1 //===- SampleProfReader.h - Read LLVM sample profile data -------*- 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 // This file contains definitions needed for reading sample profiles. 10 // 11 // NOTE: If you are making changes to this file format, please remember 12 // to document them in the Clang documentation at 13 // tools/clang/docs/UsersManual.rst. 14 // 15 // Text format 16 // ----------- 17 // 18 // Sample profiles are written as ASCII text. The file is divided into 19 // sections, which correspond to each of the functions executed at runtime. 20 // Each section has the following format 21 // 22 // function1:total_samples:total_head_samples 23 // offset1[.discriminator]: number_of_samples [fn1:num fn2:num ... ] 24 // offset2[.discriminator]: number_of_samples [fn3:num fn4:num ... ] 25 // ... 26 // offsetN[.discriminator]: number_of_samples [fn5:num fn6:num ... ] 27 // offsetA[.discriminator]: fnA:num_of_total_samples 28 // offsetA1[.discriminator]: number_of_samples [fn7:num fn8:num ... ] 29 // ... 30 // !CFGChecksum: num 31 // !Attribute: flags 32 // 33 // This is a nested tree in which the indentation represents the nesting level 34 // of the inline stack. There are no blank lines in the file. And the spacing 35 // within a single line is fixed. Additional spaces will result in an error 36 // while reading the file. 37 // 38 // Any line starting with the '#' character is completely ignored. 39 // 40 // Inlined calls are represented with indentation. The Inline stack is a 41 // stack of source locations in which the top of the stack represents the 42 // leaf function, and the bottom of the stack represents the actual 43 // symbol to which the instruction belongs. 44 // 45 // Function names must be mangled in order for the profile loader to 46 // match them in the current translation unit. The two numbers in the 47 // function header specify how many total samples were accumulated in the 48 // function (first number), and the total number of samples accumulated 49 // in the prologue of the function (second number). This head sample 50 // count provides an indicator of how frequently the function is invoked. 51 // 52 // There are three types of lines in the function body. 53 // 54 // * Sampled line represents the profile information of a source location. 55 // * Callsite line represents the profile information of a callsite. 56 // * Metadata line represents extra metadata of the function. 57 // 58 // Each sampled line may contain several items. Some are optional (marked 59 // below): 60 // 61 // a. Source line offset. This number represents the line number 62 // in the function where the sample was collected. The line number is 63 // always relative to the line where symbol of the function is 64 // defined. So, if the function has its header at line 280, the offset 65 // 13 is at line 293 in the file. 66 // 67 // Note that this offset should never be a negative number. This could 68 // happen in cases like macros. The debug machinery will register the 69 // line number at the point of macro expansion. So, if the macro was 70 // expanded in a line before the start of the function, the profile 71 // converter should emit a 0 as the offset (this means that the optimizers 72 // will not be able to associate a meaningful weight to the instructions 73 // in the macro). 74 // 75 // b. [OPTIONAL] Discriminator. This is used if the sampled program 76 // was compiled with DWARF discriminator support 77 // (http://wiki.dwarfstd.org/index.php?title=Path_Discriminators). 78 // DWARF discriminators are unsigned integer values that allow the 79 // compiler to distinguish between multiple execution paths on the 80 // same source line location. 81 // 82 // For example, consider the line of code ``if (cond) foo(); else bar();``. 83 // If the predicate ``cond`` is true 80% of the time, then the edge 84 // into function ``foo`` should be considered to be taken most of the 85 // time. But both calls to ``foo`` and ``bar`` are at the same source 86 // line, so a sample count at that line is not sufficient. The 87 // compiler needs to know which part of that line is taken more 88 // frequently. 89 // 90 // This is what discriminators provide. In this case, the calls to 91 // ``foo`` and ``bar`` will be at the same line, but will have 92 // different discriminator values. This allows the compiler to correctly 93 // set edge weights into ``foo`` and ``bar``. 94 // 95 // c. Number of samples. This is an integer quantity representing the 96 // number of samples collected by the profiler at this source 97 // location. 