1 // icf.cc -- Identical Code Folding. 2 // 3 // Copyright (C) 2009-2014 Free Software Foundation, Inc. 4 // Written by Sriraman Tallam <tmsriram (at) google.com>. 5 6 // This file is part of gold. 7 8 // This program is free software; you can redistribute it and/or modify 9 // it under the terms of the GNU General Public License as published by 10 // the Free Software Foundation; either version 3 of the License, or 11 // (at your option) any later version. 12 13 // This program is distributed in the hope that it will be useful, 14 // but WITHOUT ANY WARRANTY; without even the implied warranty of 15 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 16 // GNU General Public License for more details. 17 18 // You should have received a copy of the GNU General Public License 19 // along with this program; if not, write to the Free Software 20 // Foundation, Inc., 51 Franklin Street - Fifth Floor, Boston, 21 // MA 02110-1301, USA. 22 23 // Identical Code Folding Algorithm 24 // ---------------------------------- 25 // Detecting identical functions is done here and the basic algorithm 26 // is as follows. A checksum is computed on each foldable section using 27 // its contents and relocations. If the symbol name corresponding to 28 // a relocation is known it is used to compute the checksum. If the 29 // symbol name is not known the stringified name of the object and the 30 // section number pointed to by the relocation is used. The checksums 31 // are stored as keys in a hash map and a section is identical to some 32 // other section if its checksum is already present in the hash map. 33 // Checksum collisions are handled by using a multimap and explicitly 34 // checking the contents when two sections have the same checksum. 35 // 36 // However, two functions A and B with identical text but with 37 // relocations pointing to different foldable sections can be identical if 38 // the corresponding foldable sections to which their relocations point to 39 // turn out to be identical. Hence, this checksumming process must be 40 // done repeatedly until convergence is obtained. Here is an example for 41 // the following case : 42 // 43 // int funcA () int funcB () 44 // { { 45 // return foo(); return goo(); 46 // } } 47 // 48 // The functions funcA and funcB are identical if functions foo() and 49 // goo() are identical. 50 // 51 // Hence, as described above, we repeatedly do the checksumming, 52 // assigning identical functions to the same group, until convergence is 53 // obtained. Now, we have two different ways to do this depending on how 54 // we initialize. 55 // 56 // Algorithm I : 57 // ----------- 58 // We can start with marking all functions as different and repeatedly do 59 // the checksumming. This has the advantage that we do not need to wait 60 // for convergence. We can stop at any point and correctness will be 61 // guaranteed although not all cases would have been found. However, this 62 // has a problem that some cases can never be found even if it is run until 63 // convergence. Here is an example with mutually recursive functions : 64 // 65 // int funcA (int a) int funcB (int a) 66 // { { 67 // if (a == 1) if (a == 1) 68 // return 1; return 1; 69 // return 1 + funcB(a - 1); return 1 + funcA(a - 1); 70 // } } 71 // 72 // In this example funcA and funcB are identical and one of them could be 73 // folded into the other. However, if we start with assuming that funcA 74 // and funcB are not identical, the algorithm, even after it is run to 75 // convergence, cannot detect that they are identical. It should be noted 76 // that even if the functions were self-recursive, Algorithm I cannot catch 77 // that they are identical, at least as is. 78 // 79 // Algorithm II : 80 // ------------ 81 // Here we start with marking all functions as identical and then repeat 82 // the checksumming until convergence. This can detect the above case 83 // mentioned above. It can detect all cases that Algorithm I can and more. 84 // However, the caveat is that it has to be run to convergence. It cannot 85 // be stopped arbitrarily like Algorithm I as correctness cannot be 86 // guaranteed. Algorithm II is not implemented. 