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/Config/config.h" // plugin-api.h requires HAVE_STDINT_H
16#include "plugin-api.h"
17
18#include "llvm-c/lto.h"
19
20#include "llvm/ADT/OwningPtr.h"
21#include "llvm/Support/system_error.h"
22#include "llvm/Support/MemoryBuffer.h"
23#include "llvm/Support/ToolOutputFile.h"
24#include "llvm/Support/Errno.h"
25#include "llvm/Support/Path.h"
26#include "llvm/Support/Program.h"
27
28#include <cerrno>
29#include <cstdlib>
30#include <cstring>
31#include <fstream>
32#include <list>
33#include <vector>
34
35// Support Windows/MinGW crazyness.
36#ifdef _WIN32
37# include <io.h>
38# define lseek _lseek
39# define read _read
40#endif
41
42using namespace llvm;
43
44namespace {
45  ld_plugin_status discard_message(int level, const char *format, ...) {
46    // Die loudly. Recent versions of Gold pass ld_plugin_message as the first
47    // callback in the transfer vector. This should never be called.
48    abort();
49  }
50
51  ld_plugin_add_symbols add_symbols = NULL;
52  ld_plugin_get_symbols get_symbols = NULL;
53  ld_plugin_add_input_file add_input_file = NULL;
54  ld_plugin_add_input_library add_input_library = NULL;
55  ld_plugin_set_extra_library_path set_extra_library_path = NULL;
56  ld_plugin_get_view get_view = NULL;
57  ld_plugin_message message = discard_message;
58
59  int api_version = 0;
60  int gold_version = 0;
61
62  struct claimed_file {
63    void *handle;
64    std::vector<ld_plugin_symbol> syms;
65  };
66
67  lto_codegen_model output_type = LTO_CODEGEN_PIC_MODEL_STATIC;
68  std::string output_name = "";
69  std::list<claimed_file> Modules;
70  std::vector<sys::Path> Cleanup;
71  lto_code_gen_t code_gen = NULL;
72}
73
74namespace options {
75  enum generate_bc { BC_NO, BC_ALSO, BC_ONLY };
76  static bool generate_api_file = false;
77  static generate_bc generate_bc_file = BC_NO;
78  static std::string bc_path;
79  static std::string obj_path;
80  static std::string extra_library_path;
81  static std::string triple;
82  static std::string mcpu;
83  // Additional options to pass into the code generator.
84  // Note: This array will contain all plugin options which are not claimed
85  // as plugin exclusive to pass to the code generator.
86  // For example, "generate-api-file" and "as"options are for the plugin
87  // use only and will not be passed.
88  static std::vector<std::string> extra;
89
90  static void process_plugin_option(const char* opt_)
91  {
92    if (opt_ == NULL)
93      return;
94    llvm::StringRef opt = opt_;
95
96    if (opt == "generate-api-file") {
97      generate_api_file = true;
98    } else if (opt.startswith("mcpu=")) {
99      mcpu = opt.substr(strlen("mcpu="));
100    } else if (opt.startswith("extra-library-path=")) {
101      extra_library_path = opt.substr(strlen("extra_library_path="));
102    } else if (opt.startswith("mtriple=")) {
103      triple = opt.substr(strlen("mtriple="));
104    } else if (opt.startswith("obj-path=")) {
105      obj_path = opt.substr(strlen("obj-path="));
106    } else if (opt == "emit-llvm") {
107      generate_bc_file = BC_ONLY;
108    } else if (opt == "also-emit-llvm") {
109      generate_bc_file = BC_ALSO;
110    } else if (opt.startswith("also-emit-llvm=")) {
111      llvm::StringRef path = opt.substr(strlen("also-emit-llvm="));
112      generate_bc_file = BC_ALSO;
113      if (!bc_path.empty()) {
114        (*message)(LDPL_WARNING, "Path to the output IL file specified twice. "
115                   "Discarding %s", opt_);
116      } else {
117        bc_path = path;
118      }
119    } else {
120      // Save this option to pass to the code generator.
121      extra.push_back(opt);
122    }
123  }
124}
125
126static ld_plugin_status claim_file_hook(const ld_plugin_input_file *file,
127                                        int *claimed);
128static ld_plugin_status all_symbols_read_hook(void);
129static ld_plugin_status cleanup_hook(void);
130
131extern "C" ld_plugin_status onload(ld_plugin_tv *tv);
132ld_plugin_status onload(ld_plugin_tv *tv) {
133  // We're given a pointer to the first transfer vector. We read through them
134  // until we find one where tv_tag == LDPT_NULL. The REGISTER_* tagged values
135  // contain pointers to functions that we need to call to register our own
136  // hooks. The others are addresses of functions we can use to call into gold
137  // for services.