98 // 99 // d. [OPTIONAL] Potential call targets and samples. If present, this 100 // line contains a call instruction. This models both direct and 101 // number of samples. For example, 102 // 103 // 130: 7 foo:3 bar:2 baz:7 104 // 105 // The above means that at relative line offset 130 there is a call 106 // instruction that calls one of ``foo()``, ``bar()`` and ``baz()``, 107 // with ``baz()`` being the relatively more frequently called target. 108 // 109 // Each callsite line may contain several items. Some are optional. 110 // 111 // a. Source line offset. This number represents the line number of the 112 // callsite that is inlined in the profiled binary. 113 // 114 // b. [OPTIONAL] Discriminator. Same as the discriminator for sampled line. 115 // 116 // c. Number of samples. This is an integer quantity representing the 117 // total number of samples collected for the inlined instance at this 118 // callsite 119 // 120 // Metadata line can occur in lines with one indent only, containing extra 121 // information for the top-level function. Furthermore, metadata can only 122 // occur after all the body samples and callsite samples. 123 // Each metadata line may contain a particular type of metadata, marked by 124 // the starting characters annotated with !. We process each metadata line 125 // independently, hence each metadata line has to form an independent piece 126 // of information that does not require cross-line reference. 127 // We support the following types of metadata: 128 // 129 // a. CFG Checksum (a.k.a. function hash): 130 // !CFGChecksum: 12345 131 // b. CFG Checksum (see ContextAttributeMask): 132 // !Atribute: 1 133 // 134 // 135 // Binary format 136 // ------------- 137 // 138 // This is a more compact encoding. Numbers are encoded as ULEB128 values 139 // and all strings are encoded in a name table. The file is organized in 140 // the following sections: 141 // 142 // MAGIC (uint64_t) 143 // File identifier computed by function SPMagic() (0x5350524f463432ff) 144 // 145 // VERSION (uint32_t) 146 // File format version number computed by SPVersion() 147 // 148 // SUMMARY 149 // TOTAL_COUNT (uint64_t) 150 // Total number of samples in the profile. 151 // MAX_COUNT (uint64_t) 152 // Maximum value of samples on a line. 153 // MAX_FUNCTION_COUNT (uint64_t) 154 // Maximum number of samples at function entry (head samples). 155 // NUM_COUNTS (uint64_t) 156 // Number of lines with samples. 157 // NUM_FUNCTIONS (uint64_t) 158 // Number of functions with samples. 159 // NUM_DETAILED_SUMMARY_ENTRIES (size_t) 160 // Number of entries in detailed summary 161 // DETAILED_SUMMARY 162 // A list of detailed summary entry. Each entry consists of 163 // CUTOFF (uint32_t) 164 // Required percentile of total sample count expressed as a fraction 165 // multiplied by 1000000. 166 // MIN_COUNT (uint64_t) 167 // The minimum number of samples required to reach the target 168 // CUTOFF. 169 // NUM_COUNTS (uint64_t) 170 // Number of samples to get to the desrired percentile. 171 // 172 // NAME TABLE 173 // SIZE (uint64_t) 174 // Number of entries in the name table. 175 // NAMES 176 // A NUL-separated list of SIZE strings. 177 // 178 // FUNCTION BODY (one for each uninlined function body present in the profile) 179 // HEAD_SAMPLES (uint64_t) [only for top-level functions] 180 // Total number of samples collected at the head (prologue) of the 181 // function. 182 // NOTE: This field should only be present for top-level functions 183 // (i.e., not inlined into any caller). Inlined function calls 184 // have no prologue, so they don't need this. 185 // NAME_IDX (uint64_t) 186 // Index into the name table indicating the function name. 187 // SAMPLES (uint64_t) 188 // Total number of samples collected in this function. 189 // NRECS (uint32_t) 190 // Total number of sampling records this function's profile. 191 // BODY RECORDS 192 // A list of NRECS entries. Each entry contains: 193 // OFFSET (uint32_t) 194 // Line offset from the start of the function. 195 // DISCRIMINATOR (uint32_t) 196 // Discriminator value (see description of discriminators 197 // in the text format documentation above). 198 // SAMPLES (uint64_t) 199 // Number of samples collected at this location. 200 // NUM_CALLS (uint32_t) 201 // Number of non-inlined function calls made at this location. In the 202 // case of direct calls, this number will always be 1. For indirect 203 // calls (virtual functions and function pointers) this will 204 // represent all the actual functions called at runtime. 205 // CALL_TARGETS 206 // A list of NUM_CALLS entries for each called function: 207 // NAME_IDX (uint64_t) 208 // Index into the name table with the callee name. 209 // SAMPLES (uint64_t) 210 // Number of samples collected at the call site. 211 // NUM_INLINED_FUNCTIONS (uint32_t) 212 // Number of callees inlined into this function. 