87 // 88 // Algorithm I is used because experiments show that about three 89 // iterations are more than enough to achieve convergence. Algorithm I can 90 // handle recursive calls if it is changed to use a special common symbol 91 // for recursive relocs. This seems to be the most common case that 92 // Algorithm I could not catch as is. Mutually recursive calls are not 93 // frequent and Algorithm I wins because of its ability to be stopped 94 // arbitrarily. 95 // 96 // Caveat with using function pointers : 97 // ------------------------------------ 98 // 99 // Programs using function pointer comparisons/checks should use function 100 // folding with caution as the result of such comparisons could be different 101 // when folding takes place. This could lead to unexpected run-time 102 // behaviour. 103 // 104 // Safe Folding : 105 // ------------ 106 // 107 // ICF in safe mode folds only ctors and dtors if their function pointers can 108 // never be taken. Also, for X86-64, safe folding uses the relocation 109 // type to determine if a function's pointer is taken or not and only folds 110 // functions whose pointers are definitely not taken. 111 // 112 // Caveat with safe folding : 113 // ------------------------ 114 // 115 // This applies only to x86_64. 116 // 117 // Position independent executables are created from PIC objects (compiled 118 // with -fPIC) and/or PIE objects (compiled with -fPIE). For PIE objects, the 119 // relocation types for function pointer taken and a call are the same. 120 // Now, it is not always possible to tell if an object used in the link of 121 // a pie executable is a PIC object or a PIE object. Hence, for pie 122 // executables, using relocation types to disambiguate function pointers is 123 // currently disabled. 124 // 125 // Further, it is not correct to use safe folding to build non-pie 126 // executables using PIC/PIE objects. PIC/PIE objects have different 127 // relocation types for function pointers than non-PIC objects, and the 128 // current implementation of safe folding does not handle those relocation 129 // types. Hence, if used, functions whose pointers are taken could still be 130 // folded causing unpredictable run-time behaviour if the pointers were used 131 // in comparisons. 132 // 133 // 134 // 135 // How to run : --icf=[safe|all|none] 136 // Optional parameters : --icf-iterations <num> --print-icf-sections 137 // 138 // Performance : Less than 20 % link-time overhead on industry strength 139 // applications. Up to 6 % text size reductions. 140 141 #include "gold.h" 142 #include "object.h" 143 #include "gc.h" 144 #include "icf.h" 145 #include "symtab.h" 146 #include "libiberty.h" 147 #include "demangle.h" 148 #include "elfcpp.h" 149 #include "int_encoding.h" 150 151 namespace gold 152 { 153 154 // This function determines if a section or a group of identical 155 // sections has unique contents. Such unique sections or groups can be 156 // declared final and need not be processed any further. 157 // Parameters : 158 // ID_SECTION : Vector mapping a section index to a Section_id pair. 159 // IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical 160 // sections is already known to be unique. 161 // SECTION_CONTENTS : Contains the section's text and relocs to sections 162 // that cannot be folded. SECTION_CONTENTS are NULL 163 // implies that this function is being called for the 164 // first time before the first iteration of icf. 165 166 static void 167 preprocess_for_unique_sections(const std::vector<Section_id>& id_section, 168 std::vector<bool>* is_secn_or_group_unique, 169 std::vector<std::string>* section_contents) 170 { 171 Unordered_map<uint32_t, unsigned int> uniq_map; 172 std::pair<Unordered_map<uint32_t, unsigned int>::iterator, bool> 173 uniq_map_insert; 174 175 for (unsigned int i = 0; i < id_section.size(); i++) 176 { 177 if ((*is_secn_or_group_unique)[i]) 178 continue; 179 180 uint32_t cksum; 181 Section_id secn = id_section[i]; 182 section_size_type plen; 183 if (section_contents == NULL) 184 { 185 // Lock the object so we can read from it. This is only called 186 // single-threaded from queue_middle_tasks, so it is OK to lock. 187 // Unfortunately we have no way to pass in a Task token. 