1//===-- gold-plugin.cpp - Plugin to gold for Link Time Optimization  ------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This is a gold plugin for LLVM. It provides an LLVM implementation of the
11// interface described in http://gcc.gnu.org/wiki/whopr/driver .
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/ADT/StringSet.h"
16#include "llvm/Analysis/TargetLibraryInfo.h"
17#include "llvm/Analysis/TargetTransformInfo.h"
18#include "llvm/Bitcode/ReaderWriter.h"
19#include "llvm/CodeGen/Analysis.h"
20#include "llvm/CodeGen/CommandFlags.h"
21#include "llvm/CodeGen/ParallelCG.h"
22#include "llvm/Config/config.h" // plugin-api.h requires HAVE_STDINT_H
23#include "llvm/IR/AutoUpgrade.h"
24#include "llvm/IR/Constants.h"
25#include "llvm/IR/DiagnosticInfo.h"
26#include "llvm/IR/DiagnosticPrinter.h"
27#include "llvm/IR/LLVMContext.h"
28#include "llvm/IR/LegacyPassManager.h"
29#include "llvm/IR/Module.h"
30#include "llvm/IR/Verifier.h"
31#include "llvm/LTO/LTO.h"
32#include "llvm/Linker/IRMover.h"
33#include "llvm/MC/SubtargetFeature.h"
34#include "llvm/Object/IRObjectFile.h"
35#include "llvm/Object/ModuleSummaryIndexObjectFile.h"
36#include "llvm/Support/Host.h"
37#include "llvm/Support/ManagedStatic.h"
38#include "llvm/Support/MemoryBuffer.h"
39#include "llvm/Support/Path.h"
40#include "llvm/Support/TargetRegistry.h"
41#include "llvm/Support/TargetSelect.h"
42#include "llvm/Support/ThreadPool.h"
43#include "llvm/Support/raw_ostream.h"
44#include "llvm/Support/thread.h"
45#include "llvm/Transforms/IPO.h"
46#include "llvm/Transforms/IPO/FunctionImport.h"
47#include "llvm/Transforms/IPO/PassManagerBuilder.h"
48#include "llvm/Transforms/Utils/FunctionImportUtils.h"
49#include "llvm/Transforms/Utils/GlobalStatus.h"
50#include "llvm/Transforms/Utils/ValueMapper.h"
51#include <list>
52#include <plugin-api.h>
53#include <system_error>
54#include <utility>
55#include <vector>
56
57// FIXME: remove this declaration when we stop maintaining Ubuntu Quantal and
58// Precise and Debian Wheezy (binutils 2.23 is required)
59#define LDPO_PIE 3
60
61#define LDPT_GET_SYMBOLS_V3 28
62
63using namespace llvm;
64
65static ld_plugin_status discard_message(int level, const char *format, ...) {
66  // Die loudly. Recent versions of Gold pass ld_plugin_message as the first
67  // callback in the transfer vector. This should never be called.
68  abort();
69}
70
71static ld_plugin_release_input_file release_input_file = nullptr;
72static ld_plugin_get_input_file get_input_file = nullptr;
73static ld_plugin_message message = discard_message;
74
75namespace {
76struct claimed_file {
77  void *handle;
78  void *leader_handle;
79  std::vector<ld_plugin_symbol> syms;
80  off_t filesize;
81  std::string name;
82};
83
84/// RAII wrapper to manage opening and releasing of a ld_plugin_input_file.
85struct PluginInputFile {
86  void *Handle;
87  std::unique_ptr<ld_plugin_input_file> File;
88
89  PluginInputFile(void *Handle) : Handle(Handle) {
90    File = llvm::make_unique<ld_plugin_input_file>();
91    if (get_input_file(Handle, File.get()) != LDPS_OK)
92      message(LDPL_FATAL, "Failed to get file information");
93  }
94  ~PluginInputFile() {
95    // File would have been reset to nullptr if we moved this object
96    // to a new owner.
97    if (File)
98      if (release_input_file(Handle) != LDPS_OK)
99        message(LDPL_FATAL, "Failed to release file information");
100  }
101
102  ld_plugin_input_file &file() { return *File; }
103
104  PluginInputFile(PluginInputFile &&RHS) = default;
105  PluginInputFile &operator=(PluginInputFile &&RHS) = default;
106};
107
108struct ResolutionInfo {
109  uint64_t CommonSize = 0;
110  unsigned CommonAlign = 0;
111  bool IsLinkonceOdr = true;
112  GlobalValue::UnnamedAddr UnnamedAddr = GlobalValue::UnnamedAddr::Global;
113  GlobalValue::VisibilityTypes Visibility = GlobalValue::DefaultVisibility;
114  bool CommonInternal = false;
115  bool UseCommon = false;
116};
117
118/// Class to own information used by a task or during its cleanup for a
119/// ThinLTO backend instantiation.
120class ThinLTOTaskInfo {
121  /// The output stream the task will codegen into.
122  std::unique_ptr<raw_fd_ostream> OS;
123
124  /// The file name corresponding to the output stream, used during cleanup.
125  std::string Filename;
126
127  /// Flag indicating whether the output file is a temp file that must be
128  /// added to the cleanup list during cleanup.
129  bool TempOutFile;
130
131public:
132  ThinLTOTaskInfo(std::unique_ptr<raw_fd_ostream> OS, std::string Filename,
133                  bool TempOutFile)
134      : OS(std::move(OS)), Filename(std::move(Filename)),
135        TempOutFile(TempOutFile) {}
136
137  /// Performs task related cleanup activities that must be done
138  /// single-threaded (i.e. call backs to gold).
139  void cleanup();
140};
141}
142
143static ld_plugin_add_symbols add_symbols = nullptr;
144static ld_plugin_get_symbols get_symbols = nullptr;
145static ld_plugin_add_input_file add_input_file = nullptr;
146static ld_plugin_set_extra_library_path set_extra_library_path = nullptr;
147static ld_plugin_get_view get_view = nullptr;
148static Optional<Reloc::Model> RelocationModel;
149static std::string output_name = "";
150static std::list<claimed_file> Modules;
151static DenseMap<int, void *> FDToLeaderHandle;
152static StringMap<ResolutionInfo> ResInfo;
153static std::vector<std::string> Cleanup;
154static llvm::TargetOptions TargetOpts;
155static std::string DefaultTriple = sys::getDefaultTargetTriple();
156
157namespace options {
158  enum OutputType {
159    OT_NORMAL,
160    OT_DISABLE,
161    OT_BC_ONLY,
162    OT_SAVE_TEMPS
163  };
164  static bool generate_api_file = false;
165  static OutputType TheOutputType = OT_NORMAL;
166  static unsigned OptLevel = 2;
167  // Default parallelism of 0 used to indicate that user did not specify.
168  // Actual parallelism default value depends on implementation.
169  // Currently, code generation defaults to no parallelism, whereas
170  // ThinLTO uses the hardware_concurrency as the default.
171  static unsigned Parallelism = 0;
172#ifdef NDEBUG
173  static bool DisableVerify = true;
174#else
175  static bool DisableVerify = false;
176#endif
177  static std::string obj_path;
178  static std::string extra_library_path;
179  static std::string triple;
180  static std::string mcpu;
181  // When the thinlto plugin option is specified, only read the function
182  // the information from intermediate files and write a combined
183  // global index for the ThinLTO backends.
184  static bool thinlto = false;
185  // If false, all ThinLTO backend compilations through code gen are performed
186  // using multiple threads in the gold-plugin, before handing control back to
187  // gold. If true, write individual backend index files which reflect
188  // the import decisions, and exit afterwards. The assumption is
189  // that the build system will launch the backend processes.
190  static bool thinlto_index_only = false;
191  // If true, when generating individual index files for distributed backends,
192  // also generate a "${bitcodefile}.imports" file at the same location for each
193  // bitcode file, listing the files it imports from in plain text. This is to
194  // support distributed build file staging.
195  static bool thinlto_emit_imports_files = false;
196  // Option to control where files for a distributed backend (the individual
197  // index files and optional imports files) are created.
198  // If specified, expects a string of the form "oldprefix:newprefix", and
199  // instead of generating these files in the same directory path as the
200  // corresponding bitcode file, will use a path formed by replacing the
201  // bitcode file's path prefix matching oldprefix with newprefix.
