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 prefix of variables containing instrumented function names.
83 inline StringRef getInstrProfNameVarPrefix() { return "__profn_"; }
85 /// Return the name prefix of variables containing per-function control data.
86 inline StringRef getInstrProfDataVarPrefix() { return "__profd_"; }
88 /// Return the name prefix of profile counter variables.
89 inline StringRef getInstrProfCountersVarPrefix() { return "__profc_"; }
91 /// Return the name prefix of value profile variables.
92 inline StringRef getInstrProfValuesVarPrefix() { return "__profvp_"; }
94 /// Return the name of value profile node array variables:
95 inline StringRef getInstrProfVNodesVarName() { return "__llvm_prf_vnodes"; }
97 /// Return the name prefix of the COMDAT group for instrumentation variables
98 /// associated with a COMDAT function.
99 inline StringRef getInstrProfComdatPrefix() { return "__profv_"; }
101 /// Return the name of the variable holding the strings (possibly compressed)
102 /// of all function's PGO names.
103 inline StringRef getInstrProfNamesVarName() {
104 return "__llvm_prf_nm";
107 /// Return the name of a covarage mapping variable (internal linkage)
108 /// for each instrumented source module. Such variables are allocated
109 /// in the __llvm_covmap section.
110 inline StringRef getCoverageMappingVarName() {
111 return "__llvm_coverage_mapping";
114 /// Return the name of the internal variable recording the array
115 /// of PGO name vars referenced by the coverage mapping. The owning
116 /// functions of those names are not emitted by FE (e.g, unused inline
118 inline StringRef getCoverageUnusedNamesVarName() {
119 return "__llvm_coverage_names";
122 /// Return the name of function that registers all the per-function control
123 /// data at program startup time by calling __llvm_register_function. This
124 /// function has internal linkage and is called by __llvm_profile_init
125 /// runtime method. This function is not generated for these platforms:
126 /// Darwin, Linux, and FreeBSD.
127 inline StringRef getInstrProfRegFuncsName() {
128 return "__llvm_profile_register_functions";
131 /// Return the name of the runtime interface that registers per-function control
132 /// data for one instrumented function.
133 inline StringRef getInstrProfRegFuncName() {
134 return "__llvm_profile_register_function";
137 /// Return the name of the runtime interface that registers the PGO name strings.
138 inline StringRef getInstrProfNamesRegFuncName() {
139 return "__llvm_profile_register_names_function";
142 /// Return the name of the runtime initialization method that is generated by
143 /// the compiler. The function calls __llvm_profile_register_functions and
144 /// __llvm_profile_override_default_filename functions if needed. This function
145 /// has internal linkage and invoked at startup time via init_array.
146 inline StringRef getInstrProfInitFuncName() { return "__llvm_profile_init"; }
148 /// Return the name of the hook variable defined in profile runtime library.
149 /// A reference to the variable causes the linker to link in the runtime
150 /// initialization module (which defines the hook variable).
151 inline StringRef getInstrProfRuntimeHookVarName() {
152 return INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_RUNTIME_VAR);
155 /// Return the name of the compiler generated function that references the
156 /// runtime hook variable. The function is a weak global.
157 inline StringRef getInstrProfRuntimeHookVarUseFuncName() {
158 return "__llvm_profile_runtime_user";
161 /// Return the marker used to separate PGO names during serialization.
162 inline StringRef getInstrProfNameSeparator() { return "\01"; }
164 /// Return the modified name for function \c F suitable to be
165 /// used the key for profile lookup. Variable \c InLTO indicates if this
166 /// is called in LTO optimization passes.
167 std::string getPGOFuncName(const Function &F, bool InLTO = false,
168 uint64_t Version = INSTR_PROF_INDEX_VERSION);
170 /// Return the modified name for a function suitable to be
171 /// used the key for profile lookup. The function's original
172 /// name is \c RawFuncName and has linkage of type \c Linkage.
173 /// The function is defined in module \c FileName.
