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 {
156 unsupported_hash_type,
164 value_site_count_mismatch
167 inline std::error_code make_error_code(instrprof_error E) {
168 return std::error_code(static_cast<int>(E), instrprof_category());
171 enum InstrProfValueKind : uint32_t {
172 #define VALUE_PROF_KIND(Enumerator, Value) Enumerator = Value,
173 #include "llvm/ProfileData/InstrProfData.inc"
176 struct InstrProfStringTable {
177 // Set of string values in profiling data.
178 StringSet<> StringValueSet;
179 InstrProfStringTable() { StringValueSet.clear(); }
180 // Get a pointer to internal storage of a string in set
181 const char *getStringData(StringRef Str) {
182 auto Result = StringValueSet.find(Str);
183 return (Result == StringValueSet.end()) ? nullptr : Result->first().data();
185 // Insert a string to StringTable
186 const char *insertString(StringRef Str) {
187 auto Result = StringValueSet.insert(Str);
188 return Result.first->first().data();
192 struct InstrProfValueData {
195 // Number of times the value appears in the training run.
199 struct InstrProfValueSiteRecord {
200 /// Value profiling data pairs at a given value site.
201 std::list<InstrProfValueData> ValueData;
203 InstrProfValueSiteRecord() { ValueData.clear(); }
204 template <class InputIterator>
205 InstrProfValueSiteRecord(InputIterator F, InputIterator L)
208 /// Sort ValueData ascending by Value
209 void sortByTargetValues() {
211 [](const InstrProfValueData &left, const InstrProfValueData &right) {
212 return left.Value < right.Value;
216 /// Merge data from another InstrProfValueSiteRecord
217 void mergeValueData(InstrProfValueSiteRecord &Input) {
218 this->sortByTargetValues();
219 Input.sortByTargetValues();
220 auto I = ValueData.begin();
221 auto IE = ValueData.end();
222 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
224 while (I != IE && I->Value < J->Value)
226 if (I != IE && I->Value == J->Value) {
227 I->Count = SaturatingAdd(I->Count, J->Count);
231 ValueData.insert(I, *J);
236 /// Profiling information for a single function.
237 struct InstrProfRecord {
239 InstrProfRecord(StringRef Name, uint64_t Hash, std::vector<uint64_t> Counts)
240 : Name(Name), Hash(Hash), Counts(std::move(Counts)) {}
243 std::vector<uint64_t> Counts;
245 typedef std::vector<std::pair<uint64_t, const char *>> ValueMapType;
247 /// Return the number of value profile kinds with non-zero number
248 /// of profile sites.
249 inline uint32_t getNumValueKinds() const;
250 /// Return the number of instrumented sites for ValueKind.
251 inline uint32_t getNumValueSites(uint32_t ValueKind) const;
252 /// Return the total number of ValueData for ValueKind.
253 inline uint32_t getNumValueData(uint32_t ValueKind) const;
254 /// Return the number of value data collected for ValueKind at profiling
256 inline uint32_t getNumValueDataForSite(uint32_t ValueKind,
257 uint32_t Site) const;
258 inline std::unique_ptr<InstrProfValueData[]>
259 getValueForSite(uint32_t ValueKind, uint32_t Site) const;
260 /// Reserve space for NumValueSites sites.
261 inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites);
262 /// Add ValueData for ValueKind at value Site.
263 inline void addValueData(uint32_t ValueKind, uint32_t Site,
264 InstrProfValueData *VData, uint32_t N,
265 ValueMapType *HashKeys);
267 /// Merge the counts in \p Other into this one.
268 inline instrprof_error merge(InstrProfRecord &Other);
270 /// Used by InstrProfWriter: update the value strings to commoned strings in
271 /// the writer instance.