202  static std::string thinlto_prefix_replace;
203  // Additional options to pass into the code generator.
204  // Note: This array will contain all plugin options which are not claimed
205  // as plugin exclusive to pass to the code generator.
206  // For example, "generate-api-file" and "as"options are for the plugin
207  // use only and will not be passed.
208  static std::vector<const char *> extra;
209
210  static void process_plugin_option(const char *opt_)
211  {
212    if (opt_ == nullptr)
213      return;
214    llvm::StringRef opt = opt_;
215
216    if (opt == "generate-api-file") {
217      generate_api_file = true;
218    } else if (opt.startswith("mcpu=")) {
219      mcpu = opt.substr(strlen("mcpu="));
220    } else if (opt.startswith("extra-library-path=")) {
221      extra_library_path = opt.substr(strlen("extra_library_path="));
222    } else if (opt.startswith("mtriple=")) {
223      triple = opt.substr(strlen("mtriple="));
224    } else if (opt.startswith("obj-path=")) {
225      obj_path = opt.substr(strlen("obj-path="));
226    } else if (opt == "emit-llvm") {
227      TheOutputType = OT_BC_ONLY;
228    } else if (opt == "save-temps") {
229      TheOutputType = OT_SAVE_TEMPS;
230    } else if (opt == "disable-output") {
231      TheOutputType = OT_DISABLE;
232    } else if (opt == "thinlto") {
233      thinlto = true;
234    } else if (opt == "thinlto-index-only") {
235      thinlto_index_only = true;
236    } else if (opt == "thinlto-emit-imports-files") {
237      thinlto_emit_imports_files = true;
238    } else if (opt.startswith("thinlto-prefix-replace=")) {
239      thinlto_prefix_replace = opt.substr(strlen("thinlto-prefix-replace="));
240      if (thinlto_prefix_replace.find(";") == std::string::npos)
241        message(LDPL_FATAL, "thinlto-prefix-replace expects 'old;new' format");
242    } else if (opt.size() == 2 && opt[0] == 'O') {
243      if (opt[1] < '0' || opt[1] > '3')
244        message(LDPL_FATAL, "Optimization level must be between 0 and 3");
245      OptLevel = opt[1] - '0';
246    } else if (opt.startswith("jobs=")) {
247      if (StringRef(opt_ + 5).getAsInteger(10, Parallelism))
248        message(LDPL_FATAL, "Invalid parallelism level: %s", opt_ + 5);
249    } else if (opt == "disable-verify") {
250      DisableVerify = true;
251    } else {
252      // Save this option to pass to the code generator.
253      // ParseCommandLineOptions() expects argv[0] to be program name. Lazily
254      // add that.
255      if (extra.empty())
256        extra.push_back("LLVMgold");
257
258      extra.push_back(opt_);
259    }
260  }
261}
262
263static ld_plugin_status claim_file_hook(const ld_plugin_input_file *file,
264                                        int *claimed);
265static ld_plugin_status all_symbols_read_hook(void);
266static ld_plugin_status cleanup_hook(void);
267
268extern "C" ld_plugin_status onload(ld_plugin_tv *tv);
269ld_plugin_status onload(ld_plugin_tv *tv) {
270  InitializeAllTargetInfos();
271  InitializeAllTargets();
272  InitializeAllTargetMCs();
273  InitializeAllAsmParsers();
274  InitializeAllAsmPrinters();
275
276  // We're given a pointer to the first transfer vector. We read through them
277  // until we find one where tv_tag == LDPT_NULL. The REGISTER_* tagged values
278  // contain pointers to functions that we need to call to register our own
279  // hooks. The others are addresses of functions we can use to call into gold
280  // for services.
281
282  bool registeredClaimFile = false;
283  bool RegisteredAllSymbolsRead = false;
284
285  for (; tv->tv_tag != LDPT_NULL; ++tv) {
286    // Cast tv_tag to int to allow values not in "enum ld_plugin_tag", like, for
287    // example, LDPT_GET_SYMBOLS_V3 when building against an older plugin-api.h
288    // header.
289    switch (static_cast<int>(tv->tv_tag)) {
290    case LDPT_OUTPUT_NAME:
291      output_name = tv->tv_u.tv_string;
292      break;
293    case LDPT_LINKER_OUTPUT:
294      switch (tv->tv_u.tv_val) {
295      case LDPO_REL: // .o
296      case LDPO_DYN: // .so
297      case LDPO_PIE: // position independent executable
298        RelocationModel = Reloc::PIC_;
299        break;
300      case LDPO_EXEC: // .exe
301        RelocationModel = Reloc::Static;
302        break;
303      default:
304        message(LDPL_ERROR, "Unknown output file type %d", tv->tv_u.tv_val);
305        return LDPS_ERR;
306      }
307      break;
308    case LDPT_OPTION:
309      options::process_plugin_option(tv->tv_u.tv_string);
310      break;
311    case LDPT_REGISTER_CLAIM_FILE_HOOK: {
312      ld_plugin_register_claim_file callback;
313      callback = tv->tv_u.tv_register_claim_file;
314
315      if (callback(claim_file_hook) != LDPS_OK)
316        return LDPS_ERR;
317
318      registeredClaimFile = true;
319    } break;
320    case LDPT_REGISTER_ALL_SYMBOLS_READ_HOOK: {
321      ld_plugin_register_all_symbols_read callback;
322      callback = tv->tv_u.tv_register_all_symbols_read;
323
324      if (callback(all_symbols_read_hook) != LDPS_OK)
325        return LDPS_ERR;
326
327      RegisteredAllSymbolsRead = true;
328    } break;
329    case LDPT_REGISTER_CLEANUP_HOOK: {
330      ld_plugin_register_cleanup callback;
331      callback = tv->tv_u.tv_register_cleanup;
332
333      if (callback(cleanup_hook) != LDPS_OK)
334        return LDPS_ERR;
335    } break;
336    case LDPT_GET_INPUT_FILE:
337      get_input_file = tv->tv_u.tv_get_input_file;
338      break;
339    case LDPT_RELEASE_INPUT_FILE:
340      release_input_file = tv->tv_u.tv_release_input_file;
341      break;
342    case LDPT_ADD_SYMBOLS:
343      add_symbols = tv->tv_u.tv_add_symbols;
344      break;
345    case LDPT_GET_SYMBOLS_V2:
346      // Do not override get_symbols_v3 with get_symbols_v2.
