| 1 | // icf.cc -- Identical Code Folding. |
| 2 | // |
| 3 | // Copyright (C) 2009-2018 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 | // |
| 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 | Symbol* gsym = *it_s; |
| 331 | bool is_section_symbol = false; |
| 332 | |
| 333 | // A -1 value in the symbol vector indicates a local section symbol. |
| 334 | if (gsym == reinterpret_cast<Symbol*>(-1)) |
| 335 | { |
| 336 | is_section_symbol = true; |
| 337 | gsym = NULL; |
| 338 | } |
| 339 | |
| 340 | if (first_iteration |
| 341 | && it_v->first != NULL) |
| 342 | { |
| 343 | Symbol_location loc; |
| 344 | loc.object = it_v->first; |
| 345 | loc.shndx = it_v->second; |
| 346 | loc.offset = convert_types<off_t, long long>(it_a->first |
| 347 | + it_a->second); |
| 348 | // Look through function descriptors |
| 349 | parameters->target().function_location(&loc); |
| 350 | if (loc.shndx != it_v->second) |
| 351 | { |
| 352 | it_v->second = loc.shndx; |
| 353 | // Modify symvalue/addend to the code entry. |
| 354 | it_a->first = loc.offset; |
| 355 | it_a->second = 0; |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | // ADDEND_STR stores the symbol value and addend and offset, |
| 360 | // each at most 16 hex digits long. it_a points to a pair |
| 361 | // where first is the symbol value and second is the |
| 362 | // addend. |
| 363 | char addend_str[50]; |
| 364 | |
| 365 | // It would be nice if we could use format macros in inttypes.h |
| 366 | // here but there are not in ISO/IEC C++ 1998. |
| 367 | snprintf(addend_str, sizeof(addend_str), "%llx %llx %llx", |
| 368 | static_cast<long long>((*it_a).first), |
| 369 | static_cast<long long>((*it_a).second), |
| 370 | static_cast<unsigned long long>(*it_o)); |
| 371 | |
| 372 | // If the symbol pointed to by the reloc is not in an ordinary |
| 373 | // section or if the symbol type is not FROM_OBJECT, then the |
| 374 | // object is NULL. |
| 375 | if (it_v->first == NULL) |
| 376 | { |
| 377 | if (first_iteration) |
| 378 | { |
| 379 | // If the symbol name is available, use it. |
| 380 | if (gsym != NULL) |
| 381 | buffer.append(gsym->name()); |
| 382 | // Append the addend. |
| 383 | buffer.append(addend_str); |
| 384 | buffer.append("@"); |
| 385 | } |
| 386 | continue; |
| 387 | } |
| 388 | |
| 389 | Section_id reloc_secn(it_v->first, it_v->second); |
| 390 | |
| 391 | // If this reloc turns back and points to the same section, |
| 392 | // like a recursive call, use a special symbol to mark this. |
| 393 | if (reloc_secn.first == secn.first |
| 394 | && reloc_secn.second == secn.second) |
| 395 | { |
| 396 | if (first_iteration) |
| 397 | { |
| 398 | buffer.append("R"); |
| 399 | buffer.append(addend_str); |
| 400 | buffer.append("@"); |
| 401 | } |
| 402 | continue; |
| 403 | } |
| 404 | Icf::Uniq_secn_id_map& section_id_map = |
| 405 | symtab->icf()->section_to_int_map(); |
| 406 | Icf::Uniq_secn_id_map::iterator section_id_map_it = |
| 407 | section_id_map.find(reloc_secn); |
| 408 | bool is_sym_preemptible = (gsym != NULL |
| 409 | && !gsym->is_from_dynobj() |
| 410 | && !gsym->is_undefined() |
| 411 | && gsym->is_preemptible()); |
| 412 | if (!is_sym_preemptible |
| 413 | && section_id_map_it != section_id_map.end()) |
| 414 | { |
| 415 | // This is a reloc to a section that might be folded. |