272 inline void updateStrings(InstrProfStringTable *StrTab);
274 /// Clear value data entries
275 inline void clearValueData() {
276 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
277 getValueSitesForKind(Kind).clear();
281 std::vector<InstrProfValueSiteRecord> IndirectCallSites;
282 const std::vector<InstrProfValueSiteRecord> &
283 getValueSitesForKind(uint32_t ValueKind) const {
285 case IPVK_IndirectCallTarget:
286 return IndirectCallSites;
288 llvm_unreachable("Unknown value kind!");
290 return IndirectCallSites;
293 std::vector<InstrProfValueSiteRecord> &
294 getValueSitesForKind(uint32_t ValueKind) {
295 return const_cast<std::vector<InstrProfValueSiteRecord> &>(
296 const_cast<const InstrProfRecord *>(this)
297 ->getValueSitesForKind(ValueKind));
300 // Map indirect call target name hash to name string.
301 uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
302 ValueMapType *HashKeys) {
306 case IPVK_IndirectCallTarget: {
308 std::lower_bound(HashKeys->begin(), HashKeys->end(), Value,
309 [](const std::pair<uint64_t, const char *> &LHS,
310 uint64_t RHS) { return LHS.first < RHS; });
311 if (Result != HashKeys->end())
312 Value = (uint64_t)Result->second;
319 // Merge Value Profile data from Src record to this record for ValueKind.
320 instrprof_error mergeValueProfData(uint32_t ValueKind, InstrProfRecord &Src) {
321 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
322 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
323 if (ThisNumValueSites != OtherNumValueSites)
324 return instrprof_error::value_site_count_mismatch;
325 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
326 getValueSitesForKind(ValueKind);
327 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
328 Src.getValueSitesForKind(ValueKind);
329 for (uint32_t I = 0; I < ThisNumValueSites; I++)
330 ThisSiteRecords[I].mergeValueData(OtherSiteRecords[I]);
331 return instrprof_error::success;
335 uint32_t InstrProfRecord::getNumValueKinds() const {
336 uint32_t NumValueKinds = 0;
337 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
338 NumValueKinds += !(getValueSitesForKind(Kind).empty());
339 return NumValueKinds;
342 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
344 const std::vector<InstrProfValueSiteRecord> &SiteRecords =
345 getValueSitesForKind(ValueKind);
346 for (auto &SR : SiteRecords) {
347 N += SR.ValueData.size();
352 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
353 return getValueSitesForKind(ValueKind).size();
356 uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind,
357 uint32_t Site) const {
358 return getValueSitesForKind(ValueKind)[Site].ValueData.size();
361 std::unique_ptr<InstrProfValueData[]>
362 InstrProfRecord::getValueForSite(uint32_t ValueKind, uint32_t Site) const {
363 uint32_t N = getNumValueDataForSite(ValueKind, Site);
365 return std::unique_ptr<InstrProfValueData[]>(nullptr);
367 std::unique_ptr<InstrProfValueData[]> VD(new InstrProfValueData[N]);
369 for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) {
378 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
379 InstrProfValueData *VData, uint32_t N,
380 ValueMapType *HashKeys) {
381 for (uint32_t I = 0; I < N; I++) {
382 VData[I].Value = remapValue(VData[I].Value, ValueKind, HashKeys);
384 std::vector<InstrProfValueSiteRecord> &ValueSites =
385 getValueSitesForKind(ValueKind);
387 ValueSites.push_back(InstrProfValueSiteRecord());
389 ValueSites.emplace_back(VData, VData + N);
392 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
393 std::vector<InstrProfValueSiteRecord> &ValueSites =
394 getValueSitesForKind(ValueKind);
395 ValueSites.reserve(NumValueSites);
398 void InstrProfRecord::updateStrings(InstrProfStringTable *StrTab) {
402 Name = StrTab->insertString(Name);
403 for (auto &VSite : IndirectCallSites)
404 for (auto &VData : VSite.ValueData)
405 VData.Value = (uint64_t)StrTab->insertString((const char *)VData.Value);
408 instrprof_error InstrProfRecord::merge(InstrProfRecord &Other) {
409 // If the number of counters doesn't match we either have bad data
410 // or a hash collision.
