| 1 | //===- InstrProfReader.cpp - Instrumented profiling reader ----------------===// |
| 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 support for reading profiling data for clang's |
| 10 | // instrumentation based PGO and coverage. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/ProfileData/InstrProfReader.h" |
| 15 | #include "llvm/ADT/ArrayRef.h" |
| 16 | #include "llvm/ADT/DenseMap.h" |
| 17 | #include "llvm/ADT/StringExtras.h" |
| 18 | #include "llvm/ADT/StringRef.h" |
| 19 | #include "llvm/IR/ProfileSummary.h" |
| 20 | #include "llvm/ProfileData/InstrProf.h" |
| 21 | // #include "llvm/ProfileData/MemProf.h" |
| 22 | #include "llvm/ProfileData/MemProfRadixTree.h" |
| 23 | #include "llvm/ProfileData/ProfileCommon.h" |
| 24 | #include "llvm/ProfileData/SymbolRemappingReader.h" |
| 25 | #include "llvm/Support/Endian.h" |
| 26 | #include "llvm/Support/Error.h" |
| 27 | #include "llvm/Support/ErrorOr.h" |
| 28 | #include "llvm/Support/FormatVariadic.h" |
| 29 | #include "llvm/Support/MemoryBuffer.h" |
| 30 | #include "llvm/Support/VirtualFileSystem.h" |
| 31 | #include <algorithm> |
| 32 | #include <cstddef> |
| 33 | #include <cstdint> |
| 34 | #include <limits> |
| 35 | #include <memory> |
| 36 | #include <optional> |
| 37 | #include <system_error> |
| 38 | #include <utility> |
| 39 | #include <vector> |
| 40 | |
| 41 | using namespace llvm; |
| 42 | |
| 43 | // Extracts the variant information from the top 32 bits in the version and |
| 44 | // returns an enum specifying the variants present. |
| 45 | static InstrProfKind getProfileKindFromVersion(uint64_t Version) { |
| 46 | InstrProfKind ProfileKind = InstrProfKind::Unknown; |
| 47 | if (Version & VARIANT_MASK_IR_PROF) { |
| 48 | ProfileKind |= InstrProfKind::IRInstrumentation; |
| 49 | } |
| 50 | if (Version & VARIANT_MASK_CSIR_PROF) { |
| 51 | ProfileKind |= InstrProfKind::ContextSensitive; |
| 52 | } |
| 53 | if (Version & VARIANT_MASK_INSTR_ENTRY) { |
| 54 | ProfileKind |= InstrProfKind::FunctionEntryInstrumentation; |
| 55 | } |
| 56 | if (Version & VARIANT_MASK_INSTR_LOOP_ENTRIES) { |
| 57 | ProfileKind |= InstrProfKind::LoopEntriesInstrumentation; |
| 58 | } |
| 59 | if (Version & VARIANT_MASK_BYTE_COVERAGE) { |
| 60 | ProfileKind |= InstrProfKind::SingleByteCoverage; |
| 61 | } |
| 62 | if (Version & VARIANT_MASK_FUNCTION_ENTRY_ONLY) { |
| 63 | ProfileKind |= InstrProfKind::FunctionEntryOnly; |
| 64 | } |
| 65 | if (Version & VARIANT_MASK_MEMPROF) { |
| 66 | ProfileKind |= InstrProfKind::MemProf; |
| 67 | } |
| 68 | if (Version & VARIANT_MASK_TEMPORAL_PROF) { |
| 69 | ProfileKind |= InstrProfKind::TemporalProfile; |
| 70 | } |
| 71 | return ProfileKind; |
| 72 | } |
| 73 | |
| 74 | static Expected<std::unique_ptr<MemoryBuffer>> |
| 75 | setupMemoryBuffer(const Twine &Filename, vfs::FileSystem &FS) { |
| 76 | auto BufferOrErr = Filename.str() == "-" ? MemoryBuffer::getSTDIN() |
| 77 | : FS.getBufferForFile(Name: Filename); |
| 78 | if (std::error_code EC = BufferOrErr.getError()) |
| 79 | return errorCodeToError(EC); |
| 80 | return std::move(BufferOrErr.get()); |
| 81 | } |
| 82 | |
| 83 | static Error initializeReader(InstrProfReader &Reader) { |
| 84 | return Reader.readHeader(); |
| 85 | } |
| 86 | |
| 87 | /// Read a list of binary ids from a profile that consist of |
| 88 | /// a. uint64_t binary id length |
| 89 | /// b. uint8_t binary id data |
| 90 | /// c. uint8_t padding (if necessary) |
| 91 | /// This function is shared between raw and indexed profiles. |
| 92 | /// Raw profiles are in host-endian format, and indexed profiles are in |
| 93 | /// little-endian format. So, this function takes an argument indicating the |
| 94 | /// associated endian format to read the binary ids correctly. |
| 95 | static Error |
| 96 | readBinaryIdsInternal(const MemoryBuffer &DataBuffer, |
| 97 | ArrayRef<uint8_t> BinaryIdsBuffer, |
| 98 | std::vector<llvm::object::BuildID> &BinaryIds, |
| 99 | const llvm::endianness Endian) { |
| 100 | using namespace support; |
| 101 | |
| 102 | const uint64_t BinaryIdsSize = BinaryIdsBuffer.size(); |
| 103 | const uint8_t *BinaryIdsStart = BinaryIdsBuffer.data(); |
| 104 | |
| 105 | if (BinaryIdsSize == 0) |
| 106 | return Error::success(); |
| 107 | |
| 108 | const uint8_t *BI = BinaryIdsStart; |
| 109 | const uint8_t *BIEnd = BinaryIdsStart + BinaryIdsSize; |
| 110 | const uint8_t *End = |
| 111 | reinterpret_cast<const uint8_t *>(DataBuffer.getBufferEnd()); |
| 112 | |
| 113 | while (BI < BIEnd) { |
| 114 | size_t Remaining = BIEnd - BI; |
| 115 | // There should be enough left to read the binary id length. |
| 116 | if (Remaining < sizeof(uint64_t)) |
| 117 | return make_error<InstrProfError>( |
| 118 | Args: instrprof_error::malformed, |
| 119 | Args: "not enough data to read binary id length" ); |
| 120 | |
| 121 | uint64_t BILen = endian::readNext<uint64_t>(memory&: BI, endian: Endian); |
| 122 | if (BILen == 0) |
| 123 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
| 124 | Args: "binary id length is 0" ); |
| 125 | |
| 126 | Remaining = BIEnd - BI; |
| 127 | // There should be enough left to read the binary id data. |
| 128 | if (Remaining < alignToPowerOf2(Value: BILen, Align: sizeof(uint64_t))) |
| 129 | return make_error<InstrProfError>( |
| 130 | Args: instrprof_error::malformed, Args: "not enough data to read binary id data" ); |
| 131 | |
| 132 | // Add binary id to the binary ids list. |
| 133 | BinaryIds.push_back(x: object::BuildID(BI, BI + BILen)); |
| 134 | |
| 135 | // Increment by binary id data length, which aligned to the size of uint64. |
| 136 | BI += alignToPowerOf2(Value: BILen, Align: sizeof(uint64_t)); |
| 137 | if (BI > End) |
| 138 | return make_error<InstrProfError>( |
| 139 | Args: instrprof_error::malformed, |
| 140 | Args: "binary id section is greater than buffer size" ); |
| 141 | } |
| 142 | |
| 143 | return Error::success(); |
| 144 | } |
| 145 | |
| 146 | static void printBinaryIdsInternal(raw_ostream &OS, |
| 147 | ArrayRef<llvm::object::BuildID> BinaryIds) { |
| 148 | OS << "Binary IDs: \n" ; |
| 149 | for (const auto &BI : BinaryIds) { |
| 150 | for (auto I : BI) |
| 151 | OS << format(Fmt: "%02x" , Vals: I); |
| 152 | OS << "\n" ; |
| 153 | } |
| 154 | } |
| 155 | |
| 156 | Expected<std::unique_ptr<InstrProfReader>> InstrProfReader::create( |
| 157 | const Twine &Path, vfs::FileSystem &FS, |
| 158 | const InstrProfCorrelator *Correlator, |
| 159 | const object::BuildIDFetcher *BIDFetcher, |
| 160 | const InstrProfCorrelator::ProfCorrelatorKind BIDFetcherCorrelatorKind, |
| 161 | std::function<void(Error)> Warn) { |
| 162 | // Set up the buffer to read. |
| 163 | auto BufferOrError = setupMemoryBuffer(Filename: Path, FS); |
| 164 | if (Error E = BufferOrError.takeError()) |
| 165 | return std::move(E); |
| 166 | return InstrProfReader::create(Buffer: std::move(BufferOrError.get()), Correlator, |
| 167 | BIDFetcher, BIDFetcherCorrelatorKind, Warn); |
| 168 | } |
| 169 | |
| 170 | Expected<std::unique_ptr<InstrProfReader>> InstrProfReader::create( |
| 171 | std::unique_ptr<MemoryBuffer> Buffer, const InstrProfCorrelator *Correlator, |
| 172 | const object::BuildIDFetcher *BIDFetcher, |
| 173 | const InstrProfCorrelator::ProfCorrelatorKind BIDFetcherCorrelatorKind, |
| 174 | std::function<void(Error)> Warn) { |
| 175 | if (Buffer->getBufferSize() == 0) |
| 176 | return make_error<InstrProfError>(Args: instrprof_error::empty_raw_profile); |
| 177 | |
| 178 | std::unique_ptr<InstrProfReader> Result; |
| 179 | // Create the reader. |
| 180 | if (IndexedInstrProfReader::hasFormat(DataBuffer: *Buffer)) |
| 181 | Result.reset(p: new IndexedInstrProfReader(std::move(Buffer))); |
| 182 | else if (RawInstrProfReader64::hasFormat(DataBuffer: *Buffer)) |
| 183 | Result.reset(p: new RawInstrProfReader64(std::move(Buffer), Correlator, |
| 184 | BIDFetcher, BIDFetcherCorrelatorKind, |
| 185 | Warn)); |
| 186 | else if (RawInstrProfReader32::hasFormat(DataBuffer: *Buffer)) |
| 187 | Result.reset(p: new RawInstrProfReader32(std::move(Buffer), Correlator, |
| 188 | BIDFetcher, BIDFetcherCorrelatorKind, |
| 189 | Warn)); |
| 190 | else if (TextInstrProfReader::hasFormat(Buffer: *Buffer)) |
| 191 | Result.reset(p: new TextInstrProfReader(std::move(Buffer))); |
| 192 | else |
| 193 | return make_error<InstrProfError>(Args: instrprof_error::unrecognized_format); |
| 194 | |
| 195 | // Initialize the reader and return the result. |
| 196 | if (Error E = initializeReader(Reader&: *Result)) |
| 197 | return std::move(E); |
| 198 | |
| 199 | return std::move(Result); |
| 200 | } |
| 201 | |
| 202 | Expected<std::unique_ptr<IndexedInstrProfReader>> |
| 203 | IndexedInstrProfReader::create(const Twine &Path, vfs::FileSystem &FS, |
| 204 | const Twine &RemappingPath) { |
| 205 | // Set up the buffer to read. |
| 206 | auto BufferOrError = setupMemoryBuffer(Filename: Path, FS); |
| 207 | if (Error E = BufferOrError.takeError()) |
| 208 | return std::move(E); |
| 209 | |
| 210 | // Set up the remapping buffer if requested. |
| 211 | std::unique_ptr<MemoryBuffer> RemappingBuffer; |
| 212 | std::string RemappingPathStr = RemappingPath.str(); |
| 213 | if (!RemappingPathStr.empty()) { |