347      if (!get_symbols)
348        get_symbols = tv->tv_u.tv_get_symbols;
349      break;
350    case LDPT_GET_SYMBOLS_V3:
351      get_symbols = tv->tv_u.tv_get_symbols;
352      break;
353    case LDPT_ADD_INPUT_FILE:
354      add_input_file = tv->tv_u.tv_add_input_file;
355      break;
356    case LDPT_SET_EXTRA_LIBRARY_PATH:
357      set_extra_library_path = tv->tv_u.tv_set_extra_library_path;
358      break;
359    case LDPT_GET_VIEW:
360      get_view = tv->tv_u.tv_get_view;
361      break;
362    case LDPT_MESSAGE:
363      message = tv->tv_u.tv_message;
364      break;
365    default:
366      break;
367    }
368  }
369
370  if (!registeredClaimFile) {
371    message(LDPL_ERROR, "register_claim_file not passed to LLVMgold.");
372    return LDPS_ERR;
373  }
374  if (!add_symbols) {
375    message(LDPL_ERROR, "add_symbols not passed to LLVMgold.");
376    return LDPS_ERR;
377  }
378
379  if (!RegisteredAllSymbolsRead)
380    return LDPS_OK;
381
382  if (!get_input_file) {
383    message(LDPL_ERROR, "get_input_file not passed to LLVMgold.");
384    return LDPS_ERR;
385  }
386  if (!release_input_file) {
387    message(LDPL_ERROR, "release_input_file not passed to LLVMgold.");
388    return LDPS_ERR;
389  }
390
391  return LDPS_OK;
392}
393
394static const GlobalObject *getBaseObject(const GlobalValue &GV) {
395  if (auto *GA = dyn_cast<GlobalAlias>(&GV))
396    return GA->getBaseObject();
397  return cast<GlobalObject>(&GV);
398}
399
400static bool shouldSkip(uint32_t Symflags) {
401  if (!(Symflags & object::BasicSymbolRef::SF_Global))
402    return true;
403  if (Symflags & object::BasicSymbolRef::SF_FormatSpecific)
404    return true;
405  return false;
406}
407
408static void diagnosticHandler(const DiagnosticInfo &DI) {
409  if (const auto *BDI = dyn_cast<BitcodeDiagnosticInfo>(&DI)) {
410    std::error_code EC = BDI->getError();
411    if (EC == BitcodeError::InvalidBitcodeSignature)
412      return;
413  }
414
415  std::string ErrStorage;
416  {
417    raw_string_ostream OS(ErrStorage);
418    DiagnosticPrinterRawOStream DP(OS);
419    DI.print(DP);
420  }
421  ld_plugin_level Level;
422  switch (DI.getSeverity()) {
423  case DS_Error:
424    message(LDPL_FATAL, "LLVM gold plugin has failed to create LTO module: %s",
425            ErrStorage.c_str());
426  case DS_Warning:
427    Level = LDPL_WARNING;
428    break;
429  case DS_Note:
430  case DS_Remark:
431    Level = LDPL_INFO;
432    break;
433  }
434  message(Level, "LLVM gold plugin: %s",  ErrStorage.c_str());
435}
436
437static void diagnosticHandlerForContext(const DiagnosticInfo &DI,
438                                        void *Context) {
439  diagnosticHandler(DI);
440}
441
442static GlobalValue::VisibilityTypes
443getMinVisibility(GlobalValue::VisibilityTypes A,
444                 GlobalValue::VisibilityTypes B) {
445  if (A == GlobalValue::HiddenVisibility)
446    return A;
447  if (B == GlobalValue::HiddenVisibility)
448    return B;
449  if (A == GlobalValue::ProtectedVisibility)
450    return A;
451  return B;
452}
453
454/// Called by gold to see whether this file is one that our plugin can handle.
455/// We'll try to open it and register all the symbols with add_symbol if
456/// possible.
457static ld_plugin_status claim_file_hook(const ld_plugin_input_file *file,
458                                        int *claimed) {
459  LLVMContext Context;
460  MemoryBufferRef BufferRef;
461  std::unique_ptr<MemoryBuffer> Buffer;
462  if (get_view) {
463    const void *view;
464    if (get_view(file->handle, &view) != LDPS_OK) {
465      message(LDPL_ERROR, "Failed to get a view of %s", file->name);
466      return LDPS_ERR;
467    }
468    BufferRef =
469        MemoryBufferRef(StringRef((const char *)view, file->filesize), "");
470  } else {
471    int64_t offset = 0;
472    // Gold has found what might be IR part-way inside of a file, such as
473    // an .a archive.
474    if (file->offset) {
475      offset = file->offset;
476    }
477    ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
478        MemoryBuffer::getOpenFileSlice(file->fd, file->name, file->filesize,
479                                       offset);
480    if (std::error_code EC = BufferOrErr.getError()) {
481      message(LDPL_ERROR, EC.message().c_str());
482      return LDPS_ERR;
483    }
484    Buffer = std::move(BufferOrErr.get());
485    BufferRef = Buffer->getMemBufferRef();
486  }
487
488  Context.setDiagnosticHandler(diagnosticHandlerForContext);
489  ErrorOr<std::unique_ptr<object::IRObjectFile>> ObjOrErr =
490      object::IRObjectFile::create(BufferRef, Context);
491  std::error_code EC = ObjOrErr.getError();
492  if (EC == object::object_error::invalid_file_type ||
493      EC == object::object_error::bitcode_section_not_found)
494    return LDPS_OK;
495
496  *claimed = 1;
497
498  if (EC) {
499    message(LDPL_ERROR, "LLVM gold plugin has failed to create LTO module: %s",
500            EC.message().c_str());
501    return LDPS_ERR;
502  }
503  std::unique_ptr<object::IRObjectFile> Obj = std::move(*ObjOrErr);
504
505  Modules.resize(Modules.size() + 1);
506  claimed_file &cf = Modules.back();
507
508  cf.handle = file->handle;
509  // Keep track of the first handle for each file descriptor, since there are
510  // multiple in the case of an archive. This is used later in the case of
511  // ThinLTO parallel backends to ensure that each file is only opened and
512  // released once.
513  auto LeaderHandle =
514      FDToLeaderHandle.insert(std::make_pair(file->fd, file->handle)).first;
515  cf.leader_handle = LeaderHandle->second;
516  // Save the filesize since for parallel ThinLTO backends we can only
517  // invoke get_input_file once per archive (only for the leader handle).
518  cf.filesize = file->filesize;
519  // In the case of an archive library, all but the first member must have a
520  // non-zero offset, which we can append to the file name to obtain a
521  // unique name.
522  cf.name = file->name;
523  if (file->offset)
524    cf.name += ".llvm." + std::to_string(file->offset) + "." +
525               sys::path::filename(Obj->getModule().getSourceFileName()).str();
526
527  for (auto &Sym : Obj->symbols()) {
528    uint32_t Symflags = Sym.getFlags();
529    if (shouldSkip(Symflags))
530      continue;
531
532    cf.syms.push_back(ld_plugin_symbol());
533    ld_plugin_symbol &sym = cf.syms.back();
534    sym.version = nullptr;
535
536    SmallString<64> Name;
537    {
538      raw_svector_ostream OS(Name);
539      Sym.printName(OS);
540    }
541    sym.name = strdup(Name.c_str());
542
543    const GlobalValue *GV = Obj->getSymbolGV(Sym.getRawDataRefImpl());
544
545    ResolutionInfo &Res = ResInfo[sym.name];
546
547    sym.visibility = LDPV_DEFAULT;
548    if (GV) {
549      Res.UnnamedAddr =
550          GlobalValue::getMinUnnamedAddr(Res.UnnamedAddr, GV->getUnnamedAddr());
551      Res.IsLinkonceOdr &= GV->hasLinkOnceLinkage();
552      Res.Visibility = getMinVisibility(Res.Visibility, GV->getVisibility());
553      switch (GV->getVisibility()) {
554      case GlobalValue::DefaultVisibility:
555        break;
556      case GlobalValue::HiddenVisibility:
557        sym.visibility = LDPV_HIDDEN;
558        break;
559      case GlobalValue::ProtectedVisibility:
560        sym.visibility = LDPV_PROTECTED;
561        break;
562      }
563    }
564
565    if (Symflags & object::BasicSymbolRef::SF_Undefined) {
566      sym.def = LDPK_UNDEF;
567      if (GV && GV->hasExternalWeakLinkage())
568        sym.def = LDPK_WEAKUNDEF;
569    } else {
570      sym.def = LDPK_DEF;
571      if (GV) {
572        assert(!GV->hasExternalWeakLinkage() &&
573               !GV->hasAvailableExternallyLinkage() && "Not a declaration!");
574        if (GV->hasCommonLinkage())
575          sym.def = LDPK_COMMON;
576        else if (GV->isWeakForLinker())
577          sym.def = LDPK_WEAKDEF;
578      }
579    }
580
581    sym.size = 0;
582    sym.comdat_key = nullptr;
583    if (GV) {
584      const GlobalObject *Base = getBaseObject(*GV);
585      if (!Base)
586        message(LDPL_FATAL, "Unable to determine comdat of alias!");
587      const Comdat *C = Base->getComdat();
588      if (C)
589        sym.comdat_key = strdup(C->getName().str().c_str());
590    }
591
592    sym.resolution = LDPR_UNKNOWN;
593  }
594
595  if (!cf.syms.empty()) {
596    if (add_symbols(cf.handle, cf.syms.size(), cf.syms.data()) != LDPS_OK) {
597      message(LDPL_ERROR, "Unable to add symbols!");
598      return LDPS_ERR;
599    }
600  }
601
602  return LDPS_OK;
603}
604
605static void internalize(GlobalValue &GV) {
606  if (GV.isDeclarationForLinker())
607    return; // We get here if there is a matching asm definition.
