Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 1 | //===- InstrProf.h - Instrumented profiling format support ------*- C++ -*-===// |
| 2 | // |
||
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
||
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
||
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
||
| 6 | // |
||
| 7 | //===----------------------------------------------------------------------===// |
||
| 8 | // |
||
| 9 | // Instrumentation-based profiling data is generated by instrumented |
||
| 10 | // binaries through library functions in compiler-rt, and read by the clang |
||
| 11 | // frontend to feed PGO. |
||
| 12 | // |
||
| 13 | //===----------------------------------------------------------------------===// |
||
| 14 | |||
| 15 | #ifndef LLVM_PROFILEDATA_INSTRPROF_H |
||
| 16 | #define LLVM_PROFILEDATA_INSTRPROF_H |
||
| 17 | |||
| 18 | #include "llvm/ADT/ArrayRef.h" |
||
| 19 | #include "llvm/ADT/BitmaskEnum.h" |
||
| 20 | #include "llvm/ADT/STLExtras.h" |
||
| 21 | #include "llvm/ADT/StringRef.h" |
||
| 22 | #include "llvm/ADT/StringSet.h" |
||
| 23 | #include "llvm/ADT/Triple.h" |
||
| 24 | #include "llvm/IR/GlobalValue.h" |
||
| 25 | #include "llvm/IR/ProfileSummary.h" |
||
| 26 | #include "llvm/ProfileData/InstrProfData.inc" |
||
| 27 | #include "llvm/Support/CommandLine.h" |
||
| 28 | #include "llvm/Support/Compiler.h" |
||
| 29 | #include "llvm/Support/Endian.h" |
||
| 30 | #include "llvm/Support/Error.h" |
||
| 31 | #include "llvm/Support/ErrorHandling.h" |
||
| 32 | #include "llvm/Support/Host.h" |
||
| 33 | #include "llvm/Support/MD5.h" |
||
| 34 | #include "llvm/Support/MathExtras.h" |
||
| 35 | #include "llvm/Support/raw_ostream.h" |
||
| 36 | #include <algorithm> |
||
| 37 | #include <cassert> |
||
| 38 | #include <cstddef> |
||
| 39 | #include <cstdint> |
||
| 40 | #include <cstring> |
||
| 41 | #include <list> |
||
| 42 | #include <memory> |
||
| 43 | #include <string> |
||
| 44 | #include <system_error> |
||
| 45 | #include <utility> |
||
| 46 | #include <vector> |
||
| 47 | |||
| 48 | namespace llvm { |
||
| 49 | |||
| 50 | class Function; |
||
| 51 | class GlobalVariable; |
||
| 52 | struct InstrProfRecord; |
||
| 53 | class InstrProfSymtab; |
||
| 54 | class Instruction; |
||
| 55 | class MDNode; |
||
| 56 | class Module; |
||
| 57 | |||
| 58 | enum InstrProfSectKind { |
||
| 59 | #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) Kind, |
||
| 60 | #include "llvm/ProfileData/InstrProfData.inc" |
||
| 61 | }; |
||
| 62 | |||
| 63 | /// Return the max count value. We reserver a few large values for special use. |
||
| 64 | inline uint64_t getInstrMaxCountValue() { |
||
| 65 | return std::numeric_limits<uint64_t>::max() - 2; |
||
| 66 | } |
||
| 67 | |||
| 68 | /// Return the name of the profile section corresponding to \p IPSK. |
||
| 69 | /// |
||
| 70 | /// The name of the section depends on the object format type \p OF. If |
||
| 71 | /// \p AddSegmentInfo is true, a segment prefix and additional linker hints may |
||
| 72 | /// be added to the section name (this is the default). |
||
| 73 | std::string getInstrProfSectionName(InstrProfSectKind IPSK, |
||
| 74 | Triple::ObjectFormatType OF, |
||
| 75 | bool AddSegmentInfo = true); |
||
| 76 | |||
| 77 | /// Return the name profile runtime entry point to do value profiling |
||
| 78 | /// for a given site. |
||
| 79 | inline StringRef getInstrProfValueProfFuncName() { |
||
| 80 | return INSTR_PROF_VALUE_PROF_FUNC_STR; |
||
| 81 | } |
||
| 82 | |||
| 83 | /// Return the name profile runtime entry point to do memop size value |
||
| 84 | /// profiling. |
||
| 85 | inline StringRef getInstrProfValueProfMemOpFuncName() { |
||
| 86 | return INSTR_PROF_VALUE_PROF_MEMOP_FUNC_STR; |
||
| 87 | } |
||
| 88 | |||
| 89 | /// Return the name prefix of variables containing instrumented function names. |
||
| 90 | inline StringRef getInstrProfNameVarPrefix() { return "__profn_"; } |
||
| 91 | |||
| 92 | /// Return the name prefix of variables containing per-function control data. |
||
| 93 | inline StringRef getInstrProfDataVarPrefix() { return "__profd_"; } |
||
| 94 | |||
| 95 | /// Return the name prefix of profile counter variables. |
||
| 96 | inline StringRef getInstrProfCountersVarPrefix() { return "__profc_"; } |
||
| 97 | |||
| 98 | /// Return the name prefix of value profile variables. |
||
| 99 | inline StringRef getInstrProfValuesVarPrefix() { return "__profvp_"; } |
||
| 100 | |||
| 101 | /// Return the name of value profile node array variables: |
||
| 102 | inline StringRef getInstrProfVNodesVarName() { return "__llvm_prf_vnodes"; } |
||
| 103 | |||
| 104 | /// Return the name of the variable holding the strings (possibly compressed) |
||
| 105 | /// of all function's PGO names. |
||
| 106 | inline StringRef getInstrProfNamesVarName() { |
||
| 107 | return "__llvm_prf_nm"; |
||
| 108 | } |
||
| 109 | |||
| 110 | /// Return the name of a covarage mapping variable (internal linkage) |
||
| 111 | /// for each instrumented source module. Such variables are allocated |
||
| 112 | /// in the __llvm_covmap section. |
||
| 113 | inline StringRef getCoverageMappingVarName() { |
||
| 114 | return "__llvm_coverage_mapping"; |
||
| 115 | } |
||
| 116 | |||
| 117 | /// Return the name of the internal variable recording the array |
||
| 118 | /// of PGO name vars referenced by the coverage mapping. The owning |
||
| 119 | /// functions of those names are not emitted by FE (e.g, unused inline |
||
| 120 | /// functions.) |
||
| 121 | inline StringRef getCoverageUnusedNamesVarName() { |
||
| 122 | return "__llvm_coverage_names"; |
||
| 123 | } |
||
| 124 | |||
| 125 | /// Return the name of function that registers all the per-function control |
||
| 126 | /// data at program startup time by calling __llvm_register_function. This |
||
| 127 | /// function has internal linkage and is called by __llvm_profile_init |
||
| 128 | /// runtime method. This function is not generated for these platforms: |
||
| 129 | /// Darwin, Linux, and FreeBSD. |
||
| 130 | inline StringRef getInstrProfRegFuncsName() { |
||
| 131 | return "__llvm_profile_register_functions"; |
||
| 132 | } |
||
| 133 | |||
| 134 | /// Return the name of the runtime interface that registers per-function control |
||
| 135 | /// data for one instrumented function. |
||
| 136 | inline StringRef getInstrProfRegFuncName() { |
||
| 137 | return "__llvm_profile_register_function"; |
||
| 138 | } |
||
| 139 | |||
| 140 | /// Return the name of the runtime interface that registers the PGO name strings. |
||
| 141 | inline StringRef getInstrProfNamesRegFuncName() { |
||
| 142 | return "__llvm_profile_register_names_function"; |
||
| 143 | } |
||
| 144 | |||
| 145 | /// Return the name of the runtime initialization method that is generated by |
||
| 146 | /// the compiler. The function calls __llvm_profile_register_functions and |
||
| 147 | /// __llvm_profile_override_default_filename functions if needed. This function |
||
| 148 | /// has internal linkage and invoked at startup time via init_array. |
||
| 149 | inline StringRef getInstrProfInitFuncName() { return "__llvm_profile_init"; } |
