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