608  if (!GV.hasLocalLinkage())
609    GV.setLinkage(GlobalValue::InternalLinkage);
610}
611
612static const char *getResolutionName(ld_plugin_symbol_resolution R) {
613  switch (R) {
614  case LDPR_UNKNOWN:
615    return "UNKNOWN";
616  case LDPR_UNDEF:
617    return "UNDEF";
618  case LDPR_PREVAILING_DEF:
619    return "PREVAILING_DEF";
620  case LDPR_PREVAILING_DEF_IRONLY:
621    return "PREVAILING_DEF_IRONLY";
622  case LDPR_PREEMPTED_REG:
623    return "PREEMPTED_REG";
624  case LDPR_PREEMPTED_IR:
625    return "PREEMPTED_IR";
626  case LDPR_RESOLVED_IR:
627    return "RESOLVED_IR";
628  case LDPR_RESOLVED_EXEC:
629    return "RESOLVED_EXEC";
630  case LDPR_RESOLVED_DYN:
631    return "RESOLVED_DYN";
632  case LDPR_PREVAILING_DEF_IRONLY_EXP:
633    return "PREVAILING_DEF_IRONLY_EXP";
634  }
635  llvm_unreachable("Unknown resolution");
636}
637
638static void freeSymName(ld_plugin_symbol &Sym) {
639  free(Sym.name);
640  free(Sym.comdat_key);
641  Sym.name = nullptr;
642  Sym.comdat_key = nullptr;
643}
644
645/// Helper to get a file's symbols and a view into it via gold callbacks.
646static const void *getSymbolsAndView(claimed_file &F) {
647  ld_plugin_status status = get_symbols(F.handle, F.syms.size(), F.syms.data());
648  if (status == LDPS_NO_SYMS)
649    return nullptr;
650
651  if (status != LDPS_OK)
652    message(LDPL_FATAL, "Failed to get symbol information");
653
654  const void *View;
655  if (get_view(F.handle, &View) != LDPS_OK)
656    message(LDPL_FATAL, "Failed to get a view of file");
657
658  return View;
659}
660
661static std::unique_ptr<ModuleSummaryIndex>
662getModuleSummaryIndexForFile(claimed_file &F) {
663  const void *View = getSymbolsAndView(F);
664  if (!View)
665    return nullptr;
666
667  MemoryBufferRef BufferRef(StringRef((const char *)View, F.filesize), F.name);
668
669  // Don't bother trying to build an index if there is no summary information
670  // in this bitcode file.
671  if (!object::ModuleSummaryIndexObjectFile::hasGlobalValueSummaryInMemBuffer(
672          BufferRef, diagnosticHandler))
673    return std::unique_ptr<ModuleSummaryIndex>(nullptr);
674
675  ErrorOr<std::unique_ptr<object::ModuleSummaryIndexObjectFile>> ObjOrErr =
676      object::ModuleSummaryIndexObjectFile::create(BufferRef,
677                                                   diagnosticHandler);
678
679  if (std::error_code EC = ObjOrErr.getError())
680    message(LDPL_FATAL,
681            "Could not read module summary index bitcode from file : %s",
682            EC.message().c_str());
683
684  object::ModuleSummaryIndexObjectFile &Obj = **ObjOrErr;
685
686  return Obj.takeIndex();
687}
688
689static std::unique_ptr<Module>
690getModuleForFile(LLVMContext &Context, claimed_file &F, const void *View,
691                 StringRef Name, raw_fd_ostream *ApiFile,
692                 StringSet<> &Internalize, std::vector<GlobalValue *> &Keep,
693                 StringMap<unsigned> &Realign) {
694  MemoryBufferRef BufferRef(StringRef((const char *)View, F.filesize), Name);
695  ErrorOr<std::unique_ptr<object::IRObjectFile>> ObjOrErr =
696      object::IRObjectFile::create(BufferRef, Context);
697
698  if (std::error_code EC = ObjOrErr.getError())
699    message(LDPL_FATAL, "Could not read bitcode from file : %s",
700            EC.message().c_str());
701
702  object::IRObjectFile &Obj = **ObjOrErr;
703
704  Module &M = Obj.getModule();
705
706  M.materializeMetadata();
707  UpgradeDebugInfo(M);
708
709  SmallPtrSet<GlobalValue *, 8> Used;
710  collectUsedGlobalVariables(M, Used, /*CompilerUsed*/ false);
711
712  unsigned SymNum = 0;
713  for (auto &ObjSym : Obj.symbols()) {
714    GlobalValue *GV = Obj.getSymbolGV(ObjSym.getRawDataRefImpl());
715    if (GV && GV->hasAppendingLinkage())
716      Keep.push_back(GV);
717
718    if (shouldSkip(ObjSym.getFlags()))
719      continue;
720    ld_plugin_symbol &Sym = F.syms[SymNum];
721    ++SymNum;
722
723    ld_plugin_symbol_resolution Resolution =
724        (ld_plugin_symbol_resolution)Sym.resolution;
725
726    if (options::generate_api_file)
727      *ApiFile << Sym.name << ' ' << getResolutionName(Resolution) << '\n';
728
729    if (!GV) {
730      freeSymName(Sym);
731      continue; // Asm symbol.
732    }
733
734    ResolutionInfo &Res = ResInfo[Sym.name];
735    if (Resolution == LDPR_PREVAILING_DEF_IRONLY_EXP && !Res.IsLinkonceOdr)
736      Resolution = LDPR_PREVAILING_DEF;
737
738    GV->setUnnamedAddr(Res.UnnamedAddr);
739    GV->setVisibility(Res.Visibility);
740
741    // Override gold's resolution for common symbols. We want the largest
742    // one to win.
743    if (GV->hasCommonLinkage()) {
744      if (Resolution == LDPR_PREVAILING_DEF_IRONLY)
745        Res.CommonInternal = true;
746
747      if (Resolution == LDPR_PREVAILING_DEF_IRONLY ||
748          Resolution == LDPR_PREVAILING_DEF)
749        Res.UseCommon = true;
750
751      const DataLayout &DL = GV->getParent()->getDataLayout();
752      uint64_t Size = DL.getTypeAllocSize(GV->getType()->getElementType());
753      unsigned Align = GV->getAlignment();
754
755      if (Res.UseCommon && Size >= Res.CommonSize) {
756        // Take GV.
757        if (Res.CommonInternal)
758          Resolution = LDPR_PREVAILING_DEF_IRONLY;
759        else
760          Resolution = LDPR_PREVAILING_DEF;
761        cast<GlobalVariable>(GV)->setAlignment(
762            std::max(Res.CommonAlign, Align));
763      } else {
764        // Do not take GV, it's smaller than what we already have in the
765        // combined module.
766        Resolution = LDPR_PREEMPTED_IR;
767        if (Align > Res.CommonAlign)
768          // Need to raise the alignment though.
769          Realign[Sym.name] = Align;
770      }
771
772      Res.CommonSize = std::max(Res.CommonSize, Size);
773      Res.CommonAlign = std::max(Res.CommonAlign, Align);
774    }
775
776    switch (Resolution) {
777    case LDPR_UNKNOWN:
778      llvm_unreachable("Unexpected resolution");
779
780    case LDPR_RESOLVED_IR:
781    case LDPR_RESOLVED_EXEC:
782    case LDPR_RESOLVED_DYN:
783    case LDPR_PREEMPTED_IR:
784    case LDPR_PREEMPTED_REG:
785      break;
786
787    case LDPR_UNDEF:
788      if (!GV->isDeclarationForLinker())
789        assert(GV->hasComdat());
790      break;
791
792    case LDPR_PREVAILING_DEF_IRONLY: {
793      Keep.push_back(GV);
794      // The IR linker has to be able to map this value to a declaration,
795      // so we can only internalize after linking.
796      if (!Used.count(GV))
797        Internalize.insert(GV->getName());
798      break;
799    }
800
801    case LDPR_PREVAILING_DEF:
802      Keep.push_back(GV);
803      // There is a non IR use, so we have to force optimizations to keep this.