||
| 150 | |||
| 151 | /// Return the name of the hook variable defined in profile runtime library. |
||
| 152 | /// A reference to the variable causes the linker to link in the runtime |
||
| 153 | /// initialization module (which defines the hook variable). |
||
| 154 | inline StringRef getInstrProfRuntimeHookVarName() { |
||
| 155 | return INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_RUNTIME_VAR); |
||
| 156 | } |
||
| 157 | |||
| 158 | /// Return the name of the compiler generated function that references the |
||
| 159 | /// runtime hook variable. The function is a weak global. |
||
| 160 | inline StringRef getInstrProfRuntimeHookVarUseFuncName() { |
||
| 161 | return "__llvm_profile_runtime_user"; |
||
| 162 | } |
||
| 163 | |||
| 164 | inline StringRef getInstrProfCounterBiasVarName() { |
||
| 165 | return INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_COUNTER_BIAS_VAR); |
||
| 166 | } |
||
| 167 | |||
| 168 | /// Return the marker used to separate PGO names during serialization. |
||
| 169 | inline StringRef getInstrProfNameSeparator() { return "\01"; } |
||
| 170 | |||
| 171 | /// Return the modified name for function \c F suitable to be |
||
| 172 | /// used the key for profile lookup. Variable \c InLTO indicates if this |
||
| 173 | /// is called in LTO optimization passes. |
||
| 174 | std::string getPGOFuncName(const Function &F, bool InLTO = false, |
||
| 175 | uint64_t Version = INSTR_PROF_INDEX_VERSION); |
||
| 176 | |||
| 177 | /// Return the modified name for a function suitable to be |
||
| 178 | /// used the key for profile lookup. The function's original |
||
| 179 | /// name is \c RawFuncName and has linkage of type \c Linkage. |
||
| 180 | /// The function is defined in module \c FileName. |
||
| 181 | std::string getPGOFuncName(StringRef RawFuncName, |
||
| 182 | GlobalValue::LinkageTypes Linkage, |
||
| 183 | StringRef FileName, |
||
| 184 | uint64_t Version = INSTR_PROF_INDEX_VERSION); |
||
| 185 | |||
| 186 | /// Return the name of the global variable used to store a function |
||
| 187 | /// name in PGO instrumentation. \c FuncName is the name of the function |
||
| 188 | /// returned by the \c getPGOFuncName call. |
||
| 189 | std::string getPGOFuncNameVarName(StringRef FuncName, |
||
| 190 | GlobalValue::LinkageTypes Linkage); |
||
| 191 | |||
| 192 | /// Create and return the global variable for function name used in PGO |
||
| 193 | /// instrumentation. \c FuncName is the name of the function returned |
||
| 194 | /// by \c getPGOFuncName call. |
||
| 195 | GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName); |
||
| 196 | |||
| 197 | /// Create and return the global variable for function name used in PGO |
||
| 198 | /// instrumentation. /// \c FuncName is the name of the function |
||
| 199 | /// returned by \c getPGOFuncName call, \c M is the owning module, |
||
| 200 | /// and \c Linkage is the linkage of the instrumented function. |
||
| 201 | GlobalVariable *createPGOFuncNameVar(Module &M, |
||
| 202 | GlobalValue::LinkageTypes Linkage, |
||
| 203 | StringRef PGOFuncName); |
||
| 204 | |||
| 205 | /// Return the initializer in string of the PGO name var \c NameVar. |
||
| 206 | StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar); |
||
| 207 | |||
| 208 | /// Given a PGO function name, remove the filename prefix and return |
||
| 209 | /// the original (static) function name. |
||
| 210 | StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, |
||
| 211 | StringRef FileName = "<unknown>"); |
||
| 212 | |||
| 213 | /// Given a vector of strings (function PGO names) \c NameStrs, the |
||
| 214 | /// method generates a combined string \c Result that is ready to be |
||
| 215 | /// serialized. The \c Result string is comprised of three fields: |
||
| 216 | /// The first field is the length of the uncompressed strings, and the |
||
| 217 | /// the second field is the length of the zlib-compressed string. |
||
| 218 | /// Both fields are encoded in ULEB128. If \c doCompress is false, the |
||
| 219 | /// third field is the uncompressed strings; otherwise it is the |
||
| 220 | /// compressed string. When the string compression is off, the |
||
| 221 | /// second field will have value zero. |
||
| 222 | Error collectPGOFuncNameStrings(ArrayRef<std::string> NameStrs, |
||
| 223 | bool doCompression, std::string &Result); |
||
| 224 | |||
| 225 | /// Produce \c Result string with the same format described above. The input |
||
| 226 | /// is vector of PGO function name variables that are referenced. |
||
| 227 | Error collectPGOFuncNameStrings(ArrayRef<GlobalVariable *> NameVars, |
||
| 228 | std::string &Result, bool doCompression = true); |
||
| 229 | |||
| 230 | /// \c NameStrings is a string composed of one of more sub-strings encoded in |
||
| 231 | /// the format described above. The substrings are separated by 0 or more zero |
||
| 232 | /// bytes. This method decodes the string and populates the \c Symtab. |
||
| 233 | Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab); |
||
| 234 | |||
| 235 | /// Check if INSTR_PROF_RAW_VERSION_VAR is defined. This global is only being |
||
| 236 | /// set in IR PGO compilation. |
||
| 237 | bool isIRPGOFlagSet(const Module *M); |
||
| 238 | |||
| 239 | /// Check if we can safely rename this Comdat function. Instances of the same |
||
| 240 | /// comdat function may have different control flows thus can not share the |
||
| 241 | /// same counter variable. |
||
| 242 | bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken = false); |
||
| 243 | |||
| 244 | enum InstrProfValueKind : uint32_t { |
||
| 245 | #define VALUE_PROF_KIND(Enumerator, Value, Descr) Enumerator = Value, |
||
| 246 | #include "llvm/ProfileData/InstrProfData.inc" |
||
| 247 | }; |
||
| 248 | |||
| 249 | /// Get the value profile data for value site \p SiteIdx from \p InstrProfR |
||
| 250 | /// and annotate the instruction \p Inst with the value profile meta data. |
||
| 251 | /// Annotate up to \p MaxMDCount (default 3) number of records per value site. |
||
| 252 | void annotateValueSite(Module &M, Instruction &Inst, |
||
| 253 | const InstrProfRecord &InstrProfR, |
||
| 254 | InstrProfValueKind ValueKind, uint32_t SiteIndx, |
||
| 255 | uint32_t MaxMDCount = 3); |
||
| 256 | |||
| 257 | /// Same as the above interface but using an ArrayRef, as well as \p Sum. |
||
| 258 | void annotateValueSite(Module &M, Instruction &Inst, |
||
| 259 | ArrayRef<InstrProfValueData> VDs, uint64_t Sum, |
||
| 260 | InstrProfValueKind ValueKind, uint32_t MaxMDCount); |
||
| 261 | |||
| 262 | /// Extract the value profile data from \p Inst which is annotated with |
||
| 263 | /// value profile meta data. Return false if there is no value data annotated, |
||
| 264 | /// otherwise return true. |
||
| 265 | bool getValueProfDataFromInst(const Instruction &Inst, |
||
| 266 | InstrProfValueKind ValueKind, |
||
| 267 | uint32_t MaxNumValueData, |
||
| 268 | InstrProfValueData ValueData[], |
||
| 269 | uint32_t &ActualNumValueData, uint64_t &TotalC, |
||
| 270 | bool GetNoICPValue = false); |
||
| 271 | |||