| 416 | if (num_tracked_relocs) |
| 417 | (*num_tracked_relocs)++; |
| 418 | |
| 419 | char kept_section_str[10]; |
| 420 | unsigned int secn_id = section_id_map_it->second; |
| 421 | snprintf(kept_section_str, sizeof(kept_section_str), "%u", |
| 422 | kept_section_id[secn_id]); |
| 423 | if (first_iteration) |
| 424 | { |
| 425 | buffer.append("ICF_R"); |
| 426 | buffer.append(addend_str); |
| 427 | } |
| 428 | icf_reloc_buffer.append(kept_section_str); |
| 429 | // Append the addend. |
| 430 | icf_reloc_buffer.append(addend_str); |
| 431 | icf_reloc_buffer.append("@"); |
| 432 | } |
| 433 | else |
| 434 | { |
| 435 | // This is a reloc to a section that cannot be folded. |
| 436 | // Process it only in the first iteration. |
| 437 | if (!first_iteration) |
| 438 | continue; |
| 439 | |
| 440 | uint64_t secn_flags = (it_v->first)->section_flags(it_v->second); |
| 441 | // This reloc points to a merge section. Hash the |
| 442 | // contents of this section. |
| 443 | if ((secn_flags & elfcpp::SHF_MERGE) != 0 |
| 444 | && parameters->target().can_icf_inline_merge_sections()) |
| 445 | { |
| 446 | uint64_t entsize = |
| 447 | (it_v->first)->section_entsize(it_v->second); |
| 448 | long long offset = it_a->first; |
| 449 | |
| 450 | // Handle SHT_RELA and SHT_REL addends. Only one of these |
| 451 | // addends exists. When pointing to a merge section, the |
| 452 | // addend only matters if it's relative to a section |
| 453 | // symbol. In order to unambiguously identify the target |
| 454 | // of the relocation, the compiler (and assembler) must use |
| 455 | // a local non-section symbol unless Symbol+Addend does in |
| 456 | // fact point directly to the target. (In other words, |
| 457 | // a bias for a pc-relative reference or a non-zero based |
| 458 | // access forces the use of a local symbol, and the addend |
| 459 | // is used only to provide that bias.) |
| 460 | uint64_t reloc_addend_value = 0; |
| 461 | if (is_section_symbol) |
| 462 | { |
| 463 | // Get the SHT_RELA addend. For RELA relocations, |
| 464 | // we have the addend from the relocation. |
| 465 | reloc_addend_value = it_a->second; |
| 466 | |
| 467 | // Handle SHT_REL addends. |
| 468 | // For REL relocations, we need to fetch the addend |
| 469 | // from the section contents. |
| 470 | const unsigned char* reloc_addend_ptr = |
| 471 | contents + static_cast<unsigned long long>(*it_o); |
| 472 | |
| 473 | // Update the addend value with the SHT_REL addend if |
| 474 | // available. |
| 475 | get_rel_addend(reloc_addend_ptr, *it_addend_size, |
| 476 | &reloc_addend_value); |
| 477 | |
| 478 | // Ignore the addend when it is a negative value. |
| 479 | // See the comments in Merged_symbol_value::value |
| 480 | // in object.h. |
| 481 | if (reloc_addend_value < 0xffffff00) |
| 482 | offset = offset + reloc_addend_value; |
| 483 | } |
| 484 | |
| 485 | section_size_type secn_len; |
| 486 | |
| 487 | const unsigned char* str_contents = |
| 488 | (it_v->first)->section_contents(it_v->second, |
| 489 | &secn_len, |
| 490 | false) + offset; |
| 491 | gold_assert (offset < (long long) secn_len); |
| 492 | |
| 493 | if ((secn_flags & elfcpp::SHF_STRINGS) != 0) |
| 494 | { |
| 495 | // String merge section. |
| 496 | const char* str_char = |