804      switch (GV->getLinkage()) {
805      default:
806        break;
807      case GlobalValue::LinkOnceAnyLinkage:
808        GV->setLinkage(GlobalValue::WeakAnyLinkage);
809        break;
810      case GlobalValue::LinkOnceODRLinkage:
811        GV->setLinkage(GlobalValue::WeakODRLinkage);
812        break;
813      }
814      break;
815
816    case LDPR_PREVAILING_DEF_IRONLY_EXP: {
817      Keep.push_back(GV);
818      if (canBeOmittedFromSymbolTable(GV))
819        Internalize.insert(GV->getName());
820      break;
821    }
822    }
823
824    freeSymName(Sym);
825  }
826
827  return Obj.takeModule();
828}
829
830static void saveBCFile(StringRef Path, Module &M) {
831  std::error_code EC;
832  raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None);
833  if (EC)
834    message(LDPL_FATAL, "Failed to write the output file.");
835  WriteBitcodeToFile(&M, OS, /* ShouldPreserveUseListOrder */ false);
836}
837
838static void recordFile(std::string Filename, bool TempOutFile) {
839  if (add_input_file(Filename.c_str()) != LDPS_OK)
840    message(LDPL_FATAL,
841            "Unable to add .o file to the link. File left behind in: %s",
842            Filename.c_str());
843  if (TempOutFile)
844    Cleanup.push_back(Filename.c_str());
845}
846
847void ThinLTOTaskInfo::cleanup() {
848  // Close the output file descriptor before we pass it to gold.
849  OS->close();
850
851  recordFile(Filename, TempOutFile);
852}
853
854namespace {
855/// Class to manage optimization and code generation for a module, possibly
856/// in a thread (ThinLTO).
857class CodeGen {
858  /// The module for which this will generate code.
859  std::unique_ptr<llvm::Module> M;
860
861  /// The output stream to generate code into.
862  raw_fd_ostream *OS;
863
864  /// The task ID when this was invoked in a thread (ThinLTO).
865  int TaskID;
866
867  /// The module summary index for ThinLTO tasks.
868  const ModuleSummaryIndex *CombinedIndex;
869
870  /// The target machine for generating code for this module.
871  std::unique_ptr<TargetMachine> TM;
872
873  /// Filename to use as base when save-temps is enabled, used to get
874  /// a unique and identifiable save-temps output file for each ThinLTO backend.
875  std::string SaveTempsFilename;
876
877  /// Map from a module name to the corresponding buffer holding a view of the
878  /// bitcode provided via the get_view gold callback.
879  StringMap<MemoryBufferRef> *ModuleMap;
880
881  // Functions to import into this module.
882  FunctionImporter::ImportMapTy *ImportList;
883
884  // Map of globals defined in this module to their summary.
885  std::map<GlobalValue::GUID, GlobalValueSummary *> *DefinedGlobals;
886
887public:
888  /// Constructor used by full LTO.
889  CodeGen(std::unique_ptr<llvm::Module> M)
890      : M(std::move(M)), OS(nullptr), TaskID(-1), CombinedIndex(nullptr),
891        ModuleMap(nullptr) {
892    initTargetMachine();
893  }
894  /// Constructor used by ThinLTO.
895  CodeGen(std::unique_ptr<llvm::Module> M, raw_fd_ostream *OS, int TaskID,
896          const ModuleSummaryIndex *CombinedIndex, std::string Filename,
897          StringMap<MemoryBufferRef> *ModuleMap,
898          FunctionImporter::ImportMapTy *ImportList,
899          std::map<GlobalValue::GUID, GlobalValueSummary *> *DefinedGlobals)
900      : M(std::move(M)), OS(OS), TaskID(TaskID), CombinedIndex(CombinedIndex),
901        SaveTempsFilename(std::move(Filename)), ModuleMap(ModuleMap),
902        ImportList(ImportList), DefinedGlobals(DefinedGlobals) {
903    assert(options::thinlto == !!CombinedIndex &&
904           "Expected module summary index iff performing ThinLTO");
905    initTargetMachine();
906  }
907
908  /// Invoke LTO passes and the code generator for the module.
909  void runAll();
910
911  /// Invoke the actual code generation to emit Module's object to file.
912  void runCodegenPasses();
913
914private:
915  const Target *TheTarget;
916  std::string TripleStr;
917  std::string FeaturesString;
918  TargetOptions Options;
919
920  /// Create a target machine for the module. Must be unique for each
921  /// module/task.
922  void initTargetMachine();
923
924  std::unique_ptr<TargetMachine> createTargetMachine();
925
926  /// Run all LTO passes on the module.
927  void runLTOPasses();
928
929  /// Sets up output files necessary to perform optional multi-threaded
930  /// split code generation, and invokes the code generation implementation.
931  /// If BCFileName is not empty, saves bitcode for module partitions into
932  /// {BCFileName}0 .. {BCFileName}N.
933  void runSplitCodeGen(const SmallString<128> &BCFilename);
934};
935}
936
937static SubtargetFeatures getFeatures(Triple &TheTriple) {
938  SubtargetFeatures Features;
939  Features.getDefaultSubtargetFeatures(TheTriple);
940  for (const std::string &A : MAttrs)
941    Features.AddFeature(A);
942  return Features;
943}
944
945static CodeGenOpt::Level getCGOptLevel() {
946  switch (options::OptLevel) {
947  case 0:
948    return CodeGenOpt::None;
949  case 1:
950    return CodeGenOpt::Less;
951  case 2:
952    return CodeGenOpt::Default;
953  case 3:
954    return CodeGenOpt::Aggressive;
955  }
956  llvm_unreachable("Invalid optimization level");
957}
958
959void CodeGen::initTargetMachine() {
960  TripleStr = M->getTargetTriple();
961  Triple TheTriple(TripleStr);
962
963  std::string ErrMsg;
964  TheTarget = TargetRegistry::lookupTarget(TripleStr, ErrMsg);
965  if (!TheTarget)
966    message(LDPL_FATAL, "Target not found: %s", ErrMsg.c_str());
967
968  SubtargetFeatures Features = getFeatures(TheTriple);
969  FeaturesString = Features.getString();
970  Options = InitTargetOptionsFromCodeGenFlags();
971
972  // Disable the new X86 relax relocations since gold might not support them.
973  // FIXME: Check the gold version or add a new option to enable them.
974  Options.RelaxELFRelocations = false;
975
976  TM = createTargetMachine();
977}
978
979std::unique_ptr<TargetMachine> CodeGen::createTargetMachine() {
980  CodeGenOpt::Level CGOptLevel = getCGOptLevel();
981
982  return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
983      TripleStr, options::mcpu, FeaturesString, Options, RelocationModel,
984      CodeModel::Default, CGOptLevel));
985}
986
987void CodeGen::runLTOPasses() {
988  M->setDataLayout(TM->createDataLayout());
989
990  if (CombinedIndex) {
991    // Apply summary-based LinkOnce/Weak resolution decisions.
992    thinLTOResolveWeakForLinkerModule(*M, *DefinedGlobals);
993
994    // Apply summary-based internalization decisions. Skip if there are no
995    // defined globals from the summary since not only is it unnecessary, but
996    // if this module did not have a summary section the internalizer will
997    // assert if it finds any definitions in this module that aren't in the
998    // DefinedGlobals set.
999    if (!DefinedGlobals->empty())
1000      thinLTOInternalizeModule(*M, *DefinedGlobals);
1001
1002    // Create a loader that will parse the bitcode from the buffers
1003    // in the ModuleMap.
1004    ModuleLoader Loader(M->getContext(), *ModuleMap);
1005
1006    // Perform function importing.
1007    FunctionImporter Importer(*CombinedIndex, Loader);
1008    Importer.importFunctions(*M, *ImportList);
1009  }
1010
1011  legacy::PassManager passes;
1012  passes.add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
1013
1014  PassManagerBuilder PMB;
1015  PMB.LibraryInfo = new TargetLibraryInfoImpl(Triple(TM->getTargetTriple()));
1016  PMB.Inliner = createFunctionInliningPass();
1017  // Unconditionally verify input since it is not verified before this
1018  // point and has unknown origin.
1019  PMB.VerifyInput = true;
1020  PMB.VerifyOutput = !options::DisableVerify;
1021  PMB.LoopVectorize = true;
1022  PMB.SLPVectorize = true;
1023  PMB.OptLevel = options::OptLevel;
1024  if (options::thinlto)
1025    PMB.populateThinLTOPassManager(passes);
1026  else
1027    PMB.populateLTOPassManager(passes);
1028  passes.run(*M);
1029}
1030
1031/// Open a file and return the new file descriptor given a base input
1032/// file name, a flag indicating whether a temp file should be generated,
1033/// and an optional task id. The new filename generated is
1034/// returned in \p NewFilename.
