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/StringRef.h"
20 #include "llvm/ADT/StringSet.h"
21 #include "llvm/IR/GlobalValue.h"
22 #include "llvm/Support/Endian.h"
23 #include "llvm/Support/ErrorHandling.h"
24 #include "llvm/Support/ErrorOr.h"
25 #include "llvm/Support/MD5.h"
28 #include <system_error>
37 /// Return the name of data section containing profile counter variables.
38 inline StringRef getInstrProfCountersSectionName(bool AddSegment) {
39 return AddSegment ? "__DATA,__llvm_prf_cnts" : "__llvm_prf_cnts";
42 /// Return the name of data section containing names of instrumented
44 inline StringRef getInstrProfNameSectionName(bool AddSegment) {
45 return AddSegment ? "__DATA,__llvm_prf_names" : "__llvm_prf_names";
48 /// Return the name of the data section containing per-function control
50 inline StringRef getInstrProfDataSectionName(bool AddSegment) {
51 return AddSegment ? "__DATA,__llvm_prf_data" : "__llvm_prf_data";
54 /// Return the name of the section containing function coverage mapping
56 inline StringRef getInstrProfCoverageSectionName(bool AddSegment) {
57 return AddSegment ? "__DATA,__llvm_covmap" : "__llvm_covmap";
60 /// Return the name prefix of variables containing instrumented function names.
61 inline StringRef getInstrProfNameVarPrefix() { return "__llvm_profile_name_"; }
63 /// Return the name prefix of variables containing per-function control data.
64 inline StringRef getInstrProfDataVarPrefix() { return "__llvm_profile_data_"; }
66 /// Return the name prefix of profile counter variables.
67 inline StringRef getInstrProfCountersVarPrefix() {
68 return "__llvm_profile_counters_";
71 /// Return the name prefix of the COMDAT group for instrumentation variables
72 /// associated with a COMDAT function.
73 inline StringRef getInstrProfComdatPrefix() { return "__llvm_profile_vars_"; }
75 /// Return the name of a covarage mapping variable (internal linkage)
76 /// for each instrumented source module. Such variables are allocated
77 /// in the __llvm_covmap section.
78 inline StringRef getCoverageMappingVarName() {
79 return "__llvm_coverage_mapping";
82 /// Return the name of function that registers all the per-function control
83 /// data at program startup time by calling __llvm_register_function. This
84 /// function has internal linkage and is called by __llvm_profile_init
85 /// runtime method. This function is not generated for these platforms:
86 /// Darwin, Linux, and FreeBSD.
87 inline StringRef getInstrProfRegFuncsName() {
88 return "__llvm_profile_register_functions";
91 /// Return the name of the runtime interface that registers per-function control
92 /// data for one instrumented function.
93 inline StringRef getInstrProfRegFuncName() {
94 return "__llvm_profile_register_function";
97 /// Return the name of the runtime initialization method that is generated by
98 /// the compiler. The function calls __llvm_profile_register_functions and
99 /// __llvm_profile_override_default_filename functions if needed. This function
100 /// has internal linkage and invoked at startup time via init_array.
101 inline StringRef getInstrProfInitFuncName() { return "__llvm_profile_init"; }
103 /// Return the name of the hook variable defined in profile runtime library.
104 /// A reference to the variable causes the linker to link in the runtime
105 /// initialization module (which defines the hook variable).
106 inline StringRef getInstrProfRuntimeHookVarName() {
107 return "__llvm_profile_runtime";
110 /// Return the name of the compiler generated function that references the
111 /// runtime hook variable. The function is a weak global.
112 inline StringRef getInstrProfRuntimeHookVarUseFuncName() {
113 return "__llvm_profile_runtime_user";
116 /// Return the name of the profile runtime interface that overrides the default
117 /// profile data file name.
118 inline StringRef getInstrProfFileOverriderFuncName() {
119 return "__llvm_profile_override_default_filename";
122 /// Return the modified name for function \c F suitable to be
123 /// used the key for profile lookup.
124 std::string getPGOFuncName(const Function &F);
126 /// Return the modified name for a function suitable to be
127 /// used the key for profile lookup. The function's original
128 /// name is \c RawFuncName and has linkage of type \c Linkage.