174 std::string getPGOFuncName(StringRef RawFuncName,
175 GlobalValue::LinkageTypes Linkage,
177 uint64_t Version = INSTR_PROF_INDEX_VERSION);
179 /// Return the name of the global variable used to store a function
180 /// name in PGO instrumentation. \c FuncName is the name of the function
181 /// returned by the \c getPGOFuncName call.
182 std::string getPGOFuncNameVarName(StringRef FuncName,
183 GlobalValue::LinkageTypes Linkage);
185 /// Create and return the global variable for function name used in PGO
186 /// instrumentation. \c FuncName is the name of the function returned
187 /// by \c getPGOFuncName call.
188 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName);
190 /// Create and return the global variable for function name used in PGO
191 /// instrumentation. /// \c FuncName is the name of the function
192 /// returned by \c getPGOFuncName call, \c M is the owning module,
193 /// and \c Linkage is the linkage of the instrumented function.
194 GlobalVariable *createPGOFuncNameVar(Module &M,
195 GlobalValue::LinkageTypes Linkage,
196 StringRef PGOFuncName);
198 /// Return the initializer in string of the PGO name var \c NameVar.
199 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar);
201 /// Given a PGO function name, remove the filename prefix and return
202 /// the original (static) function name.
203 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName,
204 StringRef FileName = "<unknown>");
206 /// Given a vector of strings (function PGO names) \c NameStrs, the
207 /// method generates a combined string \c Result thatis ready to be
208 /// serialized. The \c Result string is comprised of three fields:
209 /// The first field is the legnth of the uncompressed strings, and the
210 /// the second field is the length of the zlib-compressed string.
211 /// Both fields are encoded in ULEB128. If \c doCompress is false, the
212 /// third field is the uncompressed strings; otherwise it is the
213 /// compressed string. When the string compression is off, the
214 /// second field will have value zero.
215 Error collectPGOFuncNameStrings(ArrayRef<std::string> NameStrs,
216 bool doCompression, std::string &Result);
218 /// Produce \c Result string with the same format described above. The input
219 /// is vector of PGO function name variables that are referenced.
220 Error collectPGOFuncNameStrings(ArrayRef<GlobalVariable *> NameVars,
221 std::string &Result, bool doCompression = true);
223 /// \c NameStrings is a string composed of one of more sub-strings encoded in
224 /// the format described above. The substrings are separated by 0 or more zero
225 /// bytes. This method decodes the string and populates the \c Symtab.
226 Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab);
228 /// Check if INSTR_PROF_RAW_VERSION_VAR is defined. This global is only being
229 /// set in IR PGO compilation.
230 bool isIRPGOFlagSet(const Module *M);
232 /// Check if we can safely rename this Comdat function. Instances of the same
233 /// comdat function may have different control flows thus can not share the
234 /// same counter variable.
235 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken = false);
237 enum InstrProfValueKind : uint32_t {
238 #define VALUE_PROF_KIND(Enumerator, Value) Enumerator = Value,
239 #include "llvm/ProfileData/InstrProfData.inc"
242 /// Get the value profile data for value site \p SiteIdx from \p InstrProfR
243 /// and annotate the instruction \p Inst with the value profile meta data.
244 /// Annotate up to \p MaxMDCount (default 3) number of records per value site.
245 void annotateValueSite(Module &M, Instruction &Inst,
246 const InstrProfRecord &InstrProfR,
247 InstrProfValueKind ValueKind, uint32_t SiteIndx,
248 uint32_t MaxMDCount = 3);
250 /// Same as the above interface but using an ArrayRef, as well as \p Sum.
251 void annotateValueSite(Module &M, Instruction &Inst,
252 ArrayRef<InstrProfValueData> VDs,
253 uint64_t Sum, InstrProfValueKind ValueKind,
254 uint32_t MaxMDCount);
256 /// Extract the value profile data from \p Inst which is annotated with
257 /// value profile meta data. Return false if there is no value data annotated,
258 /// otherwise return true.