| 497 | reinterpret_cast<const char*>(str_contents); |
| 498 | switch(entsize) |
| 499 | { |
| 500 | case 1: |
| 501 | { |
| 502 | buffer.append(str_char); |
| 503 | break; |
| 504 | } |
| 505 | case 2: |
| 506 | { |
| 507 | const uint16_t* ptr_16 = |
| 508 | reinterpret_cast<const uint16_t*>(str_char); |
| 509 | unsigned int strlen_16 = 0; |
| 510 | // Find the NULL character. |
| 511 | while(*(ptr_16 + strlen_16) != 0) |
| 512 | strlen_16++; |
| 513 | buffer.append(str_char, strlen_16 * 2); |
| 514 | } |
| 515 | break; |
| 516 | case 4: |
| 517 | { |
| 518 | const uint32_t* ptr_32 = |
| 519 | reinterpret_cast<const uint32_t*>(str_char); |
| 520 | unsigned int strlen_32 = 0; |
| 521 | // Find the NULL character. |
| 522 | while(*(ptr_32 + strlen_32) != 0) |
| 523 | strlen_32++; |
| 524 | buffer.append(str_char, strlen_32 * 4); |
| 525 | } |
| 526 | break; |
| 527 | default: |
| 528 | gold_unreachable(); |
| 529 | } |
| 530 | } |
| 531 | else |
| 532 | { |
| 533 | // Use the entsize to determine the length to copy. |
| 534 | uint64_t bufsize = entsize; |
| 535 | // If entsize is too big, copy all the remaining bytes. |
| 536 | if ((offset + entsize) > secn_len) |
| 537 | bufsize = secn_len - offset; |
| 538 | buffer.append(reinterpret_cast<const |
| 539 | char*>(str_contents), |
| 540 | bufsize); |
| 541 | } |
| 542 | buffer.append("@"); |
| 543 | } |
| 544 | else if (gsym != NULL) |
| 545 | { |
| 546 | // If symbol name is available use that. |
| 547 | buffer.append(gsym->name()); |
| 548 | // Append the addend. |
| 549 | buffer.append(addend_str); |
| 550 | buffer.append("@"); |
| 551 | } |
| 552 | else |
| 553 | { |
| 554 | // Symbol name is not available, like for a local symbol, |
| 555 | // use object and section id. |
| 556 | buffer.append(it_v->first->name()); |
| 557 | char secn_id[10]; |
| 558 | snprintf(secn_id, sizeof(secn_id), "%u",it_v->second); |
| 559 | buffer.append(secn_id); |
| 560 | // Append the addend. |
| 561 | buffer.append(addend_str); |
| 562 | buffer.append("@"); |
| 563 | } |
| 564 | } |
| 565 | } |
| 566 | } |
| 567 | |
| 568 | if (first_iteration) |
| 569 | { |
| 570 | buffer.append("Contents = "); |
| 571 | buffer.append(reinterpret_cast<const char*>(contents), plen); |
| 572 | // Store the section contents that don't change to avoid recomputing |
| 573 | // during the next call to this function. |
| 574 | (*section_contents)[section_num] = buffer; |
| 575 | } |
| 576 | else |
| 577 | { |
| 578 | gold_assert(buffer.empty()); |
| 579 | // Reuse the contents computed in the previous iteration. |
| 580 | buffer.append((*section_contents)[section_num]); |
| 581 | } |
| 582 | |
| 583 | buffer.append(icf_reloc_buffer); |
| 584 | return buffer; |
| 585 | } |
| 586 | |
| 587 | // This function computes a checksum on each section to detect and form |
| 588 | // groups of identical sections. The first iteration does this for all |
| 589 | // sections. |
| 590 | // Further iterations do this only for the kept sections from each group to |
| 591 | // determine if larger groups of identical sections could be formed. The |
| 592 | // first section in each group is the kept section for that group. |
| 593 | // |
| 594 | // CRC32 is the checksumming algorithm and can have collisions. That is, |