129 /// The function is defined in module \c FileName.
130 std::string getPGOFuncName(StringRef RawFuncName,
131 GlobalValue::LinkageTypes Linkage,
134 /// Create and return the global variable for function name used in PGO
135 /// instrumentation. \c FuncName is the name of the function returned
136 /// by \c getPGOFuncName call.
137 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef FuncName);
139 /// Create and return the global variable for function name used in PGO
140 /// instrumentation. /// \c FuncName is the name of the function
141 /// returned by \c getPGOFuncName call, \c M is the owning module,
142 /// and \c Linkage is the linkage of the instrumented function.
143 GlobalVariable *createPGOFuncNameVar(Module &M,
144 GlobalValue::LinkageTypes Linkage,
147 const std::error_category &instrprof_category();
149 enum class instrprof_error {
155 unsupported_hash_type,
163 value_site_count_mismatch
166 inline std::error_code make_error_code(instrprof_error E) {
167 return std::error_code(static_cast<int>(E), instrprof_category());
170 enum InstrProfValueKind : uint32_t {
171 IPVK_IndirectCallTarget = 0,
173 IPVK_First = IPVK_IndirectCallTarget,
174 IPVK_Last = IPVK_IndirectCallTarget
177 struct InstrProfStringTable {
178 // Set of string values in profiling data.
179 StringSet<> StringValueSet;
180 InstrProfStringTable() { StringValueSet.clear(); }
181 // Get a pointer to internal storage of a string in set
182 const char *getStringData(StringRef Str) {
183 auto Result = StringValueSet.find(Str);
184 return (Result == StringValueSet.end()) ? nullptr : Result->first().data();
186 // Insert a string to StringTable
187 const char *insertString(StringRef Str) {
188 auto Result = StringValueSet.insert(Str);
189 return Result.first->first().data();
193 struct InstrProfValueData {
196 // Number of times the value appears in the training run.
200 struct InstrProfValueSiteRecord {
201 /// Value profiling data pairs at a given value site.
202 std::list<InstrProfValueData> ValueData;
204 InstrProfValueSiteRecord() { ValueData.clear(); }
205 template <class InputIterator>
206 InstrProfValueSiteRecord(InputIterator F, InputIterator L)
209 /// Sort ValueData ascending by Value
210 void sortByTargetValues() {
212 [](const InstrProfValueData &left, const InstrProfValueData &right) {
213 return left.Value < right.Value;
217 /// Merge data from another InstrProfValueSiteRecord
218 void mergeValueData(InstrProfValueSiteRecord &Input) {
219 this->sortByTargetValues();
220 Input.sortByTargetValues();
221 auto I = ValueData.begin();
222 auto IE = ValueData.end();
223 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
225 while (I != IE && I->Value < J->Value)
227 if (I != IE && I->Value == J->Value) {
228 I->Count += J->Count;
232 ValueData.insert(I, *J);
237 /// Profiling information for a single function.
238 struct InstrProfRecord {
240 InstrProfRecord(StringRef Name, uint64_t Hash, std::vector<uint64_t> Counts)
241 : Name(Name), Hash(Hash), Counts(std::move(Counts)) {}
244 std::vector<uint64_t> Counts;
246 typedef std::vector<std::pair<uint64_t, const char *>> ValueMapType;
248 /// Return the number of value profile kinds with non-zero number
249 /// of profile sites.
250 inline uint32_t getNumValueKinds() const;
251 /// Return the number of instrumented sites for ValueKind.
252 inline uint32_t getNumValueSites(uint32_t ValueKind) const;
253 /// Return the total number of ValueData for ValueKind.
254 inline uint32_t getNumValueData(uint32_t ValueKind) const;
255 /// Return the number of value data collected for ValueKind at profiling
257 inline uint32_t getNumValueDataForSite(uint32_t ValueKind,
258 uint32_t Site) const;
259 inline std::unique_ptr<InstrProfValueData[]>
260 getValueForSite(uint32_t ValueKind, uint32_t Site) const;
261 /// Reserve space for NumValueSites sites.
262 inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites);
263 /// Add ValueData for ValueKind at value Site.