213 // INLINED FUNCTION RECORDS 214 // A list of NUM_INLINED_FUNCTIONS entries describing each of the inlined 215 // callees. 216 // OFFSET (uint32_t) 217 // Line offset from the start of the function. 218 // DISCRIMINATOR (uint32_t) 219 // Discriminator value (see description of discriminators 220 // in the text format documentation above). 221 // FUNCTION BODY 222 // A FUNCTION BODY entry describing the inlined function. 223 //===----------------------------------------------------------------------===// 224 225 #ifndef LLVM_PROFILEDATA_SAMPLEPROFREADER_H 226 #define LLVM_PROFILEDATA_SAMPLEPROFREADER_H 227 228 #include "llvm/ADT/SmallVector.h" 229 #include "llvm/ADT/StringRef.h" 230 #include "llvm/IR/DiagnosticInfo.h" 231 #include "llvm/IR/LLVMContext.h" 232 #include "llvm/IR/ProfileSummary.h" 233 #include "llvm/ProfileData/GCOV.h" 234 #include "llvm/ProfileData/SampleProf.h" 235 #include "llvm/ProfileData/SymbolRemappingReader.h" 236 #include "llvm/Support/Debug.h" 237 #include "llvm/Support/Discriminator.h" 238 #include "llvm/Support/ErrorOr.h" 239 #include "llvm/Support/MemoryBuffer.h" 240 #include <cstdint> 241 #include <list> 242 #include <memory> 243 #include <optional> 244 #include <string> 245 #include <system_error> 246 #include <unordered_set> 247 #include <vector> 248 249 namespace llvm { 250 251 class raw_ostream; 252 class Twine; 253 254 namespace vfs { 255 class FileSystem; 256 } // namespace vfs 257 258 namespace sampleprof { 259 260 class SampleProfileReader; 261 262 /// SampleProfileReaderItaniumRemapper remaps the profile data from a 263 /// sample profile data reader, by applying a provided set of equivalences 264 /// between components of the symbol names in the profile. 265 class SampleProfileReaderItaniumRemapper { 266 public: 267 SampleProfileReaderItaniumRemapper(std::unique_ptr<MemoryBuffer> B, 268 std::unique_ptr<SymbolRemappingReader> SRR, 269 SampleProfileReader &R) 270 : Buffer(std::move(B)), Remappings(std::move(SRR)), Reader(R) { 271 assert(Remappings && "Remappings cannot be nullptr"); 272 } 273 274 /// Create a remapper from the given remapping file. The remapper will 275 /// be used for profile read in by Reader. 276 static ErrorOr<std::unique_ptr<SampleProfileReaderItaniumRemapper>> 277 create(StringRef Filename, vfs::FileSystem &FS, SampleProfileReader &Reader, 278 LLVMContext &C); 279 280 /// Create a remapper from the given Buffer. The remapper will 281 /// be used for profile read in by Reader. 282 static ErrorOr<std::unique_ptr<SampleProfileReaderItaniumRemapper>> 283 create(std::unique_ptr<MemoryBuffer> &B, SampleProfileReader &Reader, 284 LLVMContext &C); 285 286 /// Apply remappings to the profile read by Reader. 287 void applyRemapping(LLVMContext &Ctx); 288 289 bool hasApplied() { return RemappingApplied; } 290 291 /// Insert function name into remapper. 292 void insert(StringRef FunctionName) { Remappings->insert(FunctionName); } 293 294 /// Query whether there is equivalent in the remapper which has been 295 /// inserted. 296 bool exist(StringRef FunctionName) { 297 return Remappings->lookup(FunctionName); 298 } 299 300 /// Return the equivalent name in the profile for \p FunctionName if 301 /// it exists. 302 std::optional<StringRef> lookUpNameInProfile(StringRef FunctionName); 303 304 private: 305 // The buffer holding the content read from remapping file. 306 std::unique_ptr<MemoryBuffer> Buffer; 307 std::unique_ptr<SymbolRemappingReader> Remappings; 308 // Map remapping key to the name in the profile. By looking up the 309 // key in the remapper, a given new name can be mapped to the 310 // cannonical name using the NameMap. 311 DenseMap<SymbolRemappingReader::Key, StringRef> NameMap; 312 // The Reader the remapper is servicing. 313 SampleProfileReader &Reader; 314 // Indicate whether remapping has been applied to the profile read 315 // by Reader -- by calling applyRemapping. 316 bool RemappingApplied = false; 317 }; 318 319 /// Sample-based profile reader. 320 /// 321 /// Each profile contains sample counts for all the functions 322 /// executed. Inside each function, statements are annotated with the 323 /// collected samples on all the instructions associated with that 324 /// statement. 325 /// 326 /// For this to produce meaningful data, the program needs to be 327 /// compiled with some debug information (at minimum, line numbers: 328 /// -gline-tables-only). Otherwise, it will be impossible to match IR 329 /// instructions to the line numbers collected by the profiler. 