188 const Task* dummy_task = reinterpret_cast<const Task*>(-1); 189 Task_lock_obj<Object> tl(dummy_task, secn.first); 190 const unsigned char* contents; 191 contents = secn.first->section_contents(secn.second, 192 &plen, 193 false); 194 cksum = xcrc32(contents, plen, 0xffffffff); 195 } 196 else 197 { 198 const unsigned char* contents_array = reinterpret_cast 199 <const unsigned char*>((*section_contents)[i].c_str()); 200 cksum = xcrc32(contents_array, (*section_contents)[i].length(), 201 0xffffffff); 202 } 203 uniq_map_insert = uniq_map.insert(std::make_pair(cksum, i)); 204 if (uniq_map_insert.second) 205 { 206 (*is_secn_or_group_unique)[i] = true; 207 } 208 else 209 { 210 (*is_secn_or_group_unique)[i] = false; 211 (*is_secn_or_group_unique)[uniq_map_insert.first->second] = false; 212 } 213 } 214 } 215 216 // For SHF_MERGE sections that use REL relocations, the addend is stored in 217 // the text section at the relocation offset. Read the addend value given 218 // the pointer to the addend in the text section and the addend size. 219 // Update the addend value if a valid addend is found. 220 // Parameters: 221 // RELOC_ADDEND_PTR : Pointer to the addend in the text section. 222 // ADDEND_SIZE : The size of the addend. 223 // RELOC_ADDEND_VALUE : Pointer to the addend that is updated. 224 225 inline void 226 get_rel_addend(const unsigned char* reloc_addend_ptr, 227 const unsigned int addend_size, 228 uint64_t* reloc_addend_value) 229 { 230 switch (addend_size) 231 { 232 case 0: 233 break; 234 case 1: 235 *reloc_addend_value = 236 read_from_pointer<8>(reloc_addend_ptr); 237 break; 238 case 2: 239 *reloc_addend_value = 240 read_from_pointer<16>(reloc_addend_ptr); 241 break; 242 case 4: 243 *reloc_addend_value = 244 read_from_pointer<32>(reloc_addend_ptr); 245 break; 246 case 8: 247 *reloc_addend_value = 248 read_from_pointer<64>(reloc_addend_ptr); 249 break; 250 default: 251 gold_unreachable(); 252 } 253 } 254 255 // This returns the buffer containing the section's contents, both 256 // text and relocs. Relocs are differentiated as those pointing to 257 // sections that could be folded and those that cannot. Only relocs 258 // pointing to sections that could be folded are recomputed on 259 // subsequent invocations of this function. 260 // Parameters : 261 // FIRST_ITERATION : true if it is the first invocation. 262 // SECN : Section for which contents are desired. 263 // SECTION_NUM : Unique section number of this section. 264 // NUM_TRACKED_RELOCS : Vector reference to store the number of relocs 265 // to ICF sections. 266 // KEPT_SECTION_ID : Vector which maps folded sections to kept sections. 267 // SECTION_CONTENTS : Store the section's text and relocs to non-ICF 268 // sections. 269 270 static std::string 271 get_section_contents(bool first_iteration, 272 const Section_id& secn, 273 unsigned int section_num, 274 unsigned int* num_tracked_relocs, 275 Symbol_table* symtab, 276 const std::vector<unsigned int>& kept_section_id, 277 std::vector<std::string>* section_contents) 278 { 279 // Lock the object so we can read from it. This is only called 280 // single-threaded from queue_middle_tasks, so it is OK to lock. 281 // Unfortunately we have no way to pass in a Task token. 282 const Task* dummy_task = reinterpret_cast<const Task*>(-1); 283 Task_lock_obj<Object> tl(dummy_task, secn.first); 284 285 section_size_type plen; 286 const unsigned char* contents = NULL; 287 if (first_iteration) 288 contents = secn.first->section_contents(secn.second, &plen, false); 289 290 // The buffer to hold all the contents including relocs. A checksum 291 // is then computed on this buffer. 