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 "Config.h"
11 #include "Driver.h"
12 #include "InputFiles.h"
13 #include "Symbols.h"
14 #include "Target.h"
15
16 #include "lld/Common/CommonLinkerContext.h"
17 #include "lld/Common/Filesystem.h"
18 #include "lld/Common/Strings.h"
19 #include "lld/Common/TargetOptionsCommandFlags.h"
20 #include "llvm/Bitcode/BitcodeWriter.h"
21 #include "llvm/LTO/Config.h"
22 #include "llvm/LTO/LTO.h"
23 #include "llvm/Support/Caching.h"
24 #include "llvm/Support/FileSystem.h"
25 #include "llvm/Support/Path.h"
26 #include "llvm/Support/raw_ostream.h"
27
28 using namespace lld;
29 using namespace lld::macho;
30 using namespace llvm;
31 using namespace llvm::MachO;
32 using namespace llvm::sys;
33
getThinLTOOutputFile(StringRef modulePath)34 static std::string getThinLTOOutputFile(StringRef modulePath) {
35 return lto::getThinLTOOutputFile(modulePath, config->thinLTOPrefixReplaceOld,
36 config->thinLTOPrefixReplaceNew);
37 }
38
createConfig()39 static lto::Config createConfig() {
40 lto::Config c;
41 c.Options = initTargetOptionsFromCodeGenFlags();
42 c.Options.EmitAddrsig = config->icfLevel == ICFLevel::safe;
43 for (StringRef C : config->mllvmOpts)
44 c.MllvmArgs.emplace_back(C.str());
45 for (StringRef pluginFn : config->passPlugins)
46 c.PassPlugins.push_back(std::string(pluginFn));
47 c.OptPipeline = std::string(config->ltoNewPmPasses);
48 c.CodeModel = getCodeModelFromCMModel();
49 c.CPU = getCPUStr();
50 c.MAttrs = getMAttrs();
51 c.DiagHandler = diagnosticHandler;
52
53 c.AlwaysEmitRegularLTOObj = !config->ltoObjPath.empty();
54
55 c.TimeTraceEnabled = config->timeTraceEnabled;
56 c.TimeTraceGranularity = config->timeTraceGranularity;
57 c.DebugPassManager = config->ltoDebugPassManager;
58 c.CSIRProfile = std::string(config->csProfilePath);
59 c.RunCSIRInstr = config->csProfileGenerate;
60 c.PGOWarnMismatch = config->pgoWarnMismatch;
61 c.DisableVerify = config->disableVerify;
62 c.OptLevel = config->ltoo;
63 c.CGOptLevel = config->ltoCgo;
64 if (config->saveTemps)
65 checkError(c.addSaveTemps(config->outputFile.str() + ".",
66 /*UseInputModulePath=*/true));
67 return c;
68 }
69
70 // If `originalPath` exists, hardlinks `path` to `originalPath`. If that fails,
71 // or `originalPath` is not set, saves `buffer` to `path`.
saveOrHardlinkBuffer(StringRef buffer,const Twine & path,std::optional<StringRef> originalPath)72 static void saveOrHardlinkBuffer(StringRef buffer, const Twine &path,
73 std::optional<StringRef> originalPath) {
74 if (originalPath) {
75 auto err = fs::create_hard_link(*originalPath, path);
76 if (!err)
77 return;
78 }
79 saveBuffer(buffer, path);
80 }
81
BitcodeCompiler()82 BitcodeCompiler::BitcodeCompiler() {
83 // Initialize indexFile.
84 if (!config->thinLTOIndexOnlyArg.empty())
85 indexFile = openFile(config->thinLTOIndexOnlyArg);
86
87 // Initialize ltoObj.