411 if (Counts.size() != Other.Counts.size())
412 return instrprof_error::count_mismatch;
414 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
415 if (Counts[I] + Other.Counts[I] < Counts[I])
416 return instrprof_error::counter_overflow;
417 Counts[I] += Other.Counts[I];
420 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
421 instrprof_error result = mergeValueProfData(Kind, Other);
422 if (result != instrprof_error::success)
426 return instrprof_error::success;
429 inline support::endianness getHostEndianness() {
430 return sys::IsLittleEndianHost ? support::little : support::big;
433 /// This is the header of the data structure that defines the on-disk
434 /// layout of the value profile data of a particular kind for one function.
435 struct ValueProfRecord {
436 // The kind of the value profile record.
438 // The number of value profile sites. It is guaranteed to be non-zero;
439 // otherwise the record for this kind won't be emitted.
440 uint32_t NumValueSites;
441 // The first element of the array that stores the number of profiled
442 // values for each value site. The size of the array is NumValueSites.
443 // Since NumValueSites is greater than zero, there is at least one
444 // element in the array.
445 uint8_t SiteCountArray[1];
447 // The fake declaration is for documentation purpose only.
448 // Align the start of next field to be on 8 byte boundaries.
449 // uint8_t Padding[X];
451 // The array of value profile data. The size of the array is the sum
452 // of all elements in SiteCountArray[].
453 // InstrProfValueData ValueData[];
455 /// Return the \c ValueProfRecord header size including the padding bytes.
456 static uint32_t getHeaderSize(uint32_t NumValueSites);
457 /// Return the total size of the value profile record including the
458 /// header and the value data.
459 static uint32_t getSize(uint32_t NumValueSites, uint32_t NumValueData);
460 /// Return the total size of the value profile record including the
461 /// header and the value data.
462 uint32_t getSize() const { return getSize(NumValueSites, getNumValueData()); }
463 /// Use this method to advance to the next \c ValueProfRecord.
464 ValueProfRecord *getNext();
465 /// Return the pointer to the first value profile data.
466 InstrProfValueData *getValueData();
467 /// Return the number of value sites.
468 uint32_t getNumValueSites() const { return NumValueSites; }
469 /// Return the number of value data.
470 uint32_t getNumValueData() const;
471 /// Read data from this record and save it to Record.
472 void deserializeTo(InstrProfRecord &Record,
473 InstrProfRecord::ValueMapType *VMap);
474 /// Extract data from \c Record and serialize into this instance.
475 void serializeFrom(const InstrProfRecord &Record, uint32_t ValueKind,
476 uint32_t NumValueSites);
477 /// In-place byte swap:
478 /// Do byte swap for this instance. \c Old is the original order before
479 /// the swap, and \c New is the New byte order.
480 void swapBytes(support::endianness Old, support::endianness New);
483 /// Per-function header/control data structure for value profiling
484 /// data in indexed format.
485 struct ValueProfData {
486 // Total size in bytes including this field. It must be a multiple
487 // of sizeof(uint64_t).
489 // The number of value profile kinds that has value profile data.
490 // In this implementation, a value profile kind is considered to
491 // have profile data if the number of value profile sites for the
492 // kind is not zero. More aggressively, the implementation can
493 // choose to check the actual data value: if none of the value sites
494 // has any profiled values, the kind can be skipped.
495 uint32_t NumValueKinds;
497 // Following are a sequence of variable length records. The prefix/header
498 // of each record is defined by ValueProfRecord type. The number of
499 // records is NumValueKinds.
500 // ValueProfRecord Record_1;
501 // ValueProfRecord Record_N;
503 /// Return the total size in bytes of the on-disk value profile data
504 /// given the data stored in Record.
505 static uint32_t getSize(const InstrProfRecord &Record);
506 /// Return a pointer to \c ValueProfData instance ready to be streamed.
507 static std::unique_ptr<ValueProfData>
508 serializeFrom(const InstrProfRecord &Record);
509 /// Return a pointer to \c ValueProfileData instance ready to be read.
510 /// All data in the instance are properly byte swapped. The input
511 /// data is assumed to be in little endian order.
512 static ErrorOr<std::unique_ptr<ValueProfData>>
513 getValueProfData(const unsigned char *D, const unsigned char *const BufferEnd,
514 support::endianness SrcDataEndianness);
515 /// Swap byte order from \c Endianness order to host byte order.