138
139  bool registeredClaimFile = false;
140
141  for (; tv->tv_tag != LDPT_NULL; ++tv) {
142    switch (tv->tv_tag) {
143      case LDPT_API_VERSION:
144        api_version = tv->tv_u.tv_val;
145        break;
146      case LDPT_GOLD_VERSION:  // major * 100 + minor
147        gold_version = tv->tv_u.tv_val;
148        break;
149      case LDPT_OUTPUT_NAME:
150        output_name = tv->tv_u.tv_string;
151        break;
152      case LDPT_LINKER_OUTPUT:
153        switch (tv->tv_u.tv_val) {
154          case LDPO_REL:  // .o
155          case LDPO_DYN:  // .so
156          // FIXME: Replace 3 with LDPO_PIE once that is in a released binutils.
157          case 3: // position independent executable
158            output_type = LTO_CODEGEN_PIC_MODEL_DYNAMIC;
159            break;
160          case LDPO_EXEC:  // .exe
161            output_type = LTO_CODEGEN_PIC_MODEL_STATIC;
162            break;
163          default:
164            (*message)(LDPL_ERROR, "Unknown output file type %d",
165                       tv->tv_u.tv_val);
166            return LDPS_ERR;
167        }
168        // TODO: add an option to disable PIC.
169        //output_type = LTO_CODEGEN_PIC_MODEL_DYNAMIC_NO_PIC;
170        break;
171      case LDPT_OPTION:
172        options::process_plugin_option(tv->tv_u.tv_string);
173        break;
174      case LDPT_REGISTER_CLAIM_FILE_HOOK: {
175        ld_plugin_register_claim_file callback;
176        callback = tv->tv_u.tv_register_claim_file;
177
178        if ((*callback)(claim_file_hook) != LDPS_OK)
179          return LDPS_ERR;
180
181        registeredClaimFile = true;
182      } break;
183      case LDPT_REGISTER_ALL_SYMBOLS_READ_HOOK: {
184        ld_plugin_register_all_symbols_read callback;
185        callback = tv->tv_u.tv_register_all_symbols_read;
186
187        if ((*callback)(all_symbols_read_hook) != LDPS_OK)
188          return LDPS_ERR;
189
190        code_gen = lto_codegen_create();
191      } break;
192      case LDPT_REGISTER_CLEANUP_HOOK: {
193        ld_plugin_register_cleanup callback;
194        callback = tv->tv_u.tv_register_cleanup;
195
196        if ((*callback)(cleanup_hook) != LDPS_OK)
197          return LDPS_ERR;
198      } break;
199      case LDPT_ADD_SYMBOLS:
200        add_symbols = tv->tv_u.tv_add_symbols;
201        break;
202      case LDPT_GET_SYMBOLS:
203        get_symbols = tv->tv_u.tv_get_symbols;
204        break;
205      case LDPT_ADD_INPUT_FILE:
206        add_input_file = tv->tv_u.tv_add_input_file;
207        break;
208      case LDPT_ADD_INPUT_LIBRARY:
209        add_input_library = tv->tv_u.tv_add_input_file;
210        break;
211      case LDPT_SET_EXTRA_LIBRARY_PATH:
212        set_extra_library_path = tv->tv_u.tv_set_extra_library_path;
213        break;
214      case LDPT_GET_VIEW:
215        get_view = tv->tv_u.tv_get_view;
216        break;
217      case LDPT_MESSAGE:
218        message = tv->tv_u.tv_message;
219        break;
220      default:
221        break;
222    }
223  }
224
225  if (!registeredClaimFile) {
226    (*message)(LDPL_ERROR, "register_claim_file not passed to LLVMgold.");
227    return LDPS_ERR;
228  }
229  if (!add_symbols) {
230    (*message)(LDPL_ERROR, "add_symbols not passed to LLVMgold.");
231    return LDPS_ERR;
232  }
233
234  return LDPS_OK;
235}
236
237/// claim_file_hook - called by gold to see whether this file is one that
238/// our plugin can handle. We'll try to open it and register all the symbols
239/// with add_symbol if possible.
