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 /// Given a vector of strings (function PGO names) \c NameStrs, the
164 /// method generates a combined string \c Result thatis ready to be
165 /// serialized. The \c Result string is comprised of three fields:
166 /// The first field is the legnth of the uncompressed strings, and the
167 /// the second field is the length of the zlib-compressed string.
168 /// Both fields are encoded in ULEB128. If \c doCompress is false, the
169 /// third field is the uncompressed strings; otherwise it is the
170 /// compressed string. When the string compression is off, the
171 /// second field will have value zero.
172 int collectPGOFuncNameStrings(const std::vector<std::string> &NameStrs,
173 bool doCompression, std::string &Result);
174 /// Produce \c Result string with the same format described above. The input
175 /// is vector of PGO function name variables that are referenced.
176 int collectPGOFuncNameStrings(const std::vector<GlobalVariable *> &NameVars,
177 std::string &Result);
178 class InstrProfSymtab;
179 /// \c NameStrings is a string composed of one of more sub-strings encoded in
181 /// format described above. The substrings are seperated by 0 or more zero
183 /// This method decodes the string and populates the \c Symtab.
184 int readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab);
186 const std::error_category &instrprof_category();
188 enum class instrprof_error {
195 unsupported_hash_type,
203 value_site_count_mismatch
206 inline std::error_code make_error_code(instrprof_error E) {
207 return std::error_code(static_cast<int>(E), instrprof_category());
210 inline instrprof_error MergeResult(instrprof_error &Accumulator,
211 instrprof_error Result) {
212 // Prefer first error encountered as later errors may be secondary effects of
213 // the initial problem.
214 if (Accumulator == instrprof_error::success &&
215 Result != instrprof_error::success)
216 Accumulator = Result;
220 enum InstrProfValueKind : uint32_t {
221 #define VALUE_PROF_KIND(Enumerator, Value) Enumerator = Value,
222 #include "llvm/ProfileData/InstrProfData.inc"
229 namespace IndexedInstrProf {
230 uint64_t ComputeHash(StringRef K);
233 /// A symbol table used for function PGO name look-up with keys
234 /// (such as pointers, md5hash values) to the function. A function's
235 /// PGO name or name's md5hash are used in retrieving the profile
236 /// data of the function. See \c getPGOFuncName() method for details
237 /// on how PGO name is formed.
238 class InstrProfSymtab {
240 typedef std::vector<std::pair<uint64_t, uint64_t>> AddrHashMap;
245 // A map from MD5 hash keys to function name strings.
246 std::vector<std::pair<uint64_t, std::string>> HashNameMap;
247 // A map from function runtime address to function name MD5 hash.
248 // This map is only populated and used by raw instr profile reader.
249 AddrHashMap AddrToMD5Map;
252 InstrProfSymtab() : Data(), Address(0), HashNameMap(), AddrToMD5Map() {}
254 /// Create InstrProfSymtab from an object file section which
255 /// contains function PGO names that are uncompressed.
256 /// This interface is used by CoverageMappingReader.
257 std::error_code create(object::SectionRef &Section);
258 /// This interface is used by reader of CoverageMapping test
260 inline std::error_code create(StringRef D, uint64_t BaseAddr);
261 /// \c NameStrings is a string composed of one of more sub-strings
262 /// encoded in the format described above. The substrings are
263 /// seperated by 0 or more zero bytes. This method decodes the
264 /// string and populates the \c Symtab.
265 inline std::error_code create(StringRef NameStrings);
266 /// Create InstrProfSymtab from a set of names iteratable from
267 /// \p IterRange. This interface is used by IndexedProfReader.
268 template <typename NameIterRange> void create(const NameIterRange &IterRange);
269 // If the symtab is created by a series of calls to \c addFuncName, \c
270 // finalizeSymtab needs to be called before looking up function names.
271 // This is required because the underlying map is a vector (for space
272 // efficiency) which needs to be sorted.