292 std::string buffer; 293 std::string icf_reloc_buffer; 294 295 if (num_tracked_relocs) 296 *num_tracked_relocs = 0; 297 298 Icf::Reloc_info_list& reloc_info_list = 299 symtab->icf()->reloc_info_list(); 300 301 Icf::Reloc_info_list::iterator it_reloc_info_list = 302 reloc_info_list.find(secn); 303 304 buffer.clear(); 305 icf_reloc_buffer.clear(); 306 307 // Process relocs and put them into the buffer. 308 309 if (it_reloc_info_list != reloc_info_list.end()) 310 { 311 Icf::Sections_reachable_info &v = 312 (it_reloc_info_list->second).section_info; 313 // Stores the information of the symbol pointed to by the reloc. 314 const Icf::Symbol_info &s = (it_reloc_info_list->second).symbol_info; 315 // Stores the addend and the symbol value. 316 Icf::Addend_info &a = (it_reloc_info_list->second).addend_info; 317 // Stores the offset of the reloc. 318 const Icf::Offset_info &o = (it_reloc_info_list->second).offset_info; 319 const Icf::Reloc_addend_size_info &reloc_addend_size_info = 320 (it_reloc_info_list->second).reloc_addend_size_info; 321 Icf::Sections_reachable_info::iterator it_v = v.begin(); 322 Icf::Symbol_info::const_iterator it_s = s.begin(); 323 Icf::Addend_info::iterator it_a = a.begin(); 324 Icf::Offset_info::const_iterator it_o = o.begin(); 325 Icf::Reloc_addend_size_info::const_iterator it_addend_size = 326 reloc_addend_size_info.begin(); 327 328 for (; it_v != v.end(); ++it_v, ++it_s, ++it_a, ++it_o, ++it_addend_size) 329 { 330 if (first_iteration 331 && it_v->first != NULL) 332 { 333 Symbol_location loc; 334 loc.object = it_v->first; 335 loc.shndx = it_v->second; 336 loc.offset = convert_types<off_t, long long>(it_a->first 337 + it_a->second); 338 // Look through function descriptors 339 parameters->target().function_location(&loc); 340 if (loc.shndx != it_v->second) 341 { 342 it_v->second = loc.shndx; 343 // Modify symvalue/addend to the code entry. 344 it_a->first = loc.offset; 345 it_a->second = 0; 346 } 347 } 348 349 // ADDEND_STR stores the symbol value and addend and offset, 350 // each at most 16 hex digits long. it_a points to a pair 351 // where first is the symbol value and second is the 352 // addend. 353 char addend_str[50]; 354 355 // It would be nice if we could use format macros in inttypes.h 356 // here but there are not in ISO/IEC C++ 1998. 357 snprintf(addend_str, sizeof(addend_str), "%llx %llx %llux", 358 static_cast<long long>((*it_a).first), 359 static_cast<long long>((*it_a).second), 360 static_cast<unsigned long long>(*it_o)); 361 362 // If the symbol pointed to by the reloc is not in an ordinary 363 // section or if the symbol type is not FROM_OBJECT, then the 364 // object is NULL. 365 if (it_v->first == NULL) 366 { 367 if (first_iteration) 368 { 369 // If the symbol name is available, use it. 370 if ((*it_s) != NULL) 371 buffer.append((*it_s)->name()); 372 // Append the addend. 373 buffer.append(addend_str); 374 buffer.append("@"); 375 } 376 continue; 377 } 378 379 Section_id reloc_secn(it_v->first, it_v->second); 380 381 // If this reloc turns back and points to the same section, 382 // like a recursive call, use a special symbol to mark this. 383 if (reloc_secn.first == secn.first 384 && reloc_secn.second == secn.second) 385 { 386 if (first_iteration) 387 { 388 buffer.append("R"); 389 buffer.append(addend_str); 390 buffer.append("@"); 391 } 392 continue; 393 } 394 Icf::Uniq_secn_id_map& section_id_map = 395 symtab->icf()->section_to_int_map(); 396 Icf::Uniq_secn_id_map::iterator section_id_map_it = 397 section_id_map.find(reloc_secn); 398 bool is_sym_preemptible = (*it_s != NULL 399 && !(*it_s)->is_from_dynobj() 400 && !(*it_s)->is_undefined() 401 && (*it_s)->is_preemptible()); 402 if (!is_sym_preemptible 403 && section_id_map_it != section_id_map.end()) 404 { 405 // This is a reloc to a section that might be folded. 