| 214 | auto RemappingBufferOrError = setupMemoryBuffer(Filename: RemappingPathStr, FS); |
| 215 | if (Error E = RemappingBufferOrError.takeError()) |
| 216 | return std::move(E); |
| 217 | RemappingBuffer = std::move(RemappingBufferOrError.get()); |
| 218 | } |
| 219 | |
| 220 | return IndexedInstrProfReader::create(Buffer: std::move(BufferOrError.get()), |
| 221 | RemappingBuffer: std::move(RemappingBuffer)); |
| 222 | } |
| 223 | |
| 224 | Expected<std::unique_ptr<IndexedInstrProfReader>> |
| 225 | IndexedInstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer, |
| 226 | std::unique_ptr<MemoryBuffer> RemappingBuffer) { |
| 227 | // Create the reader. |
| 228 | if (!IndexedInstrProfReader::hasFormat(DataBuffer: *Buffer)) |
| 229 | return make_error<InstrProfError>(Args: instrprof_error::bad_magic); |
| 230 | auto Result = std::make_unique<IndexedInstrProfReader>( |
| 231 | args: std::move(Buffer), args: std::move(RemappingBuffer)); |
| 232 | |
| 233 | // Initialize the reader and return the result. |
| 234 | if (Error E = initializeReader(Reader&: *Result)) |
| 235 | return std::move(E); |
| 236 | |
| 237 | return std::move(Result); |
| 238 | } |
| 239 | |
| 240 | bool TextInstrProfReader::hasFormat(const MemoryBuffer &Buffer) { |
| 241 | // Verify that this really looks like plain ASCII text by checking a |
| 242 | // 'reasonable' number of characters (up to profile magic size). |
| 243 | size_t count = std::min(a: Buffer.getBufferSize(), b: sizeof(uint64_t)); |
| 244 | StringRef buffer = Buffer.getBufferStart(); |
| 245 | return count == 0 || |
| 246 | std::all_of(first: buffer.begin(), last: buffer.begin() + count, |
| 247 | pred: [](char c) { return isPrint(C: c) || isSpace(C: c); }); |
| 248 | } |
| 249 | |
| 250 | // Read the profile variant flag from the header: ":FE" means this is a FE |
| 251 | // generated profile. ":IR" means this is an IR level profile. Other strings |
| 252 | // with a leading ':' will be reported an error format. |
| 253 | Error TextInstrProfReader::() { |
| 254 | Symtab.reset(p: new InstrProfSymtab()); |
| 255 | |
| 256 | while (Line->starts_with(Prefix: ":" )) { |
| 257 | StringRef Str = Line->substr(Start: 1); |
| 258 | if (Str.equals_insensitive(RHS: "ir" )) |
| 259 | ProfileKind |= InstrProfKind::IRInstrumentation; |
| 260 | else if (Str.equals_insensitive(RHS: "fe" )) |
| 261 | ProfileKind |= InstrProfKind::FrontendInstrumentation; |
| 262 | else if (Str.equals_insensitive(RHS: "csir" )) { |
| 263 | ProfileKind |= InstrProfKind::IRInstrumentation; |
| 264 | ProfileKind |= InstrProfKind::ContextSensitive; |
| 265 | } else if (Str.equals_insensitive(RHS: "entry_first" )) |
| 266 | ProfileKind |= InstrProfKind::FunctionEntryInstrumentation; |
| 267 | else if (Str.equals_insensitive(RHS: "not_entry_first" )) |
| 268 | ProfileKind &= ~InstrProfKind::FunctionEntryInstrumentation; |
| 269 | else if (Str.equals_insensitive(RHS: "instrument_loop_entries" )) |
| 270 | ProfileKind |= InstrProfKind::LoopEntriesInstrumentation; |
| 271 | else if (Str.equals_insensitive(RHS: "single_byte_coverage" )) |
| 272 | ProfileKind |= InstrProfKind::SingleByteCoverage; |
| 273 | else if (Str.equals_insensitive(RHS: "temporal_prof_traces" )) { |
| 274 | ProfileKind |= InstrProfKind::TemporalProfile; |
| 275 | if (auto Err = readTemporalProfTraceData()) |
| 276 | return error(E: std::move(Err)); |
| 277 | } else |
| 278 | return error(Err: instrprof_error::bad_header); |
| 279 | ++Line; |
| 280 | } |
| 281 | return success(); |
| 282 | } |
| 283 | |
| 284 | /// Temporal profile trace data is stored in the header immediately after |
| 285 | /// ":temporal_prof_traces". The first integer is the number of traces, the |
| 286 | /// second integer is the stream size, then the following lines are the actual |
| 287 | /// traces which consist of a weight and a comma separated list of function |
| 288 | /// names. |
| 289 | Error TextInstrProfReader::readTemporalProfTraceData() { |
| 290 | if ((++Line).is_at_end()) |
| 291 | return error(Err: instrprof_error::eof); |
| 292 | |
| 293 | uint32_t NumTraces; |
| 294 | if (Line->getAsInteger(Radix: 0, Result&: NumTraces)) |
| 295 | return error(Err: instrprof_error::malformed); |
| 296 | |
| 297 | if ((++Line).is_at_end()) |
| 298 | return error(Err: instrprof_error::eof); |
| 299 | |
| 300 | if (Line->getAsInteger(Radix: 0, Result&: TemporalProfTraceStreamSize)) |
| 301 | return error(Err: instrprof_error::malformed); |
| 302 | |
| 303 | for (uint32_t i = 0; i < NumTraces; i++) { |
| 304 | if ((++Line).is_at_end()) |
| 305 | return error(Err: instrprof_error::eof); |
| 306 | |
| 307 | TemporalProfTraceTy Trace; |
| 308 | if (Line->getAsInteger(Radix: 0, Result&: Trace.Weight)) |
| 309 | return error(Err: instrprof_error::malformed); |
| 310 | |
| 311 | if ((++Line).is_at_end()) |
| 312 | return error(Err: instrprof_error::eof); |
| 313 | |
| 314 | SmallVector<StringRef> FuncNames; |
| 315 | Line->split(A&: FuncNames, Separator: "," , /*MaxSplit=*/-1, /*KeepEmpty=*/false); |
| 316 | for (auto &FuncName : FuncNames) |
| 317 | Trace.FunctionNameRefs.push_back( |
| 318 | x: IndexedInstrProf::ComputeHash(K: FuncName.trim())); |
| 319 | TemporalProfTraces.push_back(Elt: std::move(Trace)); |
| 320 | } |
| 321 | return success(); |
| 322 | } |
| 323 | |
| 324 | Error |
| 325 | TextInstrProfReader::readValueProfileData(InstrProfRecord &Record) { |
| 326 | |
| 327 | #define CHECK_LINE_END(Line) \ |
| 328 | if (Line.is_at_end()) \ |
| 329 | return error(instrprof_error::truncated); |
| 330 | #define READ_NUM(Str, Dst) \ |
| 331 | if ((Str).getAsInteger(10, (Dst))) \ |
| 332 | return error(instrprof_error::malformed); |
| 333 | #define VP_READ_ADVANCE(Val) \ |
| 334 | CHECK_LINE_END(Line); \ |
| 335 | uint32_t Val; \ |
| 336 | READ_NUM((*Line), (Val)); \ |
| 337 | Line++; |
| 338 | |
| 339 | if (Line.is_at_end()) |
| 340 | return success(); |
| 341 | |
| 342 | uint32_t NumValueKinds; |
| 343 | if (Line->getAsInteger(Radix: 10, Result&: NumValueKinds)) { |
| 344 | // No value profile data |
| 345 | return success(); |
| 346 | } |
| 347 | if (NumValueKinds == 0 || NumValueKinds > IPVK_Last + 1) |
| 348 | return error(Err: instrprof_error::malformed, |
| 349 | ErrMsg: "number of value kinds is invalid" ); |
| 350 | Line++; |
| 351 | |
| 352 | for (uint32_t VK = 0; VK < NumValueKinds; VK++) { |
| 353 | VP_READ_ADVANCE(ValueKind); |
| 354 | if (ValueKind > IPVK_Last) |
| 355 | return error(Err: instrprof_error::malformed, ErrMsg: "value kind is invalid" ); |
| 356 | ; |
| 357 | VP_READ_ADVANCE(NumValueSites); |
| 358 | if (!NumValueSites) |
| 359 | continue; |
| 360 | |
| 361 | Record.reserveSites(ValueKind: VK, NumValueSites); |
| 362 | for (uint32_t S = 0; S < NumValueSites; S++) { |
| 363 | VP_READ_ADVANCE(NumValueData); |
| 364 | |
| 365 | std::vector<InstrProfValueData> CurrentValues; |
| 366 | for (uint32_t V = 0; V < NumValueData; V++) { |
| 367 | CHECK_LINE_END(Line); |
| 368 | std::pair<StringRef, StringRef> VD = Line->rsplit(Separator: ':'); |
| 369 | uint64_t TakenCount, Value; |
| 370 | if (ValueKind == IPVK_IndirectCallTarget) { |
| 371 | if (InstrProfSymtab::isExternalSymbol(Symbol: VD.first)) { |
| 372 | Value = 0; |
| 373 | } else { |
| 374 | if (Error E = Symtab->addFuncName(FuncName: VD.first)) |
| 375 | return E; |
| 376 | Value = IndexedInstrProf::ComputeHash(K: VD.first); |
| 377 | } |
| 378 | } else if (ValueKind == IPVK_VTableTarget) { |
| 379 | if (InstrProfSymtab::isExternalSymbol(Symbol: VD.first)) |
| 380 | Value = 0; |
| 381 | else { |
| 382 | if (Error E = Symtab->addVTableName(VTableName: VD.first)) |
| 383 | return E; |
| 384 | Value = IndexedInstrProf::ComputeHash(K: VD.first); |
| 385 | } |
| 386 | } else { |
| 387 | READ_NUM(VD.first, Value); |
| 388 | } |
| 389 | READ_NUM(VD.second, TakenCount); |
| 390 | CurrentValues.push_back(x: {.Value: Value, .Count: TakenCount}); |
| 391 | Line++; |
| 392 | } |
| 393 | assert(CurrentValues.size() == NumValueData); |
| 394 | Record.addValueData(ValueKind, Site: S, VData: CurrentValues, SymTab: nullptr); |
| 395 | } |
| 396 | } |
| 397 | return success(); |
| 398 | |
| 399 | #undef CHECK_LINE_END |
| 400 | #undef READ_NUM |
| 401 | #undef VP_READ_ADVANCE |
| 402 | } |
| 403 | |
| 404 | Error TextInstrProfReader::readNextRecord(NamedInstrProfRecord &Record) { |
| 405 | // Skip empty lines and comments. |
| 406 | while (!Line.is_at_end() && (Line->empty() || Line->starts_with(Prefix: "#" ))) |
| 407 | ++Line; |
| 408 | // If we hit EOF while looking for a name, we're done. |
| 409 | if (Line.is_at_end()) { |
| 410 | return error(Err: instrprof_error::eof); |
| 411 | } |
| 412 | |
| 413 | // Read the function name. |
| 414 | Record.Name = *Line++; |
| 415 | if (Error E = Symtab->addFuncName(FuncName: Record.Name)) |
| 416 | return error(E: std::move(E)); |
| 417 | |
| 418 | // Read the function hash. |
| 419 | if (Line.is_at_end()) |
| 420 | return error(Err: instrprof_error::truncated); |
| 421 | if ((Line++)->getAsInteger(Radix: 0, Result&: Record.Hash)) |
| 422 | return error(Err: instrprof_error::malformed, |
| 423 | ErrMsg: "function hash is not a valid integer" ); |
| 424 | |
| 425 | // Read the number of counters. |
| 426 | uint64_t NumCounters; |
| 427 | if (Line.is_at_end()) |
