1 // icf.cc -- Identical Code Folding. 2 // 3 // Copyright (C) 2009-2020 Free Software Foundation, Inc. 4 // Written by Sriraman Tallam <tmsriram@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 // Notes regarding C++ exception handling : 134 // -------------------------------------- 135 // 136 // It is possible for two sections to have identical text, identical 137 // relocations, but different exception handling metadata (unwind 138 // information in the .eh_frame section, and/or handler information in 139 // a .gcc_except_table section). Thus, if a foldable section is 140 // referenced from a .eh_frame FDE, we must include in its checksum 141 // the contents of that FDE as well as of the CIE that the FDE refers 142 // to. The CIE and FDE in turn probably contain relocations to the 143 // personality routine and LSDA, which are handled like any other 144 // relocation for ICF purposes. This logic is helped by the fact that 145 // gcc with -ffunction-sections puts each function's LSDA in its own 146 // .gcc_except_table.<functionname> section. Given sections for two 147 // functions with nontrivial exception handling logic, we will 148 // determine on the first iteration that their .gcc_except_table 149 // sections are identical and can be folded, and on the second 150 // iteration that their .text and .eh_frame contents (including the 151 // now-merged .gcc_except_table relocations for the LSDA) are 152 // identical and can be folded. 153 // 154 // 155 // How to run : --icf=[safe|all|none] 156 // Optional parameters : --icf-iterations <num> --print-icf-sections 157 // 158 // Performance : Less than 20 % link-time overhead on industry strength 159 // applications. Up to 6 % text size reductions. 160 161 #include "gold.h" 162 #include "object.h" 163 #include "gc.h" 164 #include "icf.h" 165 #include "symtab.h" 166 #include "libiberty.h" 167 #include "demangle.h" 168 #include "elfcpp.h" 169 #include "int_encoding.h" 170 171 #include <limits> 172 173 namespace gold 174 { 175 176 // This function determines if a section or a group of identical 177 // sections has unique contents. Such unique sections or groups can be 178 // declared final and need not be processed any further. 179 // Parameters : 180 // ID_SECTION : Vector mapping a section index to a Section_id pair. 181 // IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical 182 // sections is already known to be unique. 183 // SECTION_CONTENTS : Contains the section's text and relocs to sections 184 // that cannot be folded. SECTION_CONTENTS are NULL 185 // implies that this function is being called for the 186 // first time before the first iteration of icf. 187 188 static void 189 preprocess_for_unique_sections(const std::vector<Section_id>& id_section, 190 std::vector<bool>* is_secn_or_group_unique, 191 std::vector<std::string>* section_contents) 192 { 193 Unordered_map<uint32_t, unsigned int> uniq_map; 194 std::pair<Unordered_map<uint32_t, unsigned int>::iterator, bool> 195 uniq_map_insert; 196 197 for (unsigned int i = 0; i < id_section.size(); i++) 198 { 199 if ((*is_secn_or_group_unique)[i]) 200 continue; 201 202 uint32_t cksum; 203 Section_id secn = id_section[i]; 204 section_size_type plen; 205 if (section_contents == NULL) 206 { 207 // Lock the object so we can read from it. This is only called 208 // single-threaded from queue_middle_tasks, so it is OK to lock. 209 // Unfortunately we have no way to pass in a Task token. 210 const Task* dummy_task = reinterpret_cast<const Task*>(-1); 211 Task_lock_obj<Object> tl(dummy_task, secn.first); 212 const unsigned char* contents; 213 contents = secn.first->section_contents(secn.second, 214 &plen, 215 false); 216 cksum = xcrc32(contents, plen, 0xffffffff); 217 } 218 else 219 { 220 const unsigned char* contents_array = reinterpret_cast 221 <const unsigned char*>((*section_contents)[i].c_str()); 222 cksum = xcrc32(contents_array, (*section_contents)[i].length(), 223 0xffffffff); 224 } 225 uniq_map_insert = uniq_map.insert(std::make_pair(cksum, i)); 226 if (uniq_map_insert.second) 227 { 228 (*is_secn_or_group_unique)[i] = true; 229 } 230 else 231 { 232 (*is_secn_or_group_unique)[i] = false; 233 (*is_secn_or_group_unique)[uniq_map_insert.first->second] = false; 234 } 235 } 236 } 237 238 // For SHF_MERGE sections that use REL relocations, the addend is stored in 239 // the text section at the relocation offset. Read the addend value given 240 // the pointer to the addend in the text section and the addend size. 241 // Update the addend value if a valid addend is found. 