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 for (std::pair<const uint32_t, uint32_t> &KeyVal : Map) { 118 encodeULEB128(KeyVal.first, OS); 119 encodeULEB128(KeyVal.second, OS); 120 } 121 } 122 const uint32_t NumColdEntries = ColdPartSource.size(); 123 LLVM_DEBUG(dbgs() << "Writing " << NumColdEntries 124 << " cold part mappings.\n"); 125 encodeULEB128(NumColdEntries, OS); 126 for (std::pair<const uint64_t, uint64_t> &ColdEntry : ColdPartSource) { 127 encodeULEB128(ColdEntry.first, OS); 128 encodeULEB128(ColdEntry.second, OS); 129 LLVM_DEBUG(dbgs() << " " << Twine::utohexstr(ColdEntry.first) << " -> " 130 << Twine::utohexstr(ColdEntry.second) << "\n"); 131 } 132 133 outs() << "BOLT-INFO: Wrote " << Maps.size() << " BAT maps\n"; 134 outs() << "BOLT-INFO: Wrote " << NumColdEntries 135 << " BAT cold-to-hot entries\n"; 136 } 137 138 std::error_code BoltAddressTranslation::parse(StringRef Buf) { 139 DataExtractor DE = DataExtractor(Buf, true, 8); 140 uint64_t Offset = 0; 141 if (Buf.size() < 12) 142 return make_error_code(llvm::errc::io_error); 143 144 const uint32_t NameSz = DE.getU32(&Offset); 145 const uint32_t DescSz = DE.getU32(&Offset); 146 const uint32_t Type = DE.getU32(&Offset); 147 148 if (Type != BinarySection::NT_BOLT_BAT || 149 Buf.size() + Offset < alignTo(NameSz, 4) + DescSz) 150 return make_error_code(llvm::errc::io_error); 151 152 StringRef Name = Buf.slice(Offset, Offset + NameSz); 153 Offset = alignTo(Offset + NameSz, 4); 154 if (Name.substr(0, 4) != "BOLT") 155 return make_error_code(llvm::errc::io_error); 156 157 Error Err(Error::success()); 158 const uint32_t NumFunctions = DE.getULEB128(&Offset, &Err); 159 LLVM_DEBUG(dbgs() << "Parsing " << NumFunctions << " functions\n"); 160 for (uint32_t I = 0; I < NumFunctions; ++I) { 161 const uint64_t Address = DE.getULEB128(&Offset, &Err); 162 const uint32_t NumEntries = DE.getULEB128(&Offset, &Err); 163 MapTy Map; 164 165 LLVM_DEBUG(dbgs() << "Parsing " << NumEntries << " entries for 0x" 166 << Twine::utohexstr(Address) << "\n"); 167 for (uint32_t J = 0; J < NumEntries; ++J) { 168 const uint32_t OutputAddr = DE.getULEB128(&Offset, &Err); 169 const uint32_t InputAddr = DE.getULEB128(&Offset, &Err); 170 Map.insert(std::pair<uint32_t, uint32_t>(OutputAddr, InputAddr)); 171 LLVM_DEBUG(dbgs() << Twine::utohexstr(OutputAddr) << " -> " 172 << Twine::utohexstr(InputAddr) << "\n"); 173 } 174 Maps.insert(std::pair<uint64_t, MapTy>(Address, Map)); 175 } 176 177 const uint32_t NumColdEntries = DE.getULEB128(&Offset, &Err); 178 LLVM_DEBUG(dbgs() << "Parsing " << NumColdEntries << " cold part mappings\n"); 179 for (uint32_t I = 0; I < NumColdEntries; ++I) { 180 const uint32_t ColdAddress = DE.getULEB128(&Offset, &Err); 181 const uint32_t HotAddress = DE.getULEB128(&Offset, &Err); 182 ColdPartSource.insert( 183 std::pair<uint64_t, uint64_t>(ColdAddress, HotAddress)); 184 LLVM_DEBUG(dbgs() << Twine::utohexstr(ColdAddress) << " -> " 185 << Twine::utohexstr(HotAddress) << "\n"); 186 } 187 outs() << "BOLT-INFO: Parsed " << Maps.size() << " BAT entries\n"; 188 outs() << "BOLT-INFO: Parsed " << NumColdEntries 189 << " BAT cold-to-hot entries\n"; 190 191 return errorToErrorCode(std::move(Err)); 192 } 193 194 void BoltAddressTranslation::dump(raw_ostream &OS) { 195 const size_t NumTables = Maps.size(); 196 OS << "BAT tables for " << NumTables << " functions:\n"; 197 for (const auto &MapEntry : Maps) { 198 OS << "Function Address: 0x" << Twine::utohexstr(MapEntry.first) << "\n"; 