| 428 | return error(Err: instrprof_error::truncated); |
| 429 | if ((Line++)->getAsInteger(Radix: 10, Result&: NumCounters)) |
| 430 | return error(Err: instrprof_error::malformed, |
| 431 | ErrMsg: "number of counters is not a valid integer" ); |
| 432 | if (NumCounters == 0) |
| 433 | return error(Err: instrprof_error::malformed, ErrMsg: "number of counters is zero" ); |
| 434 | |
| 435 | // Read each counter and fill our internal storage with the values. |
| 436 | Record.Clear(); |
| 437 | Record.Counts.reserve(n: NumCounters); |
| 438 | for (uint64_t I = 0; I < NumCounters; ++I) { |
| 439 | if (Line.is_at_end()) |
| 440 | return error(Err: instrprof_error::truncated); |
| 441 | uint64_t Count; |
| 442 | if ((Line++)->getAsInteger(Radix: 10, Result&: Count)) |
| 443 | return error(Err: instrprof_error::malformed, ErrMsg: "count is invalid" ); |
| 444 | Record.Counts.push_back(x: Count); |
| 445 | } |
| 446 | |
| 447 | // Bitmap byte information is indicated with special character. |
| 448 | if (Line->starts_with(Prefix: "$" )) { |
| 449 | Record.BitmapBytes.clear(); |
| 450 | // Read the number of bitmap bytes. |
| 451 | uint64_t NumBitmapBytes; |
| 452 | if ((Line++)->drop_front(N: 1).trim().getAsInteger(Radix: 0, Result&: NumBitmapBytes)) |
| 453 | return error(Err: instrprof_error::malformed, |
| 454 | ErrMsg: "number of bitmap bytes is not a valid integer" ); |
| 455 | if (NumBitmapBytes != 0) { |
| 456 | // Read each bitmap and fill our internal storage with the values. |
| 457 | Record.BitmapBytes.reserve(n: NumBitmapBytes); |
| 458 | for (uint8_t I = 0; I < NumBitmapBytes; ++I) { |
| 459 | if (Line.is_at_end()) |
| 460 | return error(Err: instrprof_error::truncated); |
| 461 | uint8_t BitmapByte; |
| 462 | if ((Line++)->getAsInteger(Radix: 0, Result&: BitmapByte)) |
| 463 | return error(Err: instrprof_error::malformed, |
| 464 | ErrMsg: "bitmap byte is not a valid integer" ); |
| 465 | Record.BitmapBytes.push_back(x: BitmapByte); |
| 466 | } |
| 467 | } |
| 468 | } |
| 469 | |
| 470 | // Check if value profile data exists and read it if so. |
| 471 | if (Error E = readValueProfileData(Record)) |
| 472 | return error(E: std::move(E)); |
| 473 | |
| 474 | return success(); |
| 475 | } |
| 476 | |
| 477 | template <class IntPtrT> |
| 478 | InstrProfKind RawInstrProfReader<IntPtrT>::getProfileKind() const { |
| 479 | return getProfileKindFromVersion(Version); |
| 480 | } |
| 481 | |
| 482 | template <class IntPtrT> |
| 483 | SmallVector<TemporalProfTraceTy> & |
| 484 | RawInstrProfReader<IntPtrT>::getTemporalProfTraces( |
| 485 | std::optional<uint64_t> Weight) { |
| 486 | if (TemporalProfTimestamps.empty()) { |
| 487 | assert(TemporalProfTraces.empty()); |
| 488 | return TemporalProfTraces; |
| 489 | } |
| 490 | // Sort functions by their timestamps to build the trace. |
| 491 | std::sort(first: TemporalProfTimestamps.begin(), last: TemporalProfTimestamps.end()); |
| 492 | TemporalProfTraceTy Trace; |
| 493 | if (Weight) |
| 494 | Trace.Weight = *Weight; |
| 495 | for (auto &[TimestampValue, NameRef] : TemporalProfTimestamps) |
| 496 | Trace.FunctionNameRefs.push_back(x: NameRef); |
| 497 | TemporalProfTraces = {std::move(Trace)}; |
| 498 | return TemporalProfTraces; |
| 499 | } |
| 500 | |
| 501 | template <class IntPtrT> |
| 502 | bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) { |
| 503 | if (DataBuffer.getBufferSize() < sizeof(uint64_t)) |
| 504 | return false; |
| 505 | uint64_t Magic = |
| 506 | *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart()); |
| 507 | return RawInstrProf::getMagic<IntPtrT>() == Magic || |
| 508 | llvm::byteswap(RawInstrProf::getMagic<IntPtrT>()) == Magic; |
| 509 | } |
| 510 | |
| 511 | template <class IntPtrT> |
| 512 | Error RawInstrProfReader<IntPtrT>::() { |
| 513 | if (!hasFormat(DataBuffer: *DataBuffer)) |
| 514 | return error(instrprof_error::bad_magic); |
| 515 | if (DataBuffer->getBufferSize() < sizeof(RawInstrProf::Header)) |
| 516 | return error(instrprof_error::bad_header); |
| 517 | auto * = reinterpret_cast<const RawInstrProf::Header *>( |
| 518 | DataBuffer->getBufferStart()); |
| 519 | ShouldSwapBytes = Header->Magic != RawInstrProf::getMagic<IntPtrT>(); |
| 520 | return readHeader(*Header); |
| 521 | } |
| 522 | |
| 523 | template <class IntPtrT> |
| 524 | Error RawInstrProfReader<IntPtrT>::(const char *CurrentPos) { |
| 525 | const char *End = DataBuffer->getBufferEnd(); |
| 526 | // Skip zero padding between profiles. |
| 527 | while (CurrentPos != End && *CurrentPos == 0) |
| 528 | ++CurrentPos; |
| 529 | // If there's nothing left, we're done. |
| 530 | if (CurrentPos == End) |
| 531 | return make_error<InstrProfError>(Args: instrprof_error::eof); |
| 532 | // If there isn't enough space for another header, this is probably just |
| 533 | // garbage at the end of the file. |
| 534 | if (CurrentPos + sizeof(RawInstrProf::Header) > End) |
| 535 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
| 536 | Args: "not enough space for another header" ); |
| 537 | // The writer ensures each profile is padded to start at an aligned address. |
| 538 | if (reinterpret_cast<size_t>(CurrentPos) % alignof(uint64_t)) |
| 539 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
| 540 | Args: "insufficient padding" ); |
| 541 | // The magic should have the same byte order as in the previous header. |
| 542 | uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos); |
| 543 | if (Magic != swap(RawInstrProf::getMagic<IntPtrT>())) |
| 544 | return make_error<InstrProfError>(Args: instrprof_error::bad_magic); |
| 545 | |
| 546 | // There's another profile to read, so we need to process the header. |
| 547 | auto * = reinterpret_cast<const RawInstrProf::Header *>(CurrentPos); |
| 548 | return readHeader(*Header); |
| 549 | } |
| 550 | |
| 551 | template <class IntPtrT> |
| 552 | Error RawInstrProfReader<IntPtrT>::createSymtab(InstrProfSymtab &Symtab) { |
| 553 | if (Error E = Symtab.create(FuncNameStrings: StringRef(NamesStart, NamesEnd - NamesStart), |
| 554 | VTableNameStrings: StringRef(VNamesStart, VNamesEnd - VNamesStart))) |
| 555 | return error(std::move(E)); |
| 556 | for (const RawInstrProf::ProfileData<IntPtrT> *I = Data; I != DataEnd; ++I) { |
| 557 | const IntPtrT FPtr = swap(I->FunctionPointer); |
| 558 | if (!FPtr) |
| 559 | continue; |
| 560 | Symtab.mapAddress(Addr: FPtr, MD5Val: swap(I->NameRef)); |
| 561 | } |
| 562 | |
| 563 | if (VTableBegin != nullptr && VTableEnd != nullptr) { |
| 564 | for (const RawInstrProf::VTableProfileData<IntPtrT> *I = VTableBegin; |
| 565 | I != VTableEnd; ++I) { |
| 566 | const IntPtrT VPtr = swap(I->VTablePointer); |
| 567 | if (!VPtr) |
| 568 | continue; |
| 569 | // Map both begin and end address to the name hash, since the instrumented |
| 570 | // address could be somewhere in the middle. |
| 571 | // VPtr is of type uint32_t or uint64_t so 'VPtr + I->VTableSize' marks |
| 572 | // the end of vtable address. |
| 573 | Symtab.mapVTableAddress(StartAddr: VPtr, EndAddr: VPtr + swap(I->VTableSize), |
| 574 | MD5Val: swap(I->VTableNameHash)); |
| 575 | } |
| 576 | } |
| 577 | return success(); |
| 578 | } |
| 579 | |
| 580 | template <class IntPtrT> |
| 581 | Error RawInstrProfReader<IntPtrT>::( |
| 582 | const RawInstrProf::Header &) { |
| 583 | Version = swap(Header.Version); |
| 584 | if (GET_VERSION(Version) != RawInstrProf::Version) |
| 585 | return error(instrprof_error::raw_profile_version_mismatch, |
| 586 | ("Profile uses raw profile format version = " + |
| 587 | Twine(GET_VERSION(Version)) + |
| 588 | "; expected version = " + Twine(RawInstrProf::Version) + |
| 589 | "\nPLEASE update this tool to version in the raw profile, or " |
| 590 | "regenerate raw profile with expected version." ) |
| 591 | .str()); |
| 592 | |
| 593 | uint64_t BinaryIdSize = swap(Header.BinaryIdsSize); |
| 594 | // Binary id start just after the header if exists. |
| 595 | const uint8_t *BinaryIdStart = |
| 596 | reinterpret_cast<const uint8_t *>(&Header) + sizeof(RawInstrProf::Header); |
| 597 | const uint8_t *BinaryIdEnd = BinaryIdStart + BinaryIdSize; |
| 598 | const uint8_t *BufferEnd = (const uint8_t *)DataBuffer->getBufferEnd(); |
| 599 | if (BinaryIdSize % sizeof(uint64_t) || BinaryIdEnd > BufferEnd) |
| 600 | return error(instrprof_error::bad_header); |
| 601 | ArrayRef<uint8_t> BinaryIdsBuffer(BinaryIdStart, BinaryIdSize); |
| 602 | if (!BinaryIdsBuffer.empty()) { |
| 603 | if (Error Err = readBinaryIdsInternal(*DataBuffer, BinaryIdsBuffer, |
| 604 | BinaryIds, getDataEndianness())) |
| 605 | return Err; |
| 606 | } |
| 607 | |
| 608 | CountersDelta = swap(Header.CountersDelta); |
| 609 | BitmapDelta = swap(Header.BitmapDelta); |
| 610 | NamesDelta = swap(Header.NamesDelta); |
| 611 | auto NumData = swap(Header.NumData); |
| 612 | auto PaddingBytesBeforeCounters = swap(Header.PaddingBytesBeforeCounters); |
| 613 | auto = swap(Header.NumCounters) * getCounterTypeSize(); |
| 614 | auto PaddingBytesAfterCounters = swap(Header.PaddingBytesAfterCounters); |
| 615 | auto NumBitmapBytes = swap(Header.NumBitmapBytes); |
| 616 | auto PaddingBytesAfterBitmapBytes = swap(Header.PaddingBytesAfterBitmapBytes); |
| 617 | auto NamesSize = swap(Header.NamesSize); |