242 // Parameters: 243 // RELOC_ADDEND_PTR : Pointer to the addend in the text section. 244 // ADDEND_SIZE : The size of the addend. 245 // RELOC_ADDEND_VALUE : Pointer to the addend that is updated. 246 247 inline void 248 get_rel_addend(const unsigned char* reloc_addend_ptr, 249 const unsigned int addend_size, 250 uint64_t* reloc_addend_value) 251 { 252 switch (addend_size) 253 { 254 case 0: 255 break; 256 case 1: 257 *reloc_addend_value = 258 read_from_pointer<8>(reloc_addend_ptr); 259 break; 260 case 2: 261 *reloc_addend_value = 262 read_from_pointer<16>(reloc_addend_ptr); 263 break; 264 case 4: 265 *reloc_addend_value = 266 read_from_pointer<32>(reloc_addend_ptr); 267 break; 268 case 8: 269 *reloc_addend_value = 270 read_from_pointer<64>(reloc_addend_ptr); 271 break; 272 default: 273 gold_unreachable(); 274 } 275 } 276 277 // This returns the buffer containing the section's contents, both 278 // text and relocs. Relocs are differentiated as those pointing to 279 // sections that could be folded and those that cannot. Only relocs 280 // pointing to sections that could be folded are recomputed on 281 // subsequent invocations of this function. 282 // Parameters : 283 // FIRST_ITERATION : true if it is the first invocation. 284 // FIXED_CACHE : String that stores the portion of the result that 285 // does not change from iteration to iteration; 286 // written if first_iteration is true, read if it's false. 287 // SECN : Section for which contents are desired. 288 // SELF_SECN : Relocations that target this section will be 289 // considered "relocations to self" so that recursive 290 // functions can be folded. Should normally be the 291 // same as `secn` except when processing extra identity 292 // regions. 293 // NUM_TRACKED_RELOCS : Vector reference to store the number of relocs 294 // to ICF sections. 295 // KEPT_SECTION_ID : Vector which maps folded sections to kept sections. 296 // START_OFFSET : Only consider the part of the section at and after 297 // this offset. 298 // END_OFFSET : Only consider the part of the section before this 299 // offset. 300 301 static std::string 302 get_section_contents(bool first_iteration, 303 std::string* fixed_cache, 304 const Section_id& secn, 305 const Section_id& self_secn, 306 unsigned int* num_tracked_relocs, 307 Symbol_table* symtab, 308 const std::vector<unsigned int>& kept_section_id, 309 section_offset_type start_offset = 0, 310 section_offset_type end_offset = 311 std::numeric_limits<section_offset_type>::max()) 312 { 313 section_size_type plen; 314 const unsigned char* contents = NULL; 315 if (first_iteration) 316 contents = secn.first->section_contents(secn.second, &plen, false); 317 318 // The buffer to hold all the contents including relocs. A checksum 319 // is then computed on this buffer. 320 std::string buffer; 321 std::string icf_reloc_buffer; 322 323 Icf::Reloc_info_list& reloc_info_list = 324 symtab->icf()->reloc_info_list(); 325 326 Icf::Reloc_info_list::iterator it_reloc_info_list = 327 reloc_info_list.find(secn); 328 329 buffer.clear(); 330 icf_reloc_buffer.clear(); 331 332 // Process relocs and put them into the buffer. 333 334 if (it_reloc_info_list != reloc_info_list.end()) 335 { 336 Icf::Sections_reachable_info &v = 337 (it_reloc_info_list->second).section_info; 338 // Stores the information of the symbol pointed to by the reloc. 339 const Icf::Symbol_info &s = (it_reloc_info_list->second).symbol_info; 340 // Stores the addend and the symbol value. 341 Icf::Addend_info &a = (it_reloc_info_list->second).addend_info; 342 // Stores the offset of the reloc. 