 [](const InstrProfValueData &left, const InstrProfValueData &right) {
787 return left.Count > right.Count;
790 size_t max_s = INSTR_PROF_MAX_NUM_VAL_PER_SITE;
791 if (ValueData.size() > max_s)
792 ValueData.resize(max_s);
795 namespace IndexedInstrProf {
797 enum class HashT : uint32_t {
802 inline uint64_t ComputeHash(HashT Type, StringRef K) {
807 llvm_unreachable("Unhandled hash type");
810 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
813 // Version 1 is the first version. In this version, the value of
814 // a key/value pair can only include profile data of a single function.
815 // Due to this restriction, the number of block counters for a given
816 // function is not recorded but derived from the length of the value.
818 // The version 2 format supports recording profile data of multiple
819 // functions which share the same key in one value field. To support this,
820 // the number block counters is recorded as an uint64_t field right after the
821 // function structural hash.
823 // Version 3 supports value profile data. The value profile data is expected
824 // to follow the block counter profile data.
826 // In this version, profile summary data \c IndexedInstrProf::Summary is
827 // stored after the profile header.
829 // The current version is 4.
830 CurrentVersion = INSTR_PROF_INDEX_VERSION
832 const uint64_t Version = ProfVersion::CurrentVersion;
834 const HashT HashType = HashT::MD5;
836 inline uint64_t ComputeHash(StringRef K) { return ComputeHash(HashType, K); }
838 // This structure defines the file header of the LLVM profile
839 // data file in indexed-format.
843 uint64_t Unused; // Becomes unused since version 4
848 // Profile summary data recorded in the profile data file in indexed
849 // format. It is introduced in version 4. The summary data follows
850 // right after the profile file header.
853 uint64_t Cutoff; ///< The required percentile of total execution count.
855 MinBlockCount; ///< The minimum execution count for this percentile.
856 uint64_t NumBlocks; ///< Number of blocks >= the minumum execution count.
858 // The field kind enumerator to assigned value mapping should remain
859 // unchanged when a new kind is added or an old kind gets deleted in
861 enum SummaryFieldKind {
862 /// The total number of functions instrumented.
863 TotalNumFunctions = 0,
864 /// Total number of instrumented blocks/edges.
866 /// The maximal execution count among all functions.
867 /// This field does not exist for profile data from IR based
869 MaxFunctionCount = 2,
870 /// Max block count of the program.
872 /// Max internal block count of the program (excluding entry blocks).
873 MaxInternalBlockCount = 4,
874 /// The sum of all instrumented block counts.
876 NumKinds = TotalBlockCount + 1
879 // The number of summmary fields following the summary header.
880 uint64_t NumSummaryFields;
881 // The number of Cutoff Entries (Summary::Entry) following summary fields.
882 uint64_t NumCutoffEntries;
884 static uint32_t getSize(uint32_t NumSumFields, uint32_t NumCutoffEntries) {
885 return sizeof(Summary) + NumCutoffEntries * sizeof(Entry) +
886 NumSumFields * sizeof(uint64_t);
889 const uint64_t *getSummaryDataBase() const {
890 return reinterpret_cast<const uint64_t *>(this + 1);
893 uint64_t *getSummaryDataBase() {
894 return reinterpret_cast<uint64_t *>(this + 1);
897 const Entry *getCutoffEntryBase() const {
898 return reinterpret_cast<const Entry *>(
899 &getSummaryDataBase()[NumSummaryFields]);
902 Entry *getCutoffEntryBase() {
903 return reinterpret_cast<Entry *>(&getSummaryDataBase()[NumSummaryFields]);
906 uint64_t get(SummaryFieldKind K) const {
907 return getSummaryDataBase()[K];
910 void set(SummaryFieldKind K, uint64_t V) {
911 getSummaryDataBase()[K] = V;
914 const Entry &getEntry(uint32_t I) const { return getCutoffEntryBase()[I]; }
916 void setEntry(uint32_t I, const ProfileSummaryEntry &E) {
917 Entry &ER = getCutoffEntryBase()[I];
918 ER.Cutoff = E.Cutoff;
919 ER.MinBlockCount = E.MinCount;
920 ER.NumBlocks = E.NumCounts;
923 Summary(uint32_t Size) { memset(this, 0, Size); }
924 void operator delete(void *ptr) { ::operator delete(ptr); }
929 inline std::unique_ptr<Summary> allocSummary(uint32_t TotalSize) {
930 return std::unique_ptr<Summary>(new (::operator new(TotalSize))
934 } // end namespace IndexedInstrProf
936 namespace RawInstrProf {
938 // Version 1: First version
939 // Version 2: Added value profile data section. Per-function control data
940 // struct has more fields to describe value profile information.
941 // Version 3: Compressed name section support. Function PGO name reference
942 // from control data struct is changed from raw pointer to Name's MD5 value.
943 // Version 4: ValueDataBegin and ValueDataSizes fields are removed from the
945 const uint64_t Version = INSTR_PROF_RAW_VERSION;
947 template <class IntPtrT> inline uint64_t getMagic();
948 template <> inline uint64_t getMagic<uint64_t>() {
949 return INSTR_PROF_RAW_MAGIC_64;
952 template <> inline uint64_t getMagic<uint32_t>() {
953 return INSTR_PROF_RAW_MAGIC_32;
956 // Per-function profile data header/control structure.
957 // The definition should match the structure defined in
958 // compiler-rt/lib/profile/InstrProfiling.h.
959 // It should also match the synthesized type in
960 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
961 template <class IntPtrT> struct LLVM_ALIGNAS(8) ProfileData {
962 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
963 #include "llvm/ProfileData/InstrProfData.inc"
966 // File header structure of the LLVM profile data in raw format.
967 // The definition should match the header referenced in
968 // compiler-rt/lib/profile/InstrProfilingFile.c and
969 // InstrProfilingBuffer.c.
971 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) const Type Name;
972 #include "llvm/ProfileData/InstrProfData.inc"
975 } // end namespace RawInstrProf
977 // Parse MemOP Size range option.
978 void getMemOPSizeRangeFromOption(std::string Str, int64_t &RangeStart,
981 } // end namespace llvm
983 #endif // LLVM_PROFILEDATA_INSTRPROF_H