| 618 | auto VTableNameSize = swap(Header.VNamesSize); |
| 619 | auto NumVTables = swap(Header.NumVTables); |
| 620 | ValueKindLast = swap(Header.ValueKindLast); |
| 621 | |
| 622 | auto DataSize = NumData * sizeof(RawInstrProf::ProfileData<IntPtrT>); |
| 623 | auto PaddingBytesAfterNames = getNumPaddingBytes(SizeInBytes: NamesSize); |
| 624 | auto PaddingBytesAfterVTableNames = getNumPaddingBytes(SizeInBytes: VTableNameSize); |
| 625 | |
| 626 | auto VTableSectionSize = |
| 627 | NumVTables * sizeof(RawInstrProf::VTableProfileData<IntPtrT>); |
| 628 | auto PaddingBytesAfterVTableProfData = getNumPaddingBytes(SizeInBytes: VTableSectionSize); |
| 629 | |
| 630 | // Profile data starts after profile header and binary ids if exist. |
| 631 | ptrdiff_t DataOffset = sizeof(RawInstrProf::Header) + BinaryIdSize; |
| 632 | ptrdiff_t CountersOffset = DataOffset + DataSize + PaddingBytesBeforeCounters; |
| 633 | ptrdiff_t BitmapOffset = |
| 634 | CountersOffset + CountersSize + PaddingBytesAfterCounters; |
| 635 | ptrdiff_t NamesOffset = |
| 636 | BitmapOffset + NumBitmapBytes + PaddingBytesAfterBitmapBytes; |
| 637 | ptrdiff_t VTableProfDataOffset = |
| 638 | NamesOffset + NamesSize + PaddingBytesAfterNames; |
| 639 | ptrdiff_t VTableNameOffset = VTableProfDataOffset + VTableSectionSize + |
| 640 | PaddingBytesAfterVTableProfData; |
| 641 | ptrdiff_t ValueDataOffset = |
| 642 | VTableNameOffset + VTableNameSize + PaddingBytesAfterVTableNames; |
| 643 | |
| 644 | auto *Start = reinterpret_cast<const char *>(&Header); |
| 645 | if (Start + ValueDataOffset > DataBuffer->getBufferEnd()) |
| 646 | return error(instrprof_error::bad_header); |
| 647 | |
| 648 | if (BIDFetcher) { |
| 649 | std::vector<object::BuildID> BinaryIDs; |
| 650 | if (Error E = readBinaryIds(BinaryIds&: BinaryIDs)) |
| 651 | return E; |
| 652 | if (auto E = InstrProfCorrelator::get(Filename: "" , FileKind: BIDFetcherCorrelatorKind, |
| 653 | BIDFetcher, BIs: BinaryIDs) |
| 654 | .moveInto(Value&: BIDFetcherCorrelator)) { |
| 655 | return E; |
| 656 | } |
| 657 | if (auto Err = BIDFetcherCorrelator->correlateProfileData(MaxWarnings: 0)) |
| 658 | return Err; |
| 659 | } |
| 660 | |
| 661 | if (Correlator) { |
| 662 | // These sizes in the raw file are zero because we constructed them in the |
| 663 | // Correlator. |
| 664 | if (!(DataSize == 0 && NamesSize == 0 && CountersDelta == 0 && |
| 665 | NamesDelta == 0)) |
| 666 | return error(instrprof_error::unexpected_correlation_info); |
| 667 | Data = Correlator->getDataPointer(); |
| 668 | DataEnd = Data + Correlator->getDataSize(); |
| 669 | NamesStart = Correlator->getNamesPointer(); |
| 670 | NamesEnd = NamesStart + Correlator->getNamesSize(); |
| 671 | } else if (BIDFetcherCorrelator) { |
| 672 | InstrProfCorrelatorImpl<IntPtrT> *BIDFetcherCorrelatorImpl = |
| 673 | dyn_cast_or_null<InstrProfCorrelatorImpl<IntPtrT>>( |
| 674 | BIDFetcherCorrelator.get()); |
| 675 | Data = BIDFetcherCorrelatorImpl->getDataPointer(); |
| 676 | DataEnd = Data + BIDFetcherCorrelatorImpl->getDataSize(); |
| 677 | NamesStart = BIDFetcherCorrelatorImpl->getNamesPointer(); |
| 678 | NamesEnd = NamesStart + BIDFetcherCorrelatorImpl->getNamesSize(); |
| 679 | } else { |
| 680 | Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>( |
| 681 | Start + DataOffset); |
| 682 | DataEnd = Data + NumData; |
| 683 | VTableBegin = |
| 684 | reinterpret_cast<const RawInstrProf::VTableProfileData<IntPtrT> *>( |
| 685 | Start + VTableProfDataOffset); |
| 686 | VTableEnd = VTableBegin + NumVTables; |
| 687 | NamesStart = Start + NamesOffset; |
| 688 | NamesEnd = NamesStart + NamesSize; |
| 689 | VNamesStart = Start + VTableNameOffset; |
| 690 | VNamesEnd = VNamesStart + VTableNameSize; |
| 691 | } |
| 692 | |
| 693 | CountersStart = Start + CountersOffset; |
| 694 | CountersEnd = CountersStart + CountersSize; |
| 695 | BitmapStart = Start + BitmapOffset; |
| 696 | BitmapEnd = BitmapStart + NumBitmapBytes; |
| 697 | ValueDataStart = reinterpret_cast<const uint8_t *>(Start + ValueDataOffset); |
| 698 | |
| 699 | std::unique_ptr<InstrProfSymtab> NewSymtab = std::make_unique<InstrProfSymtab>(); |
| 700 | if (Error E = createSymtab(Symtab&: *NewSymtab)) |
| 701 | return E; |
| 702 | |
| 703 | Symtab = std::move(NewSymtab); |
| 704 | return success(); |
| 705 | } |
| 706 | |
| 707 | template <class IntPtrT> |
| 708 | Error RawInstrProfReader<IntPtrT>::readName(NamedInstrProfRecord &Record) { |
| 709 | Record.Name = getName(NameRef: Data->NameRef); |
| 710 | return success(); |
| 711 | } |
| 712 | |
| 713 | template <class IntPtrT> |
| 714 | Error RawInstrProfReader<IntPtrT>::readFuncHash(NamedInstrProfRecord &Record) { |
| 715 | Record.Hash = swap(Data->FuncHash); |
| 716 | return success(); |
| 717 | } |
| 718 | |
| 719 | template <class IntPtrT> |
| 720 | Error RawInstrProfReader<IntPtrT>::readRawCounts( |
| 721 | InstrProfRecord &Record) { |
| 722 | uint32_t NumCounters = swap(Data->NumCounters); |
| 723 | if (NumCounters == 0) |
| 724 | return error(instrprof_error::malformed, "number of counters is zero" ); |
| 725 | |
| 726 | ptrdiff_t CounterBaseOffset = swap(Data->CounterPtr) - CountersDelta; |
| 727 | if (CounterBaseOffset < 0) |
| 728 | return error( |
| 729 | instrprof_error::malformed, |
| 730 | ("counter offset " + Twine(CounterBaseOffset) + " is negative" ).str()); |
| 731 | |
| 732 | if (CounterBaseOffset >= CountersEnd - CountersStart) |
| 733 | return error(instrprof_error::malformed, |
| 734 | ("counter offset " + Twine(CounterBaseOffset) + |
| 735 | " is greater than the maximum counter offset " + |
| 736 | Twine(CountersEnd - CountersStart - 1)) |
| 737 | .str()); |
| 738 | |
| 739 | uint64_t MaxNumCounters = |
| 740 | (CountersEnd - (CountersStart + CounterBaseOffset)) / |
| 741 | getCounterTypeSize(); |
| 742 | if (NumCounters > MaxNumCounters) |
| 743 | return error(instrprof_error::malformed, |
| 744 | ("number of counters " + Twine(NumCounters) + |
| 745 | " is greater than the maximum number of counters " + |
| 746 | Twine(MaxNumCounters)) |
| 747 | .str()); |
| 748 | |
| 749 | Record.Counts.clear(); |
| 750 | Record.Counts.reserve(n: NumCounters); |
| 751 | for (uint32_t I = 0; I < NumCounters; I++) { |
| 752 | const char *Ptr = |
| 753 | CountersStart + CounterBaseOffset + I * getCounterTypeSize(); |
| 754 | if (I == 0 && hasTemporalProfile()) { |
| 755 | uint64_t TimestampValue = swap(*reinterpret_cast<const uint64_t *>(Ptr)); |
| 756 | if (TimestampValue != 0 && |
| 757 | TimestampValue != std::numeric_limits<uint64_t>::max()) { |
| 758 | TemporalProfTimestamps.emplace_back(TimestampValue, |
| 759 | swap(Data->NameRef)); |
| 760 | TemporalProfTraceStreamSize = 1; |
| 761 | } |
| 762 | if (hasSingleByteCoverage()) { |
| 763 | // In coverage mode, getCounterTypeSize() returns 1 byte but our |
| 764 | // timestamp field has size uint64_t. Increment I so that the next |
| 765 | // iteration of this for loop points to the byte after the timestamp |
| 766 | // field, i.e., I += 8. |
| 767 | I += 7; |
| 768 | } |
| 769 | continue; |
| 770 | } |
| 771 | if (hasSingleByteCoverage()) { |
| 772 | // A value of zero signifies the block is covered. |
| 773 | Record.Counts.push_back(x: *Ptr == 0 ? 1 : 0); |
| 774 | } else { |
| 775 | uint64_t CounterValue = swap(*reinterpret_cast<const uint64_t *>(Ptr)); |
| 776 | if (CounterValue > MaxCounterValue && Warn) |
| 777 | Warn(make_error<InstrProfError>( |
| 778 | Args: instrprof_error::counter_value_too_large, Args: Twine(CounterValue))); |
| 779 | |
| 780 | Record.Counts.push_back(x: CounterValue); |
| 781 | } |
| 782 | } |
| 783 | |
| 784 | return success(); |
| 785 | } |
| 786 | |
| 787 | template <class IntPtrT> |
| 788 | Error RawInstrProfReader<IntPtrT>::readRawBitmapBytes(InstrProfRecord &Record) { |
| 789 | uint32_t NumBitmapBytes = swap(Data->NumBitmapBytes); |
| 790 | |
| 791 | Record.BitmapBytes.clear(); |
| 792 | Record.BitmapBytes.reserve(n: NumBitmapBytes); |
| 793 | |
| 794 | // It's possible MCDC is either not enabled or only used for some functions |
| 795 | // and not others. So if we record 0 bytes, just move on. |
| 796 | if (NumBitmapBytes == 0) |
| 797 | return success(); |
| 798 | |
| 799 | // BitmapDelta decreases as we advance to the next data record. |
| 800 | ptrdiff_t BitmapOffset = swap(Data->BitmapPtr) - BitmapDelta; |
| 801 | if (BitmapOffset < 0) |
| 802 | return error( |
| 803 | instrprof_error::malformed, |
| 804 | ("bitmap offset " + Twine(BitmapOffset) + " is negative" ).str()); |
| 805 | |
| 806 | if (BitmapOffset >= BitmapEnd - BitmapStart) |
| 807 | return error(instrprof_error::malformed, |
| 808 | ("bitmap offset " + Twine(BitmapOffset) + |
| 809 | " is greater than the maximum bitmap offset " + |
| 810 | Twine(BitmapEnd - BitmapStart - 1)) |
| 811 | .str()); |
| 812 | |
| 813 | uint64_t MaxNumBitmapBytes = |
| 814 | (BitmapEnd - (BitmapStart + BitmapOffset)) / sizeof(uint8_t); |
| 815 | if (NumBitmapBytes > MaxNumBitmapBytes) |
| 816 | return error(instrprof_error::malformed, |
| 817 | ("number of bitmap bytes " + Twine(NumBitmapBytes) + |
| 818 | " is greater than the maximum number of bitmap bytes " + |
| 819 | Twine(MaxNumBitmapBytes)) |
| 820 | .str()); |
| 821 | |