330 /// 331 /// From the profile file, we are interested in collecting the 332 /// following information: 333 /// 334 /// * A list of functions included in the profile (mangled names). 335 /// 336 /// * For each function F: 337 /// 1. The total number of samples collected in F. 338 /// 339 /// 2. The samples collected at each line in F. To provide some 340 /// protection against source code shuffling, line numbers should 341 /// be relative to the start of the function. 342 /// 343 /// The reader supports two file formats: text and binary. The text format 344 /// is useful for debugging and testing, while the binary format is more 345 /// compact and I/O efficient. They can both be used interchangeably. 346 class SampleProfileReader { 347 public: 348 SampleProfileReader(std::unique_ptr<MemoryBuffer> B, LLVMContext &C, 349 SampleProfileFormat Format = SPF_None) 350 : Profiles(), Ctx(C), Buffer(std::move(B)), Format(Format) {} 351 352 virtual ~SampleProfileReader() = default; 353 354 /// Read and validate the file header. 355 virtual std::error_code readHeader() = 0; 356 357 /// Set the bits for FS discriminators. Parameter Pass specify the sequence 358 /// number, Pass == i is for the i-th round of adding FS discriminators. 359 /// Pass == 0 is for using base discriminators. 360 void setDiscriminatorMaskedBitFrom(FSDiscriminatorPass P) { 361 MaskedBitFrom = getFSPassBitEnd(P); 362 } 363 364 /// Get the bitmask the discriminators: For FS profiles, return the bit 365 /// mask for this pass. For non FS profiles, return (unsigned) -1. 366 uint32_t getDiscriminatorMask() const { 367 if (!ProfileIsFS) 368 return 0xFFFFFFFF; 369 assert((MaskedBitFrom != 0) && "MaskedBitFrom is not set properly"); 370 return getN1Bits(MaskedBitFrom); 371 } 372 373 /// The interface to read sample profiles from the associated file. 374 std::error_code read() { 375 if (std::error_code EC = readImpl()) 376 return EC; 377 if (Remapper) 378 Remapper->applyRemapping(Ctx); 379 FunctionSamples::UseMD5 = useMD5(); 380 return sampleprof_error::success; 381 } 382 383 /// Read sample profiles for the given functions. 384 std::error_code read(const DenseSet<StringRef> &FuncsToUse) { 385 DenseSet<StringRef> S; 386 for (StringRef F : FuncsToUse) 387 if (Profiles.find(FunctionId(F)) == Profiles.end()) 388 S.insert(F); 389 if (std::error_code EC = read(S, Profiles)) 390 return EC; 391 return sampleprof_error::success; 392 } 393 394 /// The implementaion to read sample profiles from the associated file. 395 virtual std::error_code readImpl() = 0; 396 397 /// Print the profile for \p FunctionSamples on stream \p OS. 398 void dumpFunctionProfile(const FunctionSamples &FS, raw_ostream &OS = dbgs()); 399 400 /// Collect functions with definitions in Module M. For reader which 401 /// support loading function profiles on demand, return true when the 402 /// reader has been given a module. Always return false for reader 403 /// which doesn't support loading function profiles on demand. 404 virtual bool collectFuncsFromModule() { return false; } 405 406 /// Print all the profiles on stream \p OS. 407 void dump(raw_ostream &OS = dbgs()); 408 409 /// Print all the profiles on stream \p OS in the JSON format. 410 void dumpJson(raw_ostream &OS = dbgs()); 411 412 /// Return the samples collected for function \p F. 413 FunctionSamples *getSamplesFor(const Function &F) { 414 // The function name may have been updated by adding suffix. Call 415 // a helper to (optionally) strip off suffixes so that we can 416 // match against the original function name in the profile. 417 StringRef CanonName = FunctionSamples::getCanonicalFnName(F); 418 return getSamplesFor(CanonName); 419 } 420 421 /// Return the samples collected for function \p F. 422 FunctionSamples *getSamplesFor(StringRef Fname) { 423 auto It = Profiles.find(FunctionId(Fname)); 424 if (It != Profiles.end()) 425 return &It->second; 426 427 if (FuncNameToProfNameMap && !FuncNameToProfNameMap->empty()) { 428 auto R = FuncNameToProfNameMap->find(FunctionId(Fname)); 429 if (R != FuncNameToProfNameMap->end()) { 430 Fname = R->second.stringRef(); 431 auto It = Profiles.find(FunctionId(Fname)); 432 if (It != Profiles.end()) 433 return &It->second; 434 } 435 } 436 437 if (Remapper) { 438 if (auto NameInProfile = Remapper->lookUpNameInProfile(Fname)) { 439 auto It = Profiles.find(FunctionId(*NameInProfile)); 440 if (It != Profiles.end()) 441 return &It->second; 442 } 443 } 444 return nullptr; 445 } 446 447 /// Return all the profiles. 448 SampleProfileMap &getProfiles() { return Profiles; } 449 450 /// Report a parse error message. 