| 272 | inline StringRef getPGOFuncNameMetadataName() { return "PGOFuncName"; } |
||
| 273 | |||
| 274 | /// Return the PGOFuncName meta data associated with a function. |
||
| 275 | MDNode *getPGOFuncNameMetadata(const Function &F); |
||
| 276 | |||
| 277 | /// Create the PGOFuncName meta data if PGOFuncName is different from |
||
| 278 | /// function's raw name. This should only apply to internal linkage functions |
||
| 279 | /// declared by users only. |
||
| 280 | void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName); |
||
| 281 | |||
| 282 | /// Check if we can use Comdat for profile variables. This will eliminate |
||
| 283 | /// the duplicated profile variables for Comdat functions. |
||
| 284 | bool needsComdatForCounter(const Function &F, const Module &M); |
||
| 285 | |||
| 286 | /// An enum describing the attributes of an instrumented profile. |
||
| 287 | enum class InstrProfKind { |
||
| 288 | Unknown = 0x0, |
||
| 289 | // A frontend clang profile, incompatible with other attrs. |
||
| 290 | FrontendInstrumentation = 0x1, |
||
| 291 | // An IR-level profile (default when -fprofile-generate is used). |
||
| 292 | IRInstrumentation = 0x2, |
||
| 293 | // A profile with entry basic block instrumentation. |
||
| 294 | FunctionEntryInstrumentation = 0x4, |
||
| 295 | // A context sensitive IR-level profile. |
||
| 296 | ContextSensitive = 0x8, |
||
| 297 | // Use single byte probes for coverage. |
||
| 298 | SingleByteCoverage = 0x10, |
||
| 299 | // Only instrument the function entry basic block. |
||
| 300 | FunctionEntryOnly = 0x20, |
||
| 301 | // A memory profile collected using -fprofile=memory. |
||
| 302 | MemProf = 0x40, |
||
| 303 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/MemProf) |
||
| 304 | }; |
||
| 305 | |||
| 306 | const std::error_category &instrprof_category(); |
||
| 307 | |||
| 308 | enum class instrprof_error { |
||
| 309 | success = 0, |
||
| 310 | eof, |
||
| 311 | unrecognized_format, |
||
| 312 | bad_magic, |
||
| 313 | bad_header, |
||
| 314 | unsupported_version, |
||
| 315 | unsupported_hash_type, |
||
| 316 | too_large, |
||
| 317 | truncated, |
||
| 318 | malformed, |
||
| 319 | missing_debug_info_for_correlation, |
||
| 320 | unexpected_debug_info_for_correlation, |
||
| 321 | unable_to_correlate_profile, |
||
| 322 | unknown_function, |
||
| 323 | invalid_prof, |
||
| 324 | hash_mismatch, |
||
| 325 | count_mismatch, |
||
| 326 | counter_overflow, |
||
| 327 | value_site_count_mismatch, |
||
| 328 | compress_failed, |
||
| 329 | uncompress_failed, |
||
| 330 | empty_raw_profile, |
||
| 331 | zlib_unavailable |
||
| 332 | }; |
||
| 333 | |||
| 334 | inline std::error_code make_error_code(instrprof_error E) { |
||
| 335 | return std::error_code(static_cast<int>(E), instrprof_category()); |
||
| 336 | } |
||
| 337 | |||
| 338 | class InstrProfError : public ErrorInfo<InstrProfError> { |
||
| 339 | public: |
||
| 340 | InstrProfError(instrprof_error Err, const Twine &ErrStr = Twine()) |
||
| 341 | : Err(Err), Msg(ErrStr.str()) { |
||
| 342 | assert(Err != instrprof_error::success && "Not an error"); |
||
| 343 | } |
||
| 344 | |||
| 345 | std::string message() const override; |
||
| 346 | |||
| 347 | void log(raw_ostream &OS) const override { OS << message(); } |
||
| 348 | |||
| 349 | std::error_code convertToErrorCode() const override { |
||
| 350 | return make_error_code(Err); |
||
| 351 | } |
||
| 352 | |||
| 353 | instrprof_error get() const { return Err; } |
||
| 354 | const std::string &getMessage() const { return Msg; } |
||
| 355 | |||
| 356 | /// Consume an Error and return the raw enum value contained within it. The |
||
| 357 | /// Error must either be a success value, or contain a single InstrProfError. |
||
| 358 | static instrprof_error take(Error E) { |
||
| 359 | auto Err = instrprof_error::success; |
||
| 360 | handleAllErrors(std::move(E), [&Err](const InstrProfError &IPE) { |
||
| 361 | assert(Err == instrprof_error::success && "Multiple errors encountered"); |
||
| 362 | Err = IPE.get(); |
||
| 363 | }); |
||
| 364 | return Err; |
||
| 365 | } |
||
| 366 | |||
| 367 | static char ID; |
||
| 368 | |||
| 369 | private: |
||
| 370 | instrprof_error Err; |
||
| 371 | std::string Msg; |
||
| 372 | }; |
||
| 373 | |||
| 374 | class SoftInstrProfErrors { |
||
| 375 | /// Count the number of soft instrprof_errors encountered and keep track of |
||
| 376 | /// the first such error for reporting purposes. |
||
| 377 | |||
| 378 | /// The first soft error encountered. |
||
| 379 | instrprof_error FirstError = instrprof_error::success; |
||
| 380 | |||
| 381 | /// The number of hash mismatches. |
||
| 382 | unsigned NumHashMismatches = 0; |
||
| 383 | |||
| 384 | /// The number of count mismatches. |
||
| 385 | unsigned NumCountMismatches = 0; |
||
| 386 | |||
| 387 | /// The number of counter overflows. |
||
| 388 | unsigned NumCounterOverflows = 0; |
||
| 389 | |||
| 390 | /// The number of value site count mismatches. |
||
| 391 | unsigned NumValueSiteCountMismatches = 0; |
||
| 392 | |||
| 393 | public: |
||
| 394 | SoftInstrProfErrors() = default; |
||
| 395 | |||
| 396 | ~SoftInstrProfErrors() { |
||
| 397 | assert(FirstError == instrprof_error::success && |
||
| 398 | "Unchecked soft error encountered"); |
||
| 399 | } |
||
| 400 | |||
| 401 | /// Track a soft error (\p IE) and increment its associated counter. |
||
| 402 | void addError(instrprof_error IE); |
||
| 403 | |||
| 404 | /// Get the number of hash mismatches. |
||
| 405 | unsigned getNumHashMismatches() const { return NumHashMismatches; } |
||
| 406 | |||
| 407 | /// Get the number of count mismatches. |
||
| 408 | unsigned getNumCountMismatches() const { return NumCountMismatches; } |
||
| 409 | |||
| 410 | /// Get the number of counter overflows. |
||
| 411 | unsigned getNumCounterOverflows() const { return NumCounterOverflows; } |
||
| 412 | |||
| 413 | /// Get the number of value site count mismatches. |
||
| 414 | unsigned getNumValueSiteCountMismatches() const { |
||
| 415 | return NumValueSiteCountMismatches; |
||
| 416 | } |
||
| 417 | |||
| 418 | /// Return the first encountered error and reset FirstError to a success |
||
| 419 | /// value. |
||
| 420 | Error takeError() { |
||
| 421 | if (FirstError == instrprof_error::success) |
||
| 422 | return Error::success(); |
||
| 423 | auto E = make_error<InstrProfError>(FirstError); |
||
| 424 | FirstError = instrprof_error::success; |
||
| 425 | return E; |
||
| 426 | } |
||
| 427 | }; |
||
| 428 | |||
| 429 | namespace object { |
||
| 430 | |||
| 431 | class SectionRef; |
||
| 432 | |||
| 433 | } // end namespace object |
||
| 434 | |||
| 435 | namespace IndexedInstrProf { |
||
| 436 | |||
| 437 | uint64_t ComputeHash(StringRef K); |
||
| 438 | |||
| 439 | } // end namespace IndexedInstrProf |
||
| 440 | |||
| 441 | /// A symbol table used for function PGO name look-up with keys |
||
| 442 | /// (such as pointers, md5hash values) to the function. A function's |
||
| 443 | /// PGO name or name's md5hash are used in retrieving the profile |
||
| 444 | /// data of the function. See \c getPGOFuncName() method for details |
||
| 445 | /// on how PGO name is formed. |