| 822 | for (uint32_t I = 0; I < NumBitmapBytes; I++) { |
| 823 | const char *Ptr = BitmapStart + BitmapOffset + I; |
| 824 | Record.BitmapBytes.push_back(swap(*Ptr)); |
| 825 | } |
| 826 | |
| 827 | return success(); |
| 828 | } |
| 829 | |
| 830 | template <class IntPtrT> |
| 831 | Error RawInstrProfReader<IntPtrT>::readValueProfilingData( |
| 832 | InstrProfRecord &Record) { |
| 833 | Record.clearValueData(); |
| 834 | CurValueDataSize = 0; |
| 835 | // Need to match the logic in value profile dumper code in compiler-rt: |
| 836 | uint32_t NumValueKinds = 0; |
| 837 | for (uint32_t I = 0; I < IPVK_Last + 1; I++) |
| 838 | NumValueKinds += (Data->NumValueSites[I] != 0); |
| 839 | |
| 840 | if (!NumValueKinds) |
| 841 | return success(); |
| 842 | |
| 843 | Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr = |
| 844 | ValueProfData::getValueProfData( |
| 845 | SrcBuffer: ValueDataStart, SrcBufferEnd: (const unsigned char *)DataBuffer->getBufferEnd(), |
| 846 | SrcDataEndianness: getDataEndianness()); |
| 847 | |
| 848 | if (Error E = VDataPtrOrErr.takeError()) |
| 849 | return E; |
| 850 | |
| 851 | // Note that besides deserialization, this also performs the conversion for |
| 852 | // indirect call targets. The function pointers from the raw profile are |
| 853 | // remapped into function name hashes. |
| 854 | VDataPtrOrErr.get()->deserializeTo(Record, SymTab: Symtab.get()); |
| 855 | CurValueDataSize = VDataPtrOrErr.get()->getSize(); |
| 856 | return success(); |
| 857 | } |
| 858 | |
| 859 | template <class IntPtrT> |
| 860 | Error RawInstrProfReader<IntPtrT>::readNextRecord(NamedInstrProfRecord &Record) { |
| 861 | // Keep reading profiles that consist of only headers and no profile data and |
| 862 | // counters. |
| 863 | while (atEnd()) |
| 864 | // At this point, ValueDataStart field points to the next header. |
| 865 | if (Error E = readNextHeader(CurrentPos: getNextHeaderPos())) |
| 866 | return error(std::move(E)); |
| 867 | |
| 868 | // Read name and set it in Record. |
| 869 | if (Error E = readName(Record)) |
| 870 | return error(std::move(E)); |
| 871 | |
| 872 | // Read FuncHash and set it in Record. |
| 873 | if (Error E = readFuncHash(Record)) |
| 874 | return error(std::move(E)); |
| 875 | |
| 876 | // Read raw counts and set Record. |
| 877 | if (Error E = readRawCounts(Record)) |
| 878 | return error(std::move(E)); |
| 879 | |
| 880 | // Read raw bitmap bytes and set Record. |
| 881 | if (Error E = readRawBitmapBytes(Record)) |
| 882 | return error(std::move(E)); |
| 883 | |
| 884 | // Read value data and set Record. |
| 885 | if (Error E = readValueProfilingData(Record)) |
| 886 | return error(std::move(E)); |
| 887 | |
| 888 | // Iterate. |
| 889 | advanceData(); |
| 890 | return success(); |
| 891 | } |
| 892 | |
| 893 | template <class IntPtrT> |
| 894 | Error RawInstrProfReader<IntPtrT>::readBinaryIds( |
| 895 | std::vector<llvm::object::BuildID> &BinaryIds) { |
| 896 | BinaryIds.insert(BinaryIds.begin(), this->BinaryIds.begin(), |
| 897 | this->BinaryIds.end()); |
| 898 | return Error::success(); |
| 899 | } |
| 900 | |
| 901 | template <class IntPtrT> |
| 902 | Error RawInstrProfReader<IntPtrT>::printBinaryIds(raw_ostream &OS) { |
| 903 | if (!BinaryIds.empty()) |
| 904 | printBinaryIdsInternal(OS, BinaryIds); |
| 905 | return Error::success(); |
| 906 | } |
| 907 | |
| 908 | namespace llvm { |
| 909 | |
| 910 | template class RawInstrProfReader<uint32_t>; |
| 911 | template class RawInstrProfReader<uint64_t>; |
| 912 | |
| 913 | } // end namespace llvm |
| 914 | |
| 915 | InstrProfLookupTrait::hash_value_type |
| 916 | InstrProfLookupTrait::ComputeHash(StringRef K) { |
| 917 | return IndexedInstrProf::ComputeHash(Type: HashType, K); |
| 918 | } |
| 919 | |
| 920 | using data_type = InstrProfLookupTrait::data_type; |
| 921 | using offset_type = InstrProfLookupTrait::offset_type; |
| 922 | |
| 923 | bool InstrProfLookupTrait::readValueProfilingData( |
| 924 | const unsigned char *&D, const unsigned char *const End) { |
| 925 | Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr = |
| 926 | ValueProfData::getValueProfData(SrcBuffer: D, SrcBufferEnd: End, SrcDataEndianness: ValueProfDataEndianness); |
| 927 | |
| 928 | if (VDataPtrOrErr.takeError()) |
| 929 | return false; |
| 930 | |
| 931 | VDataPtrOrErr.get()->deserializeTo(Record&: DataBuffer.back(), SymTab: nullptr); |
| 932 | D += VDataPtrOrErr.get()->TotalSize; |
| 933 | |
| 934 | return true; |
| 935 | } |
| 936 | |
| 937 | data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D, |
| 938 | offset_type N) { |
| 939 | using namespace support; |
| 940 | |
| 941 | // Check if the data is corrupt. If so, don't try to read it. |
| 942 | if (N % sizeof(uint64_t)) |
| 943 | return data_type(); |
| 944 | |
| 945 | DataBuffer.clear(); |
| 946 | std::vector<uint64_t> CounterBuffer; |
| 947 | std::vector<uint8_t> BitmapByteBuffer; |
| 948 | |
| 949 | const unsigned char *End = D + N; |
| 950 | while (D < End) { |
| 951 | // Read hash. |
| 952 | if (D + sizeof(uint64_t) >= End) |
| 953 | return data_type(); |
| 954 | uint64_t Hash = endian::readNext<uint64_t, llvm::endianness::little>(memory&: D); |
| 955 | |
| 956 | // Initialize number of counters for GET_VERSION(FormatVersion) == 1. |
| 957 | uint64_t CountsSize = N / sizeof(uint64_t) - 1; |
| 958 | // If format version is different then read the number of counters. |
| 959 | if (GET_VERSION(FormatVersion) != IndexedInstrProf::ProfVersion::Version1) { |
| 960 | if (D + sizeof(uint64_t) > End) |
| 961 | return data_type(); |
| 962 | CountsSize = endian::readNext<uint64_t, llvm::endianness::little>(memory&: D); |
| 963 | } |
| 964 | // Read counter values. |
| 965 | if (D + CountsSize * sizeof(uint64_t) > End) |
| 966 | return data_type(); |
| 967 | |
| 968 | CounterBuffer.clear(); |
| 969 | CounterBuffer.reserve(n: CountsSize); |
| 970 | for (uint64_t J = 0; J < CountsSize; ++J) |
| 971 | CounterBuffer.push_back( |
| 972 | x: endian::readNext<uint64_t, llvm::endianness::little>(memory&: D)); |
| 973 | |
| 974 | // Read bitmap bytes for GET_VERSION(FormatVersion) > 10. |
| 975 | if (GET_VERSION(FormatVersion) > IndexedInstrProf::ProfVersion::Version10) { |
| 976 | uint64_t BitmapBytes = 0; |
| 977 | if (D + sizeof(uint64_t) > End) |
| 978 | return data_type(); |
| 979 | BitmapBytes = endian::readNext<uint64_t, llvm::endianness::little>(memory&: D); |
| 980 | // Read bitmap byte values. |
| 981 | if (D + BitmapBytes * sizeof(uint8_t) > End) |
| 982 | return data_type(); |
| 983 | BitmapByteBuffer.clear(); |
| 984 | BitmapByteBuffer.reserve(n: BitmapBytes); |
| 985 | for (uint64_t J = 0; J < BitmapBytes; ++J) |
| 986 | BitmapByteBuffer.push_back(x: static_cast<uint8_t>( |
| 987 | endian::readNext<uint64_t, llvm::endianness::little>(memory&: D))); |
| 988 | } |
| 989 | |
| 990 | DataBuffer.emplace_back(args&: K, args&: Hash, args: std::move(CounterBuffer), |
| 991 | args: std::move(BitmapByteBuffer)); |
| 992 | |
| 993 | // Read value profiling data. |
| 994 | if (GET_VERSION(FormatVersion) > IndexedInstrProf::ProfVersion::Version2 && |
| 995 | !readValueProfilingData(D, End)) { |
| 996 | DataBuffer.clear(); |
| 997 | return data_type(); |
| 998 | } |
| 999 | } |
| 1000 | return DataBuffer; |
| 1001 | } |
| 1002 | |
| 1003 | template <typename HashTableImpl> |
| 1004 | Error InstrProfReaderIndex<HashTableImpl>::getRecords( |
| 1005 | StringRef FuncName, ArrayRef<NamedInstrProfRecord> &Data) { |
| 1006 | auto Iter = HashTable->find(FuncName); |
| 1007 | if (Iter == HashTable->end()) |
| 1008 | return make_error<InstrProfError>(Args: instrprof_error::unknown_function); |
| 1009 | |
| 1010 | Data = (*Iter); |
| 1011 | if (Data.empty()) |
| 1012 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
| 1013 | Args: "profile data is empty" ); |
| 1014 | |
| 1015 | return Error::success(); |
| 1016 | } |
| 1017 | |
| 1018 | template <typename HashTableImpl> |
| 1019 | Error InstrProfReaderIndex<HashTableImpl>::getRecords( |
| 1020 | ArrayRef<NamedInstrProfRecord> &Data) { |
| 1021 | if (atEnd()) |
| 1022 | return make_error<InstrProfError>(Args: instrprof_error::eof); |
| 1023 | |
| 1024 | Data = *RecordIterator; |
| 1025 | |
| 1026 | if (Data.empty()) |
| 1027 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
| 1028 | Args: "profile data is empty" ); |
| 1029 | |
| 1030 | return Error::success(); |
| 1031 | } |
| 1032 | |
| 1033 | template <typename HashTableImpl> |
| 1034 | InstrProfReaderIndex<HashTableImpl>::InstrProfReaderIndex( |
| 1035 | const unsigned char *Buckets, const unsigned char *const Payload, |
| 1036 | const unsigned char *const Base, IndexedInstrProf::HashT HashType, |
| 1037 | uint64_t Version) { |
| 1038 | FormatVersion = Version; |
| 1039 | HashTable.reset(HashTableImpl::Create( |
| 1040 | Buckets, Payload, Base, |
| 1041 | typename HashTableImpl::InfoType(HashType, Version))); |
| 1042 | RecordIterator = HashTable->data_begin(); |
| 1043 | } |
| 1044 | |
| 1045 | template <typename HashTableImpl> |
| 1046 | InstrProfKind InstrProfReaderIndex<HashTableImpl>::getProfileKind() const { |
| 1047 | return getProfileKindFromVersion(Version: FormatVersion); |
| 1048 | } |
| 1049 | |
| 1050 | namespace { |