451 void reportError(int64_t LineNumber, const Twine &Msg) const { 452 Ctx.diagnose(DiagnosticInfoSampleProfile(Buffer->getBufferIdentifier(), 453 LineNumber, Msg)); 454 } 455 456 /// Create a sample profile reader appropriate to the file format. 457 /// Create a remapper underlying if RemapFilename is not empty. 458 /// Parameter P specifies the FSDiscriminatorPass. 459 static ErrorOr<std::unique_ptr<SampleProfileReader>> 460 create(StringRef Filename, LLVMContext &C, vfs::FileSystem &FS, 461 FSDiscriminatorPass P = FSDiscriminatorPass::Base, 462 StringRef RemapFilename = ""); 463 464 /// Create a sample profile reader from the supplied memory buffer. 465 /// Create a remapper underlying if RemapFilename is not empty. 466 /// Parameter P specifies the FSDiscriminatorPass. 467 static ErrorOr<std::unique_ptr<SampleProfileReader>> 468 create(std::unique_ptr<MemoryBuffer> &B, LLVMContext &C, vfs::FileSystem &FS, 469 FSDiscriminatorPass P = FSDiscriminatorPass::Base, 470 StringRef RemapFilename = ""); 471 472 /// Return the profile summary. 473 ProfileSummary &getSummary() const { return *Summary; } 474 475 MemoryBuffer *getBuffer() const { return Buffer.get(); } 476 477 /// \brief Return the profile format. 478 SampleProfileFormat getFormat() const { return Format; } 479 480 /// Whether input profile is based on pseudo probes. 481 bool profileIsProbeBased() const { return ProfileIsProbeBased; } 482 483 /// Whether input profile is fully context-sensitive. 484 bool profileIsCS() const { return ProfileIsCS; } 485 486 /// Whether input profile contains ShouldBeInlined contexts. 487 bool profileIsPreInlined() const { return ProfileIsPreInlined; } 488 489 /// Whether input profile is flow-sensitive. 490 bool profileIsFS() const { return ProfileIsFS; } 491 492 virtual std::unique_ptr<ProfileSymbolList> getProfileSymbolList() { 493 return nullptr; 494 }; 495 496 /// It includes all the names that have samples either in outline instance 497 /// or inline instance. 498 virtual std::vector<FunctionId> *getNameTable() { return nullptr; } 499 virtual bool dumpSectionInfo(raw_ostream &OS = dbgs()) { return false; }; 500 501 /// Return whether names in the profile are all MD5 numbers. 502 bool useMD5() const { return ProfileIsMD5; } 503 504 /// Force the profile to use MD5 in Sample contexts, even if function names 505 /// are present. 506 virtual void setProfileUseMD5() { ProfileIsMD5 = true; } 507 508 /// Don't read profile without context if the flag is set. 509 void setSkipFlatProf(bool Skip) { SkipFlatProf = Skip; } 510 511 /// Return whether any name in the profile contains ".__uniq." suffix. 512 virtual bool hasUniqSuffix() { return false; } 513 514 SampleProfileReaderItaniumRemapper *getRemapper() { return Remapper.get(); } 515 516 void setModule(const Module *Mod) { M = Mod; } 517 518 void setFuncNameToProfNameMap( 519 const HashKeyMap<std::unordered_map, FunctionId, FunctionId> &FPMap) { 520 FuncNameToProfNameMap = &FPMap; 521 } 522 523 protected: 524 /// Map every function to its associated profile. 525 /// 526 /// The profile of every function executed at runtime is collected 527 /// in the structure FunctionSamples. This maps function objects 528 /// to their corresponding profiles. 529 SampleProfileMap Profiles; 530 531 /// LLVM context used to emit diagnostics. 532 LLVMContext &Ctx; 533 534 /// Memory buffer holding the profile file. 535 std::unique_ptr<MemoryBuffer> Buffer; 536 537 /// Profile summary information. 538 std::unique_ptr<ProfileSummary> Summary; 539 540 /// Take ownership of the summary of this reader. 541 static std::unique_ptr<ProfileSummary> 542 takeSummary(SampleProfileReader &Reader) { 543 return std::move(Reader.Summary); 544 } 545 546 /// Compute summary for this profile. 547 void computeSummary(); 548 549 /// Read sample profiles for the given functions and write them to the given 550 /// profile map. Currently it's only used for extended binary format to load 551 /// the profiles on-demand. 552 virtual std::error_code read(const DenseSet<StringRef> &FuncsToUse, 553 SampleProfileMap &Profiles) { 554 return sampleprof_error::not_implemented; 555 } 556 557 std::unique_ptr<SampleProfileReaderItaniumRemapper> Remapper; 558 559 // A map pointer to the FuncNameToProfNameMap in SampleProfileLoader, 560 // which maps the function name to the matched profile name. This is used 561 // for sample loader to look up profile using the new name. 562 const HashKeyMap<std::unordered_map, FunctionId, FunctionId> 563 *FuncNameToProfNameMap = nullptr; 564 565 // A map from a function's context hash to its meta data section range, used 566 // for on-demand read function profile metadata. 567 std::unordered_map<uint64_t, std::pair<const uint8_t *, const uint8_t *>> 568 FuncMetadataIndex; 569 570 std::pair<const uint8_t *, const uint8_t *> ProfileSecRange; 571 572 /// Whether the profile has attribute metadata. 