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/ProfileData/InstrProfData.inc"
23 #include "llvm/Support/Endian.h"
24 #include "llvm/Support/ErrorHandling.h"
25 #include "llvm/Support/ErrorOr.h"
26 #include "llvm/Support/MD5.h"
29 #include <system_error>
38 /// Return the name of data section containing profile counter variables.
39 inline StringRef getInstrProfCountersSectionName(bool AddSegment) {
40 return AddSegment ? "__DATA,__llvm_prf_cnts" : "__llvm_prf_cnts";
43 /// Return the name of data section containing names of instrumented
45 inline StringRef getInstrProfNameSectionName(bool AddSegment) {
46 return AddSegment ? "__DATA,__llvm_prf_names" : "__llvm_prf_names";
49 /// Return the name of the data section containing per-function control
51 inline StringRef getInstrProfDataSectionName(bool AddSegment) {
52 return AddSegment ? "__DATA,__llvm_prf_data" : "__llvm_prf_data";
55 /// Return the name of the section containing function coverage mapping
57 inline StringRef getInstrProfCoverageSectionName(bool AddSegment) {
58 return AddSegment ? "__DATA,__llvm_covmap" : "__llvm_covmap";
61 /// Return the name prefix of variables containing instrumented function names.
62 inline StringRef getInstrProfNameVarPrefix() { return "__llvm_profile_name_"; }
64 /// Return the name prefix of variables containing per-function control data.
65 inline StringRef getInstrProfDataVarPrefix() { return "__llvm_profile_data_"; }
67 /// Return the name prefix of profile counter variables.
68 inline StringRef getInstrProfCountersVarPrefix() {
69 return "__llvm_profile_counters_";
72 /// Return the name prefix of the COMDAT group for instrumentation variables
73 /// associated with a COMDAT function.
74 inline StringRef getInstrProfComdatPrefix() { return "__llvm_profile_vars_"; }
76 /// Return the name of a covarage mapping variable (internal linkage)
77 /// for each instrumented source module. Such variables are allocated
78 /// in the __llvm_covmap section.
79 inline StringRef getCoverageMappingVarName() {
80 return "__llvm_coverage_mapping";
83 /// Return the name of function that registers all the per-function control
84 /// data at program startup time by calling __llvm_register_function. This
85 /// function has internal linkage and is called by __llvm_profile_init
86 /// runtime method. This function is not generated for these platforms:
87 /// Darwin, Linux, and FreeBSD.
88 inline StringRef getInstrProfRegFuncsName() {
89 return "__llvm_profile_register_functions";
92 /// Return the name of the runtime interface that registers per-function control
93 /// data for one instrumented function.
94 inline StringRef getInstrProfRegFuncName() {
95 return "__llvm_profile_register_function";
98 /// Return the name of the runtime initialization method that is generated by
99 /// the compiler. The function calls __llvm_profile_register_functions and
100 /// __llvm_profile_override_default_filename functions if needed. This function
101 /// has internal linkage and invoked at startup time via init_array.
102 inline StringRef getInstrProfInitFuncName() { return "__llvm_profile_init"; }
104 /// Return the name of the hook variable defined in profile runtime library.
105 /// A reference to the variable causes the linker to link in the runtime
106 /// initialization module (which defines the hook variable).
107 inline StringRef getInstrProfRuntimeHookVarName() {
108 return "__llvm_profile_runtime";
111 /// Return the name of the compiler generated function that references the
112 /// runtime hook variable. The function is a weak global.
113 inline StringRef getInstrProfRuntimeHookVarUseFuncName() {
114 return "__llvm_profile_runtime_user";
117 /// Return the name of the profile runtime interface that overrides the default
118 /// profile data file name.
119 inline StringRef getInstrProfFileOverriderFuncName() {
120 return "__llvm_profile_override_default_filename";
123 /// Return the modified name for function \c F suitable to be
124 /// used the key for profile lookup.
125 std::string getPGOFuncName(const Function &F);
127 /// Return the modified name for a function suitable to be
128 /// used the key for profile lookup. The function's original
129 /// name is \c RawFuncName and has linkage of type \c Linkage.
130 /// The function is defined in module \c FileName.
131 std::string getPGOFuncName(StringRef RawFuncName,
132 GlobalValue::LinkageTypes Linkage,
135 /// Create and return the global variable for function name used in PGO
136 /// instrumentation. \c FuncName is the name of the function returned
137 /// by \c getPGOFuncName call.