273 inline void finalizeSymtab();
274 /// Update the symtab by adding \p FuncName to the table. This interface
275 /// is used by the raw and text profile readers.
276 void addFuncName(StringRef FuncName) {
277 HashNameMap.push_back(std::make_pair(
278 IndexedInstrProf::ComputeHash(FuncName), FuncName.str()));
280 /// Map a function address to its name's MD5 hash. This interface
281 /// is only used by the raw profiler reader.
282 void mapAddress(uint64_t Addr, uint64_t MD5Val) {
283 AddrToMD5Map.push_back(std::make_pair(Addr, MD5Val));
285 AddrHashMap &getAddrHashMap() { return AddrToMD5Map; }
286 /// Return function's PGO name from the function name's symbol
287 /// address in the object file. If an error occurs, return
289 StringRef getFuncName(uint64_t FuncNameAddress, size_t NameSize);
290 /// Return function's PGO name from the name's md5 hash value.
291 /// If not found, return an empty string.
292 inline StringRef getFuncName(uint64_t FuncMD5Hash);
295 std::error_code InstrProfSymtab::create(StringRef D, uint64_t BaseAddr) {
298 return std::error_code();
301 std::error_code InstrProfSymtab::create(StringRef NameStrings) {
302 if (readPGOFuncNameStrings(NameStrings, *this))
303 return make_error_code(instrprof_error::malformed);
304 return std::error_code();
307 template <typename NameIterRange>
308 void InstrProfSymtab::create(const NameIterRange &IterRange) {
309 for (auto Name : IterRange)
310 HashNameMap.push_back(
311 std::make_pair(IndexedInstrProf::ComputeHash(Name), Name.str()));
315 void InstrProfSymtab::finalizeSymtab() {
316 std::sort(HashNameMap.begin(), HashNameMap.end(), less_first());
317 HashNameMap.erase(std::unique(HashNameMap.begin(), HashNameMap.end()),
319 std::sort(AddrToMD5Map.begin(), AddrToMD5Map.end(), less_first());
320 AddrToMD5Map.erase(std::unique(AddrToMD5Map.begin(), AddrToMD5Map.end()),
324 StringRef InstrProfSymtab::getFuncName(uint64_t FuncMD5Hash) {
326 std::lower_bound(HashNameMap.begin(), HashNameMap.end(), FuncMD5Hash,
327 [](const std::pair<uint64_t, std::string> &LHS,
328 uint64_t RHS) { return LHS.first < RHS; });
329 if (Result != HashNameMap.end())
330 return Result->second;
334 struct InstrProfValueSiteRecord {
335 /// Value profiling data pairs at a given value site.
336 std::list<InstrProfValueData> ValueData;
338 InstrProfValueSiteRecord() { ValueData.clear(); }
339 template <class InputIterator>
340 InstrProfValueSiteRecord(InputIterator F, InputIterator L)
343 /// Sort ValueData ascending by Value
344 void sortByTargetValues() {
346 [](const InstrProfValueData &left, const InstrProfValueData &right) {
347 return left.Value < right.Value;
351 /// Merge data from another InstrProfValueSiteRecord
352 /// Optionally scale merged counts by \p Weight.
353 instrprof_error mergeValueData(InstrProfValueSiteRecord &Input,
354 uint64_t Weight = 1);
357 /// Profiling information for a single function.
358 struct InstrProfRecord {
360 InstrProfRecord(StringRef Name, uint64_t Hash, std::vector<uint64_t> Counts)
361 : Name(Name), Hash(Hash), Counts(std::move(Counts)) {}
364 std::vector<uint64_t> Counts;
366 typedef std::vector<std::pair<uint64_t, uint64_t>> ValueMapType;
368 /// Return the number of value profile kinds with non-zero number
369 /// of profile sites.
370 inline uint32_t getNumValueKinds() const;
371 /// Return the number of instrumented sites for ValueKind.
372 inline uint32_t getNumValueSites(uint32_t ValueKind) const;
373 /// Return the total number of ValueData for ValueKind.