406 if (num_tracked_relocs) 407 (*num_tracked_relocs)++; 408 409 char kept_section_str[10]; 410 unsigned int secn_id = section_id_map_it->second; 411 snprintf(kept_section_str, sizeof(kept_section_str), "%u", 412 kept_section_id[secn_id]); 413 if (first_iteration) 414 { 415 buffer.append("ICF_R"); 416 buffer.append(addend_str); 417 } 418 icf_reloc_buffer.append(kept_section_str); 419 // Append the addend. 420 icf_reloc_buffer.append(addend_str); 421 icf_reloc_buffer.append("@"); 422 } 423 else 424 { 425 // This is a reloc to a section that cannot be folded. 426 // Process it only in the first iteration. 427 if (!first_iteration) 428 continue; 429 430 uint64_t secn_flags = (it_v->first)->section_flags(it_v->second); 431 // This reloc points to a merge section. Hash the 432 // contents of this section. 433 if ((secn_flags & elfcpp::SHF_MERGE) != 0 434 && parameters->target().can_icf_inline_merge_sections()) 435 { 436 uint64_t entsize = 437 (it_v->first)->section_entsize(it_v->second); 438 long long offset = it_a->first; 439 // Handle SHT_RELA and SHT_REL addends, only one of these 440 // addends exists. 441 // Get the SHT_RELA addend. For RELA relocations, we have 442 // the addend from the relocation. 443 uint64_t reloc_addend_value = it_a->second; 444 445 // Handle SHT_REL addends. 446 // For REL relocations, we need to fetch the addend from the 447 // section contents. 448 const unsigned char* reloc_addend_ptr = 449 contents + static_cast<unsigned long long>(*it_o); 450 451 // Update the addend value with the SHT_REL addend if 452 // available. 453 get_rel_addend(reloc_addend_ptr, *it_addend_size, 454 &reloc_addend_value); 455 456 // Ignore the addend when it is a negative value. See the 457 // comments in Merged_symbol_value::value in object.h. 458 if (reloc_addend_value < 0xffffff00) 459 offset = offset + reloc_addend_value; 460 461 section_size_type secn_len; 462 463 const unsigned char* str_contents = 464 (it_v->first)->section_contents(it_v->second, 465 &secn_len, 466 false) + offset; 467 gold_assert (offset < (long long) secn_len); 468 469 if ((secn_flags & elfcpp::SHF_STRINGS) != 0) 470 { 471 // String merge section. 472 const char* str_char = 473 reinterpret_cast<const char*>(str_contents); 474 switch(entsize) 475 { 476 case 1: 477 { 478 buffer.append(str_char); 479 break; 480 } 481 case 2: 482 { 483 const uint16_t* ptr_16 = 484 reinterpret_cast<const uint16_t*>(str_char); 485 unsigned int strlen_16 = 0; 486 // Find the NULL character. 487 while(*(ptr_16 + strlen_16) != 0) 488 strlen_16++; 489 buffer.append(str_char, strlen_16 * 2); 490 } 491 break; 492 case 4: 493 { 494 const uint32_t* ptr_32 = 495 reinterpret_cast<const uint32_t*>(str_char); 496 unsigned int strlen_32 = 0; 497 // Find the NULL character. 498 while(*(ptr_32 + strlen_32) != 0) 499 strlen_32++; 500 buffer.append(str_char, strlen_32 * 4); 501 } 502 break; 503 default: 504 gold_unreachable(); 505 } 506 } 507 else 508 { 509 // Use the entsize to determine the length to copy. 510 uint64_t bufsize = entsize; 511 // If entsize is too big, copy all the remaining bytes. 512 if ((offset + entsize) > secn_len) 513 bufsize = secn_len - offset; 514 buffer.append(reinterpret_cast<const 515 char*>(str_contents), 516 bufsize); 517 } 518 buffer.append("@"); 519 } 520 else if ((*it_s) != NULL) 521 { 522 // If symbol name is available use that. 523 buffer.append((*it_s)->name()); 524 // Append the addend. 525 buffer.append(addend_str); 526 buffer.append("@"); 527 } 528 else 529 { 530 // Symbol name is not available, like for a local symbol, 531 // use object and section id. 532 buffer.append(it_v->first->name()); 533 char secn_id[10]; 534 snprintf(secn_id, sizeof(secn_id), "%u",it_v->second); 535 buffer.append(secn_id); 536 // Append the addend. 537 buffer.append(addend_str); 538 buffer.append("@"); 539 } 540 } 541 } 542 } 543 544 if (first_iteration) 545 { 546 buffer.append("Contents = "); 547 buffer.append(reinterpret_cast<const char*>(contents), plen); 548 // Store the section contents that dont change to avoid recomputing 549 // during the next call to this function. 