138 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef FuncName);
140 /// Create and return the global variable for function name used in PGO
141 /// instrumentation. /// \c FuncName is the name of the function
142 /// returned by \c getPGOFuncName call, \c M is the owning module,
143 /// and \c Linkage is the linkage of the instrumented function.
144 GlobalVariable *createPGOFuncNameVar(Module &M,
145 GlobalValue::LinkageTypes Linkage,
148 const std::error_category &instrprof_category();
150 enum class instrprof_error {
157 unsupported_hash_type,
165 value_site_count_mismatch
168 inline std::error_code make_error_code(instrprof_error E) {
169 return std::error_code(static_cast<int>(E), instrprof_category());
172 enum InstrProfValueKind : uint32_t {
173 #define VALUE_PROF_KIND(Enumerator, Value) Enumerator = Value,
174 #include "llvm/ProfileData/InstrProfData.inc"
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 = SaturatingAdd(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);
268 /// Merge the counts in \p Other into this one.
269 inline instrprof_error merge(InstrProfRecord &Other);
271 /// Used by InstrProfWriter: update the value strings to commoned strings in
272 /// the writer instance.
273 inline void updateStrings(InstrProfStringTable *StrTab);
275 /// Clear value data entries
276 inline void clearValueData() {
277 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
278 getValueSitesForKind(Kind).clear();
282 std::vector<InstrProfValueSiteRecord> IndirectCallSites;
283 const std::vector<InstrProfValueSiteRecord> &
284 getValueSitesForKind(uint32_t ValueKind) const {
286 case IPVK_IndirectCallTarget:
287 return IndirectCallSites;
289 llvm_unreachable("Unknown value kind!");
291 return IndirectCallSites;
294 std::vector<InstrProfValueSiteRecord> &
295 getValueSitesForKind(uint32_t ValueKind) {
296 return const_cast<std::vector<InstrProfValueSiteRecord> &>(
297 const_cast<const InstrProfRecord *>(this)
298 ->getValueSitesForKind(ValueKind));
301 // Map indirect call target name hash to name string.
302 uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
303 ValueMapType *HashKeys) {
307 case IPVK_IndirectCallTarget: {
309 std::lower_bound(HashKeys->begin(), HashKeys->end(), Value,
310 [](const std::pair<uint64_t, const char *> &LHS,
311 uint64_t RHS) { return LHS.first < RHS; });
312 if (Result != HashKeys->end())
313 Value = (uint64_t)Result->second;
320 // Merge Value Profile data from Src record to this record for ValueKind.
321 instrprof_error mergeValueProfData(uint32_t ValueKind, InstrProfRecord &Src) {
322 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
323 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
324 if (ThisNumValueSites != OtherNumValueSites)
325 return instrprof_error::value_site_count_mismatch;
326 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
327 getValueSitesForKind(ValueKind);
328 std::vector<InstrProfValueSiteRecord> &OtherSiteRecords =
329 Src.getValueSitesForKind(ValueKind);
330 for (uint32_t I = 0; I < ThisNumValueSites; I++)
331 ThisSiteRecords[I].mergeValueData(OtherSiteRecords[I]);
332 return instrprof_error::success;
336 uint32_t InstrProfRecord::getNumValueKinds() const {
337 uint32_t NumValueKinds = 0;
338 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
339 NumValueKinds += !(getValueSitesForKind(Kind).empty());
340 return NumValueKinds;
343 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
345 const std::vector<InstrProfValueSiteRecord> &SiteRecords =
346 getValueSitesForKind(ValueKind);
347 for (auto &SR : SiteRecords) {
348 N += SR.ValueData.size();
353 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
354 return getValueSitesForKind(ValueKind).size();
357 uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind,
358 uint32_t Site) const {
359 return getValueSitesForKind(ValueKind)[Site].ValueData.size();
362 std::unique_ptr<InstrProfValueData[]>
363 InstrProfRecord::getValueForSite(uint32_t ValueKind, uint32_t Site) const {
364 uint32_t N = getNumValueDataForSite(ValueKind, Site);
366 return std::unique_ptr<InstrProfValueData[]>(nullptr);
368 std::unique_ptr<InstrProfValueData[]> VD(new InstrProfValueData[N]);
370 for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) {
379 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
380 InstrProfValueData *VData, uint32_t N,
381 ValueMapType *HashKeys) {
382 for (uint32_t I = 0; I < N; I++) {
383 VData[I].Value = remapValue(VData[I].Value, ValueKind, HashKeys);
385 std::vector<InstrProfValueSiteRecord> &ValueSites =
386 getValueSitesForKind(ValueKind);
388 ValueSites.push_back(InstrProfValueSiteRecord());
390 ValueSites.emplace_back(VData, VData + N);
393 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
394 std::vector<InstrProfValueSiteRecord> &ValueSites =
395 getValueSitesForKind(ValueKind);
396 ValueSites.reserve(NumValueSites);
399 void InstrProfRecord::updateStrings(InstrProfStringTable *StrTab) {
403 Name = StrTab->insertString(Name);
404 for (auto &VSite : IndirectCallSites)
405 for (auto &VData : VSite.ValueData)
406 VData.Value = (uint64_t)StrTab->insertString((const char *)VData.Value);
409 instrprof_error InstrProfRecord::merge(InstrProfRecord &Other) {
410 // If the number of counters doesn't match we either have bad data
411 // or a hash collision.