264 inline void addValueData(uint32_t ValueKind, uint32_t Site,
265 InstrProfValueData *VData, uint32_t N,
266 ValueMapType *HashKeys);
267 /// Merge Value Profile ddata from Src record to this record for ValueKind.
268 inline instrprof_error mergeValueProfData(uint32_t ValueKind,
269 InstrProfRecord &Src);
271 /// Used by InstrProfWriter: update the value strings to commoned strings in
272 /// the writer instance.
273 inline void updateStrings(InstrProfStringTable *StrTab);
276 std::vector<InstrProfValueSiteRecord> IndirectCallSites;
277 const std::vector<InstrProfValueSiteRecord> &
278 getValueSitesForKind(uint32_t ValueKind) const {
280 case IPVK_IndirectCallTarget:
281 return IndirectCallSites;
283 llvm_unreachable("Unknown value kind!");
285 return IndirectCallSites;
288 std::vector<InstrProfValueSiteRecord> &
289 getValueSitesForKind(uint32_t ValueKind) {
290 return const_cast<std::vector<InstrProfValueSiteRecord> &>(
291 const_cast<const InstrProfRecord *>(this)
292 ->getValueSitesForKind(ValueKind));
294 // Map indirect call target name hash to name string.
295 uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
296 ValueMapType *HashKeys) {
300 case IPVK_IndirectCallTarget: {
302 std::lower_bound(HashKeys->begin(), HashKeys->end(), Value,
303 [](const std::pair<uint64_t, const char *> &LHS,
304 uint64_t RHS) { return LHS.first < RHS; });
305 assert(Result != HashKeys->end() &&
306 "Hash does not match any known keys\n");
307 Value = (uint64_t)Result->second;
315 uint32_t InstrProfRecord::getNumValueKinds() const {
316 uint32_t NumValueKinds = 0;
317 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
318 NumValueKinds += !(getValueSitesForKind(Kind).empty());
319 return NumValueKinds;
322 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
324 const std::vector<InstrProfValueSiteRecord> &SiteRecords =
325 getValueSitesForKind(ValueKind);
326 for (auto &SR : SiteRecords) {
327 N += SR.ValueData.size();
332 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
333 return getValueSitesForKind(ValueKind).size();
336 uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind,
337 uint32_t Site) const {
338 return getValueSitesForKind(ValueKind)[Site].ValueData.size();
341 std::unique_ptr<InstrProfValueData[]>
342 InstrProfRecord::getValueForSite(uint32_t ValueKind, uint32_t Site) const {
343 uint32_t N = getNumValueDataForSite(ValueKind, Site);
345 return std::unique_ptr<InstrProfValueData[]>(nullptr);
347 std::unique_ptr<InstrProfValueData[]> VD(new InstrProfValueData[N]);
349 for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) {
358 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
359 InstrProfValueData *VData, uint32_t N,
360 ValueMapType *HashKeys) {
361 for (uint32_t I = 0; I < N; I++) {
362 VData[I].Value = remapValue(VData[I].Value, ValueKind, HashKeys);
364 std::vector<InstrProfValueSiteRecord> &ValueSites =
365 getValueSitesForKind(ValueKind);
367 ValueSites.push_back(InstrProfValueSiteRecord());
369 ValueSites.emplace_back(VData, VData + N);
372 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
373 std::vector<InstrProfValueSiteRecord> &ValueSites =
374 getValueSitesForKind(ValueKind);
375 ValueSites.reserve(NumValueSites);
378 instrprof_error InstrProfRecord::mergeValueProfData(uint32_t ValueKind,
379 InstrProfRecord &Src) {
380 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
381 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
382 if (ThisNumValueSites != OtherNumValueSites)
383 return instrprof_error::value_site_count_mismatch;
384 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
385 getValueSitesForKind(ValueKind);
386 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
387 Src.getValueSitesForKind(ValueKind);
388 for (uint32_t I = 0; I < ThisNumValueSites; I++)
389 ThisSiteRecords[I].mergeValueData(OtherSiteRecords[I]);
390 return instrprof_error::success;
393 void InstrProfRecord::updateStrings(InstrProfStringTable *StrTab) {
397 Name = StrTab->insertString(Name);
398 for (auto &VSite : IndirectCallSites)
399 for (auto &VData : VSite.ValueData)
400 VData.Value = (uint64_t)StrTab->insertString((const char *)VData.Value);
403 namespace IndexedInstrProf {
404 enum class HashT : uint32_t {
410 static inline uint64_t MD5Hash(StringRef Str) {
413 llvm::MD5::MD5Result Result;
415 // Return the least significant 8 bytes. Our MD5 implementation returns the
416 // result in little endian, so we may need to swap bytes.