343 const Icf::Offset_info &o = (it_reloc_info_list->second).offset_info; 344 const Icf::Reloc_addend_size_info &reloc_addend_size_info = 345 (it_reloc_info_list->second).reloc_addend_size_info; 346 Icf::Sections_reachable_info::iterator it_v = v.begin(); 347 Icf::Symbol_info::const_iterator it_s = s.begin(); 348 Icf::Addend_info::iterator it_a = a.begin(); 349 Icf::Offset_info::const_iterator it_o = o.begin(); 350 Icf::Reloc_addend_size_info::const_iterator it_addend_size = 351 reloc_addend_size_info.begin(); 352 353 for (; it_v != v.end(); ++it_v, ++it_s, ++it_a, ++it_o, ++it_addend_size) 354 { 355 Symbol* gsym = *it_s; 356 bool is_section_symbol = false; 357 358 // Ignore relocations outside the region we were told to look at 359 if (static_cast<section_offset_type>(*it_o) < start_offset 360 || static_cast<section_offset_type>(*it_o) >= end_offset) 361 continue; 362 363 // A -1 value in the symbol vector indicates a local section symbol. 364 if (gsym == reinterpret_cast<Symbol*>(-1)) 365 { 366 is_section_symbol = true; 367 gsym = NULL; 368 } 369 370 if (first_iteration 371 && it_v->first != NULL) 372 { 373 Symbol_location loc; 374 loc.object = it_v->first; 375 loc.shndx = it_v->second; 376 loc.offset = convert_types<off_t, long long>(it_a->first 377 + it_a->second); 378 // Look through function descriptors 379 parameters->target().function_location(&loc); 380 if (loc.shndx != it_v->second) 381 { 382 it_v->second = loc.shndx; 383 // Modify symvalue/addend to the code entry. 384 it_a->first = loc.offset; 385 it_a->second = 0; 386 } 387 } 388 389 // ADDEND_STR stores the symbol value and addend and offset, 390 // each at most 16 hex digits long. it_a points to a pair 391 // where first is the symbol value and second is the 392 // addend. 393 char addend_str[50]; 394 395 // It would be nice if we could use format macros in inttypes.h 396 // here but there are not in ISO/IEC C++ 1998. 397 snprintf(addend_str, sizeof(addend_str), "%llx %llx %llx", 398 static_cast<long long>((*it_a).first), 399 static_cast<long long>((*it_a).second), 400 static_cast<unsigned long long>(*it_o - start_offset)); 401 402 // If the symbol pointed to by the reloc is not in an ordinary 403 // section or if the symbol type is not FROM_OBJECT, then the 404 // object is NULL. 405 if (it_v->first == NULL) 406 { 407 if (first_iteration) 408 { 409 // If the symbol name is available, use it. 410 if (gsym != NULL) 411 buffer.append(gsym->name()); 412 // Append the addend. 413 buffer.append(addend_str); 414 buffer.append("@"); 415 } 416 continue; 417 } 418 419 Section_id reloc_secn(it_v->first, it_v->second); 420 421 // If this reloc turns back and points to the same section, 422 // like a recursive call, use a special symbol to mark this. 423 if (reloc_secn.first == self_secn.first 424 && reloc_secn.second == self_secn.second) 425 { 426 if (first_iteration) 427 { 428 buffer.append("R"); 429 buffer.append(addend_str); 430 buffer.append("@"); 431 } 432 continue; 433 } 434 Icf::Uniq_secn_id_map& section_id_map = 435 symtab->icf()->section_to_int_map(); 436 Icf::Uniq_secn_id_map::iterator section_id_map_it = 437 section_id_map.find(reloc_secn); 438 bool is_sym_preemptible = (gsym != NULL 439 && !gsym->is_from_dynobj() 440 && !gsym->is_undefined() 441 && gsym->is_preemptible()); 442 if (!is_sym_preemptible 443 && section_id_map_it != section_id_map.end()) 444 { 445 // This is a reloc to a section that might be folded. 446 if (num_tracked_relocs) 447 (*num_tracked_relocs)++; 448 449 char kept_section_str[10]; 450 unsigned int secn_id = section_id_map_it->second; 451 snprintf(kept_section_str, sizeof(kept_section_str), "%u", 452 kept_section_id[secn_id]); 453 if (first_iteration) 454 { 455 buffer.append("ICF_R"); 456 buffer.append(addend_str); 457 } 458 icf_reloc_buffer.append(kept_section_str); 459 // Append the addend. 460 icf_reloc_buffer.append(addend_str); 461 icf_reloc_buffer.append("@"); 462 } 463 else 464 { 465 // This is a reloc to a section that cannot be folded. 466 // Process it only in the first iteration. 467 if (!first_iteration) 468 continue; 469 470 uint64_t secn_flags = (it_v->first)->section_flags(it_v->second); 471 // This reloc points to a merge section. Hash the 472 // contents of this section. 