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 "InputFiles.h"
12#include "LinkerScript.h"
13#include "SymbolTable.h"
14#include "Symbols.h"
15#include "lld/Common/Args.h"
16#include "lld/Common/ErrorHandler.h"
17#include "lld/Common/TargetOptionsCommandFlags.h"
18#include "llvm/ADT/STLExtras.h"
19#include "llvm/ADT/SmallString.h"
20#include "llvm/ADT/StringRef.h"
21#include "llvm/ADT/Twine.h"
22#include "llvm/BinaryFormat/ELF.h"
23#include "llvm/Bitcode/BitcodeReader.h"
24#include "llvm/Bitcode/BitcodeWriter.h"
25#include "llvm/IR/DiagnosticPrinter.h"
26#include "llvm/LTO/Caching.h"
27#include "llvm/LTO/Config.h"
28#include "llvm/LTO/LTO.h"
29#include "llvm/Object/SymbolicFile.h"
30#include "llvm/Support/CodeGen.h"
31#include "llvm/Support/Error.h"
32#include "llvm/Support/FileSystem.h"
33#include "llvm/Support/MemoryBuffer.h"
34#include <algorithm>
35#include <cstddef>
36#include <memory>
37#include <string>
38#include <system_error>
39#include <vector>
40
41using namespace llvm;
42using namespace llvm::object;
43using namespace llvm::ELF;
44using namespace lld;
45using namespace lld::elf;
46
47// Creates an empty file to store a list of object files for final
48// linking of distributed ThinLTO.
49static std::unique_ptr<raw_fd_ostream> openFile(StringRef file) {
50  std::error_code ec;
51  auto ret =
52      std::make_unique<raw_fd_ostream>(file, ec, sys::fs::OpenFlags::OF_None);
53  if (ec) {
54    error("cannot open " + file + ": " + ec.message());
55    return nullptr;
56  }
57  return ret;
58}
59
60static std::string getThinLTOOutputFile(StringRef modulePath) {
61  return lto::getThinLTOOutputFile(
62      std::string(modulePath), std::string(config->thinLTOPrefixReplace.first),
63      std::string(config->thinLTOPrefixReplace.second));
64}
65
66static lto::Config createConfig() {
67  lto::Config c;
68
69  // LLD supports the new relocations and address-significance tables.
70  c.Options = initTargetOptionsFromCodeGenFlags();
71  c.Options.RelaxELFRelocations = true;
72  c.Options.EmitAddrsig = true;
73
74  // Always emit a section per function/datum with LTO.
75  c.Options.FunctionSections = true;
76  c.Options.DataSections = true;
77
78  // Check if basic block sections must be used.
79  // Allowed values for --lto-basicblock-sections are "all", "labels",
80  // "<file name specifying basic block ids>", or none.  This is the equivalent
81  // of -fbasic-block-sections= flag in clang.
82  if (!config->ltoBasicBlockSections.empty()) {
83    if (config->ltoBasicBlockSections == "all") {
84      c.Options.BBSections = BasicBlockSection::All;
85    } else if (config->ltoBasicBlockSections == "labels") {
86      c.Options.BBSections = BasicBlockSection::Labels;
87    } else if (config->ltoBasicBlockSections == "none") {
88      c.Options.BBSections = BasicBlockSection::None;
89    } else {
90      ErrorOr<std::unique_ptr<MemoryBuffer>> MBOrErr =
91          MemoryBuffer::getFile(config->ltoBasicBlockSections.str());
92      if (!MBOrErr) {
93        error("cannot open " + config->ltoBasicBlockSections + ":" +
94              MBOrErr.getError().message());
95      } else {
96        c.Options.BBSectionsFuncListBuf = std::move(*MBOrErr);
97      }
98      c.Options.BBSections = BasicBlockSection::List;
99    }
100  }
101
102  c.Options.UniqueBasicBlockSectionNames =
103      config->ltoUniqueBasicBlockSectionNames;
104
105  if (auto relocModel = getRelocModelFromCMModel())
106    c.RelocModel = *relocModel;
107  else if (config->relocatable)
108    c.RelocModel = None;
109  else if (config->isPic)
110    c.RelocModel = Reloc::PIC_;
111  else
112    c.RelocModel = Reloc::Static;
113
114  c.CodeModel = getCodeModelFromCMModel();
115  c.DisableVerify = config->disableVerify;
116  c.DiagHandler = diagnosticHandler;
117  c.OptLevel = config->ltoo;
118  c.CPU = getCPUStr();
119  c.MAttrs = getMAttrs();
120  c.CGOptLevel = args::getCGOptLevel(config->ltoo);
121
122  c.PTO.LoopVectorization = c.OptLevel > 1;
123  c.PTO.SLPVectorization = c.OptLevel > 1;
124
125  // Set up a custom pipeline if we've been asked to.
126  c.OptPipeline = std::string(config->ltoNewPmPasses);
127  c.AAPipeline = std::string(config->ltoAAPipeline);
128
129  // Set up optimization remarks if we've been asked to.