259 bool getValueProfDataFromInst(const Instruction &Inst,
260 InstrProfValueKind ValueKind,
261 uint32_t MaxNumValueData,
262 InstrProfValueData ValueData[],
263 uint32_t &ActualNumValueData, uint64_t &TotalC);
265 inline StringRef getPGOFuncNameMetadataName() { return "PGOFuncName"; }
267 /// Return the PGOFuncName meta data associated with a function.
268 MDNode *getPGOFuncNameMetadata(const Function &F);
270 /// Create the PGOFuncName meta data if PGOFuncName is different from
271 /// function's raw name. This should only apply to internal linkage functions
272 /// declared by users only.
273 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName);
275 /// Check if we can use Comdat for profile variables. This will eliminate
276 /// the duplicated profile variables for Comdat functions.
277 bool needsComdatForCounter(const Function &F, const Module &M);
279 const std::error_category &instrprof_category();
281 enum class instrprof_error {
288 unsupported_hash_type,
296 value_site_count_mismatch,
302 inline std::error_code make_error_code(instrprof_error E) {
303 return std::error_code(static_cast<int>(E), instrprof_category());
306 class InstrProfError : public ErrorInfo<InstrProfError> {
308 InstrProfError(instrprof_error Err) : Err(Err) {
309 assert(Err != instrprof_error::success && "Not an error");
312 std::string message() const override;
314 void log(raw_ostream &OS) const override { OS << message(); }
316 std::error_code convertToErrorCode() const override {
317 return make_error_code(Err);
320 instrprof_error get() const { return Err; }
322 /// Consume an Error and return the raw enum value contained within it. The
323 /// Error must either be a success value, or contain a single InstrProfError.
324 static instrprof_error take(Error E) {
325 auto Err = instrprof_error::success;
326 handleAllErrors(std::move(E), [&Err](const InstrProfError &IPE) {
327 assert(Err == instrprof_error::success && "Multiple errors encountered");
339 class SoftInstrProfErrors {
340 /// Count the number of soft instrprof_errors encountered and keep track of
341 /// the first such error for reporting purposes.
343 /// The first soft error encountered.
344 instrprof_error FirstError = instrprof_error::success;
346 /// The number of hash mismatches.
347 unsigned NumHashMismatches = 0;
349 /// The number of count mismatches.
350 unsigned NumCountMismatches = 0;
352 /// The number of counter overflows.
353 unsigned NumCounterOverflows = 0;
355 /// The number of value site count mismatches.
356 unsigned NumValueSiteCountMismatches = 0;
359 SoftInstrProfErrors() = default;
361 ~SoftInstrProfErrors() {
362 assert(FirstError == instrprof_error::success &&
363 "Unchecked soft error encountered");
366 /// Track a soft error (\p IE) and increment its associated counter.
367 void addError(instrprof_error IE);
369 /// Get the number of hash mismatches.
370 unsigned getNumHashMismatches() const { return NumHashMismatches; }
372 /// Get the number of count mismatches.
373 unsigned getNumCountMismatches() const { return NumCountMismatches; }
375 /// Get the number of counter overflows.
376 unsigned getNumCounterOverflows() const { return NumCounterOverflows; }
378 /// Get the number of value site count mismatches.
379 unsigned getNumValueSiteCountMismatches() const {
380 return NumValueSiteCountMismatches;
383 /// Return the first encountered error and reset FirstError to a success
386 if (FirstError == instrprof_error::success)
387 return Error::success();
388 auto E = make_error<InstrProfError>(FirstError);
389 FirstError = instrprof_error::success;
398 } // end namespace object
400 namespace IndexedInstrProf {
402 uint64_t ComputeHash(StringRef K);
404 } // end namespace IndexedInstrProf
406 /// A symbol table used for function PGO name look-up with keys
407 /// (such as pointers, md5hash values) to the function. A function's
408 /// PGO name or name's md5hash are used in retrieving the profile
409 /// data of the function. See \c getPGOFuncName() method for details
410 /// on how PGO name is formed.