| 595 | // two sections with different contents can have the same checksum. Hence, |
| 596 | // a multimap is used to maintain more than one group of checksum |
| 597 | // identical sections. A section is added to a group only after its |
| 598 | // contents are explicitly compared with the kept section of the group. |
| 599 | // |
| 600 | // Parameters : |
| 601 | // ITERATION_NUM : Invocation instance of this function. |
| 602 | // NUM_TRACKED_RELOCS : Vector reference to store the number of relocs |
| 603 | // to ICF sections. |
| 604 | // KEPT_SECTION_ID : Vector which maps folded sections to kept sections. |
| 605 | // ID_SECTION : Vector mapping a section to an unique integer. |
| 606 | // IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical |
| 607 | // sections is already known to be unique. |
| 608 | // SECTION_CONTENTS : Store the section's text and relocs to non-ICF |
| 609 | // sections. |
| 610 | |
| 611 | static bool |
| 612 | match_sections(unsigned int iteration_num, |
| 613 | Symbol_table* symtab, |
| 614 | std::vector<unsigned int>* num_tracked_relocs, |
| 615 | std::vector<unsigned int>* kept_section_id, |
| 616 | const std::vector<Section_id>& id_section, |
| 617 | const std::vector<uint64_t>& section_addraligns, |
| 618 | std::vector<bool>* is_secn_or_group_unique, |
| 619 | std::vector<std::string>* section_contents) |
| 620 | { |
| 621 | Unordered_multimap<uint32_t, unsigned int> section_cksum; |
| 622 | std::pair<Unordered_multimap<uint32_t, unsigned int>::iterator, |
| 623 | Unordered_multimap<uint32_t, unsigned int>::iterator> key_range; |
| 624 | bool converged = true; |
| 625 | |
| 626 | if (iteration_num == 1) |
| 627 | preprocess_for_unique_sections(id_section, |
| 628 | is_secn_or_group_unique, |
| 629 | NULL); |
| 630 | else |
| 631 | preprocess_for_unique_sections(id_section, |
| 632 | is_secn_or_group_unique, |
| 633 | section_contents); |
| 634 | |
| 635 | std::vector<std::string> full_section_contents; |
| 636 | |
| 637 | for (unsigned int i = 0; i < id_section.size(); i++) |
| 638 | { |
| 639 | full_section_contents.push_back(""); |
| 640 | if ((*is_secn_or_group_unique)[i]) |
| 641 | continue; |
| 642 | |
| 643 | Section_id secn = id_section[i]; |
| 644 | std::string this_secn_contents; |
| 645 | uint32_t cksum; |
| 646 | if (iteration_num == 1) |
| 647 | { |
| 648 | unsigned int num_relocs = 0; |
| 649 | this_secn_contents = get_section_contents(true, secn, i, &num_relocs, |
| 650 | symtab, (*kept_section_id), |
| 651 | section_contents); |
| 652 | (*num_tracked_relocs)[i] = num_relocs; |
| 653 | } |
| 654 | else |
| 655 | { |
| 656 | if ((*kept_section_id)[i] != i) |
| 657 | { |
| 658 | // This section is already folded into something. |
| 659 | continue; |
| 660 | } |
| 661 | this_secn_contents = get_section_contents(false, secn, i, NULL, |
| 662 | symtab, (*kept_section_id), |
| 663 | section_contents); |
| 664 | } |
| 665 | |
| 666 | const unsigned char* this_secn_contents_array = |
| 667 | reinterpret_cast<const unsigned char*>(this_secn_contents.c_str()); |
| 668 | cksum = xcrc32(this_secn_contents_array, this_secn_contents.length(), |
| 669 | 0xffffffff); |
| 670 | size_t count = section_cksum.count(cksum); |
| 671 | |
| 672 | if (count == 0) |
| 673 | { |
| 674 | // Start a group with this cksum. |
