1 //===- LTO.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 "LTO.h" 10 #include "COFFLinkerContext.h" 11 #include "Config.h" 12 #include "InputFiles.h" 13 #include "Symbols.h" 14 #include "lld/Common/Args.h" 15 #include "lld/Common/CommonLinkerContext.h" 16 #include "lld/Common/Filesystem.h" 17 #include "lld/Common/Strings.h" 18 #include "lld/Common/TargetOptionsCommandFlags.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/StringRef.h" 21 #include "llvm/ADT/Twine.h" 22 #include "llvm/Bitcode/BitcodeWriter.h" 23 #include "llvm/IR/DiagnosticPrinter.h" 24 #include "llvm/LTO/Config.h" 25 #include "llvm/LTO/LTO.h" 26 #include "llvm/Support/Caching.h" 27 #include "llvm/Support/CodeGen.h" 28 #include "llvm/Support/MemoryBuffer.h" 29 #include "llvm/Support/raw_ostream.h" 30 #include <cstddef> 31 #include <memory> 32 #include <string> 33 #include <vector> 34 35 using namespace llvm; 36 using namespace llvm::object; 37 using namespace lld; 38 using namespace lld::coff; 39 40 std::string BitcodeCompiler::getThinLTOOutputFile(StringRef path) { 41 return lto::getThinLTOOutputFile(path, ctx.config.thinLTOPrefixReplaceOld, 42 ctx.config.thinLTOPrefixReplaceNew); 43 } 44 45 lto::Config BitcodeCompiler::createConfig() { 46 lto::Config c; 47 c.Options = initTargetOptionsFromCodeGenFlags(); 48 c.Options.EmitAddrsig = true; 49 for (StringRef C : ctx.config.mllvmOpts) 50 c.MllvmArgs.emplace_back(C.str()); 51 52 // Always emit a section per function/datum with LTO. LLVM LTO should get most 53 // of the benefit of linker GC, but there are still opportunities for ICF. 54 c.Options.FunctionSections = true; 55 c.Options.DataSections = true; 56 57 // Use static reloc model on 32-bit x86 because it usually results in more 58 // compact code, and because there are also known code generation bugs when 59 // using the PIC model (see PR34306). 60 if (ctx.config.machine == COFF::IMAGE_FILE_MACHINE_I386) 61 c.RelocModel = Reloc::Static; 62 else 63 c.RelocModel = Reloc::PIC_; 64 #ifndef NDEBUG 65 c.DisableVerify = false; 66 #else 67 c.DisableVerify = true; 68 #endif 69 c.DiagHandler = diagnosticHandler; 70 c.DwoDir = ctx.config.dwoDir.str(); 71 c.OptLevel = ctx.config.ltoo; 72 c.CPU = getCPUStr(); 73 c.MAttrs = getMAttrs(); 74 std::optional<CodeGenOptLevel> optLevelOrNone = CodeGenOpt::getLevel( 75 ctx.config.ltoCgo.value_or(args::getCGOptLevel(ctx.config.ltoo))); 76 assert(optLevelOrNone && "Invalid optimization level!"); 77 c.CGOptLevel = *optLevelOrNone; 78 c.AlwaysEmitRegularLTOObj = !ctx.config.ltoObjPath.empty(); 79 c.DebugPassManager = ctx.config.ltoDebugPassManager; 80 c.CSIRProfile = std::string(ctx.config.ltoCSProfileFile); 81 c.RunCSIRInstr = ctx.config.ltoCSProfileGenerate; 82 c.PGOWarnMismatch = ctx.config.ltoPGOWarnMismatch; 83 c.SampleProfile = ctx.config.ltoSampleProfileName; 84 c.TimeTraceEnabled = ctx.config.timeTraceEnabled; 85 c.TimeTraceGranularity = ctx.config.timeTraceGranularity; 86 87 if (ctx.config.emit == EmitKind::LLVM) { 88 c.PreCodeGenModuleHook = [this](size_t task, const Module &m) { 89 if (std::unique_ptr<raw_fd_ostream> os = 90 openLTOOutputFile(ctx.config.outputFile)) 91 WriteBitcodeToFile(m, *os, false); 92 return false; 93 }; 94 } else if (ctx.config.emit == EmitKind::ASM) { 95 c.CGFileType = CodeGenFileType::AssemblyFile; 96 c.Options.MCOptions.AsmVerbose = true; 97 } 98 99 if (!ctx.config.saveTempsArgs.empty()) 100 checkError(c.addSaveTemps(std::string(ctx.config.outputFile) + ".", 101 /*UseInputModulePath*/ true, 102 ctx.config.saveTempsArgs)); 103 return c; 104 } 105 106 BitcodeCompiler::BitcodeCompiler(COFFLinkerContext &c) : ctx(c) { 107 // Initialize indexFile. 