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/Support/DataExtractor.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 FuncAddress) { 26 const uint64_t BBOutputOffset = 27 BB.getOutputAddressRange().first - FuncAddress; 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 // In case of conflicts (same Key mapping to different Vals), the last 43 // update takes precedence. Of course it is not ideal to have conflicts and 44 // those happen when we have an empty BB that either contained only 45 // NOPs or a jump to the next block (successor). Either way, the successor 46 // and this deleted block will both share the same output address (the same 47 // key), and we need to map back. We choose here to privilege the successor by 48 // allowing it to overwrite the previously inserted key in the map. 49 Map[BBOutputOffset] = BBInputOffset << 1; 50 51 const auto &IOAddressMap = 52 BB.getFunction()->getBinaryContext().getIOAddressMap(); 53 54 for (const auto &[InputOffset, Sym] : BB.getLocSyms()) { 55 const auto InputAddress = BB.getFunction()->getAddress() + InputOffset; 56 const auto OutputAddress = IOAddressMap.lookup(InputAddress); 57 assert(OutputAddress && "Unknown instruction address"); 58 const auto OutputOffset = *OutputAddress - FuncAddress; 59 60 // Is this the first instruction in the BB? No need to duplicate the entry. 61 if (OutputOffset == BBOutputOffset) 62 continue; 63 64 LLVM_DEBUG(dbgs() << " Key: " << Twine::utohexstr(OutputOffset) << " Val: " 65 << Twine::utohexstr(InputOffset) << " (branch)\n"); 66 Map.insert(std::pair<uint32_t, uint32_t>(OutputOffset, 67 (InputOffset << 1) | BRANCHENTRY)); 68 } 69 } 70 71 void BoltAddressTranslation::write(const BinaryContext &BC, raw_ostream &OS) { 72 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: Writing BOLT Address Translation Tables\n"); 73 for (auto &BFI : BC.getBinaryFunctions()) { 74 const BinaryFunction &Function = BFI.second; 75 // We don't need a translation table if the body of the function hasn't 76 // changed 77 if (Function.isIgnored() || (!BC.HasRelocations && !Function.isSimple())) 78 continue; 79 80 LLVM_DEBUG(dbgs() << "Function name: " << Function.getPrintName() << "\n"); 81 LLVM_DEBUG(dbgs() << " Address reference: 0x" 82 << Twine::utohexstr(Function.getOutputAddress()) << "\n"); 83 84 MapTy Map; 85 for (const BinaryBasicBlock *const BB : 86 Function.getLayout().getMainFragment()) 87 writeEntriesForBB(Map, *BB, Function.getOutputAddress()); 88 Maps.emplace(Function.getOutputAddress(), std::move(Map)); 89 90 if (!Function.isSplit()) 91 continue; 92 93 // Split maps 94 LLVM_DEBUG(dbgs() << " Cold part\n"); 95 for (const FunctionFragment &FF : 96 Function.getLayout().getSplitFragments()) { 97 Map.clear(); 98 for (const BinaryBasicBlock *const BB : FF) 99 writeEntriesForBB(Map, *BB, FF.getAddress()); 100 101 Maps.emplace(FF.getAddress(), std::move(Map)); 102 ColdPartSource.emplace(FF.getAddress(), Function.getOutputAddress()); 103 } 104 } 105 106 const uint32_t NumFuncs = Maps.size(); 107 encodeULEB128(NumFuncs, OS); 108 LLVM_DEBUG(dbgs() << "Writing " << NumFuncs << " functions for BAT.\n"); 109 for (auto &MapEntry : Maps) { 110 const uint64_t Address = MapEntry.first; 111 MapTy &Map = MapEntry.second; 112 const uint32_t NumEntries = Map.size(); 113 LLVM_DEBUG(dbgs() << "Writing " << NumEntries << " entries for 0x" 114 << Twine::utohexstr(Address) << ".