1 //===- bolt/Profile/BoltAddressTranslation.cpp ----------------------------===// 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 #include "bolt/Profile/BoltAddressTranslation.h" 10 #include "bolt/Core/BinaryFunction.h" 11 #include "llvm/ADT/APInt.h" 12 #include "llvm/Support/Errc.h" 13 #include "llvm/Support/Error.h" 14 #include "llvm/Support/LEB128.h" 15 16 #define DEBUG_TYPE "bolt-bat" 17 18 namespace llvm { 19 namespace bolt { 20 21 const char *BoltAddressTranslation::SECTION_NAME = ".note.bolt_bat"; 22 23 void BoltAddressTranslation::writeEntriesForBB(MapTy &Map, 24 const BinaryBasicBlock &BB, 25 uint64_t FuncInputAddress, 26 uint64_t FuncOutputAddress) { 27 const uint64_t BBOutputOffset = 28 BB.getOutputAddressRange().first - FuncOutputAddress; 29 const uint32_t BBInputOffset = BB.getInputOffset(); 30 31 // Every output BB must track back to an input BB for profile collection 32 // in bolted binaries. If we are missing an offset, it means this block was 33 // created by a pass. We will skip writing any entries for it, and this means 34 // any traffic happening in this block will map to the previous block in the 35 // layout. This covers the case where an input basic block is split into two, 36 // and the second one lacks any offset. 37 if (BBInputOffset == BinaryBasicBlock::INVALID_OFFSET) 38 return; 39 40 LLVM_DEBUG(dbgs() << "BB " << BB.getName() << "\n"); 41 LLVM_DEBUG(dbgs() << " Key: " << Twine::utohexstr(BBOutputOffset) 42 << " Val: " << Twine::utohexstr(BBInputOffset) << "\n"); 43 // NB: in `writeEntriesForBB` we use the input address because hashes are 44 // saved early in `saveMetadata` before output addresses are assigned. 45 const BBHashMapTy &BBHashMap = getBBHashMap(FuncInputAddress); 46 (void)BBHashMap; 47 LLVM_DEBUG( 48 dbgs() << formatv(" Hash: {0:x}\n", BBHashMap.getBBHash(BBInputOffset))); 49 LLVM_DEBUG( 50 dbgs() << formatv(" Index: {0}\n", BBHashMap.getBBIndex(BBInputOffset))); 51 // In case of conflicts (same Key mapping to different Vals), the last 52 // update takes precedence. Of course it is not ideal to have conflicts and 53 // those happen when we have an empty BB that either contained only 54 // NOPs or a jump to the next block (successor). Either way, the successor 55 // and this deleted block will both share the same output address (the same 56 // key), and we need to map back. We choose here to privilege the successor by 57 // allowing it to overwrite the previously inserted key in the map. 58 Map[BBOutputOffset] = BBInputOffset << 1; 59 60 const auto &IOAddressMap = 61 BB.getFunction()->getBinaryContext().getIOAddressMap(); 62 63 for (const auto &[InputOffset, Sym] : BB.getLocSyms()) { 64 const auto InputAddress = BB.getFunction()->getAddress() + InputOffset; 65 const auto OutputAddress = IOAddressMap.lookup(InputAddress); 66 assert(OutputAddress && "Unknown instruction address"); 67 const auto OutputOffset = *OutputAddress - FuncOutputAddress; 68 69 // Is this the first instruction in the BB? No need to duplicate the entry. 