573 bool ProfileHasAttribute = false; 574 575 /// \brief Whether samples are collected based on pseudo probes. 576 bool ProfileIsProbeBased = false; 577 578 /// Whether function profiles are context-sensitive flat profiles. 579 bool ProfileIsCS = false; 580 581 /// Whether function profile contains ShouldBeInlined contexts. 582 bool ProfileIsPreInlined = false; 583 584 /// Number of context-sensitive profiles. 585 uint32_t CSProfileCount = 0; 586 587 /// Whether the function profiles use FS discriminators. 588 bool ProfileIsFS = false; 589 590 /// \brief The format of sample. 591 SampleProfileFormat Format = SPF_None; 592 593 /// \brief The current module being compiled if SampleProfileReader 594 /// is used by compiler. If SampleProfileReader is used by other 595 /// tools which are not compiler, M is usually nullptr. 596 const Module *M = nullptr; 597 598 /// Zero out the discriminator bits higher than bit MaskedBitFrom (0 based). 599 /// The default is to keep all the bits. 600 uint32_t MaskedBitFrom = 31; 601 602 /// Whether the profile uses MD5 for Sample Contexts and function names. This 603 /// can be one-way overriden by the user to force use MD5. 604 bool ProfileIsMD5 = false; 605 606 /// If SkipFlatProf is true, skip functions marked with !Flat in text mode or 607 /// sections with SecFlagFlat flag in ExtBinary mode. 608 bool SkipFlatProf = false; 609 }; 610 611 class SampleProfileReaderText : public SampleProfileReader { 612 public: 613 SampleProfileReaderText(std::unique_ptr<MemoryBuffer> B, LLVMContext &C) 614 : SampleProfileReader(std::move(B), C, SPF_Text) {} 615 616 /// Read and validate the file header. 617 std::error_code readHeader() override { return sampleprof_error::success; } 618 619 /// Read sample profiles from the associated file. 620 std::error_code readImpl() override; 621 622 /// Return true if \p Buffer is in the format supported by this class. 623 static bool hasFormat(const MemoryBuffer &Buffer); 624 625 /// Text format sample profile does not support MD5 for now. 626 void setProfileUseMD5() override {} 627 628 private: 629 /// CSNameTable is used to save full context vectors. This serves as an 630 /// underlying immutable buffer for all clients. 631 std::list<SampleContextFrameVector> CSNameTable; 632 }; 633 634 class SampleProfileReaderBinary : public SampleProfileReader { 635 public: 636 SampleProfileReaderBinary(std::unique_ptr<MemoryBuffer> B, LLVMContext &C, 637 SampleProfileFormat Format = SPF_None) 638 : SampleProfileReader(std::move(B), C, Format) {} 639 640 /// Read and validate the file header. 641 std::error_code readHeader() override; 642 643 /// Read sample profiles from the associated file. 644 std::error_code readImpl() override; 645 646 /// It includes all the names that have samples either in outline instance 647 /// or inline instance. 648 std::vector<FunctionId> *getNameTable() override { 649 return &NameTable; 650 } 651 652 protected: 653 /// Read a numeric value of type T from the profile. 654 /// 655 /// If an error occurs during decoding, a diagnostic message is emitted and 656 /// EC is set. 657 /// 658 /// \returns the read value. 659 template <typename T> ErrorOr<T> readNumber(); 660 661 /// Read a numeric value of type T from the profile. The value is saved 662 /// without encoded. 663 template <typename T> ErrorOr<T> readUnencodedNumber(); 664 665 /// Read a string from the profile. 666 /// 667 /// If an error occurs during decoding, a diagnostic message is emitted and 668 /// EC is set. 669 /// 670 /// \returns the read value. 671 ErrorOr<StringRef> readString(); 672 673 /// Read the string index and check whether it overflows the table. 674 template <typename T> inline ErrorOr<size_t> readStringIndex(T &Table); 675 676 /// Read the next function profile instance. 677 std::error_code readFuncProfile(const uint8_t *Start); 678 std::error_code readFuncProfile(const uint8_t *Start, 679 SampleProfileMap &Profiles); 680 681 /// Read the contents of the given profile instance. 682 std::error_code readProfile(FunctionSamples &FProfile); 683 684 /// Read the contents of Magic number and Version number. 685 std::error_code readMagicIdent(); 686 687 /// Read profile summary. 688 std::error_code readSummary(); 689 690 /// Read the whole name table. 691 std::error_code readNameTable(); 692 693 /// Read a string indirectly via the name table. Optionally return the index. 694 ErrorOr<FunctionId> readStringFromTable(size_t *RetIdx = nullptr); 695 696 /// Read a context indirectly via the CSNameTable. Optionally return the 697 /// index. 