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/STLExtras.h"
20 #include "llvm/ADT/StringRef.h"
21 #include "llvm/ADT/StringSet.h"
22 #include "llvm/IR/GlobalValue.h"
23 #include "llvm/ProfileData/InstrProfData.inc"
24 #include "llvm/Support/Endian.h"
25 #include "llvm/Support/ErrorHandling.h"
26 #include "llvm/Support/ErrorOr.h"
27 #include "llvm/Support/MD5.h"
30 #include <system_error>
33 #define INSTR_PROF_INDEX_VERSION 3
40 /// Return the name of data section containing profile counter variables.
41 inline StringRef getInstrProfCountersSectionName(bool AddSegment) {
42 return AddSegment ? "__DATA," INSTR_PROF_CNTS_SECT_NAME_STR
43 : INSTR_PROF_CNTS_SECT_NAME_STR;
46 /// Return the name of data section containing names of instrumented
48 inline StringRef getInstrProfNameSectionName(bool AddSegment) {
49 return AddSegment ? "__DATA," INSTR_PROF_NAME_SECT_NAME_STR
50 : INSTR_PROF_NAME_SECT_NAME_STR;
53 /// Return the name of the data section containing per-function control
55 inline StringRef getInstrProfDataSectionName(bool AddSegment) {
56 return AddSegment ? "__DATA," INSTR_PROF_DATA_SECT_NAME_STR
57 : INSTR_PROF_DATA_SECT_NAME_STR;
60 /// Return the name profile runtime entry point to do value profiling
62 inline StringRef getInstrProfValueProfFuncName() {
63 return INSTR_PROF_VALUE_PROF_FUNC_STR;
66 /// Return the name of the section containing function coverage mapping
68 inline StringRef getInstrProfCoverageSectionName(bool AddSegment) {
69 return AddSegment ? "__DATA,__llvm_covmap" : "__llvm_covmap";
72 /// Return the name prefix of variables containing instrumented function names.
73 inline StringRef getInstrProfNameVarPrefix() { return "__profn_"; }
75 /// Return the name prefix of variables containing per-function control data.
76 inline StringRef getInstrProfDataVarPrefix() { return "__profd_"; }
78 /// Return the name prefix of profile counter variables.
79 inline StringRef getInstrProfCountersVarPrefix() { return "__profc_"; }
81 /// Return the name prefix of the COMDAT group for instrumentation variables
82 /// associated with a COMDAT function.
83 inline StringRef getInstrProfComdatPrefix() { return "__profv_"; }
85 /// Return the name of a covarage mapping variable (internal linkage)
86 /// for each instrumented source module. Such variables are allocated
87 /// in the __llvm_covmap section.
88 inline StringRef getCoverageMappingVarName() {
89 return "__llvm_coverage_mapping";
92 /// Return the name of function that registers all the per-function control
93 /// data at program startup time by calling __llvm_register_function. This
94 /// function has internal linkage and is called by __llvm_profile_init
95 /// runtime method. This function is not generated for these platforms:
96 /// Darwin, Linux, and FreeBSD.
97 inline StringRef getInstrProfRegFuncsName() {
98 return "__llvm_profile_register_functions";
101 /// Return the name of the runtime interface that registers per-function control
102 /// data for one instrumented function.
103 inline StringRef getInstrProfRegFuncName() {
104 return "__llvm_profile_register_function";
107 /// Return the name of the runtime initialization method that is generated by
108 /// the compiler. The function calls __llvm_profile_register_functions and
109 /// __llvm_profile_override_default_filename functions if needed. This function
110 /// has internal linkage and invoked at startup time via init_array.
111 inline StringRef getInstrProfInitFuncName() { return "__llvm_profile_init"; }
113 /// Return the name of the hook variable defined in profile runtime library.
114 /// A reference to the variable causes the linker to link in the runtime
115 /// initialization module (which defines the hook variable).
116 inline StringRef getInstrProfRuntimeHookVarName() {
117 return "__llvm_profile_runtime";
120 /// Return the name of the compiler generated function that references the
121 /// runtime hook variable. The function is a weak global.