516 void swapBytesToHost(support::endianness Endianness);
517 /// Swap byte order from host byte order to \c Endianness order.
518 void swapBytesFromHost(support::endianness Endianness);
519 /// Return the total size of \c ValueProfileData.
520 uint32_t getSize() const { return TotalSize; }
521 /// Read data from this data and save it to \c Record.
522 void deserializeTo(InstrProfRecord &Record,
523 InstrProfRecord::ValueMapType *VMap);
524 /// Return the first \c ValueProfRecord instance.
525 ValueProfRecord *getFirstValueProfRecord();
528 namespace IndexedInstrProf {
530 enum class HashT : uint32_t {
536 static inline uint64_t MD5Hash(StringRef Str) {
539 llvm::MD5::MD5Result Result;
541 // Return the least significant 8 bytes. Our MD5 implementation returns the
542 // result in little endian, so we may need to swap bytes.
543 using namespace llvm::support;
544 return endian::read<uint64_t, little, unaligned>(Result);
547 static inline uint64_t ComputeHash(HashT Type, StringRef K) {
550 return IndexedInstrProf::MD5Hash(K);
552 llvm_unreachable("Unhandled hash type");
555 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
556 const uint64_t Version = 3;
557 const HashT HashType = HashT::MD5;
559 // This structure defines the file header of the LLVM profile
560 // data file in indexed-format.
564 uint64_t MaxFunctionCount;
569 } // end namespace IndexedInstrProf
571 namespace RawInstrProf {
573 const uint64_t Version = 2;
575 // Magic number to detect file format and endianness.
576 // Use 255 at one end, since no UTF-8 file can use that character. Avoid 0,
577 // so that utilities, like strings, don't grab it as a string. 129 is also
578 // invalid UTF-8, and high enough to be interesting.
579 // Use "lprofr" in the centre to stand for "LLVM Profile Raw", or "lprofR"
580 // for 32-bit platforms.
581 // The magic and version need to be kept in sync with
582 // projects/compiler-rt/lib/profile/InstrProfiling.c
584 template <class IntPtrT>
585 inline uint64_t getMagic();
587 inline uint64_t getMagic<uint64_t>() {
588 return uint64_t(255) << 56 | uint64_t('l') << 48 | uint64_t('p') << 40 |
589 uint64_t('r') << 32 | uint64_t('o') << 24 | uint64_t('f') << 16 |
590 uint64_t('r') << 8 | uint64_t(129);
594 inline uint64_t getMagic<uint32_t>() {
595 return uint64_t(255) << 56 | uint64_t('l') << 48 | uint64_t('p') << 40 |
596 uint64_t('r') << 32 | uint64_t('o') << 24 | uint64_t('f') << 16 |
597 uint64_t('R') << 8 | uint64_t(129);
600 // Per-function profile data header/control structure.
601 // The definition should match the structure defined in
602 // compiler-rt/lib/profile/InstrProfiling.h.
603 // It should also match the synthesized type in
604 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
605 template <class IntPtrT> struct LLVM_ALIGNAS(8) ProfileData {
606 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
607 #include "llvm/ProfileData/InstrProfData.inc"
610 // File header structure of the LLVM profile data in raw format.
611 // The definition should match the header referenced in
612 // compiler-rt/lib/profile/InstrProfilingFile.c and
613 // InstrProfilingBuffer.c.
615 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) Type Name;
616 #include "llvm/ProfileData/InstrProfData.inc"
619 } // end namespace RawInstrProf
623 // Profile coverage map has the following layout:
624 // [CoverageMapFileHeader]
626 // [CovMapFunctionRecord]
627 // [CovMapFunctionRecord]
630 // [Encoded Region Mapping Data]
632 template <class IntPtrT> struct CovMapFunctionRecord {
633 #define COVMAP_FUNC_RECORD(Type, LLVMType, Name, Init) Type Name;
634 #include "llvm/ProfileData/InstrProfData.inc"
640 } // end namespace llvm
644 struct is_error_code_enum<llvm::instrprof_error> : std::true_type {};
647 #endif // LLVM_PROFILEDATA_INSTRPROF_H_