473 if ((secn_flags & elfcpp::SHF_MERGE) != 0 474 && parameters->target().can_icf_inline_merge_sections()) 475 { 476 uint64_t entsize = 477 (it_v->first)->section_entsize(it_v->second); 478 long long offset = it_a->first; 479 480 // Handle SHT_RELA and SHT_REL addends. Only one of these 481 // addends exists. When pointing to a merge section, the 482 // addend only matters if it's relative to a section 483 // symbol. In order to unambiguously identify the target 484 // of the relocation, the compiler (and assembler) must use 485 // a local non-section symbol unless Symbol+Addend does in 486 // fact point directly to the target. (In other words, 487 // a bias for a pc-relative reference or a non-zero based 488 // access forces the use of a local symbol, and the addend 489 // is used only to provide that bias.) 490 uint64_t reloc_addend_value = 0; 491 if (is_section_symbol) 492 { 493 // Get the SHT_RELA addend. For RELA relocations, 494 // we have the addend from the relocation. 495 reloc_addend_value = it_a->second; 496 497 // Handle SHT_REL addends. 498 // For REL relocations, we need to fetch the addend 499 // from the section contents. 500 const unsigned char* reloc_addend_ptr = 501 contents + static_cast<unsigned long long>(*it_o); 502 503 // Update the addend value with the SHT_REL addend if 504 // available. 505 get_rel_addend(reloc_addend_ptr, *it_addend_size, 506 &reloc_addend_value); 507 508 // Ignore the addend when it is a negative value. 509 // See the comments in Merged_symbol_value::value 510 // in object.h. 511 if (reloc_addend_value < 0xffffff00) 512 offset = offset + reloc_addend_value; 513 } 514 515 section_size_type secn_len; 516 517 const unsigned char* str_contents = 518 (it_v->first)->section_contents(it_v->second, 519 &secn_len, 520 false) + offset; 521 gold_assert (offset < (long long) secn_len); 522 523 if ((secn_flags & elfcpp::SHF_STRINGS) != 0) 524 { 525 // String merge section. 526 const char* str_char = 527 reinterpret_cast<const char*>(str_contents); 528 switch(entsize) 529 { 530 case 1: 531 { 532 buffer.append(str_char); 533 break; 534 } 535 case 2: 536 { 537 const uint16_t* ptr_16 = 538 reinterpret_cast<const uint16_t*>(str_char); 539 unsigned int strlen_16 = 0; 540 // Find the NULL character. 541 while(*(ptr_16 + strlen_16) != 0) 542 strlen_16++; 543 buffer.append(str_char, strlen_16 * 2); 544 } 545 break; 546 case 4: 547 { 548 const uint32_t* ptr_32 = 549 reinterpret_cast<const uint32_t*>(str_char); 550 unsigned int strlen_32 = 0; 551 // Find the NULL character. 552 while(*(ptr_32 + strlen_32) != 0) 553 strlen_32++; 554 buffer.append(str_char, strlen_32 * 4); 555 } 556 break; 557 default: 558 gold_unreachable(); 559 } 560 } 561 else 562 { 563 // Use the entsize to determine the length to copy. 564 uint64_t bufsize = entsize; 565 // If entsize is too big, copy all the remaining bytes. 566 if ((offset + entsize) > secn_len) 567 bufsize = secn_len - offset; 568 buffer.append(reinterpret_cast<const 569 char*>(str_contents), 570 bufsize); 571 } 572 buffer.append("@"); 573 } 574 else if (gsym != NULL) 575 { 576 // If symbol name is available use that. 577 buffer.append(gsym->name()); 578 // Append the addend. 579 buffer.append(addend_str); 580 buffer.append("@"); 581 } 582 else 583 { 584 // Symbol name is not available, like for a local symbol, 585 // use object and section id. 586 buffer.append(it_v->first->name()); 587 char secn_id[10]; 588 snprintf(secn_id, sizeof(secn_id), "%u",it_v->second); 589 buffer.append(secn_id); 590 // Append the addend. 591 buffer.append(addend_str); 592 buffer.append("@"); 593 } 594 } 595 } 596 } 597 598 if (first_iteration) 599 { 600 buffer.append("Contents = "); 601 602 const unsigned char* slice_end = 603 contents + std::min<section_offset_type>(plen, end_offset); 604 605 if (contents + start_offset < slice_end) 606 { 607 buffer.append(reinterpret_cast<const char*>(contents + start_offset), 608 slice_end - (contents + start_offset)); 609 } 610 } 611 612 // Add any extra identity regions. 613 std::pair<Icf::Extra_identity_list::const_iterator, 614 Icf::Extra_identity_list::const_iterator> 615 extra_range = symtab->icf()->extra_identity_list().equal_range(secn); 616 for (Icf::Extra_identity_list::const_iterator it_ext = extra_range.first; 617 it_ext != extra_range.second; ++it_ext) 618 { 619 std::string external_fixed; 620 std::string external_all = 621 get_section_contents(first_iteration, &external_fixed, 622 it_ext->second.section, self_secn, 623 num_tracked_relocs, symtab, 624 kept_section_id, it_ext->second.offset, 625 it_ext->second.offset + it_ext->second.length); 626 buffer.append(external_fixed); 627 icf_reloc_buffer.append(external_all, external_fixed.length(), 628 std::string::npos); 629 } 630 631 if (first_iteration) 632 { 633 // Store the section contents that don't change to avoid recomputing 634 // during the next call to this function. 