1a34c753fSRafael Auler //===--- BoltAddressTranslation.cpp ---------------------------------------===// 2a34c753fSRafael Auler // 3a34c753fSRafael Auler // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4a34c753fSRafael Auler // See https://llvm.org/LICENSE.txt for license information. 5a34c753fSRafael Auler // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6a34c753fSRafael Auler // 7a34c753fSRafael Auler //===----------------------------------------------------------------------===// 8a34c753fSRafael Auler // 9a34c753fSRafael Auler //===----------------------------------------------------------------------===// 10a34c753fSRafael Auler #include "bolt/Profile/BoltAddressTranslation.h" 11a34c753fSRafael Auler #include "bolt/Core/BinaryFunction.h" 12a34c753fSRafael Auler #include "llvm/Support/DataExtractor.h" 13a34c753fSRafael Auler #include "llvm/Support/Errc.h" 14a34c753fSRafael Auler 15a34c753fSRafael Auler #define DEBUG_TYPE "bolt-bat" 16a34c753fSRafael Auler 17a34c753fSRafael Auler namespace llvm { 18a34c753fSRafael Auler namespace bolt { 19a34c753fSRafael Auler 20a34c753fSRafael Auler const char *BoltAddressTranslation::SECTION_NAME = ".note.bolt_bat"; 21a34c753fSRafael Auler 22a34c753fSRafael Auler void BoltAddressTranslation::writeEntriesForBB(MapTy &Map, 23a34c753fSRafael Auler const BinaryBasicBlock &BB, 24a34c753fSRafael Auler uint64_t FuncAddress) { 25a34c753fSRafael Auler const uint64_t BBOutputOffset = 26a34c753fSRafael Auler BB.getOutputAddressRange().first - FuncAddress; 27a34c753fSRafael Auler const uint32_t BBInputOffset = BB.getInputOffset(); 28a34c753fSRafael Auler 29a34c753fSRafael Auler assert(BBInputOffset != BinaryBasicBlock::INVALID_OFFSET && 30a34c753fSRafael Auler "Every output BB must track back to an input BB for profile " 31a34c753fSRafael Auler "collection in bolted binaries"); 32a34c753fSRafael Auler 33a34c753fSRafael Auler LLVM_DEBUG(dbgs() << "BB " << BB.getName() << "\n"); 34a34c753fSRafael Auler LLVM_DEBUG(dbgs() << " Key: " << Twine::utohexstr(BBOutputOffset) 35a34c753fSRafael Auler << " Val: " << Twine::utohexstr(BBInputOffset) << "\n"); 36a34c753fSRafael Auler // In case of conflicts (same Key mapping to different Vals), the last 37a34c753fSRafael Auler // update takes precedence. Of course it is not ideal to have conflicts and 38a34c753fSRafael Auler // those happen when we have an empty BB that either contained only 39a34c753fSRafael Auler // NOPs or a jump to the next block (successor). Either way, the successor 40a34c753fSRafael Auler // and this deleted block will both share the same output address (the same 41a34c753fSRafael Auler // key), and we need to map back. We choose here to privilege the successor by 42a34c753fSRafael Auler // allowing it to overwrite the previously inserted key in the map. 43a34c753fSRafael Auler Map[BBOutputOffset] = BBInputOffset; 44a34c753fSRafael Auler 45a34c753fSRafael Auler for (const auto &IOPair : BB.getOffsetTranslationTable()) { 46a34c753fSRafael Auler const uint64_t OutputOffset = IOPair.first + BBOutputOffset; 47a34c753fSRafael Auler const uint32_t InputOffset = IOPair.second; 48a34c753fSRafael Auler 49a34c753fSRafael Auler // Is this the first instruction in the BB? No need to duplicate the entry. 