| 675 | section_cksum.insert(std::make_pair(cksum, i)); |
| 676 | full_section_contents[i] = this_secn_contents; |
| 677 | } |
| 678 | else |
| 679 | { |
| 680 | key_range = section_cksum.equal_range(cksum); |
| 681 | Unordered_multimap<uint32_t, unsigned int>::iterator it; |
| 682 | // Search all the groups with this cksum for a match. |
| 683 | for (it = key_range.first; it != key_range.second; ++it) |
| 684 | { |
| 685 | unsigned int kept_section = it->second; |
| 686 | if (full_section_contents[kept_section].length() |
| 687 | != this_secn_contents.length()) |
| 688 | continue; |
| 689 | if (memcmp(full_section_contents[kept_section].c_str(), |
| 690 | this_secn_contents.c_str(), |
| 691 | this_secn_contents.length()) != 0) |
| 692 | continue; |
| 693 | |
| 694 | // Check section alignment here. |
| 695 | // The section with the larger alignment requirement |
| 696 | // should be kept. We assume alignment can only be |
| 697 | // zero or positive integral powers of two. |
| 698 | uint64_t align_i = section_addraligns[i]; |
| 699 | uint64_t align_kept = section_addraligns[kept_section]; |
| 700 | if (align_i <= align_kept) |
| 701 | { |
| 702 | (*kept_section_id)[i] = kept_section; |
| 703 | } |
| 704 | else |
| 705 | { |
| 706 | (*kept_section_id)[kept_section] = i; |
| 707 | it->second = i; |
| 708 | full_section_contents[kept_section].swap( |
| 709 | full_section_contents[i]); |
| 710 | } |
| 711 | |
| 712 | converged = false; |
| 713 | break; |
| 714 | } |
| 715 | if (it == key_range.second) |
| 716 | { |
| 717 | // Create a new group for this cksum. |
| 718 | section_cksum.insert(std::make_pair(cksum, i)); |
| 719 | full_section_contents[i] = this_secn_contents; |
| 720 | } |
| 721 | } |
| 722 | // If there are no relocs to foldable sections do not process |
| 723 | // this section any further. |
| 724 | if (iteration_num == 1 && (*num_tracked_relocs)[i] == 0) |
| 725 | (*is_secn_or_group_unique)[i] = true; |
| 726 | } |
| 727 | |
| 728 | // If a section was folded into another section that was later folded |
| 729 | // again then the former has to be updated. |
| 730 | for (unsigned int i = 0; i < id_section.size(); i++) |
| 731 | { |
| 732 | // Find the end of the folding chain |
| 733 | unsigned int kept = i; |
| 734 | while ((*kept_section_id)[kept] != kept) |
| 735 | { |
| 736 | kept = (*kept_section_id)[kept]; |
| 737 | } |
| 738 | // Update every element of the chain |
| 739 | unsigned int current = i; |
| 740 | while ((*kept_section_id)[current] != kept) |
| 741 | { |
| 742 | unsigned int next = (*kept_section_id)[current]; |
| 743 | (*kept_section_id)[current] = kept; |
| 744 | current = next; |
| 745 | } |
| 746 | } |
| 747 | |
| 748 | return converged; |
| 749 | } |
| 750 | |
| 751 | // During safe icf (--icf=safe), only fold functions that are ctors or dtors. |
| 752 | // This function returns true if the section name is that of a ctor or a dtor. |
| 753 | |
| 754 | static bool |
| 755 | is_function_ctor_or_dtor(const std::string& section_name) |
| 756 | { |
| 757 | const char* mangled_func_name = strrchr(section_name.c_str(), '.'); |
| 758 | gold_assert(mangled_func_name != NULL); |
| 759 | if ((is_prefix_of("._ZN", mangled_func_name) |
| 760 | || is_prefix_of("._ZZ", mangled_func_name)) |
| 761 | && (is_gnu_v3_mangled_ctor(mangled_func_name + 1) |