240static ld_plugin_status claim_file_hook(const ld_plugin_input_file *file,
241                                        int *claimed) {
242  lto_module_t M;
243  const void *view;
244  OwningPtr<MemoryBuffer> buffer;
245  if (get_view) {
246    if (get_view(file->handle, &view) != LDPS_OK) {
247      (*message)(LDPL_ERROR, "Failed to get a view of %s", file->name);
248      return LDPS_ERR;
249    }
250  } else {
251    int64_t offset = 0;
252    // Gold has found what might be IR part-way inside of a file, such as
253    // an .a archive.
254    if (file->offset) {
255      offset = file->offset;
256    }
257    if (error_code ec =
258        MemoryBuffer::getOpenFile(file->fd, file->name, buffer, file->filesize,
259                                  -1, offset, false)) {
260      (*message)(LDPL_ERROR, ec.message().c_str());
261      return LDPS_ERR;
262    }
263    view = buffer->getBufferStart();
264  }
265
266  if (!lto_module_is_object_file_in_memory(view, file->filesize))
267    return LDPS_OK;
268
269  M = lto_module_create_from_memory(view, file->filesize);
270  if (!M) {
271    if (const char* msg = lto_get_error_message()) {
272      (*message)(LDPL_ERROR,
273                 "LLVM gold plugin has failed to create LTO module: %s",
274                 msg);
275      return LDPS_ERR;
276    }
277    return LDPS_OK;
278  }
279
280  *claimed = 1;
281  Modules.resize(Modules.size() + 1);
282  claimed_file &cf = Modules.back();
283
284  if (!options::triple.empty())
285    lto_module_set_target_triple(M, options::triple.c_str());
286
287  cf.handle = file->handle;
288  unsigned sym_count = lto_module_get_num_symbols(M);
289  cf.syms.reserve(sym_count);
290
291  for (unsigned i = 0; i != sym_count; ++i) {
292    lto_symbol_attributes attrs = lto_module_get_symbol_attribute(M, i);
293    if ((attrs & LTO_SYMBOL_SCOPE_MASK) == LTO_SYMBOL_SCOPE_INTERNAL)
294      continue;
295
296    cf.syms.push_back(ld_plugin_symbol());
297    ld_plugin_symbol &sym = cf.syms.back();
298    sym.name = const_cast<char *>(lto_module_get_symbol_name(M, i));
299    sym.name = strdup(sym.name);
300    sym.version = NULL;
301
302    int scope = attrs & LTO_SYMBOL_SCOPE_MASK;
303    switch (scope) {
304      case LTO_SYMBOL_SCOPE_HIDDEN:
305        sym.visibility = LDPV_HIDDEN;
306        break;
307      case LTO_SYMBOL_SCOPE_PROTECTED:
308        sym.visibility = LDPV_PROTECTED;
309        break;
310      case 0: // extern
311      case LTO_SYMBOL_SCOPE_DEFAULT:
312        sym.visibility = LDPV_DEFAULT;
313        break;
314      default:
315        (*message)(LDPL_ERROR, "Unknown scope attribute: %d", scope);
316        return LDPS_ERR;
317    }
318
319    int definition = attrs & LTO_SYMBOL_DEFINITION_MASK;
320    sym.comdat_key = NULL;
321    switch (definition) {
322      case LTO_SYMBOL_DEFINITION_REGULAR:
323        sym.def = LDPK_DEF;
324        break;
325      case LTO_SYMBOL_DEFINITION_UNDEFINED:
326        sym.def = LDPK_UNDEF;
327        break;
328      case LTO_SYMBOL_DEFINITION_TENTATIVE:
329        sym.def = LDPK_COMMON;
330        break;
331      case LTO_SYMBOL_DEFINITION_WEAK:
332        sym.comdat_key = sym.name;
333        sym.def = LDPK_WEAKDEF;
334        break;
335      case LTO_SYMBOL_DEFINITION_WEAKUNDEF:
336        sym.def = LDPK_WEAKUNDEF;
337        break;
338      default:
339        (*message)(LDPL_ERROR, "Unknown definition attribute: %d", definition);
340        return LDPS_ERR;
341    }
342
343    sym.size = 0;
344
345    sym.resolution = LDPR_UNKNOWN;
346  }
347
348  cf.syms.reserve(cf.syms.size());
349
350  if (!cf.syms.empty()) {
351    if ((*add_symbols)(cf.handle, cf.syms.size(), &cf.syms[0]) != LDPS_OK) {
352      (*message)(LDPL_ERROR, "Unable to add symbols!");
353      return LDPS_ERR;
354    }
355  }
356
357  if (code_gen)
358    lto_codegen_add_module(code_gen, M);
359
360  lto_module_dispose(M);
361
362  return LDPS_OK;
363}
364
365/// all_symbols_read_hook - gold informs us that all symbols have been read.