1035static int openOutputFile(SmallString<128> InFilename, bool TempOutFile,
1036                          SmallString<128> &NewFilename, int TaskID = -1) {
1037  int FD;
1038  if (TempOutFile) {
1039    std::error_code EC =
1040        sys::fs::createTemporaryFile("lto-llvm", "o", FD, NewFilename);
1041    if (EC)
1042      message(LDPL_FATAL, "Could not create temporary file: %s",
1043              EC.message().c_str());
1044  } else {
1045    NewFilename = InFilename;
1046    if (TaskID >= 0)
1047      NewFilename += utostr(TaskID);
1048    std::error_code EC =
1049        sys::fs::openFileForWrite(NewFilename, FD, sys::fs::F_None);
1050    if (EC)
1051      message(LDPL_FATAL, "Could not open file: %s", EC.message().c_str());
1052  }
1053  return FD;
1054}
1055
1056void CodeGen::runCodegenPasses() {
1057  assert(OS && "Output stream must be set before emitting to file");
1058  legacy::PassManager CodeGenPasses;
1059  if (TM->addPassesToEmitFile(CodeGenPasses, *OS,
1060                              TargetMachine::CGFT_ObjectFile))
1061    report_fatal_error("Failed to setup codegen");
1062  CodeGenPasses.run(*M);
1063}
1064
1065void CodeGen::runSplitCodeGen(const SmallString<128> &BCFilename) {
1066  SmallString<128> Filename;
1067  // Note that openOutputFile will append a unique ID for each task
1068  if (!options::obj_path.empty())
1069    Filename = options::obj_path;
1070  else if (options::TheOutputType == options::OT_SAVE_TEMPS)
1071    Filename = output_name + ".o";
1072
1073  // Note that the default parallelism is 1 instead of the
1074  // hardware_concurrency, as there are behavioral differences between
1075  // parallelism levels (e.g. symbol ordering will be different, and some uses
1076  // of inline asm currently have issues with parallelism >1).
1077  unsigned int MaxThreads = options::Parallelism ? options::Parallelism : 1;
1078
1079  std::vector<SmallString<128>> Filenames(MaxThreads);
1080  std::vector<SmallString<128>> BCFilenames(MaxThreads);
1081  bool TempOutFile = Filename.empty();
1082  {
1083    // Open a file descriptor for each backend task. This is done in a block
1084    // so that the output file descriptors are closed before gold opens them.
1085    std::list<llvm::raw_fd_ostream> OSs;
1086    std::vector<llvm::raw_pwrite_stream *> OSPtrs(MaxThreads);
1087    for (unsigned I = 0; I != MaxThreads; ++I) {
1088      int FD = openOutputFile(Filename, TempOutFile, Filenames[I],
1089                              // Only append ID if there are multiple tasks.
1090                              MaxThreads > 1 ? I : -1);
1091      OSs.emplace_back(FD, true);
1092      OSPtrs[I] = &OSs.back();
1093    }
1094
1095    std::list<llvm::raw_fd_ostream> BCOSs;
1096    std::vector<llvm::raw_pwrite_stream *> BCOSPtrs;
1097    if (!BCFilename.empty() && MaxThreads > 1) {
1098      for (unsigned I = 0; I != MaxThreads; ++I) {
1099        int FD = openOutputFile(BCFilename, false, BCFilenames[I], I);
1100        BCOSs.emplace_back(FD, true);
1101        BCOSPtrs.push_back(&BCOSs.back());
1102      }
1103    }
1104
1105    // Run backend tasks.
1106    splitCodeGen(std::move(M), OSPtrs, BCOSPtrs,
1107                 [&]() { return createTargetMachine(); });
1108  }
1109
1110  for (auto &Filename : Filenames)
1111    recordFile(Filename.c_str(), TempOutFile);
1112}
1113
1114void CodeGen::runAll() {
1115  runLTOPasses();
1116
1117  SmallString<128> OptFilename;
1118  if (options::TheOutputType == options::OT_SAVE_TEMPS) {
1119    OptFilename = output_name;
1120    // If the CodeGen client provided a filename, use it. Always expect
1121    // a provided filename if we are in a task (i.e. ThinLTO backend).
1122    assert(!SaveTempsFilename.empty() || TaskID == -1);
1123    if (!SaveTempsFilename.empty())
1124      OptFilename = SaveTempsFilename;
1125    OptFilename += ".opt.bc";
1126    saveBCFile(OptFilename, *M);
1127  }
1128
1129  // If we are already in a thread (i.e. ThinLTO), just perform
1130  // codegen passes directly.
1131  if (TaskID >= 0)
1132    runCodegenPasses();
1133  // Otherwise attempt split code gen.
1134  else
1135    runSplitCodeGen(OptFilename);
1136}
1137
1138/// Links the module in \p View from file \p F into the combined module
1139/// saved in the IRMover \p L.
1140static void linkInModule(LLVMContext &Context, IRMover &L, claimed_file &F,
1141                         const void *View, StringRef Name,
1142                         raw_fd_ostream *ApiFile, StringSet<> &Internalize,
1143                         bool SetName = false) {
1144  std::vector<GlobalValue *> Keep;
1145  StringMap<unsigned> Realign;
1146  std::unique_ptr<Module> M = getModuleForFile(Context, F, View, Name, ApiFile,
1147                                               Internalize, Keep, Realign);
1148  if (!M.get())
1149    return;
1150  if (!options::triple.empty())
1151    M->setTargetTriple(options::triple.c_str());
1152  else if (M->getTargetTriple().empty()) {
1153    M->setTargetTriple(DefaultTriple);
1154  }
1155
1156  // For ThinLTO we want to propagate the source file name to ensure
1157  // we can create the correct global identifiers matching those in the
1158  // original module.
1159  if (SetName)
1160    L.getModule().setSourceFileName(M->getSourceFileName());
1161
1162  if (Error E = L.move(std::move(M), Keep,
1163                       [](GlobalValue &, IRMover::ValueAdder) {})) {
1164    handleAllErrors(std::move(E), [&](const llvm::ErrorInfoBase &EIB) {
1165      message(LDPL_FATAL, "Failed to link module %s: %s", Name.str().c_str(),
1166              EIB.message().c_str());
1167    });
1168  }
1169
1170  for (const auto &I : Realign) {
1171    GlobalValue *Dst = L.getModule().getNamedValue(I.first());
1172    if (!Dst)
1173      continue;
1174    cast<GlobalVariable>(Dst)->setAlignment(I.second);
1175  }
1176}
1177
1178/// Perform the ThinLTO backend on a single module, invoking the LTO and codegen
1179/// pipelines.
1180static void thinLTOBackendTask(claimed_file &F, const void *View,
1181                               StringRef Name, raw_fd_ostream *ApiFile,
1182                               const ModuleSummaryIndex &CombinedIndex,
1183                               raw_fd_ostream *OS, unsigned TaskID,
1184                               StringMap<MemoryBufferRef> &ModuleMap,
1185                               FunctionImporter::ImportMapTy &ImportList,
1186                               std::map<GlobalValue::GUID, GlobalValueSummary *> &DefinedGlobals) {
1187  // Need to use a separate context for each task
1188  LLVMContext Context;
1189  Context.setDiscardValueNames(options::TheOutputType !=
1190                               options::OT_SAVE_TEMPS);
1191  Context.enableDebugTypeODRUniquing(); // Merge debug info types.
1192  Context.setDiagnosticHandler(diagnosticHandlerForContext, nullptr, true);
1193
1194  std::unique_ptr<llvm::Module> NewModule(new llvm::Module(Name, Context));
1195  IRMover L(*NewModule.get());
1196
1197  StringSet<> Dummy;
1198  linkInModule(Context, L, F, View, Name, ApiFile, Dummy, true);
1199  if (renameModuleForThinLTO(*NewModule, CombinedIndex))
1200    message(LDPL_FATAL, "Failed to rename module for ThinLTO");
1201
1202  CodeGen codeGen(std::move(NewModule), OS, TaskID, &CombinedIndex, Name,
1203                  &ModuleMap, &ImportList, &DefinedGlobals);
1204  codeGen.runAll();
1205}
1206
1207/// Launch each module's backend pipeline in a separate task in a thread pool.