| 762 | || is_gnu_v3_mangled_dtor(mangled_func_name + 1))) |
| 763 | { |
| 764 | return true; |
| 765 | } |
| 766 | return false; |
| 767 | } |
| 768 | |
| 769 | // This is the main ICF function called in gold.cc. This does the |
| 770 | // initialization and calls match_sections repeatedly (twice by default) |
| 771 | // which computes the crc checksums and detects identical functions. |
| 772 | |
| 773 | void |
| 774 | Icf::find_identical_sections(const Input_objects* input_objects, |
| 775 | Symbol_table* symtab) |
| 776 | { |
| 777 | unsigned int section_num = 0; |
| 778 | std::vector<unsigned int> num_tracked_relocs; |
| 779 | std::vector<uint64_t> section_addraligns; |
| 780 | std::vector<bool> is_secn_or_group_unique; |
| 781 | std::vector<std::string> section_contents; |
| 782 | const Target& target = parameters->target(); |
| 783 | |
| 784 | // Decide which sections are possible candidates first. |
| 785 | |
| 786 | for (Input_objects::Relobj_iterator p = input_objects->relobj_begin(); |
| 787 | p != input_objects->relobj_end(); |
| 788 | ++p) |
| 789 | { |
| 790 | // Lock the object so we can read from it. This is only called |
| 791 | // single-threaded from queue_middle_tasks, so it is OK to lock. |
| 792 | // Unfortunately we have no way to pass in a Task token. |
| 793 | const Task* dummy_task = reinterpret_cast<const Task*>(-1); |
| 794 | Task_lock_obj<Object> tl(dummy_task, *p); |
| 795 | |
| 796 | for (unsigned int i = 0;i < (*p)->shnum(); ++i) |
| 797 | { |
| 798 | const std::string section_name = (*p)->section_name(i); |
| 799 | if (!is_section_foldable_candidate(section_name)) |
| 800 | continue; |
| 801 | if (!(*p)->is_section_included(i)) |
| 802 | continue; |
| 803 | if (parameters->options().gc_sections() |
| 804 | && symtab->gc()->is_section_garbage(*p, i)) |
| 805 | continue; |
| 806 | // With --icf=safe, check if the mangled function name is a ctor |
| 807 | // or a dtor. The mangled function name can be obtained from the |
| 808 | // section name by stripping the section prefix. |
| 809 | if (parameters->options().icf_safe_folding() |
| 810 | && !is_function_ctor_or_dtor(section_name) |
| 811 | && (!target.can_check_for_function_pointers() |
| 812 | || section_has_function_pointers(*p, i))) |
| 813 | { |
| 814 | continue; |
| 815 | } |
| 816 | this->id_section_.push_back(Section_id(*p, i)); |
| 817 | this->section_id_[Section_id(*p, i)] = section_num; |
| 818 | this->kept_section_id_.push_back(section_num); |
| 819 | num_tracked_relocs.push_back(0); |
| 820 | section_addraligns.push_back((*p)->section_addralign(i)); |
| 821 | is_secn_or_group_unique.push_back(false); |
| 822 | section_contents.push_back(""); |
| 823 | section_num++; |
| 824 | } |
| 825 | } |
| 826 | |
| 827 | unsigned int num_iterations = 0; |
| 828 | |
| 829 | // Default number of iterations to run ICF is 2. |
| 830 | unsigned int max_iterations = (parameters->options().icf_iterations() > 0) |
| 831 | ? parameters->options().icf_iterations() |
| 832 | : 2; |
| 833 | |
| 834 | bool converged = false; |
| 835 | |
| 836 | while (!converged && (num_iterations < max_iterations)) |
| 837 | { |
| 838 | num_iterations++; |
| 839 | converged = match_sections(num_iterations, symtab, |
| 840 | &num_tracked_relocs, &this->kept_section_id_, |