411 class InstrProfSymtab {
413 using AddrHashMap = std::vector<std::pair<uint64_t, uint64_t>>;
417 uint64_t Address = 0;
418 // Unique name strings.
420 // A map from MD5 keys to function name strings.
421 std::vector<std::pair<uint64_t, StringRef>> MD5NameMap;
422 // A map from MD5 keys to function define. We only populate this map
423 // when build the Symtab from a Module.
424 std::vector<std::pair<uint64_t, Function *>> MD5FuncMap;
425 // A map from function runtime address to function name MD5 hash.
426 // This map is only populated and used by raw instr profile reader.
427 AddrHashMap AddrToMD5Map;
430 InstrProfSymtab() = default;
432 /// Create InstrProfSymtab from an object file section which
433 /// contains function PGO names. When section may contain raw
434 /// string data or string data in compressed form. This method
435 /// only initialize the symtab with reference to the data and
436 /// the section base address. The decompression will be delayed
437 /// until before it is used. See also \c create(StringRef) method.
438 Error create(object::SectionRef &Section);
440 /// This interface is used by reader of CoverageMapping test
442 inline Error create(StringRef D, uint64_t BaseAddr);
444 /// \c NameStrings is a string composed of one of more sub-strings
445 /// encoded in the format described in \c collectPGOFuncNameStrings.
446 /// This method is a wrapper to \c readPGOFuncNameStrings method.
447 inline Error create(StringRef NameStrings);
449 /// A wrapper interface to populate the PGO symtab with functions
450 /// decls from module \c M. This interface is used by transformation
451 /// passes such as indirect function call promotion. Variable \c InLTO
452 /// indicates if this is called from LTO optimization passes.
453 Error create(Module &M, bool InLTO = false);
455 /// Create InstrProfSymtab from a set of names iteratable from
456 /// \p IterRange. This interface is used by IndexedProfReader.
457 template <typename NameIterRange> Error create(const NameIterRange &IterRange);
459 // If the symtab is created by a series of calls to \c addFuncName, \c
460 // finalizeSymtab needs to be called before looking up function names.
461 // This is required because the underlying map is a vector (for space
462 // efficiency) which needs to be sorted.
463 inline void finalizeSymtab();
465 /// Update the symtab by adding \p FuncName to the table. This interface
466 /// is used by the raw and text profile readers.
467 Error addFuncName(StringRef FuncName) {
468 if (FuncName.empty())
469 return make_error<InstrProfError>(instrprof_error::malformed);
470 auto Ins = NameTab.insert(FuncName);
472 MD5NameMap.push_back(std::make_pair(
473 IndexedInstrProf::ComputeHash(FuncName), Ins.first->getKey()));
474 return Error::success();
477 /// Map a function address to its name's MD5 hash. This interface
478 /// is only used by the raw profiler reader.
479 void mapAddress(uint64_t Addr, uint64_t MD5Val) {
480 AddrToMD5Map.push_back(std::make_pair(Addr, MD5Val));
483 AddrHashMap &getAddrHashMap() { return AddrToMD5Map; }
485 /// Return function's PGO name from the function name's symbol
486 /// address in the object file. If an error occurs, return
488 StringRef getFuncName(uint64_t FuncNameAddress, size_t NameSize);
490 /// Return function's PGO name from the name's md5 hash value.
491 /// If not found, return an empty string.
492 inline StringRef getFuncName(uint64_t FuncMD5Hash);
494 /// Return function from the name's md5 hash. Return nullptr if not found.
495 inline Function *getFunction(uint64_t FuncMD5Hash);
497 /// Return the function's original assembly name by stripping off
498 /// the prefix attached (to symbols with priviate linkage). For
499 /// global functions, it returns the same string as getFuncName.