199 OS << "BB mappings:\n"; 200 for (const auto &Entry : MapEntry.second) { 201 const bool IsBranch = Entry.second & BRANCHENTRY; 202 const uint32_t Val = Entry.second >> 1; // dropping BRANCHENTRY bit 203 OS << "0x" << Twine::utohexstr(Entry.first) << " -> " 204 << "0x" << Twine::utohexstr(Val); 205 if (IsBranch) 206 OS << " (branch)"; 207 OS << "\n"; 208 } 209 OS << "\n"; 210 } 211 const size_t NumColdParts = ColdPartSource.size(); 212 if (!NumColdParts) 213 return; 214 215 OS << NumColdParts << " cold mappings:\n"; 216 for (const auto &Entry : ColdPartSource) { 217 OS << "0x" << Twine::utohexstr(Entry.first) << " -> " 218 << Twine::utohexstr(Entry.second) << "\n"; 219 } 220 OS << "\n"; 221 } 222 223 uint64_t BoltAddressTranslation::translate(uint64_t FuncAddress, 224 uint64_t Offset, 225 bool IsBranchSrc) const { 226 auto Iter = Maps.find(FuncAddress); 227 if (Iter == Maps.end()) 228 return Offset; 229 230 const MapTy &Map = Iter->second; 231 auto KeyVal = Map.upper_bound(Offset); 232 if (KeyVal == Map.begin()) 233 return Offset; 234 235 --KeyVal; 236 237 const uint32_t Val = KeyVal->second >> 1; // dropping BRANCHENTRY bit 238 // Branch source addresses are translated to the first instruction of the 239 // source BB to avoid accounting for modifications BOLT may have made in the 240 // BB regarding deletion/addition of instructions. 241 if (IsBranchSrc) 242 return Val; 243 return Offset - KeyVal->first + Val; 244 } 245 246 std::optional<BoltAddressTranslation::FallthroughListTy> 247 BoltAddressTranslation::getFallthroughsInTrace(uint64_t FuncAddress, 248 uint64_t From, 249 uint64_t To) const { 250 SmallVector<std::pair<uint64_t, uint64_t>, 16> Res; 251 252 // Filter out trivial case 253 if (From >= To) 254 return Res; 255 256 From -= FuncAddress; 257 To -= FuncAddress; 258 259 auto Iter = Maps.find(FuncAddress); 260 if (Iter == Maps.end()) 261 return std::nullopt; 262 263 const MapTy &Map = Iter->second; 264 auto FromIter = Map.upper_bound(From); 265 if (FromIter == Map.begin()) 266 return Res; 267 // Skip instruction entries, to create fallthroughs we are only interested in 268 // BB boundaries 269 do { 270 if (FromIter == Map.begin()) 271 return Res; 272 --FromIter; 273 } while (FromIter->second & BRANCHENTRY); 274 275 auto ToIter = Map.upper_bound(To); 276 if (ToIter == Map.begin()) 277 return Res; 278 --ToIter; 279 if (FromIter->first >= ToIter->first) 280 return Res; 281 282 for (auto Iter = FromIter; Iter != ToIter;) { 283 const uint32_t Src = Iter->first; 284 if (Iter->second & BRANCHENTRY) { 285 ++Iter; 286 continue; 287 } 288 289 ++Iter; 290 while (Iter->second & BRANCHENTRY && Iter != ToIter) 291 ++Iter; 292 if (Iter->second & BRANCHENTRY) 293 break; 294 Res.emplace_back(Src, Iter->first); 295 } 296 297 return Res; 298 } 299 300 uint64_t BoltAddressTranslation::fetchParentAddress(uint64_t Address) const { 301 auto Iter = ColdPartSource.find(Address); 302 if (Iter == ColdPartSource.end()) 303 return 0; 304 return Iter->second; 305 } 306 307 bool BoltAddressTranslation::enabledFor( 308 llvm::object::ELFObjectFileBase *InputFile) const { 309 for (const SectionRef &Section : InputFile->sections()) { 310 Expected<StringRef> SectionNameOrErr = Section.getName(); 311 if (Error E = SectionNameOrErr.takeError()) 312 continue; 313 314 if (SectionNameOrErr.get() == SECTION_NAME) 315 return true; 316 } 317 return false; 318 } 319 } // namespace bolt 320 } // namespace llvm 321