||
| 446 | class InstrProfSymtab { |
||
| 447 | public: |
||
| 448 | using AddrHashMap = std::vector<std::pair<uint64_t, uint64_t>>; |
||
| 449 | |||
| 450 | private: |
||
| 451 | StringRef Data; |
||
| 452 | uint64_t Address = 0; |
||
| 453 | // Unique name strings. |
||
| 454 | StringSet<> NameTab; |
||
| 455 | // A map from MD5 keys to function name strings. |
||
| 456 | std::vector<std::pair<uint64_t, StringRef>> MD5NameMap; |
||
| 457 | // A map from MD5 keys to function define. We only populate this map |
||
| 458 | // when build the Symtab from a Module. |
||
| 459 | std::vector<std::pair<uint64_t, Function *>> MD5FuncMap; |
||
| 460 | // A map from function runtime address to function name MD5 hash. |
||
| 461 | // This map is only populated and used by raw instr profile reader. |
||
| 462 | AddrHashMap AddrToMD5Map; |
||
| 463 | bool Sorted = false; |
||
| 464 | |||
| 465 | static StringRef getExternalSymbol() { |
||
| 466 | return "** External Symbol **"; |
||
| 467 | } |
||
| 468 | |||
| 469 | // If the symtab is created by a series of calls to \c addFuncName, \c |
||
| 470 | // finalizeSymtab needs to be called before looking up function names. |
||
| 471 | // This is required because the underlying map is a vector (for space |
||
| 472 | // efficiency) which needs to be sorted. |
||
| 473 | inline void finalizeSymtab(); |
||
| 474 | |||
| 475 | public: |
||
| 476 | InstrProfSymtab() = default; |
||
| 477 | |||
| 478 | /// Create InstrProfSymtab from an object file section which |
||
| 479 | /// contains function PGO names. When section may contain raw |
||
| 480 | /// string data or string data in compressed form. This method |
||
| 481 | /// only initialize the symtab with reference to the data and |
||
| 482 | /// the section base address. The decompression will be delayed |
||
| 483 | /// until before it is used. See also \c create(StringRef) method. |
||
| 484 | Error create(object::SectionRef &Section); |
||
| 485 | |||
| 486 | /// This interface is used by reader of CoverageMapping test |
||
| 487 | /// format. |
||
| 488 | inline Error create(StringRef D, uint64_t BaseAddr); |
||
| 489 | |||
| 490 | /// \c NameStrings is a string composed of one of more sub-strings |
||
| 491 | /// encoded in the format described in \c collectPGOFuncNameStrings. |
||
| 492 | /// This method is a wrapper to \c readPGOFuncNameStrings method. |
||
| 493 | inline Error create(StringRef NameStrings); |
||
| 494 | |||
| 495 | /// A wrapper interface to populate the PGO symtab with functions |
||
| 496 | /// decls from module \c M. This interface is used by transformation |
||
| 497 | /// passes such as indirect function call promotion. Variable \c InLTO |
||
| 498 | /// indicates if this is called from LTO optimization passes. |
||
| 499 | Error create(Module &M, bool InLTO = false); |
||
| 500 | |||
| 501 | /// Create InstrProfSymtab from a set of names iteratable from |
||
| 502 | /// \p IterRange. This interface is used by IndexedProfReader. |
||
| 503 | template <typename NameIterRange> Error create(const NameIterRange &IterRange); |
||
| 504 | |||
| 505 | /// Update the symtab by adding \p FuncName to the table. This interface |
||
| 506 | /// is used by the raw and text profile readers. |
||
| 507 | Error addFuncName(StringRef FuncName) { |
||
| 508 | if (FuncName.empty()) |
||
| 509 | return make_error<InstrProfError>(instrprof_error::malformed, |
||
| 510 | "function name is empty"); |
||
| 511 | auto Ins = NameTab.insert(FuncName); |
||
| 512 | if (Ins.second) { |
||
| 513 | MD5NameMap.push_back(std::make_pair( |
||
| 514 | IndexedInstrProf::ComputeHash(FuncName), Ins.first->getKey())); |
||
| 515 | Sorted = false; |
||
| 516 | } |
||
| 517 | return Error::success(); |
||
| 518 | } |
||
| 519 | |||
| 520 | /// Map a function address to its name's MD5 hash. This interface |
||
| 521 | /// is only used by the raw profiler reader. |
||
| 522 | void mapAddress(uint64_t Addr, uint64_t MD5Val) { |
||
| 523 | AddrToMD5Map.push_back(std::make_pair(Addr, MD5Val)); |
||
| 524 | } |
||
| 525 | |||
| 526 | /// Return a function's hash, or 0, if the function isn't in this SymTab. |
||
| 527 | uint64_t getFunctionHashFromAddress(uint64_t Address); |
||
| 528 | |||
| 529 | /// Return function's PGO name from the function name's symbol |
||
| 530 | /// address in the object file. If an error occurs, return |
||
| 531 | /// an empty string. |
||
| 532 | StringRef getFuncName(uint64_t FuncNameAddress, size_t NameSize); |
||
| 533 | |||
| 534 | /// Return function's PGO name from the name's md5 hash value. |
||
| 535 | /// If not found, return an empty string. |
||
| 536 | inline StringRef getFuncName(uint64_t FuncMD5Hash); |
||
| 537 | |||
| 538 | /// Just like getFuncName, except that it will return a non-empty StringRef |
||
| 539 | /// if the function is external to this symbol table. All such cases |
||
| 540 | /// will be represented using the same StringRef value. |
||
| 541 | inline StringRef getFuncNameOrExternalSymbol(uint64_t FuncMD5Hash); |
||
| 542 | |||
| 543 | /// True if Symbol is the value used to represent external symbols. |
||
| 544 | static bool isExternalSymbol(const StringRef &Symbol) { |
||
| 545 | return Symbol == InstrProfSymtab::getExternalSymbol(); |
||
| 546 | } |
||
| 547 | |||
| 548 | /// Return function from the name's md5 hash. Return nullptr if not found. |
||
| 549 | inline Function *getFunction(uint64_t FuncMD5Hash); |
||
| 550 | |||
| 551 | /// Return the function's original assembly name by stripping off |
||
| 552 | /// the prefix attached (to symbols with priviate linkage). For |
||
| 553 | /// global functions, it returns the same string as getFuncName. |
||
| 554 | inline StringRef getOrigFuncName(uint64_t FuncMD5Hash); |
||
| 555 | |||
| 556 | /// Return the name section data. |
||
| 557 | inline StringRef getNameData() const { return Data; } |
||
| 558 | |||
| 559 | /// Dump the symbols in this table. |
||
| 560 | void dumpNames(raw_ostream &OS) const { |
||
| 561 | for (StringRef S : NameTab.keys()) |
||
| 562 | OS << S << "\n"; |
||
| 563 | } |
||
| 564 | }; |
||
| 565 | |||
| 566 | Error InstrProfSymtab::create(StringRef D, uint64_t BaseAddr) { |
||
| 567 | Data = D; |
||
| 568 | Address = BaseAddr; |
||
| 569 | return Error::success(); |
||
| 570 | } |
||
| 571 | |||
| 572 | Error InstrProfSymtab::create(StringRef NameStrings) { |
||
| 573 | return readPGOFuncNameStrings(NameStrings, *this); |
||
| 574 | } |
||
| 575 | |||
| 576 | template <typename NameIterRange> |
||
| 577 | Error InstrProfSymtab::create(const NameIterRange &IterRange) { |
||
| 578 | for (auto Name : IterRange) |
||
| 579 | if (Error E = addFuncName(Name)) |
||
| 580 | return E; |
||
| 581 | |||
| 582 | finalizeSymtab(); |
||
| 583 | return Error::success(); |
||
| 584 | } |
||
| 585 | |||
| 586 | void InstrProfSymtab::finalizeSymtab() { |
||
| 587 | if (Sorted) |
||
| 588 | return; |
||
| 589 | llvm::sort(MD5NameMap, less_first()); |
||
| 590 | llvm::sort(MD5FuncMap, less_first()); |
||
| 591 | llvm::sort(AddrToMD5Map, less_first()); |
||
| 592 | AddrToMD5Map.erase(std::unique(AddrToMD5Map.begin(), AddrToMD5Map.end()), |