122 inline StringRef getInstrProfRuntimeHookVarUseFuncName() {
123 return "__llvm_profile_runtime_user";
126 /// Return the name of the profile runtime interface that overrides the default
127 /// profile data file name.
128 inline StringRef getInstrProfFileOverriderFuncName() {
129 return "__llvm_profile_override_default_filename";
132 /// Return the modified name for function \c F suitable to be
133 /// used the key for profile lookup.
134 std::string getPGOFuncName(const Function &F,
135 uint64_t Version = INSTR_PROF_INDEX_VERSION);
137 /// Return the modified name for a function suitable to be
138 /// used the key for profile lookup. The function's original
139 /// name is \c RawFuncName and has linkage of type \c Linkage.
140 /// The function is defined in module \c FileName.
141 std::string getPGOFuncName(StringRef RawFuncName,
142 GlobalValue::LinkageTypes Linkage,
144 uint64_t Version = INSTR_PROF_INDEX_VERSION);
146 /// Create and return the global variable for function name used in PGO
147 /// instrumentation. \c FuncName is the name of the function returned
148 /// by \c getPGOFuncName call.
149 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef FuncName);
151 /// Create and return the global variable for function name used in PGO
152 /// instrumentation. /// \c FuncName is the name of the function
153 /// returned by \c getPGOFuncName call, \c M is the owning module,
154 /// and \c Linkage is the linkage of the instrumented function.
155 GlobalVariable *createPGOFuncNameVar(Module &M,
156 GlobalValue::LinkageTypes Linkage,
159 /// Given a PGO function name, remove the filename prefix and return
160 /// the original (static) function name.
161 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, StringRef FileName);
163 const std::error_category &instrprof_category();
165 enum class instrprof_error {
172 unsupported_hash_type,
180 value_site_count_mismatch
183 inline std::error_code make_error_code(instrprof_error E) {
184 return std::error_code(static_cast<int>(E), instrprof_category());
187 inline instrprof_error MergeResult(instrprof_error &Accumulator,
188 instrprof_error Result) {
189 // Prefer first error encountered as later errors may be secondary effects of
190 // the initial problem.
191 if (Accumulator == instrprof_error::success &&
192 Result != instrprof_error::success)
193 Accumulator = Result;
197 enum InstrProfValueKind : uint32_t {
198 #define VALUE_PROF_KIND(Enumerator, Value) Enumerator = Value,
199 #include "llvm/ProfileData/InstrProfData.inc"
205 /// A symbol table used for function PGO name look-up with keys
206 /// (such as pointers, md5hash values) to the function. A function's
207 /// PGO name or name's md5hash are used in retrieving the profile
208 /// data of the function. See \c getPGOFuncName() method for details
209 /// on how PGO name is formed.
210 class InstrProfSymtab {
216 InstrProfSymtab() : Data(), Address(0) {}
218 /// Create InstrProfSymtab from a object file section which
219 /// contains function PGO names that are uncompressed.
220 std::error_code create(object::SectionRef &Section);
221 std::error_code create(StringRef D, uint64_t BaseAddr) {
224 return std::error_code();
227 /// Return function's PGO name from the function name's symabol
228 /// address in the object file. If an error occurs, Return
230 StringRef getFuncName(uint64_t FuncNameAddress, size_t NameSize);
233 struct InstrProfStringTable {
234 // Set of string values in profiling data.
235 StringSet<> StringValueSet;
236 InstrProfStringTable() { StringValueSet.clear(); }
237 // Get a pointer to internal storage of a string in set
238 const char *getStringData(StringRef Str) {
239 auto Result = StringValueSet.find(Str);
240 return (Result == StringValueSet.end()) ? nullptr : Result->first().data();
242 // Insert a string to StringTable
243 const char *insertString(StringRef Str) {
244 auto Result = StringValueSet.insert(Str);
245 return Result.first->first().data();
249 struct InstrProfValueSiteRecord {
250 /// Value profiling data pairs at a given value site.
