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