417 using namespace llvm::support;
418 return endian::read<uint64_t, little, unaligned>(Result);
421 static inline uint64_t ComputeHash(HashT Type, StringRef K) {
424 return IndexedInstrProf::MD5Hash(K);
426 llvm_unreachable("Unhandled hash type");
429 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
430 const uint64_t Version = 3;
431 const HashT HashType = HashT::MD5;
433 // This structure defines the file header of the LLVM profile
434 // data file in indexed-format.
438 uint64_t MaxFunctionCount;
443 inline support::endianness getHostEndianness() {
444 return sys::IsLittleEndianHost ? support::little : support::big;
447 /// This is the header of the data structure that defines the on-disk
448 /// layout of the value profile data of a particular kind for one function.
449 struct ValueProfRecord {
450 // The kind of the value profile record.
452 // The number of value profile sites. It is guaranteed to be non-zero;
453 // otherwise the record for this kind won't be emitted.
454 uint32_t NumValueSites;
455 // The first element of the array that stores the number of profiled
456 // values for each value site. The size of the array is NumValueSites.
457 // Since NumValueSites is greater than zero, there is at least one
458 // element in the array.
459 uint8_t SiteCountArray[1];
461 // The fake declaration is for documentation purpose only.
462 // Align the start of next field to be on 8 byte boundaries.
463 // uint8_t Padding[X];
465 // The array of value profile data. The size of the array is the sum
466 // of all elements in SiteCountArray[].
467 // InstrProfValueData ValueData[];
469 /// Return the \c ValueProfRecord header size including the padding bytes.
470 static uint32_t getHeaderSize(uint32_t NumValueSites);
471 /// Return the total size of the value profile record including the
472 /// header and the value data.
473 static uint32_t getSize(uint32_t NumValueSites, uint32_t NumValueData);
474 /// Return the total size of the value profile record including the
475 /// header and the value data.
476 uint32_t getSize() const { return getSize(NumValueSites, getNumValueData()); }
477 /// Use this method to advance to the next \c ValueProfRecord.
478 ValueProfRecord *getNext();
479 /// Return the pointer to the first value profile data.
480 InstrProfValueData *getValueData();
481 /// Return the number of value sites.
482 uint32_t getNumValueSites() const { return NumValueSites; }
483 /// Return the number of value data.
484 uint32_t getNumValueData() const;
485 /// Read data from this record and save it to Record.
486 void deserializeTo(InstrProfRecord &Record,
487 InstrProfRecord::ValueMapType *VMap);
488 /// Extract data from \c Record and serialize into this instance.
489 void serializeFrom(const InstrProfRecord &Record, uint32_t ValueKind,
490 uint32_t NumValueSites);
491 /// In-place byte swap:
492 /// Do byte swap for this instance. \c Old is the original order before
493 /// the swap, and \c New is the New byte order.
494 void swapBytes(support::endianness Old, support::endianness New);
497 /// Per-function header/control data structure for value profiling
498 /// data in indexed format.
499 struct ValueProfData {
500 // Total size in bytes including this field. It must be a multiple
501 // of sizeof(uint64_t).
503 // The number of value profile kinds that has value profile data.
504 // In this implementation, a value profile kind is considered to
505 // have profile data if the number of value profile sites for the
506 // kind is not zero. More aggressively, the implemnetation can
507 // choose to check the actual data value: if none of the value sites
508 // has any profiled values, the kind can be skipped.
509 uint32_t NumValueKinds;
511 // Following are a sequence of variable length records. The prefix/header
512 // of each record is defined by ValueProfRecord type. The number of
513 // records is NumValueKinds.
514 // ValueProfRecord Record_1;
515 // ValueProfRecord Record_N;
517 /// Return the total size in bytes of the on-disk value profile data
518 /// given the data stored in Record.