\n"); 115 encodeULEB128(Address, OS); 116 encodeULEB128(NumEntries, OS); 117 uint64_t InOffset = 0, OutOffset = 0; 118 // Output and Input addresses and delta-encoded 119 for (std::pair<const uint32_t, uint32_t> &KeyVal : Map) { 120 encodeULEB128(KeyVal.first - OutOffset, OS); 121 encodeSLEB128(KeyVal.second - InOffset, OS); 122 std::tie(OutOffset, InOffset) = KeyVal; 123 } 124 } 125 const uint32_t NumColdEntries = ColdPartSource.size(); 126 LLVM_DEBUG(dbgs() << "Writing " << NumColdEntries 127 << " cold part mappings.\n"); 128 encodeULEB128(NumColdEntries, OS); 129 for (std::pair<const uint64_t, uint64_t> &ColdEntry : ColdPartSource) { 130 encodeULEB128(ColdEntry.first, OS); 131 encodeULEB128(ColdEntry.second, OS); 132 LLVM_DEBUG(dbgs() << " " << Twine::utohexstr(ColdEntry.first) << " -> " 133 << Twine::utohexstr(ColdEntry.second) << "\n"); 134 } 135 136 outs() << "BOLT-INFO: Wrote " << Maps.size() << " BAT maps\n"; 137 outs() << "BOLT-INFO: Wrote " << NumColdEntries 138 << " BAT cold-to-hot entries\n"; 139 } 140 141 std::error_code BoltAddressTranslation::parse(StringRef Buf) { 142 DataExtractor DE = DataExtractor(Buf, true, 8); 143 uint64_t Offset = 0; 144 if (Buf.size() < 12) 145 return make_error_code(llvm::errc::io_error); 146 147 const uint32_t NameSz = DE.getU32(&Offset); 148 const uint32_t DescSz = DE.getU32(&Offset); 149 const uint32_t Type = DE.getU32(&Offset); 150 151 if (Type != BinarySection::NT_BOLT_BAT || 152 Buf.size() + Offset < alignTo(NameSz, 4) + DescSz) 153 return make_error_code(llvm::errc::io_error); 154 155 StringRef Name = Buf.slice(Offset, Offset + NameSz); 156 Offset = alignTo(Offset + NameSz, 4); 157 if (Name.substr(0, 4) != "BOLT") 158 return make_error_code(llvm::errc::io_error); 159 160 Error Err(Error::success()); 161 const uint32_t NumFunctions = DE.getULEB128(&Offset, &Err); 162 LLVM_DEBUG(dbgs() << "Parsing " << NumFunctions << " functions\n"); 163 for (uint32_t I = 0; I < NumFunctions; ++I) { 164 const uint64_t Address = DE.getULEB128(&Offset, &Err); 165 const uint32_t NumEntries = DE.getULEB128(&Offset, &Err); 166 MapTy Map; 167 168 LLVM_DEBUG(dbgs() << "Parsing " << NumEntries << " entries for 0x" 169 << Twine::utohexstr(Address) << "\n"); 170 uint64_t InputOffset = 0, OutputOffset = 0; 171 for (uint32_t J = 0; J < NumEntries; ++J) { 172 const uint64_t OutputDelta = DE.getULEB128(&Offset, &Err); 173 const int64_t InputDelta = DE.getSLEB128(&Offset, &Err); 174 OutputOffset += OutputDelta; 175 InputOffset += InputDelta; 176 Map.insert(std::pair<uint32_t, uint32_t>(OutputOffset, InputOffset)); 177 LLVM_DEBUG(dbgs() << Twine::utohexstr(OutputOffset) << " -> " 178 << Twine::utohexstr(InputOffset) << " (" << OutputDelta 179 << ", " << InputDelta << ")\n"); 180 } 181 Maps.insert(std::pair<uint64_t, MapTy>(Address, Map)); 182 } 183 184 const uint32_t NumColdEntries = DE.getULEB128(&Offset, &Err); 185 LLVM_DEBUG(dbgs() << "Parsing " << NumColdEntries << " cold part mappings\n"); 186 for (uint32_t I = 0; I < NumColdEntries; ++I) { 187 const uint32_t ColdAddress = DE.getULEB128(&Offset, &Err); 188 const uint32_t HotAddress = DE.getULEB128(&Offset, &Err); 189 ColdPartSource.insert( 190 std::pair<uint64_t, uint64_t>(ColdAddress, HotAddress)); 191 LLVM_DEBUG(dbgs() << Twine::utohexstr(ColdAddress) << " -> " 192 << Twine::utohexstr(HotAddress) << "\n"); 193 } 194 outs() << "BOLT-INFO: Parsed " << Maps.size() << " BAT entries\n"; 195 outs() << "BOLT-INFO: Parsed " << NumColdEntries 196 << " BAT cold-to-hot entries\n"; 197 198 return errorToErrorCode(std::move(Err)); 199 } 200 201 void BoltAddressTranslation::dump(raw_ostream &OS) { 202 const size_t NumTables = Maps.size(); 203 OS << "BAT tables for " << NumTables << " functions:\n"; 