412 if (Counts.size() != Other.Counts.size())
413 return instrprof_error::count_mismatch;
415 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
416 if (Counts[I] + Other.Counts[I] < Counts[I])
417 return instrprof_error::counter_overflow;
418 Counts[I] += Other.Counts[I];
421 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
422 instrprof_error result = mergeValueProfData(Kind, Other);
423 if (result != instrprof_error::success)
427 return instrprof_error::success;
430 inline support::endianness getHostEndianness() {
431 return sys::IsLittleEndianHost ? support::little : support::big;
434 /// This is the header of the data structure that defines the on-disk
435 /// layout of the value profile data of a particular kind for one function.
436 struct ValueProfRecord {
437 // The kind of the value profile record.
439 // The number of value profile sites. It is guaranteed to be non-zero;
440 // otherwise the record for this kind won't be emitted.
441 uint32_t NumValueSites;
442 // The first element of the array that stores the number of profiled
443 // values for each value site. The size of the array is NumValueSites.
444 // Since NumValueSites is greater than zero, there is at least one
445 // element in the array.
446 uint8_t SiteCountArray[1];
448 // The fake declaration is for documentation purpose only.
449 // Align the start of next field to be on 8 byte boundaries.
450 // uint8_t Padding[X];
452 // The array of value profile data. The size of the array is the sum
453 // of all elements in SiteCountArray[].
454 // InstrProfValueData ValueData[];
456 /// Return the \c ValueProfRecord header size including the padding bytes.
457 static uint32_t getHeaderSize(uint32_t NumValueSites);
458 /// Return the total size of the value profile record including the
459 /// header and the value data.
460 static uint32_t getSize(uint32_t NumValueSites, uint32_t NumValueData);
461 /// Return the total size of the value profile record including the
462 /// header and the value data.
463 uint32_t getSize() const { return getSize(NumValueSites, getNumValueData()); }
464 /// Use this method to advance to the next \c ValueProfRecord.
465 ValueProfRecord *getNext();
466 /// Return the pointer to the first value profile data.
467 InstrProfValueData *getValueData();
468 /// Return the number of value sites.
469 uint32_t getNumValueSites() const { return NumValueSites; }
470 /// Return the number of value data.
471 uint32_t getNumValueData() const;
472 /// Read data from this record and save it to Record.
473 void deserializeTo(InstrProfRecord &Record,
474 InstrProfRecord::ValueMapType *VMap);
475 /// Extract data from \c Record and serialize into this instance.
476 void serializeFrom(const InstrProfRecord &Record, uint32_t ValueKind,
477 uint32_t NumValueSites);
478 /// In-place byte swap:
479 /// Do byte swap for this instance. \c Old is the original order before
480 /// the swap, and \c New is the New byte order.
481 void swapBytes(support::endianness Old, support::endianness New);
484 /// Per-function header/control data structure for value profiling
485 /// data in indexed format.
486 struct ValueProfData {
487 // Total size in bytes including this field. It must be a multiple
488 // of sizeof(uint64_t).
490 // The number of value profile kinds that has value profile data.
491 // In this implementation, a value profile kind is considered to
492 // have profile data if the number of value profile sites for the
493 // kind is not zero. More aggressively, the implementation can
494 // choose to check the actual data value: if none of the value sites
495 // has any profiled values, the kind can be skipped.