500 inline StringRef getOrigFuncName(uint64_t FuncMD5Hash);
502 /// Return the name section data.
503 inline StringRef getNameData() const { return Data; }
506 Error InstrProfSymtab::create(StringRef D, uint64_t BaseAddr) {
509 return Error::success();
512 Error InstrProfSymtab::create(StringRef NameStrings) {
513 return readPGOFuncNameStrings(NameStrings, *this);
516 template <typename NameIterRange>
517 Error InstrProfSymtab::create(const NameIterRange &IterRange) {
518 for (auto Name : IterRange)
519 if (Error E = addFuncName(Name))
523 return Error::success();
526 void InstrProfSymtab::finalizeSymtab() {
527 std::sort(MD5NameMap.begin(), MD5NameMap.end(), less_first());
528 std::sort(MD5FuncMap.begin(), MD5FuncMap.end(), less_first());
529 std::sort(AddrToMD5Map.begin(), AddrToMD5Map.end(), less_first());
530 AddrToMD5Map.erase(std::unique(AddrToMD5Map.begin(), AddrToMD5Map.end()),
534 StringRef InstrProfSymtab::getFuncName(uint64_t FuncMD5Hash) {
536 std::lower_bound(MD5NameMap.begin(), MD5NameMap.end(), FuncMD5Hash,
537 [](const std::pair<uint64_t, std::string> &LHS,
538 uint64_t RHS) { return LHS.first < RHS; });
539 if (Result != MD5NameMap.end() && Result->first == FuncMD5Hash)
540 return Result->second;
544 Function* InstrProfSymtab::getFunction(uint64_t FuncMD5Hash) {
546 std::lower_bound(MD5FuncMap.begin(), MD5FuncMap.end(), FuncMD5Hash,
547 [](const std::pair<uint64_t, Function*> &LHS,
548 uint64_t RHS) { return LHS.first < RHS; });
549 if (Result != MD5FuncMap.end() && Result->first == FuncMD5Hash)
550 return Result->second;
554 // See also getPGOFuncName implementation. These two need to be
556 StringRef InstrProfSymtab::getOrigFuncName(uint64_t FuncMD5Hash) {
557 StringRef PGOName = getFuncName(FuncMD5Hash);
558 size_t S = PGOName.find_first_of(':');
559 if (S == StringRef::npos)
561 return PGOName.drop_front(S + 1);
564 struct InstrProfValueSiteRecord {
565 /// Value profiling data pairs at a given value site.
566 std::list<InstrProfValueData> ValueData;
568 InstrProfValueSiteRecord() { ValueData.clear(); }
569 template <class InputIterator>
570 InstrProfValueSiteRecord(InputIterator F, InputIterator L)
573 /// Sort ValueData ascending by Value
574 void sortByTargetValues() {
576 [](const InstrProfValueData &left, const InstrProfValueData &right) {
577 return left.Value < right.Value;
580 /// Sort ValueData Descending by Count
581 inline void sortByCount();
583 /// Merge data from another InstrProfValueSiteRecord
584 /// Optionally scale merged counts by \p Weight.
585 void merge(SoftInstrProfErrors &SIPE, InstrProfValueSiteRecord &Input,
586 uint64_t Weight = 1);
587 /// Scale up value profile data counts.
588 void scale(SoftInstrProfErrors &SIPE, uint64_t Weight);
591 /// Profiling information for a single function.
592 struct InstrProfRecord {
595 std::vector<uint64_t> Counts;
596 SoftInstrProfErrors SIPE;
598 InstrProfRecord() = default;
599 InstrProfRecord(StringRef Name, uint64_t Hash, std::vector<uint64_t> Counts)
600 : Name(Name), Hash(Hash), Counts(std::move(Counts)) {}
602 using ValueMapType = std::vector<std::pair<uint64_t, uint64_t>>;
604 /// Return the number of value profile kinds with non-zero number
605 /// of profile sites.