130  c.RemarksFilename = std::string(config->optRemarksFilename);
131  c.RemarksPasses = std::string(config->optRemarksPasses);
132  c.RemarksWithHotness = config->optRemarksWithHotness;
133  c.RemarksFormat = std::string(config->optRemarksFormat);
134
135  c.SampleProfile = std::string(config->ltoSampleProfile);
136  c.UseNewPM = config->ltoNewPassManager;
137  c.DebugPassManager = config->ltoDebugPassManager;
138  c.DwoDir = std::string(config->dwoDir);
139
140  c.HasWholeProgramVisibility = config->ltoWholeProgramVisibility;
141  c.AlwaysEmitRegularLTOObj = !config->ltoObjPath.empty();
142
143  for (const llvm::StringRef &name : config->thinLTOModulesToCompile)
144    c.ThinLTOModulesToCompile.emplace_back(name);
145
146  c.TimeTraceEnabled = config->timeTraceEnabled;
147  c.TimeTraceGranularity = config->timeTraceGranularity;
148
149  c.CSIRProfile = std::string(config->ltoCSProfileFile);
150  c.RunCSIRInstr = config->ltoCSProfileGenerate;
151
152  if (config->emitLLVM) {
153    c.PostInternalizeModuleHook = [](size_t task, const Module &m) {
154      if (std::unique_ptr<raw_fd_ostream> os = openFile(config->outputFile))
155        WriteBitcodeToFile(m, *os, false);
156      return false;
157    };
158  }
159
160  if (config->ltoEmitAsm)
161    c.CGFileType = CGFT_AssemblyFile;
162
163  if (config->saveTemps)
164    checkError(c.addSaveTemps(config->outputFile.str() + ".",
165                              /*UseInputModulePath*/ true));
166  return c;
167}
168
169BitcodeCompiler::BitcodeCompiler() {
170  // Initialize indexFile.
171  if (!config->thinLTOIndexOnlyArg.empty())
172    indexFile = openFile(config->thinLTOIndexOnlyArg);
173
174  // Initialize ltoObj.
175  lto::ThinBackend backend;
176  if (config->thinLTOIndexOnly) {
177    auto onIndexWrite = [&](StringRef s) { thinIndices.erase(s); };
178    backend = lto::createWriteIndexesThinBackend(
179        std::string(config->thinLTOPrefixReplace.first),
180        std::string(config->thinLTOPrefixReplace.second),
181        config->thinLTOEmitImportsFiles, indexFile.get(), onIndexWrite);
182  } else {
183    backend = lto::createInProcessThinBackend(
184        llvm::heavyweight_hardware_concurrency(config->thinLTOJobs));
185  }
186
187  ltoObj = std::make_unique<lto::LTO>(createConfig(), backend,
188                                       config->ltoPartitions);
189
190  // Initialize usedStartStop.
191  for (Symbol *sym : symtab->symbols()) {
192    StringRef s = sym->getName();
193    for (StringRef prefix : {"__start_", "__stop_"})
194      if (s.startswith(prefix))
195        usedStartStop.insert(s.substr(prefix.size()));
196  }
197}
198
199BitcodeCompiler::~BitcodeCompiler() = default;
200
201void BitcodeCompiler::add(BitcodeFile &f) {
202  lto::InputFile &obj = *f.obj;
203  bool isExec = !config->shared && !config->relocatable;
204
205  if (config->thinLTOIndexOnly)
206    thinIndices.insert(obj.getName());
207
208  ArrayRef<Symbol *> syms = f.getSymbols();
209  ArrayRef<lto::InputFile::Symbol> objSyms = obj.symbols();
210  std::vector<lto::SymbolResolution> resols(syms.size());
211
212  // Provide a resolution to the LTO API for each symbol.
213  for (size_t i = 0, e = syms.size(); i != e; ++i) {
214    Symbol *sym = syms[i];
215    const lto::InputFile::Symbol &objSym = objSyms[i];
216    lto::SymbolResolution &r = resols[i];
217
218    // Ideally we shouldn't check for SF_Undefined but currently IRObjectFile
219    // reports two symbols for module ASM defined. Without this check, lld
220    // flags an undefined in IR with a definition in ASM as prevailing.
221    // Once IRObjectFile is fixed to report only one symbol this hack can
222    // be removed.
223    r.Prevailing = !objSym.isUndefined() && sym->file == &f;
224
225    // We ask LTO to preserve following global symbols:
226    // 1) All symbols when doing relocatable link, so that them can be used
227    //    for doing final link.
228    // 2) Symbols that are used in regular objects.
229    // 3) C named sections if we have corresponding __start_/__stop_ symbol.
230    // 4) Symbols that are defined in bitcode files and used for dynamic linking.
231    r.VisibleToRegularObj = config->relocatable || sym->isUsedInRegularObj ||
232                            (r.Prevailing && sym->includeInDynsym()) ||
233                            usedStartStop.count(objSym.getSectionName());
234    const auto *dr = dyn_cast<Defined>(sym);
235    r.FinalDefinitionInLinkageUnit =
236        (isExec || sym->visibility != STV_DEFAULT) && dr &&
237        // Skip absolute symbols from ELF objects, otherwise PC-rel relocations
238        // will be generated by for them, triggering linker errors.