204 for (const auto &MapEntry : Maps) { 205 OS << "Function Address: 0x" << Twine::utohexstr(MapEntry.first) << "\n"; 206 OS << "BB mappings:\n"; 207 for (const auto &Entry : MapEntry.second) { 208 const bool IsBranch = Entry.second & BRANCHENTRY; 209 const uint32_t Val = Entry.second >> 1; // dropping BRANCHENTRY bit 210 OS << "0x" << Twine::utohexstr(Entry.first) << " -> " 211 << "0x" << Twine::utohexstr(Val); 212 if (IsBranch) 213 OS << " (branch)"; 214 OS << "\n"; 215 } 216 OS << "\n"; 217 } 218 const size_t NumColdParts = ColdPartSource.size(); 219 if (!NumColdParts) 220 return; 221 222 OS << NumColdParts << " cold mappings:\n"; 223 for (const auto &Entry : ColdPartSource) { 224 OS << "0x" << Twine::utohexstr(Entry.first) << " -> " 225 << Twine::utohexstr(Entry.second) << "\n"; 226 } 227 OS << "\n"; 228 } 229 230 uint64_t BoltAddressTranslation::translate(uint64_t FuncAddress, 231 uint64_t Offset, 232 bool IsBranchSrc) const { 233 auto Iter = Maps.find(FuncAddress); 234 if (Iter == Maps.end()) 235 return Offset; 236 237 const MapTy &Map = Iter->second; 238 auto KeyVal = Map.upper_bound(Offset); 239 if (KeyVal == Map.begin()) 240 return Offset; 241 242 --KeyVal; 243 244 const uint32_t Val = KeyVal->second >> 1; // dropping BRANCHENTRY bit 245 // Branch source addresses are translated to the first instruction of the 246 // source BB to avoid accounting for modifications BOLT may have made in the 247 // BB regarding deletion/addition of instructions. 248 if (IsBranchSrc) 249 return Val; 250 return Offset - KeyVal->first + Val; 251 } 252 253 std::optional<BoltAddressTranslation::FallthroughListTy> 254 BoltAddressTranslation::getFallthroughsInTrace(uint64_t FuncAddress, 255 uint64_t From, 256 uint64_t To) const { 257 SmallVector<std::pair<uint64_t, uint64_t>, 16> Res; 258 259 // Filter out trivial case 260 if (From >= To) 261 return Res; 262 263 From -= FuncAddress; 264 To -= FuncAddress; 265 266 auto Iter = Maps.find(FuncAddress); 267 if (Iter == Maps.end()) 268 return std::nullopt; 269 270 const MapTy &Map = Iter->second; 271 auto FromIter = Map.upper_bound(From); 272 if (FromIter == Map.begin()) 273 return Res; 274 // Skip instruction entries, to create fallthroughs we are only interested in 275 // BB boundaries 276 do { 277 if (FromIter == Map.begin()) 278 return Res; 279 --FromIter; 280 } while (FromIter->second & BRANCHENTRY); 281 282 auto ToIter = Map.upper_bound(To); 283 if (ToIter == Map.begin()) 284 return Res; 285 --ToIter; 286 if (FromIter->first >= ToIter->first) 287 return Res; 288 289 for (auto Iter = FromIter; Iter != ToIter;) { 290 const uint32_t Src = Iter->first; 291 if (Iter->second & BRANCHENTRY) { 292 ++Iter; 293 continue; 294 } 295 296 ++Iter; 297 while (Iter->second & BRANCHENTRY && Iter != ToIter) 298 ++Iter; 299 if (Iter->second & BRANCHENTRY) 300 break; 301 Res.emplace_back(Src, Iter->first); 302 } 303 304 return Res; 305 } 306 307 uint64_t BoltAddressTranslation::fetchParentAddress(uint64_t Address) const { 308 auto Iter = ColdPartSource.find(Address); 309 if (Iter == ColdPartSource.end()) 310 return 0; 311 return Iter->second; 312 } 313 314 bool BoltAddressTranslation::enabledFor( 315 llvm::object::ELFObjectFileBase *InputFile) const { 316 for (const SectionRef &Section : InputFile->sections()) { 317 Expected<StringRef> SectionNameOrErr = Section.getName(); 318 if (Error E = SectionNameOrErr.takeError()) 319 continue; 320 321 if (SectionNameOrErr.get() == SECTION_NAME) 322 return true; 323 } 324 return false; 325 } 326 } // namespace bolt 327 } // namespace llvm 328