251 std::list<InstrProfValueData> ValueData;
253 InstrProfValueSiteRecord() { ValueData.clear(); }
254 template <class InputIterator>
255 InstrProfValueSiteRecord(InputIterator F, InputIterator L)
258 /// Sort ValueData ascending by Value
259 void sortByTargetValues() {
261 [](const InstrProfValueData &left, const InstrProfValueData &right) {
262 return left.Value < right.Value;
266 /// Merge data from another InstrProfValueSiteRecord
267 /// Optionally scale merged counts by \p Weight.
268 instrprof_error mergeValueData(InstrProfValueSiteRecord &Input,
269 uint64_t Weight = 1) {
270 this->sortByTargetValues();
271 Input.sortByTargetValues();
272 auto I = ValueData.begin();
273 auto IE = ValueData.end();
274 instrprof_error Result = instrprof_error::success;
275 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
277 while (I != IE && I->Value < J->Value)
279 if (I != IE && I->Value == J->Value) {
280 uint64_t JCount = J->Count;
283 JCount = SaturatingMultiply(JCount, Weight, &Overflowed);
285 Result = instrprof_error::counter_overflow;
287 I->Count = SaturatingAdd(I->Count, JCount, &Overflowed);
289 Result = instrprof_error::counter_overflow;
293 ValueData.insert(I, *J);
299 /// Profiling information for a single function.
300 struct InstrProfRecord {
302 InstrProfRecord(StringRef Name, uint64_t Hash, std::vector<uint64_t> Counts)
303 : Name(Name), Hash(Hash), Counts(std::move(Counts)) {}
306 std::vector<uint64_t> Counts;
308 typedef std::vector<std::pair<uint64_t, const char *>> ValueMapType;
310 /// Return the number of value profile kinds with non-zero number
311 /// of profile sites.
312 inline uint32_t getNumValueKinds() const;
313 /// Return the number of instrumented sites for ValueKind.
314 inline uint32_t getNumValueSites(uint32_t ValueKind) const;
315 /// Return the total number of ValueData for ValueKind.
316 inline uint32_t getNumValueData(uint32_t ValueKind) const;
317 /// Return the number of value data collected for ValueKind at profiling
319 inline uint32_t getNumValueDataForSite(uint32_t ValueKind,
320 uint32_t Site) const;
321 /// Return the array of profiled values at \p Site.
322 inline std::unique_ptr<InstrProfValueData[]>
323 getValueForSite(uint32_t ValueKind, uint32_t Site,
324 uint64_t (*ValueMapper)(uint32_t, uint64_t) = 0) const;
326 getValueForSite(InstrProfValueData Dest[], uint32_t ValueKind, uint32_t Site,
327 uint64_t (*ValueMapper)(uint32_t, uint64_t) = 0) const;
328 /// Reserve space for NumValueSites sites.
329 inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites);
330 /// Add ValueData for ValueKind at value Site.
331 inline void addValueData(uint32_t ValueKind, uint32_t Site,
332 InstrProfValueData *VData, uint32_t N,
333 ValueMapType *HashKeys);
335 /// Merge the counts in \p Other into this one.
336 /// Optionally scale merged counts by \p Weight.
337 inline instrprof_error merge(InstrProfRecord &Other, uint64_t Weight = 1);
339 /// Used by InstrProfWriter: update the value strings to commoned strings in
340 /// the writer instance.
341 inline void updateStrings(InstrProfStringTable *StrTab);
343 /// Clear value data entries
344 inline void clearValueData() {
345 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
346 getValueSitesForKind(Kind).clear();
350 std::vector<InstrProfValueSiteRecord> IndirectCallSites;
351 const std::vector<InstrProfValueSiteRecord> &
352 getValueSitesForKind(uint32_t ValueKind) const {
354 case IPVK_IndirectCallTarget:
355 return IndirectCallSites;
357 llvm_unreachable("Unknown value kind!");
359 return IndirectCallSites;
362 std::vector<InstrProfValueSiteRecord> &
363 getValueSitesForKind(uint32_t ValueKind) {
364 return const_cast<std::vector<InstrProfValueSiteRecord> &>(
365 const_cast<const InstrProfRecord *>(this)
366 ->getValueSitesForKind(ValueKind));
369 // Map indirect call target name hash to name string.