635 *fixed_cache = buffer; 636 } 637 else 638 { 639 gold_assert(buffer.empty()); 640 641 // Reuse the contents computed in the previous iteration. 642 buffer.append(*fixed_cache); 643 } 644 645 buffer.append(icf_reloc_buffer); 646 return buffer; 647 } 648 649 // This function computes a checksum on each section to detect and form 650 // groups of identical sections. The first iteration does this for all 651 // sections. 652 // Further iterations do this only for the kept sections from each group to 653 // determine if larger groups of identical sections could be formed. The 654 // first section in each group is the kept section for that group. 655 // 656 // CRC32 is the checksumming algorithm and can have collisions. That is, 657 // two sections with different contents can have the same checksum. Hence, 658 // a multimap is used to maintain more than one group of checksum 659 // identical sections. A section is added to a group only after its 660 // contents are explicitly compared with the kept section of the group. 661 // 662 // Parameters : 663 // ITERATION_NUM : Invocation instance of this function. 664 // NUM_TRACKED_RELOCS : Vector reference to store the number of relocs 665 // to ICF sections. 666 // KEPT_SECTION_ID : Vector which maps folded sections to kept sections. 667 // ID_SECTION : Vector mapping a section to an unique integer. 668 // IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical 669 // sections is already known to be unique. 670 // SECTION_CONTENTS : Store the section's text and relocs to non-ICF 671 // sections. 672 673 static bool 674 match_sections(unsigned int iteration_num, 675 Symbol_table* symtab, 676 std::vector<unsigned int>* num_tracked_relocs, 677 std::vector<unsigned int>* kept_section_id, 678 const std::vector<Section_id>& id_section, 679 const std::vector<uint64_t>& section_addraligns, 680 std::vector<bool>* is_secn_or_group_unique, 681 std::vector<std::string>* section_contents) 682 { 683 Unordered_multimap<uint32_t, unsigned int> section_cksum; 684 std::pair<Unordered_multimap<uint32_t, unsigned int>::iterator, 685 Unordered_multimap<uint32_t, unsigned int>::iterator> key_range; 686 bool converged = true; 687 688 if (iteration_num == 1) 689 preprocess_for_unique_sections(id_section, 690 is_secn_or_group_unique, 691 NULL); 692 else 693 preprocess_for_unique_sections(id_section, 694 is_secn_or_group_unique, 695 section_contents); 696 697 std::vector<std::string> full_section_contents; 698 699 for (unsigned int i = 0; i < id_section.size(); i++) 700 { 701 full_section_contents.push_back(""); 702 if ((*is_secn_or_group_unique)[i]) 703 continue; 704 705 Section_id secn = id_section[i]; 706 707 // Lock the object so we can read from it. This is only called 708 // single-threaded from queue_middle_tasks, so it is OK to lock. 709 // Unfortunately we have no way to pass in a Task token. 710 const Task* dummy_task = reinterpret_cast<const Task*>(-1); 711 Task_lock_obj<Object> tl(dummy_task, secn.first); 712 713 std::string this_secn_contents; 714 uint32_t cksum; 715 std::string* this_secn_cache = &((*section_contents)[i]); 716 if (iteration_num == 1) 717 { 718 unsigned int num_relocs = 0; 719 this_secn_contents = get_section_contents(true, this_secn_cache, 720 secn, secn, &num_relocs, 721 symtab, (*kept_section_id)); 722 (*num_tracked_relocs)[i] = num_relocs; 723 } 724 else 725 { 726 if ((*kept_section_id)[i] != i) 727 { 728 // This section is already folded into something. 729 continue; 730 } 731 this_secn_contents = get_section_contents(false, this_secn_cache, 732 secn, secn, NULL, 733 symtab, (*kept_section_id)); 734 } 735 736 const unsigned char* this_secn_contents_array = 737 reinterpret_cast<const unsigned char*>(this_secn_contents.c_str()); 738 cksum = xcrc32(this_secn_contents_array, this_secn_contents.length(), 739 0xffffffff); 740 size_t count = section_cksum.count(cksum); 741 742 if (count == 0) 743 { 744 // Start a group with this cksum. 