374 inline uint32_t getNumValueData(uint32_t ValueKind) const;
375 /// Return the number of value data collected for ValueKind at profiling
377 inline uint32_t getNumValueDataForSite(uint32_t ValueKind,
378 uint32_t Site) const;
379 /// Return the array of profiled values at \p Site.
380 inline std::unique_ptr<InstrProfValueData[]>
381 getValueForSite(uint32_t ValueKind, uint32_t Site,
382 uint64_t (*ValueMapper)(uint32_t, uint64_t) = 0) const;
384 getValueForSite(InstrProfValueData Dest[], uint32_t ValueKind, uint32_t Site,
385 uint64_t (*ValueMapper)(uint32_t, uint64_t) = 0) const;
386 /// Reserve space for NumValueSites sites.
387 inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites);
388 /// Add ValueData for ValueKind at value Site.
389 void addValueData(uint32_t ValueKind, uint32_t Site,
390 InstrProfValueData *VData, uint32_t N,
391 ValueMapType *ValueMap);
393 /// Merge the counts in \p Other into this one.
394 /// Optionally scale merged counts by \p Weight.
395 instrprof_error merge(InstrProfRecord &Other, uint64_t Weight = 1);
397 /// Clear value data entries
398 void clearValueData() {
399 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
400 getValueSitesForKind(Kind).clear();
404 std::vector<InstrProfValueSiteRecord> IndirectCallSites;
405 const std::vector<InstrProfValueSiteRecord> &
406 getValueSitesForKind(uint32_t ValueKind) const {
408 case IPVK_IndirectCallTarget:
409 return IndirectCallSites;
411 llvm_unreachable("Unknown value kind!");
413 return IndirectCallSites;
416 std::vector<InstrProfValueSiteRecord> &
417 getValueSitesForKind(uint32_t ValueKind) {
418 return const_cast<std::vector<InstrProfValueSiteRecord> &>(
419 const_cast<const InstrProfRecord *>(this)
420 ->getValueSitesForKind(ValueKind));
423 // Map indirect call target name hash to name string.
424 uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
425 ValueMapType *HashKeys);
427 // Merge Value Profile data from Src record to this record for ValueKind.
428 // Scale merged value counts by \p Weight.
429 instrprof_error mergeValueProfData(uint32_t ValueKind, InstrProfRecord &Src,
433 uint32_t InstrProfRecord::getNumValueKinds() const {
434 uint32_t NumValueKinds = 0;
435 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
436 NumValueKinds += !(getValueSitesForKind(Kind).empty());
437 return NumValueKinds;
440 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
442 const std::vector<InstrProfValueSiteRecord> &SiteRecords =
443 getValueSitesForKind(ValueKind);
444 for (auto &SR : SiteRecords) {
445 N += SR.ValueData.size();
450 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
451 return getValueSitesForKind(ValueKind).size();
454 uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind,
455 uint32_t Site) const {
456 return getValueSitesForKind(ValueKind)[Site].ValueData.size();
459 std::unique_ptr<InstrProfValueData[]> InstrProfRecord::getValueForSite(
460 uint32_t ValueKind, uint32_t Site,
461 uint64_t (*ValueMapper)(uint32_t, uint64_t)) const {
462 uint32_t N = getNumValueDataForSite(ValueKind, Site);
464 return std::unique_ptr<InstrProfValueData[]>(nullptr);
466 auto VD = llvm::make_unique<InstrProfValueData[]>(N);
467 getValueForSite(VD.get(), ValueKind, Site, ValueMapper);
472 void InstrProfRecord::getValueForSite(InstrProfValueData Dest[],
473 uint32_t ValueKind, uint32_t Site,
474 uint64_t (*ValueMapper)(uint32_t,
477 for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) {
478 Dest[I].Value = ValueMapper ? ValueMapper(ValueKind, V.Value) : V.Value;
479 Dest[I].Count = V.Count;
484 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
485 std::vector<InstrProfValueSiteRecord> &ValueSites =
486 getValueSitesForKind(ValueKind);
487 ValueSites.reserve(NumValueSites);
490 inline support::endianness getHostEndianness() {
491 return sys::IsLittleEndianHost ? support::little : support::big;
494 // Include definitions for value profile data
495 #define INSTR_PROF_VALUE_PROF_DATA
496 #include "llvm/ProfileData/InstrProfData.inc"
499 * Initialize the record for runtime value profile data.