70 if (OutputOffset == BBOutputOffset) 71 continue; 72 73 LLVM_DEBUG(dbgs() << " Key: " << Twine::utohexstr(OutputOffset) << " Val: " 74 << Twine::utohexstr(InputOffset) << " (branch)\n"); 75 Map.insert(std::pair<uint32_t, uint32_t>(OutputOffset, 76 (InputOffset << 1) | BRANCHENTRY)); 77 } 78 } 79 80 void BoltAddressTranslation::write(const BinaryContext &BC, raw_ostream &OS) { 81 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: Writing BOLT Address Translation Tables\n"); 82 for (auto &BFI : BC.getBinaryFunctions()) { 83 const BinaryFunction &Function = BFI.second; 84 const uint64_t InputAddress = Function.getAddress(); 85 const uint64_t OutputAddress = Function.getOutputAddress(); 86 // We don't need a translation table if the body of the function hasn't 87 // changed 88 if (Function.isIgnored() || (!BC.HasRelocations && !Function.isSimple())) 89 continue; 90 91 uint32_t NumSecondaryEntryPoints = 0; 92 Function.forEachEntryPoint([&](uint64_t Offset, const MCSymbol *) { 93 if (!Offset) 94 return true; 95 ++NumSecondaryEntryPoints; 96 SecondaryEntryPointsMap[OutputAddress].push_back(Offset); 97 return true; 98 }); 99 100 LLVM_DEBUG(dbgs() << "Function name: " << Function.getPrintName() << "\n"); 101 LLVM_DEBUG(dbgs() << " Address reference: 0x" 102 << Twine::utohexstr(Function.getOutputAddress()) << "\n"); 103 LLVM_DEBUG(dbgs() << formatv(" Hash: {0:x}\n", getBFHash(OutputAddress))); 104 LLVM_DEBUG(dbgs() << " Secondary Entry Points: " << NumSecondaryEntryPoints 105 << '\n'); 106 107 MapTy Map; 108 for (const BinaryBasicBlock *const BB : 109 Function.getLayout().getMainFragment()) 110 writeEntriesForBB(Map, *BB, InputAddress, OutputAddress); 111 Maps.emplace(Function.getOutputAddress(), std::move(Map)); 112 ReverseMap.emplace(OutputAddress, InputAddress); 113 114 if (!Function.isSplit()) 115 continue; 116 117 // Split maps 118 LLVM_DEBUG(dbgs() << " Cold part\n"); 119 for (const FunctionFragment &FF : 120 Function.getLayout().getSplitFragments()) { 121 ColdPartSource.emplace(FF.getAddress(), Function.getOutputAddress()); 122 Map.clear(); 123 for (const BinaryBasicBlock *const BB : FF) 124 writeEntriesForBB(Map, *BB, InputAddress, FF.getAddress()); 125 126 Maps.emplace(FF.getAddress(), std::move(Map)); 127 } 128 } 129 130 // Output addresses are delta-encoded 131 uint64_t PrevAddress = 0; 132 writeMaps</*Cold=*/false>(Maps, PrevAddress, OS); 133 writeMaps</*Cold=*/true>(Maps, PrevAddress, OS); 134 135 BC.outs() << "BOLT-INFO: Wrote " << Maps.size() << " BAT maps\n"; 136 BC.outs() << "BOLT-INFO: Wrote " << FuncHashes.getNumFunctions() 137 << " function and " << FuncHashes.getNumBasicBlocks() 138 << " basic block hashes\n"; 139 } 140 141 APInt BoltAddressTranslation::calculateBranchEntriesBitMask(MapTy &Map, 142 size_t EqualElems) { 143 APInt BitMask(alignTo(EqualElems, 8), 0); 144 size_t Index = 0; 145 for (std::pair<const uint32_t, uint32_t> &KeyVal : Map) { 146 if (Index == EqualElems) 147 break; 148 const uint32_t OutputOffset = KeyVal.second; 