| 1051 | /// A remapper that does not apply any remappings. |
| 1052 | class InstrProfReaderNullRemapper : public InstrProfReaderRemapper { |
| 1053 | InstrProfReaderIndexBase &Underlying; |
| 1054 | |
| 1055 | public: |
| 1056 | InstrProfReaderNullRemapper(InstrProfReaderIndexBase &Underlying) |
| 1057 | : Underlying(Underlying) {} |
| 1058 | |
| 1059 | Error getRecords(StringRef FuncName, |
| 1060 | ArrayRef<NamedInstrProfRecord> &Data) override { |
| 1061 | return Underlying.getRecords(FuncName, Data); |
| 1062 | } |
| 1063 | }; |
| 1064 | } // namespace |
| 1065 | |
| 1066 | /// A remapper that applies remappings based on a symbol remapping file. |
| 1067 | template <typename HashTableImpl> |
| 1068 | class llvm::InstrProfReaderItaniumRemapper |
| 1069 | : public InstrProfReaderRemapper { |
| 1070 | public: |
| 1071 | InstrProfReaderItaniumRemapper( |
| 1072 | std::unique_ptr<MemoryBuffer> RemapBuffer, |
| 1073 | InstrProfReaderIndex<HashTableImpl> &Underlying) |
| 1074 | : RemapBuffer(std::move(RemapBuffer)), Underlying(Underlying) { |
| 1075 | } |
| 1076 | |
| 1077 | /// Extract the original function name from a PGO function name. |
| 1078 | static StringRef (StringRef Name) { |
| 1079 | // We can have multiple pieces separated by kGlobalIdentifierDelimiter ( |
| 1080 | // semicolon now and colon in older profiles); there can be pieces both |
| 1081 | // before and after the mangled name. Find the first part that starts with |
| 1082 | // '_Z'; we'll assume that's the mangled name we want. |
| 1083 | std::pair<StringRef, StringRef> Parts = {StringRef(), Name}; |
| 1084 | while (true) { |
| 1085 | Parts = Parts.second.split(Separator: GlobalIdentifierDelimiter); |
| 1086 | if (Parts.first.starts_with(Prefix: "_Z" )) |
| 1087 | return Parts.first; |
| 1088 | if (Parts.second.empty()) |
| 1089 | return Name; |
| 1090 | } |
| 1091 | } |
| 1092 | |
| 1093 | /// Given a mangled name extracted from a PGO function name, and a new |
| 1094 | /// form for that mangled name, reconstitute the name. |
| 1095 | static void reconstituteName(StringRef OrigName, StringRef , |
| 1096 | StringRef Replacement, |
| 1097 | SmallVectorImpl<char> &Out) { |
| 1098 | Out.reserve(N: OrigName.size() + Replacement.size() - ExtractedName.size()); |
| 1099 | Out.insert(I: Out.end(), From: OrigName.begin(), To: ExtractedName.begin()); |
| 1100 | llvm::append_range(C&: Out, R&: Replacement); |
| 1101 | Out.insert(I: Out.end(), From: ExtractedName.end(), To: OrigName.end()); |
| 1102 | } |
| 1103 | |
| 1104 | Error populateRemappings() override { |
| 1105 | if (Error E = Remappings.read(B&: *RemapBuffer)) |
| 1106 | return E; |
| 1107 | for (StringRef Name : Underlying.HashTable->keys()) { |
| 1108 | StringRef RealName = extractName(Name); |
| 1109 | if (auto Key = Remappings.insert(FunctionName: RealName)) { |
| 1110 | // FIXME: We could theoretically map the same equivalence class to |
| 1111 | // multiple names in the profile data. If that happens, we should |
| 1112 | // return NamedInstrProfRecords from all of them. |
| 1113 | MappedNames.insert(KV: {Key, RealName}); |
| 1114 | } |
| 1115 | } |
| 1116 | return Error::success(); |
| 1117 | } |
| 1118 | |
| 1119 | Error getRecords(StringRef FuncName, |
| 1120 | ArrayRef<NamedInstrProfRecord> &Data) override { |
| 1121 | StringRef RealName = extractName(Name: FuncName); |
| 1122 | if (auto Key = Remappings.lookup(FunctionName: RealName)) { |
| 1123 | StringRef Remapped = MappedNames.lookup(Val: Key); |
| 1124 | if (!Remapped.empty()) { |
| 1125 | if (RealName.begin() == FuncName.begin() && |
| 1126 | RealName.end() == FuncName.end()) |
| 1127 | FuncName = Remapped; |
| 1128 | else { |
| 1129 | // Try rebuilding the name from the given remapping. |
| 1130 | SmallString<256> Reconstituted; |
| 1131 | reconstituteName(OrigName: FuncName, ExtractedName: RealName, Replacement: Remapped, Out&: Reconstituted); |
| 1132 | Error E = Underlying.getRecords(Reconstituted, Data); |
| 1133 | if (!E) |
| 1134 | return E; |
| 1135 | |
| 1136 | // If we failed because the name doesn't exist, fall back to asking |
| 1137 | // about the original name. |
| 1138 | if (Error Unhandled = handleErrors( |
| 1139 | std::move(E), [](std::unique_ptr<InstrProfError> Err) { |
| 1140 | return Err->get() == instrprof_error::unknown_function |
| 1141 | ? Error::success() |
| 1142 | : Error(std::move(Err)); |
| 1143 | })) |
| 1144 | return Unhandled; |
| 1145 | } |
| 1146 | } |
| 1147 | } |
| 1148 | return Underlying.getRecords(FuncName, Data); |
| 1149 | } |
| 1150 | |
| 1151 | private: |
| 1152 | /// The memory buffer containing the remapping configuration. Remappings |
| 1153 | /// holds pointers into this buffer. |
| 1154 | std::unique_ptr<MemoryBuffer> RemapBuffer; |
| 1155 | |
| 1156 | /// The mangling remapper. |
| 1157 | SymbolRemappingReader Remappings; |
| 1158 | |
| 1159 | /// Mapping from mangled name keys to the name used for the key in the |
| 1160 | /// profile data. |
| 1161 | /// FIXME: Can we store a location within the on-disk hash table instead of |
| 1162 | /// redoing lookup? |
| 1163 | DenseMap<SymbolRemappingReader::Key, StringRef> MappedNames; |
| 1164 | |
| 1165 | /// The real profile data reader. |
| 1166 | InstrProfReaderIndex<HashTableImpl> &Underlying; |
| 1167 | }; |
| 1168 | |
| 1169 | bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) { |
| 1170 | using namespace support; |
| 1171 | |
| 1172 | if (DataBuffer.getBufferSize() < 8) |
| 1173 | return false; |
| 1174 | uint64_t Magic = endian::read<uint64_t, llvm::endianness::little, aligned>( |
| 1175 | memory: DataBuffer.getBufferStart()); |
| 1176 | // Verify that it's magical. |
| 1177 | return Magic == IndexedInstrProf::Magic; |
| 1178 | } |
| 1179 | |
| 1180 | const unsigned char * |
| 1181 | IndexedInstrProfReader::readSummary(IndexedInstrProf::ProfVersion Version, |
| 1182 | const unsigned char *Cur, bool UseCS) { |
| 1183 | using namespace IndexedInstrProf; |
| 1184 | using namespace support; |
| 1185 | |
| 1186 | if (Version >= IndexedInstrProf::Version4) { |
| 1187 | const IndexedInstrProf::Summary *SummaryInLE = |
| 1188 | reinterpret_cast<const IndexedInstrProf::Summary *>(Cur); |
| 1189 | uint64_t NFields = endian::byte_swap<uint64_t, llvm::endianness::little>( |
| 1190 | value: SummaryInLE->NumSummaryFields); |
| 1191 | uint64_t NEntries = endian::byte_swap<uint64_t, llvm::endianness::little>( |
| 1192 | value: SummaryInLE->NumCutoffEntries); |
| 1193 | uint32_t SummarySize = |
| 1194 | IndexedInstrProf::Summary::getSize(NumSumFields: NFields, NumCutoffEntries: NEntries); |
| 1195 | std::unique_ptr<IndexedInstrProf::Summary> SummaryData = |
| 1196 | IndexedInstrProf::allocSummary(TotalSize: SummarySize); |
| 1197 | |
| 1198 | const uint64_t *Src = reinterpret_cast<const uint64_t *>(SummaryInLE); |
| 1199 | uint64_t *Dst = reinterpret_cast<uint64_t *>(SummaryData.get()); |
| 1200 | for (unsigned I = 0; I < SummarySize / sizeof(uint64_t); I++) |
| 1201 | Dst[I] = endian::byte_swap<uint64_t, llvm::endianness::little>(value: Src[I]); |
| 1202 | |
| 1203 | SummaryEntryVector DetailedSummary; |
| 1204 | for (unsigned I = 0; I < SummaryData->NumCutoffEntries; I++) { |
| 1205 | const IndexedInstrProf::Summary::Entry &Ent = SummaryData->getEntry(I); |
| 1206 | DetailedSummary.emplace_back(args: (uint32_t)Ent.Cutoff, args: Ent.MinBlockCount, |
| 1207 | args: Ent.NumBlocks); |
| 1208 | } |
| 1209 | std::unique_ptr<llvm::ProfileSummary> &Summary = |
| 1210 | UseCS ? this->CS_Summary : this->Summary; |
| 1211 | |
| 1212 | // initialize InstrProfSummary using the SummaryData from disk. |
| 1213 | Summary = std::make_unique<ProfileSummary>( |
| 1214 | args: UseCS ? ProfileSummary::PSK_CSInstr : ProfileSummary::PSK_Instr, |
| 1215 | args&: DetailedSummary, args: SummaryData->get(K: Summary::TotalBlockCount), |
| 1216 | args: SummaryData->get(K: Summary::MaxBlockCount), |
| 1217 | args: SummaryData->get(K: Summary::MaxInternalBlockCount), |
| 1218 | args: SummaryData->get(K: Summary::MaxFunctionCount), |
| 1219 | args: SummaryData->get(K: Summary::TotalNumBlocks), |
| 1220 | args: SummaryData->get(K: Summary::TotalNumFunctions)); |
| 1221 | return Cur + SummarySize; |
| 1222 | } else { |
| 1223 | // The older versions do not support a profile summary. This just computes |
| 1224 | // an empty summary, which will not result in accurate hot/cold detection. |
| 1225 | // We would need to call addRecord for all NamedInstrProfRecords to get the |
| 1226 | // correct summary. However, this version is old (prior to early 2016) and |
| 1227 | // has not been supporting an accurate summary for several years. |
| 1228 | InstrProfSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs); |
| 1229 | Summary = Builder.getSummary(); |
| 1230 | return Cur; |
| 1231 | } |
| 1232 | } |
| 1233 | |
| 1234 | Error IndexedInstrProfReader::() { |
| 1235 | using namespace support; |
| 1236 | |
| 1237 | const unsigned char *Start = |
| 1238 | (const unsigned char *)DataBuffer->getBufferStart(); |
| 1239 | const unsigned char *Cur = Start; |
| 1240 | if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24) |
| 1241 | return error(Err: instrprof_error::truncated); |
| 1242 | |