370 uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
371 ValueMapType *HashKeys) {
375 case IPVK_IndirectCallTarget: {
377 std::lower_bound(HashKeys->begin(), HashKeys->end(), Value,
378 [](const std::pair<uint64_t, const char *> &LHS,
379 uint64_t RHS) { return LHS.first < RHS; });
380 if (Result != HashKeys->end())
381 Value = (uint64_t)Result->second;
388 // Merge Value Profile data from Src record to this record for ValueKind.
389 // Scale merged value counts by \p Weight.
390 instrprof_error mergeValueProfData(uint32_t ValueKind, InstrProfRecord &Src,
392 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
393 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
394 if (ThisNumValueSites != OtherNumValueSites)
395 return instrprof_error::value_site_count_mismatch;
396 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
397 getValueSitesForKind(ValueKind);
398 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
399 Src.getValueSitesForKind(ValueKind);
400 instrprof_error Result = instrprof_error::success;
401 for (uint32_t I = 0; I < ThisNumValueSites; I++)
402 MergeResult(Result, ThisSiteRecords[I].mergeValueData(OtherSiteRecords[I],
408 uint32_t InstrProfRecord::getNumValueKinds() const {
409 uint32_t NumValueKinds = 0;
410 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
411 NumValueKinds += !(getValueSitesForKind(Kind).empty());
412 return NumValueKinds;
415 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
417 const std::vector<InstrProfValueSiteRecord> &SiteRecords =
418 getValueSitesForKind(ValueKind);
419 for (auto &SR : SiteRecords) {
420 N += SR.ValueData.size();
425 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
426 return getValueSitesForKind(ValueKind).size();
429 uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind,
430 uint32_t Site) const {
431 return getValueSitesForKind(ValueKind)[Site].ValueData.size();
434 std::unique_ptr<InstrProfValueData[]> InstrProfRecord::getValueForSite(
435 uint32_t ValueKind, uint32_t Site,
436 uint64_t (*ValueMapper)(uint32_t, uint64_t)) const {
437 uint32_t N = getNumValueDataForSite(ValueKind, Site);
439 return std::unique_ptr<InstrProfValueData[]>(nullptr);
441 auto VD = llvm::make_unique<InstrProfValueData[]>(N);
442 getValueForSite(VD.get(), ValueKind, Site, ValueMapper);
447 void InstrProfRecord::getValueForSite(InstrProfValueData Dest[],
448 uint32_t ValueKind, uint32_t Site,
449 uint64_t (*ValueMapper)(uint32_t,
452 for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) {
453 Dest[I].Value = ValueMapper ? ValueMapper(ValueKind, V.Value) : V.Value;
454 Dest[I].Count = V.Count;
459 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
460 InstrProfValueData *VData, uint32_t N,
461 ValueMapType *HashKeys) {
462 for (uint32_t I = 0; I < N; I++) {
463 VData[I].Value = remapValue(VData[I].Value, ValueKind, HashKeys);
465 std::vector<InstrProfValueSiteRecord> &ValueSites =
466 getValueSitesForKind(ValueKind);
468 ValueSites.push_back(InstrProfValueSiteRecord());
470 ValueSites.emplace_back(VData, VData + N);
473 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
474 std::vector<InstrProfValueSiteRecord> &ValueSites =
475 getValueSitesForKind(ValueKind);
476 ValueSites.reserve(NumValueSites);
479 void InstrProfRecord::updateStrings(InstrProfStringTable *StrTab) {
483 Name = StrTab->insertString(Name);
484 for (auto &VSite : IndirectCallSites)
485 for (auto &VData : VSite.ValueData)
486 VData.Value = (uint64_t)StrTab->insertString((const char *)VData.Value);
489 instrprof_error InstrProfRecord::merge(InstrProfRecord &Other,
491 // If the number of counters doesn't match we either have bad data
492 // or a hash collision.