366/// At this point, we use get_symbols to see if any of our definitions have
367/// been overridden by a native object file. Then, perform optimization and
368/// codegen.
369static ld_plugin_status all_symbols_read_hook(void) {
370  std::ofstream api_file;
371  assert(code_gen);
372
373  if (options::generate_api_file) {
374    api_file.open("apifile.txt", std::ofstream::out | std::ofstream::trunc);
375    if (!api_file.is_open()) {
376      (*message)(LDPL_FATAL, "Unable to open apifile.txt for writing.");
377      abort();
378    }
379  }
380
381  // If we don't preserve any symbols, libLTO will assume that all symbols are
382  // needed. Keep all symbols unless we're producing a final executable.
383  bool anySymbolsPreserved = false;
384  for (std::list<claimed_file>::iterator I = Modules.begin(),
385         E = Modules.end(); I != E; ++I) {
386    if (I->syms.empty())
387      continue;
388    (*get_symbols)(I->handle, I->syms.size(), &I->syms[0]);
389    for (unsigned i = 0, e = I->syms.size(); i != e; i++) {
390      if (I->syms[i].resolution == LDPR_PREVAILING_DEF) {
391        lto_codegen_add_must_preserve_symbol(code_gen, I->syms[i].name);
392        anySymbolsPreserved = true;
393
394        if (options::generate_api_file)
395          api_file << I->syms[i].name << "\n";
396      }
397    }
398  }
399
400  if (options::generate_api_file)
401    api_file.close();
402
403  if (!anySymbolsPreserved) {
404    // All of the IL is unnecessary!
405    lto_codegen_dispose(code_gen);
406    return LDPS_OK;
407  }
408
409  lto_codegen_set_pic_model(code_gen, output_type);
410  lto_codegen_set_debug_model(code_gen, LTO_DEBUG_MODEL_DWARF);
411  if (!options::mcpu.empty())
412    lto_codegen_set_cpu(code_gen, options::mcpu.c_str());
413
414  // Pass through extra options to the code generator.
415  if (!options::extra.empty()) {
416    for (std::vector<std::string>::iterator it = options::extra.begin();
417         it != options::extra.end(); ++it) {
418      lto_codegen_debug_options(code_gen, (*it).c_str());
419    }
420  }
421
422  if (options::generate_bc_file != options::BC_NO) {
423    std::string path;
424    if (options::generate_bc_file == options::BC_ONLY)
425      path = output_name;
426    else if (!options::bc_path.empty())
427      path = options::bc_path;
428    else
429      path = output_name + ".bc";
430    bool err = lto_codegen_write_merged_modules(code_gen, path.c_str());
431    if (err)
432      (*message)(LDPL_FATAL, "Failed to write the output file.");
433    if (options::generate_bc_file == options::BC_ONLY)
434      exit(0);
435  }
436  const char *objPath;
437  if (lto_codegen_compile_to_file(code_gen, &objPath)) {
438    (*message)(LDPL_ERROR, "Could not produce a combined object file\n");
439  }
440
441  lto_codegen_dispose(code_gen);
442  for (std::list<claimed_file>::iterator I = Modules.begin(),
443         E = Modules.end(); I != E; ++I) {
444    for (unsigned i = 0; i != I->syms.size(); ++i) {
445      ld_plugin_symbol &sym = I->syms[i];
446      free(sym.name);
447    }
448  }
449
450  if ((*add_input_file)(objPath) != LDPS_OK) {
451    (*message)(LDPL_ERROR, "Unable to add .o file to the link.");
452    (*message)(LDPL_ERROR, "File left behind in: %s", objPath);
453    return LDPS_ERR;
454  }
455
456  if (!options::extra_library_path.empty() &&
457      set_extra_library_path(options::extra_library_path.c_str()) != LDPS_OK) {
458    (*message)(LDPL_ERROR, "Unable to set the extra library path.");
459    return LDPS_ERR;
460  }
461
462  if (options::obj_path.empty())
463    Cleanup.push_back(sys::Path(objPath));
464
465  return LDPS_OK;
466}
467
468static ld_plugin_status cleanup_hook(void) {
469  std::string ErrMsg;
470
471  for (int i = 0, e = Cleanup.size(); i != e; ++i)
472    if (Cleanup[i].eraseFromDisk(false, &ErrMsg))
473      (*message)(LDPL_ERROR, "Failed to delete '%s': %s", Cleanup[i].c_str(),
474                 ErrMsg.c_str());
475
476  return LDPS_OK;
477}
478