| 1243 | auto = IndexedInstrProf::Header::readFromBuffer(Buffer: Start); |
| 1244 | if (!HeaderOr) |
| 1245 | return HeaderOr.takeError(); |
| 1246 | |
| 1247 | const IndexedInstrProf::Header * = &HeaderOr.get(); |
| 1248 | Cur += Header->size(); |
| 1249 | |
| 1250 | Cur = readSummary(Version: (IndexedInstrProf::ProfVersion)Header->Version, Cur, |
| 1251 | /* UseCS */ false); |
| 1252 | if (Header->Version & VARIANT_MASK_CSIR_PROF) |
| 1253 | Cur = readSummary(Version: (IndexedInstrProf::ProfVersion)Header->Version, Cur, |
| 1254 | /* UseCS */ true); |
| 1255 | // Read the hash type and start offset. |
| 1256 | IndexedInstrProf::HashT HashType = |
| 1257 | static_cast<IndexedInstrProf::HashT>(Header->HashType); |
| 1258 | if (HashType > IndexedInstrProf::HashT::Last) |
| 1259 | return error(Err: instrprof_error::unsupported_hash_type); |
| 1260 | |
| 1261 | // The hash table with profile counts comes next. |
| 1262 | auto IndexPtr = std::make_unique<InstrProfReaderIndex<OnDiskHashTableImplV3>>( |
| 1263 | args: Start + Header->HashOffset, args&: Cur, args&: Start, args&: HashType, args: Header->Version); |
| 1264 | |
| 1265 | // The MemProfOffset field in the header is only valid when the format |
| 1266 | // version is higher than 8 (when it was introduced). |
| 1267 | if (Header->getIndexedProfileVersion() >= 8 && |
| 1268 | Header->Version & VARIANT_MASK_MEMPROF) { |
| 1269 | if (Error E = MemProfReader.deserialize(Start, MemProfOffset: Header->MemProfOffset)) |
| 1270 | return E; |
| 1271 | } |
| 1272 | |
| 1273 | // BinaryIdOffset field in the header is only valid when the format version |
| 1274 | // is higher than 9 (when it was introduced). |
| 1275 | if (Header->getIndexedProfileVersion() >= 9) { |
| 1276 | const unsigned char *Ptr = Start + Header->BinaryIdOffset; |
| 1277 | // Read binary ids size. |
| 1278 | uint64_t BinaryIdsSize = |
| 1279 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
| 1280 | if (BinaryIdsSize % sizeof(uint64_t)) |
| 1281 | return error(Err: instrprof_error::bad_header); |
| 1282 | // Set the binary ids start. |
| 1283 | BinaryIdsBuffer = ArrayRef<uint8_t>(Ptr, BinaryIdsSize); |
| 1284 | if (Ptr > (const unsigned char *)DataBuffer->getBufferEnd()) |
| 1285 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
| 1286 | Args: "corrupted binary ids" ); |
| 1287 | } |
| 1288 | |
| 1289 | if (Header->getIndexedProfileVersion() >= 12) { |
| 1290 | const unsigned char *Ptr = Start + Header->VTableNamesOffset; |
| 1291 | |
| 1292 | uint64_t CompressedVTableNamesLen = |
| 1293 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
| 1294 | |
| 1295 | // Writer first writes the length of compressed string, and then the actual |
| 1296 | // content. |
| 1297 | const char *VTableNamePtr = (const char *)Ptr; |
| 1298 | if (VTableNamePtr > (const char *)DataBuffer->getBufferEnd()) |
| 1299 | return make_error<InstrProfError>(Args: instrprof_error::truncated); |
| 1300 | |
| 1301 | VTableName = StringRef(VTableNamePtr, CompressedVTableNamesLen); |
| 1302 | } |
| 1303 | |
| 1304 | if (Header->getIndexedProfileVersion() >= 10 && |
| 1305 | Header->Version & VARIANT_MASK_TEMPORAL_PROF) { |
| 1306 | const unsigned char *Ptr = Start + Header->TemporalProfTracesOffset; |
| 1307 | const auto *PtrEnd = (const unsigned char *)DataBuffer->getBufferEnd(); |
| 1308 | // Expect at least two 64 bit fields: NumTraces, and TraceStreamSize |
| 1309 | if (Ptr + 2 * sizeof(uint64_t) > PtrEnd) |
| 1310 | return error(Err: instrprof_error::truncated); |
| 1311 | const uint64_t NumTraces = |
| 1312 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
| 1313 | TemporalProfTraceStreamSize = |
| 1314 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
| 1315 | for (unsigned i = 0; i < NumTraces; i++) { |
| 1316 | // Expect at least two 64 bit fields: Weight and NumFunctions |
| 1317 | if (Ptr + 2 * sizeof(uint64_t) > PtrEnd) |
| 1318 | return error(Err: instrprof_error::truncated); |
| 1319 | TemporalProfTraceTy Trace; |
| 1320 | Trace.Weight = |
| 1321 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
| 1322 | const uint64_t NumFunctions = |
| 1323 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
| 1324 | // Expect at least NumFunctions 64 bit fields |
| 1325 | if (Ptr + NumFunctions * sizeof(uint64_t) > PtrEnd) |
| 1326 | return error(Err: instrprof_error::truncated); |
| 1327 | for (unsigned j = 0; j < NumFunctions; j++) { |
| 1328 | const uint64_t NameRef = |
| 1329 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
| 1330 | Trace.FunctionNameRefs.push_back(x: NameRef); |
| 1331 | } |
| 1332 | TemporalProfTraces.push_back(Elt: std::move(Trace)); |
| 1333 | } |
| 1334 | } |
| 1335 | |
| 1336 | // Load the remapping table now if requested. |
| 1337 | if (RemappingBuffer) { |
| 1338 | Remapper = |
| 1339 | std::make_unique<InstrProfReaderItaniumRemapper<OnDiskHashTableImplV3>>( |
| 1340 | args: std::move(RemappingBuffer), args&: *IndexPtr); |
| 1341 | if (Error E = Remapper->populateRemappings()) |
| 1342 | return E; |
| 1343 | } else { |
| 1344 | Remapper = std::make_unique<InstrProfReaderNullRemapper>(args&: *IndexPtr); |
| 1345 | } |
| 1346 | Index = std::move(IndexPtr); |
| 1347 | |
| 1348 | return success(); |
| 1349 | } |
| 1350 | |
| 1351 | InstrProfSymtab &IndexedInstrProfReader::getSymtab() { |
| 1352 | if (Symtab) |
| 1353 | return *Symtab; |
| 1354 | |
| 1355 | auto NewSymtab = std::make_unique<InstrProfSymtab>(); |
| 1356 | |
| 1357 | if (Error E = NewSymtab->initVTableNamesFromCompressedStrings(CompressedVTableNames: VTableName)) { |
| 1358 | auto [ErrCode, Msg] = InstrProfError::take(E: std::move(E)); |
| 1359 | consumeError(Err: error(Err: ErrCode, ErrMsg: Msg)); |
| 1360 | } |
| 1361 | |
| 1362 | // finalizeSymtab is called inside populateSymtab. |
| 1363 | if (Error E = Index->populateSymtab(*NewSymtab)) { |
| 1364 | auto [ErrCode, Msg] = InstrProfError::take(E: std::move(E)); |
| 1365 | consumeError(Err: error(Err: ErrCode, ErrMsg: Msg)); |
| 1366 | } |
| 1367 | |
| 1368 | Symtab = std::move(NewSymtab); |
| 1369 | return *Symtab; |
| 1370 | } |
| 1371 | |
| 1372 | Expected<NamedInstrProfRecord> IndexedInstrProfReader::getInstrProfRecord( |
| 1373 | StringRef FuncName, uint64_t FuncHash, StringRef DeprecatedFuncName, |
| 1374 | uint64_t *MismatchedFuncSum) { |
| 1375 | ArrayRef<NamedInstrProfRecord> Data; |
| 1376 | uint64_t FuncSum = 0; |
| 1377 | auto Err = Remapper->getRecords(FuncName, Data); |
| 1378 | if (Err) { |
| 1379 | // If we don't find FuncName, try DeprecatedFuncName to handle profiles |
| 1380 | // built by older compilers. |
| 1381 | auto Err2 = |
| 1382 | handleErrors(E: std::move(Err), Hs: [&](const InstrProfError &IE) -> Error { |
| 1383 | if (IE.get() != instrprof_error::unknown_function) |
| 1384 | return make_error<InstrProfError>(Args: IE); |
| 1385 | if (auto Err = Remapper->getRecords(FuncName: DeprecatedFuncName, Data)) |
| 1386 | return Err; |
| 1387 | return Error::success(); |
| 1388 | }); |
| 1389 | if (Err2) |
| 1390 | return std::move(Err2); |
| 1391 | } |
| 1392 | // Found it. Look for counters with the right hash. |
| 1393 | |
| 1394 | // A flag to indicate if the records are from the same type |
| 1395 | // of profile (i.e cs vs nocs). |
| 1396 | bool CSBitMatch = false; |
| 1397 | auto getFuncSum = [](ArrayRef<uint64_t> Counts) { |
| 1398 | uint64_t ValueSum = 0; |
| 1399 | for (uint64_t CountValue : Counts) { |
| 1400 | if (CountValue == (uint64_t)-1) |
| 1401 | continue; |
| 1402 | // Handle overflow -- if that happens, return max. |
| 1403 | if (std::numeric_limits<uint64_t>::max() - CountValue <= ValueSum) |
| 1404 | return std::numeric_limits<uint64_t>::max(); |
| 1405 | ValueSum += CountValue; |
| 1406 | } |
| 1407 | return ValueSum; |
| 1408 | }; |
| 1409 | |
| 1410 | for (const NamedInstrProfRecord &I : Data) { |
| 1411 | // Check for a match and fill the vector if there is one. |
| 1412 | if (I.Hash == FuncHash) |
| 1413 | return std::move(I); |
| 1414 | if (NamedInstrProfRecord::hasCSFlagInHash(FuncHash: I.Hash) == |
| 1415 | NamedInstrProfRecord::hasCSFlagInHash(FuncHash)) { |
| 1416 | CSBitMatch = true; |
| 1417 | if (MismatchedFuncSum == nullptr) |
| 1418 | continue; |
| 1419 | FuncSum = std::max(a: FuncSum, b: getFuncSum(I.Counts)); |
| 1420 | } |
| 1421 | } |
| 1422 | if (CSBitMatch) { |
| 1423 | if (MismatchedFuncSum != nullptr) |
| 1424 | *MismatchedFuncSum = FuncSum; |
| 1425 | return error(Err: instrprof_error::hash_mismatch); |
| 1426 | } |
| 1427 | return error(Err: instrprof_error::unknown_function); |
| 1428 | } |
| 1429 | |
| 1430 | static Expected<memprof::MemProfRecord> |
| 1431 | getMemProfRecordV2(const memprof::IndexedMemProfRecord &IndexedRecord, |
| 1432 | MemProfFrameHashTable &MemProfFrameTable, |
| 1433 | MemProfCallStackHashTable &MemProfCallStackTable) { |
| 1434 | memprof::FrameIdConverter<MemProfFrameHashTable> FrameIdConv( |
| 1435 | MemProfFrameTable); |
| 1436 | |
| 1437 | memprof::CallStackIdConverter<MemProfCallStackHashTable> CSIdConv( |
| 1438 | MemProfCallStackTable, FrameIdConv); |
| 1439 | |
| 1440 | memprof::MemProfRecord Record = IndexedRecord.toMemProfRecord(Callback: CSIdConv); |
| 1441 | |
| 1442 | // Check that all call stack ids were successfully converted to call stacks. |
