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"
206 namespace IndexedInstrProf {
207 uint64_t ComputeHash(StringRef K);
210 /// A symbol table used for function PGO name look-up with keys
211 /// (such as pointers, md5hash values) to the function. A function's
212 /// PGO name or name's md5hash are used in retrieving the profile
213 /// data of the function. See \c getPGOFuncName() method for details
214 /// on how PGO name is formed.
215 class InstrProfSymtab {
219 std::vector<std::pair<uint64_t, std::string>> HashNameMap;
222 InstrProfSymtab() : Data(), Address(0) {}
224 /// Create InstrProfSymtab from a object file section which
225 /// contains function PGO names that are uncompressed.
226 std::error_code create(object::SectionRef &Section);
227 std::error_code create(StringRef D, uint64_t BaseAddr) {
230 return std::error_code();
232 template <typename NameIterRange> void create(NameIterRange &IterRange) {
233 for (auto Name : IterRange)
234 HashNameMap.push_back(
235 std::make_pair(IndexedInstrProf::ComputeHash(Name), Name.str()));
239 // If the symtab is created by a series calls to \c addFuncName, \c
240 // finalizeSymtab needs to
241 // be called before function name/symbol lookup using MD5 hash. This is
243 // the underlying map is vector (for space efficiency) which needs to be
245 void finalizeSymtab() {
246 std::sort(HashNameMap.begin(), HashNameMap.end(), less_first());
247 HashNameMap.erase(std::unique(HashNameMap.begin(), HashNameMap.end()),
251 void addFuncName(StringRef FuncName) {
252 HashNameMap.push_back(std::make_pair(
253 IndexedInstrProf::ComputeHash(FuncName), FuncName.str()));
256 /// Return function's PGO name from the function name's symabol
257 /// address in the object file. If an error occurs, Return
259 StringRef getFuncName(uint64_t FuncNameAddress, size_t NameSize);
260 /// Return function's PGO name from the name's md5 hash value.
261 /// If not found, return an empty string.
262 StringRef getFuncName(uint64_t FuncMD5Hash) {
264 std::lower_bound(HashNameMap.begin(), HashNameMap.end(), FuncMD5Hash,
265 [](const std::pair<uint64_t, std::string> &LHS,
266 uint64_t RHS) { return LHS.first < RHS; });
267 if (Result != HashNameMap.end())
268 return Result->second;
273 struct InstrProfStringTable {
274 // Set of string values in profiling data.
275 StringSet<> StringValueSet;
276 InstrProfStringTable() { StringValueSet.clear(); }
277 // Get a pointer to internal storage of a string in set
278 const char *getStringData(StringRef Str) {
279 auto Result = StringValueSet.find(Str);
280 return (Result == StringValueSet.end()) ? nullptr : Result->first().data();
282 // Insert a string to StringTable
283 const char *insertString(StringRef Str) {
284 auto Result = StringValueSet.insert(Str);
285 return Result.first->first().data();
289 struct InstrProfValueSiteRecord {
290 /// Value profiling data pairs at a given value site.
291 std::list<InstrProfValueData> ValueData;
293 InstrProfValueSiteRecord() { ValueData.clear(); }
294 template <class InputIterator>
295 InstrProfValueSiteRecord(InputIterator F, InputIterator L)
298 /// Sort ValueData ascending by Value
299 void sortByTargetValues() {
301 [](const InstrProfValueData &left, const InstrProfValueData &right) {
302 return left.Value < right.Value;
306 /// Merge data from another InstrProfValueSiteRecord
307 /// Optionally scale merged counts by \p Weight.
308 instrprof_error mergeValueData(InstrProfValueSiteRecord &Input,
309 uint64_t Weight = 1) {
310 this->sortByTargetValues();
311 Input.sortByTargetValues();
312 auto I = ValueData.begin();
313 auto IE = ValueData.end();
314 instrprof_error Result = instrprof_error::success;
315 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
317 while (I != IE && I->Value < J->Value)
319 if (I != IE && I->Value == J->Value) {
320 uint64_t JCount = J->Count;
323 JCount = SaturatingMultiply(JCount, Weight, &Overflowed);
325 Result = instrprof_error::counter_overflow;
327 I->Count = SaturatingAdd(I->Count, JCount, &Overflowed);
329 Result = instrprof_error::counter_overflow;
333 ValueData.insert(I, *J);
339 /// Profiling information for a single function.