239        // Symbol section is always null for bitcode symbols, hence the check
240        // for isElf(). Skip linker script defined symbols as well: they have
241        // no File defined.
242        !(dr->section == nullptr && (!sym->file || sym->file->isElf()));
243
244    if (r.Prevailing)
245      sym->replace(Undefined{nullptr, sym->getName(), STB_GLOBAL, STV_DEFAULT,
246                             sym->type});
247
248    // We tell LTO to not apply interprocedural optimization for wrapped
249    // (with --wrap) symbols because otherwise LTO would inline them while
250    // their values are still not final.
251    r.LinkerRedefined = !sym->canInline;
252  }
253  checkError(ltoObj->add(std::move(f.obj), resols));
254}
255
256// If LazyObjFile has not been added to link, emit empty index files.
257// This is needed because this is what GNU gold plugin does and we have a
258// distributed build system that depends on that behavior.
259static void thinLTOCreateEmptyIndexFiles() {
260  for (LazyObjFile *f : lazyObjFiles) {
261    if (f->fetched || !isBitcode(f->mb))
262      continue;
263    std::string path = replaceThinLTOSuffix(getThinLTOOutputFile(f->getName()));
264    std::unique_ptr<raw_fd_ostream> os = openFile(path + ".thinlto.bc");
265    if (!os)
266      continue;
267
268    ModuleSummaryIndex m(/*HaveGVs*/ false);
269    m.setSkipModuleByDistributedBackend();
270    WriteIndexToFile(m, *os);
271    if (config->thinLTOEmitImportsFiles)
272      openFile(path + ".imports");
273  }
274}
275
276// Merge all the bitcode files we have seen, codegen the result
277// and return the resulting ObjectFile(s).
278std::vector<InputFile *> BitcodeCompiler::compile() {
279  unsigned maxTasks = ltoObj->getMaxTasks();
280  buf.resize(maxTasks);
281  files.resize(maxTasks);
282
283  // The --thinlto-cache-dir option specifies the path to a directory in which
284  // to cache native object files for ThinLTO incremental builds. If a path was
285  // specified, configure LTO to use it as the cache directory.
286  lto::NativeObjectCache cache;
287  if (!config->thinLTOCacheDir.empty())
288    cache = check(
289        lto::localCache(config->thinLTOCacheDir,
290                        [&](size_t task, std::unique_ptr<MemoryBuffer> mb) {
291                          files[task] = std::move(mb);
292                        }));
293
294  if (!bitcodeFiles.empty())
295    checkError(ltoObj->run(
296        [&](size_t task) {
297          return std::make_unique<lto::NativeObjectStream>(
298              std::make_unique<raw_svector_ostream>(buf[task]));
299        },
300        cache));
301
302  // Emit empty index files for non-indexed files but not in single-module mode.
303  if (config->thinLTOModulesToCompile.empty()) {
304    for (StringRef s : thinIndices) {
305      std::string path = getThinLTOOutputFile(s);
306      openFile(path + ".thinlto.bc");
307      if (config->thinLTOEmitImportsFiles)
308        openFile(path + ".imports");
309    }
310  }
311
312  if (config->thinLTOIndexOnly) {
313    thinLTOCreateEmptyIndexFiles();
314
315    if (!config->ltoObjPath.empty())
316      saveBuffer(buf[0], config->ltoObjPath);
317
318    // ThinLTO with index only option is required to generate only the index
319    // files. After that, we exit from linker and ThinLTO backend runs in a
320    // distributed environment.
321    if (indexFile)
322      indexFile->close();
323    return {};
324  }
325
326  if (!config->thinLTOCacheDir.empty())
327    pruneCache(config->thinLTOCacheDir, config->thinLTOCachePolicy);
328
329  if (!config->ltoObjPath.empty()) {
330    saveBuffer(buf[0], config->ltoObjPath);
331    for (unsigned i = 1; i != maxTasks; ++i)
332      saveBuffer(buf[i], config->ltoObjPath + Twine(i));
333  }
334
335  if (config->saveTemps) {
336    if (!buf[0].empty())
337      saveBuffer(buf[0], config->outputFile + ".lto.o");
338    for (unsigned i = 1; i != maxTasks; ++i)
339      saveBuffer(buf[i], config->outputFile + Twine(i) + ".lto.o");
340  }
341
342  if (config->ltoEmitAsm) {
343    saveBuffer(buf[0], config->outputFile);
344    for (unsigned i = 1; i != maxTasks; ++i)
345      saveBuffer(buf[i], config->outputFile + Twine(i));
346    return {};
347  }
348
349  std::vector<InputFile *> ret;
350  for (unsigned i = 0; i != maxTasks; ++i)
351    if (!buf[i].empty())
352      ret.push_back(createObjectFile(MemoryBufferRef(buf[i], "lto.tmp")));
353
354  for (std::unique_ptr<MemoryBuffer> &file : files)
355    if (file)
356      ret.push_back(createObjectFile(*file));
357  return ret;
358}
359