| 1443 | if (CSIdConv.LastUnmappedId) { |
| 1444 | return make_error<InstrProfError>( |
| 1445 | Args: instrprof_error::hash_mismatch, |
| 1446 | Args: "memprof call stack not found for call stack id " + |
| 1447 | Twine(*CSIdConv.LastUnmappedId)); |
| 1448 | } |
| 1449 | |
| 1450 | // Check that all frame ids were successfully converted to frames. |
| 1451 | if (FrameIdConv.LastUnmappedId) { |
| 1452 | return make_error<InstrProfError>(Args: instrprof_error::hash_mismatch, |
| 1453 | Args: "memprof frame not found for frame id " + |
| 1454 | Twine(*FrameIdConv.LastUnmappedId)); |
| 1455 | } |
| 1456 | |
| 1457 | return Record; |
| 1458 | } |
| 1459 | |
| 1460 | Expected<memprof::MemProfRecord> |
| 1461 | IndexedMemProfReader::getMemProfRecord(const uint64_t FuncNameHash) const { |
| 1462 | // TODO: Add memprof specific errors. |
| 1463 | if (MemProfRecordTable == nullptr) |
| 1464 | return make_error<InstrProfError>(Args: instrprof_error::invalid_prof, |
| 1465 | Args: "no memprof data available in profile" ); |
| 1466 | auto Iter = MemProfRecordTable->find(EKey: FuncNameHash); |
| 1467 | if (Iter == MemProfRecordTable->end()) |
| 1468 | return make_error<InstrProfError>( |
| 1469 | Args: instrprof_error::unknown_function, |
| 1470 | Args: "memprof record not found for function hash " + Twine(FuncNameHash)); |
| 1471 | |
| 1472 | const memprof::IndexedMemProfRecord &IndexedRecord = *Iter; |
| 1473 | switch (Version) { |
| 1474 | case memprof::Version2: |
| 1475 | assert(MemProfFrameTable && "MemProfFrameTable must be available" ); |
| 1476 | assert(MemProfCallStackTable && "MemProfCallStackTable must be available" ); |
| 1477 | return getMemProfRecordV2(IndexedRecord, MemProfFrameTable&: *MemProfFrameTable, |
| 1478 | MemProfCallStackTable&: *MemProfCallStackTable); |
| 1479 | // Combine V3 and V4 cases as the record conversion logic is the same. |
| 1480 | case memprof::Version3: |
| 1481 | case memprof::Version4: |
| 1482 | assert(!MemProfFrameTable && "MemProfFrameTable must not be available" ); |
| 1483 | assert(!MemProfCallStackTable && |
| 1484 | "MemProfCallStackTable must not be available" ); |
| 1485 | assert(FrameBase && "FrameBase must be available" ); |
| 1486 | assert(CallStackBase && "CallStackBase must be available" ); |
| 1487 | { |
| 1488 | memprof::LinearFrameIdConverter FrameIdConv(FrameBase); |
| 1489 | memprof::LinearCallStackIdConverter CSIdConv(CallStackBase, FrameIdConv); |
| 1490 | memprof::MemProfRecord Record = IndexedRecord.toMemProfRecord(Callback: CSIdConv); |
| 1491 | return Record; |
| 1492 | } |
| 1493 | } |
| 1494 | |
| 1495 | return make_error<InstrProfError>( |
| 1496 | Args: instrprof_error::unsupported_version, |
| 1497 | Args: formatv(Fmt: "MemProf version {} not supported; " |
| 1498 | "requires version between {} and {}, inclusive" , |
| 1499 | Vals: Version, Vals: memprof::MinimumSupportedVersion, |
| 1500 | Vals: memprof::MaximumSupportedVersion)); |
| 1501 | } |
| 1502 | |
| 1503 | DenseMap<uint64_t, SmallVector<memprof::CallEdgeTy, 0>> |
| 1504 | IndexedMemProfReader::getMemProfCallerCalleePairs() const { |
| 1505 | assert(MemProfRecordTable); |
| 1506 | assert(Version == memprof::Version3 || Version == memprof::Version4); |
| 1507 | |
| 1508 | memprof::LinearFrameIdConverter FrameIdConv(FrameBase); |
| 1509 | memprof::CallerCalleePairExtractor (CallStackBase, FrameIdConv, |
| 1510 | RadixTreeSize); |
| 1511 | |
| 1512 | // The set of linear call stack IDs that we need to traverse from. We expect |
| 1513 | // the set to be dense, so we use a BitVector. |
| 1514 | BitVector Worklist(RadixTreeSize); |
| 1515 | |
| 1516 | // Collect the set of linear call stack IDs. Since we expect a lot of |
| 1517 | // duplicates, we first collect them in the form of a bit vector before |
| 1518 | // processing them. |
| 1519 | for (const memprof::IndexedMemProfRecord &IndexedRecord : |
| 1520 | MemProfRecordTable->data()) { |
| 1521 | for (const memprof::IndexedAllocationInfo &IndexedAI : |
| 1522 | IndexedRecord.AllocSites) |
| 1523 | Worklist.set(IndexedAI.CSId); |
| 1524 | } |
| 1525 | |
| 1526 | // Collect caller-callee pairs for each linear call stack ID in Worklist. |
| 1527 | for (unsigned CS : Worklist.set_bits()) |
| 1528 | Extractor(CS); |
| 1529 | |
| 1530 | DenseMap<uint64_t, SmallVector<memprof::CallEdgeTy, 0>> Pairs = |
| 1531 | std::move(Extractor.CallerCalleePairs); |
| 1532 | |
| 1533 | // Sort each call list by the source location. |
| 1534 | for (auto &[CallerGUID, CallList] : Pairs) { |
| 1535 | llvm::sort(C&: CallList); |
| 1536 | CallList.erase(CS: llvm::unique(R&: CallList), CE: CallList.end()); |
| 1537 | } |
| 1538 | |
| 1539 | return Pairs; |
| 1540 | } |
| 1541 | |
| 1542 | memprof::AllMemProfData IndexedMemProfReader::getAllMemProfData() const { |
| 1543 | memprof::AllMemProfData AllMemProfData; |
| 1544 | AllMemProfData.HeapProfileRecords.reserve( |
| 1545 | n: MemProfRecordTable->getNumEntries()); |
| 1546 | for (uint64_t Key : MemProfRecordTable->keys()) { |
| 1547 | auto Record = getMemProfRecord(FuncNameHash: Key); |
| 1548 | if (Record.takeError()) |
| 1549 | continue; |
| 1550 | memprof::GUIDMemProfRecordPair Pair; |
| 1551 | Pair.GUID = Key; |
| 1552 | Pair.Record = std::move(*Record); |
| 1553 | AllMemProfData.HeapProfileRecords.push_back(x: std::move(Pair)); |
| 1554 | } |
| 1555 | // Populate the data access profiles for yaml output. |
| 1556 | if (DataAccessProfileData != nullptr) { |
| 1557 | for (const auto &[SymHandleRef, RecordRef] : |
| 1558 | DataAccessProfileData->getRecords()) |
| 1559 | AllMemProfData.YamlifiedDataAccessProfiles.Records.push_back( |
| 1560 | x: memprof::DataAccessProfRecord(SymHandleRef, RecordRef.AccessCount, |
| 1561 | RecordRef.Locations)); |
| 1562 | for (StringRef ColdSymbol : DataAccessProfileData->getKnownColdSymbols()) |
| 1563 | AllMemProfData.YamlifiedDataAccessProfiles.KnownColdSymbols.push_back( |
| 1564 | x: ColdSymbol.str()); |
| 1565 | for (uint64_t Hash : DataAccessProfileData->getKnownColdHashes()) |
| 1566 | AllMemProfData.YamlifiedDataAccessProfiles.KnownColdStrHashes.push_back( |
| 1567 | x: Hash); |
| 1568 | } |
| 1569 | return AllMemProfData; |
| 1570 | } |
| 1571 | |
| 1572 | Error IndexedInstrProfReader::getFunctionCounts(StringRef FuncName, |
| 1573 | uint64_t FuncHash, |
| 1574 | std::vector<uint64_t> &Counts) { |
| 1575 | auto Record = getInstrProfRecord(FuncName, FuncHash); |
| 1576 | if (Error E = Record.takeError()) |
| 1577 | return error(E: std::move(E)); |
| 1578 | |
| 1579 | Counts = Record.get().Counts; |
| 1580 | return success(); |
| 1581 | } |
| 1582 | |
| 1583 | Error IndexedInstrProfReader::getFunctionBitmap(StringRef FuncName, |
| 1584 | uint64_t FuncHash, |
| 1585 | BitVector &Bitmap) { |
| 1586 | auto Record = getInstrProfRecord(FuncName, FuncHash); |
| 1587 | if (Error E = Record.takeError()) |
| 1588 | return error(E: std::move(E)); |
| 1589 | |
| 1590 | const auto &BitmapBytes = Record.get().BitmapBytes; |
| 1591 | size_t I = 0, E = BitmapBytes.size(); |
| 1592 | Bitmap.resize(N: E * CHAR_BIT); |
| 1593 | BitVector::apply( |
| 1594 | f: [&](auto X) { |
| 1595 | using XTy = decltype(X); |
| 1596 | alignas(XTy) uint8_t W[sizeof(X)]; |
| 1597 | size_t N = std::min(a: E - I, b: sizeof(W)); |
| 1598 | std::memset(s: W, c: 0, n: sizeof(W)); |
| 1599 | std::memcpy(dest: W, src: &BitmapBytes[I], n: N); |
| 1600 | I += N; |
| 1601 | return support::endian::read<XTy, llvm::endianness::little, |
| 1602 | support::aligned>(W); |
| 1603 | }, |
| 1604 | Out&: Bitmap, Arg: Bitmap); |
| 1605 | assert(I == E); |
| 1606 | |
| 1607 | return success(); |
| 1608 | } |
| 1609 | |
| 1610 | Error IndexedInstrProfReader::readNextRecord(NamedInstrProfRecord &Record) { |
| 1611 | ArrayRef<NamedInstrProfRecord> Data; |
| 1612 | |
| 1613 | Error E = Index->getRecords(Data); |
| 1614 | if (E) |
| 1615 | return error(E: std::move(E)); |
| 1616 | |
| 1617 | Record = Data[RecordIndex++]; |
| 1618 | if (RecordIndex >= Data.size()) { |
| 1619 | Index->advanceToNextKey(); |
| 1620 | RecordIndex = 0; |
| 1621 | } |
| 1622 | return success(); |
| 1623 | } |
| 1624 | |
| 1625 | Error IndexedInstrProfReader::readBinaryIds( |
| 1626 | std::vector<llvm::object::BuildID> &BinaryIds) { |
| 1627 | return readBinaryIdsInternal(DataBuffer: *DataBuffer, BinaryIdsBuffer, BinaryIds, |
| 1628 | Endian: llvm::endianness::little); |
| 1629 | } |
| 1630 | |
| 1631 | Error IndexedInstrProfReader::printBinaryIds(raw_ostream &OS) { |
| 1632 | std::vector<llvm::object::BuildID> BinaryIds; |
| 1633 | if (Error E = readBinaryIds(BinaryIds)) |
| 1634 | return E; |
| 1635 | printBinaryIdsInternal(OS, BinaryIds); |
| 1636 | return Error::success(); |
| 1637 | } |
| 1638 | |
| 1639 | void InstrProfReader::accumulateCounts(CountSumOrPercent &Sum, bool IsCS) { |
| 1640 | uint64_t NumFuncs = 0; |
| 1641 | for (const auto &Func : *this) { |
| 1642 | if (isIRLevelProfile()) { |
| 1643 | bool FuncIsCS = NamedInstrProfRecord::hasCSFlagInHash(FuncHash: Func.Hash); |
| 1644 | if (FuncIsCS != IsCS) |
| 1645 | continue; |
| 1646 | } |
| 1647 | Func.accumulateCounts(Sum); |
| 1648 | ++NumFuncs; |
| 1649 | } |
| 1650 | Sum.NumEntries = NumFuncs; |
| 1651 | } |
| 1652 | |