340 struct InstrProfRecord {
342 InstrProfRecord(StringRef Name, uint64_t Hash, std::vector<uint64_t> Counts)
343 : Name(Name), Hash(Hash), Counts(std::move(Counts)) {}
346 std::vector<uint64_t> Counts;
348 typedef std::vector<std::pair<uint64_t, const char *>> ValueMapType;
350 /// Return the number of value profile kinds with non-zero number
351 /// of profile sites.
352 inline uint32_t getNumValueKinds() const;
353 /// Return the number of instrumented sites for ValueKind.
354 inline uint32_t getNumValueSites(uint32_t ValueKind) const;
355 /// Return the total number of ValueData for ValueKind.
356 inline uint32_t getNumValueData(uint32_t ValueKind) const;
357 /// Return the number of value data collected for ValueKind at profiling
359 inline uint32_t getNumValueDataForSite(uint32_t ValueKind,
360 uint32_t Site) const;
361 /// Return the array of profiled values at \p Site.
362 inline std::unique_ptr<InstrProfValueData[]>
363 getValueForSite(uint32_t ValueKind, uint32_t Site,
364 uint64_t (*ValueMapper)(uint32_t, uint64_t) = 0) const;
366 getValueForSite(InstrProfValueData Dest[], uint32_t ValueKind, uint32_t Site,
367 uint64_t (*ValueMapper)(uint32_t, uint64_t) = 0) const;
368 /// Reserve space for NumValueSites sites.
369 inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites);
370 /// Add ValueData for ValueKind at value Site.
371 void addValueData(uint32_t ValueKind, uint32_t Site,
372 InstrProfValueData *VData, uint32_t N,
373 ValueMapType *HashKeys);
375 /// Merge the counts in \p Other into this one.
376 /// Optionally scale merged counts by \p Weight.
377 instrprof_error merge(InstrProfRecord &Other, uint64_t Weight = 1);
379 /// Used by InstrProfWriter: update the value strings to commoned strings in
380 /// the writer instance.
381 void updateStrings(InstrProfStringTable *StrTab);
383 /// Clear value data entries
384 void clearValueData() {
385 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
386 getValueSitesForKind(Kind).clear();
390 std::vector<InstrProfValueSiteRecord> IndirectCallSites;
391 const std::vector<InstrProfValueSiteRecord> &
392 getValueSitesForKind(uint32_t ValueKind) const {
394 case IPVK_IndirectCallTarget:
395 return IndirectCallSites;
397 llvm_unreachable("Unknown value kind!");
399 return IndirectCallSites;
402 std::vector<InstrProfValueSiteRecord> &
403 getValueSitesForKind(uint32_t ValueKind) {
404 return const_cast<std::vector<InstrProfValueSiteRecord> &>(
405 const_cast<const InstrProfRecord *>(this)
406 ->getValueSitesForKind(ValueKind));
409 // Map indirect call target name hash to name string.
410 uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
411 ValueMapType *HashKeys);
413 // Merge Value Profile data from Src record to this record for ValueKind.
414 // Scale merged value counts by \p Weight.
415 instrprof_error mergeValueProfData(uint32_t ValueKind, InstrProfRecord &Src,
419 uint32_t InstrProfRecord::getNumValueKinds() const {
420 uint32_t NumValueKinds = 0;
421 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
422 NumValueKinds += !(getValueSitesForKind(Kind).empty());
423 return NumValueKinds;
426 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
428 const std::vector<InstrProfValueSiteRecord> &SiteRecords =
429 getValueSitesForKind(ValueKind);
430 for (auto &SR : SiteRecords) {
431 N += SR.ValueData.size();
436 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
437 return getValueSitesForKind(ValueKind).size();
440 uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind,
441 uint32_t Site) const {
442 return getValueSitesForKind(ValueKind)[Site].ValueData.size();
445 std::unique_ptr<InstrProfValueData[]> InstrProfRecord::getValueForSite(
446 uint32_t ValueKind, uint32_t Site,
447 uint64_t (*ValueMapper)(uint32_t, uint64_t)) const {
448 uint32_t N = getNumValueDataForSite(ValueKind, Site);
450 return std::unique_ptr<InstrProfValueData[]>(nullptr);
452 auto VD = llvm::make_unique<InstrProfValueData[]>(N);
453 getValueForSite(VD.get(), ValueKind, Site, ValueMapper);
458 void InstrProfRecord::getValueForSite(InstrProfValueData Dest[],
459 uint32_t ValueKind, uint32_t Site,
460 uint64_t (*ValueMapper)(uint32_t,
463 for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) {
464 Dest[I].Value = ValueMapper ? ValueMapper(ValueKind, V.Value) : V.Value;
465 Dest[I].Count = V.Count;
470 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
471 std::vector<InstrProfValueSiteRecord> &ValueSites =
472 getValueSitesForKind(ValueKind);
473 ValueSites.reserve(NumValueSites);
476 inline support::endianness getHostEndianness() {
477 return sys::IsLittleEndianHost ? support::little : support::big;
480 // Include definitions for value profile data
481 #define INSTR_PROF_VALUE_PROF_DATA
482 #include "llvm/ProfileData/InstrProfData.inc"
485 * Initialize the record for runtime value profile data.