698 ErrorOr<SampleContextFrames> readContextFromTable(size_t *RetIdx = nullptr); 699 700 /// Read a context indirectly via the CSNameTable if the profile has context, 701 /// otherwise same as readStringFromTable, also return its hash value. 702 ErrorOr<std::pair<SampleContext, uint64_t>> readSampleContextFromTable(); 703 704 /// Points to the current location in the buffer. 705 const uint8_t *Data = nullptr; 706 707 /// Points to the end of the buffer. 708 const uint8_t *End = nullptr; 709 710 /// Function name table. 711 std::vector<FunctionId> NameTable; 712 713 /// CSNameTable is used to save full context vectors. It is the backing buffer 714 /// for SampleContextFrames. 715 std::vector<SampleContextFrameVector> CSNameTable; 716 717 /// Table to cache MD5 values of sample contexts corresponding to 718 /// readSampleContextFromTable(), used to index into Profiles or 719 /// FuncOffsetTable. 720 std::vector<uint64_t> MD5SampleContextTable; 721 722 /// The starting address of the table of MD5 values of sample contexts. For 723 /// fixed length MD5 non-CS profile it is same as MD5NameMemStart because 724 /// hashes of non-CS contexts are already in the profile. Otherwise it points 725 /// to the start of MD5SampleContextTable. 726 const uint64_t *MD5SampleContextStart = nullptr; 727 728 private: 729 std::error_code readSummaryEntry(std::vector<ProfileSummaryEntry> &Entries); 730 virtual std::error_code verifySPMagic(uint64_t Magic) = 0; 731 }; 732 733 class SampleProfileReaderRawBinary : public SampleProfileReaderBinary { 734 private: 735 std::error_code verifySPMagic(uint64_t Magic) override; 736 737 public: 738 SampleProfileReaderRawBinary(std::unique_ptr<MemoryBuffer> B, LLVMContext &C, 739 SampleProfileFormat Format = SPF_Binary) 740 : SampleProfileReaderBinary(std::move(B), C, Format) {} 741 742 /// \brief Return true if \p Buffer is in the format supported by this class. 743 static bool hasFormat(const MemoryBuffer &Buffer); 744 }; 745 746 /// SampleProfileReaderExtBinaryBase/SampleProfileWriterExtBinaryBase defines 747 /// the basic structure of the extensible binary format. 748 /// The format is organized in sections except the magic and version number 749 /// at the beginning. There is a section table before all the sections, and 750 /// each entry in the table describes the entry type, start, size and 751 /// attributes. The format in each section is defined by the section itself. 752 /// 753 /// It is easy to add a new section while maintaining the backward 754 /// compatibility of the profile. Nothing extra needs to be done. If we want 755 /// to extend an existing section, like add cache misses information in 756 /// addition to the sample count in the profile body, we can add a new section 757 /// with the extension and retire the existing section, and we could choose 758 /// to keep the parser of the old section if we want the reader to be able 759 /// to read both new and old format profile. 760 /// 761 /// SampleProfileReaderExtBinary/SampleProfileWriterExtBinary define the 762 /// commonly used sections of a profile in extensible binary format. It is 763 /// possible to define other types of profile inherited from 764 /// SampleProfileReaderExtBinaryBase/SampleProfileWriterExtBinaryBase. 765 class SampleProfileReaderExtBinaryBase : public SampleProfileReaderBinary { 766 private: 767 std::error_code decompressSection(const uint8_t *SecStart, 768 const uint64_t SecSize, 769 const uint8_t *&DecompressBuf, 770 uint64_t &DecompressBufSize); 771 772 BumpPtrAllocator Allocator; 773 774 protected: 775 std::vector<SecHdrTableEntry> SecHdrTable; 776 std::error_code readSecHdrTableEntry(uint64_t Idx); 777 std::error_code readSecHdrTable(); 778 779 std::error_code readFuncMetadata(bool ProfileHasAttribute, 780 SampleProfileMap &Profiles); 781 std::error_code readFuncMetadata(bool ProfileHasAttribute); 782 std::error_code readFuncMetadata(bool ProfileHasAttribute, 783 FunctionSamples *FProfile); 784 std::error_code readFuncOffsetTable(); 785 std::error_code readFuncProfiles(); 786 std::error_code readFuncProfiles(const DenseSet<StringRef> &FuncsToUse, 787 SampleProfileMap &Profiles); 788 std::error_code readNameTableSec(bool IsMD5, bool FixedLengthMD5); 789 std::error_code readCSNameTableSec(); 790 std::error_code readProfileSymbolList(); 791 792 std::error_code readHeader() override; 793 std::error_code verifySPMagic(uint64_t Magic) override = 0; 794 virtual std::error_code readOneSection(const uint8_t *Start, uint64_t Size, 795 const SecHdrTableEntry &Entry); 796 // placeholder for subclasses to dispatch their own section readers. 