149 if (OutputOffset & BRANCHENTRY) 150 BitMask.setBit(Index); 151 ++Index; 152 } 153 return BitMask; 154 } 155 156 size_t BoltAddressTranslation::getNumEqualOffsets(const MapTy &Map) const { 157 size_t EqualOffsets = 0; 158 for (const std::pair<const uint32_t, uint32_t> &KeyVal : Map) { 159 const uint32_t OutputOffset = KeyVal.first; 160 const uint32_t InputOffset = KeyVal.second >> 1; 161 if (OutputOffset == InputOffset) 162 ++EqualOffsets; 163 else 164 break; 165 } 166 return EqualOffsets; 167 } 168 169 template <bool Cold> 170 void BoltAddressTranslation::writeMaps(std::map<uint64_t, MapTy> &Maps, 171 uint64_t &PrevAddress, raw_ostream &OS) { 172 const uint32_t NumFuncs = 173 llvm::count_if(llvm::make_first_range(Maps), [&](const uint64_t Address) { 174 return Cold == ColdPartSource.count(Address); 175 }); 176 encodeULEB128(NumFuncs, OS); 177 LLVM_DEBUG(dbgs() << "Writing " << NumFuncs << (Cold ? " cold" : "") 178 << " functions for BAT.\n"); 179 size_t PrevIndex = 0; 180 for (auto &MapEntry : Maps) { 181 const uint64_t Address = MapEntry.first; 182 // Only process cold fragments in cold mode, and vice versa. 183 if (Cold != ColdPartSource.count(Address)) 184 continue; 185 // NB: in `writeMaps` we use the input address because hashes are saved 186 // early in `saveMetadata` before output addresses are assigned. 187 const uint64_t HotInputAddress = 188 ReverseMap[Cold ? ColdPartSource[Address] : Address]; 189 MapTy &Map = MapEntry.second; 190 const uint32_t NumEntries = Map.size(); 191 LLVM_DEBUG(dbgs() << "Writing " << NumEntries << " entries for 0x" 192 << Twine::utohexstr(Address) << ".\n"); 193 encodeULEB128(Address - PrevAddress, OS); 194 PrevAddress = Address; 195 const uint32_t NumSecondaryEntryPoints = 196 SecondaryEntryPointsMap.count(Address) 197 ? SecondaryEntryPointsMap[Address].size() 198 : 0; 199 if (Cold) { 200 size_t HotIndex = 201 std::distance(ColdPartSource.begin(), ColdPartSource.find(Address)); 202 encodeULEB128(HotIndex - PrevIndex, OS); 203 PrevIndex = HotIndex; 204 } else { 205 // Function hash 206 size_t BFHash = getBFHash(HotInputAddress); 207 LLVM_DEBUG(dbgs() << "Hash: " << formatv("{0:x}\n", BFHash)); 208 OS.write(reinterpret_cast<char *>(&BFHash), 8); 209 // Number of basic blocks 210 size_t NumBasicBlocks = getBBHashMap(HotInputAddress).getNumBasicBlocks(); 211 LLVM_DEBUG(dbgs() << "Basic blocks: " << NumBasicBlocks << '\n'); 212 encodeULEB128(NumBasicBlocks, OS); 213 // Secondary entry points 214 encodeULEB128(NumSecondaryEntryPoints, OS); 215 LLVM_DEBUG(dbgs() << "Secondary Entry Points: " << NumSecondaryEntryPoints 216 << '\n'); 217 } 218 encodeULEB128(NumEntries, OS); 219 // For hot fragments only: encode the number of equal offsets 220 // (output = input) in the beginning of the function. Only encode one offset 221 // in these cases. 222 const size_t EqualElems = Cold ? 