108 if (!ctx.config.thinLTOIndexOnlyArg.empty()) 109 indexFile = openFile(ctx.config.thinLTOIndexOnlyArg); 110 111 // Initialize ltoObj. 112 lto::ThinBackend backend; 113 if (!ctx.config.dtltoDistributor.empty()) { 114 backend = lto::createOutOfProcessThinBackend( 115 llvm::hardware_concurrency(ctx.config.thinLTOJobs), 116 /*OnWrite=*/nullptr, 117 /*ShouldEmitIndexFiles=*/false, 118 /*ShouldEmitImportFiles=*/false, ctx.config.outputFile, 119 ctx.config.dtltoDistributor, ctx.config.dtltoDistributorArgs, 120 ctx.config.dtltoCompiler, ctx.config.dtltoCompilerArgs, 121 !ctx.config.saveTempsArgs.empty()); 122 } else if (ctx.config.thinLTOIndexOnly) { 123 auto OnIndexWrite = [&](StringRef S) { thinIndices.erase(S); }; 124 backend = lto::createWriteIndexesThinBackend( 125 llvm::hardware_concurrency(ctx.config.thinLTOJobs), 126 std::string(ctx.config.thinLTOPrefixReplaceOld), 127 std::string(ctx.config.thinLTOPrefixReplaceNew), 128 std::string(ctx.config.thinLTOPrefixReplaceNativeObject), 129 ctx.config.thinLTOEmitImportsFiles, indexFile.get(), OnIndexWrite); 130 } else { 131 backend = lto::createInProcessThinBackend( 132 llvm::heavyweight_hardware_concurrency(ctx.config.thinLTOJobs)); 133 } 134 135 ltoObj = std::make_unique<lto::LTO>(createConfig(), backend, 136 ctx.config.ltoPartitions); 137 } 138 139 BitcodeCompiler::~BitcodeCompiler() = default; 140 141 static void undefine(Symbol *s) { replaceSymbol<Undefined>(s, s->getName()); } 142 143 void BitcodeCompiler::add(BitcodeFile &f) { 144 lto::InputFile &obj = *f.obj; 145 unsigned symNum = 0; 146 std::vector<Symbol *> symBodies = f.getSymbols(); 147 std::vector<lto::SymbolResolution> resols(symBodies.size()); 148 149 if (ctx.config.thinLTOIndexOnly) 150 thinIndices.insert(obj.getName()); 151 152 // Provide a resolution to the LTO API for each symbol. 153 for (const lto::InputFile::Symbol &objSym : obj.symbols()) { 154 Symbol *sym = symBodies[symNum]; 155 lto::SymbolResolution &r = resols[symNum]; 156 ++symNum; 157 158 // Ideally we shouldn't check for SF_Undefined but currently IRObjectFile 159 // reports two symbols for module ASM defined. Without this check, lld 160 // flags an undefined in IR with a definition in ASM as prevailing. 161 // Once IRObjectFile is fixed to report only one symbol this hack can 162 // be removed. 163 r.Prevailing = !objSym.isUndefined() && sym->getFile() == &f; 164 r.VisibleToRegularObj = sym->isUsedInRegularObj; 165 if (r.Prevailing) 166 undefine(sym); 167 168 // We tell LTO to not apply interprocedural optimization for wrapped 169 // (with -wrap) symbols because otherwise LTO would inline them while 170 // their values are still not final. 171 r.LinkerRedefined = !sym->canInline; 172 } 173 checkError(ltoObj->add(std::move(f.obj), resols)); 174 } 175 176 // Merge all the bitcode files we have seen, codegen the result 177 // and return the resulting objects. 