||
| 593 | AddrToMD5Map.end()); |
||
| 594 | Sorted = true; |
||
| 595 | } |
||
| 596 | |||
| 597 | StringRef InstrProfSymtab::getFuncNameOrExternalSymbol(uint64_t FuncMD5Hash) { |
||
| 598 | StringRef ret = getFuncName(FuncMD5Hash); |
||
| 599 | if (ret.empty()) |
||
| 600 | return InstrProfSymtab::getExternalSymbol(); |
||
| 601 | return ret; |
||
| 602 | } |
||
| 603 | |||
| 604 | StringRef InstrProfSymtab::getFuncName(uint64_t FuncMD5Hash) { |
||
| 605 | finalizeSymtab(); |
||
| 606 | auto Result = llvm::lower_bound(MD5NameMap, FuncMD5Hash, |
||
| 607 | [](const std::pair<uint64_t, StringRef> &LHS, |
||
| 608 | uint64_t RHS) { return LHS.first < RHS; }); |
||
| 609 | if (Result != MD5NameMap.end() && Result->first == FuncMD5Hash) |
||
| 610 | return Result->second; |
||
| 611 | return StringRef(); |
||
| 612 | } |
||
| 613 | |||
| 614 | Function* InstrProfSymtab::getFunction(uint64_t FuncMD5Hash) { |
||
| 615 | finalizeSymtab(); |
||
| 616 | auto Result = llvm::lower_bound(MD5FuncMap, FuncMD5Hash, |
||
| 617 | [](const std::pair<uint64_t, Function *> &LHS, |
||
| 618 | uint64_t RHS) { return LHS.first < RHS; }); |
||
| 619 | if (Result != MD5FuncMap.end() && Result->first == FuncMD5Hash) |
||
| 620 | return Result->second; |
||
| 621 | return nullptr; |
||
| 622 | } |
||
| 623 | |||
| 624 | // See also getPGOFuncName implementation. These two need to be |
||
| 625 | // matched. |
||
| 626 | StringRef InstrProfSymtab::getOrigFuncName(uint64_t FuncMD5Hash) { |
||
| 627 | StringRef PGOName = getFuncName(FuncMD5Hash); |
||
| 628 | size_t S = PGOName.find_first_of(':'); |
||
| 629 | if (S == StringRef::npos) |
||
| 630 | return PGOName; |
||
| 631 | return PGOName.drop_front(S + 1); |
||
| 632 | } |
||
| 633 | |||
| 634 | // To store the sums of profile count values, or the percentage of |
||
| 635 | // the sums of the total count values. |
||
| 636 | struct CountSumOrPercent { |
||
| 637 | uint64_t NumEntries; |
||
| 638 | double CountSum; |
||
| 639 | double ValueCounts[IPVK_Last - IPVK_First + 1]; |
||
| 640 | CountSumOrPercent() : NumEntries(0), CountSum(0.0f), ValueCounts() {} |
||
| 641 | void reset() { |
||
| 642 | NumEntries = 0; |
||
| 643 | CountSum = 0.0f; |
||
| 644 | for (double &VC : ValueCounts) |
||
| 645 | VC = 0.0f; |
||
| 646 | } |
||
| 647 | }; |
||
| 648 | |||
| 649 | // Function level or program level overlap information. |
||
| 650 | struct OverlapStats { |
||
| 651 | enum OverlapStatsLevel { ProgramLevel, FunctionLevel }; |
||
| 652 | // Sum of the total count values for the base profile. |
||
| 653 | CountSumOrPercent Base; |
||
| 654 | // Sum of the total count values for the test profile. |
||
| 655 | CountSumOrPercent Test; |
||
| 656 | // Overlap lap score. Should be in range of [0.0f to 1.0f]. |
||
| 657 | CountSumOrPercent Overlap; |
||
| 658 | CountSumOrPercent Mismatch; |
||
| 659 | CountSumOrPercent Unique; |
||
| 660 | OverlapStatsLevel Level; |
||
| 661 | const std::string *BaseFilename; |
||
| 662 | const std::string *TestFilename; |
||
| 663 | StringRef FuncName; |
||
| 664 | uint64_t FuncHash; |
||
| 665 | bool Valid; |
||
| 666 | |||
| 667 | OverlapStats(OverlapStatsLevel L = ProgramLevel) |
||
| 668 | : Level(L), BaseFilename(nullptr), TestFilename(nullptr), FuncHash(0), |
||
| 669 | Valid(false) {} |
||
| 670 | |||
| 671 | void dump(raw_fd_ostream &OS) const; |
||
| 672 | |||
| 673 | void setFuncInfo(StringRef Name, uint64_t Hash) { |
||
| 674 | FuncName = Name; |
||
| 675 | FuncHash = Hash; |
||
| 676 | } |
||
| 677 | |||
| 678 | Error accumulateCounts(const std::string &BaseFilename, |
||
| 679 | const std::string &TestFilename, bool IsCS); |
||
| 680 | void addOneMismatch(const CountSumOrPercent &MismatchFunc); |
||
| 681 | void addOneUnique(const CountSumOrPercent &UniqueFunc); |
||
| 682 | |||
| 683 | static inline double score(uint64_t Val1, uint64_t Val2, double Sum1, |
||
| 684 | double Sum2) { |
||
| 685 | if (Sum1 < 1.0f || Sum2 < 1.0f) |
||
| 686 | return 0.0f; |
||
| 687 | return std::min(Val1 / Sum1, Val2 / Sum2); |
||
| 688 | } |
||
| 689 | }; |
||
| 690 | |||
| 691 | // This is used to filter the functions whose overlap information |
||
| 692 | // to be output. |
||
| 693 | struct OverlapFuncFilters { |
||
| 694 | uint64_t ValueCutoff; |
||
| 695 | const std::string NameFilter; |
||
| 696 | }; |
||
| 697 | |||
| 698 | struct InstrProfValueSiteRecord { |
||
| 699 | /// Value profiling data pairs at a given value site. |
||
| 700 | std::list<InstrProfValueData> ValueData; |
||
| 701 | |||
| 702 | InstrProfValueSiteRecord() { ValueData.clear(); } |
||
| 703 | template <class InputIterator> |
||
| 704 | InstrProfValueSiteRecord(InputIterator F, InputIterator L) |
||
| 705 | : ValueData(F, L) {} |
||
| 706 | |||
| 707 | /// Sort ValueData ascending by Value |
||
| 708 | void sortByTargetValues() { |
||
| 709 | ValueData.sort( |
||
| 710 | [](const InstrProfValueData &left, const InstrProfValueData &right) { |
||
| 711 | return left.Value < right.Value; |
||
| 712 | }); |
||
| 713 | } |
||
| 714 | /// Sort ValueData Descending by Count |
||
| 715 | inline void sortByCount(); |
||
| 716 | |||
| 717 | /// Merge data from another InstrProfValueSiteRecord |
||
| 718 | /// Optionally scale merged counts by \p Weight. |
||
| 719 | void merge(InstrProfValueSiteRecord &Input, uint64_t Weight, |
||
| 720 | function_ref<void(instrprof_error)> Warn); |
||
| 721 | /// Scale up value profile data counts by N (Numerator) / D (Denominator). |
||
| 722 | void scale(uint64_t N, uint64_t D, function_ref<void(instrprof_error)> Warn); |
||
| 723 | |||
| 724 | /// Compute the overlap b/w this record and Input record. |
||
| 725 | void overlap(InstrProfValueSiteRecord &Input, uint32_t ValueKind, |
||
| 726 | OverlapStats &Overlap, OverlapStats &FuncLevelOverlap); |
||
| 727 | }; |
||
| 728 | |||
| 729 | /// Profiling information for a single function. |
||
| 730 | struct InstrProfRecord { |
||
| 731 | std::vector<uint64_t> Counts; |
||
| 732 | |||
| 733 | InstrProfRecord() = default; |
||
| 734 | InstrProfRecord(std::vector<uint64_t> Counts) : Counts(std::move(Counts)) {} |
||
| 735 | InstrProfRecord(InstrProfRecord &&) = default; |
||
| 736 | InstrProfRecord(const InstrProfRecord &RHS) |
||
| 737 | : Counts(RHS.Counts), |
||
| 738 | ValueData(RHS.ValueData |
||
| 739 | ? std::make_unique<ValueProfData>(*RHS.ValueData) |
||
| 740 | : nullptr) {} |
||
| 741 | InstrProfRecord &operator=(InstrProfRecord &&) = default; |
||
| 742 | InstrProfRecord &operator=(const InstrProfRecord &RHS) { |
||
| 743 | Counts = RHS.Counts; |
||
| 744 | if (!RHS.ValueData) { |
||
| 745 | ValueData = nullptr; |
||
| 746 | return *this; |
||
| 747 | } |
||
| 748 | if (!ValueData) |
||
| 749 | ValueData = std::make_unique<ValueProfData>(*RHS.ValueData); |
||
| 750 | else |
||
| 751 | *ValueData = *RHS.ValueData; |
||
| 752 | return *this; |
||
| 753 | } |
||
| 754 | |||
| 755 | /// Return the number of value profile kinds with non-zero number |
||
| 756 | /// of profile sites. |
||
| 757 | inline uint32_t getNumValueKinds() const; |
||
| 758 | /// Return the number of instrumented sites for ValueKind. |
||