0 : getNumEqualOffsets(Map); 223 if (!Cold) { 224 encodeULEB128(EqualElems, OS); 225 if (EqualElems) { 226 const size_t BranchEntriesBytes = alignTo(EqualElems, 8) / 8; 227 APInt BranchEntries = calculateBranchEntriesBitMask(Map, EqualElems); 228 OS.write(reinterpret_cast<const char *>(BranchEntries.getRawData()), 229 BranchEntriesBytes); 230 LLVM_DEBUG({ 231 dbgs() << "BranchEntries: "; 232 SmallString<8> BitMaskStr; 233 BranchEntries.toString(BitMaskStr, 2, false); 234 dbgs() << BitMaskStr << '\n'; 235 }); 236 } 237 } 238 const BBHashMapTy &BBHashMap = getBBHashMap(HotInputAddress); 239 size_t Index = 0; 240 uint64_t InOffset = 0; 241 size_t PrevBBIndex = 0; 242 // Output and Input addresses and delta-encoded 243 for (std::pair<const uint32_t, uint32_t> &KeyVal : Map) { 244 const uint64_t OutputAddress = KeyVal.first + Address; 245 encodeULEB128(OutputAddress - PrevAddress, OS); 246 PrevAddress = OutputAddress; 247 if (Index++ >= EqualElems) 248 encodeSLEB128(KeyVal.second - InOffset, OS); 249 InOffset = KeyVal.second; // Keeping InOffset as if BRANCHENTRY is encoded 250 if ((InOffset & BRANCHENTRY) == 0) { 251 const bool IsBlock = BBHashMap.isInputBlock(InOffset >> 1); 252 unsigned BBIndex = IsBlock ? BBHashMap.getBBIndex(InOffset >> 1) : 0; 253 size_t BBHash = IsBlock ? BBHashMap.getBBHash(InOffset >> 1) : 0; 254 OS.write(reinterpret_cast<char *>(&BBHash), 8); 255 // Basic block index in the input binary 256 encodeULEB128(BBIndex - PrevBBIndex, OS); 257 PrevBBIndex = BBIndex; 258 LLVM_DEBUG(dbgs() << formatv("{0:x} -> {1:x} {2:x} {3}\n", KeyVal.first, 259 InOffset >> 1, BBHash, BBIndex)); 260 } 261 } 262 uint32_t PrevOffset = 0; 263 if (!Cold && NumSecondaryEntryPoints) { 264 LLVM_DEBUG(dbgs() << "Secondary entry points: "); 265 // Secondary entry point offsets, delta-encoded 266 for (uint32_t Offset : SecondaryEntryPointsMap[Address]) { 267 encodeULEB128(Offset - PrevOffset, OS); 268 LLVM_DEBUG(dbgs() << formatv("{0:x} ", Offset)); 269 PrevOffset = Offset; 270 } 271 LLVM_DEBUG(dbgs() << '\n'); 272 } 273 } 274 } 275 276 std::error_code BoltAddressTranslation::parse(raw_ostream &OS, StringRef Buf) { 277 DataExtractor DE = DataExtractor(Buf, true, 8); 278 uint64_t Offset = 0; 279 if (Buf.size() < 12) 280 return make_error_code(llvm::errc::io_error); 281 282 const uint32_t NameSz = DE.getU32(&Offset); 283 const uint32_t DescSz = DE.getU32(&Offset); 284 const uint32_t Type = DE.getU32(&Offset); 285 286 if (Type != BinarySection::NT_BOLT_BAT || 287 Buf.size() + Offset < alignTo(NameSz, 4) + DescSz) 288 return make_error_code(llvm::errc::io_error); 289 290 StringRef Name = Buf.slice(Offset, Offset + NameSz); 291 Offset = alignTo(Offset + NameSz, 4); 292 if (Name.substr(0, 4) != "BOLT") 293 return make_error_code(llvm::errc::io_error); 294 295 Error Err(Error::success()); 296 std::vector<uint64_t> HotFuncs; 297 uint64_t PrevAddress = 0; 298 parseMaps</*Cold=*/false>(HotFuncs, PrevAddress, DE, Offset, Err); 299 parseMaps</*Cold=*/true>(HotFuncs, PrevAddress, DE, Offset, Err); 300 OS << "BOLT-INFO: Parsed " << Maps.size() << " BAT entries\n"; 301 return errorToErrorCode(std::move(Err)); 302 } 303 304 