745 section_cksum.insert(std::make_pair(cksum, i)); 746 full_section_contents[i] = this_secn_contents; 747 } 748 else 749 { 750 key_range = section_cksum.equal_range(cksum); 751 Unordered_multimap<uint32_t, unsigned int>::iterator it; 752 // Search all the groups with this cksum for a match. 753 for (it = key_range.first; it != key_range.second; ++it) 754 { 755 unsigned int kept_section = it->second; 756 if (full_section_contents[kept_section].length() 757 != this_secn_contents.length()) 758 continue; 759 if (memcmp(full_section_contents[kept_section].c_str(), 760 this_secn_contents.c_str(), 761 this_secn_contents.length()) != 0) 762 continue; 763 764 // Check section alignment here. 765 // The section with the larger alignment requirement 766 // should be kept. We assume alignment can only be 767 // zero or positive integral powers of two. 768 uint64_t align_i = section_addraligns[i]; 769 uint64_t align_kept = section_addraligns[kept_section]; 770 if (align_i <= align_kept) 771 { 772 (*kept_section_id)[i] = kept_section; 773 } 774 else 775 { 776 (*kept_section_id)[kept_section] = i; 777 it->second = i; 778 full_section_contents[kept_section].swap( 779 full_section_contents[i]); 780 } 781 782 converged = false; 783 break; 784 } 785 if (it == key_range.second) 786 { 787 // Create a new group for this cksum. 788 section_cksum.insert(std::make_pair(cksum, i)); 789 full_section_contents[i] = this_secn_contents; 790 } 791 } 792 // If there are no relocs to foldable sections do not process 793 // this section any further. 794 if (iteration_num == 1 && (*num_tracked_relocs)[i] == 0) 795 (*is_secn_or_group_unique)[i] = true; 796 } 797 798 // If a section was folded into another section that was later folded 799 // again then the former has to be updated. 800 for (unsigned int i = 0; i < id_section.size(); i++) 801 { 802 // Find the end of the folding chain 803 unsigned int kept = i; 804 while ((*kept_section_id)[kept] != kept) 805 { 806 kept = (*kept_section_id)[kept]; 807 } 808 // Update every element of the chain 809 unsigned int current = i; 810 while ((*kept_section_id)[current] != kept) 811 { 812 unsigned int next = (*kept_section_id)[current]; 813 (*kept_section_id)[current] = kept; 814 current = next; 815 } 816 } 817 818 return converged; 819 } 820 821 // During safe icf (--icf=safe), only fold functions that are ctors or dtors. 822 // This function returns true if the section name is that of a ctor or a dtor. 823 824 static bool 825 is_function_ctor_or_dtor(const std::string& section_name) 826 { 827 const char* mangled_func_name = strrchr(section_name.c_str(), '.'); 828 gold_assert(mangled_func_name != NULL); 829 if ((is_prefix_of("._ZN", mangled_func_name) 830 || is_prefix_of("._ZZ", mangled_func_name)) 831 && (is_gnu_v3_mangled_ctor(mangled_func_name + 1) 832 || is_gnu_v3_mangled_dtor(mangled_func_name + 1))) 833 { 834 return true; 835 } 836 return false; 837 } 838 839 // Iterate through the .eh_frame section that has index 840 // `ehframe_shndx` in `object`, adding entries to extra_identity_list_ 841 // that will cause the contents of each FDE and its CIE to be included 842 // in the logical ICF identity of the function that the FDE refers to. 843 844 bool 845 Icf::add_ehframe_links(Relobj* object, unsigned int ehframe_shndx, 846 Reloc_info& relocs) 847 { 848 section_size_type contents_len; 849 const unsigned char* pcontents = object->section_contents(ehframe_shndx, 850 &contents_len, 851 false); 852 const unsigned char* p = pcontents; 853 const unsigned char* pend = pcontents + contents_len; 854 855 Sections_reachable_info::iterator it_target = relocs.section_info.begin(); 856 Sections_reachable_info::iterator it_target_end = relocs.section_info.end(); 857 Offset_info::iterator it_offset = relocs.offset_info.begin(); 858 Offset_info::iterator it_offset_end = relocs.offset_info.end(); 859 860 // Maps section offset to the length of the CIE defined at that offset. 861 typedef Unordered_map<section_offset_type, section_size_type> Cie_map; 862 Cie_map cies; 863 864 uint32_t (*read_swap_32)(const unsigned char*); 865 if (object->is_big_endian()) 866 read_swap_32 = &elfcpp::Swap<32, true>::readval; 867 else 868 read_swap_32 = &elfcpp::Swap<32, false>::readval; 869 870 // TODO: The logic for parsing the CIE/FDE framing is copied from 871 // Eh_frame::do_add_ehframe_input_section() and might want to be 872 // factored into a shared helper function. 