| 841 | this->id_section_, section_addraligns, |
| 842 | &is_secn_or_group_unique, §ion_contents); |
| 843 | } |
| 844 | |
| 845 | if (parameters->options().print_icf_sections()) |
| 846 | { |
| 847 | if (converged) |
| 848 | gold_info(_("%s: ICF Converged after %u iteration(s)"), |
| 849 | program_name, num_iterations); |
| 850 | else |
| 851 | gold_info(_("%s: ICF stopped after %u iteration(s)"), |
| 852 | program_name, num_iterations); |
| 853 | } |
| 854 | |
| 855 | // Unfold --keep-unique symbols. |
| 856 | for (options::String_set::const_iterator p = |
| 857 | parameters->options().keep_unique_begin(); |
| 858 | p != parameters->options().keep_unique_end(); |
| 859 | ++p) |
| 860 | { |
| 861 | const char* name = p->c_str(); |
| 862 | Symbol* sym = symtab->lookup(name); |
| 863 | if (sym == NULL) |
| 864 | { |
| 865 | gold_warning(_("Could not find symbol %s to unfold\n"), name); |
| 866 | } |
| 867 | else if (sym->source() == Symbol::FROM_OBJECT |
| 868 | && !sym->object()->is_dynamic()) |
| 869 | { |
| 870 | Relobj* obj = static_cast<Relobj*>(sym->object()); |
| 871 | bool is_ordinary; |
| 872 | unsigned int shndx = sym->shndx(&is_ordinary); |
| 873 | if (is_ordinary) |
| 874 | { |
| 875 | this->unfold_section(obj, shndx); |
| 876 | } |
| 877 | } |
| 878 | |
| 879 | } |
| 880 | |
| 881 | this->icf_ready(); |
| 882 | } |
| 883 | |
| 884 | // Unfolds the section denoted by OBJ and SHNDX if folded. |
| 885 | |
| 886 | void |
| 887 | Icf::unfold_section(Relobj* obj, unsigned int shndx) |
| 888 | { |
| 889 | Section_id secn(obj, shndx); |
| 890 | Uniq_secn_id_map::iterator it = this->section_id_.find(secn); |
| 891 | if (it == this->section_id_.end()) |
| 892 | return; |
| 893 | unsigned int section_num = it->second; |
| 894 | unsigned int kept_section_id = this->kept_section_id_[section_num]; |
| 895 | if (kept_section_id != section_num) |
| 896 | this->kept_section_id_[section_num] = section_num; |
| 897 | } |
| 898 | |
| 899 | // This function determines if the section corresponding to the |
| 900 | // given object and index is folded based on if the kept section |
| 901 | // is different from this section. |
| 902 | |
| 903 | bool |
| 904 | Icf::is_section_folded(Relobj* obj, unsigned int shndx) |
| 905 | { |
| 906 | Section_id secn(obj, shndx); |
| 907 | Uniq_secn_id_map::iterator it = this->section_id_.find(secn); |
| 908 | if (it == this->section_id_.end()) |
| 909 | return false; |
| 910 | unsigned int section_num = it->second; |
| 911 | unsigned int kept_section_id = this->kept_section_id_[section_num]; |
| 912 | return kept_section_id != section_num; |
| 913 | } |
| 914 | |
| 915 | // This function returns the folded section for the given section. |
| 916 | |
| 917 | Section_id |
| 918 | Icf::get_folded_section(Relobj* dup_obj, unsigned int dup_shndx) |
| 919 | { |
| 920 | Section_id dup_secn(dup_obj, dup_shndx); |
| 921 | Uniq_secn_id_map::iterator it = this->section_id_.find(dup_secn); |
| 922 | gold_assert(it != this->section_id_.end()); |
| 923 | unsigned int section_num = it->second; |
| 924 | unsigned int kept_section_id = this->kept_section_id_[section_num]; |
| 925 | Section_id folded_section = this->id_section_[kept_section_id]; |
| 926 | return folded_section; |
| 927 | } |
| 928 | |
| 929 | } // End of namespace gold. |