550 (*section_contents)[section_num] = buffer; 551 } 552 else 553 { 554 gold_assert(buffer.empty()); 555 // Reuse the contents computed in the previous iteration. 556 buffer.append((*section_contents)[section_num]); 557 } 558 559 buffer.append(icf_reloc_buffer); 560 return buffer; 561 } 562 563 // This function computes a checksum on each section to detect and form 564 // groups of identical sections. The first iteration does this for all 565 // sections. 566 // Further iterations do this only for the kept sections from each group to 567 // determine if larger groups of identical sections could be formed. The 568 // first section in each group is the kept section for that group. 569 // 570 // CRC32 is the checksumming algorithm and can have collisions. That is, 571 // two sections with different contents can have the same checksum. Hence, 572 // a multimap is used to maintain more than one group of checksum 573 // identical sections. A section is added to a group only after its 574 // contents are explicitly compared with the kept section of the group. 575 // 576 // Parameters : 577 // ITERATION_NUM : Invocation instance of this function. 578 // NUM_TRACKED_RELOCS : Vector reference to store the number of relocs 579 // to ICF sections. 580 // KEPT_SECTION_ID : Vector which maps folded sections to kept sections. 581 // ID_SECTION : Vector mapping a section to an unique integer. 582 // IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical 583 // sections is already known to be unique. 584 // SECTION_CONTENTS : Store the section's text and relocs to non-ICF 585 // sections. 586 587 static bool 588 match_sections(unsigned int iteration_num, 589 Symbol_table* symtab, 590 std::vector<unsigned int>* num_tracked_relocs, 591 std::vector<unsigned int>* kept_section_id, 592 const std::vector<Section_id>& id_section, 593 const std::vector<uint64_t>& section_addraligns, 594 std::vector<bool>* is_secn_or_group_unique, 595 std::vector<std::string>* section_contents) 596 { 597 Unordered_multimap<uint32_t, unsigned int> section_cksum; 598 std::pair<Unordered_multimap<uint32_t, unsigned int>::iterator, 599 Unordered_multimap<uint32_t, unsigned int>::iterator> key_range; 600 bool converged = true; 601 602 if (iteration_num == 1) 603 preprocess_for_unique_sections(id_section, 604 is_secn_or_group_unique, 605 NULL); 606 else 607 preprocess_for_unique_sections(id_section, 608 is_secn_or_group_unique, 609 section_contents); 610 611 std::vector<std::string> full_section_contents; 612 613 for (unsigned int i = 0; i < id_section.size(); i++) 614 { 615 full_section_contents.push_back(""); 616 if ((*is_secn_or_group_unique)[i]) 617 continue; 618 619 Section_id secn = id_section[i]; 620 std::string this_secn_contents; 621 uint32_t cksum; 622 if (iteration_num == 1) 623 { 624 unsigned int num_relocs = 0; 625 this_secn_contents = get_section_contents(true, secn, i, &num_relocs, 626 symtab, (*kept_section_id), 627 section_contents); 628 (*num_tracked_relocs)[i] = num_relocs; 629 } 630 else 631 { 632 if ((*kept_section_id)[i] != i) 633 { 634 // This section is already folded into something. 635 continue; 636 } 637 this_secn_contents = get_section_contents(false, secn, i, NULL, 638 symtab, (*kept_section_id), 639 section_contents); 640 } 641 642 const unsigned char* this_secn_contents_array = 643 reinterpret_cast<const unsigned char*>(this_secn_contents.c_str()); 644 cksum = xcrc32(this_secn_contents_array, this_secn_contents.length(), 645 0xffffffff); 646 size_t count = section_cksum.count(cksum); 647 648 if (count == 0) 649 { 650 // Start a group with this cksum. 651 section_cksum.insert(std::make_pair(cksum, i)); 652 full_section_contents[i] = this_secn_contents; 653 } 654 else 655 { 656 key_range = section_cksum.equal_range(cksum); 657 Unordered_multimap<uint32_t, unsigned int>::iterator it; 658 // Search all the groups with this cksum for a match. 