496 uint32_t NumValueKinds;
498 // Following are a sequence of variable length records. The prefix/header
499 // of each record is defined by ValueProfRecord type. The number of
500 // records is NumValueKinds.
501 // ValueProfRecord Record_1;
502 // ValueProfRecord Record_N;
504 /// Return the total size in bytes of the on-disk value profile data
505 /// given the data stored in Record.
506 static uint32_t getSize(const InstrProfRecord &Record);
507 /// Return a pointer to \c ValueProfData instance ready to be streamed.
508 static std::unique_ptr<ValueProfData>
509 serializeFrom(const InstrProfRecord &Record);
510 /// Return a pointer to \c ValueProfileData instance ready to be read.
511 /// All data in the instance are properly byte swapped. The input
512 /// data is assumed to be in little endian order.
513 static ErrorOr<std::unique_ptr<ValueProfData>>
514 getValueProfData(const unsigned char *D, const unsigned char *const BufferEnd,
515 support::endianness SrcDataEndianness);
516 /// Swap byte order from \c Endianness order to host byte order.
517 void swapBytesToHost(support::endianness Endianness);
518 /// Swap byte order from host byte order to \c Endianness order.
519 void swapBytesFromHost(support::endianness Endianness);
520 /// Return the total size of \c ValueProfileData.
521 uint32_t getSize() const { return TotalSize; }
522 /// Read data from this data and save it to \c Record.
523 void deserializeTo(InstrProfRecord &Record,
524 InstrProfRecord::ValueMapType *VMap);
525 /// Return the first \c ValueProfRecord instance.
526 ValueProfRecord *getFirstValueProfRecord();
529 namespace IndexedInstrProf {
531 enum class HashT : uint32_t {
537 static inline uint64_t MD5Hash(StringRef Str) {
540 llvm::MD5::MD5Result Result;
542 // Return the least significant 8 bytes. Our MD5 implementation returns the
543 // result in little endian, so we may need to swap bytes.
544 using namespace llvm::support;
545 return endian::read<uint64_t, little, unaligned>(Result);
548 static inline uint64_t ComputeHash(HashT Type, StringRef K) {
551 return IndexedInstrProf::MD5Hash(K);
553 llvm_unreachable("Unhandled hash type");
556 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
557 const uint64_t Version = 3;
558 const HashT HashType = HashT::MD5;
560 // This structure defines the file header of the LLVM profile
561 // data file in indexed-format.
565 uint64_t MaxFunctionCount;
570 } // end namespace IndexedInstrProf
572 namespace RawInstrProf {
574 const uint64_t Version = INSTR_PROF_RAW_VERSION;
576 template <class IntPtrT> inline uint64_t getMagic();
577 template <> inline uint64_t getMagic<uint64_t>() {
578 return INSTR_PROF_RAW_MAGIC_64;
581 template <> inline uint64_t getMagic<uint32_t>() {
582 return INSTR_PROF_RAW_MAGIC_32;
585 // Per-function profile data header/control structure.
586 // The definition should match the structure defined in
587 // compiler-rt/lib/profile/InstrProfiling.h.
588 // It should also match the synthesized type in
589 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
590 template <class IntPtrT> struct LLVM_ALIGNAS(8) ProfileData {
591 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
592 #include "llvm/ProfileData/InstrProfData.inc"
595 // File header structure of the LLVM profile data in raw format.
596 // The definition should match the header referenced in
597 // compiler-rt/lib/profile/InstrProfilingFile.c and
598 // InstrProfilingBuffer.c.
600 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) Type Name;
601 #include "llvm/ProfileData/InstrProfData.inc"
604 } // end namespace RawInstrProf
608 // Profile coverage map has the following layout:
609 // [CoverageMapFileHeader]
611 // [CovMapFunctionRecord]
612 // [CovMapFunctionRecord]
615 // [Encoded Region Mapping Data]
617 template <class IntPtrT> struct CovMapFunctionRecord {
618 #define COVMAP_FUNC_RECORD(Type, LLVMType, Name, Init) Type Name;
619 #include "llvm/ProfileData/InstrProfData.inc"
625 } // end namespace llvm
629 struct is_error_code_enum<llvm::instrprof_error> : std::true_type {};
632 #endif // LLVM_PROFILEDATA_INSTRPROF_H_