178 std::vector<InputFile *> BitcodeCompiler::compile() { 179 unsigned maxTasks = ltoObj->getMaxTasks(); 180 buf.resize(maxTasks); 181 files.resize(maxTasks); 182 file_names.resize(maxTasks); 183 184 // The /lldltocache option specifies the path to a directory in which to cache 185 // native object files for ThinLTO incremental builds. If a path was 186 // specified, configure LTO to use it as the cache directory. 187 FileCache cache; 188 if (!ctx.config.ltoCache.empty()) 189 cache = check(localCache("ThinLTO", "Thin", ctx.config.ltoCache, 190 [&](size_t task, const Twine &moduleName, 191 std::unique_ptr<MemoryBuffer> mb) { 192 files[task] = std::move(mb); 193 file_names[task] = moduleName.str(); 194 })); 195 196 checkError(ltoObj->run( 197 [&](size_t task, const Twine &moduleName) { 198 buf[task].first = moduleName.str(); 199 return std::make_unique<CachedFileStream>( 200 std::make_unique<raw_svector_ostream>(buf[task].second)); 201 }, 202 cache)); 203 204 // Emit empty index files for non-indexed files 205 for (StringRef s : thinIndices) { 206 std::string path = getThinLTOOutputFile(s); 207 openFile(path + ".thinlto.bc"); 208 if (ctx.config.thinLTOEmitImportsFiles) 209 openFile(path + ".imports"); 210 } 211 212 // ThinLTO with index only option is required to generate only the index 213 // files. After that, we exit from linker and ThinLTO backend runs in a 214 // distributed environment. 215 if (ctx.config.thinLTOIndexOnly) { 216 if (!ctx.config.ltoObjPath.empty()) 217 saveBuffer(buf[0].second, ctx.config.ltoObjPath); 218 if (indexFile) 219 indexFile->close(); 220 return {}; 221 } 222 223 if (!ctx.config.ltoCache.empty()) 224 pruneCache(ctx.config.ltoCache, ctx.config.ltoCachePolicy, files); 225 226 std::vector<InputFile *> ret; 227 bool emitASM = ctx.config.emit == EmitKind::ASM; 228 const char *Ext = emitASM ? ".s" : ".obj"; 229 for (unsigned i = 0; i != maxTasks; ++i) { 230 StringRef bitcodeFilePath; 231 // Get the native object contents either from the cache or from memory. Do 232 // not use the cached MemoryBuffer directly, or the PDB will not be 233 // deterministic. 234 StringRef objBuf; 235 if (files[i]) { 236 objBuf = files[i]->getBuffer(); 237 bitcodeFilePath = file_names[i]; 238 } else { 239 objBuf = buf[i].second; 240 bitcodeFilePath = buf[i].first; 241 } 242 if (objBuf.empty()) 243 continue; 244 245 // If the input bitcode file is path/to/a.obj, then the corresponding lto 246 // object file name will look something like: path/to/main.exe.lto.a.obj. 247 StringRef ltoObjName; 248 if (bitcodeFilePath == "ld-temp.o") { 249 ltoObjName = 250 saver().save(Twine(ctx.config.outputFile) + ".lto" + 251 (i == 0 ? Twine("") : Twine('.') + Twine(i)) + Ext); 252 } else { 253 StringRef directory = sys::path::parent_path(bitcodeFilePath); 254 StringRef baseName = sys::path::stem(bitcodeFilePath); 255 StringRef outputFileBaseName = sys::path::filename(ctx.config.outputFile); 256 SmallString<64> path; 257 sys::path::append(path, directory, 258 outputFileBaseName + ".lto." + baseName + Ext); 259 sys::path::remove_dots(path, true); 260 ltoObjName = saver().save(path.str()); 261 } 262 if (llvm::is_contained(ctx.config.saveTempsArgs, "prelink") || emitASM) 263 saveBuffer(buf[i].second, ltoObjName); 264 if (!emitASM) 265 ret.push_back(ObjFile::create(ctx, MemoryBufferRef(objBuf, ltoObjName))); 266 } 267 268 return ret; 269 } 270