500 * Return 0 if the initialization is successful, otherwise
503 int initializeValueProfRuntimeRecord(ValueProfRuntimeRecord *RuntimeRecord,
504 const uint16_t *NumValueSites,
505 ValueProfNode **Nodes);
507 /* Release memory allocated for the runtime record. */
508 void finalizeValueProfRuntimeRecord(ValueProfRuntimeRecord *RuntimeRecord);
510 /* Return the size of ValueProfData structure that can be used to store
511 the value profile data collected at runtime. */
512 uint32_t getValueProfDataSizeRT(const ValueProfRuntimeRecord *Record);
514 /* Return a ValueProfData instance that stores the data collected at runtime. */
516 serializeValueProfDataFromRT(const ValueProfRuntimeRecord *Record,
519 namespace IndexedInstrProf {
521 enum class HashT : uint32_t {
527 static inline uint64_t MD5Hash(StringRef Str) {
530 llvm::MD5::MD5Result Result;
532 // Return the least significant 8 bytes. Our MD5 implementation returns the
533 // result in little endian, so we may need to swap bytes.
534 using namespace llvm::support;
535 return endian::read<uint64_t, little, unaligned>(Result);
538 inline uint64_t ComputeHash(HashT Type, StringRef K) {
541 return IndexedInstrProf::MD5Hash(K);
543 llvm_unreachable("Unhandled hash type");
546 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
547 const uint64_t Version = INSTR_PROF_INDEX_VERSION;
548 const HashT HashType = HashT::MD5;
550 inline uint64_t ComputeHash(StringRef K) { return ComputeHash(HashType, K); }
552 // This structure defines the file header of the LLVM profile
553 // data file in indexed-format.
557 uint64_t MaxFunctionCount;
562 } // end namespace IndexedInstrProf
564 namespace RawInstrProf {
566 const uint64_t Version = INSTR_PROF_RAW_VERSION;
568 template <class IntPtrT> inline uint64_t getMagic();
569 template <> inline uint64_t getMagic<uint64_t>() {
570 return INSTR_PROF_RAW_MAGIC_64;
573 template <> inline uint64_t getMagic<uint32_t>() {
574 return INSTR_PROF_RAW_MAGIC_32;
577 // Per-function profile data header/control structure.
578 // The definition should match the structure defined in
579 // compiler-rt/lib/profile/InstrProfiling.h.
580 // It should also match the synthesized type in
581 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
582 template <class IntPtrT> struct LLVM_ALIGNAS(8) ProfileData {
583 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
584 #include "llvm/ProfileData/InstrProfData.inc"
587 // File header structure of the LLVM profile data in raw format.
588 // The definition should match the header referenced in
589 // compiler-rt/lib/profile/InstrProfilingFile.c and
590 // InstrProfilingBuffer.c.
592 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) const Type Name;
593 #include "llvm/ProfileData/InstrProfData.inc"
596 } // end namespace RawInstrProf
600 // Profile coverage map has the following layout:
601 // [CoverageMapFileHeader]
603 // [CovMapFunctionRecord]
604 // [CovMapFunctionRecord]
607 // [Encoded Region Mapping Data]
609 template <class IntPtrT> struct CovMapFunctionRecord {
610 #define COVMAP_FUNC_RECORD(Type, LLVMType, Name, Init) Type Name;
611 #include "llvm/ProfileData/InstrProfData.inc"
617 } // end namespace llvm
621 struct is_error_code_enum<llvm::instrprof_error> : std::true_type {};
624 #endif // LLVM_PROFILEDATA_INSTRPROF_H_