template <bool Cold> 305 void BoltAddressTranslation::parseMaps(std::vector<uint64_t> &HotFuncs, 306 uint64_t &PrevAddress, DataExtractor &DE, 307 uint64_t &Offset, Error &Err) { 308 const uint32_t NumFunctions = DE.getULEB128(&Offset, &Err); 309 LLVM_DEBUG(dbgs() << "Parsing " << NumFunctions << (Cold ? " cold" : "") 310 << " functions\n"); 311 size_t HotIndex = 0; 312 for (uint32_t I = 0; I < NumFunctions; ++I) { 313 const uint64_t Address = PrevAddress + DE.getULEB128(&Offset, &Err); 314 uint64_t HotAddress = Cold ? 0 : Address; 315 PrevAddress = Address; 316 uint32_t SecondaryEntryPoints = 0; 317 if (Cold) { 318 HotIndex += DE.getULEB128(&Offset, &Err); 319 HotAddress = HotFuncs[HotIndex]; 320 ColdPartSource.emplace(Address, HotAddress); 321 } else { 322 HotFuncs.push_back(Address); 323 // Function hash 324 const size_t FuncHash = DE.getU64(&Offset, &Err); 325 FuncHashes.addEntry(Address, FuncHash); 326 LLVM_DEBUG(dbgs() << formatv("{0:x}: hash {1:x}\n", Address, FuncHash)); 327 // Number of basic blocks 328 const size_t NumBasicBlocks = DE.getULEB128(&Offset, &Err); 329 NumBasicBlocksMap.emplace(Address, NumBasicBlocks); 330 LLVM_DEBUG(dbgs() << formatv("{0:x}: #bbs {1}, {2} bytes\n", Address, 331 NumBasicBlocks, 332 getULEB128Size(NumBasicBlocks))); 333 // Secondary entry points 334 SecondaryEntryPoints = DE.getULEB128(&Offset, &Err); 335 LLVM_DEBUG( 336 dbgs() << formatv("{0:x}: secondary entry points {1}, {2} bytes\n", 337 Address, SecondaryEntryPoints, 338 getULEB128Size(SecondaryEntryPoints))); 339 } 340 const uint32_t NumEntries = DE.getULEB128(&Offset, &Err); 341 // Equal offsets, hot fragments only. 342 size_t EqualElems = 0; 343 APInt BEBitMask; 344 if (!Cold) { 345 EqualElems = DE.getULEB128(&Offset, &Err); 346 LLVM_DEBUG(dbgs() << formatv("Equal offsets: {0}, {1} bytes\n", 347 EqualElems, getULEB128Size(EqualElems))); 348 if (EqualElems) { 349 const size_t BranchEntriesBytes = alignTo(EqualElems, 8) / 8; 350 BEBitMask = APInt(alignTo(EqualElems, 8), 0); 351 LoadIntFromMemory( 352 BEBitMask, 353 reinterpret_cast<const uint8_t *>( 354 DE.getBytes(&Offset, BranchEntriesBytes, &Err).data()), 355 BranchEntriesBytes); 356 LLVM_DEBUG({ 357 dbgs() << "BEBitMask: "; 358 SmallString<8> BitMaskStr; 359 BEBitMask.toString(BitMaskStr, 2, false); 360 dbgs() << BitMaskStr << ", " << BranchEntriesBytes << " bytes\n"; 361 }); 362 } 363 } 364 MapTy Map; 365 366 LLVM_DEBUG(dbgs() << "Parsing " << NumEntries << " entries for 0x" 367 << Twine::utohexstr(Address) << "\n"); 368 uint64_t InputOffset = 0; 369 size_t BBIndex = 0; 370 for (uint32_t J = 0; J < NumEntries; ++J) { 371 const uint64_t OutputDelta = DE.getULEB128(&Offset, &Err); 372 const uint64_t OutputAddress = PrevAddress + OutputDelta; 373 const uint64_t OutputOffset = OutputAddress - Address; 374 PrevAddress = OutputAddress; 375 int64_t InputDelta = 0; 376 if (J < EqualElems) { 377 InputOffset = (OutputOffset << 1) | BEBitMask[J]; 378 } else { 379 InputDelta = DE.getSLEB128(&Offset, &Err); 380 InputOffset += InputDelta; 381 } 382 Map.insert(std::pair<uint32_t, uint32_t>(OutputOffset, InputOffset)); 383 size_t BBHash = 0; 384 size_t BBIndexDelta = 0; 385 const bool IsBranchEntry = InputOffset & BRANCHENTRY; 386 if (!IsBranchEntry) { 387 BBHash = DE.getU64(&Offset, &Err); 388 BBIndexDelta = DE.getULEB128(&Offset, &Err); 389 BBIndex += BBIndexDelta; 390 // Map basic block hash to hot fragment by input offset 391 getBBHashMap(HotAddress).addEntry(InputOffset >> 1, BBIndex, BBHash); 392 } 393 LLVM_DEBUG({ 394 dbgs() << formatv( 395 "{0:x} -> {1:x} ({2}/{3}b -> {4}/{5}b), {6:x}", OutputOffset, 396 InputOffset, OutputDelta, getULEB128Size(OutputDelta), InputDelta, 397 (J < EqualElems) ? 0 : getSLEB128Size(InputDelta), OutputAddress); 398 if (!IsBranchEntry) { 399 dbgs() << formatv(" {0:x} {1}/{2}b", BBHash, BBIndex, 400 getULEB128Size(BBIndexDelta)); 401 } 402 dbgs() << '\n'; 403 }); 404 } 405 Maps.insert(std::pair<uint64_t, MapTy>(Address, Map)); 406 if (!Cold && SecondaryEntryPoints) { 407 uint32_t EntryPointOffset = 0; 408 LLVM_DEBUG(dbgs() << "Secondary entry points: "); 409 for (uint32_t EntryPointId = 0; EntryPointId != SecondaryEntryPoints; 410 ++EntryPointId) { 411 uint32_t OffsetDelta = DE.getULEB128(&Offset, &Err); 412 EntryPointOffset += OffsetDelta; 413 SecondaryEntryPointsMap[Address].push_back(EntryPointOffset); 414 LLVM_DEBUG(dbgs() << formatv("{0:x}/{1}b ", EntryPointOffset, 415 getULEB128Size(OffsetDelta))); 416 } 417 LLVM_DEBUG(dbgs() << '\n'); 418 } 419 } 420 } 421 422 void BoltAddressTranslation::dump(raw_ostream &OS) { 423 const size_t NumTables = Maps.size(); 424 OS << "BAT tables for " << NumTables << " functions:\n"; 425 for (const auto &MapEntry : Maps) { 426 const uint64_t Address = MapEntry.first; 427 const uint64_t HotAddress = fetchParentAddress(Address); 428 OS << "Function Address: 0x" << Twine::utohexstr(Address); 429 if (HotAddress == 0) 430 OS << formatv(", hash: {0:x}", getBFHash(Address)); 431 OS << "\n"; 432 OS << "BB mappings:\n"; 433 const BBHashMapTy &BBHashMap = 434 getBBHashMap(HotAddress ? HotAddress : Address); 435 for (const auto &Entry : MapEntry.second) { 436 const bool IsBranch = Entry.second & BRANCHENTRY; 437 const uint32_t Val = Entry.second >> 1; // dropping BRANCHENTRY bit 438 OS << "0x" << Twine::utohexstr(Entry.first) << " -> " 439 << "0x" << Twine::utohexstr(Val); 440 if (IsBranch) 441 OS << " (branch)"; 442 else 443 OS << formatv(" hash: {0:x}", BBHashMap.getBBHash(Val)); 444 OS << "\n"; 445 } 446 if (SecondaryEntryPointsMap.count(Address)) { 447 const std::vector<uint32_t> &SecondaryEntryPoints = 448 SecondaryEntryPointsMap[Address]; 449 OS << SecondaryEntryPoints.size() << " secondary entry points:\n"; 450 for (uint32_t EntryPointOffset : SecondaryEntryPoints) 451 OS << formatv("{0:x}\n", EntryPointOffset); 452 } 453 OS << "\n"; 454 } 455 const size_t NumColdParts = ColdPartSource.size(); 456 if (!NumColdParts) 457 return; 458 459 OS << NumColdParts << " cold mappings:\n"; 460 for (const auto &Entry : ColdPartSource) { 461 OS << "0x" << Twine::utohexstr(Entry.first) << " -> " 462 << Twine::utohexstr(Entry.second) << "\n"; 463 } 464 OS << "\n"; 465 } 466 467 uint64_t BoltAddressTranslation::translate(uint64_t FuncAddress, 468 uint64_t Offset, 469 bool IsBranchSrc) const { 470 auto Iter = Maps.find(FuncAddress); 471 if (Iter == Maps.end()) 472 return Offset; 473 474 const MapTy &Map = Iter->second; 475 auto KeyVal = Map.upper_bound(Offset); 476 if (KeyVal == Map.begin()) 477 return Offset; 478 479 --KeyVal; 480 481 const uint32_t Val = KeyVal->second >> 1; // dropping BRANCHENTRY bit 482 // Branch source addresses are translated to the first instruction of the 483 // source BB to avoid accounting for modifications BOLT may have made in the 484 // BB regarding deletion/addition of instructions. 485 if (IsBranchSrc) 486 return Val; 487 return Offset - KeyVal->first + Val; 488 } 489 490 std::optional<BoltAddressTranslation::FallthroughListTy> 491 BoltAddressTranslation::getFallthroughsInTrace(uint64_t FuncAddress, 492 uint64_t From, 493 uint64_t To) const { 494 SmallVector<std::pair<uint64_t, uint64_t>, 16> Res; 495 496 // Filter out trivial case 497 if (From >= To) 498 return Res; 499 500 From -= FuncAddress; 501 To -= FuncAddress; 502 503 auto Iter = Maps.find(FuncAddress); 504 if (Iter == Maps.end()) 505 return std::nullopt; 506 507 const MapTy &Map = Iter->second; 508 auto FromIter = Map.upper_bound(From); 509 if (FromIter == Map.begin()) 510 return Res; 511 // Skip instruction entries, to create fallthroughs we are only interested in 512 // BB boundaries 513 do { 514 if (FromIter == Map.begin()) 515 return Res; 516 --FromIter; 517 } while (FromIter->second & BRANCHENTRY); 518 519 auto ToIter = Map.upper_bound(To); 520 if (ToIter == Map.begin()) 521 return Res; 522 --ToIter; 523 if (FromIter->first >= ToIter->first) 524 return Res; 525 526 for (auto Iter = FromIter; Iter != ToIter;) { 527 const uint32_t Src = Iter->first; 528 if (Iter->second & BRANCHENTRY) { 529 ++Iter; 530 continue; 531 } 532 533 ++Iter; 534 while (Iter->second & BRANCHENTRY && Iter != ToIter) 535 ++Iter; 536 if (Iter->second & BRANCHENTRY) 537 break; 538 Res.emplace_back(Src, Iter->first); 539 } 540 541 return Res; 542 } 543 544 uint64_t BoltAddressTranslation::fetchParentAddress(uint64_t Address) const { 545 auto Iter = ColdPartSource.find(Address); 546 if (Iter == ColdPartSource.end()) 547 return 0; 548 return Iter->second; 549 } 550 551 bool BoltAddressTranslation::enabledFor( 552 llvm::object::ELFObjectFileBase *InputFile) const { 553 for (const SectionRef &Section : InputFile->sections()) { 554 Expected<StringRef> SectionNameOrErr = Section.getName(); 555 if (Error E = SectionNameOrErr.takeError()) 556 continue; 557 558 if (SectionNameOrErr.get() == SECTION_NAME) 559 return true; 560 } 561 return false; 562 } 563 564 void BoltAddressTranslation::saveMetadata(BinaryContext &BC) { 565 for (BinaryFunction &BF : llvm::make_second_range(BC.getBinaryFunctions())) { 566 // We don't need