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/Args.h"
17#include "lld/Common/CommonLinkerContext.h"
18#include "lld/Common/Filesystem.h"
19#include "lld/Common/Strings.h"
20#include "lld/Common/TargetOptionsCommandFlags.h"
21#include "llvm/Bitcode/BitcodeWriter.h"
22#include "llvm/LTO/Config.h"
23#include "llvm/LTO/LTO.h"
24#include "llvm/Support/Caching.h"
25#include "llvm/Support/FileSystem.h"
26#include "llvm/Support/Path.h"
27#include "llvm/Support/raw_ostream.h"
28#include "llvm/Transforms/ObjCARC.h"
29
30using namespace lld;
31using namespace lld::macho;
32using namespace llvm;
33using namespace llvm::MachO;
34using namespace llvm::sys;
35
36static std::string getThinLTOOutputFile(StringRef modulePath) {
37 return lto::getThinLTOOutputFile(Path: modulePath, OldPrefix: config->thinLTOPrefixReplaceOld,
38 NewPrefix: config->thinLTOPrefixReplaceNew);
39}
40
41static lto::Config createConfig() {
42 lto::Config c;
43 c.Options = initTargetOptionsFromCodeGenFlags();
44 c.Options.EmitAddrsig = config->icfLevel == ICFLevel::safe;
45 for (StringRef C : config->mllvmOpts)
46 c.MllvmArgs.emplace_back(args: C.str());
47 c.CodeModel = getCodeModelFromCMModel();
48 c.CPU = getCPUStr();
49 c.MAttrs = getMAttrs();
50 c.DiagHandler = diagnosticHandler;
51 c.PreCodeGenPassesHook = [](legacy::PassManager &pm) {
52 pm.add(P: createObjCARCContractPass());
53 };
54
55 c.AlwaysEmitRegularLTOObj = !config->ltoObjPath.empty();
56
57 c.TimeTraceEnabled = config->timeTraceEnabled;
58 c.TimeTraceGranularity = config->timeTraceGranularity;
59 c.DebugPassManager = config->ltoDebugPassManager;
60 c.CSIRProfile = std::string(config->csProfilePath);
61 c.RunCSIRInstr = config->csProfileGenerate;
62 c.PGOWarnMismatch = config->pgoWarnMismatch;
63 c.OptLevel = config->ltoo;
64 c.CGOptLevel = config->ltoCgo;
65 if (config->saveTemps)
66 checkError(e: c.addSaveTemps(OutputFileName: config->outputFile.str() + ".",
67 /*UseInputModulePath=*/true));
68 return c;
69}
70
71// If `originalPath` exists, hardlinks `path` to `originalPath`. If that fails,
72// or `originalPath` is not set, saves `buffer` to `path`.
73static void saveOrHardlinkBuffer(StringRef buffer, const Twine &path,
74 std::optional<StringRef> originalPath) {
75 if (originalPath) {
76 auto err = fs::create_hard_link(to: *originalPath, from: path);
77 if (!err)
78 return;
79 }
80 saveBuffer(buffer, path);
81}
82
83BitcodeCompiler::BitcodeCompiler() {
84 // Initialize indexFile.
85 if (!config->thinLTOIndexOnlyArg.empty())
86 indexFile = openFile(file: config->thinLTOIndexOnlyArg);
87
88 // Initialize ltoObj.
89 lto::ThinBackend backend;
90 auto onIndexWrite = [&](StringRef S) { thinIndices.erase(V: S); };
91 if (config->thinLTOIndexOnly) {
92 backend = lto::createWriteIndexesThinBackend(
93 OldPrefix: std::string(config->thinLTOPrefixReplaceOld),
94 NewPrefix: std::string(config->thinLTOPrefixReplaceNew),
95 NativeObjectPrefix: std::string(config->thinLTOPrefixReplaceNativeObject),
96 ShouldEmitImportsFiles: config->thinLTOEmitImportsFiles, LinkedObjectsFile: indexFile.get(), OnWrite: onIndexWrite);
97 } else {
98 backend = lto::createInProcessThinBackend(
99 Parallelism: llvm::heavyweight_hardware_concurrency(Num: config->thinLTOJobs),
100 OnWrite: onIndexWrite, ShouldEmitIndexFiles: config->thinLTOEmitIndexFiles,
101 ShouldEmitImportsFiles: config->thinLTOEmitImportsFiles);
102 }
103
104 ltoObj = std::make_unique<lto::LTO>(args: createConfig(), args&: backend);
105}
106
107void BitcodeCompiler::add(BitcodeFile &f) {
108 lto::InputFile &obj = *f.obj;
109
110 if (config->thinLTOEmitIndexFiles)
111 thinIndices.insert(V: obj.getName());
112
113 ArrayRef<lto::InputFile::Symbol> objSyms = obj.symbols();
114 std::vector<lto::SymbolResolution> resols;
115 resols.reserve(n: 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>(Val: 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>(Val: 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>(s: sym, arg: sym->getName(), arg: sym->getFile(),
150 arg: RefState::Strong, /*wasBitcodeSymbol=*/arg: true);
151
152 // TODO: set the other resolution configs properly
153 }
154 checkError(e: ltoObj->add(Obj: std::move(f.obj), Res: 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.