88 lto::ThinBackend backend;
89 auto onIndexWrite = [&](StringRef S) { thinIndices.erase(S); };
90 if (config->thinLTOIndexOnly) {
91 backend = lto::createWriteIndexesThinBackend(
92 llvm::hardware_concurrency(config->thinLTOJobs),
93 std::string(config->thinLTOPrefixReplaceOld),
94 std::string(config->thinLTOPrefixReplaceNew),
95 std::string(config->thinLTOPrefixReplaceNativeObject),
96 config->thinLTOEmitImportsFiles, indexFile.get(), onIndexWrite);
97 } else {
98 backend = lto::createInProcessThinBackend(
99 llvm::heavyweight_hardware_concurrency(config->thinLTOJobs),
100 onIndexWrite, config->thinLTOEmitIndexFiles,
101 config->thinLTOEmitImportsFiles);
102 }
103
104 ltoObj = std::make_unique<lto::LTO>(createConfig(), backend);
105 }
106
add(BitcodeFile & f)107 void BitcodeCompiler::add(BitcodeFile &f) {
108 lto::InputFile &obj = *f.obj;
109
110 if (config->thinLTOEmitIndexFiles)
111 thinIndices.insert(obj.getName());
112
113 ArrayRef<lto::InputFile::Symbol> objSyms = obj.symbols();
114 std::vector<lto::SymbolResolution> resols;
115 resols.reserve(objSyms.size());
116
117 // Provide a resolution to the LTO API for each symbol.
118 bool exportDynamic =
119 config->outputType != MH_EXECUTE || config->exportDynamic;
120 auto symIt = f.symbols.begin();
121 for (const lto::InputFile::Symbol &objSym : objSyms) {
122 resols.emplace_back();
123 lto::SymbolResolution &r = resols.back();
124 Symbol *sym = *symIt++;
125
126 // Ideally we shouldn't check for SF_Undefined but currently IRObjectFile
127 // reports two symbols for module ASM defined. Without this check, lld
128 // flags an undefined in IR with a definition in ASM as prevailing.
129 // Once IRObjectFile is fixed to report only one symbol this hack can
130 // be removed.
131 r.Prevailing = !objSym.isUndefined() && sym->getFile() == &f;
132
133 if (const auto *defined = dyn_cast<Defined>(sym)) {
134 r.ExportDynamic =
135 defined->isExternal() && !defined->privateExtern && exportDynamic;
136 r.FinalDefinitionInLinkageUnit =
137 !defined->isExternalWeakDef() && !defined->interposable;
138 } else if (const auto *common = dyn_cast<CommonSymbol>(sym)) {
139 r.ExportDynamic = !common->privateExtern && exportDynamic;
140 r.FinalDefinitionInLinkageUnit = true;
141 }
142
143 r.VisibleToRegularObj =
144 sym->isUsedInRegularObj || (r.Prevailing && r.ExportDynamic);
145
146 // Un-define the symbol so that we don't get duplicate symbol errors when we
147 // load the ObjFile emitted by LTO compilation.
148 if (r.Prevailing)
149 replaceSymbol<Undefined>(sym, sym->getName(), sym->getFile(),
150 RefState::Strong, /*wasBitcodeSymbol=*/true);
151
152 // TODO: set the other resolution configs properly
153 }
154 checkError(ltoObj->add(std::move(f.obj), resols));
155 hasFiles = true;
156 }
157
158 // If LazyObjFile has not been added to link, emit empty index files.
159 // This is needed because this is what GNU gold plugin does and we have a
160 // distributed build system that depends on that behavior.
thinLTOCreateEmptyIndexFiles()161 static void thinLTOCreateEmptyIndexFiles() {
162 DenseSet<StringRef> linkedBitCodeFiles;
163 for (InputFile *file : inputFiles)
164 if (auto *f = dyn_cast<BitcodeFile>(file))
165 if (!f->lazy)
166 linkedBitCodeFiles.insert(f->getName());
167
168 for (InputFile *file : inputFiles) {
169 if (auto *f = dyn_cast<BitcodeFile>(file)) {
170 if (!f->lazy)
171 continue;
172 if (linkedBitCodeFiles.contains(f->getName()))
173 continue;
174 std::string path =
175 replaceThinLTOSuffix(getThinLTOOutputFile(f->obj->getName()));
176 std::unique_ptr<raw_fd_ostream> os = openFile(path + ".thinlto.bc");
177 if (!os)
178 continue;
179
180 ModuleSummaryIndex m(/*HaveGVs=*/false);
181 m.setSkipModuleByDistributedBackend();
182 writeIndexToFile(m, *os);
183 if (config->thinLTOEmitImportsFiles)
184 openFile(path + ".imports");
185 }
186 }
187 }
188
189 // Merge all the bitcode files we have seen, codegen the result
190 // and return the resulting ObjectFile(s).
compile()191 std::vector<ObjFile *> BitcodeCompiler::compile() {
192 unsigned maxTasks = ltoObj->getMaxTasks();
193 buf.resize(maxTasks);
194 files.resize(maxTasks);
195
196 // The -cache_path_lto option specifies the path to a directory in which
197 // to cache native object files for ThinLTO incremental builds. If a path was
198 // specified, configure LTO to use it as the cache directory.