a translation table if the body of the function hasn't 567 // changed 568 if (BF.isIgnored() || (!BC.HasRelocations && !BF.isSimple())) 569 continue; 570 // Prepare function and block hashes 571 FuncHashes.addEntry(BF.getAddress(), BF.computeHash()); 572 BF.computeBlockHashes(); 573 BBHashMapTy &BBHashMap = getBBHashMap(BF.getAddress()); 574 // Set BF/BB metadata 575 for (const BinaryBasicBlock &BB : BF) 576 BBHashMap.addEntry(BB.getInputOffset(), BB.getIndex(), BB.getHash()); 577 } 578 } 579 580 std::unordered_map<uint32_t, std::vector<uint32_t>> 581 BoltAddressTranslation::getBFBranches(uint64_t OutputAddress) const { 582 std::unordered_map<uint32_t, std::vector<uint32_t>> Branches; 583 auto FuncIt = Maps.find(OutputAddress); 584 assert(FuncIt != Maps.end()); 585 std::vector<uint32_t> InputOffsets; 586 for (const auto &KV : FuncIt->second) 587 InputOffsets.emplace_back(KV.second); 588 // Sort with LSB BRANCHENTRY bit. 589 llvm::sort(InputOffsets); 590 uint32_t BBOffset{0}; 591 for (uint32_t InOffset : InputOffsets) { 592 if (InOffset & BRANCHENTRY) 593 Branches[BBOffset].push_back(InOffset >> 1); 594 else 595 BBOffset = InOffset >> 1; 596 } 597 return Branches; 598 } 599 600 unsigned 601 BoltAddressTranslation::getSecondaryEntryPointId(uint64_t Address, 602 uint32_t Offset) const { 603 auto FunctionIt = SecondaryEntryPointsMap.find(Address); 604 if (FunctionIt == SecondaryEntryPointsMap.end()) 605 return 0; 606 const std::vector<uint32_t> &Offsets = FunctionIt->second; 607 auto OffsetIt = std::find(Offsets.begin(), Offsets.end(), Offset); 608 if (OffsetIt == Offsets.end()) 609 return 0; 610 // Adding one here because main entry point is not stored in BAT, and 611 // enumeration for secondary entry points starts with 1. 612 return OffsetIt - Offsets.begin() + 1; 613 } 614 615 std::pair<const BinaryFunction *, unsigned> 616 BoltAddressTranslation::translateSymbol(const BinaryContext &BC, 617 const MCSymbol &Symbol, 618 uint32_t Offset) const { 619 // The symbol could be a secondary entry in a cold fragment. 620 uint64_t SymbolValue = cantFail(errorOrToExpected(BC.getSymbolValue(Symbol))); 621 622 const BinaryFunction *Callee = BC.getFunctionForSymbol(&Symbol); 623 assert(Callee); 624 625 // Containing function, not necessarily the same as symbol value. 626 const uint64_t CalleeAddress = Callee->getAddress(); 627 const uint32_t OutputOffset = SymbolValue - CalleeAddress; 628 629 const uint64_t ParentAddress = fetchParentAddress(CalleeAddress); 630 const uint64_t HotAddress = ParentAddress ? ParentAddress : CalleeAddress; 631 632 const BinaryFunction *ParentBF = BC.getBinaryFunctionAtAddress(HotAddress); 633 634 const uint32_t InputOffset = 635 translate(CalleeAddress, OutputOffset, /*IsBranchSrc*/ false) + Offset; 636 637 unsigned SecondaryEntryId{0}; 638 if (InputOffset) 639 SecondaryEntryId = getSecondaryEntryPointId(HotAddress, InputOffset); 640 641 return std::pair(ParentBF, SecondaryEntryId); 642 } 643 644 } // namespace bolt 645 } // namespace llvm 646