50a34c753fSRafael Auler if (OutputOffset == BBOutputOffset) 51a34c753fSRafael Auler continue; 52a34c753fSRafael Auler 53a34c753fSRafael Auler LLVM_DEBUG(dbgs() << " Key: " << Twine::utohexstr(OutputOffset) << " Val: " 54a34c753fSRafael Auler << Twine::utohexstr(InputOffset) << " (branch)\n"); 55a34c753fSRafael Auler Map.insert( 56a34c753fSRafael Auler std::pair<uint32_t, uint32_t>(OutputOffset, InputOffset | BRANCHENTRY)); 57a34c753fSRafael Auler } 58a34c753fSRafael Auler } 59a34c753fSRafael Auler 60a34c753fSRafael Auler void BoltAddressTranslation::write(raw_ostream &OS) { 61a34c753fSRafael Auler LLVM_DEBUG(dbgs() << "BOLT-DEBUG: Writing BOLT Address Translation Tables\n"); 62a34c753fSRafael Auler for (auto &BFI : BC.getBinaryFunctions()) { 63a34c753fSRafael Auler BinaryFunction &Function = BFI.second; 64a34c753fSRafael Auler // We don't need a translation table if the body of the function hasn't 65a34c753fSRafael Auler // changed 66a34c753fSRafael Auler if (!BC.HasRelocations && !Function.isSimple()) 67a34c753fSRafael Auler continue; 68a34c753fSRafael Auler 69a34c753fSRafael Auler LLVM_DEBUG(dbgs() << "Function name: " << Function.getPrintName() << "\n"); 70a34c753fSRafael Auler LLVM_DEBUG(dbgs() << " Address reference: 0x" 71a34c753fSRafael Auler << Twine::utohexstr(Function.getOutputAddress()) << "\n"); 72a34c753fSRafael Auler MapTy Map; 73a34c753fSRafael Auler const bool IsSplit = Function.isSplit(); 74a34c753fSRafael Auler for (BinaryBasicBlock *&BB : Function.layout()) { 75a34c753fSRafael Auler if (IsSplit && BB->isCold()) 76a34c753fSRafael Auler break; 77a34c753fSRafael Auler writeEntriesForBB(Map, *BB, Function.getOutputAddress()); 78a34c753fSRafael Auler } 79a34c753fSRafael Auler Maps.insert(std::pair<uint64_t, MapTy>(Function.getOutputAddress(), Map)); 80a34c753fSRafael Auler 81a34c753fSRafael Auler if (!IsSplit) 82a34c753fSRafael Auler continue; 83a34c753fSRafael Auler 84a34c753fSRafael Auler // Cold map 85a34c753fSRafael Auler Map.clear(); 86a34c753fSRafael Auler LLVM_DEBUG(dbgs() << " Cold part\n"); 87a34c753fSRafael Auler for (BinaryBasicBlock *&BB : Function.layout()) { 88a34c753fSRafael Auler if (!BB->isCold()) 89a34c753fSRafael Auler continue; 90a34c753fSRafael Auler writeEntriesForBB(Map, *BB, Function.cold().getAddress()); 91a34c753fSRafael Auler } 92a34c753fSRafael Auler Maps.insert(std::pair<uint64_t, MapTy>(Function.cold().getAddress(), Map)); 93a34c753fSRafael Auler ColdPartSource.insert(std::pair<uint64_t, uint64_t>( 94a34c753fSRafael Auler Function.cold().getAddress(), Function.getOutputAddress())); 95a34c753fSRafael Auler } 96a34c753fSRafael Auler 97a34c753fSRafael Auler const uint32_t NumFuncs = Maps.size(); 98a34c753fSRafael Auler OS.write(reinterpret_cast<const char *>(&NumFuncs), 4); 99a34c753fSRafael Auler LLVM_DEBUG(dbgs() << "Writing " << NumFuncs << " functions for BAT.\n"); 100a34c753fSRafael Auler for (auto &MapEntry : Maps) { 101a34c753fSRafael Auler const uint64_t Address = MapEntry.first; 102a34c753fSRafael Auler MapTy &Map = MapEntry.second; 103a34c753fSRafael Auler const uint32_t NumEntries = Map.size(); 104a34c753fSRafael Auler LLVM_DEBUG(dbgs() << "Writing " << NumEntries << " entries for 0x" 105a34c753fSRafael Auler << Twine::utohexstr(Address) << ".