199 FileCache cache;
200 if (!config->thinLTOCacheDir.empty())
201 cache = check(localCache("ThinLTO", "Thin", config->thinLTOCacheDir,
202 [&](size_t task, const Twine &moduleName,
203 std::unique_ptr<MemoryBuffer> mb) {
204 files[task] = std::move(mb);
205 }));
206
207 if (hasFiles)
208 checkError(ltoObj->run(
209 [&](size_t task, const Twine &moduleName) {
210 return std::make_unique<CachedFileStream>(
211 std::make_unique<raw_svector_ostream>(buf[task]));
212 },
213 cache));
214
215 // Emit empty index files for non-indexed files
216 for (StringRef s : thinIndices) {
217 std::string path = getThinLTOOutputFile(s);
218 openFile(path + ".thinlto.bc");
219 if (config->thinLTOEmitImportsFiles)
220 openFile(path + ".imports");
221 }
222
223 if (config->thinLTOEmitIndexFiles)
224 thinLTOCreateEmptyIndexFiles();
225
226 // In ThinLTO mode, Clang passes a temporary directory in -object_path_lto,
227 // while the argument is a single file in FullLTO mode.
228 bool objPathIsDir = true;
229 if (!config->ltoObjPath.empty()) {
230 if (std::error_code ec = fs::create_directories(config->ltoObjPath))
231 fatal("cannot create LTO object path " + config->ltoObjPath + ": " +
232 ec.message());
233
234 if (!fs::is_directory(config->ltoObjPath)) {
235 objPathIsDir = false;
236 unsigned objCount =
237 count_if(buf, [](const SmallString<0> &b) { return !b.empty(); });
238 if (objCount > 1)
239 fatal("-object_path_lto must specify a directory when using ThinLTO");
240 }
241 }
242
243 auto outputFilePath = [objPathIsDir](int i) {
244 SmallString<261> filePath("/tmp/lto.tmp");
245 if (!config->ltoObjPath.empty()) {
246 filePath = config->ltoObjPath;
247 if (objPathIsDir)
248 path::append(filePath, Twine(i) + "." +
249 getArchitectureName(config->arch()) +
250 ".lto.o");
251 }
252 return filePath;
253 };
254
255 // ThinLTO with index only option is required to generate only the index
256 // files. After that, we exit from linker and ThinLTO backend runs in a
257 // distributed environment.
258 if (config->thinLTOIndexOnly) {
259 if (!config->ltoObjPath.empty())
260 saveBuffer(buf[0], outputFilePath(0));
261 if (indexFile)
262 indexFile->close();
263 return {};
264 }
265
266 if (!config->thinLTOCacheDir.empty())
267 pruneCache(config->thinLTOCacheDir, config->thinLTOCachePolicy, files);
268
269 std::vector<ObjFile *> ret;
270 for (unsigned i = 0; i < maxTasks; ++i) {
271 // Get the native object contents either from the cache or from memory. Do
272 // not use the cached MemoryBuffer directly to ensure dsymutil does not
273 // race with the cache pruner.
274 StringRef objBuf;
275 std::optional<StringRef> cachePath;
276 if (files[i]) {
277 objBuf = files[i]->getBuffer();
278 cachePath = files[i]->getBufferIdentifier();
279 } else {
280 objBuf = buf[i];
281 }
282 if (objBuf.empty())
283 continue;
284
285 // FIXME: should `saveTemps` and `ltoObjPath` use the same file name?
286 if (config->saveTemps)
287 saveBuffer(objBuf,
288 config->outputFile + ((i == 0) ? "" : Twine(i)) + ".lto.o");
289
290 auto filePath = outputFilePath(i);
291 uint32_t modTime = 0;
292 if (!config->ltoObjPath.empty()) {
293 saveOrHardlinkBuffer(objBuf, filePath, cachePath);
294 modTime = getModTime(filePath);
295 }
296 ret.push_back(make<ObjFile>(
297 MemoryBufferRef(objBuf, saver().save(filePath.str())), modTime,
298 /*archiveName=*/"", /*lazy=*/false,
299 /*forceHidden=*/false, /*compatArch=*/true, /*builtFromBitcode=*/true));
300 }
301
302 return ret;
303 }
304