| 759 | inline uint32_t getNumValueSites(uint32_t ValueKind) const; |
||
| 760 | |||
| 761 | /// Return the total number of ValueData for ValueKind. |
||
| 762 | inline uint32_t getNumValueData(uint32_t ValueKind) const; |
||
| 763 | |||
| 764 | /// Return the number of value data collected for ValueKind at profiling |
||
| 765 | /// site: Site. |
||
| 766 | inline uint32_t getNumValueDataForSite(uint32_t ValueKind, |
||
| 767 | uint32_t Site) const; |
||
| 768 | |||
| 769 | /// Return the array of profiled values at \p Site. If \p TotalC |
||
| 770 | /// is not null, the total count of all target values at this site |
||
| 771 | /// will be stored in \c *TotalC. |
||
| 772 | inline std::unique_ptr<InstrProfValueData[]> |
||
| 773 | getValueForSite(uint32_t ValueKind, uint32_t Site, |
||
| 774 | uint64_t *TotalC = nullptr) const; |
||
| 775 | |||
| 776 | /// Get the target value/counts of kind \p ValueKind collected at site |
||
| 777 | /// \p Site and store the result in array \p Dest. Return the total |
||
| 778 | /// counts of all target values at this site. |
||
| 779 | inline uint64_t getValueForSite(InstrProfValueData Dest[], uint32_t ValueKind, |
||
| 780 | uint32_t Site) const; |
||
| 781 | |||
| 782 | /// Reserve space for NumValueSites sites. |
||
| 783 | inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites); |
||
| 784 | |||
| 785 | /// Add ValueData for ValueKind at value Site. |
||
| 786 | void addValueData(uint32_t ValueKind, uint32_t Site, |
||
| 787 | InstrProfValueData *VData, uint32_t N, |
||
| 788 | InstrProfSymtab *SymTab); |
||
| 789 | |||
| 790 | /// Merge the counts in \p Other into this one. |
||
| 791 | /// Optionally scale merged counts by \p Weight. |
||
| 792 | void merge(InstrProfRecord &Other, uint64_t Weight, |
||
| 793 | function_ref<void(instrprof_error)> Warn); |
||
| 794 | |||
| 795 | /// Scale up profile counts (including value profile data) by |
||
| 796 | /// a factor of (N / D). |
||
| 797 | void scale(uint64_t N, uint64_t D, function_ref<void(instrprof_error)> Warn); |
||
| 798 | |||
| 799 | /// Sort value profile data (per site) by count. |
||
| 800 | void sortValueData() { |
||
| 801 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) |
||
| 802 | for (auto &SR : getValueSitesForKind(Kind)) |
||
| 803 | SR.sortByCount(); |
||
| 804 | } |
||
| 805 | |||
| 806 | /// Clear value data entries and edge counters. |
||
| 807 | void Clear() { |
||
| 808 | Counts.clear(); |
||
| 809 | clearValueData(); |
||
| 810 | } |
||
| 811 | |||
| 812 | /// Clear value data entries |
||
| 813 | void clearValueData() { ValueData = nullptr; } |
||
| 814 | |||
| 815 | /// Compute the sums of all counts and store in Sum. |
||
| 816 | void accumulateCounts(CountSumOrPercent &Sum) const; |
||
| 817 | |||
| 818 | /// Compute the overlap b/w this IntrprofRecord and Other. |
||
| 819 | void overlap(InstrProfRecord &Other, OverlapStats &Overlap, |
||
| 820 | OverlapStats &FuncLevelOverlap, uint64_t ValueCutoff); |
||
| 821 | |||
| 822 | /// Compute the overlap of value profile counts. |
||
| 823 | void overlapValueProfData(uint32_t ValueKind, InstrProfRecord &Src, |
||
| 824 | OverlapStats &Overlap, |
||
| 825 | OverlapStats &FuncLevelOverlap); |
||
| 826 | |||
| 827 | enum CountPseudoKind { |
||
| 828 | NotPseudo = 0, |
||
| 829 | PseudoHot, |
||
| 830 | PseudoWarm, |
||
| 831 | }; |
||
| 832 | enum PseudoCountVal { |
||
| 833 | HotFunctionVal = -1, |
||
| 834 | WarmFunctionVal = -2, |
||
| 835 | }; |
||
| 836 | CountPseudoKind getCountPseudoKind() const { |
||
| 837 | uint64_t FirstCount = Counts[0]; |
||
| 838 | if (FirstCount == (uint64_t)HotFunctionVal) |
||
| 839 | return PseudoHot; |
||
| 840 | if (FirstCount == (uint64_t)WarmFunctionVal) |
||
| 841 | return PseudoWarm; |
||
| 842 | return NotPseudo; |
||
| 843 | } |
||
| 844 | void setPseudoCount(CountPseudoKind Kind) { |
||
| 845 | if (Kind == PseudoHot) |
||
| 846 | Counts[0] = (uint64_t)HotFunctionVal; |
||
| 847 | else if (Kind == PseudoWarm) |
||
| 848 | Counts[0] = (uint64_t)WarmFunctionVal; |
||
| 849 | } |
||
| 850 | |||
| 851 | private: |
||
| 852 | struct ValueProfData { |
||
| 853 | std::vector<InstrProfValueSiteRecord> IndirectCallSites; |
||
| 854 | std::vector<InstrProfValueSiteRecord> MemOPSizes; |
||
| 855 | }; |
||
| 856 | std::unique_ptr<ValueProfData> ValueData; |
||
| 857 | |||
| 858 | MutableArrayRef<InstrProfValueSiteRecord> |
||
| 859 | getValueSitesForKind(uint32_t ValueKind) { |
||
| 860 | // Cast to /add/ const (should be an implicit_cast, ideally, if that's ever |
||
| 861 | // implemented in LLVM) to call the const overload of this function, then |
||
| 862 | // cast away the constness from the result. |
||
| 863 | auto AR = const_cast<const InstrProfRecord *>(this)->getValueSitesForKind( |
||
| 864 | ValueKind); |
||
| 865 | return MutableArrayRef( |
||
| 866 | const_cast<InstrProfValueSiteRecord *>(AR.data()), AR.size()); |
||
| 867 | } |
||
| 868 | ArrayRef<InstrProfValueSiteRecord> |
||
| 869 | getValueSitesForKind(uint32_t ValueKind) const { |
||
| 870 | if (!ValueData) |
||
| 871 | return std::nullopt; |
||
| 872 | switch (ValueKind) { |
||
| 873 | case IPVK_IndirectCallTarget: |
||
| 874 | return ValueData->IndirectCallSites; |
||
| 875 | case IPVK_MemOPSize: |
||
| 876 | return ValueData->MemOPSizes; |
||
| 877 | default: |
||
| 878 | llvm_unreachable("Unknown value kind!"); |
||
| 879 | } |
||
| 880 | } |
||
| 881 | |||
| 882 | std::vector<InstrProfValueSiteRecord> & |
||
| 883 | getOrCreateValueSitesForKind(uint32_t ValueKind) { |
||
| 884 | if (!ValueData) |
||
| 885 | ValueData = std::make_unique<ValueProfData>(); |
||
| 886 | switch (ValueKind) { |
||
| 887 | case IPVK_IndirectCallTarget: |
||
| 888 | return ValueData->IndirectCallSites; |
||
| 889 | case IPVK_MemOPSize: |
||
| 890 | return ValueData->MemOPSizes; |
||
| 891 | default: |
||
| 892 | llvm_unreachable("Unknown value kind!"); |
||
| 893 | } |
||
| 894 | } |
||
| 895 | |||
| 896 | // Map indirect call target name hash to name string. |
||
| 897 | uint64_t remapValue(uint64_t Value, uint32_t ValueKind, |
||
| 898 | InstrProfSymtab *SymTab); |
||
| 899 | |||
| 900 | // Merge Value Profile data from Src record to this record for ValueKind. |
||
| 901 | // Scale merged value counts by \p Weight. |
||
| 902 | void mergeValueProfData(uint32_t ValkeKind, InstrProfRecord &Src, |
||
| 903 | uint64_t Weight, |
||
| 904 | function_ref<void(instrprof_error)> Warn); |
||
| 905 | |||
| 906 | // Scale up value profile data count by N (Numerator) / D (Denominator). |
||
| 907 | void scaleValueProfData(uint32_t ValueKind, uint64_t N, uint64_t D, |
||
| 908 | function_ref<void(instrprof_error)> Warn); |
||
| 909 | }; |
||
| 910 | |||
| 911 | struct NamedInstrProfRecord : InstrProfRecord { |
||
| 912 | StringRef Name; |
||
| 913 | uint64_t Hash; |
||
| 914 | |||
| 915 | // We reserve this bit as the flag for context sensitive profile record. |
||
| 916 | static const int CS_FLAG_IN_FUNC_HASH = 60; |
||
| 917 | |||
| 918 | NamedInstrProfRecord() = default; |
||
| 919 | NamedInstrProfRecord(StringRef Name, uint64_t Hash, |
||
| 920 | std::vector<uint64_t> Counts) |