519 static uint32_t getSize(const InstrProfRecord &Record);
520 /// Return a pointer to \c ValueProfData instance ready to be streamed.
521 static std::unique_ptr<ValueProfData>
522 serializeFrom(const InstrProfRecord &Record);
523 /// Return a pointer to \c ValueProfileData instance ready to be read.
524 /// All data in the instance are properly byte swapped. The input
525 /// data is assumed to be in little endian order.
526 static ErrorOr<std::unique_ptr<ValueProfData>>
527 getValueProfData(const unsigned char *D, const unsigned char *const BufferEnd,
528 support::endianness SrcDataEndianness);
529 /// Swap byte order from \c Endianness order to host byte order.
530 void swapBytesToHost(support::endianness Endianness);
531 /// Swap byte order from host byte order to \c Endianness order.
532 void swapBytesFromHost(support::endianness Endianness);
533 /// Return the total size of \c ValueProfileData.
534 uint32_t getSize() const { return TotalSize; }
535 /// Read data from this data and save it to \c Record.
536 void deserializeTo(InstrProfRecord &Record,
537 InstrProfRecord::ValueMapType *VMap);
538 /// Return the first \c ValueProfRecord instance.
539 ValueProfRecord *getFirstValueProfRecord();
542 } // end namespace IndexedInstrProf
544 namespace RawInstrProf {
546 const uint64_t Version = 1;
548 // Magic number to detect file format and endianness.
549 // Use 255 at one end, since no UTF-8 file can use that character. Avoid 0,
550 // so that utilities, like strings, don't grab it as a string. 129 is also
551 // invalid UTF-8, and high enough to be interesting.
552 // Use "lprofr" in the centre to stand for "LLVM Profile Raw", or "lprofR"
553 // for 32-bit platforms.
554 // The magic and version need to be kept in sync with
555 // projects/compiler-rt/lib/profile/InstrProfiling.c
557 template <class IntPtrT>
558 inline uint64_t getMagic();
560 inline uint64_t getMagic<uint64_t>() {
561 return uint64_t(255) << 56 | uint64_t('l') << 48 | uint64_t('p') << 40 |
562 uint64_t('r') << 32 | uint64_t('o') << 24 | uint64_t('f') << 16 |
563 uint64_t('r') << 8 | uint64_t(129);
567 inline uint64_t getMagic<uint32_t>() {
568 return uint64_t(255) << 56 | uint64_t('l') << 48 | uint64_t('p') << 40 |
569 uint64_t('r') << 32 | uint64_t('o') << 24 | uint64_t('f') << 16 |
570 uint64_t('R') << 8 | uint64_t(129);
573 // Per-function profile data header/control structure.
574 // The definition should match the structure defined in
575 // compiler-rt/lib/profile/InstrProfiling.h.
576 // It should also match the synthesized type in
577 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
578 template <class IntPtrT> struct ProfileData {
579 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
580 #include "llvm/ProfileData/InstrProfData.inc"
583 // File header structure of the LLVM profile data in raw format.
584 // The definition should match the header referenced in
585 // compiler-rt/lib/profile/InstrProfilingFile.c and
586 // InstrProfilingBuffer.c.
588 const uint64_t Magic;
589 const uint64_t Version;
590 const uint64_t DataSize;
591 const uint64_t CountersSize;
592 const uint64_t NamesSize;
593 const uint64_t CountersDelta;
594 const uint64_t NamesDelta;
597 } // end namespace RawInstrProf
601 // Profile coverage map has the following layout:
602 // [CoverageMapFileHeader]
604 // [CovMapFunctionRecord]
605 // [CovMapFunctionRecord]
608 // [Encoded Region Mapping Data]
610 template <class IntPtrT> struct CovMapFunctionRecord {
611 #define COVMAP_FUNC_RECORD(Type, LLVMType, Name, Init) Type Name;
612 #include "llvm/ProfileData/InstrProfData.inc"
618 } // end namespace llvm
622 struct is_error_code_enum<llvm::instrprof_error> : std::true_type {};
625 #endif // LLVM_PROFILEDATA_INSTRPROF_H_