606 inline uint32_t getNumValueKinds() const;
608 /// Return the number of instrumented sites for ValueKind.
609 inline uint32_t getNumValueSites(uint32_t ValueKind) const;
611 /// Return the total number of ValueData for ValueKind.
612 inline uint32_t getNumValueData(uint32_t ValueKind) const;
614 /// Return the number of value data collected for ValueKind at profiling
616 inline uint32_t getNumValueDataForSite(uint32_t ValueKind,
617 uint32_t Site) const;
619 /// Return the array of profiled values at \p Site. If \p TotalC
620 /// is not null, the total count of all target values at this site
621 /// will be stored in \c *TotalC.
622 inline std::unique_ptr<InstrProfValueData[]>
623 getValueForSite(uint32_t ValueKind, uint32_t Site,
624 uint64_t *TotalC = nullptr) const;
626 /// Get the target value/counts of kind \p ValueKind collected at site
627 /// \p Site and store the result in array \p Dest. Return the total
628 /// counts of all target values at this site.
629 inline uint64_t getValueForSite(InstrProfValueData Dest[], uint32_t ValueKind,
630 uint32_t Site) const;
632 /// Reserve space for NumValueSites sites.
633 inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites);
635 /// Add ValueData for ValueKind at value Site.
636 void addValueData(uint32_t ValueKind, uint32_t Site,
637 InstrProfValueData *VData, uint32_t N,
638 ValueMapType *ValueMap);
640 /// Merge the counts in \p Other into this one.
641 /// Optionally scale merged counts by \p Weight.
642 void merge(InstrProfRecord &Other, uint64_t Weight = 1);
644 /// Scale up profile counts (including value profile data) by
646 void scale(uint64_t Weight);
648 /// Sort value profile data (per site) by count.
649 void sortValueData() {
650 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
651 std::vector<InstrProfValueSiteRecord> &SiteRecords =
652 getValueSitesForKind(Kind);
653 for (auto &SR : SiteRecords)
658 /// Clear value data entries and edge counters.
664 /// Clear value data entries
665 void clearValueData() {
666 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
667 getValueSitesForKind(Kind).clear();
670 /// Get the error contained within the record's soft error counter.
671 Error takeError() { return SIPE.takeError(); }
674 std::vector<InstrProfValueSiteRecord> IndirectCallSites;
675 std::vector<InstrProfValueSiteRecord> MemOPSizes;
677 const std::vector<InstrProfValueSiteRecord> &
678 getValueSitesForKind(uint32_t ValueKind) const {
680 case IPVK_IndirectCallTarget:
681 return IndirectCallSites;
685 llvm_unreachable("Unknown value kind!");
687 return IndirectCallSites;
690 std::vector<InstrProfValueSiteRecord> &
691 getValueSitesForKind(uint32_t ValueKind) {
692 return const_cast<std::vector<InstrProfValueSiteRecord> &>(
693 const_cast<const InstrProfRecord *>(this)
694 ->getValueSitesForKind(ValueKind));
697 // Map indirect call target name hash to name string.
698 uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
699 ValueMapType *HashKeys);
701 // Merge Value Profile data from Src record to this record for ValueKind.
702 // Scale merged value counts by \p Weight.
703 void mergeValueProfData(uint32_t ValueKind, InstrProfRecord &Src,
706 // Scale up value profile data count.