659 for (it = key_range.first; it != key_range.second; ++it) 660 { 661 unsigned int kept_section = it->second; 662 if (full_section_contents[kept_section].length() 663 != this_secn_contents.length()) 664 continue; 665 if (memcmp(full_section_contents[kept_section].c_str(), 666 this_secn_contents.c_str(), 667 this_secn_contents.length()) != 0) 668 continue; 669 670 // Check section alignment here. 671 // The section with the larger alignment requirement 672 // should be kept. We assume alignment can only be 673 // zero or postive integral powers of two. 674 uint64_t align_i = section_addraligns[i]; 675 uint64_t align_kept = section_addraligns[kept_section]; 676 if (align_i <= align_kept) 677 { 678 (*kept_section_id)[i] = kept_section; 679 } 680 else 681 { 682 (*kept_section_id)[kept_section] = i; 683 it->second = i; 684 full_section_contents[kept_section].swap( 685 full_section_contents[i]); 686 } 687 688 converged = false; 689 break; 690 } 691 if (it == key_range.second) 692 { 693 // Create a new group for this cksum. 694 section_cksum.insert(std::make_pair(cksum, i)); 695 full_section_contents[i] = this_secn_contents; 696 } 697 } 698 // If there are no relocs to foldable sections do not process 699 // this section any further. 700 if (iteration_num == 1 && (*num_tracked_relocs)[i] == 0) 701 (*is_secn_or_group_unique)[i] = true; 702 } 703 704 // If a section was folded into another section that was later folded 705 // again then the former has to be updated. 706 for (unsigned int i = 0; i < id_section.size(); i++) 707 { 708 // Find the end of the folding chain 709 unsigned int kept = i; 710 while ((*kept_section_id)[kept] != kept) 711 { 712 kept = (*kept_section_id)[kept]; 713 } 714 // Update every element of the chain 715 unsigned int current = i; 716 while ((*kept_section_id)[current] != kept) 717 { 718 unsigned int next = (*kept_section_id)[current]; 719 (*kept_section_id)[current] = kept; 720 current = next; 721 } 722 } 723 724 return converged; 725 } 726 727 // During safe icf (--icf=safe), only fold functions that are ctors or dtors. 728 // This function returns true if the section name is that of a ctor or a dtor. 729 730 static bool 731 is_function_ctor_or_dtor(const std::string& section_name) 732 { 733 const char* mangled_func_name = strrchr(section_name.c_str(), '.'); 734 gold_assert(mangled_func_name != NULL); 735 if ((is_prefix_of("._ZN", mangled_func_name) 736 || is_prefix_of("._ZZ", mangled_func_name)) 737 && (is_gnu_v3_mangled_ctor(mangled_func_name + 1) 738 || is_gnu_v3_mangled_dtor(mangled_func_name + 1))) 739 { 740 return true; 741 } 742 return false; 743 } 744 745 // This is the main ICF function called in gold.cc. This does the 746 // initialization and calls match_sections repeatedly (twice by default) 747 // which computes the crc checksums and detects identical functions. 748 749 void 750 Icf::find_identical_sections(const Input_objects* input_objects, 751 Symbol_table* symtab) 752 { 753 unsigned int section_num = 0; 754 std::vector<unsigned int> num_tracked_relocs; 755 std::vector<uint64_t> section_addraligns; 756 std::vector<bool> is_secn_or_group_unique; 757 std::vector<std::string> section_contents; 758 const Target& target = parameters->target(); 759 760 // Decide which sections are possible candidates first. 761 762 for (Input_objects::Relobj_iterator p = input_objects->relobj_begin(); 763 p != input_objects->relobj_end(); 764 ++p) 765 { 766 // Lock the object so we can read from it. This is only called 767 // single-threaded from queue_middle_tasks, so it is OK to lock. 768 // Unfortunately we have no way to pass in a Task token. 769 const Task* dummy_task = reinterpret_cast<const Task*>(-1); 770 Task_lock_obj<Object> tl(dummy_task, *p); 771 772 for (unsigned int i = 0;i < (*p)->shnum(); ++i) 773 { 774 const std::string section_name = (*p)->section_name(i); 775 if (!is_section_foldable_candidate(section_name)) 776 continue; 777 if (!