\n"); 106a34c753fSRafael Auler OS.write(reinterpret_cast<const char *>(&Address), 8); 107a34c753fSRafael Auler OS.write(reinterpret_cast<const char *>(&NumEntries), 4); 108a34c753fSRafael Auler for (std::pair<const uint32_t, uint32_t> &KeyVal : Map) { 109a34c753fSRafael Auler OS.write(reinterpret_cast<const char *>(&KeyVal.first), 4); 110a34c753fSRafael Auler OS.write(reinterpret_cast<const char *>(&KeyVal.second), 4); 111a34c753fSRafael Auler } 112a34c753fSRafael Auler } 113a34c753fSRafael Auler const uint32_t NumColdEntries = ColdPartSource.size(); 114a34c753fSRafael Auler LLVM_DEBUG(dbgs() << "Writing " << NumColdEntries 115a34c753fSRafael Auler << " cold part mappings.\n"); 116a34c753fSRafael Auler OS.write(reinterpret_cast<const char *>(&NumColdEntries), 4); 117a34c753fSRafael Auler for (std::pair<const uint64_t, uint64_t> &ColdEntry : ColdPartSource) { 118a34c753fSRafael Auler OS.write(reinterpret_cast<const char *>(&ColdEntry.first), 8); 119a34c753fSRafael Auler OS.write(reinterpret_cast<const char *>(&ColdEntry.second), 8); 120a34c753fSRafael Auler LLVM_DEBUG(dbgs() << " " << Twine::utohexstr(ColdEntry.first) << " -> " 121a34c753fSRafael Auler << Twine::utohexstr(ColdEntry.second) << "\n"); 122a34c753fSRafael Auler } 123a34c753fSRafael Auler 124a34c753fSRafael Auler outs() << "BOLT-INFO: Wrote " << Maps.size() << " BAT maps\n"; 125a34c753fSRafael Auler outs() << "BOLT-INFO: Wrote " << NumColdEntries 126a34c753fSRafael Auler << " BAT cold-to-hot entries\n"; 127a34c753fSRafael Auler } 128a34c753fSRafael Auler 129a34c753fSRafael Auler std::error_code BoltAddressTranslation::parse(StringRef Buf) { 130a34c753fSRafael Auler DataExtractor DE = DataExtractor(Buf, true, 8); 131a34c753fSRafael Auler uint64_t Offset = 0; 132a34c753fSRafael Auler if (Buf.size() < 12) 133a34c753fSRafael Auler return make_error_code(llvm::errc::io_error); 134a34c753fSRafael Auler 135a34c753fSRafael Auler const uint32_t NameSz = DE.getU32(&Offset); 136a34c753fSRafael Auler const uint32_t DescSz = DE.getU32(&Offset); 137a34c753fSRafael Auler const uint32_t Type = DE.getU32(&Offset); 138a34c753fSRafael Auler 139a34c753fSRafael Auler if (Type != BinarySection::NT_BOLT_BAT || 140a34c753fSRafael Auler Buf.size() + Offset < alignTo(NameSz, 4) + DescSz) 141a34c753fSRafael Auler return make_error_code(llvm::errc::io_error); 142a34c753fSRafael Auler 143a34c753fSRafael Auler StringRef Name = Buf.slice(Offset, Offset + NameSz); 144a34c753fSRafael Auler Offset = alignTo(Offset + NameSz, 4); 145a34c753fSRafael Auler if (Name.substr(0, 4) != "BOLT") 146a34c753fSRafael Auler return make_error_code(llvm::errc::io_error); 147a34c753fSRafael Auler 148a34c753fSRafael Auler if (Buf.size() - Offset < 4) 149a34c753fSRafael Auler return make_error_code(llvm::errc::io_error); 150a34c753fSRafael Auler 151a34c753fSRafael Auler const uint32_t NumFunctions = DE.getU32(&Offset); 152a34c753fSRafael Auler LLVM_DEBUG(dbgs() << "Parsing " << NumFunctions << " functions\n"); 153a34c753fSRafael Auler for (uint32_t I = 0; I < NumFunctions; ++I) { 154a34c753fSRafael Auler if (Buf.size() - Offset < 12) 155a34c753fSRafael Auler return make_error_code(llvm::errc::io_error); 156a34c753fSRafael Auler 157a34c753fSRafael Auler const uint64_t Address = DE.getU64(&Offset); 158a34c753fSRafael Auler const uint32_t NumEntries = DE.getU32(&Offset); 159a34c753fSRafael Auler MapTy Map; 160a34c753fSRafael Auler 161a34c753fSRafael Auler LLVM_DEBUG(dbgs() << "Parsing " << NumEntries << " entries for 0x" 162a34c753fSRafael Auler << Twine::utohexstr(Address) << "\n"); 163a34c753fSRafael Auler if (Buf.size() - Offset < 8 * NumEntries) 164a34c753fSRafael Auler return make_error_code(llvm::errc::io_error); 165a34c753fSRafael Auler