161static void thinLTOCreateEmptyIndexFiles() {
162 DenseSet<StringRef> linkedBitCodeFiles;
163 for (InputFile *file : inputFiles)
164 if (auto *f = dyn_cast<BitcodeFile>(Val: file))
165 if (!f->lazy)
166 linkedBitCodeFiles.insert(V: f->getName());
167
168 for (InputFile *file : inputFiles) {
169 if (auto *f = dyn_cast<BitcodeFile>(Val: file)) {
170 if (!f->lazy)
171 continue;
172 if (linkedBitCodeFiles.contains(V: f->getName()))
173 continue;
174 std::string path =
175 replaceThinLTOSuffix(path: getThinLTOOutputFile(modulePath: f->obj->getName()));
176 std::unique_ptr<raw_fd_ostream> os = openFile(file: path + ".thinlto.bc");
177 if (!os)
178 continue;
179
180 ModuleSummaryIndex m(/*HaveGVs=*/false);
181 m.setSkipModuleByDistributedBackend();
182 writeIndexToFile(Index: m, Out&: *os);
183 if (config->thinLTOEmitImportsFiles)
184 openFile(file: 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).
191std::vector<ObjFile *> BitcodeCompiler::compile() {
192 unsigned maxTasks = ltoObj->getMaxTasks();
193 buf.resize(new_size: maxTasks);
194 files.resize(new_size: 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(e: localCache(CacheNameRef: "ThinLTO", TempFilePrefixRef: "Thin", CacheDirectoryPathRef: config->thinLTOCacheDir,
202 AddBuffer: [&](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(e: ltoObj->run(
209 AddStream: [&](size_t task, const Twine &moduleName) {
210 return std::make_unique<CachedFileStream>(
211 args: std::make_unique<raw_svector_ostream>(args&: buf[task]));
212 },
213 Cache: cache));
214
215 // Emit empty index files for non-indexed files
216 for (StringRef s : thinIndices) {
217 std::string path = getThinLTOOutputFile(modulePath: s);
218 openFile(file: path + ".thinlto.bc");
219 if (config->thinLTOEmitImportsFiles)
220 openFile(file: 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(path: config->ltoObjPath))
231 fatal(msg: "cannot create LTO object path " + config->ltoObjPath + ": " +
232 ec.message());
233
234 if (!fs::is_directory(Path: config->ltoObjPath)) {
235 objPathIsDir = false;
236 unsigned objCount =
237 count_if(Range&: buf, P: [](const SmallString<0> &b) { return !b.empty(); });
238 if (objCount > 1)
239 fatal(msg: "-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(path&: filePath, a: Twine(i) + "." +
249 getArchitectureName(Arch: 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(buffer: buf[0], path: outputFilePath(0));
261 if (indexFile)
262 indexFile->close();
263 return {};
264 }
265
266 if (!config->thinLTOCacheDir.empty())
267 pruneCache(Path: config->thinLTOCacheDir, Policy: config->thinLTOCachePolicy, Files: 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(buffer: objBuf,
288 path: 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(buffer: objBuf, path: filePath, originalPath: cachePath);
294 modTime = getModTime(path: filePath);
295 }
296 ret.push_back(x: make<ObjFile>(
297 args: MemoryBufferRef(objBuf, saver().save(S: filePath.str())), args&: modTime,
298 /*archiveName=*/args: "", /*lazy=*/args: false,
299 /*forceHidden=*/args: false, /*compatArch=*/args: true, /*builtFromBitcode=*/args: true));
300 }
301
302 return ret;
303}
304