873 while (p < pend) 874 { 875 if (pend - p < 4) 876 return false; 877 878 unsigned int len = read_swap_32(p); 879 p += 4; 880 if (len == 0) 881 { 882 // We should only find a zero-length entry at the end of the 883 // section. 884 if (p < pend) 885 return false; 886 break; 887 } 888 // We don't support a 64-bit .eh_frame. 889 if (len == 0xffffffff) 890 return false; 891 if (static_cast<unsigned int>(pend - p) < len) 892 return false; 893 894 const unsigned char* const pentend = p + len; 895 896 if (pend - p < 4) 897 return false; 898 899 unsigned int id = read_swap_32(p); 900 p += 4; 901 902 if (id == 0) 903 { 904 // CIE. 905 cies.insert(std::make_pair(p - pcontents, len - 4)); 906 } 907 else 908 { 909 // FDE. 910 Cie_map::const_iterator it; 911 it = cies.find((p - pcontents) - (id - 4)); 912 if (it == cies.end()) 913 return false; 914 915 // Figure out which section this FDE refers into. The word at `p` 916 // is an address, and we expect to see a relocation there. If not, 917 // this FDE isn't ICF-relevant. 918 while (it_offset != it_offset_end 919 && it_target != it_target_end 920 && static_cast<ptrdiff_t>(*it_offset) < (p - pcontents)) 921 { 922 ++it_offset; 923 ++it_target; 924 } 925 if (it_offset != it_offset_end 926 && it_target != it_target_end 927 && static_cast<ptrdiff_t>(*it_offset) == (p - pcontents)) 928 { 929 // Found a reloc. Add this FDE and its CIE as extra identity 930 // info for the section it refers to. 931 Extra_identity_info rec_fde = {Section_id(object, ehframe_shndx), 932 p - pcontents, len - 4}; 933 Extra_identity_info rec_cie = {Section_id(object, ehframe_shndx), 934 it->first, it->second}; 935 extra_identity_list_.insert(std::make_pair(*it_target, rec_fde)); 936 extra_identity_list_.insert(std::make_pair(*it_target, rec_cie)); 937 } 938 } 939 940 p = pentend; 941 } 942 943 return true; 944 } 945 946 // This is the main ICF function called in gold.cc. This does the 947 // initialization and calls match_sections repeatedly (thrice by default) 948 // which computes the crc checksums and detects identical functions. 949 950 void 951 Icf::find_identical_sections(const Input_objects* input_objects, 952 Symbol_table* symtab) 953 { 954 unsigned int section_num = 0; 955 std::vector<unsigned int> num_tracked_relocs; 956 std::vector<uint64_t> section_addraligns; 957 std::vector<bool> is_secn_or_group_unique; 958 std::vector<std::string> section_contents; 959 const Target& target = parameters->target(); 960 961 // Decide which sections are possible candidates first. 962 963 for (Input_objects::Relobj_iterator p = input_objects->relobj_begin(); 964 p != input_objects->relobj_end(); 965 ++p) 966 { 967 // Lock the object so we can read from it. This is only called 968 // single-threaded from queue_middle_tasks, so it is OK to lock. 969 // Unfortunately we have no way to pass in a Task token. 970 const Task* dummy_task = reinterpret_cast<const Task*>(-1); 971 Task_lock_obj<Object> tl(dummy_task, *p); 972 std::vector<unsigned int> eh_frame_ind; 973 974 for (unsigned int i = 0; i < (*p)->shnum(); ++i) 975 { 976 const std::string section_name = (*p)->section_name(i); 977 if (!is_section_foldable_candidate(section_name)) 978 { 979 if (is_prefix_of(".eh_frame", section_name.c_str())) 980 eh_frame_ind.push_back(i); 981 continue; 982 } 983 984 if (!(*p)->is_section_included(i)) 985 continue; 986 if (parameters->options().gc_sections() 987 && symtab->gc()->is_section_garbage(*p, i)) 988 continue; 989 // With --icf=safe, check if the mangled function name is a ctor 990 // or a dtor. The mangled function name can be obtained from the 991 // section name by stripping the section prefix. 992 if (parameters->options().icf_safe_folding() 993 && !is_function_ctor_or_dtor(section_name) 994 && (!target.can_check_for_function_pointers() 995 || section_has_function_pointers(*p, i))) 996 { 997 continue; 998 } 999 this->id_section_.push_back(Section_id(*p, i)); 1000 this->section_id_[Section_id(*p, i)] = section_num; 1001 this->kept_section_id_.push_back(section_num); 1002 num_tracked_relocs.push_back(0); 1003 section_addraligns.push_back((*p)->section_addralign(i)); 1004 is_secn_or_group_unique.push_back(false); 1005 section_contents.push_back(""); 1006 section_num++; 1007 } 1008 1009 for (std::vector<unsigned int>::iterator it_eh_ind = eh_frame_ind.begin(); 1010 it_eh_ind != eh_frame_ind.end(); ++it_eh_ind) 1011 { 1012 // gc_process_relocs() recorded relocations for this 1013 // section even though we can't fold it. We need to 1014 // use those relocations to associate other foldable 1015 // sections with the FDEs and CIEs that are relevant 1016 // to them, so we can avoid merging sections that 1017 // don't have identical exception-handling behavior. 