for (uint32_t J = 0; J < NumEntries; ++J) { 166a34c753fSRafael Auler const uint32_t OutputAddr = DE.getU32(&Offset); 167a34c753fSRafael Auler const uint32_t InputAddr = DE.getU32(&Offset); 168a34c753fSRafael Auler Map.insert(std::pair<uint32_t, uint32_t>(OutputAddr, InputAddr)); 169a34c753fSRafael Auler LLVM_DEBUG(dbgs() << Twine::utohexstr(OutputAddr) << " -> " 170a34c753fSRafael Auler << Twine::utohexstr(InputAddr) << "\n"); 171a34c753fSRafael Auler } 172a34c753fSRafael Auler Maps.insert(std::pair<uint64_t, MapTy>(Address, Map)); 173a34c753fSRafael Auler } 174a34c753fSRafael Auler 175a34c753fSRafael Auler if (Buf.size() - Offset < 4) 176a34c753fSRafael Auler return make_error_code(llvm::errc::io_error); 177a34c753fSRafael Auler 178a34c753fSRafael Auler const uint32_t NumColdEntries = DE.getU32(&Offset); 179a34c753fSRafael Auler LLVM_DEBUG(dbgs() << "Parsing " << NumColdEntries << " cold part mappings\n"); 180a34c753fSRafael Auler for (uint32_t I = 0; I < NumColdEntries; ++I) { 181a34c753fSRafael Auler if (Buf.size() - Offset < 16) 182a34c753fSRafael Auler return make_error_code(llvm::errc::io_error); 183a34c753fSRafael Auler const uint32_t ColdAddress = DE.getU64(&Offset); 184a34c753fSRafael Auler const uint32_t HotAddress = DE.getU64(&Offset); 185a34c753fSRafael Auler ColdPartSource.insert( 186a34c753fSRafael Auler std::pair<uint64_t, uint64_t>(ColdAddress, HotAddress)); 187a34c753fSRafael Auler LLVM_DEBUG(dbgs() << Twine::utohexstr(ColdAddress) << " -> " 188a34c753fSRafael Auler << Twine::utohexstr(HotAddress) << "\n"); 189a34c753fSRafael Auler } 190a34c753fSRafael Auler outs() << "BOLT-INFO: Parsed " << Maps.size() << " BAT entries\n"; 191a34c753fSRafael Auler outs() << "BOLT-INFO: Parsed " << NumColdEntries 192a34c753fSRafael Auler << " BAT cold-to-hot entries\n"; 193a34c753fSRafael Auler 194a34c753fSRafael Auler return std::error_code(); 195a34c753fSRafael Auler } 196a34c753fSRafael Auler 197a34c753fSRafael Auler uint64_t BoltAddressTranslation::translate(const BinaryFunction &Func, 198a34c753fSRafael Auler uint64_t Offset, 199a34c753fSRafael Auler bool IsBranchSrc) const { 200a34c753fSRafael Auler auto Iter = Maps.find(Func.getAddress()); 201a34c753fSRafael Auler if (Iter == Maps.end()) 202a34c753fSRafael Auler return Offset; 203a34c753fSRafael Auler 204a34c753fSRafael Auler const MapTy &Map = Iter->second; 205a34c753fSRafael Auler auto KeyVal = Map.upper_bound(Offset); 206a34c753fSRafael Auler if (KeyVal == Map.begin()) 207a34c753fSRafael Auler return Offset; 208a34c753fSRafael Auler 209a34c753fSRafael Auler --KeyVal; 210a34c753fSRafael Auler 211a34c753fSRafael Auler const uint32_t Val = KeyVal->second & ~BRANCHENTRY; 212a34c753fSRafael Auler // Branch source addresses are translated to the first instruction of the 213a34c753fSRafael Auler // source BB to avoid accounting for modifications BOLT may have made in the 214a34c753fSRafael Auler // BB regarding deletion/addition of instructions. 