||
| 921 | : InstrProfRecord(std::move(Counts)), Name(Name), Hash(Hash) {} |
||
| 922 | |||
| 923 | static bool hasCSFlagInHash(uint64_t FuncHash) { |
||
| 924 | return ((FuncHash >> CS_FLAG_IN_FUNC_HASH) & 1); |
||
| 925 | } |
||
| 926 | static void setCSFlagInHash(uint64_t &FuncHash) { |
||
| 927 | FuncHash |= ((uint64_t)1 << CS_FLAG_IN_FUNC_HASH); |
||
| 928 | } |
||
| 929 | }; |
||
| 930 | |||
| 931 | uint32_t InstrProfRecord::getNumValueKinds() const { |
||
| 932 | uint32_t NumValueKinds = 0; |
||
| 933 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) |
||
| 934 | NumValueKinds += !(getValueSitesForKind(Kind).empty()); |
||
| 935 | return NumValueKinds; |
||
| 936 | } |
||
| 937 | |||
| 938 | uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const { |
||
| 939 | uint32_t N = 0; |
||
| 940 | for (const auto &SR : getValueSitesForKind(ValueKind)) |
||
| 941 | N += SR.ValueData.size(); |
||
| 942 | return N; |
||
| 943 | } |
||
| 944 | |||
| 945 | uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const { |
||
| 946 | return getValueSitesForKind(ValueKind).size(); |
||
| 947 | } |
||
| 948 | |||
| 949 | uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind, |
||
| 950 | uint32_t Site) const { |
||
| 951 | return getValueSitesForKind(ValueKind)[Site].ValueData.size(); |
||
| 952 | } |
||
| 953 | |||
| 954 | std::unique_ptr<InstrProfValueData[]> |
||
| 955 | InstrProfRecord::getValueForSite(uint32_t ValueKind, uint32_t Site, |
||
| 956 | uint64_t *TotalC) const { |
||
| 957 | uint64_t Dummy = 0; |
||
| 958 | uint64_t &TotalCount = (TotalC == nullptr ? Dummy : *TotalC); |
||
| 959 | uint32_t N = getNumValueDataForSite(ValueKind, Site); |
||
| 960 | if (N == 0) { |
||
| 961 | TotalCount = 0; |
||
| 962 | return std::unique_ptr<InstrProfValueData[]>(nullptr); |
||
| 963 | } |
||
| 964 | |||
| 965 | auto VD = std::make_unique<InstrProfValueData[]>(N); |
||
| 966 | TotalCount = getValueForSite(VD.get(), ValueKind, Site); |
||
| 967 | |||
| 968 | return VD; |
||
| 969 | } |
||
| 970 | |||
| 971 | uint64_t InstrProfRecord::getValueForSite(InstrProfValueData Dest[], |
||
| 972 | uint32_t ValueKind, |
||
| 973 | uint32_t Site) const { |
||
| 974 | uint32_t I = 0; |
||
| 975 | uint64_t TotalCount = 0; |
||
| 976 | for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) { |
||
| 977 | Dest[I].Value = V.Value; |
||
| 978 | Dest[I].Count = V.Count; |
||
| 979 | TotalCount = SaturatingAdd(TotalCount, V.Count); |
||
| 980 | I++; |
||
| 981 | } |
||
| 982 | return TotalCount; |
||
| 983 | } |
||
| 984 | |||
| 985 | void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) { |
||
| 986 | if (!NumValueSites) |
||
| 987 | return; |
||
| 988 | getOrCreateValueSitesForKind(ValueKind).reserve(NumValueSites); |
||
| 989 | } |
||
| 990 | |||
| 991 | inline support::endianness getHostEndianness() { |
||
| 992 | return sys::IsLittleEndianHost ? support::little : support::big; |
||
| 993 | } |
||
| 994 | |||
| 995 | // Include definitions for value profile data |
||
| 996 | #define INSTR_PROF_VALUE_PROF_DATA |
||
| 997 | #include "llvm/ProfileData/InstrProfData.inc" |
||
| 998 | |||
| 999 | void InstrProfValueSiteRecord::sortByCount() { |
||
| 1000 | ValueData.sort( |
||
| 1001 | [](const InstrProfValueData &left, const InstrProfValueData &right) { |
||
| 1002 | return left.Count > right.Count; |
||
| 1003 | }); |
||
| 1004 | // Now truncate |
||
| 1005 | size_t max_s = INSTR_PROF_MAX_NUM_VAL_PER_SITE; |
||
| 1006 | if (ValueData.size() > max_s) |
||
| 1007 | ValueData.resize(max_s); |
||
| 1008 | } |
||
| 1009 | |||
| 1010 | namespace IndexedInstrProf { |
||
| 1011 | |||
| 1012 | enum class HashT : uint32_t { |
||
| 1013 | MD5, |
||
| 1014 | Last = MD5 |
||
| 1015 | }; |
||
| 1016 | |||
| 1017 | inline uint64_t ComputeHash(HashT Type, StringRef K) { |
||
| 1018 | switch (Type) { |
||
| 1019 | case HashT::MD5: |
||
| 1020 | return MD5Hash(K); |
||
| 1021 | } |
||
| 1022 | llvm_unreachable("Unhandled hash type"); |
||
| 1023 | } |
||
| 1024 | |||
| 1025 | const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81" |
||
| 1026 | |||
| 1027 | enum ProfVersion { |
||
| 1028 | // Version 1 is the first version. In this version, the value of |
||
| 1029 | // a key/value pair can only include profile data of a single function. |
||
| 1030 | // Due to this restriction, the number of block counters for a given |
||
| 1031 | // function is not recorded but derived from the length of the value. |
||
| 1032 | Version1 = 1, |
||
| 1033 | // The version 2 format supports recording profile data of multiple |
||
| 1034 | // functions which share the same key in one value field. To support this, |
||
| 1035 | // the number block counters is recorded as an uint64_t field right after the |
||
| 1036 | // function structural hash. |
||
| 1037 | Version2 = 2, |
||
| 1038 | // Version 3 supports value profile data. The value profile data is expected |
||
| 1039 | // to follow the block counter profile data. |
||
| 1040 | Version3 = 3, |
||
| 1041 | // In this version, profile summary data \c IndexedInstrProf::Summary is |
||
| 1042 | // stored after the profile header. |
||
| 1043 | Version4 = 4, |
||
| 1044 | // In this version, the frontend PGO stable hash algorithm defaults to V2. |
||
| 1045 | Version5 = 5, |
||
| 1046 | // In this version, the frontend PGO stable hash algorithm got fixed and |
||
| 1047 | // may produce hashes different from Version5. |
||
| 1048 | Version6 = 6, |
||
| 1049 | // An additional counter is added around logical operators. |
||
| 1050 | Version7 = 7, |
||
| 1051 | // An additional (optional) memory profile type is added. |
||
| 1052 | Version8 = 8, |
||
| 1053 | // Binary ids are added. |
||
| 1054 | Version9 = 9, |
||
| 1055 | // The current version is 9. |
||
| 1056 | CurrentVersion = INSTR_PROF_INDEX_VERSION |
||
| 1057 | }; |
||
| 1058 | const uint64_t Version = ProfVersion::CurrentVersion; |
||
| 1059 | |||
| 1060 | const HashT HashType = HashT::MD5; |
||
| 1061 | |||
| 1062 | inline uint64_t ComputeHash(StringRef K) { return ComputeHash(HashType, K); } |
||
| 1063 | |||
| 1064 | // This structure defines the file header of the LLVM profile |
||
| 1065 | // data file in indexed-format. |
||
| 1066 | struct Header { |
||
| 1067 | uint64_t Magic; |
||
| 1068 | uint64_t Version; |
||
| 1069 | uint64_t Unused; // Becomes unused since version 4 |
||
| 1070 | uint64_t HashType; |
||
| 1071 | uint64_t HashOffset; |
||
| 1072 | uint64_t MemProfOffset; |
||
| 1073 | uint64_t BinaryIdOffset; |
||
| 1074 | // New fields should only be added at the end to ensure that the size |
||
| 1075 | // computation is correct. The methods below need to be updated to ensure that |
||
| 1076 | // the new field is read correctly. |
||
| 1077 | |||
| 1078 | // Reads a header struct from the buffer. |
||
| 1079 | static Expected<Header> readFromBuffer(const unsigned char *Buffer); |
||
| 1080 | |||
| 1081 | // Returns the size of the header in bytes for all valid fields based on the |
||
| 1082 | // version. I.e a older version header will return a smaller size. |
||
| 1083 | size_t size() const; |
||
| 1084 | |||