797 virtual std::error_code readCustomSection(const SecHdrTableEntry &Entry) = 0; 798 799 /// Determine which container readFuncOffsetTable() should populate, the list 800 /// FuncOffsetList or the map FuncOffsetTable. 801 bool useFuncOffsetList() const; 802 803 std::unique_ptr<ProfileSymbolList> ProfSymList; 804 805 /// The table mapping from a function context's MD5 to the offset of its 806 /// FunctionSample towards file start. 807 /// At most one of FuncOffsetTable and FuncOffsetList is populated. 808 DenseMap<hash_code, uint64_t> FuncOffsetTable; 809 810 /// The list version of FuncOffsetTable. This is used if every entry is 811 /// being accessed. 812 std::vector<std::pair<SampleContext, uint64_t>> FuncOffsetList; 813 814 /// The set containing the functions to use when compiling a module. 815 DenseSet<StringRef> FuncsToUse; 816 817 public: 818 SampleProfileReaderExtBinaryBase(std::unique_ptr<MemoryBuffer> B, 819 LLVMContext &C, SampleProfileFormat Format) 820 : SampleProfileReaderBinary(std::move(B), C, Format) {} 821 822 /// Read sample profiles in extensible format from the associated file. 823 std::error_code readImpl() override; 824 825 /// Get the total size of all \p Type sections. 826 uint64_t getSectionSize(SecType Type); 827 /// Get the total size of header and all sections. 828 uint64_t getFileSize(); 829 bool dumpSectionInfo(raw_ostream &OS = dbgs()) override; 830 831 /// Collect functions with definitions in Module M. Return true if 832 /// the reader has been given a module. 833 bool collectFuncsFromModule() override; 834 835 std::unique_ptr<ProfileSymbolList> getProfileSymbolList() override { 836 return std::move(ProfSymList); 837 }; 838 839 private: 840 /// Read the profiles on-demand for the given functions. This is used after 841 /// stale call graph matching finds new functions whose profiles aren't loaded 842 /// at the beginning and we need to loaded the profiles explicitly for 843 /// potential matching. 844 std::error_code read(const DenseSet<StringRef> &FuncsToUse, 845 SampleProfileMap &Profiles) override; 846 }; 847 848 class SampleProfileReaderExtBinary : public SampleProfileReaderExtBinaryBase { 849 private: 850 std::error_code verifySPMagic(uint64_t Magic) override; 851 std::error_code readCustomSection(const SecHdrTableEntry &Entry) override { 852 // Update the data reader pointer to the end of the section. 853 Data = End; 854 return sampleprof_error::success; 855 }; 856 857 public: 858 SampleProfileReaderExtBinary(std::unique_ptr<MemoryBuffer> B, LLVMContext &C, 859 SampleProfileFormat Format = SPF_Ext_Binary) 860 : SampleProfileReaderExtBinaryBase(std::move(B), C, Format) {} 861 862 /// \brief Return true if \p Buffer is in the format supported by this class. 863 static bool hasFormat(const MemoryBuffer &Buffer); 864 }; 865 866 using InlineCallStack = SmallVector<FunctionSamples *, 10>; 867 868 // Supported histogram types in GCC. Currently, we only need support for 869 // call target histograms. 870 enum HistType { 871 HIST_TYPE_INTERVAL, 872 HIST_TYPE_POW2, 873 HIST_TYPE_SINGLE_VALUE, 874 HIST_TYPE_CONST_DELTA, 875 HIST_TYPE_INDIR_CALL, 876 HIST_TYPE_AVERAGE, 877 HIST_TYPE_IOR, 878 HIST_TYPE_INDIR_CALL_TOPN 879 }; 880 881 class SampleProfileReaderGCC : public SampleProfileReader { 882 public: 883 SampleProfileReaderGCC(std::unique_ptr<MemoryBuffer> B, LLVMContext &C) 884 : SampleProfileReader(std::move(B), C, SPF_GCC), 885 GcovBuffer(Buffer.get()) {} 886 887 /// Read and validate the file header. 888 std::error_code readHeader() override; 889 890 /// Read sample profiles from the associated file. 891 std::error_code readImpl() override; 892 893 /// Return true if \p Buffer is in the format supported by this class. 894 static bool hasFormat(const MemoryBuffer &Buffer); 895 896 protected: 897 std::error_code readNameTable(); 898 std::error_code readOneFunctionProfile(const InlineCallStack &InlineStack, 899 bool Update, uint32_t Offset); 900 std::error_code readFunctionProfiles(); 901 std::error_code skipNextWord(); 902 template <typename T> ErrorOr<T> readNumber(); 903 ErrorOr<StringRef> readString(); 904 905 /// Read the section tag and check that it's the same as \p Expected. 906 std::error_code readSectionTag(uint32_t Expected); 907 908 /// GCOV buffer containing the profile. 909 GCOVBuffer GcovBuffer; 910 911 /// Function names in this profile. 912 std::vector<std::string> Names; 913 914 /// GCOV tags used to separate sections in the profile file. 915 static const uint32_t GCOVTagAFDOFileNames = 0xaa000000; 916 static const uint32_t GCOVTagAFDOFunction = 0xac000000; 917 }; 918 919 } // end namespace sampleprof 920 921 } // end namespace llvm 922 923 #endif // LLVM_PROFILEDATA_SAMPLEPROFREADER_H 924