1208static void
1209thinLTOBackends(raw_fd_ostream *ApiFile,
1210                const ModuleSummaryIndex &CombinedIndex,
1211                StringMap<MemoryBufferRef> &ModuleMap,
1212                StringMap<FunctionImporter::ImportMapTy> &ImportLists,
1213  StringMap<std::map<GlobalValue::GUID, GlobalValueSummary *>>
1214      &ModuleToDefinedGVSummaries) {
1215  unsigned TaskCount = 0;
1216  std::vector<ThinLTOTaskInfo> Tasks;
1217  Tasks.reserve(Modules.size());
1218  unsigned int MaxThreads = options::Parallelism
1219                                ? options::Parallelism
1220                                : thread::hardware_concurrency();
1221
1222  // Create ThreadPool in nested scope so that threads will be joined
1223  // on destruction.
1224  {
1225    ThreadPool ThinLTOThreadPool(MaxThreads);
1226    for (claimed_file &F : Modules) {
1227      // Do all the gold callbacks in the main thread, since gold is not thread
1228      // safe by default.
1229      const void *View = getSymbolsAndView(F);
1230      if (!View)
1231        continue;
1232
1233      SmallString<128> Filename;
1234      if (!options::obj_path.empty())
1235        // Note that openOutputFile will append a unique ID for each task
1236        Filename = options::obj_path;
1237      else if (options::TheOutputType == options::OT_SAVE_TEMPS) {
1238        // Use the input file name so that we get a unique and identifiable
1239        // output file for each ThinLTO backend task.
1240        Filename = F.name;
1241        Filename += ".thinlto.o";
1242      }
1243      bool TempOutFile = Filename.empty();
1244
1245      SmallString<128> NewFilename;
1246      int FD = openOutputFile(Filename, TempOutFile, NewFilename,
1247                              // Only append the TaskID if we will use the
1248                              // non-unique obj_path.
1249                              !options::obj_path.empty() ? TaskCount : -1);
1250      TaskCount++;
1251      std::unique_ptr<raw_fd_ostream> OS =
1252          llvm::make_unique<raw_fd_ostream>(FD, true);
1253
1254      // Enqueue the task
1255      ThinLTOThreadPool.async(thinLTOBackendTask, std::ref(F), View, F.name,
1256                              ApiFile, std::ref(CombinedIndex), OS.get(),
1257                              TaskCount, std::ref(ModuleMap),
1258                              std::ref(ImportLists[F.name]),
1259                              std::ref(ModuleToDefinedGVSummaries[F.name]));
1260
1261      // Record the information needed by the task or during its cleanup
1262      // to a ThinLTOTaskInfo instance. For information needed by the task
1263      // the unique_ptr ownership is transferred to the ThinLTOTaskInfo.
1264      Tasks.emplace_back(std::move(OS), NewFilename.c_str(), TempOutFile);
1265    }
1266  }
1267
1268  for (auto &Task : Tasks)
1269    Task.cleanup();
1270}
1271
1272/// Parse the thinlto_prefix_replace option into the \p OldPrefix and
1273/// \p NewPrefix strings, if it was specified.
1274static void getThinLTOOldAndNewPrefix(std::string &OldPrefix,
1275                                      std::string &NewPrefix) {
1276  StringRef PrefixReplace = options::thinlto_prefix_replace;
1277  assert(PrefixReplace.empty() || PrefixReplace.find(";") != StringRef::npos);
1278  std::pair<StringRef, StringRef> Split = PrefixReplace.split(";");
1279  OldPrefix = Split.first.str();
1280  NewPrefix = Split.second.str();
1281}
1282
1283/// Given the original \p Path to an output file, replace any path
1284/// prefix matching \p OldPrefix with \p NewPrefix. Also, create the
1285/// resulting directory if it does not yet exist.
1286static std::string getThinLTOOutputFile(const std::string &Path,
1287                                        const std::string &OldPrefix,
1288                                        const std::string &NewPrefix) {
1289  if (OldPrefix.empty() && NewPrefix.empty())
1290    return Path;
1291  SmallString<128> NewPath(Path);
1292  llvm::sys::path::replace_path_prefix(NewPath, OldPrefix, NewPrefix);
1293  StringRef ParentPath = llvm::sys::path::parent_path(NewPath.str());
1294  if (!ParentPath.empty()) {
1295    // Make sure the new directory exists, creating it if necessary.
1296    if (std::error_code EC = llvm::sys::fs::create_directories(ParentPath))
1297      llvm::errs() << "warning: could not create directory '" << ParentPath
1298                   << "': " << EC.message() << '\n';
1299  }
1300  return NewPath.str();
1301}
1302
1303/// Perform ThinLTO link, which creates the combined index file.
1304/// Also, either launch backend threads or (under thinlto-index-only)
1305/// emit individual index files for distributed backends and exit.
1306static ld_plugin_status thinLTOLink(raw_fd_ostream *ApiFile) {
1307  // Map from a module name to the corresponding buffer holding a view of the
1308  // bitcode provided via the get_view gold callback.
1309  StringMap<MemoryBufferRef> ModuleMap;
1310  // Map to own RAII objects that manage the file opening and releasing
1311  // interfaces with gold.
1312  DenseMap<void *, std::unique_ptr<PluginInputFile>> HandleToInputFile;
1313
1314  // Keep track of symbols that must not be internalized because they
1315  // are referenced outside of a single IR module.
1316  DenseSet<GlobalValue::GUID> Preserve;
1317
1318  // Keep track of the prevailing copy for each GUID, for use in resolving
1319  // weak linkages.
1320  DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy;
1321
1322  ModuleSummaryIndex CombinedIndex;
1323  uint64_t NextModuleId = 0;
1324  for (claimed_file &F : Modules) {
1325    if (!HandleToInputFile.count(F.leader_handle))
1326      HandleToInputFile.insert(std::make_pair(
1327          F.leader_handle, llvm::make_unique<PluginInputFile>(F.handle)));
1328    // Pass this into getModuleSummaryIndexForFile
1329    const void *View = getSymbolsAndView(F);
1330    if (!View)
1331      continue;
1332
1333    MemoryBufferRef ModuleBuffer(StringRef((const char *)View, F.filesize),
1334                                 F.name);
1335    assert(ModuleMap.find(ModuleBuffer.getBufferIdentifier()) ==
1336               ModuleMap.end() &&
1337           "Expect unique Buffer Identifier");
1338    ModuleMap[ModuleBuffer.getBufferIdentifier()] = ModuleBuffer;
1339
1340    std::unique_ptr<ModuleSummaryIndex> Index = getModuleSummaryIndexForFile(F);
1341
1342    // Use gold's symbol resolution information to identify symbols referenced
1343    // by more than a single IR module (i.e. referenced by multiple IR modules
1344    // or by a non-IR module). Cross references introduced by importing are
1345    // checked separately via the export lists. Also track the prevailing copy
1346    // for later symbol resolution.
1347    for (auto &Sym : F.syms) {
1348      ld_plugin_symbol_resolution Resolution =
1349          (ld_plugin_symbol_resolution)Sym.resolution;
1350      GlobalValue::GUID SymGUID = GlobalValue::getGUID(Sym.name);
1351      if (Resolution != LDPR_PREVAILING_DEF_IRONLY)
1352        Preserve.insert(SymGUID);
1353
1354      if (Index && (Resolution == LDPR_PREVAILING_DEF ||
1355                    Resolution == LDPR_PREVAILING_DEF_IRONLY ||
1356                    Resolution == LDPR_PREVAILING_DEF_IRONLY_EXP))
1357        PrevailingCopy[SymGUID] = Index->getGlobalValueSummary(SymGUID);
1358    }
1359
1360    // Skip files without a module summary.
1361    if (Index)
1362      CombinedIndex.mergeFrom(std::move(Index), ++NextModuleId);
1363  }
1364
1365  // Collect for each module the list of function it defines (GUID ->
1366  // Summary).