| 1085 | // Returns the format version in little endian. The header retains the version |
||
| 1086 | // in native endian of the compiler runtime. |
||
| 1087 | uint64_t formatVersion() const; |
||
| 1088 | }; |
||
| 1089 | |||
| 1090 | // Profile summary data recorded in the profile data file in indexed |
||
| 1091 | // format. It is introduced in version 4. The summary data follows |
||
| 1092 | // right after the profile file header. |
||
| 1093 | struct Summary { |
||
| 1094 | struct Entry { |
||
| 1095 | uint64_t Cutoff; ///< The required percentile of total execution count. |
||
| 1096 | uint64_t |
||
| 1097 | MinBlockCount; ///< The minimum execution count for this percentile. |
||
| 1098 | uint64_t NumBlocks; ///< Number of blocks >= the minumum execution count. |
||
| 1099 | }; |
||
| 1100 | // The field kind enumerator to assigned value mapping should remain |
||
| 1101 | // unchanged when a new kind is added or an old kind gets deleted in |
||
| 1102 | // the future. |
||
| 1103 | enum SummaryFieldKind { |
||
| 1104 | /// The total number of functions instrumented. |
||
| 1105 | TotalNumFunctions = 0, |
||
| 1106 | /// Total number of instrumented blocks/edges. |
||
| 1107 | TotalNumBlocks = 1, |
||
| 1108 | /// The maximal execution count among all functions. |
||
| 1109 | /// This field does not exist for profile data from IR based |
||
| 1110 | /// instrumentation. |
||
| 1111 | MaxFunctionCount = 2, |
||
| 1112 | /// Max block count of the program. |
||
| 1113 | MaxBlockCount = 3, |
||
| 1114 | /// Max internal block count of the program (excluding entry blocks). |
||
| 1115 | MaxInternalBlockCount = 4, |
||
| 1116 | /// The sum of all instrumented block counts. |
||
| 1117 | TotalBlockCount = 5, |
||
| 1118 | NumKinds = TotalBlockCount + 1 |
||
| 1119 | }; |
||
| 1120 | |||
| 1121 | // The number of summmary fields following the summary header. |
||
| 1122 | uint64_t NumSummaryFields; |
||
| 1123 | // The number of Cutoff Entries (Summary::Entry) following summary fields. |
||
| 1124 | uint64_t NumCutoffEntries; |
||
| 1125 | |||
| 1126 | Summary() = delete; |
||
| 1127 | Summary(uint32_t Size) { memset(this, 0, Size); } |
||
| 1128 | |||
| 1129 | void operator delete(void *ptr) { ::operator delete(ptr); } |
||
| 1130 | |||
| 1131 | static uint32_t getSize(uint32_t NumSumFields, uint32_t NumCutoffEntries) { |
||
| 1132 | return sizeof(Summary) + NumCutoffEntries * sizeof(Entry) + |
||
| 1133 | NumSumFields * sizeof(uint64_t); |
||
| 1134 | } |
||
| 1135 | |||
| 1136 | const uint64_t *getSummaryDataBase() const { |
||
| 1137 | return reinterpret_cast<const uint64_t *>(this + 1); |
||
| 1138 | } |
||
| 1139 | |||
| 1140 | uint64_t *getSummaryDataBase() { |
||
| 1141 | return reinterpret_cast<uint64_t *>(this + 1); |
||
| 1142 | } |
||
| 1143 | |||
| 1144 | const Entry *getCutoffEntryBase() const { |
||
| 1145 | return reinterpret_cast<const Entry *>( |
||
| 1146 | &getSummaryDataBase()[NumSummaryFields]); |
||
| 1147 | } |
||
| 1148 | |||
| 1149 | Entry *getCutoffEntryBase() { |
||
| 1150 | return reinterpret_cast<Entry *>(&getSummaryDataBase()[NumSummaryFields]); |
||
| 1151 | } |
||
| 1152 | |||
| 1153 | uint64_t get(SummaryFieldKind K) const { |
||
| 1154 | return getSummaryDataBase()[K]; |
||
| 1155 | } |
||
| 1156 | |||
| 1157 | void set(SummaryFieldKind K, uint64_t V) { |
||
| 1158 | getSummaryDataBase()[K] = V; |
||
| 1159 | } |
||
| 1160 | |||
| 1161 | const Entry &getEntry(uint32_t I) const { return getCutoffEntryBase()[I]; } |
||
| 1162 | |||
| 1163 | void setEntry(uint32_t I, const ProfileSummaryEntry &E) { |
||
| 1164 | Entry &ER = getCutoffEntryBase()[I]; |
||
| 1165 | ER.Cutoff = E.Cutoff; |
||
| 1166 | ER.MinBlockCount = E.MinCount; |
||
| 1167 | ER.NumBlocks = E.NumCounts; |
||
| 1168 | } |
||
| 1169 | }; |
||
| 1170 | |||
| 1171 | inline std::unique_ptr<Summary> allocSummary(uint32_t TotalSize) { |
||
| 1172 | return std::unique_ptr<Summary>(new (::operator new(TotalSize)) |
||
| 1173 | Summary(TotalSize)); |
||
| 1174 | } |
||
| 1175 | |||
| 1176 | } // end namespace IndexedInstrProf |
||
| 1177 | |||
| 1178 | namespace RawInstrProf { |
||
| 1179 | |||
| 1180 | // Version 1: First version |
||
| 1181 | // Version 2: Added value profile data section. Per-function control data |
||
| 1182 | // struct has more fields to describe value profile information. |
||
| 1183 | // Version 3: Compressed name section support. Function PGO name reference |
||
| 1184 | // from control data struct is changed from raw pointer to Name's MD5 value. |
||
| 1185 | // Version 4: ValueDataBegin and ValueDataSizes fields are removed from the |
||
| 1186 | // raw header. |
||
| 1187 | // Version 5: Bit 60 of FuncHash is reserved for the flag for the context |
||
| 1188 | // sensitive records. |
||
| 1189 | // Version 6: Added binary id. |
||
| 1190 | // Version 7: Reorder binary id and include version in signature. |
||
| 1191 | // Version 8: Use relative counter pointer. |
||
| 1192 | const uint64_t Version = INSTR_PROF_RAW_VERSION; |
||
| 1193 | |||
| 1194 | template <class IntPtrT> inline uint64_t getMagic(); |
||
| 1195 | template <> inline uint64_t getMagic<uint64_t>() { |
||
| 1196 | return INSTR_PROF_RAW_MAGIC_64; |
||
| 1197 | } |
||
| 1198 | |||
| 1199 | template <> inline uint64_t getMagic<uint32_t>() { |
||
| 1200 | return INSTR_PROF_RAW_MAGIC_32; |
||
| 1201 | } |
||
| 1202 | |||
| 1203 | // Per-function profile data header/control structure. |
||
| 1204 | // The definition should match the structure defined in |
||
| 1205 | // compiler-rt/lib/profile/InstrProfiling.h. |
||
| 1206 | // It should also match the synthesized type in |
||
| 1207 | // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters. |
||
| 1208 | template <class IntPtrT> struct alignas(8) ProfileData { |
||
| 1209 | #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name; |
||
| 1210 | #include "llvm/ProfileData/InstrProfData.inc" |
||
| 1211 | }; |
||
| 1212 | |||
| 1213 | // File header structure of the LLVM profile data in raw format. |
||
| 1214 | // The definition should match the header referenced in |
||
| 1215 | // compiler-rt/lib/profile/InstrProfilingFile.c and |
||
| 1216 | // InstrProfilingBuffer.c. |
||
| 1217 | struct Header { |
||
| 1218 | #define INSTR_PROF_RAW_HEADER(Type, Name, Init) const Type Name; |
||
| 1219 | #include "llvm/ProfileData/InstrProfData.inc" |
||
| 1220 | }; |
||
| 1221 | |||
| 1222 | } // end namespace RawInstrProf |
||
| 1223 | |||
| 1224 | // Parse MemOP Size range option. |
||
| 1225 | void getMemOPSizeRangeFromOption(StringRef Str, int64_t &RangeStart, |
||
| 1226 | int64_t &RangeLast); |
||
| 1227 | |||
| 1228 | // Create the variable for the profile file name. |
||
| 1229 | void createProfileFileNameVar(Module &M, StringRef InstrProfileOutput); |
||
| 1230 | |||
| 1231 | // Whether to compress function names in profile records, and filenames in |
||
| 1232 | // code coverage mappings. Used by the Instrumentation library and unit tests. |
||
| 1233 | extern cl::opt<bool> DoInstrProfNameCompression; |
||
| 1234 | |||
| 1235 | } // end namespace llvm |
||
| 1236 | #endif // LLVM_PROFILEDATA_INSTRPROF_H |