215a34c753fSRafael Auler if (IsBranchSrc) 216a34c753fSRafael Auler return Val; 217a34c753fSRafael Auler return Offset - KeyVal->first + Val; 218a34c753fSRafael Auler } 219a34c753fSRafael Auler 220a34c753fSRafael Auler Optional<BoltAddressTranslation::FallthroughListTy> 221*40c2e0faSMaksim Panchenko BoltAddressTranslation::getFallthroughsInTrace(const BinaryFunction &Func, 222*40c2e0faSMaksim Panchenko uint64_t From, 223*40c2e0faSMaksim Panchenko uint64_t To) const { 224a34c753fSRafael Auler SmallVector<std::pair<uint64_t, uint64_t>, 16> Res; 225a34c753fSRafael Auler 226a34c753fSRafael Auler // Filter out trivial case 227a34c753fSRafael Auler if (From >= To) 228a34c753fSRafael Auler return Res; 229a34c753fSRafael Auler 230a34c753fSRafael Auler From -= Func.getAddress(); 231a34c753fSRafael Auler To -= Func.getAddress(); 232a34c753fSRafael Auler 233a34c753fSRafael Auler auto Iter = Maps.find(Func.getAddress()); 234a34c753fSRafael Auler if (Iter == Maps.end()) { 235a34c753fSRafael Auler return NoneType(); 236a34c753fSRafael Auler } 237a34c753fSRafael Auler 238a34c753fSRafael Auler const MapTy &Map = Iter->second; 239a34c753fSRafael Auler auto FromIter = Map.upper_bound(From); 240a34c753fSRafael Auler if (FromIter == Map.begin()) 241a34c753fSRafael Auler return Res; 242a34c753fSRafael Auler // Skip instruction entries, to create fallthroughs we are only interested in 243a34c753fSRafael Auler // BB boundaries 244a34c753fSRafael Auler do { 245a34c753fSRafael Auler if (FromIter == Map.begin()) 246a34c753fSRafael Auler return Res; 247a34c753fSRafael Auler --FromIter; 248a34c753fSRafael Auler } while (FromIter->second & BRANCHENTRY); 249a34c753fSRafael Auler 250a34c753fSRafael Auler auto ToIter = Map.upper_bound(To); 251a34c753fSRafael Auler if (ToIter == Map.begin()) 252a34c753fSRafael Auler return Res; 253a34c753fSRafael Auler --ToIter; 254a34c753fSRafael Auler if (FromIter->first >= ToIter->first) 255a34c753fSRafael Auler return Res; 256a34c753fSRafael Auler 257a34c753fSRafael Auler for (auto Iter = FromIter; Iter != ToIter;) { 258a34c753fSRafael Auler const uint32_t Src = Iter->first; 259a34c753fSRafael Auler if (Iter->second & BRANCHENTRY) { 260a34c753fSRafael Auler ++Iter; 261a34c753fSRafael Auler continue; 262a34c753fSRafael Auler } 263a34c753fSRafael Auler 264a34c753fSRafael Auler ++Iter; 265a34c753fSRafael Auler while (Iter->second & BRANCHENTRY && Iter != ToIter) { 266a34c753fSRafael Auler ++Iter; 267a34c753fSRafael Auler } 268a34c753fSRafael Auler if (Iter->second & BRANCHENTRY) 269a34c753fSRafael Auler break; 270a34c753fSRafael Auler Res.emplace_back(Src, Iter->first); 271a34c753fSRafael Auler } 272a34c753fSRafael Auler 273a34c753fSRafael Auler return Res; 274a34c753fSRafael Auler } 275a34c753fSRafael Auler 276a34c753fSRafael Auler uint64_t BoltAddressTranslation::fetchParentAddress(uint64_t Address) const { 277a34c753fSRafael Auler auto Iter = ColdPartSource.find(Address); 278a34c753fSRafael Auler if (Iter == ColdPartSource.end()) 279a34c753fSRafael Auler return 0; 280a34c753fSRafael Auler return Iter->second; 281a34c753fSRafael Auler } 282a34c753fSRafael Auler 283a34c753fSRafael Auler bool BoltAddressTranslation::enabledFor( 284a34c753fSRafael Auler llvm::object::ELFObjectFileBase *InputFile) const { 285a34c753fSRafael Auler for (const SectionRef &Section : InputFile->sections()) { 286a34c753fSRafael Auler Expected<StringRef> SectionNameOrErr = Section.getName(); 287a34c753fSRafael Auler if (Error E = SectionNameOrErr.takeError()) 288a34c753fSRafael Auler continue; 289a34c753fSRafael Auler 290a34c753fSRafael Auler if (SectionNameOrErr.get() == SECTION_NAME) 291a34c753fSRafael Auler return true; 292a34c753fSRafael Auler } 293a34c753fSRafael Auler return false; 294a34c753fSRafael Auler } 295*40c2e0faSMaksim Panchenko } // namespace bolt 296*40c2e0faSMaksim Panchenko } // namespace llvm 297