(*p)->is_section_included(i)) 778 continue; 779 if (parameters->options().gc_sections() 780 && symtab->gc()->is_section_garbage(*p, i)) 781 continue; 782 // With --icf=safe, check if the mangled function name is a ctor 783 // or a dtor. The mangled function name can be obtained from the 784 // section name by stripping the section prefix. 785 if (parameters->options().icf_safe_folding() 786 && !is_function_ctor_or_dtor(section_name) 787 && (!target.can_check_for_function_pointers() 788 || section_has_function_pointers(*p, i))) 789 { 790 continue; 791 } 792 this->id_section_.push_back(Section_id(*p, i)); 793 this->section_id_[Section_id(*p, i)] = section_num; 794 this->kept_section_id_.push_back(section_num); 795 num_tracked_relocs.push_back(0); 796 section_addraligns.push_back((*p)->section_addralign(i)); 797 is_secn_or_group_unique.push_back(false); 798 section_contents.push_back(""); 799 section_num++; 800 } 801 } 802 803 unsigned int num_iterations = 0; 804 805 // Default number of iterations to run ICF is 2. 806 unsigned int max_iterations = (parameters->options().icf_iterations() > 0) 807 ? parameters->options().icf_iterations() 808 : 2; 809 810 bool converged = false; 811 812 while (!converged && (num_iterations < max_iterations)) 813 { 814 num_iterations++; 815 converged = match_sections(num_iterations, symtab, 816 &num_tracked_relocs, &this->kept_section_id_, 817 this->id_section_, section_addraligns, 818 &is_secn_or_group_unique, §ion_contents); 819 } 820 821 if (parameters->options().print_icf_sections()) 822 { 823 if (converged) 824 gold_info(_("%s: ICF Converged after %u iteration(s)"), 825 program_name, num_iterations); 826 else 827 gold_info(_("%s: ICF stopped after %u iteration(s)"), 828 program_name, num_iterations); 829 } 830 831 // Unfold --keep-unique symbols. 832 for (options::String_set::const_iterator p = 833 parameters->options().keep_unique_begin(); 834 p != parameters->options().keep_unique_end(); 835 ++p) 836 { 837 const char* name = p->c_str(); 838 Symbol* sym = symtab->lookup(name); 839 if (sym == NULL) 840 { 841 gold_warning(_("Could not find symbol %s to unfold\n"), name); 842 } 843 else if (sym->source() == Symbol::FROM_OBJECT 844 && !sym->object()->is_dynamic()) 845 { 846 Object* obj = sym->object(); 847 bool is_ordinary; 848 unsigned int shndx = sym->shndx(&is_ordinary); 849 if (is_ordinary) 850 { 851 this->unfold_section(obj, shndx); 852 } 853 } 854 855 } 856 857 this->icf_ready(); 858 } 859 860 // Unfolds the section denoted by OBJ and SHNDX if folded. 861 862 void 863 Icf::unfold_section(Object* obj, unsigned int shndx) 864 { 865 Section_id secn(obj, shndx); 866 Uniq_secn_id_map::iterator it = this->section_id_.find(secn); 867 if (it == this->section_id_.end()) 868 return; 869 unsigned int section_num = it->second; 870 unsigned int kept_section_id = this->kept_section_id_[section_num]; 871 if (kept_section_id != section_num) 872 this->kept_section_id_[section_num] = section_num; 873 } 874 875 // This function determines if the section corresponding to the 876 // given object and index is folded based on if the kept section 877 // is different from this section. 878 879 bool 880 Icf::is_section_folded(Object* obj, unsigned int shndx) 881 { 882 Section_id secn(obj, shndx); 883 Uniq_secn_id_map::iterator it = this->section_id_.find(secn); 884 if (it == this->section_id_.end()) 885 return false; 886 unsigned int section_num = it->second; 887 unsigned int kept_section_id = this->kept_section_id_[section_num]; 888 return kept_section_id != section_num; 889 } 890 891 // This function returns the folded section for the given section. 892 893 Section_id 894 Icf::get_folded_section(Object* dup_obj, unsigned int dup_shndx) 895 { 896 Section_id dup_secn(dup_obj, dup_shndx); 897 Uniq_secn_id_map::iterator it = this->section_id_.find(dup_secn); 898 gold_assert(it != this->section_id_.end()); 899 unsigned int section_num = it->second; 900 unsigned int kept_section_id = this->kept_section_id_[section_num]; 901 Section_id folded_section = this->id_section_[kept_section_id]; 902 return folded_section; 903 } 904 905 } // End of namespace gold. 906