1367  StringMap<std::map<GlobalValue::GUID, GlobalValueSummary *>>
1368      ModuleToDefinedGVSummaries(NextModuleId);
1369  CombinedIndex.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
1370
1371  StringMap<FunctionImporter::ImportMapTy> ImportLists(NextModuleId);
1372  StringMap<FunctionImporter::ExportSetTy> ExportLists(NextModuleId);
1373  ComputeCrossModuleImport(CombinedIndex, ModuleToDefinedGVSummaries,
1374                           ImportLists, ExportLists);
1375
1376  auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) {
1377    const auto &Prevailing = PrevailingCopy.find(GUID);
1378    assert(Prevailing != PrevailingCopy.end());
1379    return Prevailing->second == S;
1380  };
1381
1382  // Callback for internalization, to prevent internalization of symbols
1383  // that were not candidates initially, and those that are being imported
1384  // (which introduces new cross references).
1385  auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
1386    const auto &ExportList = ExportLists.find(ModuleIdentifier);
1387    return (ExportList != ExportLists.end() &&
1388            ExportList->second.count(GUID)) ||
1389           Preserve.count(GUID);
1390  };
1391
1392  thinLTOResolveWeakForLinkerInIndex(
1393      CombinedIndex, isPrevailing,
1394      [](StringRef ModuleIdentifier, GlobalValue::GUID GUID,
1395         GlobalValue::LinkageTypes NewLinkage) {});
1396
1397  // Use global summary-based analysis to identify symbols that can be
1398  // internalized (because they aren't exported or preserved as per callback).
1399  // Changes are made in the index, consumed in the ThinLTO backends.
1400  thinLTOInternalizeAndPromoteInIndex(CombinedIndex, isExported);
1401
1402  if (options::thinlto_emit_imports_files && !options::thinlto_index_only)
1403    message(LDPL_WARNING,
1404            "thinlto-emit-imports-files ignored unless thinlto-index-only");
1405
1406  if (options::thinlto_index_only) {
1407    // If the thinlto-prefix-replace option was specified, parse it and
1408    // extract the old and new prefixes.
1409    std::string OldPrefix, NewPrefix;
1410    getThinLTOOldAndNewPrefix(OldPrefix, NewPrefix);
1411
1412    // For each input bitcode file, generate an individual index that
1413    // contains summaries only for its own global values, and for any that
1414    // should be imported.
1415    for (claimed_file &F : Modules) {
1416      std::error_code EC;
1417
1418      std::string NewModulePath =
1419          getThinLTOOutputFile(F.name, OldPrefix, NewPrefix);
1420      raw_fd_ostream OS((Twine(NewModulePath) + ".thinlto.bc").str(), EC,
1421                        sys::fs::OpenFlags::F_None);
1422      if (EC)
1423        message(LDPL_FATAL, "Unable to open %s.thinlto.bc for writing: %s",
1424                NewModulePath.c_str(), EC.message().c_str());
1425      // Build a map of module to the GUIDs and summary objects that should
1426      // be written to its index.
1427      std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex;
1428      gatherImportedSummariesForModule(F.name, ModuleToDefinedGVSummaries,
1429                                       ImportLists, ModuleToSummariesForIndex);
1430      WriteIndexToFile(CombinedIndex, OS, &ModuleToSummariesForIndex);
1431
1432      if (options::thinlto_emit_imports_files) {
1433        if ((EC = EmitImportsFiles(F.name,
1434                                   (Twine(NewModulePath) + ".imports").str(),
1435                                   ImportLists)))
1436          message(LDPL_FATAL, "Unable to open %s.imports",
1437                  NewModulePath.c_str(), EC.message().c_str());
1438      }
1439    }
1440
1441    cleanup_hook();
1442    exit(0);
1443  }
1444
1445  // Create OS in nested scope so that it will be closed on destruction.
1446  {
1447    std::error_code EC;
1448    raw_fd_ostream OS(output_name + ".thinlto.bc", EC,
1449                      sys::fs::OpenFlags::F_None);
1450    if (EC)
1451      message(LDPL_FATAL, "Unable to open %s.thinlto.bc for writing: %s",
1452              output_name.data(), EC.message().c_str());
1453    WriteIndexToFile(CombinedIndex, OS);
1454  }
1455
1456  thinLTOBackends(ApiFile, CombinedIndex, ModuleMap, ImportLists,
1457                  ModuleToDefinedGVSummaries);
1458  return LDPS_OK;
1459}
1460
1461/// gold informs us that all symbols have been read. At this point, we use
1462/// get_symbols to see if any of our definitions have been overridden by a
1463/// native object file. Then, perform optimization and codegen.
1464static ld_plugin_status allSymbolsReadHook(raw_fd_ostream *ApiFile) {
1465  if (Modules.empty())
1466    return LDPS_OK;
1467
1468  if (unsigned NumOpts = options::extra.size())
1469    cl::ParseCommandLineOptions(NumOpts, &options::extra[0]);
1470
1471  if (options::thinlto)
1472    return thinLTOLink(ApiFile);
1473
1474  LLVMContext Context;
1475  Context.setDiscardValueNames(options::TheOutputType !=
1476                               options::OT_SAVE_TEMPS);
1477  Context.enableDebugTypeODRUniquing(); // Merge debug info types.
1478  Context.setDiagnosticHandler(diagnosticHandlerForContext, nullptr, true);
1479
1480  std::unique_ptr<Module> Combined(new Module("ld-temp.o", Context));
1481  IRMover L(*Combined);
1482
1483  StringSet<> Internalize;
1484  for (claimed_file &F : Modules) {
1485    // RAII object to manage the file opening and releasing interfaces with
1486    // gold.
1487    PluginInputFile InputFile(F.handle);
1488    const void *View = getSymbolsAndView(F);
1489    if (!View)
1490      continue;
1491    linkInModule(Context, L, F, View, F.name, ApiFile, Internalize);
1492  }
1493
1494  for (const auto &Name : Internalize) {
1495    GlobalValue *GV = Combined->getNamedValue(Name.first());
1496    if (GV)
1497      internalize(*GV);
1498  }
1499
1500  if (options::TheOutputType == options::OT_DISABLE)
1501    return LDPS_OK;
1502
1503  if (options::TheOutputType != options::OT_NORMAL) {
1504    std::string path;
1505    if (options::TheOutputType == options::OT_BC_ONLY)
1506      path = output_name;
1507    else
1508      path = output_name + ".bc";
1509    saveBCFile(path, *Combined);
1510    if (options::TheOutputType == options::OT_BC_ONLY)
1511      return LDPS_OK;
1512  }
1513
1514  CodeGen codeGen(std::move(Combined));
1515  codeGen.runAll();
1516
1517  if (!options::extra_library_path.empty() &&
1518      set_extra_library_path(options::extra_library_path.c_str()) != LDPS_OK)
1519    message(LDPL_FATAL, "Unable to set the extra library path.");
1520
1521  return LDPS_OK;
1522}
1523
1524static ld_plugin_status all_symbols_read_hook(void) {
1525  ld_plugin_status Ret;
1526  if (!options::generate_api_file) {
1527    Ret = allSymbolsReadHook(nullptr);
1528  } else {
1529    std::error_code EC;
1530    raw_fd_ostream ApiFile("apifile.txt", EC, sys::fs::F_None);
1531    if (EC)
1532      message(LDPL_FATAL, "Unable to open apifile.txt for writing: %s",
1533              EC.message().c_str());
1534    Ret = allSymbolsReadHook(&ApiFile);
1535  }
1536
1537  llvm_shutdown();
1538
1539  if (options::TheOutputType == options::OT_BC_ONLY ||
1540      options::TheOutputType == options::OT_DISABLE) {
1541    if (options::TheOutputType == options::OT_DISABLE) {
1542      // Remove the output file here since ld.bfd creates the output file
1543      // early.
1544      std::error_code EC = sys::fs::remove(output_name);
1545      if (EC)
1546        message(LDPL_ERROR, "Failed to delete '%s': %s", output_name.c_str(),
1547                EC.message().c_str());
1548    }
1549    exit(0);
1550  }
1551
1552  return Ret;
1553}
1554
1555static ld_plugin_status cleanup_hook(void) {
1556  for (std::string &Name : Cleanup) {
1557    std::error_code EC = sys::fs::remove(Name);
1558    if (EC)
1559      message(LDPL_ERROR, "Failed to delete '%s': %s", Name.c_str(),
1560              EC.message().c_str());
1561  }
1562
1563  return LDPS_OK;
1564}
1565