707 void scaleValueProfData(uint32_t ValueKind, uint64_t Weight);
710 uint32_t InstrProfRecord::getNumValueKinds() const {
711 uint32_t NumValueKinds = 0;
712 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
713 NumValueKinds += !(getValueSitesForKind(Kind).empty());
714 return NumValueKinds;
717 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
719 const std::vector<InstrProfValueSiteRecord> &SiteRecords =
720 getValueSitesForKind(ValueKind);
721 for (auto &SR : SiteRecords) {
722 N += SR.ValueData.size();
727 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
728 return getValueSitesForKind(ValueKind).size();
731 uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind,
732 uint32_t Site) const {
733 return getValueSitesForKind(ValueKind)[Site].ValueData.size();
736 std::unique_ptr<InstrProfValueData[]>
737 InstrProfRecord::getValueForSite(uint32_t ValueKind, uint32_t Site,
738 uint64_t *TotalC) const {
740 uint64_t &TotalCount = (TotalC == nullptr ? Dummy : *TotalC);
741 uint32_t N = getNumValueDataForSite(ValueKind, Site);
744 return std::unique_ptr<InstrProfValueData[]>(nullptr);
747 auto VD = llvm::make_unique<InstrProfValueData[]>(N);
748 TotalCount = getValueForSite(VD.get(), ValueKind, Site);
753 uint64_t InstrProfRecord::getValueForSite(InstrProfValueData Dest[],
755 uint32_t Site) const {
757 uint64_t TotalCount = 0;
758 for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) {
759 Dest[I].Value = V.Value;
760 Dest[I].Count = V.Count;
761 TotalCount = SaturatingAdd(TotalCount, V.Count);
767 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
768 std::vector<InstrProfValueSiteRecord> &ValueSites =
769 getValueSitesForKind(ValueKind);
770 ValueSites.reserve(NumValueSites);
773 inline support::endianness getHostEndianness() {
774 return sys::IsLittleEndianHost ? support::little : support::big;
777 // Include definitions for value profile data
778 #define INSTR_PROF_VALUE_PROF_DATA
779 #include "llvm/ProfileData/InstrProfData.inc"
781 void InstrProfValueSiteRecord::sortByCount() {
783 [](const InstrProfValueData &left, const InstrProfValueData &right) {
784 return left.Count > right.Count;
787 size_t max_s = INSTR_PROF_MAX_NUM_VAL_PER_SITE;
788 if (ValueData.size() > max_s)
789 ValueData.resize(max_s);
792 namespace IndexedInstrProf {
794 enum class HashT : uint32_t {
799 inline uint64_t ComputeHash(HashT Type, StringRef K) {
804 llvm_unreachable("Unhandled hash type");
807 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
810 // Version 1 is the first version. In this version, the value of
811 // a key/value pair can only include profile data of a single function.
812 // Due to this restriction, the number of block counters for a given
813 // function is not recorded but derived from the length of the value.
815 // The version 2 format supports recording profile data of multiple
816 // functions which share the same key in one value field. To support this,
817 // the number block counters is recorded as an uint64_t field right after the
818 // function structural hash.
820 // Version 3 supports value profile data. The value profile data is expected
821 // to follow the block counter profile data.
823 // In this version, profile summary data \c IndexedInstrProf::Summary is
824 // stored after the profile header.
826 // The current version is 4.
827 CurrentVersion = INSTR_PROF_INDEX_VERSION
829 const uint64_t Version = ProfVersion::CurrentVersion;
831 const HashT HashType = HashT::MD5;
833 inline uint64_t ComputeHash(StringRef K) { return ComputeHash(HashType, K); }
835 // This structure defines the file header of the LLVM profile
836 // data file in indexed-format.
840 uint64_t Unused; // Becomes unused since version 4
845 // Profile summary data recorded in the profile data file in indexed
846 // format. It is introduced in version 4. The summary data follows
847 // right after the profile file header.
850 uint64_t Cutoff; ///< The required percentile of total execution count.
852 MinBlockCount; ///< The minimum execution count for this percentile.
853 uint64_t NumBlocks; ///< Number of blocks >= the minumum execution count.
855 // The field kind enumerator to assigned value mapping should remain
856 // unchanged when a new kind is added or an old kind gets deleted in
858 enum SummaryFieldKind {
859 /// The total number of functions instrumented.
860 TotalNumFunctions = 0,
861 /// Total number of instrumented blocks/edges.
863 /// The maximal execution count among all functions.
864 /// This field does not exist for profile data from IR based
866 MaxFunctionCount = 2,
867 /// Max block count of the program.