1018 1019 Section_id sect(*p, *it_eh_ind); 1020 Reloc_info_list::iterator it_rel = this->reloc_info_list().find(sect); 1021 if (it_rel != this->reloc_info_list().end()) 1022 { 1023 if (!add_ehframe_links(*p, *it_eh_ind, it_rel->second)) 1024 { 1025 gold_warning(_("could not parse eh_frame section %s(%s); ICF " 1026 "might not preserve exception handling " 1027 "behavior"), 1028 (*p)->name().c_str(), 1029 (*p)->section_name(*it_eh_ind).c_str()); 1030 } 1031 } 1032 } 1033 } 1034 1035 unsigned int num_iterations = 0; 1036 1037 // Default number of iterations to run ICF is 3. 1038 unsigned int max_iterations = (parameters->options().icf_iterations() > 0) 1039 ? parameters->options().icf_iterations() 1040 : 3; 1041 1042 bool converged = false; 1043 1044 while (!converged && (num_iterations < max_iterations)) 1045 { 1046 num_iterations++; 1047 converged = match_sections(num_iterations, symtab, 1048 &num_tracked_relocs, &this->kept_section_id_, 1049 this->id_section_, section_addraligns, 1050 &is_secn_or_group_unique, §ion_contents); 1051 } 1052 1053 if (parameters->options().print_icf_sections()) 1054 { 1055 if (converged) 1056 gold_info(_("%s: ICF Converged after %u iteration(s)"), 1057 program_name, num_iterations); 1058 else 1059 gold_info(_("%s: ICF stopped after %u iteration(s)"), 1060 program_name, num_iterations); 1061 } 1062 1063 // Unfold --keep-unique symbols. 1064 for (options::String_set::const_iterator p = 1065 parameters->options().keep_unique_begin(); 1066 p != parameters->options().keep_unique_end(); 1067 ++p) 1068 { 1069 const char* name = p->c_str(); 1070 Symbol* sym = symtab->lookup(name); 1071 if (sym == NULL) 1072 { 1073 gold_warning(_("Could not find symbol %s to unfold\n"), name); 1074 } 1075 else if (sym->source() == Symbol::FROM_OBJECT 1076 && !sym->object()->is_dynamic()) 1077 { 1078 Relobj* obj = static_cast<Relobj*>(sym->object()); 1079 bool is_ordinary; 1080 unsigned int shndx = sym->shndx(&is_ordinary); 1081 if (is_ordinary) 1082 { 1083 this->unfold_section(obj, shndx); 1084 } 1085 } 1086 1087 } 1088 1089 this->icf_ready(); 1090 } 1091 1092 // Unfolds the section denoted by OBJ and SHNDX if folded. 1093 1094 void 1095 Icf::unfold_section(Relobj* obj, unsigned int shndx) 1096 { 1097 Section_id secn(obj, shndx); 1098 Uniq_secn_id_map::iterator it = this->section_id_.find(secn); 1099 if (it == this->section_id_.end()) 1100 return; 1101 unsigned int section_num = it->second; 1102 unsigned int kept_section_id = this->kept_section_id_[section_num]; 1103 if (kept_section_id != section_num) 1104 this->kept_section_id_[section_num] = section_num; 1105 } 1106 1107 // This function determines if the section corresponding to the 1108 // given object and index is folded based on if the kept section 1109 // is different from this section. 1110 1111 bool 1112 Icf::is_section_folded(Relobj* obj, unsigned int shndx) 1113 { 1114 Section_id secn(obj, shndx); 1115 Uniq_secn_id_map::iterator it = this->section_id_.find(secn); 1116 if (it == this->section_id_.end()) 1117 return false; 1118 unsigned int section_num = it->second; 1119 unsigned int kept_section_id = this->kept_section_id_[section_num]; 1120 return kept_section_id != section_num; 1121 } 1122 1123 // This function returns the folded section for the given section. 1124 1125 Section_id 1126 Icf::get_folded_section(Relobj* dup_obj, unsigned int dup_shndx) 1127 { 1128 Section_id dup_secn(dup_obj, dup_shndx); 1129 Uniq_secn_id_map::iterator it = this->section_id_.find(dup_secn); 1130 gold_assert(it != this->section_id_.end()); 1131 unsigned int section_num = it->second; 1132 unsigned int kept_section_id = this->kept_section_id_[section_num]; 1133 Section_id folded_section = this->id_section_[kept_section_id]; 1134 return folded_section; 1135 } 1136 1137 } // End of namespace gold. 1138