493 if (Counts.size() != Other.Counts.size())
494 return instrprof_error::count_mismatch;
496 instrprof_error Result = instrprof_error::success;
498 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
500 uint64_t OtherCount = Other.Counts[I];
502 OtherCount = SaturatingMultiply(OtherCount, Weight, &Overflowed);
504 Result = instrprof_error::counter_overflow;
506 Counts[I] = SaturatingAdd(Counts[I], OtherCount, &Overflowed);
508 Result = instrprof_error::counter_overflow;
511 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
512 MergeResult(Result, mergeValueProfData(Kind, Other, Weight));
517 inline support::endianness getHostEndianness() {
518 return sys::IsLittleEndianHost ? support::little : support::big;
522 // Include definitions for value profile data
523 #define INSTR_PROF_VALUE_PROF_DATA
524 #include "llvm/ProfileData/InstrProfData.inc"
527 * Initialize the record for runtime value profile data.
528 * Return 0 if the initialization is successful, otherwise
531 int initializeValueProfRuntimeRecord(ValueProfRuntimeRecord *RuntimeRecord,
532 const uint16_t *NumValueSites,
533 ValueProfNode **Nodes);
535 /* Release memory allocated for the runtime record. */
536 void finalizeValueProfRuntimeRecord(ValueProfRuntimeRecord *RuntimeRecord);
538 /* Return the size of ValueProfData structure that can be used to store
539 the value profile data collected at runtime. */
540 uint32_t getValueProfDataSizeRT(const ValueProfRuntimeRecord *Record);
542 /* Return a ValueProfData instance that stores the data collected at runtime. */
544 serializeValueProfDataFromRT(const ValueProfRuntimeRecord *Record,
547 namespace IndexedInstrProf {
549 enum class HashT : uint32_t {
555 static inline uint64_t MD5Hash(StringRef Str) {
558 llvm::MD5::MD5Result Result;
560 // Return the least significant 8 bytes. Our MD5 implementation returns the
561 // result in little endian, so we may need to swap bytes.
562 using namespace llvm::support;
563 return endian::read<uint64_t, little, unaligned>(Result);
566 static inline uint64_t ComputeHash(HashT Type, StringRef K) {
569 return IndexedInstrProf::MD5Hash(K);
571 llvm_unreachable("Unhandled hash type");
574 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
575 const uint64_t Version = INSTR_PROF_INDEX_VERSION;
576 const HashT HashType = HashT::MD5;
578 // This structure defines the file header of the LLVM profile
579 // data file in indexed-format.
583 uint64_t MaxFunctionCount;
588 } // end namespace IndexedInstrProf
590 namespace RawInstrProf {
592 const uint64_t Version = INSTR_PROF_RAW_VERSION;
594 template <class IntPtrT> inline uint64_t getMagic();
595 template <> inline uint64_t getMagic<uint64_t>() {
596 return INSTR_PROF_RAW_MAGIC_64;
599 template <> inline uint64_t getMagic<uint32_t>() {
600 return INSTR_PROF_RAW_MAGIC_32;
603 // Per-function profile data header/control structure.
604 // The definition should match the structure defined in
605 // compiler-rt/lib/profile/InstrProfiling.h.
606 // It should also match the synthesized type in
607 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
608 template <class IntPtrT> struct LLVM_ALIGNAS(8) ProfileData {
609 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
610 #include "llvm/ProfileData/InstrProfData.inc"
613 // File header structure of the LLVM profile data in raw format.
614 // The definition should match the header referenced in
615 // compiler-rt/lib/profile/InstrProfilingFile.c and
616 // InstrProfilingBuffer.c.
618 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) const Type Name;
619 #include "llvm/ProfileData/InstrProfData.inc"
622 } // end namespace RawInstrProf
626 // Profile coverage map has the following layout:
627 // [CoverageMapFileHeader]
629 // [CovMapFunctionRecord]
630 // [CovMapFunctionRecord]
633 // [Encoded Region Mapping Data]
635 template <class IntPtrT> struct CovMapFunctionRecord {
636 #define COVMAP_FUNC_RECORD(Type, LLVMType, Name, Init) Type Name;
637 #include "llvm/ProfileData/InstrProfData.inc"
643 } // end namespace llvm
647 struct is_error_code_enum<llvm::instrprof_error> : std::true_type {};
650 #endif // LLVM_PROFILEDATA_INSTRPROF_H_