869 /// Max internal block count of the program (excluding entry blocks).
870 MaxInternalBlockCount = 4,
871 /// The sum of all instrumented block counts.
873 NumKinds = TotalBlockCount + 1
876 // The number of summmary fields following the summary header.
877 uint64_t NumSummaryFields;
878 // The number of Cutoff Entries (Summary::Entry) following summary fields.
879 uint64_t NumCutoffEntries;
882 Summary(uint32_t Size) { memset(this, 0, Size); }
884 void operator delete(void *ptr) { ::operator delete(ptr); }
886 static uint32_t getSize(uint32_t NumSumFields, uint32_t NumCutoffEntries) {
887 return sizeof(Summary) + NumCutoffEntries * sizeof(Entry) +
888 NumSumFields * sizeof(uint64_t);
891 const uint64_t *getSummaryDataBase() const {
892 return reinterpret_cast<const uint64_t *>(this + 1);
895 uint64_t *getSummaryDataBase() {
896 return reinterpret_cast<uint64_t *>(this + 1);
899 const Entry *getCutoffEntryBase() const {
900 return reinterpret_cast<const Entry *>(
901 &getSummaryDataBase()[NumSummaryFields]);
904 Entry *getCutoffEntryBase() {
905 return reinterpret_cast<Entry *>(&getSummaryDataBase()[NumSummaryFields]);
908 uint64_t get(SummaryFieldKind K) const {
909 return getSummaryDataBase()[K];
912 void set(SummaryFieldKind K, uint64_t V) {
913 getSummaryDataBase()[K] = V;
916 const Entry &getEntry(uint32_t I) const { return getCutoffEntryBase()[I]; }
918 void setEntry(uint32_t I, const ProfileSummaryEntry &E) {
919 Entry &ER = getCutoffEntryBase()[I];
920 ER.Cutoff = E.Cutoff;
921 ER.MinBlockCount = E.MinCount;
922 ER.NumBlocks = E.NumCounts;
926 inline std::unique_ptr<Summary> allocSummary(uint32_t TotalSize) {
927 return std::unique_ptr<Summary>(new (::operator new(TotalSize))
931 } // end namespace IndexedInstrProf
933 namespace RawInstrProf {
935 // Version 1: First version
936 // Version 2: Added value profile data section. Per-function control data
937 // struct has more fields to describe value profile information.
938 // Version 3: Compressed name section support. Function PGO name reference
939 // from control data struct is changed from raw pointer to Name's MD5 value.
940 // Version 4: ValueDataBegin and ValueDataSizes fields are removed from the
942 const uint64_t Version = INSTR_PROF_RAW_VERSION;
944 template <class IntPtrT> inline uint64_t getMagic();
945 template <> inline uint64_t getMagic<uint64_t>() {
946 return INSTR_PROF_RAW_MAGIC_64;
949 template <> inline uint64_t getMagic<uint32_t>() {
950 return INSTR_PROF_RAW_MAGIC_32;
953 // Per-function profile data header/control structure.
954 // The definition should match the structure defined in
955 // compiler-rt/lib/profile/InstrProfiling.h.
956 // It should also match the synthesized type in
957 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
958 template <class IntPtrT> struct LLVM_ALIGNAS(8) ProfileData {
959 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
960 #include "llvm/ProfileData/InstrProfData.inc"
963 // File header structure of the LLVM profile data in raw format.
964 // The definition should match the header referenced in
965 // compiler-rt/lib/profile/InstrProfilingFile.c and
966 // InstrProfilingBuffer.c.
968 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) const Type Name;
969 #include "llvm/ProfileData/InstrProfData.inc"
972 } // end namespace RawInstrProf
974 // Parse MemOP Size range option.
975 void getMemOPSizeRangeFromOption(StringRef Str, int64_t &RangeStart,
978 } // end namespace llvm
980 #endif // LLVM_PROFILEDATA_INSTRPROF_H