486 * Return 0 if the initialization is successful, otherwise
489 int initializeValueProfRuntimeRecord(ValueProfRuntimeRecord *RuntimeRecord,
490 const uint16_t *NumValueSites,
491 ValueProfNode **Nodes);
493 /* Release memory allocated for the runtime record. */
494 void finalizeValueProfRuntimeRecord(ValueProfRuntimeRecord *RuntimeRecord);
496 /* Return the size of ValueProfData structure that can be used to store
497 the value profile data collected at runtime. */
498 uint32_t getValueProfDataSizeRT(const ValueProfRuntimeRecord *Record);
500 /* Return a ValueProfData instance that stores the data collected at runtime. */
502 serializeValueProfDataFromRT(const ValueProfRuntimeRecord *Record,
505 namespace IndexedInstrProf {
507 enum class HashT : uint32_t {
513 static inline uint64_t MD5Hash(StringRef Str) {
516 llvm::MD5::MD5Result Result;
518 // Return the least significant 8 bytes. Our MD5 implementation returns the
519 // result in little endian, so we may need to swap bytes.
520 using namespace llvm::support;
521 return endian::read<uint64_t, little, unaligned>(Result);
524 inline uint64_t ComputeHash(HashT Type, StringRef K) {
527 return IndexedInstrProf::MD5Hash(K);
529 llvm_unreachable("Unhandled hash type");
532 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
533 const uint64_t Version = INSTR_PROF_INDEX_VERSION;
534 const HashT HashType = HashT::MD5;
536 inline uint64_t ComputeHash(StringRef K) { return ComputeHash(HashType, K); }
538 // This structure defines the file header of the LLVM profile
539 // data file in indexed-format.
543 uint64_t MaxFunctionCount;
548 } // end namespace IndexedInstrProf
550 namespace RawInstrProf {
552 const uint64_t Version = INSTR_PROF_RAW_VERSION;
554 template <class IntPtrT> inline uint64_t getMagic();
555 template <> inline uint64_t getMagic<uint64_t>() {
556 return INSTR_PROF_RAW_MAGIC_64;
559 template <> inline uint64_t getMagic<uint32_t>() {
560 return INSTR_PROF_RAW_MAGIC_32;
563 // Per-function profile data header/control structure.
564 // The definition should match the structure defined in
565 // compiler-rt/lib/profile/InstrProfiling.h.
566 // It should also match the synthesized type in
567 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
568 template <class IntPtrT> struct LLVM_ALIGNAS(8) ProfileData {
569 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
570 #include "llvm/ProfileData/InstrProfData.inc"
573 // File header structure of the LLVM profile data in raw format.
574 // The definition should match the header referenced in
575 // compiler-rt/lib/profile/InstrProfilingFile.c and
576 // InstrProfilingBuffer.c.
578 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) const Type Name;
579 #include "llvm/ProfileData/InstrProfData.inc"
582 } // end namespace RawInstrProf
586 // Profile coverage map has the following layout:
587 // [CoverageMapFileHeader]
589 // [CovMapFunctionRecord]
590 // [CovMapFunctionRecord]
593 // [Encoded Region Mapping Data]
595 template <class IntPtrT> struct CovMapFunctionRecord {
596 #define COVMAP_FUNC_RECORD(Type, LLVMType, Name, Init) Type Name;
597 #include "llvm/ProfileData/InstrProfData.inc"
603 } // end namespace llvm
607 struct is_error_code_enum<llvm::instrprof_error> : std::true_type {};
610 #endif // LLVM_PROFILEDATA_INSTRPROF_H_